diff options
Diffstat (limited to 'pkg/sentry')
159 files changed, 10620 insertions, 2088 deletions
diff --git a/pkg/sentry/arch/arch.go b/pkg/sentry/arch/arch.go index 498ca4669..81ec98a77 100644 --- a/pkg/sentry/arch/arch.go +++ b/pkg/sentry/arch/arch.go @@ -125,9 +125,9 @@ type Context interface { // SetTLS sets the current TLS pointer. Returns false if value is invalid. SetTLS(value uintptr) bool - // SetRSEQInterruptedIP sets the register that contains the old IP when a - // restartable sequence is interrupted. - SetRSEQInterruptedIP(value uintptr) + // SetOldRSeqInterruptedIP sets the register that contains the old IP + // when an "old rseq" restartable sequence is interrupted. + SetOldRSeqInterruptedIP(value uintptr) // StateData returns a pointer to underlying architecture state. StateData() *State diff --git a/pkg/sentry/arch/arch_amd64.go b/pkg/sentry/arch/arch_amd64.go index 9e7db8b30..2aa08b1a9 100644 --- a/pkg/sentry/arch/arch_amd64.go +++ b/pkg/sentry/arch/arch_amd64.go @@ -174,8 +174,8 @@ func (c *context64) SetTLS(value uintptr) bool { return true } -// SetRSEQInterruptedIP implements Context.SetRSEQInterruptedIP. -func (c *context64) SetRSEQInterruptedIP(value uintptr) { +// SetOldRSeqInterruptedIP implements Context.SetOldRSeqInterruptedIP. +func (c *context64) SetOldRSeqInterruptedIP(value uintptr) { c.Regs.R10 = uint64(value) } @@ -305,7 +305,7 @@ func (c *context64) PtracePeekUser(addr uintptr) (interface{}, error) { buf := binary.Marshal(nil, usermem.ByteOrder, c.ptraceGetRegs()) return c.Native(uintptr(usermem.ByteOrder.Uint64(buf[addr:]))), nil } - // TODO(b/34088053): debug registers + // Note: x86 debug registers are missing. return c.Native(0), nil } @@ -320,6 +320,6 @@ func (c *context64) PtracePokeUser(addr, data uintptr) error { _, err := c.PtraceSetRegs(bytes.NewBuffer(buf)) return err } - // TODO(b/34088053): debug registers + // Note: x86 debug registers are missing. return nil } diff --git a/pkg/sentry/control/pprof.go b/pkg/sentry/control/pprof.go index 1f78d54a2..e1f2fea60 100644 --- a/pkg/sentry/control/pprof.go +++ b/pkg/sentry/control/pprof.go @@ -22,6 +22,7 @@ import ( "sync" "gvisor.dev/gvisor/pkg/fd" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/urpc" ) @@ -56,6 +57,9 @@ type Profile struct { // traceFile is the current execution trace output file. traceFile *fd.FD + + // Kernel is the kernel under profile. + Kernel *kernel.Kernel } // StartCPUProfile is an RPC stub which starts recording the CPU profile in a @@ -147,6 +151,9 @@ func (p *Profile) StartTrace(o *ProfileOpts, _ *struct{}) error { return err } + // Ensure all trace contexts are registered. + p.Kernel.RebuildTraceContexts() + p.traceFile = output return nil } @@ -158,9 +165,15 @@ func (p *Profile) StopTrace(_, _ *struct{}) error { defer p.mu.Unlock() if p.traceFile == nil { - return errors.New("Execution tracing not start") + return errors.New("Execution tracing not started") } + // Similarly to the case above, if tasks have not ended traces, we will + // lose information. Thus we need to rebuild the tasks in order to have + // complete information. This will not lose information if multiple + // traces are overlapping. + p.Kernel.RebuildTraceContexts() + trace.Stop() p.traceFile.Close() p.traceFile = nil diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go index c35faeb4c..ced51c66c 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -268,14 +268,17 @@ func (proc *Proc) Ps(args *PsArgs, out *string) error { } // Process contains information about a single process in a Sandbox. -// TODO(b/117881927): Implement TTY field. type Process struct { UID auth.KUID `json:"uid"` PID kernel.ThreadID `json:"pid"` // Parent PID - PPID kernel.ThreadID `json:"ppid"` + PPID kernel.ThreadID `json:"ppid"` + Threads []kernel.ThreadID `json:"threads"` // Processor utilization C int32 `json:"c"` + // TTY name of the process. Will be of the form "pts/N" if there is a + // TTY, or "?" if there is not. + TTY string `json:"tty"` // Start time STime string `json:"stime"` // CPU time @@ -285,18 +288,19 @@ type Process struct { } // ProcessListToTable prints a table with the following format: -// UID PID PPID C STIME TIME CMD -// 0 1 0 0 14:04 505262ns tail +// UID PID PPID C TTY STIME TIME CMD +// 0 1 0 0 pty/4 14:04 505262ns tail func ProcessListToTable(pl []*Process) string { var buf bytes.Buffer tw := tabwriter.NewWriter(&buf, 10, 1, 3, ' ', 0) - fmt.Fprint(tw, "UID\tPID\tPPID\tC\tSTIME\tTIME\tCMD") + fmt.Fprint(tw, "UID\tPID\tPPID\tC\tTTY\tSTIME\tTIME\tCMD") for _, d := range pl { - fmt.Fprintf(tw, "\n%d\t%d\t%d\t%d\t%s\t%s\t%s", + fmt.Fprintf(tw, "\n%d\t%d\t%d\t%d\t%s\t%s\t%s\t%s", d.UID, d.PID, d.PPID, d.C, + d.TTY, d.STime, d.Time, d.Cmd) @@ -307,7 +311,7 @@ func ProcessListToTable(pl []*Process) string { // ProcessListToJSON will return the JSON representation of ps. func ProcessListToJSON(pl []*Process) (string, error) { - b, err := json.Marshal(pl) + b, err := json.MarshalIndent(pl, "", " ") if err != nil { return "", fmt.Errorf("couldn't marshal process list %v: %v", pl, err) } @@ -334,7 +338,9 @@ func Processes(k *kernel.Kernel, containerID string, out *[]*Process) error { ts := k.TaskSet() now := k.RealtimeClock().Now() for _, tg := range ts.Root.ThreadGroups() { - pid := tg.PIDNamespace().IDOfThreadGroup(tg) + pidns := tg.PIDNamespace() + pid := pidns.IDOfThreadGroup(tg) + // If tg has already been reaped ignore it. if pid == 0 { continue @@ -345,16 +351,19 @@ func Processes(k *kernel.Kernel, containerID string, out *[]*Process) error { ppid := kernel.ThreadID(0) if p := tg.Leader().Parent(); p != nil { - ppid = p.PIDNamespace().IDOfThreadGroup(p.ThreadGroup()) + ppid = pidns.IDOfThreadGroup(p.ThreadGroup()) } + threads := tg.MemberIDs(pidns) *out = append(*out, &Process{ - UID: tg.Leader().Credentials().EffectiveKUID, - PID: pid, - PPID: ppid, - STime: formatStartTime(now, tg.Leader().StartTime()), - C: percentCPU(tg.CPUStats(), tg.Leader().StartTime(), now), - Time: tg.CPUStats().SysTime.String(), - Cmd: tg.Leader().Name(), + UID: tg.Leader().Credentials().EffectiveKUID, + PID: pid, + PPID: ppid, + Threads: threads, + STime: formatStartTime(now, tg.Leader().StartTime()), + C: percentCPU(tg.CPUStats(), tg.Leader().StartTime(), now), + Time: tg.CPUStats().SysTime.String(), + Cmd: tg.Leader().Name(), + TTY: ttyName(tg.TTY()), }) } sort.Slice(*out, func(i, j int) bool { return (*out)[i].PID < (*out)[j].PID }) @@ -395,3 +404,10 @@ func percentCPU(stats usage.CPUStats, startTime, now ktime.Time) int32 { } return int32(percentCPU) } + +func ttyName(tty *kernel.TTY) string { + if tty == nil { + return "?" + } + return fmt.Sprintf("pts/%d", tty.Index) +} diff --git a/pkg/sentry/control/proc_test.go b/pkg/sentry/control/proc_test.go index d8ada2694..0a88459b2 100644 --- a/pkg/sentry/control/proc_test.go +++ b/pkg/sentry/control/proc_test.go @@ -34,7 +34,7 @@ func TestProcessListTable(t *testing.T) { }{ { pl: []*Process{}, - expected: "UID PID PPID C STIME TIME CMD", + expected: "UID PID PPID C TTY STIME TIME CMD", }, { pl: []*Process{ @@ -43,6 +43,7 @@ func TestProcessListTable(t *testing.T) { PID: 0, PPID: 0, C: 0, + TTY: "?", STime: "0", Time: "0", Cmd: "zero", @@ -52,14 +53,15 @@ func TestProcessListTable(t *testing.T) { PID: 1, PPID: 1, C: 1, + TTY: "pts/4", STime: "1", Time: "1", Cmd: "one", }, }, - expected: `UID PID PPID C STIME TIME CMD -0 0 0 0 0 0 zero -1 1 1 1 1 1 one`, + expected: `UID PID PPID C TTY STIME TIME CMD +0 0 0 0 ? 0 0 zero +1 1 1 1 pts/4 1 1 one`, }, } diff --git a/pkg/sentry/fs/file.go b/pkg/sentry/fs/file.go index c0a6e884b..a2f966cb6 100644 --- a/pkg/sentry/fs/file.go +++ b/pkg/sentry/fs/file.go @@ -555,6 +555,10 @@ type lockedWriter struct { // // This applies only to Write, not WriteAt. Offset int64 + + // Err contains the first error encountered while copying. This is + // useful to determine whether Writer or Reader failed during io.Copy. + Err error } // Write implements io.Writer.Write. @@ -590,5 +594,8 @@ func (w *lockedWriter) WriteAt(buf []byte, offset int64) (int, error) { break } } + if w.Err == nil { + w.Err = err + } return written, err } diff --git a/pkg/sentry/fs/fsutil/BUILD b/pkg/sentry/fs/fsutil/BUILD index b2e8d9c77..9ca695a95 100644 --- a/pkg/sentry/fs/fsutil/BUILD +++ b/pkg/sentry/fs/fsutil/BUILD @@ -53,7 +53,7 @@ go_template_instance( "Key": "uint64", "Range": "memmap.MappableRange", "Value": "uint64", - "Functions": "fileRangeSetFunctions", + "Functions": "FileRangeSetFunctions", }, ) diff --git a/pkg/sentry/fs/fsutil/file_range_set.go b/pkg/sentry/fs/fsutil/file_range_set.go index 0a5466b0a..f52d712e3 100644 --- a/pkg/sentry/fs/fsutil/file_range_set.go +++ b/pkg/sentry/fs/fsutil/file_range_set.go @@ -34,25 +34,25 @@ import ( // // type FileRangeSet <generated by go_generics> -// fileRangeSetFunctions implements segment.Functions for FileRangeSet. -type fileRangeSetFunctions struct{} +// FileRangeSetFunctions implements segment.Functions for FileRangeSet. +type FileRangeSetFunctions struct{} // MinKey implements segment.Functions.MinKey. -func (fileRangeSetFunctions) MinKey() uint64 { +func (FileRangeSetFunctions) MinKey() uint64 { return 0 } // MaxKey implements segment.Functions.MaxKey. -func (fileRangeSetFunctions) MaxKey() uint64 { +func (FileRangeSetFunctions) MaxKey() uint64 { return math.MaxUint64 } // ClearValue implements segment.Functions.ClearValue. -func (fileRangeSetFunctions) ClearValue(_ *uint64) { +func (FileRangeSetFunctions) ClearValue(_ *uint64) { } // Merge implements segment.Functions.Merge. -func (fileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ memmap.MappableRange, frstart2 uint64) (uint64, bool) { +func (FileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ memmap.MappableRange, frstart2 uint64) (uint64, bool) { if frstart1+mr1.Length() != frstart2 { return 0, false } @@ -60,7 +60,7 @@ func (fileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ } // Split implements segment.Functions.Split. -func (fileRangeSetFunctions) Split(mr memmap.MappableRange, frstart uint64, split uint64) (uint64, uint64) { +func (FileRangeSetFunctions) Split(mr memmap.MappableRange, frstart uint64, split uint64) (uint64, uint64) { return frstart, frstart + (split - mr.Start) } diff --git a/pkg/sentry/fs/gofer/path.go b/pkg/sentry/fs/gofer/path.go index 8c17603f8..c09f3b71c 100644 --- a/pkg/sentry/fs/gofer/path.go +++ b/pkg/sentry/fs/gofer/path.go @@ -234,6 +234,8 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, if err != nil { return nil, err } + // We're not going to use newFile after return. + defer newFile.close(ctx) // Stabilize the endpoint map while creation is in progress. unlock := i.session().endpoints.lock() @@ -254,7 +256,6 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, // Get the attributes of the file to create inode key. qid, mask, attr, err := getattr(ctx, newFile) if err != nil { - newFile.close(ctx) return nil, err } @@ -270,7 +271,6 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, // cloned and re-opened multiple times after creation. _, unopened, err := i.fileState.file.walk(ctx, []string{name}) if err != nil { - newFile.close(ctx) return nil, err } diff --git a/pkg/sentry/fs/gofer/session.go b/pkg/sentry/fs/gofer/session.go index 0da608548..4e358a46a 100644 --- a/pkg/sentry/fs/gofer/session.go +++ b/pkg/sentry/fs/gofer/session.go @@ -143,9 +143,9 @@ type session struct { // socket files. This allows unix domain sockets to be used with paths that // belong to a gofer. // - // TODO(b/77154739): there are few possible races with someone stat'ing the - // file and another deleting it concurrently, where the file will not be - // reported as socket file. + // TODO(gvisor.dev/issue/1200): there are few possible races with someone + // stat'ing the file and another deleting it concurrently, where the file + // will not be reported as socket file. endpoints *endpointMaps `state:"wait"` } diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go index 2d43dff1d..91e2fde2f 100644 --- a/pkg/sentry/fs/inode.go +++ b/pkg/sentry/fs/inode.go @@ -270,6 +270,14 @@ func (i *Inode) Getxattr(name string) (string, error) { return i.InodeOperations.Getxattr(i, name) } +// Setxattr calls i.InodeOperations.Setxattr with i as the Inode. +func (i *Inode) Setxattr(name, value string) error { + if i.overlay != nil { + return overlaySetxattr(i.overlay, name, value) + } + return i.InodeOperations.Setxattr(i, name, value) +} + // Listxattr calls i.InodeOperations.Listxattr with i as the Inode. func (i *Inode) Listxattr() (map[string]struct{}, error) { if i.overlay != nil { diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go index a09147080..13d11e001 100644 --- a/pkg/sentry/fs/inode_overlay.go +++ b/pkg/sentry/fs/inode_overlay.go @@ -436,7 +436,7 @@ func overlayRename(ctx context.Context, o *overlayEntry, oldParent *Dirent, rena } func overlayBind(ctx context.Context, o *overlayEntry, parent *Dirent, name string, data transport.BoundEndpoint, perm FilePermissions) (*Dirent, error) { - if err := copyUp(ctx, parent); err != nil { + if err := copyUpLockedForRename(ctx, parent); err != nil { return nil, err } @@ -552,6 +552,11 @@ func overlayGetxattr(o *overlayEntry, name string) (string, error) { return s, err } +// TODO(b/146028302): Support setxattr for overlayfs. +func overlaySetxattr(o *overlayEntry, name, value string) error { + return syserror.EOPNOTSUPP +} + func overlayListxattr(o *overlayEntry) (map[string]struct{}, error) { o.copyMu.RLock() defer o.copyMu.RUnlock() diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index f3b63dfc2..bd93f83fa 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -64,7 +64,7 @@ var _ fs.InodeOperations = (*tcpMemInode)(nil) func newTCPMemInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack, dir tcpMemDir) *fs.Inode { tm := &tcpMemInode{ - SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0444), linux.PROC_SUPER_MAGIC), + SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), s: s, dir: dir, } @@ -77,6 +77,11 @@ func newTCPMemInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack, dir return fs.NewInode(ctx, tm, msrc, sattr) } +// Truncate implements fs.InodeOperations.Truncate. +func (tcpMemInode) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + // GetFile implements fs.InodeOperations.GetFile. func (m *tcpMemInode) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { flags.Pread = true @@ -168,14 +173,15 @@ func writeSize(dirType tcpMemDir, s inet.Stack, size inet.TCPBufferSize) error { // +stateify savable type tcpSack struct { + fsutil.SimpleFileInode + stack inet.Stack `state:"wait"` enabled *bool - fsutil.SimpleFileInode } func newTCPSackInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { ts := &tcpSack{ - SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0444), linux.PROC_SUPER_MAGIC), + SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), stack: s, } sattr := fs.StableAttr{ @@ -187,6 +193,11 @@ func newTCPSackInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *f return fs.NewInode(ctx, ts, msrc, sattr) } +// Truncate implements fs.InodeOperations.Truncate. +func (tcpSack) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + // GetFile implements fs.InodeOperations.GetFile. func (s *tcpSack) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { flags.Pread = true diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 87184ec67..9bf4b4527 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -67,29 +67,28 @@ type taskDir struct { var _ fs.InodeOperations = (*taskDir)(nil) // newTaskDir creates a new proc task entry. -func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, showSubtasks bool) *fs.Inode { +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), - // FIXME(b/123511468): create the correct io file for threads. - "io": newIO(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), "ns": newNamespaceDir(t, msrc), "smaps": newSmaps(t, msrc), - "stat": newTaskStat(t, msrc, showSubtasks, p.pidns), + "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), "statm": newStatm(t, msrc), "status": newStatus(t, msrc, p.pidns), "uid_map": newUIDMap(t, msrc), } - if showSubtasks { + if isThreadGroup { contents["task"] = p.newSubtasks(t, msrc) } if len(p.cgroupControllers) > 0 { @@ -605,6 +604,10 @@ func (s *statusData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ( fmt.Fprintf(&buf, "CapEff:\t%016x\n", creds.EffectiveCaps) fmt.Fprintf(&buf, "CapBnd:\t%016x\n", creds.BoundingCaps) fmt.Fprintf(&buf, "Seccomp:\t%d\n", s.t.SeccompMode()) + // We unconditionally report a single NUMA node. See + // pkg/sentry/syscalls/linux/sys_mempolicy.go. + fmt.Fprintf(&buf, "Mems_allowed:\t1\n") + fmt.Fprintf(&buf, "Mems_allowed_list:\t0\n") return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*statusData)(nil)}}, 0 } @@ -619,8 +622,11 @@ type ioData struct { ioUsage } -func newIO(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { - return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t.ThreadGroup()}), msrc, fs.SpecialFile, t) +func newIO(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode { + if isThreadGroup { + return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t.ThreadGroup()}), msrc, fs.SpecialFile, t) + } + return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t}), msrc, fs.SpecialFile, t) } // NeedsUpdate returns whether the generation is old or not. @@ -639,7 +645,7 @@ func (i *ioData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se io.Accumulate(i.IOUsage()) var buf bytes.Buffer - fmt.Fprintf(&buf, "char: %d\n", io.CharsRead) + fmt.Fprintf(&buf, "rchar: %d\n", io.CharsRead) fmt.Fprintf(&buf, "wchar: %d\n", io.CharsWritten) fmt.Fprintf(&buf, "syscr: %d\n", io.ReadSyscalls) fmt.Fprintf(&buf, "syscw: %d\n", io.WriteSyscalls) diff --git a/pkg/sentry/fs/splice.go b/pkg/sentry/fs/splice.go index 311798811..389c330a0 100644 --- a/pkg/sentry/fs/splice.go +++ b/pkg/sentry/fs/splice.go @@ -167,6 +167,11 @@ func Splice(ctx context.Context, dst *File, src *File, opts SpliceOpts) (int64, if !srcPipe && !opts.SrcOffset { atomic.StoreInt64(&src.offset, src.offset+n) } + + // Don't report any errors if we have some progress without data loss. + if w.Err == nil { + err = nil + } } // Drop locks. diff --git a/pkg/sentry/fs/tty/master.go b/pkg/sentry/fs/tty/master.go index bc56be696..6b07f6bf2 100644 --- a/pkg/sentry/fs/tty/master.go +++ b/pkg/sentry/fs/tty/master.go @@ -32,7 +32,6 @@ import ( // +stateify savable type masterInodeOperations struct { fsutil.SimpleFileInode - fsutil.InodeNoopTruncate // d is the containing dir. d *dirInodeOperations @@ -77,6 +76,11 @@ func newMasterInode(ctx context.Context, d *dirInodeOperations, owner fs.FileOwn func (mi *masterInodeOperations) Release(ctx context.Context) { } +// Truncate implements fs.InodeOperations.Truncate. +func (*masterInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + // GetFile implements fs.InodeOperations.GetFile. // // It allocates a new terminal. diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go index 4cbea0367..2a51e6bab 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/slave.go @@ -31,7 +31,6 @@ import ( // +stateify savable type slaveInodeOperations struct { fsutil.SimpleFileInode - fsutil.InodeNoopTruncate // d is the containing dir. d *dirInodeOperations @@ -73,6 +72,11 @@ func (si *slaveInodeOperations) Release(ctx context.Context) { si.t.DecRef() } +// Truncate implements fs.InodeOperations.Truncate. +func (slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + // GetFile implements fs.InodeOperations.GetFile. // // This may race with destruction of the terminal. If the terminal is gone, it diff --git a/pkg/sentry/fs/tty/terminal.go b/pkg/sentry/fs/tty/terminal.go index ff8138820..917f90cc0 100644 --- a/pkg/sentry/fs/tty/terminal.go +++ b/pkg/sentry/fs/tty/terminal.go @@ -53,8 +53,8 @@ func newTerminal(ctx context.Context, d *dirInodeOperations, n uint32) *Terminal d: d, n: n, ld: newLineDiscipline(termios), - masterKTTY: &kernel.TTY{}, - slaveKTTY: &kernel.TTY{}, + masterKTTY: &kernel.TTY{Index: n}, + slaveKTTY: &kernel.TTY{Index: n}, } t.EnableLeakCheck("tty.Terminal") return &t diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index 7ccff8b0d..bc90330bc 100644 --- a/pkg/sentry/fsimpl/ext/BUILD +++ b/pkg/sentry/fsimpl/ext/BUILD @@ -38,6 +38,7 @@ go_library( "//pkg/abi/linux", "//pkg/binary", "//pkg/fd", + "//pkg/fspath", "//pkg/log", "//pkg/sentry/arch", "//pkg/sentry/context", @@ -73,6 +74,7 @@ go_test( deps = [ "//pkg/abi/linux", "//pkg/binary", + "//pkg/fspath", "//pkg/sentry/context", "//pkg/sentry/context/contexttest", "//pkg/sentry/fsimpl/ext/disklayout", diff --git a/pkg/sentry/fsimpl/ext/benchmark/BUILD b/pkg/sentry/fsimpl/ext/benchmark/BUILD index bfc46dfa6..4fc8296ef 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/BUILD +++ b/pkg/sentry/fsimpl/ext/benchmark/BUILD @@ -7,6 +7,7 @@ go_test( size = "small", srcs = ["benchmark_test.go"], deps = [ + "//pkg/fspath", "//pkg/sentry/context", "//pkg/sentry/context/contexttest", "//pkg/sentry/fsimpl/ext", diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go index 94cd74095..a56b03711 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go +++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go @@ -24,6 +24,7 @@ import ( "strings" "testing" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/context/contexttest" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext" @@ -49,7 +50,9 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys // Create VFS. vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}) + vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) mntns, err := vfsObj.NewMountNamespace(ctx, creds, imagePath, "extfs", &vfs.GetFilesystemOptions{InternalData: int(f.Fd())}) if err != nil { f.Close() @@ -81,7 +84,11 @@ func mount(b *testing.B, imagePath string, vfsfs *vfs.VirtualFilesystem, pop *vf ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) - if err := vfsfs.NewMount(ctx, creds, imagePath, pop, "extfs", &vfs.GetFilesystemOptions{InternalData: int(f.Fd())}); err != nil { + if err := vfsfs.MountAt(ctx, creds, imagePath, pop, "extfs", &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: int(f.Fd()), + }, + }); err != nil { b.Fatalf("failed to mount tmpfs submount: %v", err) } return func() { @@ -117,7 +124,7 @@ func BenchmarkVFS2Ext4fsStat(b *testing.B) { stat, err := vfsfs.StatAt(ctx, creds, &vfs.PathOperation{ Root: *root, Start: *root, - Pathname: filePath, + Path: fspath.Parse(filePath), FollowFinalSymlink: true, }, &vfs.StatOptions{}) if err != nil { @@ -146,9 +153,9 @@ func BenchmarkVFS2ExtfsMountStat(b *testing.B) { creds := auth.CredentialsFromContext(ctx) mountPointName := "/1/" pop := vfs.PathOperation{ - Root: *root, - Start: *root, - Pathname: mountPointName, + Root: *root, + Start: *root, + Path: fspath.Parse(mountPointName), } // Save the mount point for later use. @@ -177,7 +184,7 @@ func BenchmarkVFS2ExtfsMountStat(b *testing.B) { stat, err := vfsfs.StatAt(ctx, creds, &vfs.PathOperation{ Root: *root, Start: *root, - Pathname: filePath, + Path: fspath.Parse(filePath), FollowFinalSymlink: true, }, &vfs.StatOptions{}) if err != nil { diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go index 307e4d68c..6c14a1e2d 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -25,6 +25,7 @@ import ( "github.com/google/go-cmp/cmp" "github.com/google/go-cmp/cmp/cmpopts" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/context/contexttest" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" @@ -65,7 +66,9 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys // Create VFS. vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}) + vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) mntns, err := vfsObj.NewMountNamespace(ctx, creds, localImagePath, "extfs", &vfs.GetFilesystemOptions{InternalData: int(f.Fd())}) if err != nil { f.Close() @@ -140,62 +143,61 @@ func TestSeek(t *testing.T) { fd, err := vfsfs.OpenAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: *root, Start: *root, Pathname: test.path}, + &vfs.PathOperation{Root: *root, Start: *root, Path: fspath.Parse(test.path)}, &vfs.OpenOptions{}, ) if err != nil { t.Fatalf("vfsfs.OpenAt failed: %v", err) } - if n, err := fd.Impl().Seek(ctx, 0, linux.SEEK_SET); n != 0 || err != nil { + if n, err := fd.Seek(ctx, 0, linux.SEEK_SET); n != 0 || err != nil { t.Errorf("expected seek position 0, got %d and error %v", n, err) } - stat, err := fd.Impl().Stat(ctx, vfs.StatOptions{}) + stat, err := fd.Stat(ctx, vfs.StatOptions{}) if err != nil { t.Errorf("fd.stat failed for file %s in image %s: %v", test.path, test.image, err) } // We should be able to seek beyond the end of file. size := int64(stat.Size) - if n, err := fd.Impl().Seek(ctx, size, linux.SEEK_SET); n != size || err != nil { + if n, err := fd.Seek(ctx, size, linux.SEEK_SET); n != size || err != nil { t.Errorf("expected seek position %d, got %d and error %v", size, n, err) } // EINVAL should be returned if the resulting offset is negative. - if _, err := fd.Impl().Seek(ctx, -1, linux.SEEK_SET); err != syserror.EINVAL { + if _, err := fd.Seek(ctx, -1, linux.SEEK_SET); err != syserror.EINVAL { t.Errorf("expected error EINVAL but got %v", err) } - if n, err := fd.Impl().Seek(ctx, 3, linux.SEEK_CUR); n != size+3 || err != nil { + if n, err := fd.Seek(ctx, 3, linux.SEEK_CUR); n != size+3 || err != nil { t.Errorf("expected seek position %d, got %d and error %v", size+3, n, err) } // Make sure negative offsets work with SEEK_CUR. - if n, err := fd.Impl().Seek(ctx, -2, linux.SEEK_CUR); n != size+1 || err != nil { + if n, err := fd.Seek(ctx, -2, linux.SEEK_CUR); n != size+1 || err != nil { t.Errorf("expected seek position %d, got %d and error %v", size+1, n, err) } // EINVAL should be returned if the resulting offset is negative. - if _, err := fd.Impl().Seek(ctx, -(size + 2), linux.SEEK_CUR); err != syserror.EINVAL { + if _, err := fd.Seek(ctx, -(size + 2), linux.SEEK_CUR); err != syserror.EINVAL { t.Errorf("expected error EINVAL but got %v", err) } // Make sure SEEK_END works with regular files. - switch fd.Impl().(type) { - case *regularFileFD: + if _, ok := fd.Impl().(*regularFileFD); ok { // Seek back to 0. - if n, err := fd.Impl().Seek(ctx, -size, linux.SEEK_END); n != 0 || err != nil { + if n, err := fd.Seek(ctx, -size, linux.SEEK_END); n != 0 || err != nil { t.Errorf("expected seek position %d, got %d and error %v", 0, n, err) } // Seek forward beyond EOF. - if n, err := fd.Impl().Seek(ctx, 1, linux.SEEK_END); n != size+1 || err != nil { + if n, err := fd.Seek(ctx, 1, linux.SEEK_END); n != size+1 || err != nil { t.Errorf("expected seek position %d, got %d and error %v", size+1, n, err) } // EINVAL should be returned if the resulting offset is negative. - if _, err := fd.Impl().Seek(ctx, -(size + 1), linux.SEEK_END); err != syserror.EINVAL { + if _, err := fd.Seek(ctx, -(size + 1), linux.SEEK_END); err != syserror.EINVAL { t.Errorf("expected error EINVAL but got %v", err) } } @@ -360,7 +362,7 @@ func TestStatAt(t *testing.T) { got, err := vfsfs.StatAt(ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: *root, Start: *root, Pathname: test.path}, + &vfs.PathOperation{Root: *root, Start: *root, Path: fspath.Parse(test.path)}, &vfs.StatOptions{}, ) if err != nil { @@ -430,7 +432,7 @@ func TestRead(t *testing.T) { fd, err := vfsfs.OpenAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: *root, Start: *root, Pathname: test.absPath}, + &vfs.PathOperation{Root: *root, Start: *root, Path: fspath.Parse(test.absPath)}, &vfs.OpenOptions{}, ) if err != nil { @@ -456,7 +458,7 @@ func TestRead(t *testing.T) { want := make([]byte, 1) for { n, err := f.Read(want) - fd.Impl().Read(ctx, usermem.BytesIOSequence(got), vfs.ReadOptions{}) + fd.Read(ctx, usermem.BytesIOSequence(got), vfs.ReadOptions{}) if diff := cmp.Diff(got, want); diff != "" { t.Errorf("file data mismatch (-want +got):\n%s", diff) @@ -464,7 +466,7 @@ func TestRead(t *testing.T) { // Make sure there is no more file data left after getting EOF. if n == 0 || err == io.EOF { - if n, _ := fd.Impl().Read(ctx, usermem.BytesIOSequence(got), vfs.ReadOptions{}); n != 0 { + if n, _ := fd.Read(ctx, usermem.BytesIOSequence(got), vfs.ReadOptions{}); n != 0 { t.Errorf("extra unexpected file data in file %s in image %s", test.absPath, test.image) } @@ -566,7 +568,7 @@ func TestIterDirents(t *testing.T) { fd, err := vfsfs.OpenAt( ctx, auth.CredentialsFromContext(ctx), - &vfs.PathOperation{Root: *root, Start: *root, Pathname: test.path}, + &vfs.PathOperation{Root: *root, Start: *root, Path: fspath.Parse(test.path)}, &vfs.OpenOptions{}, ) if err != nil { @@ -574,7 +576,7 @@ func TestIterDirents(t *testing.T) { } cb := &iterDirentsCb{} - if err = fd.Impl().IterDirents(ctx, cb); err != nil { + if err = fd.IterDirents(ctx, cb); err != nil { t.Fatalf("dir fd.IterDirents() failed: %v", err) } diff --git a/pkg/sentry/fsimpl/ext/file_description.go b/pkg/sentry/fsimpl/ext/file_description.go index 5eca2b83f..841274daf 100644 --- a/pkg/sentry/fsimpl/ext/file_description.go +++ b/pkg/sentry/fsimpl/ext/file_description.go @@ -26,13 +26,6 @@ import ( type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl - - // flags is the same as vfs.OpenOptions.Flags which are passed to - // vfs.FilesystemImpl.OpenAt. - // TODO(b/134676337): syscalls like read(2), write(2), fchmod(2), fchown(2), - // fgetxattr(2), ioctl(2), mmap(2) should fail with EBADF if O_PATH is set. - // Only close(2), fstat(2), fstatfs(2) should work. - flags uint32 } func (fd *fileDescription) filesystem() *filesystem { @@ -43,18 +36,6 @@ func (fd *fileDescription) inode() *inode { return fd.vfsfd.Dentry().Impl().(*dentry).inode } -// StatusFlags implements vfs.FileDescriptionImpl.StatusFlags. -func (fd *fileDescription) StatusFlags(ctx context.Context) (uint32, error) { - return fd.flags, nil -} - -// SetStatusFlags implements vfs.FileDescriptionImpl.SetStatusFlags. -func (fd *fileDescription) SetStatusFlags(ctx context.Context, flags uint32) error { - // None of the flags settable by fcntl(F_SETFL) are supported, so this is a - // no-op. - return nil -} - // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index 2d15e8aaf..616fc002a 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -20,6 +20,7 @@ import ( "sync" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -274,6 +275,16 @@ func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op return vfsd, nil } +// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt. +func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { + vfsd, inode, err := fs.walk(rp, true) + if err != nil { + return nil, err + } + inode.incRef() + return vfsd, nil +} + // OpenAt implements vfs.FilesystemImpl.OpenAt. func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { vfsd, inode, err := fs.walk(rp, false) @@ -377,7 +388,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v } // RenameAt implements vfs.FilesystemImpl.RenameAt. -func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry, opts vfs.RenameOptions) error { +func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error { if rp.Done() { return syserror.ENOENT } @@ -441,3 +452,46 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + _, _, err := fs.walk(rp, false) + if err != nil { + return nil, err + } + return nil, syserror.ENOTSUP +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + _, _, err := fs.walk(rp, false) + if err != nil { + return "", err + } + return "", syserror.ENOTSUP +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + _, _, err := fs.walk(rp, false) + if err != nil { + return err + } + return syserror.ENOTSUP +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + _, _, err := fs.walk(rp, false) + if err != nil { + return err + } + return syserror.ENOTSUP +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/ext/inode.go b/pkg/sentry/fsimpl/ext/inode.go index 24249525c..8608805bf 100644 --- a/pkg/sentry/fsimpl/ext/inode.go +++ b/pkg/sentry/fsimpl/ext/inode.go @@ -157,10 +157,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v switch in.impl.(type) { case *regularFile: var fd regularFileFD - fd.flags = flags - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) + fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) return &fd.vfsfd, nil case *directory: // Can't open directories writably. This check is not necessary for a read @@ -169,10 +166,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v return nil, syserror.EISDIR } var fd directoryFD - fd.flags = flags - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) + fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) return &fd.vfsfd, nil case *symlink: if flags&linux.O_PATH == 0 { @@ -180,10 +174,7 @@ func (in *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*v return nil, syserror.ELOOP } var fd symlinkFD - fd.flags = flags - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) + fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) return &fd.vfsfd, nil default: panic(fmt.Sprintf("unknown inode type: %T", in.impl)) diff --git a/pkg/sentry/fsimpl/kernfs/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD new file mode 100644 index 000000000..39c03ee9d --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -0,0 +1,62 @@ +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +package(licenses = ["notice"]) + +go_template_instance( + name = "slot_list", + out = "slot_list.go", + package = "kernfs", + prefix = "slot", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*slot", + "Linker": "*slot", + }, +) + +go_library( + name = "kernfs", + srcs = [ + "dynamic_bytes_file.go", + "fd_impl_util.go", + "filesystem.go", + "inode_impl_util.go", + "kernfs.go", + "slot_list.go", + "symlink.go", + ], + importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs", + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/fspath", + "//pkg/log", + "//pkg/refs", + "//pkg/sentry/context", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/usermem", + "//pkg/sentry/vfs", + "//pkg/syserror", + ], +) + +go_test( + name = "kernfs_test", + size = "small", + srcs = ["kernfs_test.go"], + deps = [ + ":kernfs", + "//pkg/abi/linux", + "//pkg/fspath", + "//pkg/sentry/context", + "//pkg/sentry/context/contexttest", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/usermem", + "//pkg/sentry/vfs", + "//pkg/syserror", + "@com_github_google_go-cmp//cmp:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go new file mode 100644 index 000000000..606ca692d --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -0,0 +1,127 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// DynamicBytesFile implements kernfs.Inode and represents a read-only +// file whose contents are backed by a vfs.DynamicBytesSource. +// +// Must be instantiated with NewDynamicBytesFile or initialized with Init +// before first use. +// +// +stateify savable +type DynamicBytesFile struct { + InodeAttrs + InodeNoopRefCount + InodeNotDirectory + InodeNotSymlink + + data vfs.DynamicBytesSource +} + +var _ Inode = (*DynamicBytesFile)(nil) + +// Init initializes a dynamic bytes file. +func (f *DynamicBytesFile) Init(creds *auth.Credentials, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) { + if perm&^linux.PermissionsMask != 0 { + panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) + } + f.InodeAttrs.Init(creds, ino, linux.ModeRegular|perm) + f.data = data +} + +// Open implements Inode.Open. +func (f *DynamicBytesFile) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + fd := &DynamicBytesFD{} + fd.Init(rp.Mount(), vfsd, f.data, flags) + return &fd.vfsfd, nil +} + +// SetStat implements Inode.SetStat. +func (f *DynamicBytesFile) SetStat(*vfs.Filesystem, vfs.SetStatOptions) error { + // DynamicBytesFiles are immutable. + return syserror.EPERM +} + +// DynamicBytesFD implements vfs.FileDescriptionImpl for an FD backed by a +// DynamicBytesFile. +// +// Must be initialized with Init before first use. +// +// +stateify savable +type DynamicBytesFD struct { + vfs.FileDescriptionDefaultImpl + vfs.DynamicBytesFileDescriptionImpl + + vfsfd vfs.FileDescription + inode Inode +} + +// Init initializes a DynamicBytesFD. +func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *vfs.Dentry, data vfs.DynamicBytesSource, flags uint32) { + fd.inode = d.Impl().(*Dentry).inode + fd.SetDataSource(data) + fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}) +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *DynamicBytesFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + return fd.DynamicBytesFileDescriptionImpl.Seek(ctx, offset, whence) +} + +// Read implmenets vfs.FileDescriptionImpl.Read. +func (fd *DynamicBytesFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return fd.DynamicBytesFileDescriptionImpl.Read(ctx, dst, opts) +} + +// PRead implmenets vfs.FileDescriptionImpl.PRead. +func (fd *DynamicBytesFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return fd.DynamicBytesFileDescriptionImpl.PRead(ctx, dst, offset, opts) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *DynamicBytesFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return fd.FileDescriptionDefaultImpl.Write(ctx, src, opts) +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *DynamicBytesFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return fd.FileDescriptionDefaultImpl.PWrite(ctx, src, offset, opts) +} + +// Release implements vfs.FileDescriptionImpl.Release. +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 +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *DynamicBytesFD) SetStat(context.Context, vfs.SetStatOptions) error { + // DynamicBytesFiles are immutable. + return syserror.EPERM +} diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go new file mode 100644 index 000000000..bcf069b5f --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -0,0 +1,194 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// GenericDirectoryFD implements vfs.FileDescriptionImpl for a generic directory +// inode that uses OrderChildren to track child nodes. GenericDirectoryFD is not +// compatible with dynamic directories. +// +// Note that GenericDirectoryFD holds a lock over OrderedChildren while calling +// IterDirents callback. The IterDirents callback therefore cannot hash or +// unhash children, or recursively call IterDirents on the same underlying +// inode. +// +// Must be initialize with Init before first use. +type GenericDirectoryFD struct { + vfs.FileDescriptionDefaultImpl + vfs.DirectoryFileDescriptionDefaultImpl + + vfsfd vfs.FileDescription + children *OrderedChildren + off int64 +} + +// Init initializes a GenericDirectoryFD. +func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, flags uint32) { + fd.children = children + fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}) +} + +// VFSFileDescription returns a pointer to the vfs.FileDescription representing +// this object. +func (fd *GenericDirectoryFD) VFSFileDescription() *vfs.FileDescription { + return &fd.vfsfd +} + +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *GenericDirectoryFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return fd.FileDescriptionDefaultImpl.ConfigureMMap(ctx, opts) +} + +// Read implmenets vfs.FileDescriptionImpl.Read. +func (fd *GenericDirectoryFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return fd.DirectoryFileDescriptionDefaultImpl.Read(ctx, dst, opts) +} + +// PRead implmenets vfs.FileDescriptionImpl.PRead. +func (fd *GenericDirectoryFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return fd.DirectoryFileDescriptionDefaultImpl.PRead(ctx, dst, offset, opts) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *GenericDirectoryFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return fd.DirectoryFileDescriptionDefaultImpl.Write(ctx, src, opts) +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *GenericDirectoryFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return fd.DirectoryFileDescriptionDefaultImpl.PWrite(ctx, src, offset, opts) +} + +// Release implements vfs.FileDecriptionImpl.Release. +func (fd *GenericDirectoryFD) Release() {} + +func (fd *GenericDirectoryFD) filesystem() *vfs.Filesystem { + return fd.vfsfd.VirtualDentry().Mount().Filesystem() +} + +func (fd *GenericDirectoryFD) inode() Inode { + return fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode +} + +// IterDirents implements vfs.FileDecriptionImpl.IterDirents. IterDirents holds +// o.mu when calling cb. +func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + vfsFS := fd.filesystem() + fs := vfsFS.Impl().(*Filesystem) + vfsd := fd.vfsfd.VirtualDentry().Dentry() + + fs.mu.Lock() + defer fs.mu.Unlock() + + // Handle ".". + if fd.off == 0 { + stat := fd.inode().Stat(vfsFS) + dirent := vfs.Dirent{ + Name: ".", + Type: linux.DT_DIR, + Ino: stat.Ino, + NextOff: 1, + } + if !cb.Handle(dirent) { + return nil + } + fd.off++ + } + + // Handle "..". + if fd.off == 1 { + parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode + stat := parentInode.Stat(vfsFS) + dirent := vfs.Dirent{ + Name: "..", + Type: linux.FileMode(stat.Mode).DirentType(), + Ino: stat.Ino, + NextOff: 2, + } + if !cb.Handle(dirent) { + return nil + } + fd.off++ + } + + // Handle static children. + fd.children.mu.RLock() + defer fd.children.mu.RUnlock() + // fd.off accounts for "." and "..", but fd.children do not track + // these. + 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) + dirent := vfs.Dirent{ + Name: it.Name, + Type: linux.FileMode(stat.Mode).DirentType(), + Ino: stat.Ino, + NextOff: fd.off + 1, + } + if !cb.Handle(dirent) { + return nil + } + fd.off++ + } + + var err error + relOffset := fd.off - int64(len(fd.children.set)) - 2 + fd.off, err = fd.inode().IterDirents(ctx, cb, fd.off, relOffset) + return err +} + +// Seek implements vfs.FileDecriptionImpl.Seek. +func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fs := fd.filesystem().Impl().(*Filesystem) + fs.mu.Lock() + defer fs.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + // Use offset as given. + case linux.SEEK_CUR: + offset += fd.off + default: + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := fd.filesystem() + inode := fd.inode() + return inode.Stat(fs), nil +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *GenericDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + fs := fd.filesystem() + inode := fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode + return inode.SetStat(fs, opts) +} diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go new file mode 100644 index 000000000..79759e0fc --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -0,0 +1,793 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// This file implements vfs.FilesystemImpl for kernfs. + +package kernfs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// stepExistingLocked resolves rp.Component() in parent directory vfsd. +// +// stepExistingLocked is loosely analogous to fs/namei.c:walk_component(). +// +// Preconditions: Filesystem.mu must be locked for at least reading. !rp.Done(). +// +// Postcondition: Caller must call fs.processDeferredDecRefs*. +func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) (*vfs.Dentry, error) { + d := vfsd.Impl().(*Dentry) + if !d.isDir() { + return nil, syserror.ENOTDIR + } + // Directory searchable? + if err := d.inode.CheckPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } +afterSymlink: + name := rp.Component() + // Revalidation must be skipped if name is "." or ".."; d or its parent + // respectively can't be expected to transition from invalidated back to + // valid, so detecting invalidation and retrying would loop forever. This + // is consistent with Linux: fs/namei.c:walk_component() => lookup_fast() + // calls d_revalidate(), but walk_component() => handle_dots() does not. + if name == "." { + rp.Advance() + return vfsd, nil + } + if name == ".." { + nextVFSD, err := rp.ResolveParent(vfsd) + if err != nil { + return nil, err + } + rp.Advance() + return nextVFSD, nil + } + d.dirMu.Lock() + nextVFSD, err := rp.ResolveChild(vfsd, name) + if err != nil { + d.dirMu.Unlock() + return nil, err + } + next, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, nextVFSD) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + // Resolve any symlink at current path component. + if rp.ShouldFollowSymlink() && next.isSymlink() { + // TODO: VFS2 needs something extra for /proc/[pid]/fd/ "magic symlinks". + target, err := next.inode.Readlink(ctx) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + goto afterSymlink + + } + rp.Advance() + return &next.vfsd, nil +} + +// revalidateChildLocked must be called after a call to parent.vfsd.Child(name) +// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be +// nil) to verify that the returned child (or lack thereof) is correct. +// +// Preconditions: Filesystem.mu must be locked for at least reading. +// parent.dirMu must be locked. parent.isDir(). name is not "." or "..". +// +// Postconditions: Caller must call fs.processDeferredDecRefs*. +func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, childVFSD *vfs.Dentry) (*Dentry, error) { + if childVFSD != nil { + // Cached dentry exists, revalidate. + child := childVFSD.Impl().(*Dentry) + if !child.inode.Valid(ctx) { + vfsObj.ForceDeleteDentry(childVFSD) + fs.deferDecRef(childVFSD) // Reference from Lookup. + childVFSD = nil + } + } + if childVFSD == nil { + // Dentry isn't cached; it either doesn't exist or failed + // revalidation. Attempt to resolve it via Lookup. + // + // FIXME(b/144498111): Inode.Lookup() should return *(kernfs.)Dentry, + // not *vfs.Dentry, since (kernfs.)Filesystem assumes that all dentries + // in the filesystem are (kernfs.)Dentry and performs vfs.DentryImpl + // casts accordingly. + var err error + childVFSD, err = parent.inode.Lookup(ctx, name) + if err != nil { + return nil, err + } + // Reference on childVFSD dropped by a corresponding Valid. + parent.insertChildLocked(name, childVFSD) + } + return childVFSD.Impl().(*Dentry), nil +} + +// walkExistingLocked resolves rp to an existing file. +// +// walkExistingLocked is loosely analogous to Linux's +// fs/namei.c:path_lookupat(). +// +// Preconditions: Filesystem.mu must be locked for at least reading. +// +// Postconditions: Caller must call fs.processDeferredDecRefs*. +func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { + vfsd := rp.Start() + for !rp.Done() { + var err error + vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd) + if err != nil { + return nil, nil, err + } + } + d := vfsd.Impl().(*Dentry) + if rp.MustBeDir() && !d.isDir() { + return nil, nil, syserror.ENOTDIR + } + return vfsd, d.inode, nil +} + +// walkParentDirLocked resolves all but the last path component of rp to an +// existing directory. It does not check that the returned directory is +// searchable by the provider of rp. +// +// walkParentDirLocked is loosely analogous to Linux's +// fs/namei.c:path_parentat(). +// +// Preconditions: Filesystem.mu must be locked for at least reading. !rp.Done(). +// +// Postconditions: Caller must call fs.processDeferredDecRefs*. +func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { + vfsd := rp.Start() + for !rp.Final() { + var err error + vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd) + if err != nil { + return nil, nil, err + } + } + d := vfsd.Impl().(*Dentry) + if !d.isDir() { + return nil, nil, syserror.ENOTDIR + } + return vfsd, d.inode, nil +} + +// checkCreateLocked checks that a file named rp.Component() may be created in +// directory parentVFSD, then returns rp.Component(). +// +// Preconditions: Filesystem.mu must be locked for at least reading. parentInode +// == parentVFSD.Impl().(*Dentry).Inode. isDir(parentInode) == true. +func checkCreateLocked(rp *vfs.ResolvingPath, parentVFSD *vfs.Dentry, parentInode Inode) (string, error) { + if err := parentInode.CheckPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { + return "", err + } + pc := rp.Component() + if pc == "." || pc == ".." { + return "", syserror.EEXIST + } + childVFSD, err := rp.ResolveChild(parentVFSD, pc) + if err != nil { + return "", err + } + if childVFSD != nil { + return "", syserror.EEXIST + } + if parentVFSD.IsDisowned() { + return "", syserror.ENOENT + } + return pc, nil +} + +// checkDeleteLocked checks that the file represented by vfsd may be deleted. +// +// Preconditions: Filesystem.mu must be locked for at least reading. +func checkDeleteLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { + parentVFSD := vfsd.Parent() + if parentVFSD == nil { + return syserror.EBUSY + } + if parentVFSD.IsDisowned() { + return syserror.ENOENT + } + if err := parentVFSD.Impl().(*Dentry).inode.CheckPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + return nil +} + +// checkRenameLocked checks that a rename operation may be performed on the +// target dentry across the given set of parent directories. The target dentry +// may be nil. +// +// Precondition: isDir(dstInode) == true. +func checkRenameLocked(creds *auth.Credentials, src, dstDir *vfs.Dentry, dstInode Inode) error { + srcDir := src.Parent() + if srcDir == nil { + return syserror.EBUSY + } + if srcDir.IsDisowned() { + return syserror.ENOENT + } + if dstDir.IsDisowned() { + return syserror.ENOENT + } + // Check for creation permissions on dst dir. + if err := dstInode.CheckPermissions(creds, vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + + return nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *Filesystem) Release() { +} + +// Sync implements vfs.FilesystemImpl.Sync. +func (fs *Filesystem) Sync(ctx context.Context) error { + // All filesystem state is in-memory. + return nil +} + +// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. +func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { + fs.mu.RLock() + defer fs.processDeferredDecRefs() + defer fs.mu.RUnlock() + vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + if err != nil { + return nil, err + } + + if opts.CheckSearchable { + d := vfsd.Impl().(*Dentry) + if !d.isDir() { + return nil, syserror.ENOTDIR + } + if err := inode.CheckPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } + } + vfsd.IncRef() // Ownership transferred to caller. + return vfsd, nil +} + +// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt. +func (fs *Filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { + fs.mu.RLock() + defer fs.processDeferredDecRefs() + defer fs.mu.RUnlock() + vfsd, _, err := fs.walkParentDirLocked(ctx, rp) + if err != nil { + return nil, err + } + vfsd.IncRef() // Ownership transferred to caller. + return vfsd, nil +} + +// LinkAt implements vfs.FilesystemImpl.LinkAt. +func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { + if rp.Done() { + return syserror.EEXIST + } + fs.mu.Lock() + defer fs.mu.Unlock() + parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + pc, err := checkCreateLocked(rp, parentVFSD, parentInode) + if err != nil { + return err + } + if rp.Mount() != vd.Mount() { + return syserror.EXDEV + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + + d := vd.Dentry().Impl().(*Dentry) + if d.isDir() { + return syserror.EPERM + } + + child, err := parentInode.NewLink(ctx, pc, d.inode) + if err != nil { + return err + } + parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + return nil +} + +// MkdirAt implements vfs.FilesystemImpl.MkdirAt. +func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + if rp.Done() { + return syserror.EEXIST + } + fs.mu.Lock() + defer fs.mu.Unlock() + parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + pc, err := checkCreateLocked(rp, parentVFSD, parentInode) + if err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + child, err := parentInode.NewDir(ctx, pc, opts) + if err != nil { + return err + } + parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + return nil +} + +// MknodAt implements vfs.FilesystemImpl.MknodAt. +func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { + if rp.Done() { + return syserror.EEXIST + } + fs.mu.Lock() + defer fs.mu.Unlock() + parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + pc, err := checkCreateLocked(rp, parentVFSD, parentInode) + if err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + new, err := parentInode.NewNode(ctx, pc, opts) + if err != nil { + return err + } + parentVFSD.Impl().(*Dentry).InsertChild(pc, new) + return nil +} + +// OpenAt implements vfs.FilesystemImpl.OpenAt. +func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + // Filter out flags that are not supported by kernfs. O_DIRECTORY and + // O_NOFOLLOW have no effect here (they're handled by VFS by setting + // appropriate bits in rp), but are returned by + // FileDescriptionImpl.StatusFlags(). + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW + ats := vfs.AccessTypesForOpenFlags(opts.Flags) + + // Do not create new file. + if opts.Flags&linux.O_CREAT == 0 { + fs.mu.RLock() + defer fs.processDeferredDecRefs() + defer fs.mu.RUnlock() + vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + if err != nil { + return nil, err + } + if err := inode.CheckPermissions(rp.Credentials(), ats); err != nil { + return nil, err + } + return inode.Open(rp, vfsd, opts.Flags) + } + + // May create new file. + mustCreate := opts.Flags&linux.O_EXCL != 0 + vfsd := rp.Start() + inode := vfsd.Impl().(*Dentry).inode + fs.mu.Lock() + defer fs.mu.Unlock() + if rp.Done() { + if rp.MustBeDir() { + return nil, syserror.EISDIR + } + if mustCreate { + return nil, syserror.EEXIST + } + if err := inode.CheckPermissions(rp.Credentials(), ats); err != nil { + return nil, err + } + return inode.Open(rp, vfsd, opts.Flags) + } +afterTrailingSymlink: + parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return nil, err + } + // Check for search permission in the parent directory. + if err := parentInode.CheckPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } + // Reject attempts to open directories with O_CREAT. + if rp.MustBeDir() { + return nil, syserror.EISDIR + } + pc := rp.Component() + if pc == "." || pc == ".." { + return nil, syserror.EISDIR + } + // Determine whether or not we need to create a file. + childVFSD, err := rp.ResolveChild(parentVFSD, pc) + if err != nil { + return nil, err + } + if childVFSD == nil { + // Already checked for searchability above; now check for writability. + if err := parentInode.CheckPermissions(rp.Credentials(), vfs.MayWrite); err != nil { + return nil, err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return nil, err + } + defer rp.Mount().EndWrite() + // Create and open the child. + child, err := parentInode.NewFile(ctx, pc, opts) + if err != nil { + return nil, err + } + parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + return child.Impl().(*Dentry).inode.Open(rp, child, opts.Flags) + } + // Open existing file or follow symlink. + if mustCreate { + return nil, syserror.EEXIST + } + childDentry := childVFSD.Impl().(*Dentry) + childInode := childDentry.inode + if rp.ShouldFollowSymlink() { + if childDentry.isSymlink() { + target, err := childInode.Readlink(ctx) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + // rp.Final() may no longer be true since we now need to resolve the + // symlink target. + goto afterTrailingSymlink + } + } + if err := childInode.CheckPermissions(rp.Credentials(), ats); err != nil { + return nil, err + } + return childInode.Open(rp, childVFSD, opts.Flags) +} + +// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. +func (fs *Filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { + fs.mu.RLock() + d, inode, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return "", err + } + if !d.Impl().(*Dentry).isSymlink() { + return "", syserror.EINVAL + } + return inode.Readlink(ctx) +} + +// RenameAt implements vfs.FilesystemImpl.RenameAt. +func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error { + // Only RENAME_NOREPLACE is supported. + if opts.Flags&^linux.RENAME_NOREPLACE != 0 { + return syserror.EINVAL + } + noReplace := opts.Flags&linux.RENAME_NOREPLACE != 0 + + fs.mu.Lock() + defer fs.mu.Lock() + + // Resolve the destination directory first to verify that it's on this + // Mount. + dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + mnt := rp.Mount() + if mnt != oldParentVD.Mount() { + return syserror.EXDEV + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + + srcDirVFSD := oldParentVD.Dentry() + srcDir := srcDirVFSD.Impl().(*Dentry) + srcDir.dirMu.Lock() + src, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), srcDir, oldName, srcDirVFSD.Child(oldName)) + srcDir.dirMu.Unlock() + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + srcVFSD := &src.vfsd + + // Can we remove the src dentry? + if err := checkDeleteLocked(rp, srcVFSD); err != nil { + return err + } + + // Can we create the dst dentry? + var dstVFSD *vfs.Dentry + pc, err := checkCreateLocked(rp, dstDirVFSD, dstDirInode) + switch err { + case nil: + // Ok, continue with rename as replacement. + case syserror.EEXIST: + if noReplace { + // Won't overwrite existing node since RENAME_NOREPLACE was requested. + return syserror.EEXIST + } + dstVFSD, err = rp.ResolveChild(dstDirVFSD, pc) + if err != nil { + panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) + } + default: + return err + } + + mntns := vfs.MountNamespaceFromContext(ctx) + virtfs := rp.VirtualFilesystem() + + srcDirDentry := srcDirVFSD.Impl().(*Dentry) + dstDirDentry := dstDirVFSD.Impl().(*Dentry) + + // We can't deadlock here due to lock ordering because we're protected from + // concurrent renames by fs.mu held for writing. + srcDirDentry.dirMu.Lock() + defer srcDirDentry.dirMu.Unlock() + dstDirDentry.dirMu.Lock() + defer dstDirDentry.dirMu.Unlock() + + if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { + return err + } + srcDirInode := srcDirDentry.inode + replaced, err := srcDirInode.Rename(ctx, srcVFSD.Name(), pc, srcVFSD, dstDirVFSD) + if err != nil { + virtfs.AbortRenameDentry(srcVFSD, dstVFSD) + return err + } + virtfs.CommitRenameReplaceDentry(srcVFSD, dstDirVFSD, pc, replaced) + return nil +} + +// RmdirAt implements vfs.FilesystemImpl.RmdirAt. +func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { + fs.mu.Lock() + defer fs.mu.Unlock() + vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + if err := checkDeleteLocked(rp, vfsd); err != nil { + return err + } + if !vfsd.Impl().(*Dentry).isDir() { + return syserror.ENOTDIR + } + if inode.HasChildren() { + return syserror.ENOTEMPTY + } + virtfs := rp.VirtualFilesystem() + parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry.dirMu.Lock() + defer parentDentry.dirMu.Unlock() + if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + return err + } + if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { + virtfs.AbortDeleteDentry(vfsd) + return err + } + virtfs.CommitDeleteDentry(vfsd) + return nil +} + +// SetStatAt implements vfs.FilesystemImpl.SetStatAt. +func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { + fs.mu.RLock() + _, inode, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return err + } + if opts.Stat.Mask == 0 { + return nil + } + return inode.SetStat(fs.VFSFilesystem(), opts) +} + +// StatAt implements vfs.FilesystemImpl.StatAt. +func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { + fs.mu.RLock() + _, inode, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return linux.Statx{}, err + } + return inode.Stat(fs.VFSFilesystem()), nil +} + +// StatFSAt implements vfs.FilesystemImpl.StatFSAt. +func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return linux.Statfs{}, err + } + // TODO: actually implement statfs + return linux.Statfs{}, syserror.ENOSYS +} + +// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. +func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { + if rp.Done() { + return syserror.EEXIST + } + fs.mu.Lock() + defer fs.mu.Unlock() + parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + pc, err := checkCreateLocked(rp, parentVFSD, parentInode) + if err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + child, err := parentInode.NewSymlink(ctx, pc, target) + if err != nil { + return err + } + parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + return nil +} + +// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. +func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { + fs.mu.Lock() + defer fs.mu.Unlock() + vfsd, _, err := fs.walkExistingLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + if err := checkDeleteLocked(rp, vfsd); err != nil { + return err + } + if vfsd.Impl().(*Dentry).isDir() { + return syserror.EISDIR + } + virtfs := rp.VirtualFilesystem() + parentDentry := vfsd.Parent().Impl().(*Dentry) + parentDentry.dirMu.Lock() + defer parentDentry.dirMu.Unlock() + if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + return err + } + if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { + virtfs.AbortDeleteDentry(vfsd) + return err + } + virtfs.CommitDeleteDentry(vfsd) + return nil +} + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return nil, err + } + // kernfs currently does not support extended attributes. + return nil, syserror.ENOTSUP +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return "", err + } + // kernfs currently does not support extended attributes. + return "", syserror.ENOTSUP +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *Filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return err + } + // kernfs currently does not support extended attributes. + return syserror.ENOTSUP +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *Filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + fs.mu.RLock() + _, _, err := fs.walkExistingLocked(ctx, rp) + fs.mu.RUnlock() + fs.processDeferredDecRefs() + if err != nil { + return err + } + // kernfs currently does not support extended attributes. + return syserror.ENOTSUP +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *Filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go new file mode 100644 index 000000000..752e0f659 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -0,0 +1,512 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs + +import ( + "fmt" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// InodeNoopRefCount partially implements the Inode interface, specifically the +// inodeRefs sub interface. InodeNoopRefCount implements a simple reference +// count for inodes, performing no extra actions when references are obtained or +// released. This is suitable for simple file inodes that don't reference any +// resources. +type InodeNoopRefCount struct { +} + +// IncRef implements Inode.IncRef. +func (n *InodeNoopRefCount) IncRef() { +} + +// DecRef implements Inode.DecRef. +func (n *InodeNoopRefCount) DecRef() { +} + +// TryIncRef implements Inode.TryIncRef. +func (n *InodeNoopRefCount) TryIncRef() bool { + return true +} + +// Destroy implements Inode.Destroy. +func (n *InodeNoopRefCount) Destroy() { +} + +// InodeDirectoryNoNewChildren partially implements the Inode interface. +// InodeDirectoryNoNewChildren represents a directory inode which does not +// support creation of new children. +type InodeDirectoryNoNewChildren struct{} + +// NewFile implements Inode.NewFile. +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) { + return nil, syserror.EPERM +} + +// NewLink implements Inode.NewLink. +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) { + return nil, syserror.EPERM +} + +// NewNode implements Inode.NewNode. +func (*InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { + return nil, syserror.EPERM +} + +// InodeNotDirectory partially implements the Inode interface, specifically the +// inodeDirectory and inodeDynamicDirectory sub interfaces. Inodes that do not +// represent directories can embed this to provide no-op implementations for +// directory-related functions. +type InodeNotDirectory struct { +} + +// HasChildren implements Inode.HasChildren. +func (*InodeNotDirectory) HasChildren() bool { + return false +} + +// NewFile implements Inode.NewFile. +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) { + panic("NewDir called on non-directory inode") +} + +// NewLink implements Inode.NewLinkink. +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) { + panic("NewSymlink called on non-directory inode") +} + +// NewNode implements Inode.NewNode. +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 { + panic("Unlink called on non-directory inode") +} + +// RmDir implements Inode.RmDir. +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) { + panic("Rename called on non-directory inode") +} + +// Lookup implements Inode.Lookup. +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) { + panic("IterDirents called on non-directory inode") +} + +// Valid implements Inode.Valid. +func (*InodeNotDirectory) Valid(context.Context) bool { + return true +} + +// InodeNoDynamicLookup partially implements the Inode interface, specifically +// the inodeDynamicLookup sub interface. Directory inodes that do not support +// dymanic entries (i.e. entries that are not "hashed" into the +// vfs.Dentry.children) can embed this to provide no-op implementations for +// functions related to dynamic entries. +type InodeNoDynamicLookup struct{} + +// Lookup implements Inode.Lookup. +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) { + return offset, nil +} + +// Valid implements Inode.Valid. +func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { + return true +} + +// InodeNotSymlink partially implements the Inode interface, specifically the +// inodeSymlink sub interface. All inodes that are not symlinks may embed this +// to return the appropriate errors from symlink-related functions. +type InodeNotSymlink struct{} + +// Readlink implements Inode.Readlink. +func (*InodeNotSymlink) Readlink(context.Context) (string, error) { + return "", syserror.EINVAL +} + +// InodeAttrs partially implements the Inode interface, specifically the +// inodeMetadata sub interface. InodeAttrs provides functionality related to +// inode attributes. +// +// Must be initialized by Init prior to first use. +type InodeAttrs struct { + ino uint64 + mode uint32 + uid uint32 + gid uint32 + nlink uint32 +} + +// Init initializes this InodeAttrs. +func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMode) { + if mode.FileType() == 0 { + panic(fmt.Sprintf("No file type specified in 'mode' for InodeAttrs.Init(): mode=0%o", mode)) + } + + nlink := uint32(1) + if mode.FileType() == linux.ModeDirectory { + nlink = 2 + } + atomic.StoreUint64(&a.ino, ino) + atomic.StoreUint32(&a.mode, uint32(mode)) + atomic.StoreUint32(&a.uid, uint32(creds.EffectiveKUID)) + atomic.StoreUint32(&a.gid, uint32(creds.EffectiveKGID)) + atomic.StoreUint32(&a.nlink, nlink) +} + +// Mode implements Inode.Mode. +func (a *InodeAttrs) Mode() linux.FileMode { + return linux.FileMode(atomic.LoadUint32(&a.mode)) +} + +// 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 { + 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) + stat.Mode = uint16(a.Mode()) + stat.UID = atomic.LoadUint32(&a.uid) + stat.GID = atomic.LoadUint32(&a.gid) + stat.Nlink = atomic.LoadUint32(&a.nlink) + + // TODO: Implement other stat fields like timestamps. + + return stat +} + +// SetStat implements Inode.SetStat. +func (a *InodeAttrs) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { + stat := opts.Stat + if stat.Mask&linux.STATX_MODE != 0 { + for { + old := atomic.LoadUint32(&a.mode) + new := old | uint32(stat.Mode & ^uint16(linux.S_IFMT)) + if swapped := atomic.CompareAndSwapUint32(&a.mode, old, new); swapped { + break + } + } + } + + if stat.Mask&linux.STATX_UID != 0 { + atomic.StoreUint32(&a.uid, stat.UID) + } + if stat.Mask&linux.STATX_GID != 0 { + atomic.StoreUint32(&a.gid, stat.GID) + } + + // 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. + + return nil +} + +// CheckPermissions implements Inode.CheckPermissions. +func (a *InodeAttrs) CheckPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + mode := a.Mode() + return vfs.GenericCheckPermissions( + creds, + ats, + mode.FileType() == linux.ModeDirectory, + uint16(mode), + auth.KUID(atomic.LoadUint32(&a.uid)), + auth.KGID(atomic.LoadUint32(&a.gid)), + ) +} + +// IncLinks implements Inode.IncLinks. +func (a *InodeAttrs) IncLinks(n uint32) { + if atomic.AddUint32(&a.nlink, n) <= n { + panic("InodeLink.IncLinks called with no existing links") + } +} + +// DecLinks implements Inode.DecLinks. +func (a *InodeAttrs) DecLinks() { + if nlink := atomic.AddUint32(&a.nlink, ^uint32(0)); nlink == ^uint32(0) { + // Negative overflow + panic("Inode.DecLinks called at 0 links") + } +} + +type slot struct { + Name string + Dentry *vfs.Dentry + slotEntry +} + +// OrderedChildrenOptions contains initialization options for OrderedChildren. +type OrderedChildrenOptions struct { + // Writable indicates whether vfs.FilesystemImpl methods implemented by + // OrderedChildren may modify the tracked children. This applies to + // operations related to rename, unlink and rmdir. If an OrderedChildren is + // not writable, these operations all fail with EPERM. + Writable bool +} + +// OrderedChildren partially implements the Inode interface. OrderedChildren can +// be embedded in directory inodes to keep track of the children in the +// directory, and can then be used to implement a generic directory FD -- see +// GenericDirectoryFD. OrderedChildren is not compatible with dynamic +// directories. +// +// Must be initialize with Init before first use. +type OrderedChildren struct { + refs.AtomicRefCount + + // Can children be modified by user syscalls? It set to false, interface + // methods that would modify the children return EPERM. Immutable. + writable bool + + mu sync.RWMutex + order slotList + set map[string]*slot +} + +// Init initializes an OrderedChildren. +func (o *OrderedChildren) Init(opts OrderedChildrenOptions) { + o.writable = opts.Writable + o.set = make(map[string]*slot) +} + +// DecRef implements Inode.DecRef. +func (o *OrderedChildren) DecRef() { + o.AtomicRefCount.DecRefWithDestructor(o.Destroy) +} + +// Destroy cleans up resources referenced by this OrderedChildren. +func (o *OrderedChildren) Destroy() { + o.mu.Lock() + defer o.mu.Unlock() + o.order.Reset() + o.set = nil +} + +// Populate inserts children into this OrderedChildren, and d's dentry +// cache. Populate returns the number of directories inserted, which the caller +// may use to update the link count for the parent directory. +// +// Precondition: d.Impl() must be a kernfs Dentry. d must represent a directory +// inode. children must not contain any conflicting entries already in o. +func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint32 { + var links uint32 + for name, child := range children { + if child.isDir() { + links++ + } + if err := o.Insert(name, child.VFSDentry()); err != nil { + panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) + } + d.InsertChild(name, child.VFSDentry()) + } + return links +} + +// HasChildren implements Inode.HasChildren. +func (o *OrderedChildren) HasChildren() bool { + o.mu.RLock() + defer o.mu.RUnlock() + return len(o.set) > 0 +} + +// Insert inserts child into o. This ignores the writability of o, as this is +// not part of the vfs.FilesystemImpl interface, and is a lower-level operation. +func (o *OrderedChildren) Insert(name string, child *vfs.Dentry) error { + o.mu.Lock() + defer o.mu.Unlock() + if _, ok := o.set[name]; ok { + return syserror.EEXIST + } + s := &slot{ + Name: name, + Dentry: child, + } + o.order.PushBack(s) + o.set[name] = s + return nil +} + +// Precondition: caller must hold o.mu for writing. +func (o *OrderedChildren) removeLocked(name string) { + if s, ok := o.set[name]; ok { + delete(o.set, name) + o.order.Remove(s) + } +} + +// Precondition: caller must hold o.mu for writing. +func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs.Dentry { + if s, ok := o.set[name]; ok { + // Existing slot with given name, simply replace the dentry. + var old *vfs.Dentry + old, s.Dentry = s.Dentry, new + return old + } + + // No existing slot with given name, create and hash new slot. + s := &slot{ + Name: name, + Dentry: new, + } + o.order.PushBack(s) + o.set[name] = s + return nil +} + +// Precondition: caller must hold o.mu for reading or writing. +func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) error { + s, ok := o.set[name] + if !ok { + return syserror.ENOENT + } + if s.Dentry != child { + panic(fmt.Sprintf("Dentry hashed into inode doesn't match what vfs thinks! OrderedChild: %+v, vfs: %+v", s.Dentry, child)) + } + return nil +} + +// Unlink implements Inode.Unlink. +func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.Dentry) error { + if !o.writable { + return syserror.EPERM + } + o.mu.Lock() + defer o.mu.Unlock() + if err := o.checkExistingLocked(name, child); err != nil { + return err + } + o.removeLocked(name) + return nil +} + +// Rmdir implements Inode.Rmdir. +func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *vfs.Dentry) error { + // We're not responsible for checking that child is a directory, that it's + // empty, or updating any link counts; so this is the same as unlink. + return o.Unlink(ctx, name, child) +} + +type renameAcrossDifferentImplementationsError struct{} + +func (renameAcrossDifferentImplementationsError) Error() string { + return "rename across inodes with different implementations" +} + +// Rename implements Inode.Rename. +// +// Precondition: Rename may only be called across two directory inodes with +// identical implementations of Rename. Practically, this means filesystems that +// implement Rename by embedding OrderedChildren for any directory +// implementation must use OrderedChildren for all directory implementations +// that will support Rename. +// +// Postcondition: reference on any replaced dentry transferred to caller. +func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (*vfs.Dentry, error) { + dst, ok := dstDir.Impl().(*Dentry).inode.(interface{}).(*OrderedChildren) + if !ok { + return nil, renameAcrossDifferentImplementationsError{} + } + if !o.writable || !dst.writable { + return nil, syserror.EPERM + } + // Note: There's a potential deadlock below if concurrent calls to Rename + // refer to the same src and dst directories in reverse. We avoid any + // ordering issues because the caller is required to serialize concurrent + // calls to Rename in accordance with the interface declaration. + o.mu.Lock() + defer o.mu.Unlock() + if dst != o { + dst.mu.Lock() + defer dst.mu.Unlock() + } + if err := o.checkExistingLocked(oldname, child); err != nil { + return nil, err + } + replaced := dst.replaceChildLocked(newname, child) + return replaced, nil +} + +// nthLocked returns an iterator to the nth child tracked by this object. The +// iterator is valid until the caller releases o.mu. Returns nil if the +// requested index falls out of bounds. +// +// Preconditon: Caller must hold o.mu for reading. +func (o *OrderedChildren) nthLocked(i int64) *slot { + for it := o.order.Front(); it != nil && i >= 0; it = it.Next() { + if i == 0 { + return it + } + i-- + } + return nil +} + +// InodeSymlink partially implements Inode interface for symlinks. +type InodeSymlink struct { + InodeNotDirectory +} + +// Open implements Inode.Open. +func (InodeSymlink) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + return nil, syserror.ELOOP +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go new file mode 100644 index 000000000..d69b299ae --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -0,0 +1,422 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package kernfs provides the tools to implement inode-based filesystems. +// Kernfs has two main features: +// +// 1. The Inode interface, which maps VFS2's path-based filesystem operations to +// specific filesystem nodes. Kernfs uses the Inode interface to provide a +// blanket implementation for the vfs.FilesystemImpl. Kernfs also serves as +// the synchronization mechanism for all filesystem operations by holding a +// filesystem-wide lock across all operations. +// +// 2. Various utility types which provide generic implementations for various +// parts of the Inode and vfs.FileDescription interfaces. Client filesystems +// based on kernfs can embed the appropriate set of these to avoid having to +// reimplement common filesystem operations. See inode_impl_util.go and +// fd_impl_util.go. +// +// Reference Model: +// +// Kernfs dentries represents named pointers to inodes. Dentries and inode have +// independent lifetimes and reference counts. A child dentry unconditionally +// holds a reference on its parent directory's dentry. A dentry also holds a +// reference on the inode it points to. Multiple dentries can point to the same +// inode (for example, in the case of hardlinks). File descriptors hold a +// reference to the dentry they're opened on. +// +// Dentries are guaranteed to exist while holding Filesystem.mu for +// reading. Dropping dentries require holding Filesystem.mu for writing. To +// queue dentries for destruction from a read critical section, see +// Filesystem.deferDecRef. +// +// Lock ordering: +// +// kernfs.Filesystem.mu +// kernfs.Dentry.dirMu +// vfs.VirtualFilesystem.mountMu +// vfs.Dentry.mu +// kernfs.Filesystem.droppedDentriesMu +// (inode implementation locks, if any) +package kernfs + +import ( + "fmt" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// FilesystemType implements vfs.FilesystemType. +type FilesystemType struct{} + +// Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory +// filesystem. Concrete implementations are expected to embed this in their own +// Filesystem type. +type Filesystem struct { + vfsfs vfs.Filesystem + + droppedDentriesMu sync.Mutex + + // droppedDentries is a list of dentries waiting to be DecRef()ed. This is + // used to defer dentry destruction until mu can be acquired for + // writing. Protected by droppedDentriesMu. + droppedDentries []*vfs.Dentry + + // mu synchronizes the lifetime of Dentries on this filesystem. Holding it + // for reading guarantees continued existence of any resolved dentries, but + // the dentry tree may be modified. + // + // Kernfs dentries can only be DecRef()ed while holding mu for writing. For + // example: + // + // fs.mu.Lock() + // defer fs.mu.Unlock() + // ... + // dentry1.DecRef() + // defer dentry2.DecRef() // Ok, will run before Unlock. + // + // If discarding dentries in a read context, use Filesystem.deferDecRef. For + // example: + // + // fs.mu.RLock() + // fs.mu.processDeferredDecRefs() + // defer fs.mu.RUnlock() + // ... + // fs.deferDecRef(dentry) + mu sync.RWMutex + + // nextInoMinusOne is used to to allocate inode numbers on this + // filesystem. Must be accessed by atomic operations. + nextInoMinusOne uint64 +} + +// deferDecRef defers dropping a dentry ref until the next call to +// processDeferredDecRefs{,Locked}. See comment on Filesystem.mu. +// +// Precondition: d must not already be pending destruction. +func (fs *Filesystem) deferDecRef(d *vfs.Dentry) { + fs.droppedDentriesMu.Lock() + fs.droppedDentries = append(fs.droppedDentries, d) + fs.droppedDentriesMu.Unlock() +} + +// processDeferredDecRefs calls vfs.Dentry.DecRef on all dentries in the +// droppedDentries list. See comment on Filesystem.mu. +func (fs *Filesystem) processDeferredDecRefs() { + fs.mu.Lock() + fs.processDeferredDecRefsLocked() + fs.mu.Unlock() +} + +// Precondition: fs.mu must be held for writing. +func (fs *Filesystem) processDeferredDecRefsLocked() { + fs.droppedDentriesMu.Lock() + for _, d := range fs.droppedDentries { + d.DecRef() + } + fs.droppedDentries = fs.droppedDentries[:0] // Keep slice memory for reuse. + fs.droppedDentriesMu.Unlock() +} + +// Init initializes a kernfs filesystem. This should be called from during +// vfs.FilesystemType.NewFilesystem for the concrete filesystem embedding +// kernfs. +func (fs *Filesystem) Init(vfsObj *vfs.VirtualFilesystem) { + fs.vfsfs.Init(vfsObj, fs) +} + +// VFSFilesystem returns the generic vfs filesystem object. +func (fs *Filesystem) VFSFilesystem() *vfs.Filesystem { + return &fs.vfsfs +} + +// NextIno allocates a new inode number on this filesystem. +func (fs *Filesystem) NextIno() uint64 { + return atomic.AddUint64(&fs.nextInoMinusOne, 1) +} + +// These consts are used in the Dentry.flags field. +const ( + // Dentry points to a directory inode. + dflagsIsDir = 1 << iota + + // Dentry points to a symlink inode. + dflagsIsSymlink +) + +// Dentry implements vfs.DentryImpl. +// +// A kernfs dentry is similar to a dentry in a traditional filesystem: it's a +// named reference to an inode. A dentry generally lives as long as it's part of +// a mounted filesystem tree. Kernfs doesn't cache dentries once all references +// to them are removed. Dentries hold a single reference to the inode they point +// to, and child dentries hold a reference on their parent. +// +// Must be initialized by Init prior to first use. +type Dentry struct { + refs.AtomicRefCount + + 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 + + // dirMu protects vfsd.children for directory dentries. + dirMu sync.Mutex +} + +// Init initializes this dentry. +// +// Precondition: Caller must hold a reference on inode. +// +// Postcondition: Caller's reference on inode is transferred to the dentry. +func (d *Dentry) Init(inode Inode) { + d.vfsd.Init(d) + d.inode = inode + ftype := inode.Mode().FileType() + if ftype == linux.ModeDirectory { + d.flags |= dflagsIsDir + } + if ftype == linux.ModeSymlink { + d.flags |= dflagsIsSymlink + } +} + +// VFSDentry returns the generic vfs dentry for this kernfs dentry. +func (d *Dentry) VFSDentry() *vfs.Dentry { + return &d.vfsd +} + +// isDir checks whether the dentry points to a directory inode. +func (d *Dentry) isDir() bool { + return atomic.LoadUint32(&d.flags)&dflagsIsDir != 0 +} + +// isSymlink checks whether the dentry points to a symlink inode. +func (d *Dentry) isSymlink() bool { + return atomic.LoadUint32(&d.flags)&dflagsIsSymlink != 0 +} + +// DecRef implements vfs.DentryImpl.DecRef. +func (d *Dentry) DecRef() { + d.AtomicRefCount.DecRefWithDestructor(d.destroy) +} + +// Precondition: Dentry must be removed from VFS' dentry cache. +func (d *Dentry) destroy() { + d.inode.DecRef() // IncRef from Init. + d.inode = nil + if parent := d.vfsd.Parent(); parent != nil { + parent.DecRef() // IncRef from Dentry.InsertChild. + } +} + +// InsertChild inserts child into the vfs dentry cache with the given name under +// this dentry. This does not update the directory inode, so calling this on +// it's own isn't sufficient to insert a child into a directory. InsertChild +// updates the link count on d if required. +// +// Precondition: d must represent a directory inode. +func (d *Dentry) InsertChild(name string, child *vfs.Dentry) { + d.dirMu.Lock() + d.insertChildLocked(name, child) + d.dirMu.Unlock() +} + +// insertChildLocked is equivalent to InsertChild, with additional +// preconditions. +// +// Precondition: d.dirMu must be locked. +func (d *Dentry) insertChildLocked(name string, child *vfs.Dentry) { + if !d.isDir() { + panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) + } + vfsDentry := d.VFSDentry() + vfsDentry.IncRef() // DecRef in child's Dentry.destroy. + vfsDentry.InsertChild(child, name) +} + +// The Inode interface maps filesystem-level operations that operate on paths to +// equivalent operations on specific filesystem nodes. +// +// The interface methods are groups into logical categories as sub interfaces +// below. Generally, an implementation for each sub interface can be provided by +// embedding an appropriate type from inode_impl_utils.go. The sub interfaces +// are purely organizational. Methods declared directly in the main interface +// have no generic implementations, and should be explicitly provided by the +// client filesystem. +// +// Generally, implementations are not responsible for tasks that are common to +// all filesystems. These include: +// +// - Checking that dentries passed to methods are of the appropriate file type. +// - Checking permissions. +// - Updating link and reference counts. +// +// Specific responsibilities of implementations are documented below. +type Inode interface { + // Methods related to reference counting. A generic implementation is + // provided by InodeNoopRefCount. These methods are generally called by the + // equivalent Dentry methods. + inodeRefs + + // Methods related to node metadata. A generic implementation is provided by + // InodeAttrs. + inodeMetadata + + // Method for inodes that represent symlink. InodeNotSymlink provides a + // blanket implementation for all non-symlink inodes. + inodeSymlink + + // Method for inodes that represent directories. InodeNotDirectory provides + // a blanket implementation for all non-directory inodes. + inodeDirectory + + // Method for inodes that represent dynamic directories and their + // children. InodeNoDynamicLookup provides a blanket implementation for all + // non-dynamic-directory inodes. + inodeDynamicLookup + + // Open creates a file description for the filesystem object represented by + // this inode. The returned file description should hold a reference on the + // inode for its lifetime. + // + // Precondition: !rp.Done(). vfsd.Impl() must be a kernfs Dentry. + Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) +} + +type inodeRefs interface { + IncRef() + DecRef() + TryIncRef() bool + // Destroy is called when the inode reaches zero references. Destroy release + // all resources (references) on objects referenced by the inode, including + // any child dentries. + Destroy() +} + +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(creds *auth.Credentials, atx vfs.AccessTypes) error + + // Mode returns the (struct stat)::st_mode value for this inode. This is + // separated from Stat for performance. + Mode() linux.FileMode + + // Stat returns the metadata for this inode. This corresponds to + // vfs.FilesystemImpl.StatAt. + Stat(fs *vfs.Filesystem) linux.Statx + + // SetStat updates the metadata for this inode. This corresponds to + // vfs.FilesystemImpl.SetStatAt. + SetStat(fs *vfs.Filesystem, opts vfs.SetStatOptions) error +} + +// Precondition: All methods in this interface may only be called on directory +// inodes. +type inodeDirectory interface { + // The New{File,Dir,Node,Symlink} methods below should return a new inode + // hashed into this inode. + // + // These inode constructors are inode-level operations rather than + // filesystem-level operations to allow client filesystems to mix different + // implementations based on the new node's location in the + // filesystem. + + // HasChildren returns true if the directory inode has any children. + HasChildren() bool + + // NewFile creates a new regular file inode. + NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) + + // NewDir creates a new directory inode. + NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) + + // NewLink creates a new hardlink to a specified inode in this + // directory. Implementations should create a new kernfs Dentry pointing to + // target, and update target's link count. + NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) + + // NewSymlink creates a new symbolic link inode. + NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) + + // NewNode creates a new filesystem node for a mknod syscall. + NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) + + // Unlink removes a child dentry from this directory inode. + Unlink(ctx context.Context, name string, child *vfs.Dentry) error + + // RmDir removes an empty child directory from this directory + // inode. Implementations must update the parent directory's link count, + // if required. Implementations are not responsible for checking that child + // is a directory, checking for an empty directory. + RmDir(ctx context.Context, name string, child *vfs.Dentry) error + + // Rename is called on the source directory containing an inode being + // renamed. child should point to the resolved child in the source + // directory. If Rename replaces a dentry in the destination directory, it + // should return the replaced dentry or nil otherwise. + // + // Precondition: Caller must serialize concurrent calls to Rename. + Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (replaced *vfs.Dentry, err error) +} + +type inodeDynamicLookup interface { + // Lookup should return an appropriate dentry if name should resolve to a + // child of this dynamic directory inode. This gives the directory an + // opportunity on every lookup to resolve additional entries that aren't + // hashed into the directory. This is only called when the inode is a + // directory. If the inode is not a directory, or if the directory only + // contains a static set of children, the implementer can unconditionally + // return an appropriate error (ENOTDIR and ENOENT respectively). + // + // The child returned by Lookup will be hashed into the VFS dentry tree. Its + // lifetime can be controlled by the filesystem implementation with an + // appropriate implementation of Valid. + // + // Lookup returns the child with an extra reference and the caller owns this + // reference. + Lookup(ctx context.Context, name string) (*vfs.Dentry, error) + + // Valid should return true if this inode is still valid, or needs to + // be resolved again by a call to Lookup. + Valid(ctx context.Context) bool + + // IterDirents is used to iterate over dynamically created entries. It invokes + // cb on each entry in the directory represented by the FileDescription. + // 'offset' is the offset for the entire IterDirents call, which may include + // results from the caller. 'relOffset' is the offset inside the entries + // returned by this IterDirents invocation. In other words, + // 'offset+relOffset+1' is the value that should be set in vfs.Dirent.NextOff, + // while 'relOffset' is the place where iteration should start from. + IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) +} + +type inodeSymlink interface { + // Readlink resolves the target of a symbolic link. If an inode is not a + // symlink, the implementation should return EINVAL. + Readlink(ctx context.Context) (string, error) +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go new file mode 100644 index 000000000..4b6b95f5f --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -0,0 +1,426 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs_test + +import ( + "bytes" + "fmt" + "io" + "runtime" + "sync" + "testing" + + "github.com/google/go-cmp/cmp" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +const defaultMode linux.FileMode = 01777 +const staticFileContent = "This is sample content for a static test file." + +// RootDentryFn is a generator function for creating the root dentry of a test +// filesystem. See newTestSystem. +type RootDentryFn func(*auth.Credentials, *filesystem) *kernfs.Dentry + +// TestSystem represents the context for a single test. +type TestSystem struct { + t *testing.T + ctx context.Context + creds *auth.Credentials + vfs *vfs.VirtualFilesystem + mns *vfs.MountNamespace + root vfs.VirtualDentry +} + +// newTestSystem sets up a minimal environment for running a test, including an +// instance of a test filesystem. Tests can control the contents of the +// filesystem by providing an appropriate rootFn, which should return a +// pre-populated root dentry. +func newTestSystem(t *testing.T, rootFn RootDentryFn) *TestSystem { + ctx := contexttest.Context(t) + creds := auth.CredentialsFromContext(ctx) + v := vfs.New() + v.MustRegisterFilesystemType("testfs", &fsType{rootFn: rootFn}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mns, err := v.NewMountNamespace(ctx, creds, "", "testfs", &vfs.GetFilesystemOptions{}) + if err != nil { + t.Fatalf("Failed to create testfs root mount: %v", err) + } + + s := &TestSystem{ + t: t, + ctx: ctx, + creds: creds, + vfs: v, + mns: mns, + root: mns.Root(), + } + runtime.SetFinalizer(s, func(s *TestSystem) { s.root.DecRef() }) + return s +} + +// PathOpAtRoot constructs a vfs.PathOperation for a path from the +// root of the test filesystem. +// +// Precondition: path should be relative path. +func (s *TestSystem) PathOpAtRoot(path string) vfs.PathOperation { + return vfs.PathOperation{ + Root: s.root, + Start: s.root, + Path: fspath.Parse(path), + } +} + +// GetDentryOrDie attempts to resolve a dentry referred to by the +// provided path operation. If unsuccessful, the test fails. +func (s *TestSystem) GetDentryOrDie(pop vfs.PathOperation) vfs.VirtualDentry { + vd, err := s.vfs.GetDentryAt(s.ctx, s.creds, &pop, &vfs.GetDentryOptions{}) + if err != nil { + s.t.Fatalf("GetDentryAt(pop:%+v) failed: %v", pop, err) + } + return vd +} + +func (s *TestSystem) ReadToEnd(fd *vfs.FileDescription) (string, error) { + buf := make([]byte, usermem.PageSize) + bufIOSeq := usermem.BytesIOSequence(buf) + opts := vfs.ReadOptions{} + + var content bytes.Buffer + for { + n, err := fd.Impl().Read(s.ctx, bufIOSeq, opts) + if n == 0 || err != nil { + if err == io.EOF { + err = nil + } + return content.String(), err + } + content.Write(buf[:n]) + } +} + +type fsType struct { + rootFn RootDentryFn +} + +type filesystem struct { + kernfs.Filesystem +} + +type file struct { + kernfs.DynamicBytesFile + content string +} + +func (fs *filesystem) newFile(creds *auth.Credentials, content string) *kernfs.Dentry { + f := &file{} + f.content = content + f.DynamicBytesFile.Init(creds, fs.NextIno(), f, 0777) + + d := &kernfs.Dentry{} + d.Init(f) + return d +} + +func (f *file) Generate(ctx context.Context, buf *bytes.Buffer) error { + fmt.Fprintf(buf, "%s", f.content) + return nil +} + +type attrs struct { + kernfs.InodeAttrs +} + +func (a *attrs) SetStat(fs *vfs.Filesystem, opt vfs.SetStatOptions) error { + return syserror.EPERM +} + +type readonlyDir struct { + attrs + kernfs.InodeNotSymlink + kernfs.InodeNoDynamicLookup + kernfs.InodeDirectoryNoNewChildren + + kernfs.OrderedChildren + dentry kernfs.Dentry +} + +func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { + dir := &readonlyDir{} + dir.attrs.Init(creds, fs.NextIno(), linux.ModeDirectory|mode) + dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + dir.dentry.Init(dir) + + dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) + + return &dir.dentry +} + +func (d *readonlyDir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, flags) + return fd.VFSFileDescription(), nil +} + +type dir struct { + attrs + kernfs.InodeNotSymlink + kernfs.InodeNoDynamicLookup + + fs *filesystem + dentry kernfs.Dentry + kernfs.OrderedChildren +} + +func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { + dir := &dir{} + dir.fs = fs + dir.attrs.Init(creds, fs.NextIno(), linux.ModeDirectory|mode) + dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{Writable: true}) + dir.dentry.Init(dir) + + dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) + + return &dir.dentry +} + +func (d *dir) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &d.OrderedChildren, flags) + return fd.VFSFileDescription(), nil +} + +func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { + creds := auth.CredentialsFromContext(ctx) + dir := d.fs.newDir(creds, opts.Mode, nil) + dirVFSD := dir.VFSDentry() + if err := d.OrderedChildren.Insert(name, dirVFSD); err != nil { + dir.DecRef() + return nil, err + } + d.IncLinks(1) + return dirVFSD, nil +} + +func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { + creds := auth.CredentialsFromContext(ctx) + f := d.fs.newFile(creds, "") + fVFSD := f.VFSDentry() + if err := d.OrderedChildren.Insert(name, fVFSD); err != nil { + f.DecRef() + return nil, err + } + return fVFSD, nil +} + +func (*dir) NewLink(context.Context, string, kernfs.Inode) (*vfs.Dentry, error) { + return nil, syserror.EPERM +} + +func (*dir) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { + return nil, syserror.EPERM +} + +func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { + return nil, syserror.EPERM +} + +func (fst *fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + fs := &filesystem{} + fs.Init(vfsObj) + root := fst.rootFn(creds, fs) + return fs.VFSFilesystem(), root.VFSDentry(), nil +} + +// -------------------- Remainder of the file are test cases -------------------- + +func TestBasic(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + "file1": fs.newFile(creds, staticFileContent), + }) + }) + sys.GetDentryOrDie(sys.PathOpAtRoot("file1")).DecRef() +} + +func TestMkdirGetDentry(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + "dir1": fs.newDir(creds, 0755, nil), + }) + }) + + pop := sys.PathOpAtRoot("dir1/a new directory") + if err := sys.vfs.MkdirAt(sys.ctx, sys.creds, &pop, &vfs.MkdirOptions{Mode: 0755}); err != nil { + t.Fatalf("MkdirAt for PathOperation %+v failed: %v", pop, err) + } + sys.GetDentryOrDie(pop).DecRef() +} + +func TestReadStaticFile(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + "file1": fs.newFile(creds, staticFileContent), + }) + }) + + pop := sys.PathOpAtRoot("file1") + fd, err := sys.vfs.OpenAt(sys.ctx, sys.creds, &pop, &vfs.OpenOptions{}) + if err != nil { + sys.t.Fatalf("OpenAt for PathOperation %+v failed: %v", pop, err) + } + defer fd.DecRef() + + content, err := sys.ReadToEnd(fd) + if err != nil { + sys.t.Fatalf("Read failed: %v", err) + } + if diff := cmp.Diff(staticFileContent, content); diff != "" { + sys.t.Fatalf("Read returned unexpected data:\n--- want\n+++ got\n%v", diff) + } +} + +func TestCreateNewFileInStaticDir(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + "dir1": fs.newDir(creds, 0755, nil), + }) + }) + + pop := sys.PathOpAtRoot("dir1/newfile") + opts := &vfs.OpenOptions{Flags: linux.O_CREAT | linux.O_EXCL, Mode: defaultMode} + fd, err := sys.vfs.OpenAt(sys.ctx, sys.creds, &pop, opts) + if err != nil { + sys.t.Fatalf("OpenAt(pop:%+v, opts:%+v) failed: %v", pop, opts, err) + } + + // Close the file. The file should persist. + fd.DecRef() + + fd, err = sys.vfs.OpenAt(sys.ctx, sys.creds, &pop, &vfs.OpenOptions{}) + if err != nil { + sys.t.Fatalf("OpenAt(pop:%+v) = %+v failed: %v", pop, fd, err) + } + fd.DecRef() +} + +// direntCollector provides an implementation for vfs.IterDirentsCallback for +// testing. It simply iterates to the end of a given directory FD and collects +// all dirents emitted by the callback. +type direntCollector struct { + mu sync.Mutex + dirents map[string]vfs.Dirent +} + +// Handle implements vfs.IterDirentsCallback.Handle. +func (d *direntCollector) Handle(dirent vfs.Dirent) bool { + d.mu.Lock() + if d.dirents == nil { + d.dirents = make(map[string]vfs.Dirent) + } + d.dirents[dirent.Name] = dirent + d.mu.Unlock() + return true +} + +// count returns the number of dirents currently in the collector. +func (d *direntCollector) count() int { + d.mu.Lock() + defer d.mu.Unlock() + return len(d.dirents) +} + +// contains checks whether the collector has a dirent with the given name and +// type. +func (d *direntCollector) contains(name string, typ uint8) error { + d.mu.Lock() + defer d.mu.Unlock() + dirent, ok := d.dirents[name] + if !ok { + return fmt.Errorf("No dirent named %q found", name) + } + if dirent.Type != typ { + return fmt.Errorf("Dirent named %q found, but was expecting type %d, got: %+v", name, typ, dirent) + } + return nil +} + +func TestDirFDReadWrite(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, nil) + }) + + pop := sys.PathOpAtRoot("/") + fd, err := sys.vfs.OpenAt(sys.ctx, sys.creds, &pop, &vfs.OpenOptions{}) + if err != nil { + sys.t.Fatalf("OpenAt for PathOperation %+v failed: %v", pop, err) + } + defer fd.DecRef() + + // Read/Write should fail for directory FDs. + if _, err := fd.Read(sys.ctx, usermem.BytesIOSequence([]byte{}), vfs.ReadOptions{}); err != syserror.EISDIR { + sys.t.Fatalf("Read for directory FD failed with unexpected error: %v", err) + } + if _, err := fd.Write(sys.ctx, usermem.BytesIOSequence([]byte{}), vfs.WriteOptions{}); err != syserror.EISDIR { + sys.t.Fatalf("Wrire for directory FD failed with unexpected error: %v", err) + } +} + +func TestDirFDIterDirents(t *testing.T) { + sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { + return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + // Fill root with nodes backed by various inode implementations. + "dir1": fs.newReadonlyDir(creds, 0755, nil), + "dir2": fs.newDir(creds, 0755, map[string]*kernfs.Dentry{ + "dir3": fs.newDir(creds, 0755, nil), + }), + "file1": fs.newFile(creds, staticFileContent), + }) + }) + + pop := sys.PathOpAtRoot("/") + fd, err := sys.vfs.OpenAt(sys.ctx, sys.creds, &pop, &vfs.OpenOptions{}) + if err != nil { + sys.t.Fatalf("OpenAt for PathOperation %+v failed: %v", pop, err) + } + defer fd.DecRef() + + collector := &direntCollector{} + if err := fd.IterDirents(sys.ctx, collector); err != nil { + sys.t.Fatalf("IterDirent failed: %v", err) + } + + // Root directory should contain ".", ".." and 3 children: + if collector.count() != 5 { + sys.t.Fatalf("IterDirent returned too many dirents") + } + for _, dirName := range []string{".", "..", "dir1", "dir2"} { + if err := collector.contains(dirName, linux.DT_DIR); err != nil { + sys.t.Fatalf("IterDirent had unexpected results: %v", err) + } + } + if err := collector.contains("file1", linux.DT_REG); err != nil { + sys.t.Fatalf("IterDirent had unexpected results: %v", err) + } + +} diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go new file mode 100644 index 000000000..068063f4e --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -0,0 +1,45 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" +) + +type staticSymlink struct { + InodeAttrs + InodeNoopRefCount + InodeSymlink + + target string +} + +var _ Inode = (*staticSymlink)(nil) + +// NewStaticSymlink creates a new symlink file pointing to 'target'. +func NewStaticSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, target string) *Dentry { + inode := &staticSymlink{target: target} + inode.Init(creds, ino, linux.ModeSymlink|perm) + + d := &Dentry{} + d.Init(inode) + return d +} + +func (s *staticSymlink) Readlink(_ context.Context) (string, error) { + return s.target, nil +} diff --git a/pkg/sentry/fsimpl/memfs/filesystem.go b/pkg/sentry/fsimpl/memfs/filesystem.go deleted file mode 100644 index 08a9cb8ef..000000000 --- a/pkg/sentry/fsimpl/memfs/filesystem.go +++ /dev/null @@ -1,584 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "fmt" - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -// stepLocked resolves rp.Component() in parent directory vfsd. -// -// stepLocked is loosely analogous to fs/namei.c:walk_component(). -// -// Preconditions: filesystem.mu must be locked. !rp.Done(). inode == -// vfsd.Impl().(*dentry).inode. -func stepLocked(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode) (*vfs.Dentry, *inode, error) { - if !inode.isDir() { - return nil, nil, syserror.ENOTDIR - } - if err := inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { - return nil, nil, err - } -afterSymlink: - nextVFSD, err := rp.ResolveComponent(vfsd) - if err != nil { - return nil, nil, err - } - if nextVFSD == nil { - // Since the Dentry tree is the sole source of truth for memfs, if it's - // not in the Dentry tree, it doesn't exist. - return nil, nil, syserror.ENOENT - } - nextInode := nextVFSD.Impl().(*dentry).inode - if symlink, ok := nextInode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { - // TODO: symlink traversals update access time - if err := rp.HandleSymlink(symlink.target); err != nil { - return nil, nil, err - } - goto afterSymlink // don't check the current directory again - } - rp.Advance() - return nextVFSD, nextInode, nil -} - -// walkExistingLocked resolves rp to an existing file. -// -// walkExistingLocked is loosely analogous to Linux's -// fs/namei.c:path_lookupat(). -// -// Preconditions: filesystem.mu must be locked. -func walkExistingLocked(rp *vfs.ResolvingPath) (*vfs.Dentry, *inode, error) { - vfsd := rp.Start() - inode := vfsd.Impl().(*dentry).inode - for !rp.Done() { - var err error - vfsd, inode, err = stepLocked(rp, vfsd, inode) - if err != nil { - return nil, nil, err - } - } - if rp.MustBeDir() && !inode.isDir() { - return nil, nil, syserror.ENOTDIR - } - return vfsd, inode, nil -} - -// walkParentDirLocked resolves all but the last path component of rp to an -// existing directory. It does not check that the returned directory is -// searchable by the provider of rp. -// -// walkParentDirLocked is loosely analogous to Linux's -// fs/namei.c:path_parentat(). -// -// Preconditions: filesystem.mu must be locked. !rp.Done(). -func walkParentDirLocked(rp *vfs.ResolvingPath) (*vfs.Dentry, *inode, error) { - vfsd := rp.Start() - inode := vfsd.Impl().(*dentry).inode - for !rp.Final() { - var err error - vfsd, inode, err = stepLocked(rp, vfsd, inode) - if err != nil { - return nil, nil, err - } - } - if !inode.isDir() { - return nil, nil, syserror.ENOTDIR - } - return vfsd, inode, nil -} - -// checkCreateLocked checks that a file named rp.Component() may be created in -// directory parentVFSD, then returns rp.Component(). -// -// Preconditions: filesystem.mu must be locked. parentInode == -// parentVFSD.Impl().(*dentry).inode. parentInode.isDir() == true. -func checkCreateLocked(rp *vfs.ResolvingPath, parentVFSD *vfs.Dentry, parentInode *inode) (string, error) { - if err := parentInode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { - return "", err - } - pc := rp.Component() - if pc == "." || pc == ".." { - return "", syserror.EEXIST - } - childVFSD, err := rp.ResolveChild(parentVFSD, pc) - if err != nil { - return "", err - } - if childVFSD != nil { - return "", syserror.EEXIST - } - if parentVFSD.IsDisowned() { - return "", syserror.ENOENT - } - return pc, nil -} - -// checkDeleteLocked checks that the file represented by vfsd may be deleted. -func checkDeleteLocked(vfsd *vfs.Dentry) error { - parentVFSD := vfsd.Parent() - if parentVFSD == nil { - return syserror.EBUSY - } - if parentVFSD.IsDisowned() { - return syserror.ENOENT - } - return nil -} - -// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. -func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { - fs.mu.RLock() - defer fs.mu.RUnlock() - vfsd, inode, err := walkExistingLocked(rp) - if err != nil { - return nil, err - } - if opts.CheckSearchable { - if !inode.isDir() { - return nil, syserror.ENOTDIR - } - if err := inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { - return nil, err - } - } - inode.incRef() - return vfsd, nil -} - -// LinkAt implements vfs.FilesystemImpl.LinkAt. -func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { - if rp.Done() { - return syserror.EEXIST - } - fs.mu.Lock() - defer fs.mu.Unlock() - parentVFSD, parentInode, err := walkParentDirLocked(rp) - if err != nil { - return err - } - pc, err := checkCreateLocked(rp, parentVFSD, parentInode) - if err != nil { - return err - } - if rp.Mount() != vd.Mount() { - return syserror.EXDEV - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - d := vd.Dentry().Impl().(*dentry) - if d.inode.isDir() { - return syserror.EPERM - } - d.inode.incLinksLocked() - child := fs.newDentry(d.inode) - parentVFSD.InsertChild(&child.vfsd, pc) - parentInode.impl.(*directory).childList.PushBack(child) - return nil -} - -// MkdirAt implements vfs.FilesystemImpl.MkdirAt. -func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { - if rp.Done() { - return syserror.EEXIST - } - fs.mu.Lock() - defer fs.mu.Unlock() - parentVFSD, parentInode, err := walkParentDirLocked(rp) - if err != nil { - return err - } - pc, err := checkCreateLocked(rp, parentVFSD, parentInode) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) - parentVFSD.InsertChild(&child.vfsd, pc) - parentInode.impl.(*directory).childList.PushBack(child) - parentInode.incLinksLocked() // from child's ".." - return nil -} - -// MknodAt implements vfs.FilesystemImpl.MknodAt. -func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { - if rp.Done() { - return syserror.EEXIST - } - fs.mu.Lock() - defer fs.mu.Unlock() - parentVFSD, parentInode, err := walkParentDirLocked(rp) - if err != nil { - return err - } - pc, err := checkCreateLocked(rp, parentVFSD, parentInode) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - - switch opts.Mode.FileType() { - case 0: - // "Zero file type is equivalent to type S_IFREG." - mknod(2) - fallthrough - case linux.ModeRegular: - // TODO(b/138862511): Implement. - return syserror.EINVAL - - case linux.ModeNamedPipe: - child := fs.newDentry(fs.newNamedPipe(rp.Credentials(), opts.Mode)) - parentVFSD.InsertChild(&child.vfsd, pc) - parentInode.impl.(*directory).childList.PushBack(child) - return nil - - case linux.ModeSocket: - // TODO(b/138862511): Implement. - return syserror.EINVAL - - case linux.ModeCharacterDevice: - fallthrough - case linux.ModeBlockDevice: - // TODO(b/72101894): We don't support creating block or character - // devices at the moment. - // - // When we start supporting block and character devices, we'll - // need to check for CAP_MKNOD here. - return syserror.EPERM - - default: - // "EINVAL - mode requested creation of something other than a - // regular file, device special file, FIFO or socket." - mknod(2) - return syserror.EINVAL - } -} - -// OpenAt implements vfs.FilesystemImpl.OpenAt. -func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - // Filter out flags that are not supported by memfs. O_DIRECTORY and - // O_NOFOLLOW have no effect here (they're handled by VFS by setting - // appropriate bits in rp), but are returned by - // FileDescriptionImpl.StatusFlags(). O_NONBLOCK is supported only by - // pipes. - opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NONBLOCK - - if opts.Flags&linux.O_CREAT == 0 { - fs.mu.RLock() - defer fs.mu.RUnlock() - vfsd, inode, err := walkExistingLocked(rp) - if err != nil { - return nil, err - } - return inode.open(ctx, rp, vfsd, opts.Flags, false) - } - - mustCreate := opts.Flags&linux.O_EXCL != 0 - vfsd := rp.Start() - inode := vfsd.Impl().(*dentry).inode - fs.mu.Lock() - defer fs.mu.Unlock() - if rp.Done() { - if rp.MustBeDir() { - return nil, syserror.EISDIR - } - if mustCreate { - return nil, syserror.EEXIST - } - return inode.open(ctx, rp, vfsd, opts.Flags, false) - } -afterTrailingSymlink: - // Walk to the parent directory of the last path component. - for !rp.Final() { - var err error - vfsd, inode, err = stepLocked(rp, vfsd, inode) - if err != nil { - return nil, err - } - } - if !inode.isDir() { - return nil, syserror.ENOTDIR - } - // Check for search permission in the parent directory. - if err := inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { - return nil, err - } - // Reject attempts to open directories with O_CREAT. - if rp.MustBeDir() { - return nil, syserror.EISDIR - } - pc := rp.Component() - if pc == "." || pc == ".." { - return nil, syserror.EISDIR - } - // Determine whether or not we need to create a file. - childVFSD, err := rp.ResolveChild(vfsd, pc) - if err != nil { - return nil, err - } - if childVFSD == nil { - // Already checked for searchability above; now check for writability. - if err := inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { - return nil, err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return nil, err - } - defer rp.Mount().EndWrite() - // Create and open the child. - childInode := fs.newRegularFile(rp.Credentials(), opts.Mode) - child := fs.newDentry(childInode) - vfsd.InsertChild(&child.vfsd, pc) - inode.impl.(*directory).childList.PushBack(child) - return childInode.open(ctx, rp, &child.vfsd, opts.Flags, true) - } - // Open existing file or follow symlink. - if mustCreate { - return nil, syserror.EEXIST - } - childInode := childVFSD.Impl().(*dentry).inode - if symlink, ok := childInode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { - // TODO: symlink traversals update access time - if err := rp.HandleSymlink(symlink.target); err != nil { - return nil, err - } - // rp.Final() may no longer be true since we now need to resolve the - // symlink target. - goto afterTrailingSymlink - } - return childInode.open(ctx, rp, childVFSD, opts.Flags, false) -} - -func (i *inode) open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32, afterCreate bool) (*vfs.FileDescription, error) { - ats := vfs.AccessTypesForOpenFlags(flags) - if !afterCreate { - if err := i.checkPermissions(rp.Credentials(), ats, i.isDir()); err != nil { - return nil, err - } - } - mnt := rp.Mount() - switch impl := i.impl.(type) { - case *regularFile: - var fd regularFileFD - fd.flags = flags - fd.readable = vfs.MayReadFileWithOpenFlags(flags) - fd.writable = vfs.MayWriteFileWithOpenFlags(flags) - if fd.writable { - if err := mnt.CheckBeginWrite(); err != nil { - return nil, err - } - // mnt.EndWrite() is called by regularFileFD.Release(). - } - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) - if flags&linux.O_TRUNC != 0 { - impl.mu.Lock() - impl.data = impl.data[:0] - atomic.StoreInt64(&impl.dataLen, 0) - impl.mu.Unlock() - } - return &fd.vfsfd, nil - case *directory: - // Can't open directories writably. - if ats&vfs.MayWrite != 0 { - return nil, syserror.EISDIR - } - var fd directoryFD - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) - fd.flags = flags - return &fd.vfsfd, nil - case *symlink: - // Can't open symlinks without O_PATH (which is unimplemented). - return nil, syserror.ELOOP - case *namedPipe: - return newNamedPipeFD(ctx, impl, rp, vfsd, flags) - default: - panic(fmt.Sprintf("unknown inode type: %T", i.impl)) - } -} - -// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. -func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { - fs.mu.RLock() - _, inode, err := walkExistingLocked(rp) - fs.mu.RUnlock() - if err != nil { - return "", err - } - symlink, ok := inode.impl.(*symlink) - if !ok { - return "", syserror.EINVAL - } - return symlink.target, nil -} - -// RenameAt implements vfs.FilesystemImpl.RenameAt. -func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry, opts vfs.RenameOptions) error { - if rp.Done() { - return syserror.ENOENT - } - fs.mu.Lock() - defer fs.mu.Unlock() - parentVFSD, parentInode, err := walkParentDirLocked(rp) - if err != nil { - return err - } - _, err = checkCreateLocked(rp, parentVFSD, parentInode) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - // TODO: actually implement RenameAt - return syserror.EPERM -} - -// RmdirAt implements vfs.FilesystemImpl.RmdirAt. -func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { - fs.mu.Lock() - defer fs.mu.Unlock() - vfsd, inode, err := walkExistingLocked(rp) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - if err := checkDeleteLocked(vfsd); err != nil { - return err - } - if !inode.isDir() { - return syserror.ENOTDIR - } - if vfsd.HasChildren() { - return syserror.ENOTEMPTY - } - if err := rp.VirtualFilesystem().DeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { - return err - } - // Remove from parent directory's childList. - vfsd.Parent().Impl().(*dentry).inode.impl.(*directory).childList.Remove(vfsd.Impl().(*dentry)) - inode.decRef() - return nil -} - -// SetStatAt implements vfs.FilesystemImpl.SetStatAt. -func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { - fs.mu.RLock() - _, _, err := walkExistingLocked(rp) - fs.mu.RUnlock() - if err != nil { - return err - } - if opts.Stat.Mask == 0 { - return nil - } - // TODO: implement inode.setStat - return syserror.EPERM -} - -// StatAt implements vfs.FilesystemImpl.StatAt. -func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { - fs.mu.RLock() - _, inode, err := walkExistingLocked(rp) - fs.mu.RUnlock() - if err != nil { - return linux.Statx{}, err - } - var stat linux.Statx - inode.statTo(&stat) - return stat, nil -} - -// StatFSAt implements vfs.FilesystemImpl.StatFSAt. -func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { - fs.mu.RLock() - _, _, err := walkExistingLocked(rp) - fs.mu.RUnlock() - if err != nil { - return linux.Statfs{}, err - } - // TODO: actually implement statfs - return linux.Statfs{}, syserror.ENOSYS -} - -// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. -func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { - if rp.Done() { - return syserror.EEXIST - } - fs.mu.Lock() - defer fs.mu.Unlock() - parentVFSD, parentInode, err := walkParentDirLocked(rp) - if err != nil { - return err - } - pc, err := checkCreateLocked(rp, parentVFSD, parentInode) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) - parentVFSD.InsertChild(&child.vfsd, pc) - parentInode.impl.(*directory).childList.PushBack(child) - return nil -} - -// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. -func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { - fs.mu.Lock() - defer fs.mu.Unlock() - vfsd, inode, err := walkExistingLocked(rp) - if err != nil { - return err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return err - } - defer rp.Mount().EndWrite() - if err := checkDeleteLocked(vfsd); err != nil { - return err - } - if inode.isDir() { - return syserror.EISDIR - } - if err := rp.VirtualFilesystem().DeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { - return err - } - // Remove from parent directory's childList. - vfsd.Parent().Impl().(*dentry).inode.impl.(*directory).childList.Remove(vfsd.Impl().(*dentry)) - inode.decLinksLocked() - return nil -} diff --git a/pkg/sentry/fsimpl/memfs/regular_file.go b/pkg/sentry/fsimpl/memfs/regular_file.go deleted file mode 100644 index b7f4853b3..000000000 --- a/pkg/sentry/fsimpl/memfs/regular_file.go +++ /dev/null @@ -1,154 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "io" - "sync" - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/usermem" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -type regularFile struct { - inode inode - - mu sync.RWMutex - data []byte - // dataLen is len(data), but accessed using atomic memory operations to - // avoid locking in inode.stat(). - dataLen int64 -} - -func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { - file := ®ularFile{} - file.inode.init(file, fs, creds, mode) - file.inode.nlink = 1 // from parent directory - return &file.inode -} - -type regularFileFD struct { - fileDescription - - // These are immutable. - readable bool - writable bool - - // off is the file offset. off is accessed using atomic memory operations. - // offMu serializes operations that may mutate off. - off int64 - offMu sync.Mutex -} - -// Release implements vfs.FileDescriptionImpl.Release. -func (fd *regularFileFD) Release() { - if fd.writable { - fd.vfsfd.VirtualDentry().Mount().EndWrite() - } -} - -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { - if !fd.readable { - return 0, syserror.EINVAL - } - f := fd.inode().impl.(*regularFile) - f.mu.RLock() - if offset >= int64(len(f.data)) { - f.mu.RUnlock() - return 0, io.EOF - } - n, err := dst.CopyOut(ctx, f.data[offset:]) - f.mu.RUnlock() - return int64(n), err -} - -// Read implements vfs.FileDescriptionImpl.Read. -func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - fd.offMu.Lock() - n, err := fd.PRead(ctx, dst, fd.off, opts) - fd.off += n - fd.offMu.Unlock() - return n, err -} - -// PWrite implements vfs.FileDescriptionImpl.PWrite. -func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { - if !fd.writable { - return 0, syserror.EINVAL - } - if offset < 0 { - return 0, syserror.EINVAL - } - srclen := src.NumBytes() - if srclen == 0 { - return 0, nil - } - f := fd.inode().impl.(*regularFile) - f.mu.Lock() - end := offset + srclen - if end < offset { - // Overflow. - f.mu.Unlock() - return 0, syserror.EFBIG - } - if end > f.dataLen { - f.data = append(f.data, make([]byte, end-f.dataLen)...) - atomic.StoreInt64(&f.dataLen, end) - } - n, err := src.CopyIn(ctx, f.data[offset:end]) - f.mu.Unlock() - return int64(n), err -} - -// Write implements vfs.FileDescriptionImpl.Write. -func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { - fd.offMu.Lock() - n, err := fd.PWrite(ctx, src, fd.off, opts) - fd.off += n - fd.offMu.Unlock() - return n, err -} - -// Seek implements vfs.FileDescriptionImpl.Seek. -func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { - fd.offMu.Lock() - defer fd.offMu.Unlock() - switch whence { - case linux.SEEK_SET: - // use offset as specified - case linux.SEEK_CUR: - offset += fd.off - case linux.SEEK_END: - offset += atomic.LoadInt64(&fd.inode().impl.(*regularFile).dataLen) - default: - return 0, syserror.EINVAL - } - if offset < 0 { - return 0, syserror.EINVAL - } - fd.off = offset - return offset, nil -} - -// Sync implements vfs.FileDescriptionImpl.Sync. -func (fd *regularFileFD) Sync(ctx context.Context) error { - return nil -} diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index ade6ac946..1f44b3217 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -6,15 +6,17 @@ package(licenses = ["notice"]) go_library( name = "proc", srcs = [ - "filesystems.go", + "filesystem.go", "loadavg.go", "meminfo.go", "mounts.go", "net.go", - "proc.go", "stat.go", "sys.go", "task.go", + "task_files.go", + "tasks.go", + "tasks_files.go", "version.go", ], importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/proc", @@ -24,8 +26,10 @@ go_library( "//pkg/log", "//pkg/sentry/context", "//pkg/sentry/fs", + "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/inet", "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", "//pkg/sentry/limits", "//pkg/sentry/mm", "//pkg/sentry/socket", @@ -34,17 +38,40 @@ go_library( "//pkg/sentry/usage", "//pkg/sentry/usermem", "//pkg/sentry/vfs", + "//pkg/syserror", ], ) go_test( name = "proc_test", size = "small", - srcs = ["net_test.go"], + srcs = [ + "boot_test.go", + "net_test.go", + "tasks_test.go", + ], embed = [":proc"], deps = [ "//pkg/abi/linux", + "//pkg/cpuid", + "//pkg/fspath", + "//pkg/memutil", + "//pkg/sentry/context", "//pkg/sentry/context/contexttest", + "//pkg/sentry/fs", "//pkg/sentry/inet", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/sched", + "//pkg/sentry/limits", + "//pkg/sentry/loader", + "//pkg/sentry/pgalloc", + "//pkg/sentry/platform", + "//pkg/sentry/platform/kvm", + "//pkg/sentry/platform/ptrace", + "//pkg/sentry/time", + "//pkg/sentry/usermem", + "//pkg/sentry/vfs", + "//pkg/syserror", ], ) diff --git a/pkg/sentry/fsimpl/proc/boot_test.go b/pkg/sentry/fsimpl/proc/boot_test.go new file mode 100644 index 000000000..84a93ee56 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/boot_test.go @@ -0,0 +1,149 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "flag" + "fmt" + "os" + "runtime" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/memutil" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "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/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/time" + + // Platforms are plugable. + _ "gvisor.dev/gvisor/pkg/sentry/platform/kvm" + _ "gvisor.dev/gvisor/pkg/sentry/platform/ptrace" +) + +var ( + platformFlag = flag.String("platform", "ptrace", "specify which platform to use") +) + +// boot initializes a new bare bones kernel for test. +func boot() (*kernel.Kernel, error) { + platformCtr, err := platform.Lookup(*platformFlag) + if err != nil { + return nil, fmt.Errorf("platform not found: %v", err) + } + deviceFile, err := platformCtr.OpenDevice() + if err != nil { + return nil, fmt.Errorf("creating platform: %v", err) + } + plat, err := platformCtr.New(deviceFile) + if err != nil { + return nil, fmt.Errorf("creating platform: %v", err) + } + + k := &kernel.Kernel{ + Platform: plat, + } + + mf, err := createMemoryFile() + if err != nil { + return nil, err + } + k.SetMemoryFile(mf) + + // Pass k as the platform since it is savable, unlike the actual platform. + vdso, err := loader.PrepareVDSO(nil, k) + if err != nil { + return nil, fmt.Errorf("creating vdso: %v", err) + } + + // Create timekeeper. + tk, err := kernel.NewTimekeeper(k, vdso.ParamPage.FileRange()) + if err != nil { + return nil, fmt.Errorf("creating timekeeper: %v", err) + } + tk.SetClocks(time.NewCalibratedClocks()) + + creds := auth.NewRootCredentials(auth.NewRootUserNamespace()) + + // Initiate the Kernel object, which is required by the Context passed + // to createVFS in order to mount (among other things) procfs. + if err = k.Init(kernel.InitKernelArgs{ + ApplicationCores: uint(runtime.GOMAXPROCS(-1)), + FeatureSet: cpuid.HostFeatureSet(), + Timekeeper: tk, + RootUserNamespace: creds.UserNamespace, + Vdso: vdso, + RootUTSNamespace: kernel.NewUTSNamespace("hostname", "domain", creds.UserNamespace), + RootIPCNamespace: kernel.NewIPCNamespace(creds.UserNamespace), + RootAbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), + PIDNamespace: kernel.NewRootPIDNamespace(creds.UserNamespace), + }); err != nil { + return nil, fmt.Errorf("initializing kernel: %v", err) + } + + ctx := k.SupervisorContext() + + // Create mount namespace without root as it's the minimum required to create + // the global thread group. + mntns, err := fs.NewMountNamespace(ctx, nil) + if err != nil { + return nil, err + } + ls, err := limits.NewLinuxLimitSet() + if err != nil { + return nil, err + } + tg := k.NewThreadGroup(mntns, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) + k.TestOnly_SetGlobalInit(tg) + + return k, nil +} + +// createTask creates a new bare bones task for tests. +func createTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kernel.Task, error) { + k := kernel.KernelFromContext(ctx) + config := &kernel.TaskConfig{ + Kernel: k, + ThreadGroup: tc, + TaskContext: &kernel.TaskContext{Name: name}, + Credentials: auth.CredentialsFromContext(ctx), + AllowedCPUMask: sched.NewFullCPUSet(k.ApplicationCores()), + UTSNamespace: kernel.UTSNamespaceFromContext(ctx), + IPCNamespace: kernel.IPCNamespaceFromContext(ctx), + AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), + } + return k.TaskSet().NewTask(config) +} + +func createMemoryFile() (*pgalloc.MemoryFile, error) { + const memfileName = "test-memory" + memfd, err := memutil.CreateMemFD(memfileName, 0) + if err != nil { + return nil, fmt.Errorf("error creating memfd: %v", err) + } + memfile := os.NewFile(uintptr(memfd), memfileName) + mf, err := pgalloc.NewMemoryFile(memfile, pgalloc.MemoryFileOpts{}) + if err != nil { + memfile.Close() + return nil, fmt.Errorf("error creating pgalloc.MemoryFile: %v", err) + } + return mf, nil +} diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go new file mode 100644 index 000000000..d09182c77 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -0,0 +1,69 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package proc implements a partial in-memory file system for procfs. +package proc + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// procFSType is the factory class for procfs. +// +// +stateify savable +type procFSType struct{} + +var _ vfs.FilesystemType = (*procFSType)(nil) + +// GetFilesystem implements vfs.FilesystemType. +func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + k := kernel.KernelFromContext(ctx) + if k == nil { + return nil, nil, fmt.Errorf("procfs requires a kernel") + } + pidns := kernel.PIDNamespaceFromContext(ctx) + if pidns == nil { + return nil, nil, fmt.Errorf("procfs requires a PID namespace") + } + + procfs := &kernfs.Filesystem{} + procfs.VFSFilesystem().Init(vfsObj, procfs) + + _, dentry := newTasksInode(procfs, k, pidns) + return procfs.VFSFilesystem(), dentry.VFSDentry(), nil +} + +// dynamicInode is an overfitted interface for common Inodes with +// dynamicByteSource types used in procfs. +type dynamicInode interface { + kernfs.Inode + vfs.DynamicBytesSource + + Init(creds *auth.Credentials, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) +} + +func newDentry(creds *auth.Credentials, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { + inode.Init(creds, ino, inode, perm) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} diff --git a/pkg/sentry/fsimpl/proc/loadavg.go b/pkg/sentry/fsimpl/proc/loadavg.go index 9135afef1..5351d86e8 100644 --- a/pkg/sentry/fsimpl/proc/loadavg.go +++ b/pkg/sentry/fsimpl/proc/loadavg.go @@ -19,15 +19,17 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" ) // loadavgData backs /proc/loadavg. // // +stateify savable -type loadavgData struct{} +type loadavgData struct { + kernfs.DynamicBytesFile +} -var _ vfs.DynamicBytesSource = (*loadavgData)(nil) +var _ dynamicInode = (*loadavgData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. func (d *loadavgData) Generate(ctx context.Context, buf *bytes.Buffer) error { diff --git a/pkg/sentry/fsimpl/proc/meminfo.go b/pkg/sentry/fsimpl/proc/meminfo.go index 9a827cd66..cbdd4f3fc 100644 --- a/pkg/sentry/fsimpl/proc/meminfo.go +++ b/pkg/sentry/fsimpl/proc/meminfo.go @@ -19,21 +19,23 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/usermem" - "gvisor.dev/gvisor/pkg/sentry/vfs" ) // meminfoData implements vfs.DynamicBytesSource for /proc/meminfo. // // +stateify savable type meminfoData struct { + kernfs.DynamicBytesFile + // k is the owning Kernel. k *kernel.Kernel } -var _ vfs.DynamicBytesSource = (*meminfoData)(nil) +var _ dynamicInode = (*meminfoData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. func (d *meminfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { diff --git a/pkg/sentry/fsimpl/proc/mounts.go b/pkg/sentry/fsimpl/proc/mounts.go index e81b1e910..8683cf677 100644 --- a/pkg/sentry/fsimpl/proc/mounts.go +++ b/pkg/sentry/fsimpl/proc/mounts.go @@ -16,7 +16,7 @@ package proc import "gvisor.dev/gvisor/pkg/sentry/kernel" -// TODO(b/138862512): Implement mountInfoFile and mountsFile. +// TODO(gvisor.dev/issue/1195): Implement mountInfoFile and mountsFile. // mountInfoFile implements vfs.DynamicBytesSource for /proc/[pid]/mountinfo. // diff --git a/pkg/sentry/fsimpl/proc/stat.go b/pkg/sentry/fsimpl/proc/stat.go index 720db3828..50894a534 100644 --- a/pkg/sentry/fsimpl/proc/stat.go +++ b/pkg/sentry/fsimpl/proc/stat.go @@ -20,8 +20,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/vfs" ) // cpuStats contains the breakdown of CPU time for /proc/stat. @@ -66,11 +66,13 @@ func (c cpuStats) String() string { // // +stateify savable type statData struct { + kernfs.DynamicBytesFile + // k is the owning Kernel. k *kernel.Kernel } -var _ vfs.DynamicBytesSource = (*statData)(nil) +var _ dynamicInode = (*statData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. func (s *statData) Generate(ctx context.Context, buf *bytes.Buffer) error { diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index c46e05c3a..11a64c777 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -15,247 +15,176 @@ package proc import ( - "bytes" - "fmt" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" - "gvisor.dev/gvisor/pkg/sentry/usage" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" ) -// mapsCommon is embedded by mapsData and smapsData. -type mapsCommon struct { - t *kernel.Task -} - -// mm gets the kernel task's MemoryManager. No additional reference is taken on -// mm here. This is safe because MemoryManager.destroy is required to leave the -// MemoryManager in a state where it's still usable as a DynamicBytesSource. -func (md *mapsCommon) mm() *mm.MemoryManager { - var tmm *mm.MemoryManager - md.t.WithMuLocked(func(t *kernel.Task) { - if mm := t.MemoryManager(); mm != nil { - tmm = mm - } - }) - return tmm -} - -// mapsData implements vfs.DynamicBytesSource for /proc/[pid]/maps. +// taskInode represents the inode for /proc/PID/ directory. // // +stateify savable -type mapsData struct { - mapsCommon +type taskInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNoDynamicLookup + kernfs.InodeAttrs + kernfs.OrderedChildren + + task *kernel.Task } -var _ vfs.DynamicBytesSource = (*mapsData)(nil) - -// Generate implements vfs.DynamicBytesSource.Generate. -func (md *mapsData) Generate(ctx context.Context, buf *bytes.Buffer) error { - if mm := md.mm(); mm != nil { - mm.ReadMapsDataInto(ctx, buf) +var _ kernfs.Inode = (*taskInode)(nil) + +func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool) *kernfs.Dentry { + contents := map[string]*kernfs.Dentry{ + //"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": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, newIO(task, isThreadGroup)), + "maps": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, &mapsData{task: task}), + //"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), + //"mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + //"ns": newNamespaceDir(t, msrc), + "smaps": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, &smapsData{task: task}), + "stat": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, &taskStatData{t: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, &statmData{t: task}), + "status": newTaskOwnedFile(task, inoGen.NextIno(), defaultPermission, &statusData{t: task, pidns: pidns}), + //"uid_map": newUIDMap(t, msrc), } - return nil -} - -// smapsData implements vfs.DynamicBytesSource for /proc/[pid]/smaps. -// -// +stateify savable -type smapsData struct { - mapsCommon -} - -var _ vfs.DynamicBytesSource = (*smapsData)(nil) - -// Generate implements vfs.DynamicBytesSource.Generate. -func (sd *smapsData) Generate(ctx context.Context, buf *bytes.Buffer) error { - if mm := sd.mm(); mm != nil { - mm.ReadSmapsDataInto(ctx, buf) + if isThreadGroup { + //contents["task"] = p.newSubtasks(t, msrc) } - return nil -} - -// +stateify savable -type taskStatData struct { - t *kernel.Task + //if len(p.cgroupControllers) > 0 { + // contents["cgroup"] = newCGroupInode(t, msrc, p.cgroupControllers) + //} - // If tgstats is true, accumulate fault stats (not implemented) and CPU - // time across all tasks in t's thread group. - tgstats bool + taskInode := &taskInode{task: task} + // Note: credentials are overridden by taskOwnedInode. + taskInode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555) - // pidns is the PID namespace associated with the proc filesystem that - // includes the file using this statData. - pidns *kernel.PIDNamespace -} - -var _ vfs.DynamicBytesSource = (*taskStatData)(nil) - -// Generate implements vfs.DynamicBytesSource.Generate. -func (s *taskStatData) Generate(ctx context.Context, buf *bytes.Buffer) error { - fmt.Fprintf(buf, "%d ", s.pidns.IDOfTask(s.t)) - fmt.Fprintf(buf, "(%s) ", s.t.Name()) - fmt.Fprintf(buf, "%c ", s.t.StateStatus()[0]) - ppid := kernel.ThreadID(0) - if parent := s.t.Parent(); parent != nil { - ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup()) - } - fmt.Fprintf(buf, "%d ", ppid) - fmt.Fprintf(buf, "%d ", s.pidns.IDOfProcessGroup(s.t.ThreadGroup().ProcessGroup())) - fmt.Fprintf(buf, "%d ", s.pidns.IDOfSession(s.t.ThreadGroup().Session())) - fmt.Fprintf(buf, "0 0 " /* tty_nr tpgid */) - fmt.Fprintf(buf, "0 " /* flags */) - fmt.Fprintf(buf, "0 0 0 0 " /* minflt cminflt majflt cmajflt */) - var cputime usage.CPUStats - if s.tgstats { - cputime = s.t.ThreadGroup().CPUStats() - } else { - cputime = s.t.CPUStats() - } - fmt.Fprintf(buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime)) - cputime = s.t.ThreadGroup().JoinedChildCPUStats() - fmt.Fprintf(buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime)) - fmt.Fprintf(buf, "%d %d ", s.t.Priority(), s.t.Niceness()) - fmt.Fprintf(buf, "%d ", s.t.ThreadGroup().Count()) + inode := &taskOwnedInode{Inode: taskInode, owner: task} + dentry := &kernfs.Dentry{} + dentry.Init(inode) - // itrealvalue. Since kernel 2.6.17, this field is no longer - // maintained, and is hard coded as 0. - fmt.Fprintf(buf, "0 ") + taskInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + links := taskInode.OrderedChildren.Populate(dentry, contents) + taskInode.IncLinks(links) - // Start time is relative to boot time, expressed in clock ticks. - fmt.Fprintf(buf, "%d ", linux.ClockTFromDuration(s.t.StartTime().Sub(s.t.Kernel().Timekeeper().BootTime()))) + return dentry +} - var vss, rss uint64 - s.t.WithMuLocked(func(t *kernel.Task) { - if mm := t.MemoryManager(); mm != nil { - vss = mm.VirtualMemorySize() - rss = mm.ResidentSetSize() - } - }) - fmt.Fprintf(buf, "%d %d ", vss, rss/usermem.PageSize) +// Valid implements kernfs.inodeDynamicLookup. This inode remains valid as long +// as the task is still running. When it's dead, another tasks with the same +// PID could replace it. +func (i *taskInode) Valid(ctx context.Context) bool { + return i.task.ExitState() != kernel.TaskExitDead +} - // rsslim. - fmt.Fprintf(buf, "%d ", s.t.ThreadGroup().Limits().Get(limits.Rss).Cur) +// Open implements kernfs.Inode. +func (i *taskInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, flags) + return fd.VFSFileDescription(), nil +} - fmt.Fprintf(buf, "0 0 0 0 0 " /* startcode endcode startstack kstkesp kstkeip */) - fmt.Fprintf(buf, "0 0 0 0 0 " /* signal blocked sigignore sigcatch wchan */) - fmt.Fprintf(buf, "0 0 " /* nswap cnswap */) - terminationSignal := linux.Signal(0) - if s.t == s.t.ThreadGroup().Leader() { - terminationSignal = s.t.ThreadGroup().TerminationSignal() +// 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 } - fmt.Fprintf(buf, "%d ", terminationSignal) - fmt.Fprintf(buf, "0 0 0 " /* processor rt_priority policy */) - fmt.Fprintf(buf, "0 0 0 " /* delayacct_blkio_ticks guest_time cguest_time */) - fmt.Fprintf(buf, "0 0 0 0 0 0 0 " /* start_data end_data start_brk arg_start arg_end env_start env_end */) - fmt.Fprintf(buf, "0\n" /* exit_code */) - return nil } -// statmData implements vfs.DynamicBytesSource for /proc/[pid]/statm. -// -// +stateify savable -type statmData struct { - t *kernel.Task +// taskOwnedInode implements kernfs.Inode and overrides inode owner with task +// effective user and group. +type taskOwnedInode struct { + kernfs.Inode + + // owner is the task that owns this inode. + owner *kernel.Task } -var _ vfs.DynamicBytesSource = (*statmData)(nil) +var _ kernfs.Inode = (*taskOwnedInode)(nil) -// Generate implements vfs.DynamicBytesSource.Generate. -func (s *statmData) Generate(ctx context.Context, buf *bytes.Buffer) error { - var vss, rss uint64 - s.t.WithMuLocked(func(t *kernel.Task) { - if mm := t.MemoryManager(); mm != nil { - vss = mm.VirtualMemorySize() - rss = mm.ResidentSetSize() - } - }) +func newTaskOwnedFile(task *kernel.Task, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { + // Note: credentials are overridden by taskOwnedInode. + inode.Init(task.Credentials(), ino, inode, perm) - fmt.Fprintf(buf, "%d %d 0 0 0 0 0\n", vss/usermem.PageSize, rss/usermem.PageSize) - return nil + taskInode := &taskOwnedInode{Inode: inode, owner: task} + d := &kernfs.Dentry{} + d.Init(taskInode) + return d } -// statusData implements vfs.DynamicBytesSource for /proc/[pid]/status. -// -// +stateify savable -type statusData struct { - t *kernel.Task - pidns *kernel.PIDNamespace +// 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 } -var _ vfs.DynamicBytesSource = (*statusData)(nil) +// CheckPermissions implements kernfs.Inode. +func (i *taskOwnedInode) CheckPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error { + mode := i.Mode() + uid, gid := i.getOwner(mode) + return vfs.GenericCheckPermissions( + creds, + ats, + mode.FileType() == linux.ModeDirectory, + uint16(mode), + uid, + gid, + ) +} -// Generate implements vfs.DynamicBytesSource.Generate. -func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { - fmt.Fprintf(buf, "Name:\t%s\n", s.t.Name()) - fmt.Fprintf(buf, "State:\t%s\n", s.t.StateStatus()) - fmt.Fprintf(buf, "Tgid:\t%d\n", s.pidns.IDOfThreadGroup(s.t.ThreadGroup())) - fmt.Fprintf(buf, "Pid:\t%d\n", s.pidns.IDOfTask(s.t)) - ppid := kernel.ThreadID(0) - if parent := s.t.Parent(); parent != nil { - ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup()) +func (i *taskOwnedInode) getOwner(mode linux.FileMode) (auth.KUID, auth.KGID) { + // By default, set the task owner as the file owner. + creds := i.owner.Credentials() + uid := creds.EffectiveKUID + gid := creds.EffectiveKGID + + // Linux doesn't apply dumpability adjustments to world readable/executable + // directories so that applications can stat /proc/PID to determine the + // effective UID of a process. See fs/proc/base.c:task_dump_owner. + if mode.FileType() == linux.ModeDirectory && mode.Permissions() == 0555 { + return uid, gid } - fmt.Fprintf(buf, "PPid:\t%d\n", ppid) - tpid := kernel.ThreadID(0) - if tracer := s.t.Tracer(); tracer != nil { - tpid = s.pidns.IDOfTask(tracer) + + // If the task is not dumpable, then root (in the namespace preferred) + // owns the file. + m := getMM(i.owner) + if m == nil { + return auth.RootKUID, auth.RootKGID } - fmt.Fprintf(buf, "TracerPid:\t%d\n", tpid) - var fds int - var vss, rss, data uint64 - s.t.WithMuLocked(func(t *kernel.Task) { - if fdTable := t.FDTable(); fdTable != nil { - fds = fdTable.Size() + if m.Dumpability() != mm.UserDumpable { + uid = auth.RootKUID + if kuid := creds.UserNamespace.MapToKUID(auth.RootUID); kuid.Ok() { + uid = kuid } - if mm := t.MemoryManager(); mm != nil { - vss = mm.VirtualMemorySize() - rss = mm.ResidentSetSize() - data = mm.VirtualDataSize() + gid = auth.RootKGID + if kgid := creds.UserNamespace.MapToKGID(auth.RootGID); kgid.Ok() { + gid = kgid } - }) - fmt.Fprintf(buf, "FDSize:\t%d\n", fds) - fmt.Fprintf(buf, "VmSize:\t%d kB\n", vss>>10) - fmt.Fprintf(buf, "VmRSS:\t%d kB\n", rss>>10) - fmt.Fprintf(buf, "VmData:\t%d kB\n", data>>10) - fmt.Fprintf(buf, "Threads:\t%d\n", s.t.ThreadGroup().Count()) - creds := s.t.Credentials() - fmt.Fprintf(buf, "CapInh:\t%016x\n", creds.InheritableCaps) - fmt.Fprintf(buf, "CapPrm:\t%016x\n", creds.PermittedCaps) - fmt.Fprintf(buf, "CapEff:\t%016x\n", creds.EffectiveCaps) - fmt.Fprintf(buf, "CapBnd:\t%016x\n", creds.BoundingCaps) - fmt.Fprintf(buf, "Seccomp:\t%d\n", s.t.SeccompMode()) - return nil -} - -// 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 -} - -// +stateify savable -type ioData struct { - ioUsage + } + return uid, gid } -var _ vfs.DynamicBytesSource = (*ioData)(nil) - -// Generate implements vfs.DynamicBytesSource.Generate. -func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { - io := usage.IO{} - io.Accumulate(i.IOUsage()) - - fmt.Fprintf(buf, "char: %d\n", io.CharsRead) - fmt.Fprintf(buf, "wchar: %d\n", io.CharsWritten) - fmt.Fprintf(buf, "syscr: %d\n", io.ReadSyscalls) - fmt.Fprintf(buf, "syscw: %d\n", io.WriteSyscalls) - fmt.Fprintf(buf, "read_bytes: %d\n", io.BytesRead) - fmt.Fprintf(buf, "write_bytes: %d\n", io.BytesWritten) - fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) - return nil +func newIO(t *kernel.Task, isThreadGroup bool) *ioData { + if isThreadGroup { + return &ioData{ioUsage: t.ThreadGroup()} + } + return &ioData{ioUsage: t} } diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go new file mode 100644 index 000000000..93f0e1aa8 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -0,0 +1,272 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "bytes" + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/usermem" +) + +// mm gets the kernel task's MemoryManager. No additional reference is taken on +// mm here. This is safe because MemoryManager.destroy is required to leave the +// MemoryManager in a state where it's still usable as a DynamicBytesSource. +func getMM(task *kernel.Task) *mm.MemoryManager { + var tmm *mm.MemoryManager + task.WithMuLocked(func(t *kernel.Task) { + if mm := t.MemoryManager(); mm != nil { + tmm = mm + } + }) + return tmm +} + +// mapsData implements vfs.DynamicBytesSource for /proc/[pid]/maps. +// +// +stateify savable +type mapsData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*mapsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *mapsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + if mm := getMM(d.task); mm != nil { + mm.ReadMapsDataInto(ctx, buf) + } + return nil +} + +// smapsData implements vfs.DynamicBytesSource for /proc/[pid]/smaps. +// +// +stateify savable +type smapsData struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ dynamicInode = (*smapsData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (d *smapsData) Generate(ctx context.Context, buf *bytes.Buffer) error { + if mm := getMM(d.task); mm != nil { + mm.ReadSmapsDataInto(ctx, buf) + } + return nil +} + +// +stateify savable +type taskStatData struct { + kernfs.DynamicBytesFile + + t *kernel.Task + + // If tgstats is true, accumulate fault stats (not implemented) and CPU + // time across all tasks in t's thread group. + tgstats bool + + // pidns is the PID namespace associated with the proc filesystem that + // includes the file using this statData. + pidns *kernel.PIDNamespace +} + +var _ dynamicInode = (*taskStatData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (s *taskStatData) Generate(ctx context.Context, buf *bytes.Buffer) error { + fmt.Fprintf(buf, "%d ", s.pidns.IDOfTask(s.t)) + fmt.Fprintf(buf, "(%s) ", s.t.Name()) + fmt.Fprintf(buf, "%c ", s.t.StateStatus()[0]) + ppid := kernel.ThreadID(0) + if parent := s.t.Parent(); parent != nil { + ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup()) + } + fmt.Fprintf(buf, "%d ", ppid) + fmt.Fprintf(buf, "%d ", s.pidns.IDOfProcessGroup(s.t.ThreadGroup().ProcessGroup())) + fmt.Fprintf(buf, "%d ", s.pidns.IDOfSession(s.t.ThreadGroup().Session())) + fmt.Fprintf(buf, "0 0 " /* tty_nr tpgid */) + fmt.Fprintf(buf, "0 " /* flags */) + fmt.Fprintf(buf, "0 0 0 0 " /* minflt cminflt majflt cmajflt */) + var cputime usage.CPUStats + if s.tgstats { + cputime = s.t.ThreadGroup().CPUStats() + } else { + cputime = s.t.CPUStats() + } + fmt.Fprintf(buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime)) + cputime = s.t.ThreadGroup().JoinedChildCPUStats() + fmt.Fprintf(buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime)) + fmt.Fprintf(buf, "%d %d ", s.t.Priority(), s.t.Niceness()) + fmt.Fprintf(buf, "%d ", s.t.ThreadGroup().Count()) + + // itrealvalue. Since kernel 2.6.17, this field is no longer + // maintained, and is hard coded as 0. + fmt.Fprintf(buf, "0 ") + + // Start time is relative to boot time, expressed in clock ticks. + fmt.Fprintf(buf, "%d ", linux.ClockTFromDuration(s.t.StartTime().Sub(s.t.Kernel().Timekeeper().BootTime()))) + + var vss, rss uint64 + s.t.WithMuLocked(func(t *kernel.Task) { + if mm := t.MemoryManager(); mm != nil { + vss = mm.VirtualMemorySize() + rss = mm.ResidentSetSize() + } + }) + fmt.Fprintf(buf, "%d %d ", vss, rss/usermem.PageSize) + + // rsslim. + fmt.Fprintf(buf, "%d ", s.t.ThreadGroup().Limits().Get(limits.Rss).Cur) + + fmt.Fprintf(buf, "0 0 0 0 0 " /* startcode endcode startstack kstkesp kstkeip */) + fmt.Fprintf(buf, "0 0 0 0 0 " /* signal blocked sigignore sigcatch wchan */) + fmt.Fprintf(buf, "0 0 " /* nswap cnswap */) + terminationSignal := linux.Signal(0) + if s.t == s.t.ThreadGroup().Leader() { + terminationSignal = s.t.ThreadGroup().TerminationSignal() + } + fmt.Fprintf(buf, "%d ", terminationSignal) + fmt.Fprintf(buf, "0 0 0 " /* processor rt_priority policy */) + fmt.Fprintf(buf, "0 0 0 " /* delayacct_blkio_ticks guest_time cguest_time */) + fmt.Fprintf(buf, "0 0 0 0 0 0 0 " /* start_data end_data start_brk arg_start arg_end env_start env_end */) + fmt.Fprintf(buf, "0\n" /* exit_code */) + + return nil +} + +// statmData implements vfs.DynamicBytesSource for /proc/[pid]/statm. +// +// +stateify savable +type statmData struct { + kernfs.DynamicBytesFile + + t *kernel.Task +} + +var _ dynamicInode = (*statmData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (s *statmData) Generate(ctx context.Context, buf *bytes.Buffer) error { + var vss, rss uint64 + s.t.WithMuLocked(func(t *kernel.Task) { + if mm := t.MemoryManager(); mm != nil { + vss = mm.VirtualMemorySize() + rss = mm.ResidentSetSize() + } + }) + + fmt.Fprintf(buf, "%d %d 0 0 0 0 0\n", vss/usermem.PageSize, rss/usermem.PageSize) + return nil +} + +// statusData implements vfs.DynamicBytesSource for /proc/[pid]/status. +// +// +stateify savable +type statusData struct { + kernfs.DynamicBytesFile + + t *kernel.Task + pidns *kernel.PIDNamespace +} + +var _ dynamicInode = (*statusData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { + fmt.Fprintf(buf, "Name:\t%s\n", s.t.Name()) + fmt.Fprintf(buf, "State:\t%s\n", s.t.StateStatus()) + fmt.Fprintf(buf, "Tgid:\t%d\n", s.pidns.IDOfThreadGroup(s.t.ThreadGroup())) + fmt.Fprintf(buf, "Pid:\t%d\n", s.pidns.IDOfTask(s.t)) + ppid := kernel.ThreadID(0) + if parent := s.t.Parent(); parent != nil { + ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup()) + } + fmt.Fprintf(buf, "PPid:\t%d\n", ppid) + tpid := kernel.ThreadID(0) + if tracer := s.t.Tracer(); tracer != nil { + tpid = s.pidns.IDOfTask(tracer) + } + fmt.Fprintf(buf, "TracerPid:\t%d\n", tpid) + var fds int + var vss, rss, data uint64 + s.t.WithMuLocked(func(t *kernel.Task) { + if fdTable := t.FDTable(); fdTable != nil { + fds = fdTable.Size() + } + if mm := t.MemoryManager(); mm != nil { + vss = mm.VirtualMemorySize() + rss = mm.ResidentSetSize() + data = mm.VirtualDataSize() + } + }) + fmt.Fprintf(buf, "FDSize:\t%d\n", fds) + fmt.Fprintf(buf, "VmSize:\t%d kB\n", vss>>10) + fmt.Fprintf(buf, "VmRSS:\t%d kB\n", rss>>10) + fmt.Fprintf(buf, "VmData:\t%d kB\n", data>>10) + fmt.Fprintf(buf, "Threads:\t%d\n", s.t.ThreadGroup().Count()) + creds := s.t.Credentials() + fmt.Fprintf(buf, "CapInh:\t%016x\n", creds.InheritableCaps) + fmt.Fprintf(buf, "CapPrm:\t%016x\n", creds.PermittedCaps) + fmt.Fprintf(buf, "CapEff:\t%016x\n", creds.EffectiveCaps) + fmt.Fprintf(buf, "CapBnd:\t%016x\n", creds.BoundingCaps) + fmt.Fprintf(buf, "Seccomp:\t%d\n", s.t.SeccompMode()) + // We unconditionally report a single NUMA node. See + // pkg/sentry/syscalls/linux/sys_mempolicy.go. + fmt.Fprintf(buf, "Mems_allowed:\t1\n") + fmt.Fprintf(buf, "Mems_allowed_list:\t0\n") + return nil +} + +// 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 +} + +// +stateify savable +type ioData struct { + kernfs.DynamicBytesFile + + ioUsage +} + +var _ dynamicInode = (*ioData)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { + io := usage.IO{} + io.Accumulate(i.IOUsage()) + + fmt.Fprintf(buf, "char: %d\n", io.CharsRead) + fmt.Fprintf(buf, "wchar: %d\n", io.CharsWritten) + fmt.Fprintf(buf, "syscr: %d\n", io.ReadSyscalls) + fmt.Fprintf(buf, "syscw: %d\n", io.WriteSyscalls) + fmt.Fprintf(buf, "read_bytes: %d\n", io.BytesRead) + fmt.Fprintf(buf, "write_bytes: %d\n", io.BytesWritten) + fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) + return nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go new file mode 100644 index 000000000..d8f92d52f --- /dev/null +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -0,0 +1,218 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "sort" + "strconv" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +const ( + defaultPermission = 0444 + selfName = "self" + threadSelfName = "thread-self" +) + +// InoGenerator generates unique inode numbers for a given filesystem. +type InoGenerator interface { + NextIno() uint64 +} + +// tasksInode represents the inode for /proc/ directory. +// +// +stateify savable +type tasksInode struct { + kernfs.InodeNotSymlink + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeAttrs + kernfs.OrderedChildren + + inoGen InoGenerator + pidns *kernel.PIDNamespace + + // '/proc/self' and '/proc/thread-self' have custom directory offsets in + // Linux. So handle them outside of OrderedChildren. + selfSymlink *vfs.Dentry + threadSelfSymlink *vfs.Dentry +} + +var _ kernfs.Inode = (*tasksInode)(nil) + +func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace) (*tasksInode, *kernfs.Dentry) { + root := auth.NewRootCredentials(pidns.UserNamespace()) + contents := map[string]*kernfs.Dentry{ + //"cpuinfo": newCPUInfo(ctx, msrc), + //"filesystems": seqfile.NewSeqFileInode(ctx, &filesystemsData{}, msrc), + "loadavg": newDentry(root, inoGen.NextIno(), defaultPermission, &loadavgData{}), + "meminfo": newDentry(root, inoGen.NextIno(), defaultPermission, &meminfoData{k: k}), + "mounts": kernfs.NewStaticSymlink(root, inoGen.NextIno(), defaultPermission, "self/mounts"), + "stat": newDentry(root, inoGen.NextIno(), defaultPermission, &statData{k: k}), + //"uptime": newUptime(ctx, msrc), + //"version": newVersionData(root, inoGen.NextIno(), k), + "version": newDentry(root, inoGen.NextIno(), defaultPermission, &versionData{k: k}), + } + + inode := &tasksInode{ + pidns: pidns, + inoGen: inoGen, + selfSymlink: newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), + threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(), + } + inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555) + + dentry := &kernfs.Dentry{} + dentry.Init(inode) + + inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + links := inode.OrderedChildren.Populate(dentry, contents) + inode.IncLinks(links) + + return inode, dentry +} + +// Lookup implements kernfs.inodeDynamicLookup. +func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { + // Try to lookup a corresponding task. + tid, err := strconv.ParseUint(name, 10, 64) + if err != nil { + // If it failed to parse, check if it's one of the special handled files. + switch name { + case selfName: + return i.selfSymlink, nil + case threadSelfName: + return i.threadSelfSymlink, nil + } + return nil, syserror.ENOENT + } + + task := i.pidns.TaskWithID(kernel.ThreadID(tid)) + if task == nil { + return nil, syserror.ENOENT + } + + taskDentry := newTaskInode(i.inoGen, task, i.pidns, true) + 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 + const FIRST_PROCESS_ENTRY = 256 + + // Use maxTaskID to shortcut searches that will result in 0 entries. + const maxTaskID = kernel.TasksLimit + 1 + if offset >= maxTaskID { + return offset, nil + } + + // According to Linux (fs/proc/base.c:proc_pid_readdir()), process directories + // start at offset FIRST_PROCESS_ENTRY with '/proc/self', followed by + // '/proc/thread-self' and then '/proc/[pid]'. + if offset < FIRST_PROCESS_ENTRY { + offset = FIRST_PROCESS_ENTRY + } + + if offset == FIRST_PROCESS_ENTRY { + dirent := vfs.Dirent{ + Name: selfName, + Type: linux.DT_LNK, + Ino: i.inoGen.NextIno(), + NextOff: offset + 1, + } + if !cb.Handle(dirent) { + return offset, nil + } + offset++ + } + if offset == FIRST_PROCESS_ENTRY+1 { + dirent := vfs.Dirent{ + Name: threadSelfName, + Type: linux.DT_LNK, + Ino: i.inoGen.NextIno(), + NextOff: offset + 1, + } + if !cb.Handle(dirent) { + return offset, nil + } + offset++ + } + + // Collect all tasks that TGIDs are greater than the offset specified. Per + // Linux we only include in directory listings if it's the leader. But for + // whatever crazy reason, you can still walk to the given node. + var tids []int + startTid := offset - FIRST_PROCESS_ENTRY - 2 + for _, tg := range i.pidns.ThreadGroups() { + tid := i.pidns.IDOfThreadGroup(tg) + if int64(tid) < startTid { + continue + } + if leader := tg.Leader(); leader != nil { + tids = append(tids, int(tid)) + } + } + + if len(tids) == 0 { + return offset, nil + } + + sort.Ints(tids) + for _, tid := range tids { + dirent := vfs.Dirent{ + Name: strconv.FormatUint(uint64(tid), 10), + Type: linux.DT_DIR, + Ino: i.inoGen.NextIno(), + NextOff: FIRST_PROCESS_ENTRY + 2 + int64(tid) + 1, + } + if !cb.Handle(dirent) { + return offset, nil + } + offset++ + } + return maxTaskID, nil +} + +// Open implements kernfs.Inode. +func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + fd := &kernfs.GenericDirectoryFD{} + fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, flags) + return fd.VFSFileDescription(), nil +} + +func (i *tasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { + stat := i.InodeAttrs.Stat(vsfs) + + // Add dynamic children to link count. + for _, tg := range i.pidns.ThreadGroups() { + if leader := tg.Leader(); leader != nil { + stat.Nlink++ + } + } + + return stat +} diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go new file mode 100644 index 000000000..91f30a798 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -0,0 +1,92 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "fmt" + "strconv" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" +) + +type selfSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + pidns *kernel.PIDNamespace +} + +var _ kernfs.Inode = (*selfSymlink)(nil) + +func newSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { + inode := &selfSymlink{pidns: pidns} + inode.Init(creds, ino, linux.ModeSymlink|perm) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // Who is reading this link? + return "", syserror.EINVAL + } + tgid := s.pidns.IDOfThreadGroup(t.ThreadGroup()) + if tgid == 0 { + return "", syserror.ENOENT + } + return strconv.FormatUint(uint64(tgid), 10), nil +} + +type threadSelfSymlink struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + pidns *kernel.PIDNamespace +} + +var _ kernfs.Inode = (*threadSelfSymlink)(nil) + +func newThreadSelfSymlink(creds *auth.Credentials, ino uint64, perm linux.FileMode, pidns *kernel.PIDNamespace) *kernfs.Dentry { + inode := &threadSelfSymlink{pidns: pidns} + inode.Init(creds, ino, linux.ModeSymlink|perm) + + d := &kernfs.Dentry{} + d.Init(inode) + return d +} + +func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // Who is reading this link? + return "", syserror.EINVAL + } + tgid := s.pidns.IDOfThreadGroup(t.ThreadGroup()) + tid := s.pidns.IDOfTask(t) + if tid == 0 || tgid == 0 { + return "", syserror.ENOENT + } + return fmt.Sprintf("%d/task/%d", tgid, tid), nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go new file mode 100644 index 000000000..ca8c87ec2 --- /dev/null +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -0,0 +1,555 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package proc + +import ( + "fmt" + "math" + "path" + "strconv" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +var ( + // Next offset 256 by convention. Adds 1 for the next offset. + selfLink = vfs.Dirent{Type: linux.DT_LNK, NextOff: 256 + 0 + 1} + threadSelfLink = vfs.Dirent{Type: linux.DT_LNK, NextOff: 256 + 1 + 1} + + // /proc/[pid] next offset starts at 256+2 (files above), then adds the + // PID, and adds 1 for the next offset. + proc1 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 1 + 1} + proc2 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 2 + 1} + proc3 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 3 + 1} +) + +type testIterDirentsCallback struct { + dirents []vfs.Dirent +} + +func (t *testIterDirentsCallback) Handle(d vfs.Dirent) bool { + t.dirents = append(t.dirents, d) + return true +} + +func checkDots(dirs []vfs.Dirent) ([]vfs.Dirent, error) { + if got := len(dirs); got < 2 { + return dirs, fmt.Errorf("wrong number of dirents, want at least: 2, got: %d: %v", got, dirs) + } + for i, want := range []string{".", ".."} { + if got := dirs[i].Name; got != want { + return dirs, fmt.Errorf("wrong name, want: %s, got: %s", want, got) + } + if got := dirs[i].Type; got != linux.DT_DIR { + return dirs, fmt.Errorf("wrong type, want: %d, got: %d", linux.DT_DIR, got) + } + } + return dirs[2:], nil +} + +func checkTasksStaticFiles(gots []vfs.Dirent) ([]vfs.Dirent, error) { + wants := map[string]vfs.Dirent{ + "loadavg": {Type: linux.DT_REG}, + "meminfo": {Type: linux.DT_REG}, + "mounts": {Type: linux.DT_LNK}, + "self": selfLink, + "stat": {Type: linux.DT_REG}, + "thread-self": threadSelfLink, + "version": {Type: linux.DT_REG}, + } + return checkFiles(gots, wants) +} + +func checkTaskStaticFiles(gots []vfs.Dirent) ([]vfs.Dirent, error) { + wants := map[string]vfs.Dirent{ + "io": {Type: linux.DT_REG}, + "maps": {Type: linux.DT_REG}, + "smaps": {Type: linux.DT_REG}, + "stat": {Type: linux.DT_REG}, + "statm": {Type: linux.DT_REG}, + "status": {Type: linux.DT_REG}, + } + return checkFiles(gots, wants) +} + +func checkFiles(gots []vfs.Dirent, wants map[string]vfs.Dirent) ([]vfs.Dirent, error) { + // Go over all files, when there is a match, the file is removed from both + // 'gots' and 'wants'. wants is expected to reach 0, as all files must + // be present. Remaining files in 'gots', is returned to caller to decide + // whether this is valid or not. + for i := 0; i < len(gots); i++ { + got := gots[i] + want, ok := wants[got.Name] + if !ok { + continue + } + if want.Type != got.Type { + return gots, fmt.Errorf("wrong file type, want: %v, got: %v: %+v", want.Type, got.Type, got) + } + if want.NextOff != 0 && want.NextOff != got.NextOff { + return gots, fmt.Errorf("wrong dirent offset, want: %v, got: %v: %+v", want.NextOff, got.NextOff, got) + } + + delete(wants, got.Name) + gots = append(gots[0:i], gots[i+1:]...) + i-- + } + if len(wants) != 0 { + return gots, fmt.Errorf("not all files were found, missing: %+v", wants) + } + return gots, nil +} + +func setup() (context.Context, *vfs.VirtualFilesystem, vfs.VirtualDentry, error) { + k, err := boot() + if err != nil { + return nil, nil, vfs.VirtualDentry{}, fmt.Errorf("creating kernel: %v", err) + } + + ctx := k.SupervisorContext() + creds := auth.CredentialsFromContext(ctx) + + vfsObj := vfs.New() + vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &vfs.GetFilesystemOptions{}) + if err != nil { + return nil, nil, vfs.VirtualDentry{}, fmt.Errorf("NewMountNamespace(): %v", err) + } + return ctx, vfsObj, mntns.Root(), nil +} + +func TestTasksEmpty(t *testing.T) { + ctx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/")}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt failed: %v", err) + } + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + cb.dirents, err = checkDots(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + cb.dirents, err = checkTasksStaticFiles(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + if len(cb.dirents) != 0 { + t.Errorf("found more files than expected: %+v", cb.dirents) + } +} + +func TestTasks(t *testing.T) { + ctx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + k := kernel.KernelFromContext(ctx) + var tasks []*kernel.Task + for i := 0; i < 5; i++ { + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + task, err := createTask(ctx, fmt.Sprintf("name-%d", i), tc) + if err != nil { + t.Fatalf("CreateTask(): %v", err) + } + tasks = append(tasks, task) + } + + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/")}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) + } + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + cb.dirents, err = checkDots(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + cb.dirents, err = checkTasksStaticFiles(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + lastPid := 0 + for _, d := range cb.dirents { + pid, err := strconv.Atoi(d.Name) + if err != nil { + t.Fatalf("Invalid process directory %q", d.Name) + } + if lastPid > pid { + t.Errorf("pids not in order: %v", cb.dirents) + } + found := false + for _, t := range tasks { + if k.TaskSet().Root.IDOfTask(t) == kernel.ThreadID(pid) { + found = true + } + } + if !found { + t.Errorf("Additional task ID %d listed: %v", pid, tasks) + } + // Next offset starts at 256+2 ('self' and 'thread-self'), then adds the + // PID, and adds 1 for the next offset. + if want := int64(256 + 2 + pid + 1); d.NextOff != want { + t.Errorf("Wrong dirent offset want: %d got: %d: %+v", want, d.NextOff, d) + } + } + + // Test lookup. + for _, path := range []string{"/1", "/2"} { + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse(path)}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(%q) failed: %v", path, err) + } + buf := make([]byte, 1) + bufIOSeq := usermem.BytesIOSequence(buf) + if _, err := fd.Read(ctx, bufIOSeq, vfs.ReadOptions{}); err != syserror.EISDIR { + t.Errorf("wrong error reading directory: %v", err) + } + } + + if _, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/9999")}, + &vfs.OpenOptions{}, + ); err != syserror.ENOENT { + t.Fatalf("wrong error from vfsfs.OpenAt(/9999): %v", err) + } +} + +func TestTasksOffset(t *testing.T) { + ctx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + k := kernel.KernelFromContext(ctx) + for i := 0; i < 3; i++ { + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + if _, err := createTask(ctx, fmt.Sprintf("name-%d", i), tc); err != nil { + t.Fatalf("CreateTask(): %v", err) + } + } + + for _, tc := range []struct { + name string + offset int64 + wants map[string]vfs.Dirent + }{ + { + name: "small offset", + offset: 100, + wants: map[string]vfs.Dirent{ + "self": selfLink, + "thread-self": threadSelfLink, + "1": proc1, + "2": proc2, + "3": proc3, + }, + }, + { + name: "offset at start", + offset: 256, + wants: map[string]vfs.Dirent{ + "self": selfLink, + "thread-self": threadSelfLink, + "1": proc1, + "2": proc2, + "3": proc3, + }, + }, + { + name: "skip /proc/self", + offset: 257, + wants: map[string]vfs.Dirent{ + "thread-self": threadSelfLink, + "1": proc1, + "2": proc2, + "3": proc3, + }, + }, + { + name: "skip symlinks", + offset: 258, + wants: map[string]vfs.Dirent{ + "1": proc1, + "2": proc2, + "3": proc3, + }, + }, + { + name: "skip first process", + offset: 260, + wants: map[string]vfs.Dirent{ + "2": proc2, + "3": proc3, + }, + }, + { + name: "last process", + offset: 261, + wants: map[string]vfs.Dirent{ + "3": proc3, + }, + }, + { + name: "after last", + offset: 262, + wants: nil, + }, + { + name: "TaskLimit+1", + offset: kernel.TasksLimit + 1, + wants: nil, + }, + { + name: "max", + offset: math.MaxInt64, + wants: nil, + }, + } { + t.Run(tc.name, func(t *testing.T) { + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/")}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) + } + if _, err := fd.Impl().Seek(ctx, tc.offset, linux.SEEK_SET); err != nil { + t.Fatalf("Seek(%d, SEEK_SET): %v", tc.offset, err) + } + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + if cb.dirents, err = checkFiles(cb.dirents, tc.wants); err != nil { + t.Error(err.Error()) + } + if len(cb.dirents) != 0 { + t.Errorf("found more files than expected: %+v", cb.dirents) + } + }) + } +} + +func TestTask(t *testing.T) { + ctx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + k := kernel.KernelFromContext(ctx) + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + _, err = createTask(ctx, "name", tc) + if err != nil { + t.Fatalf("CreateTask(): %v", err) + } + + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/1")}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(/1) failed: %v", err) + } + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + cb.dirents, err = checkDots(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + cb.dirents, err = checkTaskStaticFiles(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + if len(cb.dirents) != 0 { + t.Errorf("found more files than expected: %+v", cb.dirents) + } +} + +func TestProcSelf(t *testing.T) { + ctx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + k := kernel.KernelFromContext(ctx) + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + task, err := createTask(ctx, "name", tc) + if err != nil { + t.Fatalf("CreateTask(): %v", err) + } + + fd, err := vfsObj.OpenAt( + task, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/self/"), FollowFinalSymlink: true}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(/self/) failed: %v", err) + } + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + cb.dirents, err = checkDots(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + cb.dirents, err = checkTaskStaticFiles(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + if len(cb.dirents) != 0 { + t.Errorf("found more files than expected: %+v", cb.dirents) + } +} + +func iterateDir(ctx context.Context, t *testing.T, vfsObj *vfs.VirtualFilesystem, root vfs.VirtualDentry, fd *vfs.FileDescription) { + t.Logf("Iterating: /proc%s", fd.MappedName(ctx)) + + cb := testIterDirentsCallback{} + if err := fd.Impl().IterDirents(ctx, &cb); err != nil { + t.Fatalf("IterDirents(): %v", err) + } + var err error + cb.dirents, err = checkDots(cb.dirents) + if err != nil { + t.Error(err.Error()) + } + for _, d := range cb.dirents { + childPath := path.Join(fd.MappedName(ctx), d.Name) + if d.Type == linux.DT_LNK { + link, err := vfsObj.ReadlinkAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse(childPath)}, + ) + if err != nil { + t.Errorf("vfsfs.ReadlinkAt(%v) failed: %v", childPath, err) + } else { + t.Logf("Skipping symlink: /proc%s => %s", childPath, link) + } + continue + } + + t.Logf("Opening: /proc%s", childPath) + child, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(ctx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse(childPath)}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Errorf("vfsfs.OpenAt(%v) failed: %v", childPath, err) + continue + } + stat, err := child.Stat(ctx, vfs.StatOptions{}) + if err != nil { + t.Errorf("Stat(%v) failed: %v", childPath, err) + } + if got := linux.FileMode(stat.Mode).DirentType(); got != d.Type { + t.Errorf("wrong file mode, stat: %v, dirent: %v", got, d.Type) + } + if d.Type == linux.DT_DIR { + // Found another dir, let's do it again! + iterateDir(ctx, t, vfsObj, root, child) + } + } +} + +// TestTree iterates all directories and stats every file. +func TestTree(t *testing.T) { + uberCtx, vfsObj, root, err := setup() + if err != nil { + t.Fatalf("Setup failed: %v", err) + } + defer root.DecRef() + + k := kernel.KernelFromContext(uberCtx) + var tasks []*kernel.Task + for i := 0; i < 5; i++ { + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + task, err := createTask(uberCtx, fmt.Sprintf("name-%d", i), tc) + if err != nil { + t.Fatalf("CreateTask(): %v", err) + } + tasks = append(tasks, task) + } + + ctx := tasks[0] + fd, err := vfsObj.OpenAt( + ctx, + auth.CredentialsFromContext(uberCtx), + &vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/")}, + &vfs.OpenOptions{}, + ) + if err != nil { + t.Fatalf("vfsfs.OpenAt(/) failed: %v", err) + } + iterateDir(ctx, t, vfsObj, root, fd) +} diff --git a/pkg/sentry/fsimpl/proc/version.go b/pkg/sentry/fsimpl/proc/version.go index e1643d4e0..367f2396b 100644 --- a/pkg/sentry/fsimpl/proc/version.go +++ b/pkg/sentry/fsimpl/proc/version.go @@ -19,19 +19,21 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/vfs" ) // versionData implements vfs.DynamicBytesSource for /proc/version. // // +stateify savable type versionData struct { + kernfs.DynamicBytesFile + // k is the owning Kernel. k *kernel.Kernel } -var _ vfs.DynamicBytesSource = (*versionData)(nil) +var _ dynamicInode = (*versionData)(nil) // Generate implements vfs.DynamicBytesSource.Generate. func (v *versionData) Generate(ctx context.Context, buf *bytes.Buffer) error { diff --git a/pkg/sentry/fsimpl/memfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD index bc5c0b591..a5b285987 100644 --- a/pkg/sentry/fsimpl/memfs/BUILD +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -1,14 +1,13 @@ load("//tools/go_stateify:defs.bzl", "go_library") load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -load("//tools/go_generics:defs.bzl", "go_template_instance") - go_template_instance( name = "dentry_list", out = "dentry_list.go", - package = "memfs", + package = "tmpfs", prefix = "dentry", template = "//pkg/ilist:generic_list", types = { @@ -18,24 +17,34 @@ go_template_instance( ) go_library( - name = "memfs", + name = "tmpfs", srcs = [ "dentry_list.go", "directory.go", "filesystem.go", - "memfs.go", "named_pipe.go", "regular_file.go", "symlink.go", + "tmpfs.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/memfs", + importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs", deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/fspath", + "//pkg/log", "//pkg/sentry/arch", "//pkg/sentry/context", + "//pkg/sentry/fs", + "//pkg/sentry/fs/fsutil", + "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/pipe", + "//pkg/sentry/memmap", + "//pkg/sentry/pgalloc", + "//pkg/sentry/platform", + "//pkg/sentry/safemem", + "//pkg/sentry/usage", "//pkg/sentry/usermem", "//pkg/sentry/vfs", "//pkg/syserror", @@ -47,8 +56,9 @@ go_test( size = "small", srcs = ["benchmark_test.go"], deps = [ - ":memfs", + ":tmpfs", "//pkg/abi/linux", + "//pkg/fspath", "//pkg/refs", "//pkg/sentry/context", "//pkg/sentry/context/contexttest", @@ -61,15 +71,20 @@ go_test( ) go_test( - name = "memfs_test", + name = "tmpfs_test", size = "small", - srcs = ["pipe_test.go"], - embed = [":memfs"], + srcs = [ + "pipe_test.go", + "regular_file_test.go", + ], + embed = [":tmpfs"], deps = [ "//pkg/abi/linux", + "//pkg/fspath", "//pkg/sentry/context", "//pkg/sentry/context/contexttest", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/contexttest", "//pkg/sentry/usermem", "//pkg/sentry/vfs", "//pkg/syserror", diff --git a/pkg/sentry/fsimpl/memfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go index ea6417ce7..d88c83499 100644 --- a/pkg/sentry/fsimpl/memfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -21,12 +21,13 @@ import ( "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/context/contexttest" "gvisor.dev/gvisor/pkg/sentry/fs" _ "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" - "gvisor.dev/gvisor/pkg/sentry/fsimpl/memfs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" @@ -175,8 +176,10 @@ func BenchmarkVFS2MemfsStat(b *testing.B) { // Create VFS. vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("memfs", memfs.FilesystemType{}) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.GetFilesystemOptions{}) + vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) if err != nil { b.Fatalf("failed to create tmpfs root mount: %v", err) } @@ -193,9 +196,9 @@ func BenchmarkVFS2MemfsStat(b *testing.B) { for i := depth; i > 0; i-- { name := fmt.Sprintf("%d", i) pop := vfs.PathOperation{ - Root: root, - Start: vd, - Pathname: name, + Root: root, + Start: vd, + Path: fspath.Parse(name), } if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{ Mode: 0755, @@ -216,7 +219,7 @@ func BenchmarkVFS2MemfsStat(b *testing.B) { fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ Root: root, Start: vd, - Pathname: filename, + Path: fspath.Parse(filename), FollowFinalSymlink: true, }, &vfs.OpenOptions{ Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, @@ -237,7 +240,7 @@ func BenchmarkVFS2MemfsStat(b *testing.B) { stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ Root: root, Start: root, - Pathname: filePath, + Path: fspath.Parse(filePath), FollowFinalSymlink: true, }, &vfs.StatOptions{}) if err != nil { @@ -364,8 +367,10 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { // Create VFS. vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("memfs", memfs.FilesystemType{}) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.GetFilesystemOptions{}) + vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) if err != nil { b.Fatalf("failed to create tmpfs root mount: %v", err) } @@ -378,9 +383,9 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { root := mntns.Root() defer root.DecRef() pop := vfs.PathOperation{ - Root: root, - Start: root, - Pathname: mountPointName, + Root: root, + Start: root, + Path: fspath.Parse(mountPointName), } if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{ Mode: 0755, @@ -394,7 +399,7 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { } defer mountPoint.DecRef() // Create and mount the submount. - if err := vfsObj.NewMount(ctx, creds, "", &pop, "memfs", &vfs.GetFilesystemOptions{}); err != nil { + if err := vfsObj.MountAt(ctx, creds, "", &pop, "tmpfs", &vfs.MountOptions{}); err != nil { b.Fatalf("failed to mount tmpfs submount: %v", err) } filePathBuilder.WriteString(mountPointName) @@ -408,9 +413,9 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { for i := depth; i > 0; i-- { name := fmt.Sprintf("%d", i) pop := vfs.PathOperation{ - Root: root, - Start: vd, - Pathname: name, + Root: root, + Start: vd, + Path: fspath.Parse(name), } if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{ Mode: 0755, @@ -438,7 +443,7 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ Root: root, Start: vd, - Pathname: filename, + Path: fspath.Parse(filename), FollowFinalSymlink: true, }, &vfs.OpenOptions{ Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, @@ -458,7 +463,7 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) { stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ Root: root, Start: root, - Pathname: filePath, + Path: fspath.Parse(filePath), FollowFinalSymlink: true, }, &vfs.StatOptions{}) if err != nil { diff --git a/pkg/sentry/fsimpl/memfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index 0bd82e480..887ca2619 100644 --- a/pkg/sentry/fsimpl/memfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -package memfs +package tmpfs import ( "gvisor.dev/gvisor/pkg/abi/linux" diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go new file mode 100644 index 000000000..26979729e --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -0,0 +1,698 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tmpfs + +import ( + "fmt" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Sync implements vfs.FilesystemImpl.Sync. +func (fs *filesystem) Sync(ctx context.Context) error { + // All filesystem state is in-memory. + return nil +} + +// stepLocked resolves rp.Component() to an existing file, starting from the +// given directory. +// +// stepLocked is loosely analogous to fs/namei.c:walk_component(). +// +// Preconditions: filesystem.mu must be locked. !rp.Done(). +func stepLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { + if !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } +afterSymlink: + nextVFSD, err := rp.ResolveComponent(&d.vfsd) + if err != nil { + return nil, err + } + if nextVFSD == nil { + // Since the Dentry tree is the sole source of truth for tmpfs, if it's + // not in the Dentry tree, it doesn't exist. + return nil, syserror.ENOENT + } + next := nextVFSD.Impl().(*dentry) + if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // TODO: symlink traversals update access time + if err := rp.HandleSymlink(symlink.target); err != nil { + return nil, err + } + goto afterSymlink // don't check the current directory again + } + rp.Advance() + return next, nil +} + +// walkParentDirLocked resolves all but the last path component of rp to an +// existing directory, starting from the given directory (which is usually +// rp.Start().Impl().(*dentry)). It does not check that the returned directory +// is searchable by the provider of rp. +// +// walkParentDirLocked is loosely analogous to Linux's +// fs/namei.c:path_parentat(). +// +// Preconditions: filesystem.mu must be locked. !rp.Done(). +func walkParentDirLocked(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { + for !rp.Final() { + next, err := stepLocked(rp, d) + if err != nil { + return nil, err + } + d = next + } + if !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// resolveLocked resolves rp to an existing file. +// +// resolveLocked is loosely analogous to Linux's fs/namei.c:path_lookupat(). +// +// Preconditions: filesystem.mu must be locked. +func resolveLocked(rp *vfs.ResolvingPath) (*dentry, error) { + d := rp.Start().Impl().(*dentry) + for !rp.Done() { + next, err := stepLocked(rp, d) + if err != nil { + return nil, err + } + d = next + } + if rp.MustBeDir() && !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// doCreateAt checks that creating a file at rp is permitted, then invokes +// create to do so. +// +// doCreateAt is loosely analogous to a conjunction of Linux's +// fs/namei.c:filename_create() and done_path_create(). +// +// Preconditions: !rp.Done(). For the final path component in rp, +// !rp.ShouldFollowSymlink(). +func (fs *filesystem) doCreateAt(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { + fs.mu.Lock() + defer fs.mu.Unlock() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EEXIST + } + // Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(), + // because if the child exists we want to return EEXIST immediately instead + // of attempting symlink/mount traversal. + if parent.vfsd.Child(name) != nil { + return syserror.EEXIST + } + if !dir && rp.MustBeDir() { + return syserror.ENOENT + } + // In memfs, the only way to cause a dentry to be disowned is by removing + // it from the filesystem, so this check is equivalent to checking if + // parent has been removed. + if parent.vfsd.IsDisowned() { + return syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + return create(parent, name) +} + +// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. +func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return nil, err + } + if opts.CheckSearchable { + if !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true /* isDir */); err != nil { + return nil, err + } + } + d.IncRef() + return &d.vfsd, nil +} + +// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt. +func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return nil, err + } + d.IncRef() + return &d.vfsd, nil +} + +// LinkAt implements vfs.FilesystemImpl.LinkAt. +func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + if rp.Mount() != vd.Mount() { + return syserror.EXDEV + } + d := vd.Dentry().Impl().(*dentry) + if d.inode.isDir() { + return syserror.EPERM + } + if d.inode.nlink == 0 { + return syserror.ENOENT + } + if d.inode.nlink == maxLinks { + return syserror.EMLINK + } + d.inode.incLinksLocked() + child := fs.newDentry(d.inode) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + }) +} + +// MkdirAt implements vfs.FilesystemImpl.MkdirAt. +func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error { + if parent.inode.nlink == maxLinks { + return syserror.EMLINK + } + parent.inode.incLinksLocked() // from child's ".." + child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + }) +} + +// MknodAt implements vfs.FilesystemImpl.MknodAt. +func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + switch opts.Mode.FileType() { + case 0, linux.S_IFREG: + child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + case linux.S_IFIFO: + child := fs.newDentry(fs.newNamedPipe(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + case linux.S_IFBLK, linux.S_IFCHR, linux.S_IFSOCK: + // Not yet supported. + return syserror.EPERM + default: + return syserror.EINVAL + } + }) +} + +// OpenAt implements vfs.FilesystemImpl.OpenAt. +func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + if opts.Flags&linux.O_TMPFILE != 0 { + // Not yet supported. + return nil, syserror.EOPNOTSUPP + } + + // Handle O_CREAT and !O_CREAT separately, since in the latter case we + // don't need fs.mu for writing. + if opts.Flags&linux.O_CREAT == 0 { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return nil, err + } + return d.open(ctx, rp, opts.Flags, false /* afterCreate */) + } + + mustCreate := opts.Flags&linux.O_EXCL != 0 + start := rp.Start().Impl().(*dentry) + fs.mu.Lock() + defer fs.mu.Unlock() + if rp.Done() { + // Reject attempts to open directories with O_CREAT. + if rp.MustBeDir() { + return nil, syserror.EISDIR + } + if mustCreate { + return nil, syserror.EEXIST + } + return start.open(ctx, rp, opts.Flags, false /* afterCreate */) + } +afterTrailingSymlink: + parent, err := walkParentDirLocked(rp, start) + if err != nil { + return nil, err + } + // Check for search permission in the parent directory. + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } + // Reject attempts to open directories with O_CREAT. + if rp.MustBeDir() { + return nil, syserror.EISDIR + } + name := rp.Component() + if name == "." || name == ".." { + return nil, syserror.EISDIR + } + // Determine whether or not we need to create a file. + child, err := stepLocked(rp, parent) + if err == syserror.ENOENT { + // Already checked for searchability above; now check for writability. + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return nil, err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return nil, err + } + defer rp.Mount().EndWrite() + // Create and open the child. + child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return child.open(ctx, rp, opts.Flags, true) + } + if err != nil { + return nil, err + } + // Do we need to resolve a trailing symlink? + if !rp.Done() { + start = parent + goto afterTrailingSymlink + } + // Open existing file. + if mustCreate { + return nil, syserror.EEXIST + } + return child.open(ctx, rp, opts.Flags, false) +} + +func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, flags uint32, afterCreate bool) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(flags) + if !afterCreate { + if err := d.inode.checkPermissions(rp.Credentials(), ats, d.inode.isDir()); err != nil { + return nil, err + } + } + mnt := rp.Mount() + switch impl := d.inode.impl.(type) { + case *regularFile: + var fd regularFileFD + fd.readable = vfs.MayReadFileWithOpenFlags(flags) + fd.writable = vfs.MayWriteFileWithOpenFlags(flags) + if fd.writable { + if err := mnt.CheckBeginWrite(); err != nil { + return nil, err + } + // mnt.EndWrite() is called by regularFileFD.Release(). + } + fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) + if flags&linux.O_TRUNC != 0 { + impl.mu.Lock() + impl.data.Truncate(0, impl.memFile) + atomic.StoreUint64(&impl.size, 0) + impl.mu.Unlock() + } + return &fd.vfsfd, nil + case *directory: + // Can't open directories writably. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EISDIR + } + var fd directoryFD + fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) + return &fd.vfsfd, nil + case *symlink: + // Can't open symlinks without O_PATH (which is unimplemented). + return nil, syserror.ELOOP + case *namedPipe: + return newNamedPipeFD(ctx, impl, rp, &d.vfsd, flags) + default: + panic(fmt.Sprintf("unknown inode type: %T", d.inode.impl)) + } +} + +// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. +func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return "", err + } + symlink, ok := d.inode.impl.(*symlink) + if !ok { + return "", syserror.EINVAL + } + return symlink.target, nil +} + +// RenameAt implements vfs.FilesystemImpl.RenameAt. +func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error { + if opts.Flags != 0 { + // TODO(b/145974740): Support renameat2 flags. + return syserror.EINVAL + } + + // Resolve newParent first to verify that it's on this Mount. + fs.mu.Lock() + defer fs.mu.Unlock() + newParent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + newName := rp.Component() + if newName == "." || newName == ".." { + return syserror.EBUSY + } + mnt := rp.Mount() + if mnt != oldParentVD.Mount() { + return syserror.EXDEV + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + + oldParent := oldParentVD.Dentry().Impl().(*dentry) + if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + // Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(), + // because if the existing child is a symlink or mount point then we want + // to rename over it rather than follow it. + renamedVFSD := oldParent.vfsd.Child(oldName) + if renamedVFSD == nil { + return syserror.ENOENT + } + renamed := renamedVFSD.Impl().(*dentry) + if renamed.inode.isDir() { + if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) { + return syserror.EINVAL + } + if oldParent != newParent { + // Writability is needed to change renamed's "..". + if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true /* isDir */); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + replacedVFSD := newParent.vfsd.Child(newName) + var replaced *dentry + if replacedVFSD != nil { + replaced = replacedVFSD.Impl().(*dentry) + if replaced.inode.isDir() { + if !renamed.inode.isDir() { + return syserror.EISDIR + } + if replaced.vfsd.HasChildren() { + return syserror.ENOTEMPTY + } + } else { + if rp.MustBeDir() { + return syserror.ENOTDIR + } + if renamed.inode.isDir() { + return syserror.ENOTDIR + } + } + } else { + if renamed.inode.isDir() && newParent.inode.nlink == maxLinks { + return syserror.EMLINK + } + } + if newParent.vfsd.IsDisowned() { + return syserror.ENOENT + } + + // Linux places this check before some of those above; we do it here for + // simplicity, under the assumption that applications are not intentionally + // doing noop renames expecting them to succeed where non-noop renames + // would fail. + if renamedVFSD == replacedVFSD { + return nil + } + vfsObj := rp.VirtualFilesystem() + oldParentDir := oldParent.inode.impl.(*directory) + newParentDir := newParent.inode.impl.(*directory) + if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { + return err + } + if replaced != nil { + newParentDir.childList.Remove(replaced) + if replaced.inode.isDir() { + newParent.inode.decLinksLocked() // from replaced's ".." + } + replaced.inode.decLinksLocked() + } + oldParentDir.childList.Remove(renamed) + newParentDir.childList.PushBack(renamed) + if renamed.inode.isDir() { + oldParent.inode.decLinksLocked() + newParent.inode.incLinksLocked() + } + // TODO: update timestamps and parent directory sizes + vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD) + return nil +} + +// RmdirAt implements vfs.FilesystemImpl.RmdirAt. +func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { + fs.mu.Lock() + defer fs.mu.Unlock() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." { + return syserror.EINVAL + } + if name == ".." { + return syserror.ENOTEMPTY + } + childVFSD := parent.vfsd.Child(name) + if childVFSD == nil { + return syserror.ENOENT + } + child := childVFSD.Impl().(*dentry) + if !child.inode.isDir() { + return syserror.ENOTDIR + } + if childVFSD.HasChildren() { + return syserror.ENOTEMPTY + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + vfsObj := rp.VirtualFilesystem() + if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + return err + } + parent.inode.impl.(*directory).childList.Remove(child) + parent.inode.decLinksLocked() // from child's ".." + child.inode.decLinksLocked() + vfsObj.CommitDeleteDentry(childVFSD) + return nil +} + +// SetStatAt implements vfs.FilesystemImpl.SetStatAt. +func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + if opts.Stat.Mask == 0 { + return nil + } + // TODO: implement inode.setStat + return syserror.EPERM +} + +// StatAt implements vfs.FilesystemImpl.StatAt. +func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + d, err := resolveLocked(rp) + if err != nil { + return linux.Statx{}, err + } + var stat linux.Statx + d.inode.statTo(&stat) + return stat, nil +} + +// StatFSAt implements vfs.FilesystemImpl.StatFSAt. +func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return linux.Statfs{}, err + } + // TODO: actually implement statfs + return linux.Statfs{}, syserror.ENOSYS +} + +// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. +func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + }) +} + +// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. +func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { + fs.mu.Lock() + defer fs.mu.Unlock() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EISDIR + } + childVFSD := parent.vfsd.Child(name) + if childVFSD == nil { + return syserror.ENOENT + } + child := childVFSD.Impl().(*dentry) + if child.inode.isDir() { + return syserror.EISDIR + } + if !rp.MustBeDir() { + return syserror.ENOTDIR + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + vfsObj := rp.VirtualFilesystem() + if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + return err + } + parent.inode.impl.(*directory).childList.Remove(child) + child.inode.decLinksLocked() + vfsObj.CommitDeleteDentry(childVFSD) + return nil +} + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return nil, err + } + // TODO(b/127675828): support extended attributes + return nil, syserror.ENOTSUP +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return "", err + } + // TODO(b/127675828): support extended attributes + return "", syserror.ENOTSUP +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + // TODO(b/127675828): support extended attributes + return syserror.ENOTSUP +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + // TODO(b/127675828): support extended attributes + return syserror.ENOTSUP +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/memfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go index 91cb4b1fc..40bde54de 100644 --- a/pkg/sentry/fsimpl/memfs/named_pipe.go +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -package memfs +package tmpfs import ( "gvisor.dev/gvisor/pkg/abi/linux" @@ -55,8 +55,6 @@ func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, v return nil, err } mnt := rp.Mount() - mnt.IncRef() - vfsd.IncRef() - fd.vfsfd.Init(&fd, mnt, vfsd) + fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) return &fd.vfsfd, nil } diff --git a/pkg/sentry/fsimpl/memfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go index a3a870571..70b42a6ec 100644 --- a/pkg/sentry/fsimpl/memfs/pipe_test.go +++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go @@ -12,13 +12,14 @@ // See the License for the specific language governing permissions and // limitations under the License. -package memfs +package tmpfs import ( "bytes" "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/context/contexttest" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -38,7 +39,7 @@ func TestSeparateFDs(t *testing.T) { pop := vfs.PathOperation{ Root: root, Start: root, - Pathname: fileName, + Path: fspath.Parse(fileName), FollowFinalSymlink: true, } rfdchan := make(chan *vfs.FileDescription) @@ -76,7 +77,7 @@ func TestNonblockingRead(t *testing.T) { pop := vfs.PathOperation{ Root: root, Start: root, - Pathname: fileName, + Path: fspath.Parse(fileName), FollowFinalSymlink: true, } openOpts := vfs.OpenOptions{Flags: linux.O_RDONLY | linux.O_NONBLOCK} @@ -108,7 +109,7 @@ func TestNonblockingWriteError(t *testing.T) { pop := vfs.PathOperation{ Root: root, Start: root, - Pathname: fileName, + Path: fspath.Parse(fileName), FollowFinalSymlink: true, } openOpts := vfs.OpenOptions{Flags: linux.O_WRONLY | linux.O_NONBLOCK} @@ -126,7 +127,7 @@ func TestSingleFD(t *testing.T) { pop := vfs.PathOperation{ Root: root, Start: root, - Pathname: fileName, + Path: fspath.Parse(fileName), FollowFinalSymlink: true, } openOpts := vfs.OpenOptions{Flags: linux.O_RDWR} @@ -151,8 +152,10 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy // Create VFS. vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("memfs", FilesystemType{}) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.GetFilesystemOptions{}) + vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) if err != nil { t.Fatalf("failed to create tmpfs root mount: %v", err) } @@ -160,10 +163,9 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy // Create the pipe. root := mntns.Root() pop := vfs.PathOperation{ - Root: root, - Start: root, - Pathname: fileName, - FollowFinalSymlink: true, + Root: root, + Start: root, + Path: fspath.Parse(fileName), } mknodOpts := vfs.MknodOptions{Mode: linux.ModeNamedPipe | 0644} if err := vfsObj.MknodAt(ctx, creds, &pop, &mknodOpts); err != nil { @@ -174,7 +176,7 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ Root: root, Start: root, - Pathname: fileName, + Path: fspath.Parse(fileName), FollowFinalSymlink: true, }, &vfs.StatOptions{}) if err != nil { @@ -194,7 +196,7 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy func checkEmpty(ctx context.Context, t *testing.T, fd *vfs.FileDescription) { readData := make([]byte, 1) dst := usermem.BytesIOSequence(readData) - bytesRead, err := fd.Impl().Read(ctx, dst, vfs.ReadOptions{}) + bytesRead, err := fd.Read(ctx, dst, vfs.ReadOptions{}) if err != syserror.ErrWouldBlock { t.Fatalf("expected ErrWouldBlock reading from empty pipe %q, but got: %v", fileName, err) } @@ -207,7 +209,7 @@ func checkEmpty(ctx context.Context, t *testing.T, fd *vfs.FileDescription) { func checkWrite(ctx context.Context, t *testing.T, fd *vfs.FileDescription, msg string) { writeData := []byte(msg) src := usermem.BytesIOSequence(writeData) - bytesWritten, err := fd.Impl().Write(ctx, src, vfs.WriteOptions{}) + bytesWritten, err := fd.Write(ctx, src, vfs.WriteOptions{}) if err != nil { t.Fatalf("error writing to pipe %q: %v", fileName, err) } @@ -220,7 +222,7 @@ func checkWrite(ctx context.Context, t *testing.T, fd *vfs.FileDescription, msg func checkRead(ctx context.Context, t *testing.T, fd *vfs.FileDescription, msg string) { readData := make([]byte, len(msg)) dst := usermem.BytesIOSequence(readData) - bytesRead, err := fd.Impl().Read(ctx, dst, vfs.ReadOptions{}) + bytesRead, err := fd.Read(ctx, dst, vfs.ReadOptions{}) if err != nil { t.Fatalf("error reading from pipe %q: %v", fileName, err) } diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go new file mode 100644 index 000000000..f51e247a7 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -0,0 +1,357 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tmpfs + +import ( + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/safemem" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +type regularFile struct { + inode inode + + // memFile is a platform.File used to allocate pages to this regularFile. + memFile *pgalloc.MemoryFile + + // mu protects the fields below. + mu sync.RWMutex + + // data maps offsets into the file to offsets into memFile that store + // the file's data. + data fsutil.FileRangeSet + + // size is the size of data, but accessed using atomic memory + // operations to avoid locking in inode.stat(). + size uint64 + + // seals represents file seals on this inode. + seals uint32 +} + +func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { + file := ®ularFile{ + memFile: fs.memFile, + } + file.inode.init(file, fs, creds, mode) + file.inode.nlink = 1 // from parent directory + return &file.inode +} + +type regularFileFD struct { + fileDescription + + // These are immutable. + readable bool + writable bool + + // off is the file offset. off is accessed using atomic memory operations. + // offMu serializes operations that may mutate off. + off int64 + offMu sync.Mutex +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *regularFileFD) Release() { + if fd.writable { + fd.vfsfd.VirtualDentry().Mount().EndWrite() + } +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if !fd.readable { + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + if dst.NumBytes() == 0 { + return 0, nil + } + f := fd.inode().impl.(*regularFile) + rw := getRegularFileReadWriter(f, offset) + n, err := dst.CopyOutFrom(ctx, rw) + putRegularFileReadWriter(rw) + return int64(n), err +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.offMu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.offMu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + if !fd.writable { + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + srclen := src.NumBytes() + if srclen == 0 { + return 0, nil + } + f := fd.inode().impl.(*regularFile) + end := offset + srclen + if end < offset { + // Overflow. + return 0, syserror.EFBIG + } + rw := getRegularFileReadWriter(f, offset) + n, err := src.CopyInTo(ctx, rw) + putRegularFileReadWriter(rw) + return n, err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.offMu.Lock() + n, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + fd.offMu.Unlock() + return n, err +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.offMu.Lock() + defer fd.offMu.Unlock() + switch whence { + case linux.SEEK_SET: + // use offset as specified + case linux.SEEK_CUR: + offset += fd.off + case linux.SEEK_END: + offset += int64(atomic.LoadUint64(&fd.inode().impl.(*regularFile).size)) + default: + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *regularFileFD) Sync(ctx context.Context) error { + return nil +} + +// regularFileReadWriter implements safemem.Reader and Safemem.Writer. +type regularFileReadWriter struct { + file *regularFile + + // Offset into the file to read/write at. Note that this may be + // different from the FD offset if PRead/PWrite is used. + off uint64 +} + +var regularFileReadWriterPool = sync.Pool{ + New: func() interface{} { + return ®ularFileReadWriter{} + }, +} + +func getRegularFileReadWriter(file *regularFile, offset int64) *regularFileReadWriter { + rw := regularFileReadWriterPool.Get().(*regularFileReadWriter) + rw.file = file + rw.off = uint64(offset) + return rw +} + +func putRegularFileReadWriter(rw *regularFileReadWriter) { + rw.file = nil + regularFileReadWriterPool.Put(rw) +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + rw.file.mu.RLock() + + // Compute the range to read (limited by file size and overflow-checked). + if rw.off >= rw.file.size { + rw.file.mu.RUnlock() + return 0, io.EOF + } + end := rw.file.size + if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end { + end = rend + } + + var done uint64 + seg, gap := rw.file.data.Find(uint64(rw.off)) + for rw.off < end { + mr := memmap.MappableRange{uint64(rw.off), uint64(end)} + switch { + case seg.Ok(): + // Get internal mappings. + ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read) + if err != nil { + rw.file.mu.RUnlock() + return done, err + } + + // Copy from internal mappings. + n, err := safemem.CopySeq(dsts, ims) + done += n + rw.off += uint64(n) + dsts = dsts.DropFirst64(n) + if err != nil { + rw.file.mu.RUnlock() + return done, err + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + // Tmpfs holes are zero-filled. + gapmr := gap.Range().Intersect(mr) + dst := dsts.TakeFirst64(gapmr.Length()) + n, err := safemem.ZeroSeq(dst) + done += n + rw.off += uint64(n) + dsts = dsts.DropFirst64(n) + if err != nil { + rw.file.mu.RUnlock() + return done, err + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } + rw.file.mu.RUnlock() + return done, nil +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + rw.file.mu.Lock() + + // Compute the range to write (overflow-checked). + end := rw.off + srcs.NumBytes() + if end <= rw.off { + end = math.MaxInt64 + } + + // Check if seals prevent either file growth or all writes. + switch { + case rw.file.seals&linux.F_SEAL_WRITE != 0: // Write sealed + rw.file.mu.Unlock() + return 0, syserror.EPERM + case end > rw.file.size && rw.file.seals&linux.F_SEAL_GROW != 0: // Grow sealed + // When growth is sealed, Linux effectively allows writes which would + // normally grow the file to partially succeed up to the current EOF, + // rounded down to the page boundary before the EOF. + // + // This happens because writes (and thus the growth check) for tmpfs + // files proceed page-by-page on Linux, and the final write to the page + // containing EOF fails, resulting in a partial write up to the start of + // that page. + // + // To emulate this behaviour, artifically truncate the write to the + // start of the page containing the current EOF. + // + // See Linux, mm/filemap.c:generic_perform_write() and + // mm/shmem.c:shmem_write_begin(). + if pgstart := uint64(usermem.Addr(rw.file.size).RoundDown()); end > pgstart { + end = pgstart + } + if end <= rw.off { + // Truncation would result in no data being written. + rw.file.mu.Unlock() + return 0, syserror.EPERM + } + } + + // Page-aligned mr for when we need to allocate memory. RoundUp can't + // overflow since end is an int64. + pgstartaddr := usermem.Addr(rw.off).RoundDown() + pgendaddr, _ := usermem.Addr(end).RoundUp() + pgMR := memmap.MappableRange{uint64(pgstartaddr), uint64(pgendaddr)} + + var ( + done uint64 + retErr error + ) + seg, gap := rw.file.data.Find(uint64(rw.off)) + for rw.off < end { + mr := memmap.MappableRange{uint64(rw.off), uint64(end)} + switch { + case seg.Ok(): + // Get internal mappings. + ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Write) + if err != nil { + retErr = err + goto exitLoop + } + + // Copy to internal mappings. + n, err := safemem.CopySeq(ims, srcs) + done += n + rw.off += uint64(n) + srcs = srcs.DropFirst64(n) + if err != nil { + retErr = err + goto exitLoop + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + // Allocate memory for the write. + gapMR := gap.Range().Intersect(pgMR) + fr, err := rw.file.memFile.Allocate(gapMR.Length(), usage.Tmpfs) + if err != nil { + retErr = err + goto exitLoop + } + + // Write to that memory as usual. + seg, gap = rw.file.data.Insert(gap, gapMR, fr.Start), fsutil.FileRangeGapIterator{} + } + } +exitLoop: + // If the write ends beyond the file's previous size, it causes the + // file to grow. + if rw.off > rw.file.size { + atomic.StoreUint64(&rw.file.size, rw.off) + } + + rw.file.mu.Unlock() + return done, retErr +} diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go new file mode 100644 index 000000000..3731c5b6f --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go @@ -0,0 +1,224 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tmpfs + +import ( + "bytes" + "fmt" + "io" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// newFileFD creates a new file in a new tmpfs mount, and returns the FD. If +// the returned err is not nil, then cleanup should be called when the FD is no +// longer needed. +func newFileFD(ctx context.Context, filename string) (*vfs.FileDescription, func(), error) { + creds := auth.CredentialsFromContext(ctx) + + vfsObj := vfs.New() + vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + if err != nil { + return nil, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err) + } + root := mntns.Root() + + // Create the file that will be write/read. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + FollowFinalSymlink: true, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, + Mode: 0644, + }) + if err != nil { + root.DecRef() + mntns.DecRef(vfsObj) + return nil, nil, fmt.Errorf("failed to create file %q: %v", filename, err) + } + + return fd, func() { + root.DecRef() + mntns.DecRef(vfsObj) + }, nil +} + +// Test that we can write some data to a file and read it back.` +func TestSimpleWriteRead(t *testing.T) { + ctx := contexttest.Context(t) + fd, cleanup, err := newFileFD(ctx, "simpleReadWrite") + if err != nil { + t.Fatal(err) + } + defer cleanup() + + // Write. + data := []byte("foobarbaz") + n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + t.Fatalf("fd.Write failed: %v", err) + } + if n != int64(len(data)) { + t.Errorf("fd.Write got short write length %d, want %d", n, len(data)) + } + if got, want := fd.Impl().(*regularFileFD).off, int64(len(data)); got != want { + t.Errorf("fd.Write left offset at %d, want %d", got, want) + } + + // Seek back to beginning. + if _, err := fd.Seek(ctx, 0, linux.SEEK_SET); err != nil { + t.Fatalf("fd.Seek failed: %v", err) + } + if got, want := fd.Impl().(*regularFileFD).off, int64(0); got != want { + t.Errorf("fd.Seek(0) left offset at %d, want %d", got, want) + } + + // Read. + buf := make([]byte, len(data)) + n, err = fd.Read(ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) + if err != nil && err != io.EOF { + t.Fatalf("fd.Read failed: %v", err) + } + if n != int64(len(data)) { + t.Errorf("fd.Read got short read length %d, want %d", n, len(data)) + } + if got, want := string(buf), string(data); got != want { + t.Errorf("Read got %q want %s", got, want) + } + if got, want := fd.Impl().(*regularFileFD).off, int64(len(data)); got != want { + t.Errorf("fd.Write left offset at %d, want %d", got, want) + } +} + +func TestPWrite(t *testing.T) { + ctx := contexttest.Context(t) + fd, cleanup, err := newFileFD(ctx, "PRead") + if err != nil { + t.Fatal(err) + } + defer cleanup() + + // Fill file with 1k 'a's. + data := bytes.Repeat([]byte{'a'}, 1000) + n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + t.Fatalf("fd.Write failed: %v", err) + } + if n != int64(len(data)) { + t.Errorf("fd.Write got short write length %d, want %d", n, len(data)) + } + + // Write "gVisor is awesome" at various offsets. + buf := []byte("gVisor is awesome") + offsets := []int{0, 1, 2, 10, 20, 50, 100, len(data) - 100, len(data) - 1, len(data), len(data) + 1} + for _, offset := range offsets { + name := fmt.Sprintf("PWrite offset=%d", offset) + t.Run(name, func(t *testing.T) { + n, err := fd.PWrite(ctx, usermem.BytesIOSequence(buf), int64(offset), vfs.WriteOptions{}) + if err != nil { + t.Errorf("fd.PWrite got err %v want nil", err) + } + if n != int64(len(buf)) { + t.Errorf("fd.PWrite got %d bytes want %d", n, len(buf)) + } + + // Update data to reflect expected file contents. + if len(data) < offset+len(buf) { + data = append(data, make([]byte, (offset+len(buf))-len(data))...) + } + copy(data[offset:], buf) + + // Read the whole file and compare with data. + readBuf := make([]byte, len(data)) + n, err = fd.PRead(ctx, usermem.BytesIOSequence(readBuf), 0, vfs.ReadOptions{}) + if err != nil { + t.Fatalf("fd.PRead failed: %v", err) + } + if n != int64(len(data)) { + t.Errorf("fd.PRead got short read length %d, want %d", n, len(data)) + } + if got, want := string(readBuf), string(data); got != want { + t.Errorf("PRead got %q want %s", got, want) + } + + }) + } +} + +func TestPRead(t *testing.T) { + ctx := contexttest.Context(t) + fd, cleanup, err := newFileFD(ctx, "PRead") + if err != nil { + t.Fatal(err) + } + defer cleanup() + + // Write 100 sequences of 'gVisor is awesome'. + data := bytes.Repeat([]byte("gVisor is awsome"), 100) + n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + t.Fatalf("fd.Write failed: %v", err) + } + if n != int64(len(data)) { + t.Errorf("fd.Write got short write length %d, want %d", n, len(data)) + } + + // Read various sizes from various offsets. + sizes := []int{0, 1, 2, 10, 20, 50, 100, 1000} + offsets := []int{0, 1, 2, 10, 20, 50, 100, 1000, len(data) - 100, len(data) - 1, len(data), len(data) + 1} + + for _, size := range sizes { + for _, offset := range offsets { + name := fmt.Sprintf("PRead offset=%d size=%d", offset, size) + t.Run(name, func(t *testing.T) { + var ( + wantRead []byte + wantErr error + ) + if offset < len(data) { + wantRead = data[offset:] + } else if size > 0 { + wantErr = io.EOF + } + if offset+size < len(data) { + wantRead = wantRead[:size] + } + buf := make([]byte, size) + n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), int64(offset), vfs.ReadOptions{}) + if err != wantErr { + t.Errorf("fd.PRead got err %v want %v", err, wantErr) + } + if n != int64(len(wantRead)) { + t.Errorf("fd.PRead got %d bytes want %d", n, len(wantRead)) + } + if got := string(buf[:n]); got != string(wantRead) { + t.Errorf("fd.PRead got %q want %q", got, string(wantRead)) + } + }) + } + } +} diff --git a/pkg/sentry/fsimpl/memfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go index b2ac2cbeb..5246aca84 100644 --- a/pkg/sentry/fsimpl/memfs/symlink.go +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -package memfs +package tmpfs import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" diff --git a/pkg/sentry/fsimpl/memfs/memfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 4cb2a4e0f..7be6faa5b 100644 --- a/pkg/sentry/fsimpl/memfs/memfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -12,29 +12,27 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package memfs provides a filesystem implementation that behaves like tmpfs: +// Package tmpfs provides a filesystem implementation that behaves like tmpfs: // the Dentry tree is the sole source of truth for the state of the filesystem. // -// memfs is intended primarily to demonstrate filesystem implementation -// patterns. Real uses cases for an in-memory filesystem should use tmpfs -// instead. -// // Lock order: // // filesystem.mu // regularFileFD.offMu // regularFile.mu // inode.mu -package memfs +package tmpfs import ( "fmt" + "math" "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -46,6 +44,9 @@ type FilesystemType struct{} type filesystem struct { vfsfs vfs.Filesystem + // memFile is used to allocate pages to for regular files. + memFile *pgalloc.MemoryFile + // mu serializes changes to the Dentry tree. mu sync.RWMutex @@ -54,7 +55,13 @@ type filesystem struct { // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { - var fs filesystem + memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx) + if memFileProvider == nil { + panic("MemoryFileProviderFromContext returned nil") + } + fs := filesystem{ + memFile: memFileProvider.MemoryFile(), + } fs.vfsfs.Init(vfsObj, &fs) root := fs.newDentry(fs.newDirectory(creds, 01777)) return &fs.vfsfs, &root.vfsd, nil @@ -64,12 +71,6 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt func (fs *filesystem) Release() { } -// Sync implements vfs.FilesystemImpl.Sync. -func (fs *filesystem) Sync(ctx context.Context) error { - // All filesystem state is in-memory. - return nil -} - // dentry implements vfs.DentryImpl. type dentry struct { vfsd vfs.Dentry @@ -79,11 +80,11 @@ type dentry struct { // immutable. inode *inode - // memfs doesn't count references on dentries; because the dentry tree is + // tmpfs doesn't count references on dentries; because the dentry tree is // the sole source of truth, it is by definition always consistent with the // state of the filesystem. However, it does count references on inodes, // because inode resources are released when all references are dropped. - // (memfs doesn't really have resources to release, but we implement + // (tmpfs doesn't really have resources to release, but we implement // reference counting because tmpfs regular files will.) // dentryEntry (ugh) links dentries into their parent directory.childList. @@ -137,6 +138,8 @@ type inode struct { impl interface{} // immutable } +const maxLinks = math.MaxUint32 + func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) { i.refs = 1 i.mode = uint32(mode) @@ -147,25 +150,33 @@ func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, i.impl = impl } -// Preconditions: filesystem.mu must be locked for writing. +// incLinksLocked increments i's link count. +// +// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. +// i.nlink < maxLinks. func (i *inode) incLinksLocked() { - if atomic.AddUint32(&i.nlink, 1) <= 1 { - panic("memfs.inode.incLinksLocked() called with no existing links") + if i.nlink == 0 { + panic("tmpfs.inode.incLinksLocked() called with no existing links") } + if i.nlink == maxLinks { + panic("memfs.inode.incLinksLocked() called with maximum link count") + } + atomic.AddUint32(&i.nlink, 1) } -// Preconditions: filesystem.mu must be locked for writing. +// decLinksLocked decrements i's link count. +// +// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. func (i *inode) decLinksLocked() { - if nlink := atomic.AddUint32(&i.nlink, ^uint32(0)); nlink == 0 { - i.decRef() - } else if nlink == ^uint32(0) { // negative overflow - panic("memfs.inode.decLinksLocked() called with no existing links") + if i.nlink == 0 { + panic("tmpfs.inode.decLinksLocked() called with no existing links") } + atomic.AddUint32(&i.nlink, ^uint32(0)) } func (i *inode) incRef() { if atomic.AddInt64(&i.refs, 1) <= 1 { - panic("memfs.inode.incRef() called without holding a reference") + panic("tmpfs.inode.incRef() called without holding a reference") } } @@ -184,14 +195,14 @@ func (i *inode) tryIncRef() bool { func (i *inode) decRef() { if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { // This is unnecessary; it's mostly to simulate what tmpfs would do. - if regfile, ok := i.impl.(*regularFile); ok { - regfile.mu.Lock() - regfile.data = nil - atomic.StoreInt64(®file.dataLen, 0) - regfile.mu.Unlock() + if regFile, ok := i.impl.(*regularFile); ok { + regFile.mu.Lock() + regFile.data.DropAll(regFile.memFile) + atomic.StoreUint64(®File.size, 0) + regFile.mu.Unlock() } } else if refs < 0 { - panic("memfs.inode.decRef() called without holding a reference") + panic("tmpfs.inode.decRef() called without holding a reference") } } @@ -215,7 +226,7 @@ func (i *inode) statTo(stat *linux.Statx) { case *regularFile: stat.Mode |= linux.S_IFREG stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS - stat.Size = uint64(atomic.LoadInt64(&impl.dataLen)) + stat.Size = uint64(atomic.LoadUint64(&impl.size)) // In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in // a uint64 accessed using atomic memory operations to avoid taking // locks). @@ -256,13 +267,11 @@ func (i *inode) direntType() uint8 { } } -// fileDescription is embedded by memfs implementations of +// fileDescription is embedded by tmpfs implementations of // vfs.FileDescriptionImpl. type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl - - flags uint32 // status flags; immutable } func (fd *fileDescription) filesystem() *filesystem { @@ -273,18 +282,6 @@ func (fd *fileDescription) inode() *inode { return fd.vfsfd.Dentry().Impl().(*dentry).inode } -// StatusFlags implements vfs.FileDescriptionImpl.StatusFlags. -func (fd *fileDescription) StatusFlags(ctx context.Context) (uint32, error) { - return fd.flags, nil -} - -// SetStatusFlags implements vfs.FileDescriptionImpl.SetStatusFlags. -func (fd *fileDescription) SetStatusFlags(ctx context.Context, flags uint32) error { - // None of the flags settable by fcntl(F_SETFL) are supported, so this is a - // no-op. - return nil -} - // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 28ba950bd..8653d2f63 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -762,7 +762,7 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, mounts.IncRef() } - tg := k.newThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits, k.monotonicClock) + tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) // Get the root directory from the MountNamespace. @@ -841,9 +841,11 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, AbstractSocketNamespace: args.AbstractSocketNamespace, ContainerID: args.ContainerID, } - if _, err := k.tasks.NewTask(config); err != nil { + t, err := k.tasks.NewTask(config) + if err != nil { return nil, 0, err } + t.traceExecEvent(tc) // Simulate exec for tracing. // Success. tgid := k.tasks.Root.IDOfThreadGroup(tg) @@ -1118,6 +1120,22 @@ func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { return lastErr } +// RebuildTraceContexts rebuilds the trace context for all tasks. +// +// Unfortunately, if these are built while tracing is not enabled, then we will +// not have meaningful trace data. Rebuilding here ensures that we can do so +// after tracing has been enabled. +func (k *Kernel) RebuildTraceContexts() { + k.extMu.Lock() + defer k.extMu.Unlock() + k.tasks.mu.RLock() + defer k.tasks.mu.RUnlock() + + for t, tid := range k.tasks.Root.tids { + t.rebuildTraceContext(tid) + } +} + // FeatureSet returns the FeatureSet. func (k *Kernel) FeatureSet() *cpuid.FeatureSet { return k.featureSet @@ -1173,6 +1191,11 @@ func (k *Kernel) GlobalInit() *ThreadGroup { return k.globalInit } +// TestOnly_SetGlobalInit sets the thread group with ID 1 in the root PID namespace. +func (k *Kernel) TestOnly_SetGlobalInit(tg *ThreadGroup) { + k.globalInit = tg +} + // ApplicationCores returns the number of CPUs visible to sandboxed // applications. func (k *Kernel) ApplicationCores() uint { diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index 24ea002ba..b14429854 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -15,17 +15,29 @@ package kernel import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" ) -// Restartable sequences, as described in https://lwn.net/Articles/650333/. +// Restartable sequences. +// +// We support two different APIs for restartable sequences. +// +// 1. The upstream interface added in v4.18. +// 2. The interface described in https://lwn.net/Articles/650333/. +// +// Throughout this file and other parts of the kernel, the latter is referred +// to as "old rseq". This interface was never merged upstream, but is supported +// for a limited set of applications that use it regardless. -// RSEQCriticalRegion describes a restartable sequence critical region. +// OldRSeqCriticalRegion describes an old rseq critical region. // // +stateify savable -type RSEQCriticalRegion struct { +type OldRSeqCriticalRegion struct { // When a task in this thread group has its CPU preempted (as defined by // platform.ErrContextCPUPreempted) or has a signal delivered to an // application handler while its instruction pointer is in CriticalSection, @@ -35,86 +47,359 @@ type RSEQCriticalRegion struct { Restart usermem.Addr } -// RSEQAvailable returns true if t supports restartable sequences. -func (t *Task) RSEQAvailable() bool { +// RSeqAvailable returns true if t supports (old and new) restartable sequences. +func (t *Task) RSeqAvailable() bool { return t.k.useHostCores && t.k.Platform.DetectsCPUPreemption() } -// RSEQCriticalRegion returns a copy of t's thread group's current restartable -// sequence. -func (t *Task) RSEQCriticalRegion() RSEQCriticalRegion { - return *t.tg.rscr.Load().(*RSEQCriticalRegion) +// SetRSeq registers addr as this thread's rseq structure. +// +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) SetRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr != 0 { + if t.rseqAddr != addr { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EINVAL + } + return syserror.EBUSY + } + + // rseq must be aligned and correctly sized. + if addr&(linux.AlignOfRSeq-1) != 0 { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if _, ok := t.MemoryManager().CheckIORange(addr, linux.SizeOfRSeq); !ok { + return syserror.EFAULT + } + + t.rseqAddr = addr + t.rseqSignature = signature + + // Initialize the CPUID. + // + // Linux implicitly does this on return from userspace, where failure + // would cause SIGSEGV. + if err := t.rseqUpdateCPU(); err != nil { + t.rseqAddr = 0 + t.rseqSignature = 0 + + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return syserror.EFAULT + } + + return nil } -// SetRSEQCriticalRegion replaces t's thread group's restartable sequence. +// ClearRSeq unregisters addr as this thread's rseq structure. // -// Preconditions: t.RSEQAvailable() == true. -func (t *Task) SetRSEQCriticalRegion(rscr RSEQCriticalRegion) error { +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) ClearRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr == 0 { + return syserror.EINVAL + } + if t.rseqAddr != addr { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EPERM + } + + if err := t.rseqClearCPU(); err != nil { + return err + } + + t.rseqAddr = 0 + t.rseqSignature = 0 + + if t.oldRSeqCPUAddr == 0 { + // rseqCPU no longer needed. + t.rseqCPU = -1 + } + + return nil +} + +// OldRSeqCriticalRegion returns a copy of t's thread group's current +// old restartable sequence. +func (t *Task) OldRSeqCriticalRegion() OldRSeqCriticalRegion { + return *t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) +} + +// SetOldRSeqCriticalRegion replaces t's thread group's old restartable +// sequence. +// +// Preconditions: t.RSeqAvailable() == true. +func (t *Task) SetOldRSeqCriticalRegion(r OldRSeqCriticalRegion) error { // These checks are somewhat more lenient than in Linux, which (bizarrely) - // requires rscr.CriticalSection to be non-empty and rscr.Restart to be - // outside of rscr.CriticalSection, even if rscr.CriticalSection.Start == 0 + // requires r.CriticalSection to be non-empty and r.Restart to be + // outside of r.CriticalSection, even if r.CriticalSection.Start == 0 // (which disables the critical region). - if rscr.CriticalSection.Start == 0 { - rscr.CriticalSection.End = 0 - rscr.Restart = 0 - t.tg.rscr.Store(&rscr) + if r.CriticalSection.Start == 0 { + r.CriticalSection.End = 0 + r.Restart = 0 + t.tg.oldRSeqCritical.Store(&r) return nil } - if rscr.CriticalSection.Start >= rscr.CriticalSection.End { + if r.CriticalSection.Start >= r.CriticalSection.End { return syserror.EINVAL } - if rscr.CriticalSection.Contains(rscr.Restart) { + if r.CriticalSection.Contains(r.Restart) { return syserror.EINVAL } - // TODO(jamieliu): check that rscr.CriticalSection and rscr.Restart are in - // the application address range, for consistency with Linux - t.tg.rscr.Store(&rscr) + // TODO(jamieliu): check that r.CriticalSection and r.Restart are in + // the application address range, for consistency with Linux. + t.tg.oldRSeqCritical.Store(&r) return nil } -// RSEQCPUAddr returns the address that RSEQ will keep updated with t's CPU -// number. +// OldRSeqCPUAddr returns the address that old rseq will keep updated with t's +// CPU number. // // Preconditions: The caller must be running on the task goroutine. -func (t *Task) RSEQCPUAddr() usermem.Addr { - return t.rseqCPUAddr +func (t *Task) OldRSeqCPUAddr() usermem.Addr { + return t.oldRSeqCPUAddr } -// SetRSEQCPUAddr replaces the address that RSEQ will keep updated with t's CPU -// number. +// SetOldRSeqCPUAddr replaces the address that old rseq will keep updated with +// t's CPU number. // -// Preconditions: t.RSEQAvailable() == true. The caller must be running on the +// Preconditions: t.RSeqAvailable() == true. The caller must be running on the // task goroutine. t's AddressSpace must be active. -func (t *Task) SetRSEQCPUAddr(addr usermem.Addr) error { - t.rseqCPUAddr = addr - if addr != 0 { - t.rseqCPU = int32(hostcpu.GetCPU()) - if err := t.rseqCopyOutCPU(); err != nil { - t.rseqCPUAddr = 0 - t.rseqCPU = -1 - return syserror.EINVAL // yes, EINVAL, not err or EFAULT - } - } else { - t.rseqCPU = -1 +func (t *Task) SetOldRSeqCPUAddr(addr usermem.Addr) error { + t.oldRSeqCPUAddr = addr + + // Check that addr is writable. + // + // N.B. rseqUpdateCPU may fail on a bad t.rseqAddr as well. That's + // unfortunate, but unlikely in a correct program. + if err := t.rseqUpdateCPU(); err != nil { + t.oldRSeqCPUAddr = 0 + return syserror.EINVAL // yes, EINVAL, not err or EFAULT } return nil } // Preconditions: The caller must be running on the task goroutine. t's // AddressSpace must be active. -func (t *Task) rseqCopyOutCPU() error { +func (t *Task) rseqUpdateCPU() error { + if t.rseqAddr == 0 && t.oldRSeqCPUAddr == 0 { + t.rseqCPU = -1 + return nil + } + + t.rseqCPU = int32(hostcpu.GetCPU()) + + // Update both CPUs, even if one fails. + rerr := t.rseqCopyOutCPU() + oerr := t.oldRSeqCopyOutCPU() + + if rerr != nil { + return rerr + } + return oerr +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) oldRSeqCopyOutCPU() error { + if t.oldRSeqCPUAddr == 0 { + return nil + } + buf := t.CopyScratchBuffer(4) usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) - _, err := t.CopyOutBytes(t.rseqCPUAddr, buf) + _, err := t.CopyOutBytes(t.oldRSeqCPUAddr, buf) + return err +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqCopyOutCPU() error { + if t.rseqAddr == 0 { + return nil + } + + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], uint32(t.rseqCPU)) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) + return err +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqClearCPU() error { + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, 0) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], linux.RSEQ_CPU_ID_UNINITIALIZED) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) return err } +// rseqAddrInterrupt checks if IP is in a critical section, and aborts if so. +// +// This is a bit complex since both the RSeq and RSeqCriticalSection structs +// are stored in userspace. So we must: +// +// 1. Copy in the address of RSeqCriticalSection from RSeq. +// 2. Copy in RSeqCriticalSection itself. +// 3. Validate critical section struct version, address range, abort address. +// 4. Validate the abort signature (4 bytes preceding abort IP match expected +// signature). +// 5. Clear address of RSeqCriticalSection from RSeq. +// 6. Finally, conditionally abort. +// +// See kernel/rseq.c:rseq_ip_fixup for reference. +// +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqAddrInterrupt() { + if t.rseqAddr == 0 { + return + } + + critAddrAddr, ok := t.rseqAddr.AddLength(linux.OffsetOfRSeqCriticalSection) + if !ok { + // SetRSeq should validate this. + panic(fmt.Sprintf("t.rseqAddr (%#x) not large enough", t.rseqAddr)) + } + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + t.Debugf("Only 64-bit rseq supported.") + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + buf := t.CopyScratchBuffer(8) + if _, err := t.CopyInBytes(critAddrAddr, buf); err != nil { + t.Debugf("Failed to copy critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + critAddr := usermem.Addr(usermem.ByteOrder.Uint64(buf)) + if critAddr == 0 { + return + } + + buf = t.CopyScratchBuffer(linux.SizeOfRSeqCriticalSection) + if _, err := t.CopyInBytes(critAddr, buf); err != nil { + t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Manually marshal RSeqCriticalSection as this is in the hot path when + // rseq is enabled. It must be as fast as possible. + // + // TODO(b/130243041): Replace with go_marshal. + cs := linux.RSeqCriticalSection{ + Version: usermem.ByteOrder.Uint32(buf[0:4]), + Flags: usermem.ByteOrder.Uint32(buf[4:8]), + Start: usermem.ByteOrder.Uint64(buf[8:16]), + PostCommitOffset: usermem.ByteOrder.Uint64(buf[16:24]), + Abort: usermem.ByteOrder.Uint64(buf[24:32]), + } + + if cs.Version != 0 { + t.Debugf("Unknown version in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + start := usermem.Addr(cs.Start) + critRange, ok := start.ToRange(cs.PostCommitOffset) + if !ok { + t.Debugf("Invalid start and offset in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + abort := usermem.Addr(cs.Abort) + if critRange.Contains(abort) { + t.Debugf("Abort in critical section in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Verify signature. + sigAddr := abort - linux.SizeOfRSeqSignature + + buf = t.CopyScratchBuffer(linux.SizeOfRSeqSignature) + if _, err := t.CopyInBytes(sigAddr, buf); err != nil { + t.Debugf("Failed to copy critical section signature from %#x for rseq: %v", sigAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + sig := usermem.ByteOrder.Uint32(buf) + if sig != t.rseqSignature { + t.Debugf("Mismatched rseq signature %d != %d", sig, t.rseqSignature) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Clear the critical section address. + // + // NOTE(b/143949567): We don't support any rseq flags, so we always + // restart if we are in the critical section, and thus *always* clear + // critAddrAddr. + if _, err := t.MemoryManager().ZeroOut(t, critAddrAddr, int64(t.Arch().Width()), usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + t.Debugf("Failed to clear critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Finally we can actually decide whether or not to restart. + if !critRange.Contains(usermem.Addr(t.Arch().IP())) { + return + } + + t.Arch().SetIP(uintptr(cs.Abort)) +} + // Preconditions: The caller must be running on the task goroutine. -func (t *Task) rseqInterrupt() { - rscr := t.tg.rscr.Load().(*RSEQCriticalRegion) - if ip := t.Arch().IP(); rscr.CriticalSection.Contains(usermem.Addr(ip)) { - t.Debugf("Interrupted RSEQ critical section at %#x; restarting at %#x", ip, rscr.Restart) - t.Arch().SetIP(uintptr(rscr.Restart)) - t.Arch().SetRSEQInterruptedIP(ip) +func (t *Task) oldRSeqInterrupt() { + r := t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) + if ip := t.Arch().IP(); r.CriticalSection.Contains(usermem.Addr(ip)) { + t.Debugf("Interrupted rseq critical section at %#x; restarting at %#x", ip, r.Restart) + t.Arch().SetIP(uintptr(r.Restart)) + t.Arch().SetOldRSeqInterruptedIP(ip) } } + +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) rseqInterrupt() { + t.rseqAddrInterrupt() + t.oldRSeqInterrupt() +} diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go index 93fe68a3e..de9617e9d 100644 --- a/pkg/sentry/kernel/semaphore/semaphore.go +++ b/pkg/sentry/kernel/semaphore/semaphore.go @@ -302,7 +302,7 @@ func (s *Set) SetVal(ctx context.Context, num int32, val int16, creds *auth.Cred return syserror.ERANGE } - // TODO(b/29354920): Clear undo entries in all processes + // TODO(gvisor.dev/issue/137): Clear undo entries in all processes. sem.value = val sem.pid = pid s.changeTime = ktime.NowFromContext(ctx) @@ -336,7 +336,7 @@ func (s *Set) SetValAll(ctx context.Context, vals []uint16, creds *auth.Credenti for i, val := range vals { sem := &s.sems[i] - // TODO(b/29354920): Clear undo entries in all processes + // TODO(gvisor.dev/issue/137): Clear undo entries in all processes. sem.value = int16(val) sem.pid = pid sem.wakeWaiters() @@ -481,7 +481,7 @@ func (s *Set) executeOps(ctx context.Context, ops []linux.Sembuf, pid int32) (ch } // All operations succeeded, apply them. - // TODO(b/29354920): handle undo operations. + // TODO(gvisor.dev/issue/137): handle undo operations. for i, v := range tmpVals { s.sems[i].value = v s.sems[i].wakeWaiters() diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 5bd610f68..19034a21e 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -71,9 +71,20 @@ type Registry struct { mu sync.Mutex `state:"nosave"` // shms maps segment ids to segments. + // + // shms holds all referenced segments, which are removed on the last + // DecRef. Thus, it cannot itself hold a reference on the Shm. + // + // Since removal only occurs after the last (unlocked) DecRef, there + // exists a short window during which a Shm still exists in Shm, but is + // unreferenced. Users must use TryIncRef to determine if the Shm is + // still valid. shms map[ID]*Shm // keysToShms maps segment keys to segments. + // + // Shms in keysToShms are guaranteed to be referenced, as they are + // removed by disassociateKey before the last DecRef. keysToShms map[Key]*Shm // Sum of the sizes of all existing segments rounded up to page size, in @@ -95,10 +106,18 @@ func NewRegistry(userNS *auth.UserNamespace) *Registry { } // FindByID looks up a segment given an ID. +// +// FindByID returns a reference on Shm. func (r *Registry) FindByID(id ID) *Shm { r.mu.Lock() defer r.mu.Unlock() - return r.shms[id] + s := r.shms[id] + // Take a reference on s. If TryIncRef fails, s has reached the last + // DecRef, but hasn't quite been removed from r.shms yet. + if s != nil && s.TryIncRef() { + return s + } + return nil } // dissociateKey removes the association between a segment and its key, @@ -119,6 +138,8 @@ func (r *Registry) dissociateKey(s *Shm) { // FindOrCreate looks up or creates a segment in the registry. It's functionally // analogous to open(2). +// +// FindOrCreate returns a reference on Shm. func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size uint64, mode linux.FileMode, private, create, exclusive bool) (*Shm, error) { if (create || private) && (size < linux.SHMMIN || size > linux.SHMMAX) { // "A new segment was to be created and size is less than SHMMIN or @@ -166,6 +187,7 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui return nil, syserror.EEXIST } + shm.IncRef() return shm, nil } @@ -193,7 +215,14 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui // Need to create a new segment. creator := fs.FileOwnerFromContext(ctx) perms := fs.FilePermsFromMode(mode) - return r.newShm(ctx, pid, key, creator, perms, size) + s, err := r.newShm(ctx, pid, key, creator, perms, size) + if err != nil { + return nil, err + } + // The initial reference is held by s itself. Take another to return to + // the caller. + s.IncRef() + return s, nil } // newShm creates a new segment in the registry. @@ -296,22 +325,26 @@ func (r *Registry) remove(s *Shm) { // Shm represents a single shared memory segment. // -// Shm segment are backed directly by an allocation from platform -// memory. Segments are always mapped as a whole, greatly simplifying how -// mappings are tracked. However note that mremap and munmap calls may cause the -// vma for a segment to become fragmented; which requires special care when -// unmapping a segment. See mm/shm.go. +// Shm segment are backed directly by an allocation from platform memory. +// Segments are always mapped as a whole, greatly simplifying how mappings are +// tracked. However note that mremap and munmap calls may cause the vma for a +// segment to become fragmented; which requires special care when unmapping a +// segment. See mm/shm.go. // // Segments persist until they are explicitly marked for destruction via -// shmctl(SHM_RMID). +// MarkDestroyed(). // // Shm implements memmap.Mappable and memmap.MappingIdentity. // // +stateify savable type Shm struct { - // AtomicRefCount tracks the number of references to this segment from - // maps. A segment always holds a reference to itself, until it's marked for + // AtomicRefCount tracks the number of references to this segment. + // + // A segment holds a reference to itself until it is marked for // destruction. + // + // In addition to direct users, the MemoryManager will hold references + // via MappingIdentity. refs.AtomicRefCount mfp pgalloc.MemoryFileProvider @@ -484,9 +517,8 @@ type AttachOpts struct { // ConfigureAttach creates an mmap configuration for the segment with the // requested attach options. // -// ConfigureAttach returns with a ref on s on success. The caller should drop -// this once the map is installed. This reference prevents s from being -// destroyed before the returned configuration is used. +// Postconditions: The returned MMapOpts are valid only as long as a reference +// continues to be held on s. func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts AttachOpts) (memmap.MMapOpts, error) { s.mu.Lock() defer s.mu.Unlock() @@ -504,7 +536,6 @@ func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts Attac // in the user namespace that governs its IPC namespace." - man shmat(2) return memmap.MMapOpts{}, syserror.EACCES } - s.IncRef() return memmap.MMapOpts{ Length: s.size, Offset: 0, @@ -549,10 +580,15 @@ func (s *Shm) IPCStat(ctx context.Context) (*linux.ShmidDS, error) { } creds := auth.CredentialsFromContext(ctx) - nattach := uint64(s.ReadRefs()) - // Don't report the self-reference we keep prior to being marked for - // destruction. However, also don't report a count of -1 for segments marked - // as destroyed, with no mappings. + // Use the reference count as a rudimentary count of the number of + // attaches. We exclude: + // + // 1. The reference the caller holds. + // 2. The self-reference held by s prior to destruction. + // + // Note that this may still overcount by including transient references + // used in concurrent calls. + nattach := uint64(s.ReadRefs()) - 1 if !s.pendingDestruction { nattach-- } @@ -620,18 +656,17 @@ func (s *Shm) MarkDestroyed() { s.registry.dissociateKey(s) s.mu.Lock() - // Only drop the segment's self-reference once, when destruction is - // requested. Otherwise, repeated calls to shmctl(IPC_RMID) would force a - // segment to be destroyed prematurely, potentially with active maps to the - // segment's address range. Remaining references are dropped when the - // segment is detached or unmaped. + defer s.mu.Unlock() if !s.pendingDestruction { s.pendingDestruction = true - s.mu.Unlock() // Must release s.mu before calling s.DecRef. + // Drop the self-reference so destruction occurs when all + // external references are gone. + // + // N.B. This cannot be the final DecRef, as the caller also + // holds a reference. s.DecRef() return } - s.mu.Unlock() } // checkOwnership verifies whether a segment may be accessed by ctx as an diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go index 220fa73a2..2fdee0282 100644 --- a/pkg/sentry/kernel/syscalls.go +++ b/pkg/sentry/kernel/syscalls.go @@ -339,6 +339,14 @@ func (s *SyscallTable) Lookup(sysno uintptr) SyscallFn { return nil } +// LookupName looks up a syscall name. +func (s *SyscallTable) LookupName(sysno uintptr) string { + if sc, ok := s.Table[sysno]; ok { + return sc.Name + } + return fmt.Sprintf("sys_%d", sysno) // Unlikely. +} + // LookupEmulate looks up an emulation syscall number. func (s *SyscallTable) LookupEmulate(addr usermem.Addr) (uintptr, bool) { sysno, ok := s.Emulate[addr] diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 80c8e5464..d25a7903b 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -15,6 +15,8 @@ package kernel import ( + gocontext "context" + "runtime/trace" "sync" "sync/atomic" @@ -390,7 +392,14 @@ type Task struct { // logPrefix is a string containing the task's thread ID in the root PID // namespace, and is prepended to log messages emitted by Task.Infof etc. - logPrefix atomic.Value `state:".(string)"` + logPrefix atomic.Value `state:"nosave"` + + // traceContext and traceTask are both used for tracing, and are + // updated along with the logPrefix in updateInfoLocked. + // + // These are exclusive to the task goroutine. + traceContext gocontext.Context `state:"nosave"` + traceTask *trace.Task `state:"nosave"` // creds is the task's credentials. // @@ -480,18 +489,43 @@ type Task struct { // netns is protected by mu. netns is owned by the task goroutine. netns bool - // If rseqPreempted is true, before the next call to p.Switch(), interrupt - // RSEQ critical regions as defined by tg.rseq and write the task - // goroutine's CPU number to rseqCPUAddr. rseqCPU is the last CPU number - // written to rseqCPUAddr. + // If rseqPreempted is true, before the next call to p.Switch(), + // interrupt rseq critical regions as defined by rseqAddr and + // tg.oldRSeqCritical and write the task goroutine's CPU number to + // rseqAddr/oldRSeqCPUAddr. // - // If rseqCPUAddr is 0, rseqCPU is -1. + // We support two ABIs for restartable sequences: // - // rseqCPUAddr, rseqCPU, and rseqPreempted are exclusive to the task - // goroutine. + // 1. The upstream interface added in v4.18, + // 2. An "old" interface never merged upstream. In the implementation, + // this is referred to as "old rseq". + // + // rseqPreempted is exclusive to the task goroutine. rseqPreempted bool `state:"nosave"` - rseqCPUAddr usermem.Addr - rseqCPU int32 + + // rseqCPU is the last CPU number written to rseqAddr/oldRSeqCPUAddr. + // + // If rseq is unused, rseqCPU is -1 for convenient use in + // platform.Context.Switch. + // + // rseqCPU is exclusive to the task goroutine. + rseqCPU int32 + + // oldRSeqCPUAddr is a pointer to the userspace old rseq CPU variable. + // + // oldRSeqCPUAddr is exclusive to the task goroutine. + oldRSeqCPUAddr usermem.Addr + + // rseqAddr is a pointer to the userspace linux.RSeq structure. + // + // rseqAddr is exclusive to the task goroutine. + rseqAddr usermem.Addr + + // rseqSignature is the signature that the rseq abort IP must be signed + // with. + // + // rseqSignature is exclusive to the task goroutine. + rseqSignature uint32 // copyScratchBuffer is a buffer available to CopyIn/CopyOut // implementations that require an intermediate buffer to copy data @@ -528,14 +562,6 @@ func (t *Task) loadPtraceTracer(tracer *Task) { t.ptraceTracer.Store(tracer) } -func (t *Task) saveLogPrefix() string { - return t.logPrefix.Load().(string) -} - -func (t *Task) loadLogPrefix(prefix string) { - t.logPrefix.Store(prefix) -} - func (t *Task) saveSyscallFilters() []bpf.Program { if f := t.syscallFilters.Load(); f != nil { return f.([]bpf.Program) @@ -549,6 +575,7 @@ func (t *Task) loadSyscallFilters(filters []bpf.Program) { // afterLoad is invoked by stateify. func (t *Task) afterLoad() { + t.updateInfoLocked() t.interruptChan = make(chan struct{}, 1) t.gosched.State = TaskGoroutineNonexistent if t.stop != nil { diff --git a/pkg/sentry/kernel/task_block.go b/pkg/sentry/kernel/task_block.go index dd69939f9..4a4a69ee2 100644 --- a/pkg/sentry/kernel/task_block.go +++ b/pkg/sentry/kernel/task_block.go @@ -16,6 +16,7 @@ package kernel import ( "runtime" + "runtime/trace" "time" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -133,19 +134,24 @@ func (t *Task) block(C <-chan struct{}, timerChan <-chan struct{}) error { runtime.Gosched() } + region := trace.StartRegion(t.traceContext, blockRegion) select { case <-C: + region.End() t.SleepFinish(true) + // Woken by event. return nil case <-interrupt: + region.End() t.SleepFinish(false) // Return the indicated error on interrupt. return syserror.ErrInterrupted case <-timerChan: - // We've timed out. + region.End() t.SleepFinish(true) + // We've timed out. return syserror.ETIMEDOUT } } diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 0916fd658..247bd4aba 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -236,14 +236,19 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else if opts.NewPIDNamespace { pidns = pidns.NewChild(userns) } + tg := t.tg + rseqAddr := usermem.Addr(0) + rseqSignature := uint32(0) if opts.NewThreadGroup { tg.mounts.IncRef() sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() } - tg = t.k.newThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy(), t.k.monotonicClock) + tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + rseqAddr = t.rseqAddr + rseqSignature = t.rseqSignature } cfg := &TaskConfig{ @@ -260,6 +265,8 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + RSeqAddr: rseqAddr, + RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), } if opts.NewThreadGroup { @@ -299,6 +306,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { // nt that it must receive before its task goroutine starts running. tid := nt.k.tasks.Root.IDOfTask(nt) defer nt.Start(tid) + t.traceCloneEvent(tid) // "If fork/clone and execve are allowed by @prog, any child processes will // be constrained to the same filters and system call ABI as the parent." - diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 17a089b90..fa6528386 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -129,6 +129,7 @@ type runSyscallAfterExecStop struct { } func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { + t.traceExecEvent(r.tc) t.tg.pidns.owner.mu.Lock() t.tg.execing = nil if t.killed() { @@ -189,9 +190,11 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.updateRSSLocked() // Restartable sequence state is discarded. t.rseqPreempted = false - t.rseqCPUAddr = 0 t.rseqCPU = -1 - t.tg.rscr.Store(&RSEQCriticalRegion{}) + t.rseqAddr = 0 + t.rseqSignature = 0 + t.oldRSeqCPUAddr = 0 + t.tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) t.tg.pidns.owner.mu.Unlock() // Remove FDs with the CloseOnExec flag set. @@ -253,7 +256,7 @@ func (t *Task) promoteLocked() { t.tg.leader = t t.Infof("Becoming TID %d (in root PID namespace)", t.tg.pidns.owner.Root.tids[t]) - t.updateLogPrefixLocked() + t.updateInfoLocked() // Reap the original leader. If it has a tracer, detach it instead of // waiting for it to acknowledge the original leader's death. oldLeader.exitParentNotified = true diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 535f03e50..435761e5a 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -236,6 +236,7 @@ func (*runExit) execute(t *Task) taskRunState { type runExitMain struct{} func (*runExitMain) execute(t *Task) taskRunState { + t.traceExitEvent() lastExiter := t.exitThreadGroup() // If the task has a cleartid, and the thread group wasn't killed by a diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index a29e9b9eb..0fb3661de 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -16,6 +16,7 @@ package kernel import ( "fmt" + "runtime/trace" "sort" "gvisor.dev/gvisor/pkg/log" @@ -127,11 +128,88 @@ func (t *Task) debugDumpStack() { } } -// updateLogPrefix updates the task's cached log prefix to reflect its -// current thread ID. +// trace definitions. +// +// Note that all region names are prefixed by ':' in order to ensure that they +// are lexically ordered before all system calls, which use the naked system +// call name (e.g. "read") for maximum clarity. +const ( + traceCategory = "task" + runRegion = ":run" + blockRegion = ":block" + cpuidRegion = ":cpuid" + faultRegion = ":fault" +) + +// updateInfoLocked updates the task's cached log prefix and tracing +// information to reflect its current thread ID. // // Preconditions: The task's owning TaskSet.mu must be locked. -func (t *Task) updateLogPrefixLocked() { +func (t *Task) updateInfoLocked() { // Use the task's TID in the root PID namespace for logging. - t.logPrefix.Store(fmt.Sprintf("[% 4d] ", t.tg.pidns.owner.Root.tids[t])) + tid := t.tg.pidns.owner.Root.tids[t] + t.logPrefix.Store(fmt.Sprintf("[% 4d] ", tid)) + t.rebuildTraceContext(tid) +} + +// rebuildTraceContext rebuilds the trace context. +// +// Precondition: the passed tid must be the tid in the root namespace. +func (t *Task) rebuildTraceContext(tid ThreadID) { + // Re-initialize the trace context. + if t.traceTask != nil { + t.traceTask.End() + } + + // Note that we define the "task type" to be the dynamic TID. This does + // not align perfectly with the documentation for "tasks" in the + // tracing package. Tasks may be assumed to be bounded by analysis + // tools. However, if we just use a generic "task" type here, then the + // "user-defined tasks" page on the tracing dashboard becomes nearly + // unusable, as it loads all traces from all tasks. + // + // We can assume that the number of tasks in the system is not + // arbitrarily large (in general it won't be, especially for cases + // where we're collecting a brief profile), so using the TID is a + // reasonable compromise in this case. + t.traceContext, t.traceTask = trace.NewTask(t, fmt.Sprintf("tid:%d", tid)) +} + +// traceCloneEvent is called when a new task is spawned. +// +// ntid must be the new task's ThreadID in the root namespace. +func (t *Task) traceCloneEvent(ntid ThreadID) { + if !trace.IsEnabled() { + return + } + trace.Logf(t.traceContext, traceCategory, "spawn: %d", ntid) +} + +// traceExitEvent is called when a task exits. +func (t *Task) traceExitEvent() { + if !trace.IsEnabled() { + return + } + trace.Logf(t.traceContext, traceCategory, "exit status: 0x%x", t.exitStatus.Status()) +} + +// traceExecEvent is called when a task calls exec. +func (t *Task) traceExecEvent(tc *TaskContext) { + if !trace.IsEnabled() { + return + } + d := tc.MemoryManager.Executable() + if d == nil { + trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") + return + } + defer d.DecRef() + root := t.fsContext.RootDirectory() + if root == nil { + trace.Logf(t.traceContext, traceCategory, "exec: << no root directory >>") + return + } + defer root.DecRef() + n, _ := d.FullName(root) + trace.Logf(t.traceContext, traceCategory, "exec: %s", n) } diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index c92266c59..6357273d3 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -17,6 +17,7 @@ package kernel import ( "bytes" "runtime" + "runtime/trace" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -168,12 +169,22 @@ func (*runApp) execute(t *Task) taskRunState { // Apply restartable sequences. if t.rseqPreempted { t.rseqPreempted = false - if t.rseqCPUAddr != 0 { + if t.rseqAddr != 0 || t.oldRSeqCPUAddr != 0 { + // Linux writes the CPU on every preemption. We only do + // so if it changed. Thus we may delay delivery of + // SIGSEGV if rseqAddr/oldRSeqCPUAddr is invalid. cpu := int32(hostcpu.GetCPU()) if t.rseqCPU != cpu { t.rseqCPU = cpu if err := t.rseqCopyOutCPU(); err != nil { - t.Warningf("Failed to copy CPU to %#x for RSEQ: %v", t.rseqCPUAddr, err) + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + // Re-enter the task run loop for signal delivery. + return (*runApp)(nil) + } + if err := t.oldRSeqCopyOutCPU(); err != nil { + t.Debugf("Failed to copy CPU to %#x for old rseq: %v", t.oldRSeqCPUAddr, err) t.forceSignal(linux.SIGSEGV, false) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) // Re-enter the task run loop for signal delivery. @@ -205,9 +216,11 @@ func (*runApp) execute(t *Task) taskRunState { t.tg.pidns.owner.mu.RUnlock() } + region := trace.StartRegion(t.traceContext, runRegion) t.accountTaskGoroutineEnter(TaskGoroutineRunningApp) info, at, err := t.p.Switch(t.MemoryManager().AddressSpace(), t.Arch(), t.rseqCPU) t.accountTaskGoroutineLeave(TaskGoroutineRunningApp) + region.End() if clearSinglestep { t.Arch().ClearSingleStep() @@ -225,6 +238,7 @@ func (*runApp) execute(t *Task) taskRunState { 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) @@ -232,10 +246,12 @@ func (*runApp) execute(t *Task) taskRunState { // Skip the cpuid instruction. t.Arch().CPUIDEmulate(t) t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() // 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. @@ -251,8 +267,10 @@ func (*runApp) execute(t *Task) taskRunState { // an application-generated signal and we should continue execution // normally. if at.Any() { + region := trace.StartRegion(t.traceContext, faultRegion) addr := usermem.Addr(info.Addr()) err := t.MemoryManager().HandleUserFault(t, addr, at, usermem.Addr(t.Arch().Stack())) + region.End() if err == nil { // The fault was handled appropriately. // We can resume running the application. @@ -260,6 +278,12 @@ func (*runApp) execute(t *Task) taskRunState { } // Is this a vsyscall that we need emulate? + // + // Note that we don't track vsyscalls as part of a + // specific trace region. This is because regions don't + // stack, and the actual system call will count as a + // region. We should be able to easily identify + // vsyscalls by having a <fault><syscall> pair. if at.Execute { if sysno, ok := t.tc.st.LookupEmulate(addr); ok { return t.doVsyscall(addr, sysno) @@ -306,7 +330,7 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextCPUPreempted: - // Ensure that RSEQ critical sections are interrupted and per-thread + // Ensure that rseq critical sections are interrupted and per-thread // CPU values are updated before the next platform.Context.Switch(). t.rseqPreempted = true return (*runApp)(nil) diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index ae6fc4025..58af16ee2 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" ) @@ -79,6 +80,13 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. + RSeqAddr usermem.Addr + + // RSeqSignature is the signature that the rseq abort IP must be signed + // with. + RSeqSignature uint32 + // ContainerID is the container the new task belongs to. ContainerID string } @@ -126,6 +134,8 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { ipcns: cfg.IPCNamespace, abstractSockets: cfg.AbstractSocketNamespace, rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, futexWaiter: futex.NewWaiter(), containerID: cfg.ContainerID, } @@ -154,10 +164,10 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { // Below this point, newTask is expected not to fail (there is no rollback // of assignTIDsLocked or any of the following). - // Logging on t's behalf will panic if t.logPrefix hasn't been initialized. - // This is the earliest point at which we can do so (since t now has thread - // IDs). - t.updateLogPrefixLocked() + // Logging on t's behalf will panic if t.logPrefix hasn't been + // initialized. This is the earliest point at which we can do so + // (since t now has thread IDs). + t.updateInfoLocked() if cfg.InheritParent != nil { t.parent = cfg.InheritParent.parent diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go index b543d536a..3180f5560 100644 --- a/pkg/sentry/kernel/task_syscall.go +++ b/pkg/sentry/kernel/task_syscall.go @@ -17,6 +17,7 @@ package kernel import ( "fmt" "os" + "runtime/trace" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -160,6 +161,10 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u ctrl = ctrlStopAndReinvokeSyscall } else { fn := s.Lookup(sysno) + var region *trace.Region // Only non-nil if tracing == true. + if trace.IsEnabled() { + region = trace.StartRegion(t.traceContext, s.LookupName(sysno)) + } if fn != nil { // Call our syscall implementation. rval, ctrl, err = fn(t, args) @@ -167,6 +172,9 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u // Use the missing function if not found. rval, err = t.SyscallTable().Missing(t, sysno, args) } + if region != nil { + region.End() + } } if bits.IsOn32(fe, ExternalAfterEnable) && (s.ExternalFilterAfter == nil || s.ExternalFilterAfter(t, sysno, args)) { diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 72568d296..c0197a563 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -238,8 +238,8 @@ type ThreadGroup struct { // execed is protected by the TaskSet mutex. execed bool - // rscr is the thread group's RSEQ critical region. - rscr atomic.Value `state:".(*RSEQCriticalRegion)"` + // oldRSeqCritical is the thread group's old rseq critical region. + oldRSeqCritical atomic.Value `state:".(*OldRSeqCriticalRegion)"` // mounts is the thread group's mount namespace. This does not really // correspond to a "mount namespace" in Linux, but is more like a @@ -256,35 +256,35 @@ type ThreadGroup struct { tty *TTY } -// newThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace ns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. -func (k *Kernel) newThreadGroup(mounts *fs.MountNamespace, ns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet, monotonicClock *timekeeperClock) *ThreadGroup { +func (k *Kernel) NewThreadGroup(mntns *fs.MountNamespace, pidns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet) *ThreadGroup { tg := &ThreadGroup{ threadGroupNode: threadGroupNode{ - pidns: ns, + pidns: pidns, }, signalHandlers: sh, terminationSignal: terminationSignal, ioUsage: &usage.IO{}, limits: limits, - mounts: mounts, + mounts: mntns, } tg.itimerRealTimer = ktime.NewTimer(k.monotonicClock, &itimerRealListener{tg: tg}) tg.timers = make(map[linux.TimerID]*IntervalTimer) - tg.rscr.Store(&RSEQCriticalRegion{}) + tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) return tg } -// saveRscr is invoked by stateify. -func (tg *ThreadGroup) saveRscr() *RSEQCriticalRegion { - return tg.rscr.Load().(*RSEQCriticalRegion) +// saveOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) saveOldRSeqCritical() *OldRSeqCriticalRegion { + return tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) } -// loadRscr is invoked by stateify. -func (tg *ThreadGroup) loadRscr(rscr *RSEQCriticalRegion) { - tg.rscr.Store(rscr) +// loadOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) loadOldRSeqCritical(r *OldRSeqCriticalRegion) { + tg.oldRSeqCritical.Store(r) } // SignalHandlers returns the signal handlers used by tg. diff --git a/pkg/sentry/kernel/tty.go b/pkg/sentry/kernel/tty.go index 34f84487a..048de26dc 100644 --- a/pkg/sentry/kernel/tty.go +++ b/pkg/sentry/kernel/tty.go @@ -21,8 +21,19 @@ import "sync" // // +stateify savable type TTY struct { + // Index is the terminal index. It is immutable. + Index uint32 + mu sync.Mutex `state:"nosave"` // tg is protected by mu. tg *ThreadGroup } + +// TTY returns the thread group's controlling terminal. If nil, there is no +// controlling terminal. +func (tg *ThreadGroup) TTY() *TTY { + tg.signalHandlers.mu.Lock() + defer tg.signalHandlers.mu.Unlock() + return tg.tty +} diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index c2c3ec06e..6299a3e2f 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -408,6 +408,8 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el start = vaddr } if vaddr < end { + // NOTE(b/37474556): Linux allows out-of-order + // segments, in violation of the spec. ctx.Infof("PT_LOAD headers out-of-order. %#x < %#x", vaddr, end) return loadedELF{}, syserror.ENOEXEC } diff --git a/pkg/sentry/memmap/BUILD b/pkg/sentry/memmap/BUILD index 3ef84245b..112794e9c 100644 --- a/pkg/sentry/memmap/BUILD +++ b/pkg/sentry/memmap/BUILD @@ -41,7 +41,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/log", - "//pkg/refs", "//pkg/sentry/context", "//pkg/sentry/platform", "//pkg/sentry/usermem", diff --git a/pkg/sentry/memmap/memmap.go b/pkg/sentry/memmap/memmap.go index 03b99aaea..16a722a13 100644 --- a/pkg/sentry/memmap/memmap.go +++ b/pkg/sentry/memmap/memmap.go @@ -18,7 +18,6 @@ package memmap import ( "fmt" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/usermem" @@ -235,8 +234,11 @@ type InvalidateOpts struct { // coincidental; fs.File implements MappingIdentity, and some // fs.InodeOperations implement Mappable.) type MappingIdentity interface { - // MappingIdentity is reference-counted. - refs.RefCounter + // IncRef increments the MappingIdentity's reference count. + IncRef() + + // DecRef decrements the MappingIdentity's reference count. + DecRef() // MappedName returns the application-visible name shown in // /proc/[pid]/maps. diff --git a/pkg/sentry/mm/procfs.go b/pkg/sentry/mm/procfs.go index 8c2246bb4..79610acb7 100644 --- a/pkg/sentry/mm/procfs.go +++ b/pkg/sentry/mm/procfs.go @@ -66,8 +66,6 @@ func (mm *MemoryManager) ReadMapsDataInto(ctx context.Context, buf *bytes.Buffer var start usermem.Addr for vseg := mm.vmas.LowerBoundSegment(start); vseg.Ok(); vseg = vseg.NextSegment() { - // FIXME(b/30793614): If we use a usermem.Addr for the handle, we get - // "panic: autosave error: type usermem.Addr is not registered". mm.appendVMAMapsEntryLocked(ctx, vseg, buf) } @@ -81,7 +79,6 @@ func (mm *MemoryManager) ReadMapsDataInto(ctx context.Context, buf *bytes.Buffer // // Artifically adjust the seqfile handle so we only output vsyscall entry once. if start != vsyscallEnd { - // FIXME(b/30793614): Can't get a pointer to constant vsyscallEnd. buf.WriteString(vsyscallMapsEntry) } } @@ -97,8 +94,6 @@ func (mm *MemoryManager) ReadMapsSeqFileData(ctx context.Context, handle seqfile start = *handle.(*usermem.Addr) } for vseg := mm.vmas.LowerBoundSegment(start); vseg.Ok(); vseg = vseg.NextSegment() { - // FIXME(b/30793614): If we use a usermem.Addr for the handle, we get - // "panic: autosave error: type usermem.Addr is not registered". vmaAddr := vseg.End() data = append(data, seqfile.SeqData{ Buf: mm.vmaMapsEntryLocked(ctx, vseg), @@ -116,7 +111,6 @@ func (mm *MemoryManager) ReadMapsSeqFileData(ctx context.Context, handle seqfile // // Artifically adjust the seqfile handle so we only output vsyscall entry once. if start != vsyscallEnd { - // FIXME(b/30793614): Can't get a pointer to constant vsyscallEnd. vmaAddr := vsyscallEnd data = append(data, seqfile.SeqData{ Buf: []byte(vsyscallMapsEntry), @@ -187,15 +181,12 @@ func (mm *MemoryManager) ReadSmapsDataInto(ctx context.Context, buf *bytes.Buffe var start usermem.Addr for vseg := mm.vmas.LowerBoundSegment(start); vseg.Ok(); vseg = vseg.NextSegment() { - // FIXME(b/30793614): If we use a usermem.Addr for the handle, we get - // "panic: autosave error: type usermem.Addr is not registered". mm.vmaSmapsEntryIntoLocked(ctx, vseg, buf) } // We always emulate vsyscall, so advertise it here. See // ReadMapsSeqFileData for additional commentary. if start != vsyscallEnd { - // FIXME(b/30793614): Can't get a pointer to constant vsyscallEnd. buf.WriteString(vsyscallSmapsEntry) } } @@ -211,8 +202,6 @@ func (mm *MemoryManager) ReadSmapsSeqFileData(ctx context.Context, handle seqfil start = *handle.(*usermem.Addr) } for vseg := mm.vmas.LowerBoundSegment(start); vseg.Ok(); vseg = vseg.NextSegment() { - // FIXME(b/30793614): If we use a usermem.Addr for the handle, we get - // "panic: autosave error: type usermem.Addr is not registered". vmaAddr := vseg.End() data = append(data, seqfile.SeqData{ Buf: mm.vmaSmapsEntryLocked(ctx, vseg), @@ -223,7 +212,6 @@ func (mm *MemoryManager) ReadSmapsSeqFileData(ctx context.Context, handle seqfil // We always emulate vsyscall, so advertise it here. See // ReadMapsSeqFileData for additional commentary. if start != vsyscallEnd { - // FIXME(b/30793614): Can't get a pointer to constant vsyscallEnd. vmaAddr := vsyscallEnd data = append(data, seqfile.SeqData{ Buf: []byte(vsyscallSmapsEntry), diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD index 6803d488c..f3afd98da 100644 --- a/pkg/sentry/platform/kvm/BUILD +++ b/pkg/sentry/platform/kvm/BUILD @@ -12,20 +12,30 @@ go_library( "bluepill_amd64.go", "bluepill_amd64.s", "bluepill_amd64_unsafe.go", + "bluepill_arm64.go", + "bluepill_arm64.s", + "bluepill_arm64_unsafe.go", "bluepill_fault.go", "bluepill_unsafe.go", "context.go", - "filters.go", + "filters_amd64.go", + "filters_arm64.go", "kvm.go", "kvm_amd64.go", "kvm_amd64_unsafe.go", + "kvm_arm64.go", + "kvm_arm64_unsafe.go", "kvm_const.go", + "kvm_const_arm64.go", "machine.go", "machine_amd64.go", "machine_amd64_unsafe.go", "machine_arm64.go", + "machine_arm64_unsafe.go", "machine_unsafe.go", "physical_map.go", + "physical_map_amd64.go", + "physical_map_arm64.go", "virtual_map.go", ], importpath = "gvisor.dev/gvisor/pkg/sentry/platform/kvm", diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go index 043de51b3..30dbb74d6 100644 --- a/pkg/sentry/platform/kvm/bluepill.go +++ b/pkg/sentry/platform/kvm/bluepill.go @@ -20,6 +20,7 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/safecopy" ) @@ -36,6 +37,18 @@ func sighandler() func dieTrampoline() var ( + // bounceSignal is the signal used for bouncing KVM. + // + // We use SIGCHLD because it is not masked by the runtime, and + // it will be ignored properly by other parts of the kernel. + bounceSignal = syscall.SIGCHLD + + // bounceSignalMask has only bounceSignal set. + bounceSignalMask = uint64(1 << (uint64(bounceSignal) - 1)) + + // bounce is the interrupt vector used to return to the kernel. + bounce = uint32(ring0.VirtualizationException) + // savedHandler is a pointer to the previous handler. // // This is called by bluepillHandler. @@ -45,6 +58,13 @@ var ( dieTrampolineAddr uintptr ) +// redpill invokes a syscall with -1. +// +//go:nosplit +func redpill() { + syscall.RawSyscall(^uintptr(0), 0, 0, 0) +} + // dieHandler is called by dieTrampoline. // //go:nosplit @@ -73,8 +93,8 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) { func init() { // Install the handler. - if err := safecopy.ReplaceSignalHandler(syscall.SIGSEGV, reflect.ValueOf(sighandler).Pointer(), &savedHandler); err != nil { - panic(fmt.Sprintf("Unable to set handler for signal %d: %v", syscall.SIGSEGV, err)) + if err := safecopy.ReplaceSignalHandler(bluepillSignal, reflect.ValueOf(sighandler).Pointer(), &savedHandler); err != nil { + panic(fmt.Sprintf("Unable to set handler for signal %d: %v", bluepillSignal, err)) } // Extract the address for the trampoline. diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.go b/pkg/sentry/platform/kvm/bluepill_amd64.go index 421c88220..133c2203d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64.go @@ -24,26 +24,10 @@ import ( ) var ( - // bounceSignal is the signal used for bouncing KVM. - // - // We use SIGCHLD because it is not masked by the runtime, and - // it will be ignored properly by other parts of the kernel. - bounceSignal = syscall.SIGCHLD - - // bounceSignalMask has only bounceSignal set. - bounceSignalMask = uint64(1 << (uint64(bounceSignal) - 1)) - - // bounce is the interrupt vector used to return to the kernel. - bounce = uint32(ring0.VirtualizationException) + // The action for bluepillSignal is changed by sigaction(). + bluepillSignal = syscall.SIGSEGV ) -// redpill on amd64 invokes a syscall with -1. -// -//go:nosplit -func redpill() { - syscall.RawSyscall(^uintptr(0), 0, 0, 0) -} - // bluepillArchEnter is called during bluepillEnter. // //go:nosplit diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index 9d8af143e..a63a6a071 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -23,13 +23,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) -// bluepillArchContext returns the arch-specific context. -// -//go:nosplit -func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 { - return &((*arch.UContext64)(context).MContext) -} - // dieArchSetup initializes the state for dieTrampoline. // // The amd64 dieTrampoline requires the vCPU to be set in BX, and the last RIP diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.go b/pkg/sentry/platform/kvm/bluepill_arm64.go new file mode 100644 index 000000000..552341721 --- /dev/null +++ b/pkg/sentry/platform/kvm/bluepill_arm64.go @@ -0,0 +1,79 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package kvm + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" +) + +var ( + // The action for bluepillSignal is changed by sigaction(). + bluepillSignal = syscall.SIGILL +) + +// bluepillArchEnter is called during bluepillEnter. +// +//go:nosplit +func bluepillArchEnter(context *arch.SignalContext64) (c *vCPU) { + c = vCPUPtr(uintptr(context.Regs[8])) + regs := c.CPU.Registers() + regs.Regs = context.Regs + regs.Sp = context.Sp + regs.Pc = context.Pc + regs.Pstate = context.Pstate + regs.Pstate &^= uint64(ring0.KernelFlagsClear) + regs.Pstate |= ring0.KernelFlagsSet + return +} + +// bluepillArchExit is called during bluepillEnter. +// +//go:nosplit +func bluepillArchExit(c *vCPU, context *arch.SignalContext64) { + regs := c.CPU.Registers() + context.Regs = regs.Regs + context.Sp = regs.Sp + context.Pc = regs.Pc + context.Pstate = regs.Pstate + context.Pstate &^= uint64(ring0.UserFlagsClear) + context.Pstate |= ring0.UserFlagsSet +} + +// KernelSyscall handles kernel syscalls. +// +//go:nosplit +func (c *vCPU) KernelSyscall() { + regs := c.Registers() + if regs.Regs[8] != ^uint64(0) { + regs.Pc -= 4 // Rewind. + } + ring0.Halt() +} + +// KernelException handles kernel exceptions. +// +//go:nosplit +func (c *vCPU) KernelException(vector ring0.Vector) { + regs := c.Registers() + if vector == ring0.Vector(bounce) { + regs.Pc = 0 + } + ring0.Halt() +} diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s new file mode 100644 index 000000000..c61700892 --- /dev/null +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -0,0 +1,87 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "textflag.h" + +// VCPU_CPU is the location of the CPU in the vCPU struct. +// +// This is guaranteed to be zero. +#define VCPU_CPU 0x0 + +// CPU_SELF is the self reference in ring0's percpu. +// +// This is guaranteed to be zero. +#define CPU_SELF 0x0 + +// Context offsets. +// +// Only limited use of the context is done in the assembly stub below, most is +// done in the Go handlers. +#define SIGINFO_SIGNO 0x0 +#define CONTEXT_PC 0x1B8 +#define CONTEXT_R0 0xB8 + +// See bluepill.go. +TEXT ·bluepill(SB),NOSPLIT,$0 +begin: + MOVD vcpu+0(FP), R8 + MOVD $VCPU_CPU(R8), R9 + ORR $0xffff000000000000, R9, R9 + // Trigger sigill. + // In ring0.Start(), the value of R8 will be stored into tpidr_el1. + // When the context was loaded into vcpu successfully, + // we will check if the value of R10 and R9 are the same. + WORD $0xd538d08a // MRS TPIDR_EL1, R10 +check_vcpu: + CMP R10, R9 + BEQ right_vCPU +wrong_vcpu: + CALL ·redpill(SB) + B begin +right_vCPU: + RET + +// sighandler: see bluepill.go for documentation. +// +// The arguments are the following: +// +// R0 - The signal number. +// R1 - Pointer to siginfo_t structure. +// R2 - Pointer to ucontext structure. +// +TEXT ·sighandler(SB),NOSPLIT,$0 + // si_signo should be sigill. + MOVD SIGINFO_SIGNO(R1), R7 + CMPW $4, R7 + BNE fallback + + MOVD CONTEXT_PC(R2), R7 + CMPW $0, R7 + BEQ fallback + + MOVD R2, 8(RSP) + BL ·bluepillHandler(SB) // Call the handler. + + RET + +fallback: + // Jump to the previous signal handler. + MOVD ·savedHandler(SB), R7 + B (R7) + +// 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) diff --git a/pkg/sentry/fsimpl/proc/filesystems.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go index c36c4aff5..e5fac0d6a 100644 --- a/pkg/sentry/fsimpl/proc/filesystems.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -12,14 +12,17 @@ // See the License for the specific language governing permissions and // limitations under the License. -package proc +// +build arm64 -// filesystemsData implements vfs.DynamicBytesSource for /proc/filesystems. -// -// +stateify savable -type filesystemsData struct{} +package kvm + +import ( + "unsafe" + + "gvisor.dev/gvisor/pkg/sentry/arch" +) -// TODO(b/138862512): Implement vfs.DynamicBytesSource.Generate for -// filesystemsData. We would need to retrive filesystem names from -// vfs.VirtualFilesystem. Also needs vfs replacement for -// fs.Filesystem.AllowUserList() and fs.FilesystemRequiresDev. +//go:nosplit +func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) { + // TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64. +} diff --git a/pkg/sentry/platform/kvm/bluepill_unsafe.go b/pkg/sentry/platform/kvm/bluepill_unsafe.go index ca011ef78..9add7c944 100644 --- a/pkg/sentry/platform/kvm/bluepill_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_unsafe.go @@ -23,6 +23,8 @@ import ( "sync/atomic" "syscall" "unsafe" + + "gvisor.dev/gvisor/pkg/sentry/arch" ) //go:linkname throw runtime.throw @@ -49,6 +51,13 @@ func uintptrValue(addr *byte) uintptr { return (uintptr)(unsafe.Pointer(addr)) } +// bluepillArchContext returns the UContext64. +// +//go:nosplit +func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 { + return &((*arch.UContext64)(context).MContext) +} + // bluepillHandler is called from the signal stub. // // The world may be stopped while this is executing, and it executes on the diff --git a/pkg/sentry/platform/kvm/filters.go b/pkg/sentry/platform/kvm/filters_amd64.go index 7d949f1dd..7d949f1dd 100644 --- a/pkg/sentry/platform/kvm/filters.go +++ b/pkg/sentry/platform/kvm/filters_amd64.go diff --git a/pkg/sentry/platform/kvm/filters_arm64.go b/pkg/sentry/platform/kvm/filters_arm64.go new file mode 100644 index 000000000..9245d07c2 --- /dev/null +++ b/pkg/sentry/platform/kvm/filters_arm64.go @@ -0,0 +1,32 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kvm + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/seccomp" +) + +// SyscallFilters returns syscalls made exclusively by the KVM platform. +func (*KVM) SyscallFilters() seccomp.SyscallRules { + return seccomp.SyscallRules{ + syscall.SYS_IOCTL: {}, + syscall.SYS_MMAP: {}, + syscall.SYS_RT_SIGSUSPEND: {}, + syscall.SYS_RT_SIGTIMEDWAIT: {}, + 0xffffffffffffffff: {}, // KVM uses syscall -1 to transition to host. + } +} diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index ee4cd2f4d..f2c2c059e 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -21,7 +21,6 @@ import ( "sync" "syscall" - "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" @@ -56,9 +55,7 @@ func New(deviceFile *os.File) (*KVM, error) { // Ensure global initialization is done. globalOnce.Do(func() { - physicalInit() - globalErr = updateSystemValues(int(fd)) - ring0.Init(cpuid.HostFeatureSet()) + globalErr = updateGlobalOnce(int(fd)) }) if globalErr != nil { return nil, globalErr diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go index 5d8ef4761..c5a6f9c7d 100644 --- a/pkg/sentry/platform/kvm/kvm_amd64.go +++ b/pkg/sentry/platform/kvm/kvm_amd64.go @@ -17,6 +17,7 @@ package kvm import ( + "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/sentry/platform/ring0" ) @@ -211,3 +212,11 @@ type cpuidEntries struct { _ uint32 entries [_KVM_NR_CPUID_ENTRIES]cpuidEntry } + +// updateGlobalOnce does global initialization. It has to be called only once. +func updateGlobalOnce(fd int) error { + physicalInit() + err := updateSystemValues(int(fd)) + ring0.Init(cpuid.HostFeatureSet()) + return err +} diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go new file mode 100644 index 000000000..2319c86d3 --- /dev/null +++ b/pkg/sentry/platform/kvm/kvm_arm64.go @@ -0,0 +1,83 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package kvm + +import ( + "syscall" +) + +// userMemoryRegion is a region of physical memory. +// +// This mirrors kvm_memory_region. +type userMemoryRegion struct { + slot uint32 + flags uint32 + guestPhysAddr uint64 + memorySize uint64 + userspaceAddr uint64 +} + +type kvmOneReg struct { + id uint64 + addr uint64 +} + +const KVM_NR_SPSR = 5 + +type userFpsimdState struct { + vregs [64]uint64 + fpsr uint32 + fpcr uint32 + reserved [2]uint32 +} + +type userRegs struct { + Regs syscall.PtraceRegs + sp_el1 uint64 + elr_el1 uint64 + spsr [KVM_NR_SPSR]uint64 + fpRegs userFpsimdState +} + +// runData is the run structure. This may be mapped for synchronous register +// access (although that doesn't appear to be supported by my kernel at least). +// +// This mirrors kvm_run. +type runData struct { + requestInterruptWindow uint8 + _ [7]uint8 + + exitReason uint32 + readyForInterruptInjection uint8 + ifFlag uint8 + _ [2]uint8 + + cr8 uint64 + apicBase uint64 + + // This is the union data for exits. Interpretation depends entirely on + // the exitReason above (see vCPU code for more information). + data [32]uint64 +} + +// updateGlobalOnce does global initialization. It has to be called only once. +func updateGlobalOnce(fd int) error { + physicalInit() + err := updateSystemValues(int(fd)) + updateVectorTable() + return err +} diff --git a/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go b/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go new file mode 100644 index 000000000..6531bae1d --- /dev/null +++ b/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go @@ -0,0 +1,39 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package kvm + +import ( + "fmt" + "syscall" +) + +var ( + runDataSize int +) + +func updateSystemValues(fd int) error { + // Extract the mmap size. + sz, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(fd), _KVM_GET_VCPU_MMAP_SIZE, 0) + if errno != 0 { + return fmt.Errorf("getting VCPU mmap size: %v", errno) + } + // Save the data. + runDataSize = int(sz) + + // Success. + return nil +} diff --git a/pkg/sentry/platform/kvm/kvm_const.go b/pkg/sentry/platform/kvm/kvm_const.go index 766131d60..1d5c77ff4 100644 --- a/pkg/sentry/platform/kvm/kvm_const.go +++ b/pkg/sentry/platform/kvm/kvm_const.go @@ -49,11 +49,13 @@ const ( _KVM_EXIT_SHUTDOWN = 0x8 _KVM_EXIT_FAIL_ENTRY = 0x9 _KVM_EXIT_INTERNAL_ERROR = 0x11 + _KVM_EXIT_SYSTEM_EVENT = 0x18 ) // KVM capability options. const ( - _KVM_CAP_MAX_VCPUS = 0x42 + _KVM_CAP_MAX_VCPUS = 0x42 + _KVM_CAP_ARM_VM_IPA_SIZE = 0xa5 ) // KVM limits. diff --git a/pkg/sentry/platform/kvm/kvm_const_arm64.go b/pkg/sentry/platform/kvm/kvm_const_arm64.go new file mode 100644 index 000000000..5a74c6e36 --- /dev/null +++ b/pkg/sentry/platform/kvm/kvm_const_arm64.go @@ -0,0 +1,132 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kvm + +// KVM ioctls for Arm64. +const ( + _KVM_GET_ONE_REG = 0x4010aeab + _KVM_SET_ONE_REG = 0x4010aeac + + _KVM_ARM_PREFERRED_TARGET = 0x8020aeaf + _KVM_ARM_VCPU_INIT = 0x4020aeae + _KVM_ARM64_REGS_PSTATE = 0x6030000000100042 + _KVM_ARM64_REGS_SP_EL1 = 0x6030000000100044 + _KVM_ARM64_REGS_R0 = 0x6030000000100000 + _KVM_ARM64_REGS_R1 = 0x6030000000100002 + _KVM_ARM64_REGS_R2 = 0x6030000000100004 + _KVM_ARM64_REGS_R3 = 0x6030000000100006 + _KVM_ARM64_REGS_R8 = 0x6030000000100010 + _KVM_ARM64_REGS_R18 = 0x6030000000100024 + _KVM_ARM64_REGS_PC = 0x6030000000100040 + _KVM_ARM64_REGS_MAIR_EL1 = 0x603000000013c510 + _KVM_ARM64_REGS_TCR_EL1 = 0x603000000013c102 + _KVM_ARM64_REGS_TTBR0_EL1 = 0x603000000013c100 + _KVM_ARM64_REGS_TTBR1_EL1 = 0x603000000013c101 + _KVM_ARM64_REGS_SCTLR_EL1 = 0x603000000013c080 + _KVM_ARM64_REGS_CPACR_EL1 = 0x603000000013c082 + _KVM_ARM64_REGS_VBAR_EL1 = 0x603000000013c600 +) + +// Arm64: Architectural Feature Access Control Register EL1. +const ( + _FPEN_NOTRAP = 0x3 + _FPEN_SHIFT = 0x20 +) + +// Arm64: System Control Register EL1. +const ( + _SCTLR_M = 1 << 0 + _SCTLR_C = 1 << 2 + _SCTLR_I = 1 << 12 +) + +// Arm64: Translation Control Register EL1. +const ( + _TCR_IPS_40BITS = 2 << 32 // PA=40 + _TCR_IPS_48BITS = 5 << 32 // PA=48 + + _TCR_T0SZ_OFFSET = 0 + _TCR_T1SZ_OFFSET = 16 + _TCR_IRGN0_SHIFT = 8 + _TCR_IRGN1_SHIFT = 24 + _TCR_ORGN0_SHIFT = 10 + _TCR_ORGN1_SHIFT = 26 + _TCR_SH0_SHIFT = 12 + _TCR_SH1_SHIFT = 28 + _TCR_TG0_SHIFT = 14 + _TCR_TG1_SHIFT = 30 + + _TCR_T0SZ_VA48 = 64 - 48 // VA=48 + _TCR_T1SZ_VA48 = 64 - 48 // VA=48 + + _TCR_ASID16 = 1 << 36 + _TCR_TBI0 = 1 << 37 + + _TCR_TXSZ_VA48 = (_TCR_T0SZ_VA48 << _TCR_T0SZ_OFFSET) | (_TCR_T1SZ_VA48 << _TCR_T1SZ_OFFSET) + + _TCR_TG0_4K = 0 << _TCR_TG0_SHIFT // 4K + _TCR_TG0_64K = 1 << _TCR_TG0_SHIFT // 64K + + _TCR_TG1_4K = 2 << _TCR_TG1_SHIFT + + _TCR_TG_FLAGS = _TCR_TG0_4K | _TCR_TG1_4K + + _TCR_IRGN0_WBWA = 1 << _TCR_IRGN0_SHIFT + _TCR_IRGN1_WBWA = 1 << _TCR_IRGN1_SHIFT + _TCR_IRGN_WBWA = _TCR_IRGN0_WBWA | _TCR_IRGN1_WBWA + + _TCR_ORGN0_WBWA = 1 << _TCR_ORGN0_SHIFT + _TCR_ORGN1_WBWA = 1 << _TCR_ORGN1_SHIFT + + _TCR_ORGN_WBWA = _TCR_ORGN0_WBWA | _TCR_ORGN1_WBWA + + _TCR_SHARED = (3 << _TCR_SH0_SHIFT) | (3 << _TCR_SH1_SHIFT) + + _TCR_CACHE_FLAGS = _TCR_IRGN_WBWA | _TCR_ORGN_WBWA +) + +// Arm64: Memory Attribute Indirection Register EL1. +const ( + _MT_DEVICE_nGnRnE = 0 + _MT_DEVICE_nGnRE = 1 + _MT_DEVICE_GRE = 2 + _MT_NORMAL_NC = 3 + _MT_NORMAL = 4 + _MT_NORMAL_WT = 5 + _MT_EL1_INIT = (0 << _MT_DEVICE_nGnRnE) | (0x4 << _MT_DEVICE_nGnRE * 8) | (0xc << _MT_DEVICE_GRE * 8) | (0x44 << _MT_NORMAL_NC * 8) | (0xff << _MT_NORMAL * 8) | (0xbb << _MT_NORMAL_WT * 8) +) + +const ( + _KVM_ARM_VCPU_POWER_OFF = 0 // CPU is started in OFF state + _KVM_ARM_VCPU_PSCI_0_2 = 2 // CPU uses PSCI v0.2 +) + +// Arm64: Exception Syndrome Register EL1. +const ( + _ESR_ELx_FSC = 0x3F + + _ESR_SEGV_MAPERR_L0 = 0x4 + _ESR_SEGV_MAPERR_L1 = 0x5 + _ESR_SEGV_MAPERR_L2 = 0x6 + _ESR_SEGV_MAPERR_L3 = 0x7 + + _ESR_SEGV_ACCERR_L1 = 0x9 + _ESR_SEGV_ACCERR_L2 = 0xa + _ESR_SEGV_ACCERR_L3 = 0xb + + _ESR_SEGV_PEMERR_L1 = 0xd + _ESR_SEGV_PEMERR_L2 = 0xe + _ESR_SEGV_PEMERR_L3 = 0xf +) diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go index 61227cafb..7156c245f 100644 --- a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go @@ -135,3 +135,43 @@ func (c *vCPU) setSignalMask() error { } return nil } + +// setUserRegisters sets user registers in the vCPU. +func (c *vCPU) setUserRegisters(uregs *userRegs) error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_SET_REGS, + uintptr(unsafe.Pointer(uregs))); errno != 0 { + return fmt.Errorf("error setting user registers: %v", errno) + } + return nil +} + +// getUserRegisters reloads user registers in the vCPU. +// +// This is safe to call from a nosplit context. +// +//go:nosplit +func (c *vCPU) getUserRegisters(uregs *userRegs) syscall.Errno { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_GET_REGS, + uintptr(unsafe.Pointer(uregs))); errno != 0 { + return errno + } + return 0 +} + +// setSystemRegisters sets system registers. +func (c *vCPU) setSystemRegisters(sregs *systemRegs) error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_SET_SREGS, + uintptr(unsafe.Pointer(sregs))); errno != 0 { + return fmt.Errorf("error setting system registers: %v", errno) + } + return nil +} diff --git a/pkg/sentry/platform/kvm/machine_arm64.go b/pkg/sentry/platform/kvm/machine_arm64.go index b7e2cfb9d..7ae47f291 100644 --- a/pkg/sentry/platform/kvm/machine_arm64.go +++ b/pkg/sentry/platform/kvm/machine_arm64.go @@ -12,8 +12,38 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package kvm +import ( + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" + "gvisor.dev/gvisor/pkg/sentry/usermem" +) + +type vCPUArchState struct { + // PCIDs is the set of PCIDs for this vCPU. + // + // This starts above fixedKernelPCID. + PCIDs *pagetables.PCIDs +} + +const ( + // fixedKernelPCID is a fixed kernel PCID used for the kernel page + // tables. We must start allocating user PCIDs above this in order to + // avoid any conflict (see below). + fixedKernelPCID = 1 + + // poolPCIDs is the number of PCIDs to record in the database. As this + // grows, assignment can take longer, since it is a simple linear scan. + // Beyond a relatively small number, there are likely few perform + // benefits, since the TLB has likely long since lost any translations + // from more than a few PCIDs past. + poolPCIDs = 8 +) + // Get all read-only physicalRegions. func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { var rdonlyRegions []region @@ -59,3 +89,95 @@ func availableRegionsForSetMem() (phyRegions []physicalRegion) { return phyRegions } + +// dropPageTables drops cached page table entries. +func (m *machine) dropPageTables(pt *pagetables.PageTables) { + m.mu.Lock() + defer m.mu.Unlock() + + // Clear from all PCIDs. + for _, c := range m.vCPUs { + c.PCIDs.Drop(pt) + } +} + +// nonCanonical generates a canonical address return. +// +//go:nosplit +func nonCanonical(addr uint64, signal int32, info *arch.SignalInfo) (usermem.AccessType, error) { + *info = arch.SignalInfo{ + Signo: signal, + Code: arch.SignalInfoKernel, + } + info.SetAddr(addr) // Include address. + return usermem.NoAccess, platform.ErrContextSignal +} + +// fault generates an appropriate fault return. +// +//go:nosplit +func (c *vCPU) fault(signal int32, info *arch.SignalInfo) (usermem.AccessType, error) { + faultAddr := c.GetFaultAddr() + code, user := c.ErrorCode() + + // Reset the pointed SignalInfo. + *info = arch.SignalInfo{Signo: signal} + info.SetAddr(uint64(faultAddr)) + + read := true + write := false + execute := true + + ret := code & _ESR_ELx_FSC + switch ret { + case _ESR_SEGV_MAPERR_L0, _ESR_SEGV_MAPERR_L1, _ESR_SEGV_MAPERR_L2, _ESR_SEGV_MAPERR_L3: + info.Code = 1 //SEGV_MAPERR + read = false + write = true + execute = false + case _ESR_SEGV_ACCERR_L1, _ESR_SEGV_ACCERR_L2, _ESR_SEGV_ACCERR_L3, _ESR_SEGV_PEMERR_L1, _ESR_SEGV_PEMERR_L2, _ESR_SEGV_PEMERR_L3: + info.Code = 2 // SEGV_ACCERR. + read = true + write = false + execute = false + default: + info.Code = 2 + } + + if !user { + read = true + write = false + execute = true + + } + accessType := usermem.AccessType{ + Read: read, + Write: write, + Execute: execute, + } + + return accessType, platform.ErrContextSignal +} + +// retryInGuest runs the given function in guest mode. +// +// If the function does not complete in guest mode (due to execution of a +// system call due to a GC stall, for example), then it will be retried. The +// given function must be idempotent as a result of the retry mechanism. +func (m *machine) retryInGuest(fn func()) { + c := m.Get() + defer m.Put(c) + for { + c.ClearErrorCode() // See below. + bluepill(c) // Force guest mode. + fn() // Execute the given function. + _, user := c.ErrorCode() + if user { + // If user is set, then we haven't bailed back to host + // mode via a kernel exception or system call. We + // consider the full function to have executed in guest + // mode and we can return. + break + } + } +} diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go new file mode 100644 index 000000000..3f2f97a6b --- /dev/null +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -0,0 +1,362 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package kvm + +import ( + "fmt" + "reflect" + "sync/atomic" + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" + "gvisor.dev/gvisor/pkg/sentry/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 +} + +var vcpuInit kvmVcpuInit + +// initArchState initializes architecture-specific state. +func (m *machine) initArchState() error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(m.fd), + _KVM_ARM_PREFERRED_TARGET, + uintptr(unsafe.Pointer(&vcpuInit))); errno != 0 { + panic(fmt.Sprintf("error setting KVM_ARM_PREFERRED_TARGET failed: %v", errno)) + } + return nil +} + +func getPageWithReflect(p uintptr) []byte { + return (*(*[0xFFFFFF]byte)(unsafe.Pointer(p & ^uintptr(syscall.Getpagesize()-1))))[:syscall.Getpagesize()] +} + +// Work around: move ring0.Vectors() into a specific address with 11-bits alignment. +// +// According to the design documentation of Arm64, +// the start address of exception vector table should be 11-bits aligned. +// Please see the code in linux kernel as reference: arch/arm64/kernel/entry.S +// But, we can't align a function's start address to a specific address by using golang. +// We have raised this question in golang community: +// https://groups.google.com/forum/m/#!topic/golang-dev/RPj90l5x86I +// This function will be removed when golang supports this feature. +// +// There are 2 jobs were implemented in this function: +// 1, move the start address of exception vector table into the specific address. +// 2, modify the offset of each instruction. +func updateVectorTable() { + fromLocation := reflect.ValueOf(ring0.Vectors).Pointer() + offset := fromLocation & (1<<11 - 1) + if offset != 0 { + offset = 1<<11 - offset + } + + toLocation := fromLocation + offset + page := getPageWithReflect(toLocation) + if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { + panic(err) + } + + page = getPageWithReflect(toLocation + 4096) + if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_WRITE|syscall.PROT_EXEC); err != nil { + panic(err) + } + + // Move exception-vector-table into the specific address. + var entry *uint32 + var entryFrom *uint32 + for i := 1; i <= 0x800; i++ { + entry = (*uint32)(unsafe.Pointer(toLocation + 0x800 - uintptr(i))) + entryFrom = (*uint32)(unsafe.Pointer(fromLocation + 0x800 - uintptr(i))) + *entry = *entryFrom + } + + // The offset from the address of each unconditionally branch is changed. + // We should modify the offset of each instruction. + nums := []uint32{0x0, 0x80, 0x100, 0x180, 0x200, 0x280, 0x300, 0x380, 0x400, 0x480, 0x500, 0x580, 0x600, 0x680, 0x700, 0x780} + for _, num := range nums { + entry = (*uint32)(unsafe.Pointer(toLocation + uintptr(num))) + *entry = *entry - (uint32)(offset/4) + } + + page = getPageWithReflect(toLocation) + if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { + panic(err) + } + + page = getPageWithReflect(toLocation + 4096) + if err := syscall.Mprotect(page, syscall.PROT_READ|syscall.PROT_EXEC); err != nil { + panic(err) + } +} + +// initArchState initializes architecture-specific state. +func (c *vCPU) initArchState() error { + var ( + reg kvmOneReg + data uint64 + regGet kvmOneReg + dataGet uint64 + ) + + reg.addr = uint64(reflect.ValueOf(&data).Pointer()) + regGet.addr = uint64(reflect.ValueOf(&dataGet).Pointer()) + + vcpuInit.features[0] |= (1 << _KVM_ARM_VCPU_PSCI_0_2) + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_ARM_VCPU_INIT, + uintptr(unsafe.Pointer(&vcpuInit))); errno != 0 { + panic(fmt.Sprintf("error setting KVM_ARM_VCPU_INIT failed: %v", errno)) + } + + // cpacr_el1 + reg.id = _KVM_ARM64_REGS_CPACR_EL1 + data = (_FPEN_NOTRAP << _FPEN_SHIFT) + if err := c.setOneRegister(®); err != nil { + return err + } + + // sctlr_el1 + regGet.id = _KVM_ARM64_REGS_SCTLR_EL1 + if err := c.getOneRegister(®Get); err != nil { + return err + } + + dataGet |= (_SCTLR_M | _SCTLR_C | _SCTLR_I) + data = dataGet + reg.id = _KVM_ARM64_REGS_SCTLR_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + // tcr_el1 + data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS + reg.id = _KVM_ARM64_REGS_TCR_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + // mair_el1 + data = _MT_EL1_INIT + reg.id = _KVM_ARM64_REGS_MAIR_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + // ttbr0_el1 + data = c.machine.kernel.PageTables.TTBR0_EL1(false, 0) + + reg.id = _KVM_ARM64_REGS_TTBR0_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + c.SetTtbr0Kvm(uintptr(data)) + + // ttbr1_el1 + data = c.machine.kernel.PageTables.TTBR1_EL1(false, 0) + + reg.id = _KVM_ARM64_REGS_TTBR1_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + // sp_el1 + data = c.CPU.StackTop() + reg.id = _KVM_ARM64_REGS_SP_EL1 + if err := c.setOneRegister(®); err != nil { + return err + } + + // pc + reg.id = _KVM_ARM64_REGS_PC + data = uint64(reflect.ValueOf(ring0.Start).Pointer()) + if err := c.setOneRegister(®); err != nil { + return err + } + + // r8 + reg.id = _KVM_ARM64_REGS_R8 + data = uint64(reflect.ValueOf(&c.CPU).Pointer()) + if err := c.setOneRegister(®); err != nil { + return err + } + + // vbar_el1 + reg.id = _KVM_ARM64_REGS_VBAR_EL1 + + fromLocation := reflect.ValueOf(ring0.Vectors).Pointer() + offset := fromLocation & (1<<11 - 1) + if offset != 0 { + offset = 1<<11 - offset + } + + toLocation := fromLocation + offset + data = uint64(ring0.KernelStartAddress | toLocation) + if err := c.setOneRegister(®); err != nil { + return err + } + + data = ring0.PsrDefaultSet | ring0.KernelFlagsSet + reg.id = _KVM_ARM64_REGS_PSTATE + if err := c.setOneRegister(®); err != nil { + return err + } + + return nil +} + +//go:nosplit +func (c *vCPU) loadSegments(tid uint64) { + // TODO(gvisor.dev/issue/1238): TLS is not supported. + // Get TLS from tpidr_el0. + atomic.StoreUint64(&c.tid, tid) +} + +func (c *vCPU) setOneRegister(reg *kvmOneReg) error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_SET_ONE_REG, + uintptr(unsafe.Pointer(reg))); errno != 0 { + return fmt.Errorf("error setting one register: %v", errno) + } + return nil +} + +func (c *vCPU) getOneRegister(reg *kvmOneReg) error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_GET_ONE_REG, + uintptr(unsafe.Pointer(reg))); errno != 0 { + return fmt.Errorf("error setting one register: %v", errno) + } + return nil +} + +// setCPUID sets the CPUID to be used by the guest. +func (c *vCPU) setCPUID() error { + return nil +} + +// setSystemTime sets the TSC for the vCPU. +func (c *vCPU) setSystemTime() error { + return nil +} + +// setSignalMask sets the vCPU signal mask. +// +// This must be called prior to running the vCPU. +func (c *vCPU) setSignalMask() error { + // The layout of this structure implies that it will not necessarily be + // the same layout chosen by the Go compiler. It gets fudged here. + var data struct { + length uint32 + mask1 uint32 + mask2 uint32 + _ uint32 + } + data.length = 8 // Fixed sigset size. + data.mask1 = ^uint32(bounceSignalMask & 0xffffffff) + data.mask2 = ^uint32(bounceSignalMask >> 32) + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_SET_SIGNAL_MASK, + uintptr(unsafe.Pointer(&data))); errno != 0 { + return fmt.Errorf("error setting signal mask: %v", errno) + } + + return nil +} + +// SwitchToUser unpacks architectural-details. +func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo) (usermem.AccessType, error) { + // Check for canonical addresses. + if regs := switchOpts.Registers; !ring0.IsCanonical(regs.Pc) { + return nonCanonical(regs.Pc, int32(syscall.SIGSEGV), info) + } else if !ring0.IsCanonical(regs.Sp) { + return nonCanonical(regs.Sp, int32(syscall.SIGBUS), info) + } + + var vector ring0.Vector + ttbr0App := switchOpts.PageTables.TTBR0_EL1(false, 0) + c.SetTtbr0App(uintptr(ttbr0App)) + + // TODO(gvisor.dev/issue/1238): full context-switch supporting for Arm64. + // The Arm64 user-mode execution state consists of: + // x0-x30 + // PC, SP, PSTATE + // V0-V31: 32 128-bit registers for floating point, and simd + // FPSR + // TPIDR_EL0, used for TLS + appRegs := switchOpts.Registers + c.SetAppAddr(ring0.KernelStartAddress | uintptr(unsafe.Pointer(appRegs))) + + entersyscall() + bluepill(c) + vector = c.CPU.SwitchToUser(switchOpts) + exitsyscall() + + switch vector { + case ring0.Syscall: + // Fast path: system call executed. + return usermem.NoAccess, nil + + case ring0.PageFault: + return c.fault(int32(syscall.SIGSEGV), info) + case 0xaa: + return usermem.NoAccess, nil + default: + return usermem.NoAccess, platform.ErrContextSignal + } + +} diff --git a/pkg/sentry/platform/kvm/machine_unsafe.go b/pkg/sentry/platform/kvm/machine_unsafe.go index ed9433311..f04be2ab5 100644 --- a/pkg/sentry/platform/kvm/machine_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_unsafe.go @@ -87,46 +87,6 @@ func unmapRunData(r *runData) error { return nil } -// setUserRegisters sets user registers in the vCPU. -func (c *vCPU) setUserRegisters(uregs *userRegs) error { - if _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(c.fd), - _KVM_SET_REGS, - uintptr(unsafe.Pointer(uregs))); errno != 0 { - return fmt.Errorf("error setting user registers: %v", errno) - } - return nil -} - -// getUserRegisters reloads user registers in the vCPU. -// -// This is safe to call from a nosplit context. -// -//go:nosplit -func (c *vCPU) getUserRegisters(uregs *userRegs) syscall.Errno { - if _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(c.fd), - _KVM_GET_REGS, - uintptr(unsafe.Pointer(uregs))); errno != 0 { - return errno - } - return 0 -} - -// setSystemRegisters sets system registers. -func (c *vCPU) setSystemRegisters(sregs *systemRegs) error { - if _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(c.fd), - _KVM_SET_SREGS, - uintptr(unsafe.Pointer(sregs))); errno != 0 { - return fmt.Errorf("error setting system registers: %v", errno) - } - return nil -} - // atomicAddressSpace is an atomic address space pointer. type atomicAddressSpace struct { pointer unsafe.Pointer diff --git a/pkg/sentry/platform/kvm/physical_map.go b/pkg/sentry/platform/kvm/physical_map.go index 586e91bb2..91de5dab1 100644 --- a/pkg/sentry/platform/kvm/physical_map.go +++ b/pkg/sentry/platform/kvm/physical_map.go @@ -24,15 +24,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usermem" ) -const ( - // reservedMemory is a chunk of physical memory reserved starting at - // physical address zero. There are some special pages in this region, - // so we just call the whole thing off. - // - // Other architectures may define this to be zero. - reservedMemory = 0x100000000 -) - type region struct { virtual uintptr length uintptr @@ -59,8 +50,7 @@ func fillAddressSpace() (excludedRegions []region) { // We can cut vSize in half, because the kernel will be using the top // half and we ignore it while constructing mappings. It's as if we've // already excluded half the possible addresses. - vSize := uintptr(1) << ring0.VirtualAddressBits() - vSize = vSize >> 1 + vSize := ring0.UserspaceSize // We exclude reservedMemory below from our physical memory size, so it // needs to be dropped here as well. Otherwise, we could end up with diff --git a/pkg/sentry/platform/kvm/physical_map_amd64.go b/pkg/sentry/platform/kvm/physical_map_amd64.go new file mode 100644 index 000000000..c5adfb577 --- /dev/null +++ b/pkg/sentry/platform/kvm/physical_map_amd64.go @@ -0,0 +1,22 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kvm + +const ( + // reservedMemory is a chunk of physical memory reserved starting at + // physical address zero. There are some special pages in this region, + // so we just call the whole thing off. + reservedMemory = 0x100000000 +) diff --git a/pkg/sentry/fsimpl/proc/proc.go b/pkg/sentry/platform/kvm/physical_map_arm64.go index 31dec36de..4d8561453 100644 --- a/pkg/sentry/fsimpl/proc/proc.go +++ b/pkg/sentry/platform/kvm/physical_map_arm64.go @@ -12,5 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package proc implements a partial in-memory file system for procfs. -package proc +package kvm + +const ( + reservedMemory = 0 +) diff --git a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s index 2cd28b2d2..0bebee852 100644 --- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s +++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s @@ -50,6 +50,21 @@ TEXT ·SpinLoop(SB),NOSPLIT,$0 start: B start +TEXT ·FloatingPointWorks(SB),NOSPLIT,$0-8 + NO_LOCAL_POINTERS + FMOVD $(9.9), F0 + MOVD $SYS_GETPID, R8 // getpid + SVC + FMOVD $(9.9), F1 + FCMPD F0, F1 + BNE isNaN + MOVD $1, R0 + MOVD R0, ret+0(FP) + RET +isNaN: + MOVD $0, ret+0(FP) + RET + // MVN: bitwise logical NOT // This case simulates an application that modified R0-R30. #define TWIDDLE_REGS() \ diff --git a/pkg/sentry/platform/ptrace/stub_amd64.s b/pkg/sentry/platform/ptrace/stub_amd64.s index 64c718d21..16f9c523e 100644 --- a/pkg/sentry/platform/ptrace/stub_amd64.s +++ b/pkg/sentry/platform/ptrace/stub_amd64.s @@ -64,6 +64,8 @@ begin: CMPQ AX, $0 JL error + MOVQ $0, BX + // SIGSTOP to wait for attach. // // The SYSCALL instruction will be used for future syscall injection by @@ -73,23 +75,26 @@ begin: MOVQ $SIGSTOP, SI SYSCALL - // The tracer may "detach" and/or allow code execution here in three cases: - // - // 1. New (traced) stub threads are explicitly detached by the - // goroutine in newSubprocess. However, they are detached while in - // group-stop, so they do not execute code here. - // - // 2. If a tracer thread exits, it implicitly detaches from the stub, - // potentially allowing code execution here. However, the Go runtime - // never exits individual threads, so this case never occurs. - // - // 3. subprocess.createStub clones a new stub process that is untraced, + // The sentry sets BX to 1 when creating stub process. + CMPQ BX, $1 + JE clone + + // Notify the Sentry that syscall exited. +done: + INT $3 + // Be paranoid. + JMP done +clone: + // subprocess.createStub clones a new stub process that is untraced, // thus executing this code. We setup the PDEATHSIG before SIGSTOPing // ourselves for attach by the tracer. // // R15 has been updated with the expected PPID. - JMP begin + CMPQ AX, $0 + JE begin + // The clone syscall returns a non-zero value. + JMP done error: // Exit with -errno. MOVQ AX, DI diff --git a/pkg/sentry/platform/ptrace/stub_arm64.s b/pkg/sentry/platform/ptrace/stub_arm64.s index 2c5e4d5cb..6162df02a 100644 --- a/pkg/sentry/platform/ptrace/stub_arm64.s +++ b/pkg/sentry/platform/ptrace/stub_arm64.s @@ -59,6 +59,8 @@ begin: CMP $0x0, R0 BLT error + MOVD $0, R9 + // SIGSTOP to wait for attach. // // The SYSCALL instruction will be used for future syscall injection by @@ -66,22 +68,26 @@ begin: MOVD $SYS_KILL, R8 MOVD $SIGSTOP, R1 SVC - // The tracer may "detach" and/or allow code execution here in three cases: - // - // 1. New (traced) stub threads are explicitly detached by the - // goroutine in newSubprocess. However, they are detached while in - // group-stop, so they do not execute code here. - // - // 2. If a tracer thread exits, it implicitly detaches from the stub, - // potentially allowing code execution here. However, the Go runtime - // never exits individual threads, so this case never occurs. - // - // 3. subprocess.createStub clones a new stub process that is untraced, + + // The sentry sets R9 to 1 when creating stub process. + CMP $1, R9 + BEQ clone + +done: + // Notify the Sentry that syscall exited. + BRK $3 + B done // Be paranoid. +clone: + // subprocess.createStub clones a new stub process that is untraced, // thus executing this code. We setup the PDEATHSIG before SIGSTOPing // ourselves for attach by the tracer. // // R7 has been updated with the expected PPID. - B begin + CMP $0, R0 + BEQ begin + + // The clone system call returned a non-zero value. + B done error: // Exit with -errno. diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index ddb1f41e3..20244fd95 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -21,6 +21,7 @@ import ( "sync" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -429,13 +430,15 @@ func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { } for { - // Execute the syscall instruction. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + // Execute the syscall instruction. The task has to stop on the + // trap instruction which is right after the syscall + // instruction. + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_CONT, 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) { + if sig == syscall.SIGTRAP { // Reached syscall-enter-stop. break } else { @@ -447,18 +450,6 @@ func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { } } - // Complete the actual system call. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { - panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) - } - - // Wait for syscall-exit-stop. "[Signal-delivery-stop] never happens - // between syscall-enter-stop and syscall-exit-stop; it happens *after* - // syscall-exit-stop.)" - ptrace(2), "Syscall-stops" - if sig := t.wait(stopped); sig != (syscallEvent | syscall.SIGTRAP) { - t.dumpAndPanic(fmt.Sprintf("wait failed: expected SIGTRAP, got %v [%d]", sig, sig)) - } - // Grab registers. if err := t.getRegs(regs); err != nil { panic(fmt.Sprintf("ptrace get regs failed: %v", err)) @@ -541,14 +532,14 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { if isSingleStepping(regs) { if _, _, errno := syscall.RawSyscall6( syscall.SYS_PTRACE, - syscall.PTRACE_SYSEMU_SINGLESTEP, + unix.PTRACE_SYSEMU_SINGLESTEP, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace sysemu failed: %v", errno)) } } else { if _, _, errno := syscall.RawSyscall6( syscall.SYS_PTRACE, - syscall.PTRACE_SYSEMU, + unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace sysemu failed: %v", errno)) } diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index 4649a94a7..e99798c56 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,8 @@ import ( "strings" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" ) @@ -139,7 +141,55 @@ func (t *thread) adjustInitRegsRip() { t.initRegs.Rip -= initRegsRipAdjustment } -// Pass the expected PPID to the child via R15 when creating stub process +// Pass the expected PPID to the child via R15 when creating stub process. func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.R15 = uint64(ppid) + // Rbx has to be set to 1 when creating stub process. + initregs.Rbx = 1 +} + +// patchSignalInfo patches the signal info to account for hitting the seccomp +// filters from vsyscall emulation, specified below. We allow for SIGSYS as a +// synchronous trap, but patch the structure to appear like a SIGSEGV with the +// Rip as the faulting address. +// +// Note that this should only be called after verifying that the signalInfo has +// been generated by the kernel. +func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { + if linux.Signal(signalInfo.Signo) == linux.SIGSYS { + signalInfo.Signo = int32(linux.SIGSEGV) + + // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered + // with the si_call_addr field pointing to the current RIP. This field + // aligns with the si_addr field for a SIGSEGV, so we don't need to touch + // anything there. We do need to unwind emulation however, so we set the + // instruction pointer to the faulting value, and "unpop" the stack. + regs.Rip = signalInfo.Addr() + regs.Rsp -= 8 + } +} + +// enableCpuidFault enables cpuid-faulting. +// +// This may fail on older kernels or hardware, so we just disregard the result. +// Host CPUID will be enabled. +// +// This is safe to call in an afterFork context. +// +//go:nosplit +func enableCpuidFault() { + syscall.RawSyscall6(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0, 0, 0, 0) +} + +// 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)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) } diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index bec884ba5..7b975137f 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -17,8 +17,12 @@ package ptrace import ( + "fmt" + "strings" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" ) @@ -37,7 +41,7 @@ const ( // resetSysemuRegs sets up emulation registers. // // This should be called prior to calling sysemu. -func (s *subprocess) resetSysemuRegs(regs *syscall.PtraceRegs) { +func (t *thread) resetSysemuRegs(regs *syscall.PtraceRegs) { } // createSyscallRegs sets up syscall registers. @@ -123,4 +127,39 @@ func (t *thread) adjustInitRegsRip() { // Pass the expected PPID to the child via X7 when creating stub process func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.Regs[7] = uint64(ppid) + // R9 has to be set to 1 when creating stub process. + initregs.Regs[9] = 1 +} + +// patchSignalInfo patches the signal info to account for hitting the seccomp +// filters from vsyscall emulation, specified below. We allow for SIGSYS as a +// synchronous trap, but patch the structure to appear like a SIGSEGV with the +// Rip as the faulting address. +// +// Note that this should only be called after verifying that the signalInfo has +// been generated by the kernel. +func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { + if linux.Signal(signalInfo.Signo) == linux.SIGSYS { + signalInfo.Signo = int32(linux.SIGSEGV) + + // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered + // with the si_call_addr field pointing to the current RIP. This field + // aligns with the si_addr field for a SIGSEGV, so we don't need to touch + // anything there. We do need to unwind emulation however, so we set the + // instruction pointer to the faulting value, and "unpop" the stack. + regs.Pc = signalInfo.Addr() + regs.Sp -= 8 + } +} + +// Noop on arm64. +// +//go:nosplit +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 rules } diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 3782d4332..74968dfdf 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,6 +20,7 @@ import ( "fmt" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -53,7 +54,7 @@ func probeSeccomp() bool { for { // Attempt an emulation. - if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { + 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)) } @@ -77,27 +78,6 @@ func probeSeccomp() bool { } } -// patchSignalInfo patches the signal info to account for hitting the seccomp -// filters from vsyscall emulation, specified below. We allow for SIGSYS as a -// synchronous trap, but patch the structure to appear like a SIGSEGV with the -// Rip as the faulting address. -// -// Note that this should only be called after verifying that the signalInfo has -// been generated by the kernel. -func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { - if linux.Signal(signalInfo.Signo) == linux.SIGSYS { - signalInfo.Signo = int32(linux.SIGSEGV) - - // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered - // with the si_call_addr field pointing to the current RIP. This field - // aligns with the si_addr field for a SIGSEGV, so we don't need to touch - // anything there. We do need to unwind emulation however, so we set the - // instruction pointer to the faulting value, and "unpop" the stack. - regs.Rip = signalInfo.Addr() - regs.Rsp -= 8 - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -149,7 +129,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro Rules: seccomp.SyscallRules{ syscall.SYS_GETTIMEOFDAY: {}, syscall.SYS_TIME: {}, - 309: {}, // SYS_GETCPU. + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, Action: linux.SECCOMP_RET_TRAP, Vsyscall: true, @@ -173,10 +153,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // For the initial process creation. syscall.SYS_WAIT4: {}, - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, - }, - syscall.SYS_EXIT: {}, + syscall.SYS_EXIT: {}, // For the stub prctl dance (all). syscall.SYS_PRCTL: []seccomp.Rule{ @@ -196,6 +173,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) + + rules = appendArchSeccompRules(rules) } instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { @@ -267,9 +246,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro syscall.RawSyscall(syscall.SYS_EXIT, uintptr(errno), 0, 0) } - // Enable cpuid-faulting; this may fail on older kernels or hardware, - // so we just disregard the result. Host CPUID will be enabled. - syscall.RawSyscall6(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0, 0, 0, 0) + // Enable cpuid-faulting. + enableCpuidFault() // Call the stub; should not return. stubCall(stubStart, ppid) diff --git a/pkg/sentry/platform/ring0/BUILD b/pkg/sentry/platform/ring0/BUILD index f1af18265..87f4552b5 100644 --- a/pkg/sentry/platform/ring0/BUILD +++ b/pkg/sentry/platform/ring0/BUILD @@ -71,6 +71,7 @@ go_library( "lib_amd64.go", "lib_amd64.s", "lib_arm64.go", + "lib_arm64.s", "ring0.go", ], importpath = "gvisor.dev/gvisor/pkg/sentry/platform/ring0", diff --git a/pkg/sentry/platform/ring0/defs_arm64.go b/pkg/sentry/platform/ring0/defs_arm64.go index fbfbd9bab..dc0eeec01 100644 --- a/pkg/sentry/platform/ring0/defs_arm64.go +++ b/pkg/sentry/platform/ring0/defs_arm64.go @@ -73,6 +73,9 @@ type CPUArchState struct { // application context pointer. appAddr uintptr + + // lazyVFP is the value of cpacr_el1. + lazyVFP uintptr } // ErrorCode returns the last error code. diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index 674c569e1..64e9c0845 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -31,6 +31,11 @@ #define RSV_REG R18_PLATFORM #define RSV_REG_APP R9 +#define FPEN_NOTRAP 0x3 +#define FPEN_SHIFT 20 + +#define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT) + #define REGISTERS_SAVE(reg, offset) \ MOVD R0, offset+PTRACE_R0(reg); \ MOVD R1, offset+PTRACE_R1(reg); \ @@ -279,6 +284,16 @@ #define IRQ_DISABLE \ MSR $2, DAIFClr; +#define VFP_ENABLE \ + MOVD $FPEN_ENABLE, R0; \ + WORD $0xd5181040; \ //MSR R0, CPACR_EL1 + ISB $15; + +#define VFP_DISABLE \ + MOVD $0x0, R0; \ + WORD $0xd5181040; \ //MSR R0, CPACR_EL1 + ISB $15; + #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); \ @@ -318,6 +333,11 @@ TEXT ·Halt(SB),NOSPLIT,$0 BNE mmio_exit MOVD $0, CPU_REGISTERS+PTRACE_R9(RSV_REG) mmio_exit: + // Disable fpsimd. + WORD $0xd5381041 // MRS CPACR_EL1, R1 + MOVD R1, CPU_LAZY_VFP(RSV_REG) + VFP_DISABLE + // MMIO_EXIT. MOVD $0, R9 MOVD R0, 0xffff000000001000(R9) @@ -449,6 +469,8 @@ TEXT ·El1_sync(SB),NOSPLIT,$0 BEQ el1_svc CMP $ESR_ELx_EC_BREAKPT_CUR, R24 BGE el1_dbg + CMP $ESR_ELx_EC_FP_ASIMD, R24 + BEQ el1_fpsimd_acc B el1_invalid el1_da: @@ -469,6 +491,10 @@ el1_svc: el1_dbg: B ·Shutdown(SB) +el1_fpsimd_acc: + VFP_ENABLE + B ·kernelExitToEl1(SB) // Resume. + el1_invalid: B ·Shutdown(SB) @@ -595,6 +621,10 @@ TEXT ·Vectors(SB),NOSPLIT,$0 B ·El0_error_invalid(SB) nop31Instructions() + // The exception-vector-table is required to be 11-bits aligned. + // Please see Linux source code as reference: arch/arm64/kernel/entry.s. + // For gvisor, I defined it as 4K in length, filled the 2nd 2K part with NOPs. + // So that, I can safely move the 1st 2K part into the address with 11-bits alignment. WORD $0xd503201f //nop nop31Instructions() WORD $0xd503201f diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index 900ee6380..8bcfe1032 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -16,10 +16,24 @@ package ring0 -// LoadFloatingPoint loads floating point state by the most efficient mechanism -// available (set by Init). -var LoadFloatingPoint func(*byte) +// CPACREL1 returns the value of the CPACR_EL1 register. +func CPACREL1() (value uintptr) -// SaveFloatingPoint saves floating point state by the most efficient mechanism -// available (set by Init). -var SaveFloatingPoint func(*byte) +// FPCR returns the value of FPCR register. +func FPCR() (value uintptr) + +// SetFPCR writes the FPCR value. +func SetFPCR(value uintptr) + +// FPSR returns the value of FPSR register. +func FPSR() (value uintptr) + +// SetFPSR writes the FPSR value. +func SetFPSR(value uintptr) + +// SaveVRegs saves V0-V31 registers. +// V0-V31: 32 128-bit registers for floating point and simd. +func SaveVRegs(*byte) + +// LoadVRegs loads V0-V31 registers. +func LoadVRegs(*byte) diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s new file mode 100644 index 000000000..0e6a6235b --- /dev/null +++ b/pkg/sentry/platform/ring0/lib_arm64.s @@ -0,0 +1,121 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "funcdata.h" +#include "textflag.h" + +TEXT ·CPACREL1(SB),NOSPLIT,$0-8 + WORD $0xd5381041 // MRS CPACR_EL1, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·GetFPCR(SB),NOSPLIT,$0-8 + WORD $0xd53b4201 // MRS NZCV, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·GetFPSR(SB),NOSPLIT,$0-8 + WORD $0xd53b4421 // MRS FPSR, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·SetFPCR(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + WORD $0xd51b4201 // MSR R1, NZCV + RET + +TEXT ·SetFPSR(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + WORD $0xd51b4421 // MSR R1, FPSR + RET + +TEXT ·SaveVRegs(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + // Skip aarch64_ctx, fpsr, fpcr. + FMOVD F0, 16*1(R0) + FMOVD F1, 16*2(R0) + FMOVD F2, 16*3(R0) + FMOVD F3, 16*4(R0) + FMOVD F4, 16*5(R0) + FMOVD F5, 16*6(R0) + FMOVD F6, 16*7(R0) + FMOVD F7, 16*8(R0) + FMOVD F8, 16*9(R0) + FMOVD F9, 16*10(R0) + FMOVD F10, 16*11(R0) + FMOVD F11, 16*12(R0) + FMOVD F12, 16*13(R0) + FMOVD F13, 16*14(R0) + FMOVD F14, 16*15(R0) + FMOVD F15, 16*16(R0) + FMOVD F16, 16*17(R0) + FMOVD F17, 16*18(R0) + FMOVD F18, 16*19(R0) + FMOVD F19, 16*20(R0) + FMOVD F20, 16*21(R0) + FMOVD F21, 16*22(R0) + FMOVD F22, 16*23(R0) + FMOVD F23, 16*24(R0) + FMOVD F24, 16*25(R0) + FMOVD F25, 16*26(R0) + FMOVD F26, 16*27(R0) + FMOVD F27, 16*28(R0) + FMOVD F28, 16*29(R0) + FMOVD F29, 16*30(R0) + FMOVD F30, 16*31(R0) + FMOVD F31, 16*32(R0) + ISB $15 + + RET + +TEXT ·LoadVRegs(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + // Skip aarch64_ctx, fpsr, fpcr. + FMOVD 16*1(R0), F0 + FMOVD 16*2(R0), F1 + FMOVD 16*3(R0), F2 + FMOVD 16*4(R0), F3 + FMOVD 16*5(R0), F4 + FMOVD 16*6(R0), F5 + FMOVD 16*7(R0), F6 + FMOVD 16*8(R0), F7 + FMOVD 16*9(R0), F8 + FMOVD 16*10(R0), F9 + FMOVD 16*11(R0), F10 + FMOVD 16*12(R0), F11 + FMOVD 16*13(R0), F12 + FMOVD 16*14(R0), F13 + FMOVD 16*15(R0), F14 + FMOVD 16*16(R0), F15 + FMOVD 16*17(R0), F16 + FMOVD 16*18(R0), F17 + FMOVD 16*19(R0), F18 + FMOVD 16*20(R0), F19 + FMOVD 16*21(R0), F20 + FMOVD 16*22(R0), F21 + FMOVD 16*23(R0), F22 + FMOVD 16*24(R0), F23 + FMOVD 16*25(R0), F24 + FMOVD 16*26(R0), F25 + FMOVD 16*27(R0), F26 + FMOVD 16*28(R0), F27 + FMOVD 16*29(R0), F28 + FMOVD 16*30(R0), F29 + FMOVD 16*31(R0), F30 + FMOVD 16*32(R0), F31 + ISB $15 + + RET diff --git a/pkg/sentry/platform/ring0/offsets_arm64.go b/pkg/sentry/platform/ring0/offsets_arm64.go index d7aa1c7cc..cd2a65f97 100644 --- a/pkg/sentry/platform/ring0/offsets_arm64.go +++ b/pkg/sentry/platform/ring0/offsets_arm64.go @@ -39,6 +39,7 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define CPU_TTBR0_APP 0x%02x\n", reflect.ValueOf(&c.ttbr0App).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_VECTOR_CODE 0x%02x\n", reflect.ValueOf(&c.vecCode).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_APP_ADDR 0x%02x\n", reflect.ValueOf(&c.appAddr).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_LAZY_VFP 0x%02x\n", reflect.ValueOf(&c.lazyVFP).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "\n// Bits.\n") fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) diff --git a/pkg/sentry/sighandling/sighandling.go b/pkg/sentry/sighandling/sighandling.go index 2f65db70b..ba1f9043d 100644 --- a/pkg/sentry/sighandling/sighandling.go +++ b/pkg/sentry/sighandling/sighandling.go @@ -16,7 +16,6 @@ package sighandling import ( - "fmt" "os" "os/signal" "reflect" @@ -31,37 +30,25 @@ const numSignals = 32 // handleSignals listens for incoming signals and calls the given handler // function. // -// It starts when the start channel is closed, stops when the stop channel -// is closed, and closes done once it will no longer deliver signals to k. -func handleSignals(sigchans []chan os.Signal, handler func(linux.Signal), start, stop, done chan struct{}) { +// It stops when the stop channel is closed. The done channel is closed once it +// will no longer deliver signals to k. +func handleSignals(sigchans []chan os.Signal, handler func(linux.Signal), stop, done chan struct{}) { // Build a select case. - sc := []reflect.SelectCase{{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(start)}} + sc := []reflect.SelectCase{{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(stop)}} for _, sigchan := range sigchans { sc = append(sc, reflect.SelectCase{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(sigchan)}) } - started := false for { // Wait for a notification. index, _, ok := reflect.Select(sc) - // Was it the start / stop channel? + // Was it the stop channel? if index == 0 { if !ok { - if !started { - // start channel; start forwarding and - // swap this case for the stop channel - // to select stop requests. - started = true - sc[0] = reflect.SelectCase{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(stop)} - } else { - // stop channel; stop forwarding and - // clear this case so it is never - // selected again. - started = false - close(done) - sc[0].Chan = reflect.Value{} - } + // Stop forwarding and notify that it's done. + close(done) + return } continue } @@ -73,44 +60,17 @@ func handleSignals(sigchans []chan os.Signal, handler func(linux.Signal), start, // Otherwise, it was a signal on channel N. Index 0 represents the stop // channel, so index N represents the channel for signal N. - signal := linux.Signal(index) - - if !started { - // Kernel cannot receive signals, either because it is - // not ready yet or is shutting down. - // - // Kill ourselves if this signal would have killed the - // process before PrepareForwarding was called. i.e., all - // _SigKill signals; see Go - // src/runtime/sigtab_linux_generic.go. - // - // Otherwise ignore the signal. - // - // TODO(b/114489875): Drop in Go 1.12, which uses tgkill - // in runtime.raise. - switch signal { - case linux.SIGHUP, linux.SIGINT, linux.SIGTERM: - dieFromSignal(signal) - panic(fmt.Sprintf("Failed to die from signal %d", signal)) - default: - continue - } - } - - // Pass the signal to the handler. - handler(signal) + handler(linux.Signal(index)) } } -// PrepareHandler ensures that synchronous signals are passed to the given -// handler function and returns a callback that starts signal delivery, which -// itself returns a callback that stops signal handling. +// StartSignalForwarding ensures that synchronous signals are passed to the +// given handler function and returns a callback that stops signal delivery. // // Note that this function permanently takes over signal handling. After the // stop callback, signals revert to the default Go runtime behavior, which // cannot be overridden with external calls to signal.Notify. -func PrepareHandler(handler func(linux.Signal)) func() func() { - start := make(chan struct{}) +func StartSignalForwarding(handler func(linux.Signal)) func() { stop := make(chan struct{}) done := make(chan struct{}) @@ -128,13 +88,10 @@ func PrepareHandler(handler func(linux.Signal)) func() func() { signal.Notify(sigchan, syscall.Signal(sig)) } // Start up our listener. - go handleSignals(sigchans, handler, start, stop, done) // S/R-SAFE: synchronized by Kernel.extMu. + go handleSignals(sigchans, handler, stop, done) // S/R-SAFE: synchronized by Kernel.extMu. - return func() func() { - close(start) - return func() { - close(stop) - <-done - } + return func() { + close(stop) + <-done } } diff --git a/pkg/sentry/sighandling/sighandling_unsafe.go b/pkg/sentry/sighandling/sighandling_unsafe.go index c303435d5..1ebe22d34 100644 --- a/pkg/sentry/sighandling/sighandling_unsafe.go +++ b/pkg/sentry/sighandling/sighandling_unsafe.go @@ -15,8 +15,6 @@ package sighandling import ( - "fmt" - "runtime" "syscall" "unsafe" @@ -48,27 +46,3 @@ func IgnoreChildStop() error { return nil } - -// dieFromSignal kills the current process with sig. -// -// Preconditions: The default action of sig is termination. -func dieFromSignal(sig linux.Signal) { - runtime.LockOSThread() - defer runtime.UnlockOSThread() - - sa := sigaction{handler: linux.SIG_DFL} - if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(sig), uintptr(unsafe.Pointer(&sa)), 0, linux.SignalSetSize, 0, 0); e != 0 { - panic(fmt.Sprintf("rt_sigaction failed: %v", e)) - } - - set := linux.MakeSignalSet(sig) - if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGPROCMASK, linux.SIG_UNBLOCK, uintptr(unsafe.Pointer(&set)), 0, linux.SignalSetSize, 0, 0); e != 0 { - panic(fmt.Sprintf("rt_sigprocmask failed: %v", e)) - } - - if err := syscall.Tgkill(syscall.Getpid(), syscall.Gettid(), syscall.Signal(sig)); err != nil { - panic(fmt.Sprintf("tgkill failed: %v", err)) - } - - panic("failed to die") -} diff --git a/pkg/sentry/socket/control/BUILD b/pkg/sentry/socket/control/BUILD index 4a6e83a8b..357517ed4 100644 --- a/pkg/sentry/socket/control/BUILD +++ b/pkg/sentry/socket/control/BUILD @@ -17,6 +17,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", + "//pkg/sentry/socket", "//pkg/sentry/socket/unix/transport", "//pkg/sentry/usermem", "//pkg/syserror", diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go index 4e95101b7..af1a4e95f 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" @@ -194,15 +195,15 @@ func putCmsg(buf []byte, flags int, msgType uint32, align uint, data []int32) ([ // the available space, we must align down. // // align must be >= 4 and each data int32 is 4 bytes. The length of the - // header is already aligned, so if we align to the with of the data there + // header is already aligned, so if we align to the width of the data there // are two cases: // 1. The aligned length is less than the length of the header. The // unaligned length was also less than the length of the header, so we // can't write anything. // 2. The aligned length is greater than or equal to the length of the - // header. We can write the header plus zero or more datas. We can't write - // a partial int32, so the length of the message will be - // min(aligned length, header + datas). + // header. We can write the header plus zero or more bytes of data. We can't + // write a partial int32, so the length of the message will be + // min(aligned length, header + data). if space < linux.SizeOfControlMessageHeader { flags |= linux.MSG_CTRUNC return buf, flags @@ -239,12 +240,12 @@ func putCmsgStruct(buf []byte, msgLevel, msgType uint32, align uint, data interf buf = binary.Marshal(buf, usermem.ByteOrder, data) - // Check if we went over. + // If the control message data brought us over capacity, omit it. if cap(buf) != cap(ob) { return hdrBuf } - // Fix up length. + // Update control message length to include data. putUint64(ob, uint64(len(buf)-len(ob))) return alignSlice(buf, align) @@ -320,35 +321,109 @@ func PackInq(t *kernel.Task, inq int32, buf []byte) []byte { buf, linux.SOL_TCP, linux.TCP_INQ, - 4, + t.Arch().Width(), inq, ) } +// PackTOS packs an IP_TOS socket control message. +func PackTOS(t *kernel.Task, tos int8, buf []byte) []byte { + return putCmsgStruct( + buf, + linux.SOL_IP, + linux.IP_TOS, + t.Arch().Width(), + tos, + ) +} + +// PackTClass packs an IPV6_TCLASS socket control message. +func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { + return putCmsgStruct( + buf, + linux.SOL_IPV6, + linux.IPV6_TCLASS, + t.Arch().Width(), + tClass, + ) +} + +// PackControlMessages packs control messages into the given buffer. +// +// We skip control messages specific to Unix domain sockets. +// +// Note that some control messages may be truncated if they do not fit under +// the capacity of buf. +func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages, buf []byte) []byte { + if cmsgs.IP.HasTimestamp { + buf = PackTimestamp(t, cmsgs.IP.Timestamp, buf) + } + + if cmsgs.IP.HasInq { + // In Linux, TCP_CM_INQ is added after SO_TIMESTAMP. + buf = PackInq(t, cmsgs.IP.Inq, buf) + } + + if cmsgs.IP.HasTOS { + buf = PackTOS(t, cmsgs.IP.TOS, buf) + } + + if cmsgs.IP.HasTClass { + buf = PackTClass(t, cmsgs.IP.TClass, buf) + } + + return buf +} + +// cmsgSpace is equivalent to CMSG_SPACE in Linux. +func cmsgSpace(t *kernel.Task, dataLen int) int { + return linux.SizeOfControlMessageHeader + AlignUp(dataLen, t.Arch().Width()) +} + +// CmsgsSpace returns the number of bytes needed to fit the control messages +// represented in cmsgs. +func CmsgsSpace(t *kernel.Task, cmsgs socket.ControlMessages) int { + space := 0 + + if cmsgs.IP.HasTimestamp { + space += cmsgSpace(t, linux.SizeOfTimeval) + } + + if cmsgs.IP.HasInq { + space += cmsgSpace(t, linux.SizeOfControlMessageInq) + } + + if cmsgs.IP.HasTOS { + space += cmsgSpace(t, linux.SizeOfControlMessageTOS) + } + + if cmsgs.IP.HasTClass { + space += cmsgSpace(t, linux.SizeOfControlMessageTClass) + } + + return space +} + // Parse parses a raw socket control message into portable objects. -func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (transport.ControlMessages, error) { +func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.ControlMessages, error) { var ( - fds linux.ControlMessageRights - haveCreds bool - creds linux.ControlMessageCredentials + cmsgs socket.ControlMessages + fds linux.ControlMessageRights ) for i := 0; i < len(buf); { if i+linux.SizeOfControlMessageHeader > len(buf) { - return transport.ControlMessages{}, syserror.EINVAL + return cmsgs, syserror.EINVAL } var h linux.ControlMessageHeader binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageHeader], usermem.ByteOrder, &h) if h.Length < uint64(linux.SizeOfControlMessageHeader) { - return transport.ControlMessages{}, syserror.EINVAL + return socket.ControlMessages{}, syserror.EINVAL } if h.Length > uint64(len(buf)-i) { - return transport.ControlMessages{}, syserror.EINVAL - } - if h.Level != linux.SOL_SOCKET { - return transport.ControlMessages{}, syserror.EINVAL + return socket.ControlMessages{}, syserror.EINVAL } i += linux.SizeOfControlMessageHeader @@ -358,59 +433,79 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (transport. // sizeof(long) in CMSG_ALIGN. width := t.Arch().Width() - switch h.Type { - case linux.SCM_RIGHTS: - rightsSize := AlignDown(length, linux.SizeOfControlMessageRight) - numRights := rightsSize / linux.SizeOfControlMessageRight - - if len(fds)+numRights > linux.SCM_MAX_FD { - return transport.ControlMessages{}, syserror.EINVAL + switch h.Level { + case linux.SOL_SOCKET: + switch h.Type { + case linux.SCM_RIGHTS: + rightsSize := AlignDown(length, linux.SizeOfControlMessageRight) + numRights := rightsSize / linux.SizeOfControlMessageRight + + if len(fds)+numRights > linux.SCM_MAX_FD { + return socket.ControlMessages{}, syserror.EINVAL + } + + for j := i; j < i+rightsSize; j += linux.SizeOfControlMessageRight { + fds = append(fds, int32(usermem.ByteOrder.Uint32(buf[j:j+linux.SizeOfControlMessageRight]))) + } + + i += AlignUp(length, width) + + case linux.SCM_CREDENTIALS: + if length < linux.SizeOfControlMessageCredentials { + return socket.ControlMessages{}, syserror.EINVAL + } + + var creds linux.ControlMessageCredentials + binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageCredentials], usermem.ByteOrder, &creds) + scmCreds, err := NewSCMCredentials(t, creds) + if err != nil { + return socket.ControlMessages{}, err + } + cmsgs.Unix.Credentials = scmCreds + i += AlignUp(length, width) + + default: + // Unknown message type. + return socket.ControlMessages{}, syserror.EINVAL } - - for j := i; j < i+rightsSize; j += linux.SizeOfControlMessageRight { - fds = append(fds, int32(usermem.ByteOrder.Uint32(buf[j:j+linux.SizeOfControlMessageRight]))) + case linux.SOL_IP: + switch h.Type { + case linux.IP_TOS: + cmsgs.IP.HasTOS = true + binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTOS], usermem.ByteOrder, &cmsgs.IP.TOS) + i += AlignUp(length, width) + + default: + return socket.ControlMessages{}, syserror.EINVAL } - - i += AlignUp(length, width) - - case linux.SCM_CREDENTIALS: - if length < linux.SizeOfControlMessageCredentials { - return transport.ControlMessages{}, syserror.EINVAL + case linux.SOL_IPV6: + switch h.Type { + case linux.IPV6_TCLASS: + cmsgs.IP.HasTClass = true + binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageTClass], usermem.ByteOrder, &cmsgs.IP.TClass) + i += AlignUp(length, width) + + default: + return socket.ControlMessages{}, syserror.EINVAL } - - binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageCredentials], usermem.ByteOrder, &creds) - haveCreds = true - i += AlignUp(length, width) - default: - // Unknown message type. - return transport.ControlMessages{}, syserror.EINVAL + return socket.ControlMessages{}, syserror.EINVAL } } - var credentials SCMCredentials - if haveCreds { - var err error - if credentials, err = NewSCMCredentials(t, creds); err != nil { - return transport.ControlMessages{}, err - } - } else { - credentials = makeCreds(t, socketOrEndpoint) + if cmsgs.Unix.Credentials == nil { + cmsgs.Unix.Credentials = makeCreds(t, socketOrEndpoint) } - var rights SCMRights if len(fds) > 0 { - var err error - if rights, err = NewSCMRights(t, fds); err != nil { - return transport.ControlMessages{}, err + rights, err := NewSCMRights(t, fds) + if err != nil { + return socket.ControlMessages{}, err } + cmsgs.Unix.Rights = rights } - if credentials == nil && rights == nil { - return transport.ControlMessages{}, nil - } - - return transport.ControlMessages{Credentials: credentials, Rights: rights}, nil + return cmsgs, nil } func makeCreds(t *kernel.Task, socketOrEndpoint interface{}) SCMCredentials { diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index 8b66a719d..4c44c7c0f 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -29,10 +29,12 @@ go_library( "//pkg/sentry/kernel/time", "//pkg/sentry/safemem", "//pkg/sentry/socket", + "//pkg/sentry/socket/control", "//pkg/sentry/usermem", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip/stack", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index 92beb1bcf..c957b0f1d 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -18,6 +18,7 @@ import ( "fmt" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/fdnotifier" @@ -29,6 +30,7 @@ import ( ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/safemem" "gvisor.dev/gvisor/pkg/sentry/socket" + "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" @@ -41,6 +43,10 @@ const ( // sizeofSockaddr is the size in bytes of the largest sockaddr type // supported by this package. sizeofSockaddr = syscall.SizeofSockaddrInet6 // sizeof(sockaddr_in6) > sizeof(sockaddr_in) + + // maxControlLen is the maximum size of a control message buffer used in a + // recvmsg or sendmsg syscall. + maxControlLen = 1024 ) // socketOperations implements fs.FileOperations and socket.Socket for a socket @@ -281,26 +287,32 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt // Whitelist options and constrain option length. var optlen int switch level { - case syscall.SOL_IPV6: + case linux.SOL_IP: + switch name { + case linux.IP_TOS, linux.IP_RECVTOS: + optlen = sizeofInt32 + } + case linux.SOL_IPV6: switch name { - case syscall.IPV6_V6ONLY: + case linux.IPV6_TCLASS, linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY: optlen = sizeofInt32 } - case syscall.SOL_SOCKET: + case linux.SOL_SOCKET: switch name { - case syscall.SO_ERROR, syscall.SO_KEEPALIVE, syscall.SO_SNDBUF, syscall.SO_RCVBUF, syscall.SO_REUSEADDR: + case linux.SO_ERROR, linux.SO_KEEPALIVE, linux.SO_SNDBUF, linux.SO_RCVBUF, linux.SO_REUSEADDR: optlen = sizeofInt32 - case syscall.SO_LINGER: + case linux.SO_LINGER: optlen = syscall.SizeofLinger } - case syscall.SOL_TCP: + case linux.SOL_TCP: switch name { - case syscall.TCP_NODELAY: + case linux.TCP_NODELAY: optlen = sizeofInt32 - case syscall.TCP_INFO: + case linux.TCP_INFO: optlen = int(linux.SizeOfTCPInfo) } } + if optlen == 0 { return nil, syserr.ErrProtocolNotAvailable // ENOPROTOOPT } @@ -320,19 +332,24 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ // Whitelist options and constrain option length. var optlen int switch level { - case syscall.SOL_IPV6: + case linux.SOL_IP: + switch name { + case linux.IP_TOS, linux.IP_RECVTOS: + optlen = sizeofInt32 + } + case linux.SOL_IPV6: switch name { - case syscall.IPV6_V6ONLY: + case linux.IPV6_TCLASS, linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY: optlen = sizeofInt32 } - case syscall.SOL_SOCKET: + case linux.SOL_SOCKET: switch name { - case syscall.SO_SNDBUF, syscall.SO_RCVBUF, syscall.SO_REUSEADDR: + case linux.SO_SNDBUF, linux.SO_RCVBUF, linux.SO_REUSEADDR: optlen = sizeofInt32 } - case syscall.SOL_TCP: + case linux.SOL_TCP: switch name { - case syscall.TCP_NODELAY: + case linux.TCP_NODELAY: optlen = sizeofInt32 } } @@ -354,11 +371,11 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ } // RecvMsg implements socket.Socket.RecvMsg. -func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { +func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) { // Whitelist flags. // // FIXME(jamieliu): We can't support MSG_ERRQUEUE because it uses ancillary - // messages that netstack/tcpip/transport/unix doesn't understand. Kill the + // messages that gvisor/pkg/tcpip/transport/unix doesn't understand. Kill the // Socket interface's dependence on netstack. if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_PEEK|syscall.MSG_TRUNC) != 0 { return 0, 0, nil, 0, socket.ControlMessages{}, syserr.ErrInvalidArgument @@ -370,6 +387,7 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags senderAddrBuf = make([]byte, sizeofSockaddr) } + var controlBuf []byte var msgFlags int recvmsgToBlocks := safemem.ReaderFunc(func(dsts safemem.BlockSeq) (uint64, error) { @@ -384,11 +402,6 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags // We always do a non-blocking recv*(). sysflags := flags | syscall.MSG_DONTWAIT - if dsts.NumBlocks() == 1 { - // Skip allocating []syscall.Iovec. - return recvfrom(s.fd, dsts.Head().ToSlice(), sysflags, &senderAddrBuf) - } - iovs := iovecsFromBlockSeq(dsts) msg := syscall.Msghdr{ Iov: &iovs[0], @@ -398,12 +411,21 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags msg.Name = &senderAddrBuf[0] msg.Namelen = uint32(len(senderAddrBuf)) } + if controlLen > 0 { + if controlLen > maxControlLen { + controlLen = maxControlLen + } + controlBuf = make([]byte, controlLen) + msg.Control = &controlBuf[0] + msg.Controllen = controlLen + } n, err := recvmsg(s.fd, &msg, sysflags) if err != nil { return 0, err } senderAddrBuf = senderAddrBuf[:msg.Namelen] msgFlags = int(msg.Flags) + controlLen = uint64(msg.Controllen) return n, nil }) @@ -429,14 +451,38 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags n, err = dst.CopyOutFrom(t, recvmsgToBlocks) } } - - // We don't allow control messages. - msgFlags &^= linux.MSG_CTRUNC + if err != nil { + return 0, 0, nil, 0, socket.ControlMessages{}, syserr.FromError(err) + } if senderRequested { senderAddr = socket.UnmarshalSockAddr(s.family, senderAddrBuf) } - return int(n), msgFlags, senderAddr, uint32(len(senderAddrBuf)), socket.ControlMessages{}, syserr.FromError(err) + + unixControlMessages, err := unix.ParseSocketControlMessage(controlBuf[:controlLen]) + if err != nil { + return 0, 0, nil, 0, socket.ControlMessages{}, syserr.FromError(err) + } + + controlMessages := socket.ControlMessages{} + for _, unixCmsg := range unixControlMessages { + switch unixCmsg.Header.Level { + case syscall.SOL_IP: + switch unixCmsg.Header.Type { + case syscall.IP_TOS: + controlMessages.IP.HasTOS = true + binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageTOS], usermem.ByteOrder, &controlMessages.IP.TOS) + } + case syscall.SOL_IPV6: + switch unixCmsg.Header.Type { + case syscall.IPV6_TCLASS: + controlMessages.IP.HasTClass = true + binary.Unmarshal(unixCmsg.Data[:linux.SizeOfControlMessageTClass], usermem.ByteOrder, &controlMessages.IP.TClass) + } + } + } + + return int(n), msgFlags, senderAddr, uint32(len(senderAddrBuf)), controlMessages, nil } // SendMsg implements socket.Socket.SendMsg. @@ -446,6 +492,14 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return 0, syserr.ErrInvalidArgument } + space := uint64(control.CmsgsSpace(t, controlMessages)) + if space > maxControlLen { + space = maxControlLen + } + controlBuf := make([]byte, 0, space) + // PackControlMessages will append up to space bytes to controlBuf. + controlBuf = control.PackControlMessages(t, controlMessages, controlBuf) + sendmsgFromBlocks := safemem.WriterFunc(func(srcs safemem.BlockSeq) (uint64, error) { // Refuse to do anything if any part of src.Addrs was unusable. if uint64(src.NumBytes()) != srcs.NumBytes() { @@ -458,7 +512,7 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] // We always do a non-blocking send*(). sysflags := flags | syscall.MSG_DONTWAIT - if srcs.NumBlocks() == 1 { + if srcs.NumBlocks() == 1 && len(controlBuf) == 0 { // Skip allocating []syscall.Iovec. src := srcs.Head() n, _, errno := syscall.Syscall6(syscall.SYS_SENDTO, uintptr(s.fd), src.Addr(), uintptr(src.Len()), uintptr(sysflags), uintptr(firstBytePtr(to)), uintptr(len(to))) @@ -477,6 +531,10 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] msg.Name = &to[0] msg.Namelen = uint32(len(to)) } + if len(controlBuf) != 0 { + msg.Control = &controlBuf[0] + msg.Controllen = uint64(len(controlBuf)) + } return sendmsg(s.fd, &msg, sysflags) }) diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index d92399efd..140851c17 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -151,6 +151,8 @@ var Metrics = tcpip.Stats{ PassiveConnectionOpenings: mustCreateMetric("/netstack/tcp/passive_connection_openings", "Number of connections opened successfully via Listen."), CurrentEstablished: mustCreateMetric("/netstack/tcp/current_established", "Number of connections in either ESTABLISHED or CLOSE-WAIT state now."), EstablishedResets: mustCreateMetric("/netstack/tcp/established_resets", "Number of times TCP connections have made a direct transition to the CLOSED state from either the ESTABLISHED state or the CLOSE-WAIT state"), + EstablishedClosed: mustCreateMetric("/netstack/tcp/established_closed", "number of times established TCP connections made a transition to CLOSED state."), + EstablishedTimedout: mustCreateMetric("/netstack/tcp/established_timedout", "Number of times an established connection was reset because of keep-alive time out."), ListenOverflowSynDrop: mustCreateMetric("/netstack/tcp/listen_overflow_syn_drop", "Number of times the listen queue overflowed and a SYN was dropped."), ListenOverflowAckDrop: mustCreateMetric("/netstack/tcp/listen_overflow_ack_drop", "Number of times the listen queue overflowed and the final ACK in the handshake was dropped."), ListenOverflowSynCookieSent: mustCreateMetric("/netstack/tcp/listen_overflow_syn_cookie_sent", "Number of times a SYN cookie was sent."), @@ -324,7 +326,7 @@ func AddressAndFamily(sfamily int, addr []byte, strict bool) (tcpip.FullAddress, } family := usermem.ByteOrder.Uint16(addr) - if family != uint16(sfamily) && (!strict && family != linux.AF_UNSPEC) { + if family != uint16(sfamily) && (strict || family != linux.AF_UNSPEC) { return tcpip.FullAddress{}, family, syserr.ErrAddressFamilyNotSupported } @@ -1125,6 +1127,18 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa return int32(time.Duration(v) / time.Second), nil + case linux.TCP_USER_TIMEOUT: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + var v tcpip.TCPUserTimeoutOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return int32(time.Duration(v) / time.Millisecond), nil + case linux.TCP_INFO: var v tcpip.TCPInfoOption if err := ep.GetSockOpt(&v); err != nil { @@ -1561,6 +1575,17 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.KeepaliveIntervalOption(time.Second * time.Duration(v)))) + case linux.TCP_USER_TIMEOUT: + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + + v := int32(usermem.ByteOrder.Uint32(optVal)) + if v < 0 { + return syserr.ErrInvalidArgument + } + return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPUserTimeoutOption(time.Millisecond * time.Duration(v)))) + case linux.TCP_CONGESTION: v := tcpip.CongestionControlOption(optVal) if err := ep.SetSockOpt(v); err != nil { diff --git a/pkg/sentry/socket/rpcinet/syscall_rpc.proto b/pkg/sentry/socket/rpcinet/syscall_rpc.proto index 9586f5923..b677e9eb3 100644 --- a/pkg/sentry/socket/rpcinet/syscall_rpc.proto +++ b/pkg/sentry/socket/rpcinet/syscall_rpc.proto @@ -3,7 +3,6 @@ syntax = "proto3"; // package syscall_rpc is a set of networking related system calls that can be // forwarded to a socket gofer. // -// TODO(b/77963526): Document individual RPCs. package syscall_rpc; message SendmsgRequest { diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go index 8c250c325..2389a9cdb 100644 --- a/pkg/sentry/socket/socket.go +++ b/pkg/sentry/socket/socket.go @@ -43,6 +43,11 @@ type ControlMessages struct { IP tcpip.ControlMessages } +// Release releases Unix domain socket credentials and rights. +func (c *ControlMessages) Release() { + c.Unix.Release() +} + // Socket is the interface containing socket syscalls used by the syscall layer // to redirect them to the appropriate implementation. type Socket interface { diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go index 1aaae8487..885758054 100644 --- a/pkg/sentry/socket/unix/unix.go +++ b/pkg/sentry/socket/unix/unix.go @@ -118,6 +118,9 @@ func (s *SocketOperations) Endpoint() transport.Endpoint { func extractPath(sockaddr []byte) (string, *syserr.Error) { addr, _, err := netstack.AddressAndFamily(linux.AF_UNIX, sockaddr, true /* strict */) if err != nil { + if err == syserr.ErrAddressFamilyNotSupported { + err = syserr.ErrInvalidArgument + } return "", err } diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD index 72ebf766d..d46421199 100644 --- a/pkg/sentry/strace/BUILD +++ b/pkg/sentry/strace/BUILD @@ -14,6 +14,7 @@ go_library( "open.go", "poll.go", "ptrace.go", + "select.go", "signal.go", "socket.go", "strace.go", diff --git a/pkg/sentry/strace/linux64.go b/pkg/sentry/strace/linux64.go index 5d57b75af..e603f858f 100644 --- a/pkg/sentry/strace/linux64.go +++ b/pkg/sentry/strace/linux64.go @@ -40,7 +40,7 @@ var linuxAMD64 = SyscallMap{ 20: makeSyscallInfo("writev", FD, WriteIOVec, Hex), 21: makeSyscallInfo("access", Path, Oct), 22: makeSyscallInfo("pipe", PipeFDs), - 23: makeSyscallInfo("select", Hex, Hex, Hex, Hex, Timeval), + 23: makeSyscallInfo("select", Hex, SelectFDSet, SelectFDSet, SelectFDSet, Timeval), 24: makeSyscallInfo("sched_yield"), 25: makeSyscallInfo("mremap", Hex, Hex, Hex, Hex, Hex), 26: makeSyscallInfo("msync", Hex, Hex, Hex), @@ -287,7 +287,7 @@ var linuxAMD64 = SyscallMap{ 267: makeSyscallInfo("readlinkat", FD, Path, ReadBuffer, Hex), 268: makeSyscallInfo("fchmodat", FD, Path, Mode), 269: makeSyscallInfo("faccessat", FD, Path, Oct, Hex), - 270: makeSyscallInfo("pselect6", Hex, Hex, Hex, Hex, Hex, Hex), + 270: makeSyscallInfo("pselect6", Hex, SelectFDSet, SelectFDSet, SelectFDSet, Timespec, SigSet), 271: makeSyscallInfo("ppoll", PollFDs, Hex, Timespec, SigSet, Hex), 272: makeSyscallInfo("unshare", CloneFlags), 273: makeSyscallInfo("set_robust_list", Hex, Hex), @@ -335,5 +335,33 @@ var linuxAMD64 = SyscallMap{ 315: makeSyscallInfo("sched_getattr", Hex, Hex, Hex), 316: makeSyscallInfo("renameat2", FD, Path, Hex, Path, Hex), 317: makeSyscallInfo("seccomp", Hex, Hex, Hex), + 318: makeSyscallInfo("getrandom", Hex, Hex, Hex), + 319: makeSyscallInfo("memfd_create", Path, Hex), // Not quite a path, but close. + 320: makeSyscallInfo("kexec_file_load", FD, FD, Hex, Hex, Hex), + 321: makeSyscallInfo("bpf", Hex, Hex, Hex), + 322: makeSyscallInfo("execveat", FD, Path, ExecveStringVector, ExecveStringVector, Hex), + 323: makeSyscallInfo("userfaultfd", Hex), + 324: makeSyscallInfo("membarrier", Hex, Hex), + 325: makeSyscallInfo("mlock2", Hex, Hex, Hex), + 326: makeSyscallInfo("copy_file_range", FD, Hex, FD, Hex, Hex, Hex), + 327: makeSyscallInfo("preadv2", FD, ReadIOVec, Hex, Hex, Hex), + 328: makeSyscallInfo("pwritev2", FD, WriteIOVec, Hex, Hex, Hex), + 329: makeSyscallInfo("pkey_mprotect", Hex, Hex, Hex, Hex), + 330: makeSyscallInfo("pkey_alloc", Hex, Hex), + 331: makeSyscallInfo("pkey_free", Hex), 332: makeSyscallInfo("statx", FD, Path, Hex, Hex, Hex), + 333: makeSyscallInfo("io_pgetevents", Hex, Hex, Hex, Hex, Timespec, SigSet), + 334: makeSyscallInfo("rseq", Hex, Hex, Hex, Hex), + 424: makeSyscallInfo("pidfd_send_signal", FD, Signal, Hex, Hex), + 425: makeSyscallInfo("io_uring_setup", Hex, Hex), + 426: makeSyscallInfo("io_uring_enter", FD, Hex, Hex, Hex, SigSet, Hex), + 427: makeSyscallInfo("io_uring_register", FD, Hex, Hex, Hex), + 428: makeSyscallInfo("open_tree", FD, Path, Hex), + 429: makeSyscallInfo("move_mount", FD, Path, FD, Path, Hex), + 430: makeSyscallInfo("fsopen", Path, Hex), // Not quite a path, but close. + 431: makeSyscallInfo("fsconfig", FD, Hex, Hex, Hex, Hex), + 432: makeSyscallInfo("fsmount", FD, Hex, Hex), + 433: makeSyscallInfo("fspick", FD, Path, Hex), + 434: makeSyscallInfo("pidfd_open", Hex, Hex), + 435: makeSyscallInfo("clone3", Hex, Hex), } diff --git a/pkg/sentry/strace/select.go b/pkg/sentry/strace/select.go new file mode 100644 index 000000000..c77d418e6 --- /dev/null +++ b/pkg/sentry/strace/select.go @@ -0,0 +1,56 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package strace + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" + "gvisor.dev/gvisor/pkg/sentry/usermem" +) + +func fdsFromSet(t *kernel.Task, set []byte) []int { + var fds []int + // Append n if the n-th bit is 1. + for i, v := range set { + for j := 0; j < 8; j++ { + if (v>>j)&1 == 1 { + fds = append(fds, i*8+j) + } + } + } + return fds +} + +func fdSet(t *kernel.Task, nfds int, addr usermem.Addr) string { + if nfds < 0 { + return fmt.Sprintf("%#x (negative nfds)", addr) + } + if addr == 0 { + return "null" + } + + // Calculate the size of the fd set (one bit per fd). + nBytes := (nfds + 7) / 8 + nBitsInLastPartialByte := nfds % 8 + + set, err := linux.CopyInFDSet(t, addr, nBytes, nBitsInLastPartialByte) + if err != nil { + return fmt.Sprintf("%#x (error decoding fdset: %s)", addr, err) + } + + return fmt.Sprintf("%#x %v", addr, fdsFromSet(t, set)) +} diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index 94334f6d2..51f2efb39 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -208,6 +208,15 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) i += linux.SizeOfControlMessageHeader width := t.Arch().Width() length := int(h.Length) - linux.SizeOfControlMessageHeader + if length < 0 { + strs = append(strs, fmt.Sprintf( + "{level=%s, type=%s, length=%d, content too short}", + level, + typ, + h.Length, + )) + break + } if skipData { strs = append(strs, fmt.Sprintf("{level=%s, type=%s, length=%d}", level, typ, h.Length)) diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index 311389547..629c1f308 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -439,6 +439,8 @@ func (i *SyscallInfo) pre(t *kernel.Task, args arch.SyscallArguments, maximumBlo output = append(output, capData(t, args[arg-1].Pointer(), args[arg].Pointer())) case PollFDs: output = append(output, pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), false)) + case SelectFDSet: + output = append(output, fdSet(t, int(args[0].Int()), args[arg].Pointer())) case Oct: output = append(output, "0o"+strconv.FormatUint(args[arg].Uint64(), 8)) case Hex: diff --git a/pkg/sentry/strace/syscalls.go b/pkg/sentry/strace/syscalls.go index 3c389d375..e5d486c4e 100644 --- a/pkg/sentry/strace/syscalls.go +++ b/pkg/sentry/strace/syscalls.go @@ -206,6 +206,10 @@ const ( // PollFDs is an array of struct pollfd. The number of entries in the // array is in the next argument. PollFDs + + // SelectFDSet is an fd_set argument in select(2)/pselect(2). The number of + // fds represented must be the first argument. + SelectFDSet ) // defaultFormat is the syscall argument format to use if the actual format is diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index 4c0bf96e4..a76975cee 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -30,6 +30,7 @@ go_library( "sys_random.go", "sys_read.go", "sys_rlimit.go", + "sys_rseq.go", "sys_rusage.go", "sys_sched.go", "sys_seccomp.go", @@ -49,6 +50,7 @@ go_library( "sys_tls.go", "sys_utsname.go", "sys_write.go", + "sys_xattr.go", "timespec.go", ], importpath = "gvisor.dev/gvisor/pkg/sentry/syscalls/linux", diff --git a/pkg/sentry/syscalls/linux/linux64_amd64.go b/pkg/sentry/syscalls/linux/linux64_amd64.go index 81e4f93a6..479c5f6ff 100644 --- a/pkg/sentry/syscalls/linux/linux64_amd64.go +++ b/pkg/sentry/syscalls/linux/linux64_amd64.go @@ -228,10 +228,10 @@ var AMD64 = &kernel.SyscallTable{ 185: syscalls.Error("security", syserror.ENOSYS, "Not implemented in Linux.", nil), 186: syscalls.Supported("gettid", Gettid), 187: syscalls.Supported("readahead", Readahead), - 188: syscalls.Error("setxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), + 188: syscalls.PartiallySupported("setxattr", Setxattr, "Only supported for tmpfs.", nil), 189: syscalls.Error("lsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 190: syscalls.Error("fsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), - 191: syscalls.ErrorWithEvent("getxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), + 191: syscalls.PartiallySupported("getxattr", Getxattr, "Only supported for tmpfs.", nil), 192: syscalls.ErrorWithEvent("lgetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 193: syscalls.ErrorWithEvent("fgetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 194: syscalls.ErrorWithEvent("listxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), @@ -260,7 +260,7 @@ var AMD64 = &kernel.SyscallTable{ 217: syscalls.Supported("getdents64", Getdents64), 218: syscalls.Supported("set_tid_address", SetTidAddress), 219: syscalls.Supported("restart_syscall", RestartSyscall), - 220: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), // TODO(b/29354920) + 220: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), 221: syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil), 222: syscalls.Supported("timer_create", TimerCreate), 223: syscalls.Supported("timer_settime", TimerSettime), @@ -367,11 +367,31 @@ var AMD64 = &kernel.SyscallTable{ 324: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) 325: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), - // Syscalls after 325 are "backports" from versions of Linux after 4.4. + // Syscalls implemented after 325 are "backports" from versions + // of Linux after 4.4. 326: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), 327: syscalls.Supported("preadv2", Preadv2), 328: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), + 329: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), + 330: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), + 331: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), 332: syscalls.Supported("statx", Statx), + 333: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), + 334: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), + + // Linux skips ahead to syscall 424 to sync numbers between arches. + 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), + 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), + 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), + 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), + 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), + 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), + 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), + 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), + 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), + 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), + 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), + 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), }, Emulate: map[usermem.Addr]uintptr{ diff --git a/pkg/sentry/syscalls/linux/linux64_arm64.go b/pkg/sentry/syscalls/linux/linux64_arm64.go index a809115e0..d3f61f5e8 100644 --- a/pkg/sentry/syscalls/linux/linux64_arm64.go +++ b/pkg/sentry/syscalls/linux/linux64_arm64.go @@ -41,10 +41,10 @@ var ARM64 = &kernel.SyscallTable{ 2: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), 3: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), 4: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 5: syscalls.Error("setxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), + 5: syscalls.PartiallySupported("setxattr", Setxattr, "Only supported for tmpfs.", nil), 6: syscalls.Error("lsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 7: syscalls.Error("fsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), - 8: syscalls.ErrorWithEvent("getxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), + 8: syscalls.PartiallySupported("getxattr", Getxattr, "Only supported for tmpfs.", nil), 9: syscalls.ErrorWithEvent("lgetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 10: syscalls.ErrorWithEvent("fgetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 11: syscalls.ErrorWithEvent("listxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), @@ -224,7 +224,7 @@ var ARM64 = &kernel.SyscallTable{ 189: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) 190: syscalls.Supported("semget", Semget), 191: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), - 192: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), // TODO(b/29354920) + 192: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), 193: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), 194: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), 195: syscalls.PartiallySupported("shmctl", Shmctl, "Options SHM_LOCK, SHM_UNLOCK are not supported.", nil), @@ -295,14 +295,33 @@ var ARM64 = &kernel.SyscallTable{ 278: syscalls.Supported("getrandom", GetRandom), 279: syscalls.Supported("memfd_create", MemfdCreate), 280: syscalls.CapError("bpf", linux.CAP_SYS_ADMIN, "", nil), - 281: syscalls.ErrorWithEvent("execveat", syserror.ENOSYS, "", []string{"gvisor.dev/issue/265"}), // TODO(b/118901836) + 281: syscalls.Supported("execveat", Execveat), 282: syscalls.ErrorWithEvent("userfaultfd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/266"}), // TODO(b/118906345) 283: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) 284: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), 285: syscalls.ErrorWithEvent("copy_file_range", syserror.ENOSYS, "", nil), 286: syscalls.Supported("preadv2", Preadv2), 287: syscalls.PartiallySupported("pwritev2", Pwritev2, "Flag RWF_HIPRI is not supported.", nil), + 288: syscalls.ErrorWithEvent("pkey_mprotect", syserror.ENOSYS, "", nil), + 289: syscalls.ErrorWithEvent("pkey_alloc", syserror.ENOSYS, "", nil), + 290: syscalls.ErrorWithEvent("pkey_free", syserror.ENOSYS, "", nil), 291: syscalls.Supported("statx", Statx), + 292: syscalls.ErrorWithEvent("io_pgetevents", syserror.ENOSYS, "", nil), + 293: syscalls.PartiallySupported("rseq", RSeq, "Not supported on all platforms.", nil), + + // Linux skips ahead to syscall 424 to sync numbers between arches. + 424: syscalls.ErrorWithEvent("pidfd_send_signal", syserror.ENOSYS, "", nil), + 425: syscalls.ErrorWithEvent("io_uring_setup", syserror.ENOSYS, "", nil), + 426: syscalls.ErrorWithEvent("io_uring_enter", syserror.ENOSYS, "", nil), + 427: syscalls.ErrorWithEvent("io_uring_register", syserror.ENOSYS, "", nil), + 428: syscalls.ErrorWithEvent("open_tree", syserror.ENOSYS, "", nil), + 429: syscalls.ErrorWithEvent("move_mount", syserror.ENOSYS, "", nil), + 430: syscalls.ErrorWithEvent("fsopen", syserror.ENOSYS, "", nil), + 431: syscalls.ErrorWithEvent("fsconfig", syserror.ENOSYS, "", nil), + 432: syscalls.ErrorWithEvent("fsmount", syserror.ENOSYS, "", nil), + 433: syscalls.ErrorWithEvent("fspick", syserror.ENOSYS, "", nil), + 434: syscalls.ErrorWithEvent("pidfd_open", syserror.ENOSYS, "", nil), + 435: syscalls.ErrorWithEvent("clone3", syserror.ENOSYS, "", nil), }, Emulate: map[usermem.Addr]uintptr{}, diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index 167c2b60b..9bc2445a5 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -171,6 +171,9 @@ func openAt(t *kernel.Task, dirFD int32, addr usermem.Addr, flags uint) (fd uint } } + // Truncate is called when O_TRUNC is specified for any kind of + // existing Dirent. Behavior is delegated to the entry's Truncate + // implementation. if flags&linux.O_TRUNC != 0 { if err := d.Inode.Truncate(t, d, 0); err != nil { return err @@ -397,7 +400,9 @@ func createAt(t *kernel.Task, dirFD int32, addr usermem.Addr, flags uint, mode l return err } - // Should we truncate the file? + // Truncate is called when O_TRUNC is specified for any kind of + // existing Dirent. Behavior is delegated to the entry's Truncate + // implementation. if flags&linux.O_TRUNC != 0 { if err := found.Inode.Truncate(t, found, 0); err != nil { return err @@ -835,25 +840,42 @@ func Dup3(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(newfd), nil, nil } -func fGetOwn(t *kernel.Task, file *fs.File) int32 { +func fGetOwnEx(t *kernel.Task, file *fs.File) linux.FOwnerEx { ma := file.Async(nil) if ma == nil { - return 0 + return linux.FOwnerEx{} } a := ma.(*fasync.FileAsync) ot, otg, opg := a.Owner() switch { case ot != nil: - return int32(t.PIDNamespace().IDOfTask(ot)) + return linux.FOwnerEx{ + Type: linux.F_OWNER_TID, + PID: int32(t.PIDNamespace().IDOfTask(ot)), + } case otg != nil: - return int32(t.PIDNamespace().IDOfThreadGroup(otg)) + return linux.FOwnerEx{ + Type: linux.F_OWNER_PID, + PID: int32(t.PIDNamespace().IDOfThreadGroup(otg)), + } case opg != nil: - return int32(-t.PIDNamespace().IDOfProcessGroup(opg)) + return linux.FOwnerEx{ + Type: linux.F_OWNER_PGRP, + PID: int32(t.PIDNamespace().IDOfProcessGroup(opg)), + } default: - return 0 + return linux.FOwnerEx{} } } +func fGetOwn(t *kernel.Task, file *fs.File) int32 { + owner := fGetOwnEx(t, file) + if owner.Type == linux.F_OWNER_PGRP { + return -owner.PID + } + return owner.PID +} + // fSetOwn sets the file's owner with the semantics of F_SETOWN in Linux. // // If who is positive, it represents a PID. If negative, it represents a PGID. @@ -896,11 +918,13 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall t.FDTable().SetFlags(fd, kernel.FDFlags{ CloseOnExec: flags&linux.FD_CLOEXEC != 0, }) + return 0, nil, nil case linux.F_GETFL: return uintptr(file.Flags().ToLinux()), nil, nil case linux.F_SETFL: flags := uint(args[2].Uint()) file.SetFlags(linuxToFlags(flags).Settable()) + return 0, nil, nil case linux.F_SETLK, linux.F_SETLKW: // In Linux the file system can choose to provide lock operations for an inode. // Normally pipe and socket types lack lock operations. We diverge and use a heavy @@ -1003,6 +1027,44 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.F_SETOWN: fSetOwn(t, file, args[2].Int()) return 0, nil, nil + case linux.F_GETOWN_EX: + addr := args[2].Pointer() + owner := fGetOwnEx(t, file) + _, err := t.CopyOut(addr, &owner) + return 0, nil, err + case linux.F_SETOWN_EX: + addr := args[2].Pointer() + var owner linux.FOwnerEx + n, err := t.CopyIn(addr, &owner) + if err != nil { + return 0, nil, err + } + a := file.Async(fasync.New).(*fasync.FileAsync) + switch owner.Type { + case linux.F_OWNER_TID: + task := t.PIDNamespace().TaskWithID(kernel.ThreadID(owner.PID)) + if task == nil { + return 0, nil, syserror.ESRCH + } + a.SetOwnerTask(t, task) + return uintptr(n), nil, nil + case linux.F_OWNER_PID: + tg := t.PIDNamespace().ThreadGroupWithID(kernel.ThreadID(owner.PID)) + if tg == nil { + return 0, nil, syserror.ESRCH + } + a.SetOwnerThreadGroup(t, tg) + return uintptr(n), nil, nil + case linux.F_OWNER_PGRP: + pg := t.PIDNamespace().ProcessGroupWithID(kernel.ProcessGroupID(owner.PID)) + if pg == nil { + return 0, nil, syserror.ESRCH + } + a.SetOwnerProcessGroup(t, pg) + return uintptr(n), nil, nil + default: + return 0, nil, syserror.EINVAL + } case linux.F_GET_SEALS: val, err := tmpfs.GetSeals(file.Dirent.Inode) return uintptr(val), nil, err @@ -1030,7 +1092,6 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Everything else is not yet supported. return 0, nil, syserror.EINVAL } - return 0, nil, nil } const ( @@ -1484,6 +1545,8 @@ func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if fs.IsDir(d.Inode.StableAttr) { return syserror.EISDIR } + // In contrast to open(O_TRUNC), truncate(2) is only valid for file + // types. if !fs.IsFile(d.Inode.StableAttr) { return syserror.EINVAL } @@ -1522,7 +1585,8 @@ func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, syserror.EINVAL } - // Note that this is different from truncate(2) above, where a + // In contrast to open(O_TRUNC), truncate(2) is only valid for file + // types. Note that this is different from truncate(2) above, where a // directory returns EISDIR. if !fs.IsFile(file.Dirent.Inode.StableAttr) { return 0, nil, syserror.EINVAL diff --git a/pkg/sentry/syscalls/linux/sys_futex.go b/pkg/sentry/syscalls/linux/sys_futex.go index b9bd25464..bde17a767 100644 --- a/pkg/sentry/syscalls/linux/sys_futex.go +++ b/pkg/sentry/syscalls/linux/sys_futex.go @@ -226,6 +226,11 @@ func Futex(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if mask == 0 { return 0, nil, syserror.EINVAL } + if val <= 0 { + // The Linux kernel wakes one waiter even if val is + // non-positive. + val = 1 + } n, err := t.Futex().Wake(t, addr, private, mask, val) return uintptr(n), nil, err @@ -242,6 +247,11 @@ func Futex(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.FUTEX_WAKE_OP: op := uint32(val3) + if val <= 0 { + // The Linux kernel wakes one waiter even if val is + // non-positive. + val = 1 + } n, err := t.Futex().WakeOp(t, addr, naddr, private, val, nreq, op) return uintptr(n), nil, err diff --git a/pkg/sentry/syscalls/linux/sys_poll.go b/pkg/sentry/syscalls/linux/sys_poll.go index 7a13beac2..2b2df989a 100644 --- a/pkg/sentry/syscalls/linux/sys_poll.go +++ b/pkg/sentry/syscalls/linux/sys_poll.go @@ -197,53 +197,51 @@ func doPoll(t *kernel.Task, addr usermem.Addr, nfds uint, timeout time.Duration) 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 } - // Capture all the provided input vectors. - // - // N.B. This only works on little-endian architectures. - byteCount := (nfds + 7) / 8 - - bitsInLastPartialByte := uint(nfds % 8) - r := make([]byte, byteCount) - w := make([]byte, byteCount) - e := make([]byte, byteCount) + // Calculate the size of the fd sets (one bit per fd). + nBytes := (nfds + 7) / 8 + nBitsInLastPartialByte := nfds % 8 - if readFDs != 0 { - if _, err := t.CopyIn(readFDs, &r); err != nil { - return 0, err - } - // Mask out bits above nfds. - if bitsInLastPartialByte != 0 { - r[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte - } + // Capture all the provided input vectors. + r, err := CopyInFDSet(t, readFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err } - - if writeFDs != 0 { - if _, err := t.CopyIn(writeFDs, &w); err != nil { - return 0, err - } - if bitsInLastPartialByte != 0 { - w[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte - } + w, err := CopyInFDSet(t, writeFDs, nBytes, nBitsInLastPartialByte) + if err != nil { + return 0, err } - - if exceptFDs != 0 { - if _, err := t.CopyIn(exceptFDs, &e); err != nil { - return 0, err - } - if bitsInLastPartialByte != 0 { - e[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte - } + 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 < byteCount; i++ { + for i := 0; i < nBytes; i++ { v := r[i] | w[i] | e[i] for v != 0 { v &= (v - 1) @@ -254,7 +252,7 @@ func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Add // Build the PollFD array. pfd := make([]linux.PollFD, 0, fdCount) var fd int32 - for i := 0; i < byteCount; i++ { + for i := 0; i < nBytes; i++ { rV, wV, eV := r[i], w[i], e[i] v := rV | wV | eV m := byte(1) @@ -295,8 +293,7 @@ func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Add } // Do the syscall, then count the number of bits set. - _, _, err := pollBlock(t, pfd, timeout) - if err != nil { + if _, _, err = pollBlock(t, pfd, timeout); err != nil { return 0, syserror.ConvertIntr(err, syserror.EINTR) } diff --git a/pkg/sentry/syscalls/linux/sys_rseq.go b/pkg/sentry/syscalls/linux/sys_rseq.go new file mode 100644 index 000000000..90db10ea6 --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_rseq.go @@ -0,0 +1,48 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// RSeq implements syscall rseq(2). +func RSeq(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + addr := args[0].Pointer() + length := args[1].Uint() + flags := args[2].Int() + signature := args[3].Uint() + + if !t.RSeqAvailable() { + // Event for applications that want rseq on a configuration + // that doesn't support them. + t.Kernel().EmitUnimplementedEvent(t) + return 0, nil, syserror.ENOSYS + } + + switch flags { + case 0: + // Register. + return 0, nil, t.SetRSeq(addr, length, signature) + case linux.RSEQ_FLAG_UNREGISTER: + return 0, nil, t.ClearRSeq(addr, length, signature) + default: + // Unknown flag. + return 0, nil, syserror.EINVAL + } +} diff --git a/pkg/sentry/syscalls/linux/sys_shm.go b/pkg/sentry/syscalls/linux/sys_shm.go index d57ffb3a1..4a8bc24a2 100644 --- a/pkg/sentry/syscalls/linux/sys_shm.go +++ b/pkg/sentry/syscalls/linux/sys_shm.go @@ -39,10 +39,13 @@ func Shmget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if err != nil { return 0, nil, err } + defer segment.DecRef() return uintptr(segment.ID), nil, nil } // findSegment retrives a shm segment by the given id. +// +// findSegment returns a reference on Shm. func findSegment(t *kernel.Task, id shm.ID) (*shm.Shm, error) { r := t.IPCNamespace().ShmRegistry() segment := r.FindByID(id) @@ -63,6 +66,7 @@ func Shmat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if err != nil { return 0, nil, syserror.EINVAL } + defer segment.DecRef() opts, err := segment.ConfigureAttach(t, addr, shm.AttachOpts{ Execute: flag&linux.SHM_EXEC == linux.SHM_EXEC, @@ -72,7 +76,6 @@ func Shmat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if err != nil { return 0, nil, err } - defer segment.DecRef() addr, err = t.MemoryManager().MMap(t, opts) return uintptr(addr), nil, err } @@ -105,6 +108,7 @@ func Shmctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if err != nil { return 0, nil, syserror.EINVAL } + defer segment.DecRef() stat, err := segment.IPCStat(t) if err == nil { @@ -128,6 +132,7 @@ func Shmctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if err != nil { return 0, nil, syserror.EINVAL } + defer segment.DecRef() switch cmd { case linux.IPC_SET: diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index ab1001f16..4b5aafcc0 100644 --- a/pkg/sentry/syscalls/linux/sys_socket.go +++ b/pkg/sentry/syscalls/linux/sys_socket.go @@ -764,7 +764,7 @@ func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags i } if !cms.Unix.Empty() { mflags |= linux.MSG_CTRUNC - cms.Unix.Release() + cms.Release() } if int(msg.Flags) != mflags { @@ -784,24 +784,16 @@ func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags i if e != nil { return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) } - defer cms.Unix.Release() + defer cms.Release() controlData := make([]byte, 0, msg.ControlLen) + controlData = control.PackControlMessages(t, cms, controlData) if cr, ok := s.(transport.Credentialer); ok && cr.Passcred() { creds, _ := cms.Unix.Credentials.(control.SCMCredentials) controlData, mflags = control.PackCredentials(t, creds, controlData, mflags) } - if cms.IP.HasTimestamp { - controlData = control.PackTimestamp(t, cms.IP.Timestamp, controlData) - } - - if cms.IP.HasInq { - // In Linux, TCP_CM_INQ is added after SO_TIMESTAMP. - controlData = control.PackInq(t, cms.IP.Inq, controlData) - } - if cms.Unix.Rights != nil { controlData, mflags = control.PackRights(t, cms.Unix.Rights.(control.SCMRights), flags&linux.MSG_CMSG_CLOEXEC != 0, controlData, mflags) } @@ -877,7 +869,7 @@ func recvFrom(t *kernel.Task, fd int32, bufPtr usermem.Addr, bufLen uint64, flag } n, _, sender, senderLen, cm, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, nameLenPtr != 0, 0) - cm.Unix.Release() + cm.Release() if e != nil { return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS) } @@ -1060,7 +1052,7 @@ func sendSingleMsg(t *kernel.Task, s socket.Socket, file *fs.File, msgPtr userme } // Call the syscall implementation. - n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, socket.ControlMessages{Unix: controlMessages}) + n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, controlMessages) err = handleIOError(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendmsg", file) if err != nil { controlMessages.Release() diff --git a/pkg/sentry/syscalls/linux/sys_xattr.go b/pkg/sentry/syscalls/linux/sys_xattr.go new file mode 100644 index 000000000..97d9a65ea --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_xattr.go @@ -0,0 +1,169 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +import ( + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Getxattr implements linux syscall getxattr(2). +func Getxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + + path, dirPath, err := copyInPath(t, pathAddr, false /* allowEmpty */) + if err != nil { + return 0, nil, err + } + + valueLen := 0 + err = fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + value, err := getxattr(t, d, dirPath, nameAddr) + if err != nil { + return err + } + + valueLen = len(value) + if size == 0 { + return nil + } + if size > linux.XATTR_SIZE_MAX { + size = linux.XATTR_SIZE_MAX + } + if valueLen > int(size) { + return syserror.ERANGE + } + + _, err = t.CopyOutBytes(valueAddr, []byte(value)) + return err + }) + if err != nil { + return 0, nil, err + } + return uintptr(valueLen), nil, nil +} + +// getxattr implements getxattr from the given *fs.Dirent. +func getxattr(t *kernel.Task, d *fs.Dirent, dirPath bool, nameAddr usermem.Addr) (string, error) { + if dirPath && !fs.IsDir(d.Inode.StableAttr) { + return "", syserror.ENOTDIR + } + + if err := checkXattrPermissions(t, d.Inode, fs.PermMask{Read: true}); err != nil { + return "", err + } + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return "", err + } + + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return "", syserror.EOPNOTSUPP + } + + return d.Inode.Getxattr(name) +} + +// Setxattr implements linux syscall setxattr(2). +func Setxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + valueAddr := args[2].Pointer() + size := args[3].SizeT() + flags := args[4].Uint() + + path, dirPath, err := copyInPath(t, pathAddr, false /* allowEmpty */) + if err != nil { + return 0, nil, err + } + + if flags&^(linux.XATTR_CREATE|linux.XATTR_REPLACE) != 0 { + return 0, nil, syserror.EINVAL + } + + return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + return setxattr(t, d, dirPath, nameAddr, valueAddr, size, flags) + }) +} + +// setxattr implements setxattr from the given *fs.Dirent. +func setxattr(t *kernel.Task, d *fs.Dirent, dirPath bool, nameAddr, valueAddr usermem.Addr, size uint, flags uint32) error { + if dirPath && !fs.IsDir(d.Inode.StableAttr) { + return syserror.ENOTDIR + } + + if err := checkXattrPermissions(t, d.Inode, fs.PermMask{Write: true}); err != nil { + return err + } + + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + + if size > linux.XATTR_SIZE_MAX { + return syserror.E2BIG + } + buf := make([]byte, size) + if _, err = t.CopyInBytes(valueAddr, buf); err != nil { + return err + } + value := string(buf) + + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + + return d.Inode.Setxattr(name, value) +} + +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 checkXattrPermissions(t *kernel.Task, i *fs.Inode, perms fs.PermMask) error { + // Restrict xattrs to regular files and directories. + // + // In Linux, this restriction technically only applies to xattrs in the + // "user.*" namespace, but we don't allow any other xattr prefixes anyway. + if !fs.IsRegular(i.StableAttr) && !fs.IsDir(i.StableAttr) { + if perms.Write { + return syserror.EPERM + } + return syserror.ENODATA + } + + return i.CheckPermission(t, perms) +} diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 74a325309..4c6aa04a1 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -9,6 +9,7 @@ go_library( "context.go", "debug.go", "dentry.go", + "device.go", "file_description.go", "file_description_impl_util.go", "filesystem.go", @@ -17,9 +18,9 @@ go_library( "mount.go", "mount_unsafe.go", "options.go", + "pathname.go", "permissions.go", "resolving_path.go", - "syscalls.go", "testutil.go", "vfs.go", ], diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go index 32cf9151b..705194ebc 100644 --- a/pkg/sentry/vfs/context.go +++ b/pkg/sentry/vfs/context.go @@ -24,6 +24,9 @@ type contextID int const ( // CtxMountNamespace is a Context.Value key for a MountNamespace. CtxMountNamespace contextID = iota + + // CtxRoot is a Context.Value key for a VFS root. + CtxRoot ) // MountNamespaceFromContext returns the MountNamespace used by ctx. It does @@ -35,3 +38,13 @@ func MountNamespaceFromContext(ctx context.Context) *MountNamespace { } return nil } + +// RootFromContext returns the VFS root used by ctx. It takes a reference on +// the returned VirtualDentry. If ctx does not have a specific VFS root, +// RootFromContext returns a zero-value VirtualDentry. +func RootFromContext(ctx context.Context) VirtualDentry { + if v := ctx.Value(CtxRoot); v != nil { + return v.(VirtualDentry) + } + return VirtualDentry{} +} diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index 40f4c1d09..1bc9c4a38 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -85,12 +85,12 @@ type Dentry struct { // mounts is accessed using atomic memory operations. mounts uint32 - // mu synchronizes disowning and mounting over this Dentry. - mu sync.Mutex - // children are child Dentries. children map[string]*Dentry + // mu synchronizes disowning and mounting over this Dentry. + mu sync.Mutex + // impl is the DentryImpl associated with this Dentry. impl is immutable. // This should be the last field in Dentry. impl DentryImpl @@ -199,6 +199,18 @@ func (d *Dentry) HasChildren() bool { return len(d.children) != 0 } +// Children returns a map containing all of d's children. +func (d *Dentry) Children() map[string]*Dentry { + if !d.HasChildren() { + return nil + } + m := make(map[string]*Dentry) + for name, child := range d.children { + m[name] = child + } + return m +} + // InsertChild makes child a child of d with the given name. // // InsertChild is a mutator of d and child. @@ -222,6 +234,18 @@ func (d *Dentry) InsertChild(child *Dentry, name string) { child.name = name } +// IsAncestorOf returns true if d is an ancestor of d2; that is, d is either +// d2's parent or an ancestor of d2's parent. +func (d *Dentry) IsAncestorOf(d2 *Dentry) bool { + for d2.parent != nil { + if d2.parent == d { + return true + } + d2 = d2.parent + } + return false +} + // PrepareDeleteDentry must be called before attempting to delete the file // represented by d. If PrepareDeleteDentry succeeds, the caller must call // AbortDeleteDentry or CommitDeleteDentry depending on the deletion's outcome. @@ -271,21 +295,6 @@ func (vfs *VirtualFilesystem) CommitDeleteDentry(d *Dentry) { } } -// DeleteDentry combines PrepareDeleteDentry and CommitDeleteDentry, as -// appropriate for in-memory filesystems that don't need to ensure that some -// external state change succeeds before committing the deletion. -// -// DeleteDentry is a mutator of d and d.Parent(). -// -// Preconditions: d is a child Dentry. -func (vfs *VirtualFilesystem) DeleteDentry(mntns *MountNamespace, d *Dentry) error { - if err := vfs.PrepareDeleteDentry(mntns, d); err != nil { - return err - } - vfs.CommitDeleteDentry(d) - return nil -} - // ForceDeleteDentry causes d to become disowned. It should only be used in // cases where VFS has no ability to stop the deletion (e.g. d represents the // local state of a file on a remote filesystem on which the file has already @@ -314,7 +323,7 @@ func (vfs *VirtualFilesystem) ForceDeleteDentry(d *Dentry) { // CommitRenameExchangeDentry depending on the rename's outcome. // // Preconditions: from is a child Dentry. If to is not nil, it must be a child -// Dentry from the same Filesystem. +// Dentry from the same Filesystem. from != to. func (vfs *VirtualFilesystem) PrepareRenameDentry(mntns *MountNamespace, from, to *Dentry) error { if checkInvariants { if from.parent == nil { diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go new file mode 100644 index 000000000..cb672e36f --- /dev/null +++ b/pkg/sentry/vfs/device.go @@ -0,0 +1,100 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/syserror" +) + +// DeviceKind indicates whether a device is a block or character device. +type DeviceKind uint32 + +const ( + // BlockDevice indicates a block device. + BlockDevice DeviceKind = iota + + // CharDevice indicates a character device. + CharDevice +) + +// String implements fmt.Stringer.String. +func (kind DeviceKind) String() string { + switch kind { + case BlockDevice: + return "block" + case CharDevice: + return "character" + default: + return fmt.Sprintf("invalid device kind %d", kind) + } +} + +type devTuple struct { + kind DeviceKind + major uint32 + minor uint32 +} + +// A Device backs device special files. +type Device interface { + // Open returns a FileDescription representing this device. + Open(ctx context.Context, mnt *Mount, d *Dentry, opts OpenOptions) (*FileDescription, error) +} + +type registeredDevice struct { + dev Device + opts RegisterDeviceOptions +} + +// RegisterDeviceOptions contains options to +// VirtualFilesystem.RegisterDevice(). +type RegisterDeviceOptions struct { + // GroupName is the name shown for this device registration in + // /proc/devices. If GroupName is empty, this registration will not be + // shown in /proc/devices. + GroupName string +} + +// RegisterDevice registers the given Device in vfs with the given major and +// minor device numbers. +func (vfs *VirtualFilesystem) RegisterDevice(kind DeviceKind, major, minor uint32, dev Device, opts *RegisterDeviceOptions) error { + tup := devTuple{kind, major, minor} + vfs.devicesMu.Lock() + defer vfs.devicesMu.Unlock() + if existing, ok := vfs.devices[tup]; ok { + return fmt.Errorf("%s device number (%d, %d) is already registered to device type %T", kind, major, minor, existing.dev) + } + vfs.devices[tup] = ®isteredDevice{ + dev: dev, + opts: *opts, + } + return nil +} + +// OpenDeviceSpecialFile returns a FileDescription representing the given +// device. +func (vfs *VirtualFilesystem) OpenDeviceSpecialFile(ctx context.Context, mnt *Mount, d *Dentry, kind DeviceKind, major, minor uint32, opts *OpenOptions) (*FileDescription, error) { + tup := devTuple{kind, major, minor} + vfs.devicesMu.RLock() + defer vfs.devicesMu.RUnlock() + rd, ok := vfs.devices[tup] + if !ok { + return nil, syserror.ENXIO + } + return rd.dev.Open(ctx, mnt, d, *opts) +} diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 34007eb57..6afe280bc 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -20,8 +20,10 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" ) @@ -38,49 +40,58 @@ type FileDescription struct { // operations. refs int64 + // statusFlags contains status flags, "initialized by open(2) and possibly + // modified by fcntl()" - fcntl(2). statusFlags is accessed using atomic + // memory operations. + statusFlags uint32 + // vd is the filesystem location at which this FileDescription was opened. // A reference is held on vd. vd is immutable. vd VirtualDentry + opts FileDescriptionOptions + // impl is the FileDescriptionImpl associated with this Filesystem. impl is // immutable. This should be the last field in FileDescription. impl FileDescriptionImpl } -// Init must be called before first use of fd. It takes ownership of references -// on mnt and d held by the caller. -func (fd *FileDescription) Init(impl FileDescriptionImpl, mnt *Mount, d *Dentry) { +// FileDescriptionOptions contains options to FileDescription.Init(). +type FileDescriptionOptions struct { + // If AllowDirectIO is true, allow O_DIRECT to be set on the file. This is + // usually only the case if O_DIRECT would actually have an effect. + AllowDirectIO bool + + // If UseDentryMetadata is true, calls to FileDescription methods that + // interact with file and filesystem metadata (Stat, SetStat, StatFS, + // Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling + // the corresponding FilesystemImpl methods instead of the corresponding + // FileDescriptionImpl methods. + // + // UseDentryMetadata is intended for file descriptions that are implemented + // outside of individual filesystems, such as pipes, sockets, and device + // special files. FileDescriptions for which UseDentryMetadata is true may + // embed DentryMetadataFileDescriptionImpl to obtain appropriate + // implementations of FileDescriptionImpl methods that should not be + // called. + UseDentryMetadata bool +} + +// Init must be called before first use of fd. It takes references on mnt and +// d. statusFlags is the initial file description status flags, which is +// usually the full set of flags passed to open(2). +func (fd *FileDescription) Init(impl FileDescriptionImpl, statusFlags uint32, mnt *Mount, d *Dentry, opts *FileDescriptionOptions) { fd.refs = 1 + fd.statusFlags = statusFlags | linux.O_LARGEFILE fd.vd = VirtualDentry{ mount: mnt, dentry: d, } + fd.vd.IncRef() + fd.opts = *opts fd.impl = impl } -// Impl returns the FileDescriptionImpl associated with fd. -func (fd *FileDescription) Impl() FileDescriptionImpl { - return fd.impl -} - -// Mount returns the mount on which fd was opened. It does not take a reference -// on the returned Mount. -func (fd *FileDescription) Mount() *Mount { - return fd.vd.mount -} - -// Dentry returns the dentry at which fd was opened. It does not take a -// reference on the returned Dentry. -func (fd *FileDescription) Dentry() *Dentry { - return fd.vd.dentry -} - -// VirtualDentry returns the location at which fd was opened. It does not take -// a reference on the returned VirtualDentry. -func (fd *FileDescription) VirtualDentry() VirtualDentry { - return fd.vd -} - // IncRef increments fd's reference count. func (fd *FileDescription) IncRef() { atomic.AddInt64(&fd.refs, 1) @@ -112,6 +123,82 @@ func (fd *FileDescription) DecRef() { } } +// Mount returns the mount on which fd was opened. It does not take a reference +// on the returned Mount. +func (fd *FileDescription) Mount() *Mount { + return fd.vd.mount +} + +// Dentry returns the dentry at which fd was opened. It does not take a +// reference on the returned Dentry. +func (fd *FileDescription) Dentry() *Dentry { + return fd.vd.dentry +} + +// VirtualDentry returns the location at which fd was opened. It does not take +// a reference on the returned VirtualDentry. +func (fd *FileDescription) VirtualDentry() VirtualDentry { + return fd.vd +} + +// StatusFlags returns file description status flags, as for fcntl(F_GETFL). +func (fd *FileDescription) StatusFlags() uint32 { + return atomic.LoadUint32(&fd.statusFlags) +} + +// SetStatusFlags sets file description status flags, as for fcntl(F_SETFL). +func (fd *FileDescription) SetStatusFlags(ctx context.Context, creds *auth.Credentials, flags uint32) error { + // Compare Linux's fs/fcntl.c:setfl(). + oldFlags := fd.StatusFlags() + // Linux documents this check as "O_APPEND cannot be cleared if the file is + // marked as append-only and the file is open for write", which would make + // sense. However, the check as actually implemented seems to be "O_APPEND + // cannot be changed if the file is marked as append-only". + if (flags^oldFlags)&linux.O_APPEND != 0 { + stat, err := fd.Stat(ctx, StatOptions{ + // There is no mask bit for stx_attributes. + Mask: 0, + // Linux just reads inode::i_flags directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil { + return err + } + if (stat.AttributesMask&linux.STATX_ATTR_APPEND != 0) && (stat.Attributes&linux.STATX_ATTR_APPEND != 0) { + return syserror.EPERM + } + } + if (flags&linux.O_NOATIME != 0) && (oldFlags&linux.O_NOATIME == 0) { + stat, err := fd.Stat(ctx, StatOptions{ + Mask: linux.STATX_UID, + // Linux's inode_owner_or_capable() just reads inode::i_uid + // directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil { + return err + } + if stat.Mask&linux.STATX_UID == 0 { + return syserror.EPERM + } + if !CanActAsOwner(creds, auth.KUID(stat.UID)) { + return syserror.EPERM + } + } + if flags&linux.O_DIRECT != 0 && !fd.opts.AllowDirectIO { + return syserror.EINVAL + } + // TODO(jamieliu): FileDescriptionImpl.SetOAsync()? + const settableFlags = linux.O_APPEND | linux.O_ASYNC | linux.O_DIRECT | linux.O_NOATIME | linux.O_NONBLOCK + atomic.StoreUint32(&fd.statusFlags, (oldFlags&^settableFlags)|(flags&settableFlags)) + return nil +} + +// Impl returns the FileDescriptionImpl associated with fd. +func (fd *FileDescription) Impl() FileDescriptionImpl { + return fd.impl +} + // FileDescriptionImpl contains implementation details for an FileDescription. // Implementations of FileDescriptionImpl should contain their associated // FileDescription by value as their first field. @@ -120,6 +207,8 @@ func (fd *FileDescription) DecRef() { // be interpreted as IDs in the root UserNamespace (i.e. as auth.KUID and // auth.KGID respectively). // +// All methods may return errors not specified. +// // FileDescriptionImpl is analogous to Linux's struct file_operations. type FileDescriptionImpl interface { // Release is called when the associated FileDescription reaches zero @@ -131,14 +220,6 @@ type FileDescriptionImpl interface { // prevent the file descriptor from being closed. OnClose(ctx context.Context) error - // StatusFlags returns file description status flags, as for - // fcntl(F_GETFL). - StatusFlags(ctx context.Context) (uint32, error) - - // SetStatusFlags sets file description status flags, as for - // fcntl(F_SETFL). - SetStatusFlags(ctx context.Context, flags uint32) error - // Stat returns metadata for the file represented by the FileDescription. Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) @@ -156,6 +237,10 @@ type FileDescriptionImpl interface { // PRead reads from the file into dst, starting at the given offset, and // returns the number of bytes read. PRead is permitted to return partial // reads with a nil error. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, PRead returns EOPNOTSUPP. PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) // Read is similar to PRead, but does not specify an offset. @@ -165,6 +250,10 @@ type FileDescriptionImpl interface { // the number of bytes read; note that POSIX 2.9.7 "Thread Interactions // with Regular File Operations" requires that all operations that may // mutate the FileDescription offset are serialized. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, Read returns EOPNOTSUPP. Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) // PWrite writes src to the file, starting at the given offset, and returns @@ -174,6 +263,11 @@ type FileDescriptionImpl interface { // As in Linux (but not POSIX), if O_APPEND is in effect for the // FileDescription, PWrite should ignore the offset and append data to the // end of the file. + // + // Errors: + // + // - If opts.Flags specifies unsupported options, PWrite returns + // EOPNOTSUPP. PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) // Write is similar to PWrite, but does not specify an offset, which is @@ -183,6 +277,10 @@ type FileDescriptionImpl interface { // PWrite that uses a FileDescription offset, to make it possible for // remote filesystems to implement O_APPEND correctly (i.e. atomically with // respect to writers outside the scope of VFS). + // + // Errors: + // + // - If opts.Flags specifies unsupported options, Write returns EOPNOTSUPP. Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) // IterDirents invokes cb on each entry in the directory represented by the @@ -212,7 +310,21 @@ type FileDescriptionImpl interface { // Ioctl implements the ioctl(2) syscall. Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) - // TODO: extended attributes; file locking + // Listxattr returns all extended attribute names for the file. + Listxattr(ctx context.Context) ([]string, error) + + // Getxattr returns the value associated with the given extended attribute + // for the file. + Getxattr(ctx context.Context, name string) (string, error) + + // Setxattr changes the value associated with the given extended attribute + // for the file. + Setxattr(ctx context.Context, opts SetxattrOptions) error + + // Removexattr removes the given extended attribute from the file. + Removexattr(ctx context.Context, name string) error + + // TODO: file locking } // Dirent holds the information contained in struct linux_dirent64. @@ -241,3 +353,230 @@ type IterDirentsCallback interface { // called. Handle(dirent Dirent) bool } + +// OnClose is called when a file descriptor representing the FileDescription is +// closed. Returning a non-nil error should not prevent the file descriptor +// from being closed. +func (fd *FileDescription) OnClose(ctx context.Context) error { + return fd.impl.OnClose(ctx) +} + +// Stat returns metadata for the file represented by fd. +func (fd *FileDescription) Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + stat, err := fd.vd.mount.fs.impl.StatAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return stat, err + } + return fd.impl.Stat(ctx, opts) +} + +// SetStat updates metadata for the file represented by fd. +func (fd *FileDescription) SetStat(ctx context.Context, opts SetStatOptions) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.SetStatAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return err + } + return fd.impl.SetStat(ctx, opts) +} + +// StatFS returns metadata for the filesystem containing the file represented +// by fd. +func (fd *FileDescription) StatFS(ctx context.Context) (linux.Statfs, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + statfs, err := fd.vd.mount.fs.impl.StatFSAt(ctx, rp) + vfsObj.putResolvingPath(rp) + return statfs, err + } + return fd.impl.StatFS(ctx) +} + +// PRead reads from the file represented by fd into dst, starting at the given +// offset, and returns the number of bytes read. PRead is permitted to return +// partial reads with a nil error. +func (fd *FileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) { + return fd.impl.PRead(ctx, dst, offset, opts) +} + +// Read is similar to PRead, but does not specify an offset. +func (fd *FileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts ReadOptions) (int64, error) { + return fd.impl.Read(ctx, dst, opts) +} + +// PWrite writes src to the file represented by fd, starting at the given +// offset, and returns the number of bytes written. PWrite is permitted to +// return partial writes with a nil error. +func (fd *FileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) { + return fd.impl.PWrite(ctx, src, offset, opts) +} + +// Write is similar to PWrite, but does not specify an offset. +func (fd *FileDescription) Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) { + return fd.impl.Write(ctx, src, opts) +} + +// IterDirents invokes cb on each entry in the directory represented by fd. If +// IterDirents has been called since the last call to Seek, it continues +// iteration from the end of the last call. +func (fd *FileDescription) IterDirents(ctx context.Context, cb IterDirentsCallback) error { + return fd.impl.IterDirents(ctx, cb) +} + +// Seek changes fd's offset (assuming one exists) and returns its new value. +func (fd *FileDescription) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + return fd.impl.Seek(ctx, offset, whence) +} + +// Sync has the semantics of fsync(2). +func (fd *FileDescription) Sync(ctx context.Context) error { + return fd.impl.Sync(ctx) +} + +// ConfigureMMap mutates opts to implement mmap(2) for the file represented by +// fd. +func (fd *FileDescription) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return fd.impl.ConfigureMMap(ctx, opts) +} + +// Ioctl implements the ioctl(2) syscall. +func (fd *FileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + return fd.impl.Ioctl(ctx, uio, args) +} + +// Listxattr returns all extended attribute names for the file represented by +// fd. +func (fd *FileDescription) Listxattr(ctx context.Context) ([]string, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp) + vfsObj.putResolvingPath(rp) + return names, err + } + names, err := fd.impl.Listxattr(ctx) + if err == syserror.ENOTSUP { + // Linux doesn't actually return ENOTSUP in this case; instead, + // fs/xattr.c:vfs_listxattr() falls back to allowing the security + // subsystem to return security extended attributes, which by default + // don't exist. + return nil, nil + } + return names, err +} + +// Getxattr returns the value associated with the given extended attribute for +// the file represented by fd. +func (fd *FileDescription) Getxattr(ctx context.Context, name string) (string, error) { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, name) + vfsObj.putResolvingPath(rp) + return val, err + } + return fd.impl.Getxattr(ctx, name) +} + +// Setxattr changes the value associated with the given extended attribute for +// the file represented by fd. +func (fd *FileDescription) Setxattr(ctx context.Context, opts SetxattrOptions) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, opts) + vfsObj.putResolvingPath(rp) + return err + } + return fd.impl.Setxattr(ctx, opts) +} + +// Removexattr removes the given extended attribute from the file represented +// by fd. +func (fd *FileDescription) Removexattr(ctx context.Context, name string) error { + if fd.opts.UseDentryMetadata { + vfsObj := fd.vd.mount.vfs + rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ + Root: fd.vd, + Start: fd.vd, + }) + err := fd.vd.mount.fs.impl.RemovexattrAt(ctx, rp, name) + vfsObj.putResolvingPath(rp) + return err + } + return fd.impl.Removexattr(ctx, name) +} + +// SyncFS instructs the filesystem containing fd to execute the semantics of +// syncfs(2). +func (fd *FileDescription) SyncFS(ctx context.Context) error { + return fd.vd.mount.fs.impl.Sync(ctx) +} + +// MappedName implements memmap.MappingIdentity.MappedName. +func (fd *FileDescription) MappedName(ctx context.Context) string { + vfsroot := RootFromContext(ctx) + s, _ := fd.vd.mount.vfs.PathnameWithDeleted(ctx, vfsroot, fd.vd) + if vfsroot.Ok() { + vfsroot.DecRef() + } + return s +} + +// DeviceID implements memmap.MappingIdentity.DeviceID. +func (fd *FileDescription) DeviceID() uint64 { + stat, err := fd.Stat(context.Background(), StatOptions{ + // There is no STATX_DEV; we assume that Stat will return it if it's + // available regardless of mask. + Mask: 0, + // fs/proc/task_mmu.c:show_map_vma() just reads inode::i_sb->s_dev + // directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil { + return 0 + } + return uint64(linux.MakeDeviceID(uint16(stat.DevMajor), stat.DevMinor)) +} + +// InodeID implements memmap.MappingIdentity.InodeID. +func (fd *FileDescription) InodeID() uint64 { + stat, err := fd.Stat(context.Background(), StatOptions{ + Mask: linux.STATX_INO, + // fs/proc/task_mmu.c:show_map_vma() just reads inode::i_ino directly. + Sync: linux.AT_STATX_DONT_SYNC, + }) + if err != nil || stat.Mask&linux.STATX_INO == 0 { + return 0 + } + return stat.Ino +} + +// Msync implements memmap.MappingIdentity.Msync. +func (fd *FileDescription) Msync(ctx context.Context, mr memmap.MappableRange) error { + return fd.Sync(ctx) +} diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index 4fbad7840..66eb57bc2 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -127,6 +127,31 @@ func (FileDescriptionDefaultImpl) Ioctl(ctx context.Context, uio usermem.IO, arg return 0, syserror.ENOTTY } +// Listxattr implements FileDescriptionImpl.Listxattr analogously to +// inode_operations::listxattr == NULL in Linux. +func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context) ([]string, error) { + // This isn't exactly accurate; see FileDescription.Listxattr. + return nil, syserror.ENOTSUP +} + +// Getxattr implements FileDescriptionImpl.Getxattr analogously to +// inode::i_opflags & IOP_XATTR == 0 in Linux. +func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, name string) (string, error) { + return "", syserror.ENOTSUP +} + +// Setxattr implements FileDescriptionImpl.Setxattr analogously to +// inode::i_opflags & IOP_XATTR == 0 in Linux. +func (FileDescriptionDefaultImpl) Setxattr(ctx context.Context, opts SetxattrOptions) error { + return syserror.ENOTSUP +} + +// Removexattr implements FileDescriptionImpl.Removexattr analogously to +// inode::i_opflags & IOP_XATTR == 0 in Linux. +func (FileDescriptionDefaultImpl) Removexattr(ctx context.Context, name string) error { + return syserror.ENOTSUP +} + // DirectoryFileDescriptionDefaultImpl may be embedded by implementations of // FileDescriptionImpl that always represent directories to obtain // implementations of non-directory I/O methods that return EISDIR. @@ -152,6 +177,21 @@ func (DirectoryFileDescriptionDefaultImpl) Write(ctx context.Context, src userme return 0, syserror.EISDIR } +// DentryMetadataFileDescriptionImpl may be embedded by implementations of +// FileDescriptionImpl for which FileDescriptionOptions.UseDentryMetadata is +// true to obtain implementations of Stat and SetStat that panic. +type DentryMetadataFileDescriptionImpl struct{} + +// Stat implements FileDescriptionImpl.Stat. +func (DentryMetadataFileDescriptionImpl) Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) { + panic("illegal call to DentryMetadataFileDescriptionImpl.Stat") +} + +// SetStat implements FileDescriptionImpl.SetStat. +func (DentryMetadataFileDescriptionImpl) SetStat(ctx context.Context, opts SetStatOptions) error { + panic("illegal call to DentryMetadataFileDescriptionImpl.SetStat") +} + // DynamicBytesFileDescriptionImpl may be embedded by implementations of // FileDescriptionImpl that represent read-only regular files whose contents // are backed by a bytes.Buffer that is regenerated when necessary, consistent @@ -174,6 +214,17 @@ type DynamicBytesSource interface { Generate(ctx context.Context, buf *bytes.Buffer) error } +// StaticData implements DynamicBytesSource over a static string. +type StaticData struct { + Data string +} + +// Generate implements DynamicBytesSource. +func (s *StaticData) Generate(ctx context.Context, buf *bytes.Buffer) error { + buf.WriteString(s.Data) + return nil +} + // SetDataSource must be called exactly once on fd before first use. func (fd *DynamicBytesFileDescriptionImpl) SetDataSource(data DynamicBytesSource) { fd.data = data @@ -252,3 +303,12 @@ func (fd *DynamicBytesFileDescriptionImpl) Seek(ctx context.Context, offset int6 fd.off = offset return offset, nil } + +// GenericConfigureMMap may be used by most implementations of +// FileDescriptionImpl.ConfigureMMap. +func GenericConfigureMMap(fd *FileDescription, m memmap.Mappable, opts *memmap.MMapOpts) error { + opts.Mappable = m + opts.MappingIdentity = fd + fd.IncRef() + return nil +} diff --git a/pkg/sentry/vfs/file_description_impl_util_test.go b/pkg/sentry/vfs/file_description_impl_util_test.go index a5561dcbe..9ed58512f 100644 --- a/pkg/sentry/vfs/file_description_impl_util_test.go +++ b/pkg/sentry/vfs/file_description_impl_util_test.go @@ -48,7 +48,7 @@ type genCountFD struct { func newGenCountFD(mnt *Mount, vfsd *Dentry) *FileDescription { var fd genCountFD - fd.vfsfd.Init(&fd, mnt, vfsd) + fd.vfsfd.Init(&fd, 0 /* statusFlags */, mnt, vfsd, &FileDescriptionOptions{}) fd.DynamicBytesFileDescriptionImpl.SetDataSource(&fd) return &fd.vfsfd } @@ -89,7 +89,7 @@ func TestGenCountFD(t *testing.T) { creds := auth.CredentialsFromContext(ctx) vfsObj := New() // vfs.New() - vfsObj.MustRegisterFilesystemType("testfs", FDTestFilesystemType{}) + vfsObj.MustRegisterFilesystemType("testfs", FDTestFilesystemType{}, &RegisterFilesystemTypeOptions{}) mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "testfs", &GetFilesystemOptions{}) if err != nil { t.Fatalf("failed to create testfs root mount: %v", err) @@ -103,7 +103,7 @@ func TestGenCountFD(t *testing.T) { // The first read causes Generate to be called to fill the FD's buffer. buf := make([]byte, 2) ioseq := usermem.BytesIOSequence(buf) - n, err := fd.Impl().Read(ctx, ioseq, ReadOptions{}) + n, err := fd.Read(ctx, ioseq, ReadOptions{}) if n != 1 || (err != nil && err != io.EOF) { t.Fatalf("first Read: got (%d, %v), wanted (1, nil or EOF)", n, err) } @@ -112,17 +112,17 @@ func TestGenCountFD(t *testing.T) { } // A second read without seeking is still at EOF. - n, err = fd.Impl().Read(ctx, ioseq, ReadOptions{}) + n, err = fd.Read(ctx, ioseq, ReadOptions{}) if n != 0 || err != io.EOF { t.Fatalf("second Read: got (%d, %v), wanted (0, EOF)", n, err) } // Seeking to the beginning of the file causes it to be regenerated. - n, err = fd.Impl().Seek(ctx, 0, linux.SEEK_SET) + n, err = fd.Seek(ctx, 0, linux.SEEK_SET) if n != 0 || err != nil { t.Fatalf("Seek: got (%d, %v), wanted (0, nil)", n, err) } - n, err = fd.Impl().Read(ctx, ioseq, ReadOptions{}) + n, err = fd.Read(ctx, ioseq, ReadOptions{}) if n != 1 || (err != nil && err != io.EOF) { t.Fatalf("Read after Seek: got (%d, %v), wanted (1, nil or EOF)", n, err) } @@ -131,7 +131,7 @@ func TestGenCountFD(t *testing.T) { } // PRead at the beginning of the file also causes it to be regenerated. - n, err = fd.Impl().PRead(ctx, ioseq, 0, ReadOptions{}) + n, err = fd.PRead(ctx, ioseq, 0, ReadOptions{}) if n != 1 || (err != nil && err != io.EOF) { t.Fatalf("PRead: got (%d, %v), wanted (1, nil or EOF)", n, err) } diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index 76ff8cf51..ea78f555b 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -18,6 +18,7 @@ import ( "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" ) @@ -47,6 +48,9 @@ func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, impl FilesystemImpl) { fs.refs = 1 fs.vfs = vfsObj fs.impl = impl + vfsObj.filesystemsMu.Lock() + vfsObj.filesystems[fs] = struct{}{} + vfsObj.filesystemsMu.Unlock() } // VirtualFilesystem returns the containing VirtualFilesystem. @@ -66,9 +70,28 @@ func (fs *Filesystem) IncRef() { } } +// TryIncRef increments fs' reference count and returns true. If fs' reference +// count is zero, TryIncRef does nothing and returns false. +// +// TryIncRef does not require that a reference is held on fs. +func (fs *Filesystem) TryIncRef() bool { + for { + refs := atomic.LoadInt64(&fs.refs) + if refs <= 0 { + return false + } + if atomic.CompareAndSwapInt64(&fs.refs, refs, refs+1) { + return true + } + } +} + // DecRef decrements fs' reference count. func (fs *Filesystem) DecRef() { if refs := atomic.AddInt64(&fs.refs, -1); refs == 0 { + fs.vfs.filesystemsMu.Lock() + delete(fs.vfs.filesystems, fs) + fs.vfs.filesystemsMu.Unlock() fs.impl.Release() } else if refs < 0 { panic("Filesystem.decRef() called without holding a reference") @@ -85,6 +108,24 @@ func (fs *Filesystem) DecRef() { // (responsible for actually implementing the operation) isn't known until path // resolution is complete. // +// Unless otherwise specified, FilesystemImpl methods are responsible for +// performing permission checks. In many cases, vfs package functions in +// permissions.go may be used to help perform these checks. +// +// When multiple specified error conditions apply to a given method call, the +// implementation may return any applicable errno unless otherwise specified, +// but returning the earliest error specified is preferable to maximize +// compatibility with Linux. +// +// All methods may return errors not specified, notably including: +// +// - ENOENT if a required path component does not exist. +// +// - ENOTDIR if an intermediate path component is not a directory. +// +// - Errors from vfs-package functions (ResolvingPath.Resolve*(), +// Mount.CheckBeginWrite(), permission-checking functions, etc.) +// // For all methods that take or return linux.Statx, Statx.Uid and Statx.Gid // should be interpreted as IDs in the root UserNamespace (i.e. as auth.KUID // and auth.KGID respectively). @@ -107,46 +148,223 @@ type FilesystemImpl interface { // GetDentryAt does not correspond directly to a Linux syscall; it is used // in the implementation of: // - // - Syscalls that need to resolve two paths: rename(), renameat(), - // renameat2(), link(), linkat(). + // - Syscalls that need to resolve two paths: link(), linkat(). // // - Syscalls that need to refer to a filesystem position outside the // context of a file description: chdir(), fchdir(), chroot(), mount(), // umount(). GetDentryAt(ctx context.Context, rp *ResolvingPath, opts GetDentryOptions) (*Dentry, error) + // GetParentDentryAt returns a Dentry representing the directory at the + // second-to-last path component in rp. (Note that, despite the name, this + // is not necessarily the parent directory of the file at rp, since the + // last path component in rp may be "." or "..".) A reference is taken on + // the returned Dentry. + // + // GetParentDentryAt does not correspond directly to a Linux syscall; it is + // used in the implementation of the rename() family of syscalls, which + // must resolve the parent directories of two paths. + // + // Preconditions: !rp.Done(). + // + // Postconditions: If GetParentDentryAt returns a nil error, then + // rp.Final(). If GetParentDentryAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). + GetParentDentryAt(ctx context.Context, rp *ResolvingPath) (*Dentry, error) + // LinkAt creates a hard link at rp representing the same file as vd. It // does not take ownership of references on vd. // - // The implementation is responsible for checking that vd.Mount() == - // rp.Mount(), and that vd does not represent a directory. + // Errors: + // + // - If the last path component in rp is "." or "..", LinkAt returns + // EEXIST. + // + // - If a file already exists at rp, LinkAt returns EEXIST. + // + // - If rp.MustBeDir(), LinkAt returns ENOENT. + // + // - If the directory in which the link would be created has been removed + // by RmdirAt or RenameAt, LinkAt returns ENOENT. + // + // - If rp.Mount != vd.Mount(), LinkAt returns EXDEV. + // + // - If vd represents a directory, LinkAt returns EPERM. + // + // - If vd represents a file for which all existing links have been + // removed, or a file created by open(O_TMPFILE|O_EXCL), LinkAt returns + // ENOENT. Equivalently, if vd represents a file with a link count of 0 not + // created by open(O_TMPFILE) without O_EXCL, LinkAt returns ENOENT. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If LinkAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). LinkAt(ctx context.Context, rp *ResolvingPath, vd VirtualDentry) error // MkdirAt creates a directory at rp. + // + // Errors: + // + // - If the last path component in rp is "." or "..", MkdirAt returns + // EEXIST. + // + // - If a file already exists at rp, MkdirAt returns EEXIST. + // + // - If the directory in which the new directory would be created has been + // removed by RmdirAt or RenameAt, MkdirAt returns ENOENT. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If MkdirAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). MkdirAt(ctx context.Context, rp *ResolvingPath, opts MkdirOptions) error // MknodAt creates a regular file, device special file, or named pipe at // rp. + // + // Errors: + // + // - If the last path component in rp is "." or "..", MknodAt returns + // EEXIST. + // + // - If a file already exists at rp, MknodAt returns EEXIST. + // + // - If rp.MustBeDir(), MknodAt returns ENOENT. + // + // - If the directory in which the file would be created has been removed + // by RmdirAt or RenameAt, MknodAt returns ENOENT. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If MknodAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). MknodAt(ctx context.Context, rp *ResolvingPath, opts MknodOptions) error // OpenAt returns an FileDescription providing access to the file at rp. A // reference is taken on the returned FileDescription. + // + // Errors: + // + // - If opts.Flags specifies O_TMPFILE and this feature is unsupported by + // the implementation, OpenAt returns EOPNOTSUPP. (All other unsupported + // features are silently ignored, consistently with Linux's open*(2).) OpenAt(ctx context.Context, rp *ResolvingPath, opts OpenOptions) (*FileDescription, error) // ReadlinkAt returns the target of the symbolic link at rp. + // + // Errors: + // + // - If the file at rp is not a symbolic link, ReadlinkAt returns EINVAL. ReadlinkAt(ctx context.Context, rp *ResolvingPath) (string, error) - // RenameAt renames the Dentry represented by vd to rp. It does not take - // ownership of references on vd. + // RenameAt renames the file named oldName in directory oldParentVD to rp. + // It does not take ownership of references on oldParentVD. + // + // Errors [1]: + // + // - If opts.Flags specifies unsupported options, RenameAt returns EINVAL. + // + // - If the last path component in rp is "." or "..", and opts.Flags + // contains RENAME_NOREPLACE, RenameAt returns EEXIST. + // + // - If the last path component in rp is "." or "..", and opts.Flags does + // not contain RENAME_NOREPLACE, RenameAt returns EBUSY. + // + // - If rp.Mount != oldParentVD.Mount(), RenameAt returns EXDEV. + // + // - If the renamed file is not a directory, and opts.MustBeDir is true, + // RenameAt returns ENOTDIR. + // + // - If renaming would replace an existing file and opts.Flags contains + // RENAME_NOREPLACE, RenameAt returns EEXIST. + // + // - If there is no existing file at rp and opts.Flags contains + // RENAME_EXCHANGE, RenameAt returns ENOENT. + // + // - If there is an existing non-directory file at rp, and rp.MustBeDir() + // is true, RenameAt returns ENOTDIR. + // + // - If the renamed file is not a directory, opts.Flags does not contain + // RENAME_EXCHANGE, and rp.MustBeDir() is true, RenameAt returns ENOTDIR. + // (This check is not subsumed by the check for directory replacement below + // since it applies even if there is no file to replace.) + // + // - If the renamed file is a directory, and the new parent directory of + // the renamed file is either the renamed directory or a descendant + // subdirectory of the renamed directory, RenameAt returns EINVAL. + // + // - If renaming would exchange the renamed file with an ancestor directory + // of the renamed file, RenameAt returns EINVAL. + // + // - If renaming would replace an ancestor directory of the renamed file, + // RenameAt returns ENOTEMPTY. (This check would be subsumed by the + // non-empty directory check below; however, this check takes place before + // the self-rename check.) + // + // - If the renamed file would replace or exchange with itself (i.e. the + // source and destination paths resolve to the same file), RenameAt returns + // nil, skipping the checks described below. + // + // - If the source or destination directory is not writable by the provider + // of rp.Credentials(), RenameAt returns EACCES. + // + // - If the renamed file is a directory, and renaming would replace a + // non-directory file, RenameAt returns ENOTDIR. // - // The implementation is responsible for checking that vd.Mount() == - // rp.Mount(). - RenameAt(ctx context.Context, rp *ResolvingPath, vd VirtualDentry, opts RenameOptions) error + // - If the renamed file is not a directory, and renaming would replace a + // directory, RenameAt returns EISDIR. + // + // - If the new parent directory of the renamed file has been removed by + // RmdirAt or a preceding call to RenameAt, RenameAt returns ENOENT. + // + // - If the renamed file is a directory, it is not writable by the + // provider of rp.Credentials(), and the source and destination parent + // directories are different, RenameAt returns EACCES. (This is nominally + // required to change the ".." entry in the renamed directory.) + // + // - If renaming would replace a non-empty directory, RenameAt returns + // ENOTEMPTY. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). oldName is not "." or "..". + // + // Postconditions: If RenameAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). + // + // [1] "The worst of all namespace operations - renaming directory. + // "Perverted" doesn't even start to describe it. Somebody in UCB had a + // heck of a trip..." - fs/namei.c:vfs_rename() + RenameAt(ctx context.Context, rp *ResolvingPath, oldParentVD VirtualDentry, oldName string, opts RenameOptions) error // RmdirAt removes the directory at rp. + // + // Errors: + // + // - If the last path component in rp is ".", RmdirAt returns EINVAL. + // + // - If the last path component in rp is "..", RmdirAt returns ENOTEMPTY. + // + // - If no file exists at rp, RmdirAt returns ENOENT. + // + // - If the file at rp exists but is not a directory, RmdirAt returns + // ENOTDIR. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If RmdirAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). RmdirAt(ctx context.Context, rp *ResolvingPath) error // SetStatAt updates metadata for the file at the given path. + // + // Errors: + // + // - If opts specifies unsupported options, SetStatAt returns EINVAL. SetStatAt(ctx context.Context, rp *ResolvingPath, opts SetStatOptions) error // StatAt returns metadata for the file at rp. @@ -158,10 +376,132 @@ type FilesystemImpl interface { StatFSAt(ctx context.Context, rp *ResolvingPath) (linux.Statfs, error) // SymlinkAt creates a symbolic link at rp referring to the given target. + // + // Errors: + // + // - If the last path component in rp is "." or "..", SymlinkAt returns + // EEXIST. + // + // - If a file already exists at rp, SymlinkAt returns EEXIST. + // + // - If rp.MustBeDir(), SymlinkAt returns ENOENT. + // + // - If the directory in which the symbolic link would be created has been + // removed by RmdirAt or RenameAt, SymlinkAt returns ENOENT. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If SymlinkAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). SymlinkAt(ctx context.Context, rp *ResolvingPath, target string) error - // UnlinkAt removes the non-directory file at rp. + // UnlinkAt removes the file at rp. + // + // Errors: + // + // - If the last path component in rp is "." or "..", UnlinkAt returns + // EISDIR. + // + // - If no file exists at rp, UnlinkAt returns ENOENT. + // + // - If rp.MustBeDir(), and the file at rp exists and is not a directory, + // UnlinkAt returns ENOTDIR. + // + // - If the file at rp exists but is a directory, UnlinkAt returns EISDIR. + // + // Preconditions: !rp.Done(). For the final path component in rp, + // !rp.ShouldFollowSymlink(). + // + // Postconditions: If UnlinkAt returns an error returned by + // ResolvingPath.Resolve*(), then !rp.Done(). UnlinkAt(ctx context.Context, rp *ResolvingPath) error - // TODO: d_path(); extended attributes; inotify_add_watch(); bind() + // ListxattrAt returns all extended attribute names for the file at rp. + // + // Errors: + // + // - If extended attributes are not supported by the filesystem, + // ListxattrAt returns nil. (See FileDescription.Listxattr for an + // explanation.) + ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) + + // GetxattrAt returns the value associated with the given extended + // attribute for the file at rp. + // + // Errors: + // + // - If extended attributes are not supported by the filesystem, GetxattrAt + // returns ENOTSUP. + GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) + + // SetxattrAt changes the value associated with the given extended + // attribute for the file at rp. + // + // Errors: + // + // - If extended attributes are not supported by the filesystem, SetxattrAt + // returns ENOTSUP. + SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error + + // RemovexattrAt removes the given extended attribute from the file at rp. + // + // Errors: + // + // - If extended attributes are not supported by the filesystem, + // RemovexattrAt returns ENOTSUP. + RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error + + // PrependPath prepends a path from vd to vd.Mount().Root() to b. + // + // If vfsroot.Ok(), it is the contextual VFS root; if it is encountered + // before vd.Mount().Root(), PrependPath should stop prepending path + // components and return a PrependPathAtVFSRootError. + // + // If traversal of vd.Dentry()'s ancestors encounters an independent + // ("root") Dentry that is not vd.Mount().Root() (i.e. vd.Dentry() is not a + // descendant of vd.Mount().Root()), PrependPath should stop prepending + // path components and return a PrependPathAtNonMountRootError. + // + // Filesystems for which Dentries do not have meaningful paths may prepend + // an arbitrary descriptive string to b and then return a + // PrependPathSyntheticError. + // + // Most implementations can acquire the appropriate locks to ensure that + // Dentry.Name() and Dentry.Parent() are fixed for vd.Dentry() and all of + // its ancestors, then call GenericPrependPath. + // + // Preconditions: vd.Mount().Filesystem().Impl() == this FilesystemImpl. + PrependPath(ctx context.Context, vfsroot, vd VirtualDentry, b *fspath.Builder) error + + // TODO: inotify_add_watch(); bind() +} + +// PrependPathAtVFSRootError is returned by implementations of +// FilesystemImpl.PrependPath() when they encounter the contextual VFS root. +type PrependPathAtVFSRootError struct{} + +// Error implements error.Error. +func (PrependPathAtVFSRootError) Error() string { + return "vfs.FilesystemImpl.PrependPath() reached VFS root" +} + +// PrependPathAtNonMountRootError is returned by implementations of +// FilesystemImpl.PrependPath() when they encounter an independent ancestor +// Dentry that is not the Mount root. +type PrependPathAtNonMountRootError struct{} + +// Error implements error.Error. +func (PrependPathAtNonMountRootError) Error() string { + return "vfs.FilesystemImpl.PrependPath() reached root other than Mount root" +} + +// PrependPathSyntheticError is returned by implementations of +// FilesystemImpl.PrependPath() for which prepended names do not represent real +// paths. +type PrependPathSyntheticError struct{} + +// Error implements error.Error. +func (PrependPathSyntheticError) Error() string { + return "vfs.FilesystemImpl.PrependPath() prepended synthetic name" } diff --git a/pkg/sentry/vfs/filesystem_impl_util.go b/pkg/sentry/vfs/filesystem_impl_util.go index 465e610e0..7315a588e 100644 --- a/pkg/sentry/vfs/filesystem_impl_util.go +++ b/pkg/sentry/vfs/filesystem_impl_util.go @@ -16,6 +16,8 @@ package vfs import ( "strings" + + "gvisor.dev/gvisor/pkg/fspath" ) // GenericParseMountOptions parses a comma-separated list of options of the @@ -41,3 +43,27 @@ func GenericParseMountOptions(str string) map[string]string { } return m } + +// GenericPrependPath may be used by implementations of +// FilesystemImpl.PrependPath() for which a single statically-determined lock +// or set of locks is sufficient to ensure its preconditions (as opposed to +// e.g. per-Dentry locks). +// +// Preconditions: Dentry.Name() and Dentry.Parent() must be held constant for +// vd.Dentry() and all of its ancestors. +func GenericPrependPath(vfsroot, vd VirtualDentry, b *fspath.Builder) error { + mnt, d := vd.mount, vd.dentry + for { + if mnt == vfsroot.mount && d == vfsroot.dentry { + return PrependPathAtVFSRootError{} + } + if d == mnt.root { + return nil + } + if d.parent == nil { + return PrependPathAtNonMountRootError{} + } + b.PrependComponent(d.name) + d = d.parent + } +} diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go index c335e206d..023301780 100644 --- a/pkg/sentry/vfs/filesystem_type.go +++ b/pkg/sentry/vfs/filesystem_type.go @@ -15,6 +15,7 @@ package vfs import ( + "bytes" "fmt" "gvisor.dev/gvisor/pkg/sentry/context" @@ -43,28 +44,70 @@ type GetFilesystemOptions struct { InternalData interface{} } +type registeredFilesystemType struct { + fsType FilesystemType + opts RegisterFilesystemTypeOptions +} + +// RegisterFilesystemTypeOptions contains options to +// VirtualFilesystem.RegisterFilesystem(). +type RegisterFilesystemTypeOptions struct { + // If AllowUserMount is true, allow calls to VirtualFilesystem.MountAt() + // for which MountOptions.InternalMount == false to use this filesystem + // type. + AllowUserMount bool + + // If AllowUserList is true, make this filesystem type visible in + // /proc/filesystems. + AllowUserList bool + + // If RequiresDevice is true, indicate that mounting this filesystem + // requires a block device as the mount source in /proc/filesystems. + RequiresDevice bool +} + // RegisterFilesystemType registers the given FilesystemType in vfs with the // given name. -func (vfs *VirtualFilesystem) RegisterFilesystemType(name string, fsType FilesystemType) error { +func (vfs *VirtualFilesystem) RegisterFilesystemType(name string, fsType FilesystemType, opts *RegisterFilesystemTypeOptions) error { vfs.fsTypesMu.Lock() defer vfs.fsTypesMu.Unlock() if existing, ok := vfs.fsTypes[name]; ok { - return fmt.Errorf("name %q is already registered to filesystem type %T", name, existing) + return fmt.Errorf("name %q is already registered to filesystem type %T", name, existing.fsType) + } + vfs.fsTypes[name] = ®isteredFilesystemType{ + fsType: fsType, + opts: *opts, } - vfs.fsTypes[name] = fsType return nil } // MustRegisterFilesystemType is equivalent to RegisterFilesystemType but // panics on failure. -func (vfs *VirtualFilesystem) MustRegisterFilesystemType(name string, fsType FilesystemType) { - if err := vfs.RegisterFilesystemType(name, fsType); err != nil { +func (vfs *VirtualFilesystem) MustRegisterFilesystemType(name string, fsType FilesystemType, opts *RegisterFilesystemTypeOptions) { + if err := vfs.RegisterFilesystemType(name, fsType, opts); err != nil { panic(fmt.Sprintf("failed to register filesystem type %T: %v", fsType, err)) } } -func (vfs *VirtualFilesystem) getFilesystemType(name string) FilesystemType { +func (vfs *VirtualFilesystem) getFilesystemType(name string) *registeredFilesystemType { vfs.fsTypesMu.RLock() defer vfs.fsTypesMu.RUnlock() return vfs.fsTypes[name] } + +// GenerateProcFilesystems emits the contents of /proc/filesystems for vfs to +// buf. +func (vfs *VirtualFilesystem) GenerateProcFilesystems(buf *bytes.Buffer) { + vfs.fsTypesMu.RLock() + defer vfs.fsTypesMu.RUnlock() + for name, rft := range vfs.fsTypes { + if !rft.opts.AllowUserList { + continue + } + var nodev string + if !rft.opts.RequiresDevice { + nodev = "nodev" + } + fmt.Fprintf(buf, "%s\t%s\n", nodev, name) + } +} diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 1c3b2e987..00177b371 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -18,6 +18,7 @@ import ( "math" "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/syserror" @@ -111,11 +112,11 @@ type MountNamespace struct { // configured by the given arguments. A reference is taken on the returned // MountNamespace. func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { - fsType := vfs.getFilesystemType(fsTypeName) - if fsType == nil { + rft := vfs.getFilesystemType(fsTypeName) + if rft == nil { return nil, syserror.ENODEV } - fs, root, err := fsType.GetFilesystem(ctx, vfs, creds, source, *opts) + fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) if err != nil { return nil, err } @@ -133,13 +134,16 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth return mntns, nil } -// NewMount creates and mounts a Filesystem configured by the given arguments. -func (vfs *VirtualFilesystem) NewMount(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *GetFilesystemOptions) error { - fsType := vfs.getFilesystemType(fsTypeName) - if fsType == 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) + if rft == nil { return syserror.ENODEV } - fs, root, err := fsType.GetFilesystem(ctx, vfs, creds, source, *opts) + if !opts.InternalMount && !rft.opts.AllowUserMount { + return syserror.ENODEV + } + fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, opts.GetFilesystemOptions) if err != nil { return err } @@ -207,6 +211,68 @@ func (vfs *VirtualFilesystem) NewMount(ctx context.Context, creds *auth.Credenti return nil } +// UmountAt removes the Mount at the given path. +func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *UmountOptions) error { + if opts.Flags&^(linux.MNT_FORCE|linux.MNT_DETACH) != 0 { + return syserror.EINVAL + } + + // MNT_FORCE is currently unimplemented except for the permission check. + if opts.Flags&linux.MNT_FORCE != 0 && creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, creds.UserNamespace.Root()) { + return syserror.EPERM + } + + vd, err := vfs.GetDentryAt(ctx, creds, pop, &GetDentryOptions{}) + if err != nil { + return err + } + defer vd.DecRef() + if vd.dentry != vd.mount.root { + return syserror.EINVAL + } + vfs.mountMu.Lock() + if mntns := MountNamespaceFromContext(ctx); mntns != nil && mntns != vd.mount.ns { + vfs.mountMu.Unlock() + return syserror.EINVAL + } + + // TODO(jamieliu): Linux special-cases umount of the caller's root, which + // we don't implement yet (we'll just fail it since the caller holds a + // reference on it). + + vfs.mounts.seq.BeginWrite() + if opts.Flags&linux.MNT_DETACH == 0 { + if len(vd.mount.children) != 0 { + vfs.mounts.seq.EndWrite() + vfs.mountMu.Unlock() + return syserror.EBUSY + } + // We are holding a reference on vd.mount. + expectedRefs := int64(1) + if !vd.mount.umounted { + expectedRefs = 2 + } + if atomic.LoadInt64(&vd.mount.refs)&^math.MinInt64 != expectedRefs { // mask out MSB + vfs.mounts.seq.EndWrite() + vfs.mountMu.Unlock() + return syserror.EBUSY + } + } + vdsToDecRef, mountsToDecRef := vfs.umountRecursiveLocked(vd.mount, &umountRecursiveOptions{ + eager: opts.Flags&linux.MNT_DETACH == 0, + disconnectHierarchy: true, + }, nil, nil) + vfs.mounts.seq.EndWrite() + vfs.mountMu.Unlock() + for _, vd := range vdsToDecRef { + vd.DecRef() + } + for _, mnt := range mountsToDecRef { + mnt.DecRef() + } + return nil +} + type umountRecursiveOptions struct { // If eager is true, ensure that future calls to Mount.tryIncMountedRef() // on umounted mounts fail. diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index 3aa73d911..b7774bf28 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -46,6 +46,16 @@ type MknodOptions struct { DevMinor uint32 } +// MountOptions contains options to VirtualFilesystem.MountAt(). +type MountOptions struct { + // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). + GetFilesystemOptions GetFilesystemOptions + + // If InternalMount is true, allow the use of filesystem types for which + // RegisterFilesystemTypeOptions.AllowUserMount == false. + InternalMount bool +} + // OpenOptions contains options to VirtualFilesystem.OpenAt() and // FilesystemImpl.OpenAt(). type OpenOptions struct { @@ -77,6 +87,9 @@ type ReadOptions struct { type RenameOptions struct { // Flags contains flags as specified for renameat2(2). Flags uint32 + + // If MustBeDir is true, the renamed file must be a directory. + MustBeDir bool } // SetStatOptions contains options to VirtualFilesystem.SetStatAt(), @@ -95,6 +108,20 @@ type SetStatOptions struct { Stat linux.Statx } +// SetxattrOptions contains options to VirtualFilesystem.SetxattrAt(), +// FilesystemImpl.SetxattrAt(), FileDescription.Setxattr(), and +// FileDescriptionImpl.Setxattr(). +type SetxattrOptions struct { + // Name is the name of the extended attribute being mutated. + Name string + + // Value is the extended attribute's new value. + Value string + + // Flags contains flags as specified for setxattr/lsetxattr/fsetxattr(2). + Flags uint32 +} + // StatOptions contains options to VirtualFilesystem.StatAt(), // FilesystemImpl.StatAt(), FileDescription.Stat(), and // FileDescriptionImpl.Stat(). @@ -114,6 +141,12 @@ type StatOptions struct { Sync uint32 } +// UmountOptions contains options to VirtualFilesystem.UmountAt(). +type UmountOptions struct { + // Flags contains flags as specified for umount2(2). + Flags uint32 +} + // WriteOptions contains options to FileDescription.PWrite(), // FileDescriptionImpl.PWrite(), FileDescription.Write(), and // FileDescriptionImpl.Write(). diff --git a/pkg/sentry/vfs/pathname.go b/pkg/sentry/vfs/pathname.go new file mode 100644 index 000000000..8e155654f --- /dev/null +++ b/pkg/sentry/vfs/pathname.go @@ -0,0 +1,153 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs + +import ( + "sync" + + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/syserror" +) + +var fspathBuilderPool = sync.Pool{ + New: func() interface{} { + return &fspath.Builder{} + }, +} + +func getFSPathBuilder() *fspath.Builder { + return fspathBuilderPool.Get().(*fspath.Builder) +} + +func putFSPathBuilder(b *fspath.Builder) { + // No methods can be called on b after b.String(), so reset it to its zero + // value (as returned by fspathBuilderPool.New) instead. + *b = fspath.Builder{} + fspathBuilderPool.Put(b) +} + +// PathnameWithDeleted returns an absolute pathname to vd, consistent with +// Linux's d_path(). In particular, if vd.Dentry() has been disowned, +// PathnameWithDeleted appends " (deleted)" to the returned pathname. +func (vfs *VirtualFilesystem) PathnameWithDeleted(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { + b := getFSPathBuilder() + defer putFSPathBuilder(b) + haveRef := false + defer func() { + if haveRef { + vd.DecRef() + } + }() + + origD := vd.dentry +loop: + for { + err := vd.mount.fs.impl.PrependPath(ctx, vfsroot, vd, b) + switch err.(type) { + case nil: + if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { + // GenericPrependPath() will have returned + // PrependPathAtVFSRootError in this case since it checks + // against vfsroot before mnt.root, but other implementations + // of FilesystemImpl.PrependPath() may return nil instead. + break loop + } + nextVD := vfs.getMountpointAt(vd.mount, vfsroot) + if !nextVD.Ok() { + break loop + } + if haveRef { + vd.DecRef() + } + vd = nextVD + haveRef = true + // continue loop + case PrependPathSyntheticError: + // Skip prepending "/" and appending " (deleted)". + return b.String(), nil + case PrependPathAtVFSRootError, PrependPathAtNonMountRootError: + break loop + default: + return "", err + } + } + b.PrependByte('/') + if origD.IsDisowned() { + b.AppendString(" (deleted)") + } + return b.String(), nil +} + +// PathnameForGetcwd returns an absolute pathname to vd, consistent with +// Linux's sys_getcwd(). +func (vfs *VirtualFilesystem) PathnameForGetcwd(ctx context.Context, vfsroot, vd VirtualDentry) (string, error) { + if vd.dentry.IsDisowned() { + return "", syserror.ENOENT + } + + b := getFSPathBuilder() + defer putFSPathBuilder(b) + haveRef := false + defer func() { + if haveRef { + vd.DecRef() + } + }() + unreachable := false +loop: + for { + err := vd.mount.fs.impl.PrependPath(ctx, vfsroot, vd, b) + switch err.(type) { + case nil: + if vd.mount == vfsroot.mount && vd.mount.root == vfsroot.dentry { + break loop + } + nextVD := vfs.getMountpointAt(vd.mount, vfsroot) + if !nextVD.Ok() { + unreachable = true + break loop + } + if haveRef { + vd.DecRef() + } + vd = nextVD + haveRef = true + case PrependPathAtVFSRootError: + break loop + case PrependPathAtNonMountRootError, PrependPathSyntheticError: + unreachable = true + break loop + default: + return "", err + } + } + b.PrependByte('/') + if unreachable { + b.PrependString("(unreachable)") + } + return b.String(), nil +} + +// As of this writing, we do not have equivalents to: +// +// - d_absolute_path(), which returns EINVAL if (effectively) any call to +// FilesystemImpl.PrependPath() would return PrependPathAtNonMountRootError. +// +// - dentry_path(), which does not walk up mounts (and only returns the path +// relative to Filesystem root), but also appends "//deleted" for disowned +// Dentries. +// +// These should be added as necessary. diff --git a/pkg/sentry/vfs/permissions.go b/pkg/sentry/vfs/permissions.go index f8e74355c..f1edb0680 100644 --- a/pkg/sentry/vfs/permissions.go +++ b/pkg/sentry/vfs/permissions.go @@ -119,3 +119,65 @@ func MayWriteFileWithOpenFlags(flags uint32) bool { return false } } + +// CheckSetStat checks that creds has permission to change the metadata of a +// file with the given permissions, UID, and GID as specified by stat, subject +// to the rules of Linux's fs/attr.c:setattr_prepare(). +func CheckSetStat(creds *auth.Credentials, stat *linux.Statx, mode uint16, kuid auth.KUID, kgid auth.KGID) error { + if stat.Mask&linux.STATX_MODE != 0 { + if !CanActAsOwner(creds, kuid) { + return syserror.EPERM + } + // TODO(b/30815691): "If the calling process is not privileged (Linux: + // does not have the CAP_FSETID capability), and the group of the file + // does not match the effective group ID of the process or one of its + // supplementary group IDs, the S_ISGID bit will be turned off, but + // this will not cause an error to be returned." - chmod(2) + } + if stat.Mask&linux.STATX_UID != 0 { + if !((creds.EffectiveKUID == kuid && auth.KUID(stat.UID) == kuid) || + HasCapabilityOnFile(creds, linux.CAP_CHOWN, kuid, kgid)) { + return syserror.EPERM + } + } + if stat.Mask&linux.STATX_GID != 0 { + if !((creds.EffectiveKUID == kuid && creds.InGroup(auth.KGID(stat.GID))) || + HasCapabilityOnFile(creds, linux.CAP_CHOWN, kuid, kgid)) { + return syserror.EPERM + } + } + if stat.Mask&(linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME) != 0 { + if !CanActAsOwner(creds, kuid) { + if (stat.Mask&linux.STATX_ATIME != 0 && stat.Atime.Nsec != linux.UTIME_NOW) || + (stat.Mask&linux.STATX_MTIME != 0 && stat.Mtime.Nsec != linux.UTIME_NOW) || + (stat.Mask&linux.STATX_CTIME != 0 && stat.Ctime.Nsec != linux.UTIME_NOW) { + return syserror.EPERM + } + // isDir is irrelevant in the following call to + // GenericCheckPermissions since ats == MayWrite means that + // CAP_DAC_READ_SEARCH does not apply, and CAP_DAC_OVERRIDE + // applies, regardless of isDir. + if err := GenericCheckPermissions(creds, MayWrite, false /* isDir */, mode, kuid, kgid); err != nil { + return err + } + } + } + return nil +} + +// CanActAsOwner returns true if creds can act as the owner of a file with the +// given owning UID, consistent with Linux's +// fs/inode.c:inode_owner_or_capable(). +func CanActAsOwner(creds *auth.Credentials, kuid auth.KUID) bool { + if creds.EffectiveKUID == kuid { + return true + } + return creds.HasCapability(linux.CAP_FOWNER) && creds.UserNamespace.MapFromKUID(kuid).Ok() +} + +// HasCapabilityOnFile returns true if creds has the given capability with +// respect to a file with the given owning UID and GID, consistent with Linux's +// kernel/capability.c:capable_wrt_inode_uidgid(). +func HasCapabilityOnFile(creds *auth.Credentials, cp linux.Capability, kuid auth.KUID, kgid auth.KGID) bool { + return creds.HasCapability(cp) && creds.UserNamespace.MapFromKUID(kuid).Ok() && creds.UserNamespace.MapFromKGID(kgid).Ok() +} diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 621f5a6f8..f0641d314 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -85,11 +85,11 @@ func init() { // so error "constants" are really mutable vars, necessitating somewhat // expensive interface object comparisons. -type resolveMountRootError struct{} +type resolveMountRootOrJumpError struct{} // Error implements error.Error. -func (resolveMountRootError) Error() string { - return "resolving mount root" +func (resolveMountRootOrJumpError) Error() string { + return "resolving mount root or jump" } type resolveMountPointError struct{} @@ -112,30 +112,26 @@ var resolvingPathPool = sync.Pool{ }, } -func (vfs *VirtualFilesystem) getResolvingPath(creds *auth.Credentials, pop *PathOperation) (*ResolvingPath, error) { - path, err := fspath.Parse(pop.Pathname) - if err != nil { - return nil, err - } +func (vfs *VirtualFilesystem) getResolvingPath(creds *auth.Credentials, pop *PathOperation) *ResolvingPath { rp := resolvingPathPool.Get().(*ResolvingPath) rp.vfs = vfs rp.root = pop.Root rp.mount = pop.Start.mount rp.start = pop.Start.dentry - rp.pit = path.Begin + rp.pit = pop.Path.Begin rp.flags = 0 if pop.FollowFinalSymlink { rp.flags |= rpflagsFollowFinalSymlink } - rp.mustBeDir = path.Dir - rp.mustBeDirOrig = path.Dir + rp.mustBeDir = pop.Path.Dir + rp.mustBeDirOrig = pop.Path.Dir rp.symlinks = 0 rp.curPart = 0 rp.numOrigParts = 1 rp.creds = creds - rp.parts[0] = path.Begin - rp.origParts[0] = path.Begin - return rp, nil + rp.parts[0] = pop.Path.Begin + rp.origParts[0] = pop.Path.Begin + return rp } func (vfs *VirtualFilesystem) putResolvingPath(rp *ResolvingPath) { @@ -274,7 +270,7 @@ func (rp *ResolvingPath) ResolveParent(d *Dentry) (*Dentry, error) { // ... of non-root mount. rp.nextMount = vd.mount rp.nextStart = vd.dentry - return nil, resolveMountRootError{} + return nil, resolveMountRootOrJumpError{} } // ... of root mount. parent = d @@ -345,29 +341,34 @@ func (rp *ResolvingPath) ShouldFollowSymlink() bool { // symlink target and returns nil. Otherwise it returns a non-nil error. // // Preconditions: !rp.Done(). +// +// Postconditions: If HandleSymlink returns a nil error, then !rp.Done(). func (rp *ResolvingPath) HandleSymlink(target string) error { if rp.symlinks >= linux.MaxSymlinkTraversals { return syserror.ELOOP } - targetPath, err := fspath.Parse(target) - if err != nil { - return err + if len(target) == 0 { + return syserror.ENOENT } rp.symlinks++ + targetPath := fspath.Parse(target) if targetPath.Absolute { rp.absSymlinkTarget = targetPath return resolveAbsSymlinkError{} } - if !targetPath.Begin.Ok() { - panic(fmt.Sprintf("symbolic link has non-empty target %q that is both relative and has no path components?", target)) - } // Consume the path component that represented the symlink. rp.Advance() // Prepend the symlink target to the relative path. + if checkInvariants { + if !targetPath.HasComponents() { + panic(fmt.Sprintf("non-empty pathname %q parsed to relative path with no components", target)) + } + } rp.relpathPrepend(targetPath) return nil } +// Preconditions: path.HasComponents(). func (rp *ResolvingPath) relpathPrepend(path fspath.Path) { if rp.pit.Ok() { rp.parts[rp.curPart] = rp.pit @@ -385,11 +386,32 @@ func (rp *ResolvingPath) relpathPrepend(path fspath.Path) { } } +// HandleJump is called when the current path component is a "magic" link to +// the given VirtualDentry, like /proc/[pid]/fd/[fd]. If the calling Filesystem +// method should continue path traversal, HandleMagicSymlink updates the path +// component stream to reflect the magic link target and returns nil. Otherwise +// it returns a non-nil error. +// +// Preconditions: !rp.Done(). +func (rp *ResolvingPath) HandleJump(target VirtualDentry) error { + if rp.symlinks >= linux.MaxSymlinkTraversals { + return syserror.ELOOP + } + rp.symlinks++ + // Consume the path component that represented the magic link. + rp.Advance() + // Unconditionally return a resolveMountRootOrJumpError, even if the Mount + // isn't changing, to force restarting at the new Dentry. + target.IncRef() + rp.nextMount = target.mount + rp.nextStart = target.dentry + return resolveMountRootOrJumpError{} +} + func (rp *ResolvingPath) handleError(err error) bool { switch err.(type) { - case resolveMountRootError: - // Switch to the new Mount. We hold references on the Mount and Dentry - // (from VFS.getMountpointAt()). + case resolveMountRootOrJumpError: + // Switch to the new Mount. We hold references on the Mount and Dentry. rp.decRefStartAndMount() rp.mount = rp.nextMount rp.start = rp.nextStart @@ -407,9 +429,8 @@ func (rp *ResolvingPath) handleError(err error) bool { return true case resolveMountPointError: - // Switch to the new Mount. We hold a reference on the Mount (from - // VFS.getMountAt()), but borrow the reference on the mount root from - // the Mount. + // Switch to the new Mount. We hold a reference on the Mount, but + // borrow the reference on the mount root from the Mount. rp.decRefStartAndMount() rp.mount = rp.nextMount rp.start = rp.nextMount.root @@ -447,6 +468,17 @@ func (rp *ResolvingPath) handleError(err error) bool { } } +// canHandleError returns true if err is an error returned by rp.Resolve*() +// that rp.handleError() may attempt to handle. +func (rp *ResolvingPath) canHandleError(err error) bool { + switch err.(type) { + case resolveMountRootOrJumpError, resolveMountPointError, resolveAbsSymlinkError: + return true + default: + return false + } +} + // MustBeDir returns true if the file traversed by rp must be a directory. func (rp *ResolvingPath) MustBeDir() bool { return rp.mustBeDir diff --git a/pkg/sentry/vfs/syscalls.go b/pkg/sentry/vfs/syscalls.go deleted file mode 100644 index 436151afa..000000000 --- a/pkg/sentry/vfs/syscalls.go +++ /dev/null @@ -1,237 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package vfs - -import ( - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/syserror" -) - -// PathOperation specifies the path operated on by a VFS method. -// -// PathOperation is passed to VFS methods by pointer to reduce memory copying: -// it's somewhat large and should never escape. (Options structs are passed by -// pointer to VFS and FileDescription methods for the same reason.) -type PathOperation struct { - // Root is the VFS root. References on Root are borrowed from the provider - // of the PathOperation. - // - // Invariants: Root.Ok(). - Root VirtualDentry - - // Start is the starting point for the path traversal. References on Start - // are borrowed from the provider of the PathOperation (i.e. the caller of - // the VFS method to which the PathOperation was passed). - // - // Invariants: Start.Ok(). If Pathname.Absolute, then Start == Root. - Start VirtualDentry - - // Path is the pathname traversed by this operation. - Pathname string - - // If FollowFinalSymlink is true, and the Dentry traversed by the final - // path component represents a symbolic link, the symbolic link should be - // followed. - FollowFinalSymlink bool -} - -// 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) { - rp, err := vfs.getResolvingPath(creds, pop) - if err != nil { - return VirtualDentry{}, err - } - for { - d, err := rp.mount.fs.impl.GetDentryAt(ctx, rp, *opts) - if err == nil { - vd := VirtualDentry{ - mount: rp.mount, - dentry: d, - } - rp.mount.IncRef() - vfs.putResolvingPath(rp) - return vd, nil - } - if !rp.handleError(err) { - vfs.putResolvingPath(rp) - return VirtualDentry{}, err - } - } -} - -// MkdirAt creates a directory at the given path. -func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MkdirOptions) error { - // "Under Linux, apart from the permission bits, the S_ISVTX mode bit is - // also honored." - mkdir(2) - opts.Mode &= 01777 - rp, err := vfs.getResolvingPath(creds, pop) - if err != nil { - return err - } - for { - err := rp.mount.fs.impl.MkdirAt(ctx, rp, *opts) - if err == nil { - vfs.putResolvingPath(rp) - return nil - } - if !rp.handleError(err) { - vfs.putResolvingPath(rp) - return err - } - } -} - -// MknodAt creates a file of the given mode at the given path. It returns an -// error from the syserror package. -func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MknodOptions) error { - rp, err := vfs.getResolvingPath(creds, pop) - if err != nil { - return nil - } - for { - if err = rp.mount.fs.impl.MknodAt(ctx, rp, *opts); err == nil { - vfs.putResolvingPath(rp) - return nil - } - // Handle mount traversals. - if !rp.handleError(err) { - vfs.putResolvingPath(rp) - return err - } - } -} - -// OpenAt returns a FileDescription providing access to the file at the given -// path. A reference is taken on the returned FileDescription. -func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) { - // Remove: - // - // - O_LARGEFILE, which we always report in FileDescription status flags - // since only 64-bit architectures are supported at this time. - // - // - O_CLOEXEC, which affects file descriptors and therefore must be - // handled outside of VFS. - // - // - Unknown flags. - opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE - // Linux's __O_SYNC (which we call linux.O_SYNC) implies O_DSYNC. - if opts.Flags&linux.O_SYNC != 0 { - opts.Flags |= linux.O_DSYNC - } - // Linux's __O_TMPFILE (which we call linux.O_TMPFILE) must be specified - // with O_DIRECTORY and a writable access mode (to ensure that it fails on - // filesystem implementations that do not support it). - if opts.Flags&linux.O_TMPFILE != 0 { - if opts.Flags&linux.O_DIRECTORY == 0 { - return nil, syserror.EINVAL - } - if opts.Flags&linux.O_CREAT != 0 { - return nil, syserror.EINVAL - } - if opts.Flags&linux.O_ACCMODE == linux.O_RDONLY { - return nil, syserror.EINVAL - } - } - // O_PATH causes most other flags to be ignored. - if opts.Flags&linux.O_PATH != 0 { - opts.Flags &= linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_PATH - } - // "On Linux, the following bits are also honored in mode: [S_ISUID, - // S_ISGID, S_ISVTX]" - open(2) - opts.Mode &= 07777 - - if opts.Flags&linux.O_NOFOLLOW != 0 { - pop.FollowFinalSymlink = false - } - rp, err := vfs.getResolvingPath(creds, pop) - if err != nil { - return nil, err - } - if opts.Flags&linux.O_DIRECTORY != 0 { - rp.mustBeDir = true - rp.mustBeDirOrig = true - } - for { - fd, err := rp.mount.fs.impl.OpenAt(ctx, rp, *opts) - if err == nil { - vfs.putResolvingPath(rp) - return fd, nil - } - if !rp.handleError(err) { - vfs.putResolvingPath(rp) - return nil, err - } - } -} - -// StatAt returns metadata for the file at the given path. -func (vfs *VirtualFilesystem) StatAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *StatOptions) (linux.Statx, error) { - rp, err := vfs.getResolvingPath(creds, pop) - if err != nil { - return linux.Statx{}, err - } - for { - stat, err := rp.mount.fs.impl.StatAt(ctx, rp, *opts) - if err == nil { - vfs.putResolvingPath(rp) - return stat, nil - } - if !rp.handleError(err) { - vfs.putResolvingPath(rp) - return linux.Statx{}, err - } - } -} - -// StatusFlags returns file description status flags. -func (fd *FileDescription) StatusFlags(ctx context.Context) (uint32, error) { - flags, err := fd.impl.StatusFlags(ctx) - flags |= linux.O_LARGEFILE - return flags, err -} - -// SetStatusFlags sets file description status flags. -func (fd *FileDescription) SetStatusFlags(ctx context.Context, flags uint32) error { - return fd.impl.SetStatusFlags(ctx, flags) -} - -// TODO: -// -// - VFS.SyncAllFilesystems() for sync(2) -// -// - Something for syncfs(2) -// -// - VFS.LinkAt() -// -// - VFS.ReadlinkAt() -// -// - VFS.RenameAt() -// -// - VFS.RmdirAt() -// -// - VFS.SetStatAt() -// -// - VFS.StatFSAt() -// -// - VFS.SymlinkAt() -// -// - VFS.UmountAt() -// -// - VFS.UnlinkAt() -// -// - FileDescription.(almost everything) diff --git a/pkg/sentry/vfs/testutil.go b/pkg/sentry/vfs/testutil.go index 593144cb7..ee5c8b9e2 100644 --- a/pkg/sentry/vfs/testutil.go +++ b/pkg/sentry/vfs/testutil.go @@ -15,7 +15,10 @@ package vfs import ( + "fmt" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/syserror" @@ -54,6 +57,11 @@ func (fs *FDTestFilesystem) GetDentryAt(ctx context.Context, rp *ResolvingPath, return nil, syserror.EPERM } +// GetParentDentryAt implements FilesystemImpl.GetParentDentryAt. +func (fs *FDTestFilesystem) GetParentDentryAt(ctx context.Context, rp *ResolvingPath) (*Dentry, error) { + return nil, syserror.EPERM +} + // LinkAt implements FilesystemImpl.LinkAt. func (fs *FDTestFilesystem) LinkAt(ctx context.Context, rp *ResolvingPath, vd VirtualDentry) error { return syserror.EPERM @@ -80,7 +88,7 @@ func (fs *FDTestFilesystem) ReadlinkAt(ctx context.Context, rp *ResolvingPath) ( } // RenameAt implements FilesystemImpl.RenameAt. -func (fs *FDTestFilesystem) RenameAt(ctx context.Context, rp *ResolvingPath, vd VirtualDentry, opts RenameOptions) error { +func (fs *FDTestFilesystem) RenameAt(ctx context.Context, rp *ResolvingPath, oldParentVD VirtualDentry, oldName string, opts RenameOptions) error { return syserror.EPERM } @@ -114,6 +122,32 @@ func (fs *FDTestFilesystem) UnlinkAt(ctx context.Context, rp *ResolvingPath) err return syserror.EPERM } +// ListxattrAt implements FilesystemImpl.ListxattrAt. +func (fs *FDTestFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath) ([]string, error) { + return nil, syserror.EPERM +} + +// GetxattrAt implements FilesystemImpl.GetxattrAt. +func (fs *FDTestFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, name string) (string, error) { + return "", syserror.EPERM +} + +// SetxattrAt implements FilesystemImpl.SetxattrAt. +func (fs *FDTestFilesystem) SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error { + return syserror.EPERM +} + +// RemovexattrAt implements FilesystemImpl.RemovexattrAt. +func (fs *FDTestFilesystem) RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error { + return syserror.EPERM +} + +// PrependPath implements FilesystemImpl.PrependPath. +func (fs *FDTestFilesystem) PrependPath(ctx context.Context, vfsroot, vd VirtualDentry, b *fspath.Builder) error { + b.PrependComponent(fmt.Sprintf("vfs.fdTestDentry:%p", vd.dentry.impl.(*fdTestDentry))) + return PrependPathSyntheticError{} +} + type fdTestDentry struct { vfsd Dentry } diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index f0cd3ffe5..ea2db7031 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -20,6 +20,7 @@ // VirtualFilesystem.mountMu // Dentry.mu // Locks acquired by FilesystemImpls between Prepare{Delete,Rename}Dentry and Commit{Delete,Rename*}Dentry +// VirtualFilesystem.filesystemsMu // VirtualFilesystem.fsTypesMu // // Locking Dentry.mu in multiple Dentries requires holding @@ -27,7 +28,14 @@ package vfs import ( + "fmt" "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" ) // A VirtualFilesystem (VFS for short) combines Filesystems in trees of Mounts. @@ -67,22 +75,598 @@ type VirtualFilesystem struct { // mountpoints is analogous to Linux's mountpoint_hashtable. mountpoints map[*Dentry]map[*Mount]struct{} - // fsTypes contains all FilesystemTypes that are usable in the - // VirtualFilesystem. fsTypes is protected by fsTypesMu. + // devices contains all registered Devices. devices is protected by + // devicesMu. + devicesMu sync.RWMutex + devices map[devTuple]*registeredDevice + + // fsTypes contains all registered FilesystemTypes. fsTypes is protected by + // fsTypesMu. fsTypesMu sync.RWMutex - fsTypes map[string]FilesystemType + fsTypes map[string]*registeredFilesystemType + + // filesystems contains all Filesystems. filesystems is protected by + // filesystemsMu. + filesystemsMu sync.Mutex + filesystems map[*Filesystem]struct{} } // New returns a new VirtualFilesystem with no mounts or FilesystemTypes. func New() *VirtualFilesystem { vfs := &VirtualFilesystem{ mountpoints: make(map[*Dentry]map[*Mount]struct{}), - fsTypes: make(map[string]FilesystemType), + devices: make(map[devTuple]*registeredDevice), + fsTypes: make(map[string]*registeredFilesystemType), + filesystems: make(map[*Filesystem]struct{}), } vfs.mounts.Init() return vfs } +// PathOperation specifies the path operated on by a VFS method. +// +// PathOperation is passed to VFS methods by pointer to reduce memory copying: +// it's somewhat large and should never escape. (Options structs are passed by +// pointer to VFS and FileDescription methods for the same reason.) +type PathOperation struct { + // Root is the VFS root. References on Root are borrowed from the provider + // of the PathOperation. + // + // Invariants: Root.Ok(). + Root VirtualDentry + + // Start is the starting point for the path traversal. References on Start + // are borrowed from the provider of the PathOperation (i.e. the caller of + // the VFS method to which the PathOperation was passed). + // + // Invariants: Start.Ok(). If Path.Absolute, then Start == Root. + Start VirtualDentry + + // Path is the pathname traversed by this operation. + Path fspath.Path + + // If FollowFinalSymlink is true, and the Dentry traversed by the final + // path component represents a symbolic link, the symbolic link should be + // followed. + FollowFinalSymlink bool +} + +// 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) { + rp := vfs.getResolvingPath(creds, pop) + for { + d, err := rp.mount.fs.impl.GetDentryAt(ctx, rp, *opts) + if err == nil { + vd := VirtualDentry{ + mount: rp.mount, + dentry: d, + } + rp.mount.IncRef() + vfs.putResolvingPath(rp) + return vd, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return VirtualDentry{}, err + } + } +} + +// Preconditions: pop.Path.Begin.Ok(). +func (vfs *VirtualFilesystem) getParentDirAndName(ctx context.Context, creds *auth.Credentials, pop *PathOperation) (VirtualDentry, string, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + parent, err := rp.mount.fs.impl.GetParentDentryAt(ctx, rp) + if err == nil { + parentVD := VirtualDentry{ + mount: rp.mount, + dentry: parent, + } + rp.mount.IncRef() + name := rp.Component() + vfs.putResolvingPath(rp) + return parentVD, name, nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.GetParentDentryAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return VirtualDentry{}, "", err + } + } +} + +// LinkAt creates a hard link at newpop representing the existing file at +// oldpop. +func (vfs *VirtualFilesystem) LinkAt(ctx context.Context, creds *auth.Credentials, oldpop, newpop *PathOperation) error { + oldVD, err := vfs.GetDentryAt(ctx, creds, oldpop, &GetDentryOptions{}) + if err != nil { + return err + } + + if !newpop.Path.Begin.Ok() { + oldVD.DecRef() + if newpop.Path.Absolute { + return syserror.EEXIST + } + return syserror.ENOENT + } + if newpop.FollowFinalSymlink { + oldVD.DecRef() + ctx.Warningf("VirtualFilesystem.LinkAt: file creation paths can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, newpop) + for { + err := rp.mount.fs.impl.LinkAt(ctx, rp, oldVD) + if err == nil { + vfs.putResolvingPath(rp) + oldVD.DecRef() + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.LinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + oldVD.DecRef() + return err + } + } +} + +// MkdirAt creates a directory at the given path. +func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MkdirOptions) error { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return syserror.EEXIST + } + return syserror.ENOENT + } + if pop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.MkdirAt: file creation paths can't follow final symlink") + return syserror.EINVAL + } + // "Under Linux, apart from the permission bits, the S_ISVTX mode bit is + // also honored." - mkdir(2) + opts.Mode &= 0777 | linux.S_ISVTX + + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.MkdirAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.MkdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// MknodAt creates a file of the given mode at the given path. It returns an +// error from the syserror package. +func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MknodOptions) error { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return syserror.EEXIST + } + return syserror.ENOENT + } + if pop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.MknodAt: file creation paths can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.MknodAt(ctx, rp, *opts) + if err != nil { + vfs.putResolvingPath(rp) + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.MknodAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// OpenAt returns a FileDescription providing access to the file at the given +// path. A reference is taken on the returned FileDescription. +func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) { + // Remove: + // + // - O_LARGEFILE, which we always report in FileDescription status flags + // since only 64-bit architectures are supported at this time. + // + // - O_CLOEXEC, which affects file descriptors and therefore must be + // handled outside of VFS. + // + // - Unknown flags. + opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_NOCTTY | linux.O_TRUNC | linux.O_APPEND | linux.O_NONBLOCK | linux.O_DSYNC | linux.O_ASYNC | linux.O_DIRECT | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NOATIME | linux.O_SYNC | linux.O_PATH | linux.O_TMPFILE + // Linux's __O_SYNC (which we call linux.O_SYNC) implies O_DSYNC. + if opts.Flags&linux.O_SYNC != 0 { + opts.Flags |= linux.O_DSYNC + } + // Linux's __O_TMPFILE (which we call linux.O_TMPFILE) must be specified + // with O_DIRECTORY and a writable access mode (to ensure that it fails on + // filesystem implementations that do not support it). + if opts.Flags&linux.O_TMPFILE != 0 { + if opts.Flags&linux.O_DIRECTORY == 0 { + return nil, syserror.EINVAL + } + if opts.Flags&linux.O_CREAT != 0 { + return nil, syserror.EINVAL + } + if opts.Flags&linux.O_ACCMODE == linux.O_RDONLY { + return nil, syserror.EINVAL + } + } + // O_PATH causes most other flags to be ignored. + if opts.Flags&linux.O_PATH != 0 { + opts.Flags &= linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_PATH + } + // "On Linux, the following bits are also honored in mode: [S_ISUID, + // S_ISGID, S_ISVTX]" - open(2) + opts.Mode &= 0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX + + if opts.Flags&linux.O_NOFOLLOW != 0 { + pop.FollowFinalSymlink = false + } + rp := vfs.getResolvingPath(creds, pop) + if opts.Flags&linux.O_DIRECTORY != 0 { + rp.mustBeDir = true + rp.mustBeDirOrig = true + } + for { + fd, err := rp.mount.fs.impl.OpenAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return fd, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return nil, err + } + } +} + +// ReadlinkAt returns the target of the symbolic link at the given path. +func (vfs *VirtualFilesystem) ReadlinkAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) (string, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + target, err := rp.mount.fs.impl.ReadlinkAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return target, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return "", err + } + } +} + +// RenameAt renames the file at oldpop to newpop. +func (vfs *VirtualFilesystem) RenameAt(ctx context.Context, creds *auth.Credentials, oldpop, newpop *PathOperation, opts *RenameOptions) error { + if !oldpop.Path.Begin.Ok() { + if oldpop.Path.Absolute { + return syserror.EBUSY + } + return syserror.ENOENT + } + if oldpop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.RenameAt: source path can't follow final symlink") + return syserror.EINVAL + } + + oldParentVD, oldName, err := vfs.getParentDirAndName(ctx, creds, oldpop) + if err != nil { + return err + } + if oldName == "." || oldName == ".." { + oldParentVD.DecRef() + return syserror.EBUSY + } + + if !newpop.Path.Begin.Ok() { + oldParentVD.DecRef() + if newpop.Path.Absolute { + return syserror.EBUSY + } + return syserror.ENOENT + } + if newpop.FollowFinalSymlink { + oldParentVD.DecRef() + ctx.Warningf("VirtualFilesystem.RenameAt: destination path can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, newpop) + renameOpts := *opts + if oldpop.Path.Dir { + renameOpts.MustBeDir = true + } + for { + err := rp.mount.fs.impl.RenameAt(ctx, rp, oldParentVD, oldName, renameOpts) + if err == nil { + vfs.putResolvingPath(rp) + oldParentVD.DecRef() + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.RenameAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + oldParentVD.DecRef() + return err + } + } +} + +// RmdirAt removes the directory at the given path. +func (vfs *VirtualFilesystem) RmdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) error { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return syserror.EBUSY + } + return syserror.ENOENT + } + if pop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.RmdirAt: file deletion paths can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.RmdirAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.RmdirAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// SetStatAt changes metadata for the file at the given path. +func (vfs *VirtualFilesystem) SetStatAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *SetStatOptions) error { + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.SetStatAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// StatAt returns metadata for the file at the given path. +func (vfs *VirtualFilesystem) StatAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *StatOptions) (linux.Statx, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + stat, err := rp.mount.fs.impl.StatAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return stat, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return linux.Statx{}, err + } + } +} + +// StatFSAt returns metadata for the filesystem containing the file at the +// given path. +func (vfs *VirtualFilesystem) StatFSAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) (linux.Statfs, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + statfs, err := rp.mount.fs.impl.StatFSAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return statfs, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return linux.Statfs{}, err + } + } +} + +// SymlinkAt creates a symbolic link at the given path with the given target. +func (vfs *VirtualFilesystem) SymlinkAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, target string) error { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return syserror.EEXIST + } + return syserror.ENOENT + } + if pop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.SymlinkAt: file creation paths can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.SymlinkAt(ctx, rp, target) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.SymlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// UnlinkAt deletes the non-directory file at the given path. +func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) error { + if !pop.Path.Begin.Ok() { + if pop.Path.Absolute { + return syserror.EBUSY + } + return syserror.ENOENT + } + if pop.FollowFinalSymlink { + ctx.Warningf("VirtualFilesystem.UnlinkAt: file deletion paths can't follow final symlink") + return syserror.EINVAL + } + + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.UnlinkAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if checkInvariants { + if rp.canHandleError(err) && rp.Done() { + panic(fmt.Sprintf("%T.UnlinkAt() consumed all path components and returned %T", rp.mount.fs.impl, err)) + } + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// ListxattrAt returns all extended attribute names for the file at the given +// path. +func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) ([]string, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return names, nil + } + if err == syserror.ENOTSUP { + // Linux doesn't actually return ENOTSUP in this case; instead, + // fs/xattr.c:vfs_listxattr() falls back to allowing the security + // subsystem to return security extended attributes, which by + // default don't exist. + vfs.putResolvingPath(rp) + return nil, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return nil, err + } + } +} + +// GetxattrAt returns the value associated with the given extended attribute +// for the file at the given path. +func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) (string, error) { + rp := vfs.getResolvingPath(creds, pop) + for { + val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, name) + if err == nil { + vfs.putResolvingPath(rp) + return val, nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return "", err + } + } +} + +// SetxattrAt changes the value associated with the given extended attribute +// for the file at the given path. +func (vfs *VirtualFilesystem) SetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *SetxattrOptions) error { + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.SetxattrAt(ctx, rp, *opts) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// RemovexattrAt removes the given extended attribute from the file at rp. +func (vfs *VirtualFilesystem) RemovexattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) error { + rp := vfs.getResolvingPath(creds, pop) + for { + err := rp.mount.fs.impl.RemovexattrAt(ctx, rp, name) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + vfs.putResolvingPath(rp) + return err + } + } +} + +// SyncAllFilesystems has the semantics of Linux's sync(2). +func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error { + fss := make(map[*Filesystem]struct{}) + vfs.filesystemsMu.Lock() + for fs := range vfs.filesystems { + if !fs.TryIncRef() { + continue + } + fss[fs] = struct{}{} + } + vfs.filesystemsMu.Unlock() + var retErr error + for fs := range fss { + if err := fs.impl.Sync(ctx); err != nil && retErr == nil { + retErr = err + } + fs.DecRef() + } + return retErr +} + // A VirtualDentry represents a node in a VFS tree, by combining a Dentry // (which represents a node in a Filesystem's tree) and a Mount (which // represents the Filesystem's position in a VFS mount tree). diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index ecce6c69f..5e4611333 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -287,7 +287,9 @@ func (w *Watchdog) runTurn() { if !ok { // New stuck task detected. // - // TODO(b/65849403): Tasks blocked doing IO may be considered stuck in kernel. + // Note that tasks blocked doing IO may be considered stuck in kernel, + // unless they are surrounded b + // Task.UninterruptibleSleepStart/Finish. tc = &offender{lastUpdateTime: lastUpdateTime} stuckTasks.Increment() newTaskFound = true |