diff options
Diffstat (limited to 'pkg/sentry')
101 files changed, 5360 insertions, 635 deletions
diff --git a/pkg/sentry/arch/arch_amd64.go b/pkg/sentry/arch/arch_amd64.go index 9e7db8b30..67daa6c24 100644 --- a/pkg/sentry/arch/arch_amd64.go +++ b/pkg/sentry/arch/arch_amd64.go @@ -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/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/task.go b/pkg/sentry/fs/proc/task.go index 2a598149d..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. 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..880b7bcd3 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", diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go index 94cd74095..177ce2cb9 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go +++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go @@ -81,7 +81,11 @@ 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() { diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go index 307e4d68c..e9f756732 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -147,55 +147,54 @@ func TestSeek(t *testing.T) { 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) } } @@ -456,7 +455,7 @@ 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 +463,7 @@ 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) } @@ -574,7 +573,7 @@ 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/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index 2d15e8aaf..e7aa3b41b 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" @@ -441,3 +442,10 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } + +// 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/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD new file mode 100644 index 000000000..52596c090 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -0,0 +1,60 @@ +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", + ], + 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/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..30c06baf0 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -0,0 +1,131 @@ +// 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" + "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 initialized with Init before first use. +type DynamicBytesFile struct { + InodeAttrs + InodeNoopRefCount + InodeNotDirectory + InodeNotSymlink + + data vfs.DynamicBytesSource +} + +// Init intializes a dynamic bytes file. +func (f *DynamicBytesFile) Init(creds *auth.Credentials, ino uint64, data vfs.DynamicBytesSource) { + f.InodeAttrs.Init(creds, ino, linux.ModeRegular|0444) + 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. +type DynamicBytesFD struct { + vfs.FileDescriptionDefaultImpl + vfs.DynamicBytesFileDescriptionImpl + + vfsfd vfs.FileDescription + inode Inode + flags uint32 +} + +// Init initializes a DynamicBytesFD. +func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *vfs.Dentry, data vfs.DynamicBytesSource, flags uint32) { + m.IncRef() // DecRef in vfs.FileDescription.vd.DecRef on final ref. + d.IncRef() // DecRef in vfs.FileDescription.vd.DecRef on final ref. + fd.flags = flags + fd.inode = d.Impl().(*Dentry).inode + fd.SetDataSource(data) + fd.vfsfd.Init(fd, m, d) +} + +// 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 +} + +// StatusFlags implements vfs.FileDescriptionImpl.StatusFlags. +func (fd *DynamicBytesFD) StatusFlags(ctx context.Context) (uint32, error) { + return fd.flags, nil +} + +// SetStatusFlags implements vfs.FileDescriptionImpl.SetStatusFlags. +func (fd *DynamicBytesFD) 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 +} 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..d6c18937a --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -0,0 +1,207 @@ +// 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 + flags uint32 + off int64 +} + +// Init initializes a GenericDirectoryFD. +func (fd *GenericDirectoryFD) Init(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, flags uint32) { + m.IncRef() // DecRef in vfs.FileDescription.vd.DecRef on final ref. + d.IncRef() // DecRef in vfs.FileDescription.vd.DecRef on final ref. + fd.children = children + fd.flags = flags + fd.vfsfd.Init(fd, m, d) +} + +// 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++ + } + + return nil +} + +// 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 +} + +// StatusFlags implements vfs.FileDescriptionImpl.StatusFlags. +func (fd *GenericDirectoryFD) StatusFlags(ctx context.Context) (uint32, error) { + return fd.flags, nil +} + +// SetStatusFlags implements vfs.FileDescriptionImpl.SetStatusFlags. +func (fd *GenericDirectoryFD) 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 *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..db486b6c1 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -0,0 +1,691 @@ +// 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: + d.dirMu.Lock() + nextVFSD, err := rp.ResolveComponent(vfsd) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + if nextVFSD != nil { + // Cached dentry exists, revalidate. + next := nextVFSD.Impl().(*Dentry) + if !next.inode.Valid(ctx) { + d.dirMu.Lock() + rp.VirtualFilesystem().ForceDeleteDentry(nextVFSD) + d.dirMu.Unlock() + fs.deferDecRef(nextVFSD) // Reference from Lookup. + nextVFSD = nil + } + } + if nextVFSD == nil { + // Dentry isn't cached; it either doesn't exist or failed + // revalidation. Attempt to resolve it via Lookup. + name := rp.Component() + var err error + nextVFSD, err = d.inode.Lookup(ctx, name) + // Reference on nextVFSD dropped by a corresponding Valid. + if err != nil { + return nil, err + } + d.InsertChild(name, nextVFSD) + } + next := nextVFSD.Impl().(*Dentry) + + // Resolve any symlink at current path component. + if rp.ShouldFollowSymlink() && d.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 nextVFSD, 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 +} + +// 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, vd vfs.VirtualDentry, opts vfs.RenameOptions) error { + noReplace := opts.Flags&linux.RENAME_NOREPLACE != 0 + exchange := opts.Flags&linux.RENAME_EXCHANGE != 0 + whiteout := opts.Flags&linux.RENAME_WHITEOUT != 0 + if exchange && (noReplace || whiteout) { + // Can't specify RENAME_NOREPLACE or RENAME_WHITEOUT with RENAME_EXCHANGE. + return syserror.EINVAL + } + if exchange || whiteout { + // Exchange and Whiteout flags are not supported on kernfs. + return syserror.EINVAL + } + + fs.mu.Lock() + defer fs.mu.Lock() + + mnt := rp.Mount() + if mnt != vd.Mount() { + return syserror.EXDEV + } + + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + + dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) + fs.processDeferredDecRefsLocked() + if err != nil { + return err + } + + srcVFSD := vd.Dentry() + srcDirVFSD := srcVFSD.Parent() + + // 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 +} + +// 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..7b45b702a --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -0,0 +1,492 @@ +// 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") +} + +// 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 +} + +// 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 +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go new file mode 100644 index 000000000..bb01c3d01 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -0,0 +1,405 @@ +// 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) { + 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. + d.dirMu.Lock() + vfsDentry.InsertChild(child, name) + d.dirMu.Unlock() +} + +// 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 +} + +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..f78bb7b04 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -0,0 +1,423 @@ +// 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/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}) + 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, + Pathname: 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) + + 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/memfs/BUILD b/pkg/sentry/fsimpl/memfs/BUILD index bc5c0b591..0cc751eb8 100644 --- a/pkg/sentry/fsimpl/memfs/BUILD +++ b/pkg/sentry/fsimpl/memfs/BUILD @@ -32,6 +32,7 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/fspath", "//pkg/sentry/arch", "//pkg/sentry/context", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/memfs/benchmark_test.go b/pkg/sentry/fsimpl/memfs/benchmark_test.go index ea6417ce7..4a7a94a52 100644 --- a/pkg/sentry/fsimpl/memfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/memfs/benchmark_test.go @@ -394,7 +394,7 @@ 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, "memfs", &vfs.MountOptions{}); err != nil { b.Fatalf("failed to mount tmpfs submount: %v", err) } filePathBuilder.WriteString(mountPointName) diff --git a/pkg/sentry/fsimpl/memfs/filesystem.go b/pkg/sentry/fsimpl/memfs/filesystem.go index 08a9cb8ef..1f2a5122a 100644 --- a/pkg/sentry/fsimpl/memfs/filesystem.go +++ b/pkg/sentry/fsimpl/memfs/filesystem.go @@ -19,6 +19,7 @@ import ( "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" @@ -582,3 +583,10 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error inode.decLinksLocked() return nil } + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/memfs/pipe_test.go b/pkg/sentry/fsimpl/memfs/pipe_test.go index a3a870571..5bf527c80 100644 --- a/pkg/sentry/fsimpl/memfs/pipe_test.go +++ b/pkg/sentry/fsimpl/memfs/pipe_test.go @@ -194,7 +194,7 @@ 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 +207,7 @@ 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 +220,7 @@ 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/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index c46e05c3a..0d87be52b 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -229,6 +229,10 @@ func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { 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 } diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 28ba950bd..bd3fb4c03 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -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 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/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..ab0c6c4aa 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. // @@ -528,14 +537,6 @@ 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 +550,7 @@ 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..3eadfedb4 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -299,6 +299,7 @@ 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..90a6190f1 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() { @@ -253,7 +254,7 @@ 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..d97f8c189 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" @@ -205,9 +206,11 @@ 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 +228,7 @@ 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 +236,12 @@ 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 +257,10 @@ 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 +268,12 @@ 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) diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index ae6fc4025..3522a4ae5 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -154,10 +154,10 @@ 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/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/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD index 6803d488c..5c52d4007 100644 --- a/pkg/sentry/platform/kvm/BUILD +++ b/pkg/sentry/platform/kvm/BUILD @@ -12,18 +12,26 @@ 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", "virtual_map.go", 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/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go new file mode 100644 index 000000000..e5fac0d6a --- /dev/null +++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go @@ -0,0 +1,28 @@ +// 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 ( + "unsafe" + + "gvisor.dev/gvisor/pkg/sentry/arch" +) + +//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/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/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index 4649a94a7..606dc2b1d 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" ) @@ -143,3 +145,49 @@ func (t *thread) adjustInitRegsRip() { func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.R15 = uint64(ppid) } + +// 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..62a686ee7 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. @@ -124,3 +128,36 @@ func (t *thread) adjustInitRegsRip() { func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.Regs[7] = uint64(ppid) } + +// 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..cf13ea5e4 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" @@ -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/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index 3f3ab5cfb..8bcfe1032 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -19,14 +19,14 @@ package ring0 // CPACREL1 returns the value of the CPACR_EL1 register. func CPACREL1() (value uintptr) -// GetFPCR returns the value of FPCR register. -func GetFPCR() (value uintptr) +// FPCR returns the value of FPCR register. +func FPCR() (value uintptr) // SetFPCR writes the FPCR value. func SetFPCR(value uintptr) -// GetFPSR returns the value of FPSR register. -func GetFPSR() (value uintptr) +// FPSR returns the value of FPSR register. +func FPSR() (value uintptr) // SetFPSR writes the FPSR value. func SetFPSR(value uintptr) diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s index 7c96e8e9c..1c9171004 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.s +++ b/pkg/sentry/platform/ring0/lib_arm64.s @@ -17,7 +17,7 @@ TEXT ·CPACREL1(SB),NOSPLIT,$0-8 MOVD R1, ret+0(FP) RET -TEXT ·GetFPCR(SB),NOSPLIT,$0-8 +TEXT ·FPCR(SB),NOSPLIT,$0-8 WORD $0xd53b4201 // MRS NZCV, R1 MOVD R1, ret+0(FP) RET @@ -27,7 +27,7 @@ TEXT ·GetFPSR(SB),NOSPLIT,$0-8 MOVD R1, ret+0(FP) RET -TEXT ·SetFPCR(SB),NOSPLIT,$0-8 +TEXT ·FPCR(SB),NOSPLIT,$0-8 MOVD addr+0(FP), R1 WORD $0xd51b4201 // MSR R1, NZCV RET 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/control.go b/pkg/sentry/socket/control/control.go index 782a3cb92..af1a4e95f 100644 --- a/pkg/sentry/socket/control/control.go +++ b/pkg/sentry/socket/control/control.go @@ -195,15 +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 @@ -240,12 +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) @@ -348,43 +348,62 @@ func PackTClass(t *kernel.Task, tClass int32, buf []byte) []byte { ) } -func addSpaceForCmsg(cmsgDataLen int, buf []byte) []byte { - newBuf := make([]byte, 0, len(buf)+linux.SizeOfControlMessageHeader+cmsgDataLen) - return append(newBuf, buf...) -} - -// PackControlMessages converts the given ControlMessages struct into a buffer. +// PackControlMessages packs control messages into the given buffer. +// // We skip control messages specific to Unix domain sockets. -func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages) []byte { - var buf []byte - // The use of t.Arch().Width() is analogous to Linux's use of sizeof(long) in - // CMSG_ALIGN. - width := t.Arch().Width() - +// +// 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 = addSpaceForCmsg(int(width), buf) buf = PackTimestamp(t, cmsgs.IP.Timestamp, buf) } if cmsgs.IP.HasInq { // In Linux, TCP_CM_INQ is added after SO_TIMESTAMP. - buf = addSpaceForCmsg(AlignUp(linux.SizeOfControlMessageInq, width), buf) buf = PackInq(t, cmsgs.IP.Inq, buf) } if cmsgs.IP.HasTOS { - buf = addSpaceForCmsg(AlignUp(linux.SizeOfControlMessageTOS, width), buf) buf = PackTOS(t, cmsgs.IP.TOS, buf) } if cmsgs.IP.HasTClass { - buf = addSpaceForCmsg(AlignUp(linux.SizeOfControlMessageTClass, width), buf) 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) (socket.ControlMessages, error) { var ( diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index a8c152b54..c957b0f1d 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -45,7 +45,7 @@ const ( sizeofSockaddr = syscall.SizeofSockaddrInet6 // sizeof(sockaddr_in6) > sizeof(sockaddr_in) // maxControlLen is the maximum size of a control message buffer used in a - // recvmsg syscall. + // recvmsg or sendmsg syscall. maxControlLen = 1024 ) @@ -412,9 +412,12 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags msg.Namelen = uint32(len(senderAddrBuf)) } if controlLen > 0 { - controlBuf = make([]byte, maxControlLen) + if controlLen > maxControlLen { + controlLen = maxControlLen + } + controlBuf = make([]byte, controlLen) msg.Control = &controlBuf[0] - msg.Controllen = maxControlLen + msg.Controllen = controlLen } n, err := recvmsg(s.fd, &msg, sysflags) if err != nil { @@ -489,7 +492,14 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return 0, syserr.ErrInvalidArgument } - controlBuf := control.PackControlMessages(t, controlMessages) + 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() { 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/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..dea309fda --- /dev/null +++ b/pkg/sentry/strace/select.go @@ -0,0 +1,53 @@ +// 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 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/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..6766ba587 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -49,6 +49,7 @@ 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..272ae9991 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.ErrorWithEvent("rseq", syserror.ENOSYS, "", 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 f1dd4b0c0..3b584eed9 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), @@ -302,7 +302,26 @@ var ARM64 = &kernel.SyscallTable{ 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.ErrorWithEvent("rseq", syserror.ENOSYS, "", 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 3b9181002..9bc2445a5 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -840,25 +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. @@ -901,11 +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 @@ -1008,6 +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 @@ -1035,7 +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 ( 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_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index d8acae063..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 { @@ -784,32 +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.IP.HasTOS { - controlData = control.PackTOS(t, cms.IP.TOS, controlData) - } - - if cms.IP.HasTClass { - controlData = control.PackTClass(t, cms.IP.TClass, controlData) - } - if cms.Unix.Rights != nil { controlData, mflags = control.PackRights(t, cms.Unix.Rights.(control.SCMRights), flags&linux.MSG_CMSG_CLOEXEC != 0, controlData, mflags) } @@ -885,7 +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) } @@ -1071,7 +1055,7 @@ func sendSingleMsg(t *kernel.Task, s socket.Socket, file *fs.File, msgPtr userme 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.Unix.Release() + controlMessages.Release() } return uintptr(n), err } 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..e3e554b88 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -17,9 +17,9 @@ 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/file_description.go b/pkg/sentry/vfs/file_description.go index 34007eb57..6575afd16 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -241,3 +241,140 @@ 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) +} + +// StatusFlags returns file description status flags, as for fcntl(F_GETFL). +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, as for fcntl(F_SETFL). +func (fd *FileDescription) SetStatusFlags(ctx context.Context, flags uint32) error { + return fd.impl.SetStatusFlags(ctx, flags) +} + +// Stat returns metadata for the file represented by fd. +func (fd *FileDescription) Stat(ctx context.Context, opts StatOptions) (linux.Statx, error) { + 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 { + 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) { + 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) +} + +// 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.impl.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.impl.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.impl.Sync(ctx) +} diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index 4fbad7840..aae023254 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -252,3 +252,12 @@ 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..ac7799296 100644 --- a/pkg/sentry/vfs/file_description_impl_util_test.go +++ b/pkg/sentry/vfs/file_description_impl_util_test.go @@ -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..8011eba3f 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") @@ -163,5 +186,56 @@ type FilesystemImpl interface { // UnlinkAt removes the non-directory file at rp. UnlinkAt(ctx context.Context, rp *ResolvingPath) error - // TODO: d_path(); extended attributes; inotify_add_watch(); bind() + // 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: extended attributes; 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/mount.go b/pkg/sentry/vfs/mount.go index 1c3b2e987..ec23ab0dd 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" @@ -133,13 +134,13 @@ 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 { +// 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 { fsType := vfs.getFilesystemType(fsTypeName) if fsType == nil { return syserror.ENODEV } - fs, root, err := fsType.GetFilesystem(ctx, vfs, creds, source, *opts) + fs, root, err := fsType.GetFilesystem(ctx, vfs, creds, source, opts.GetFilesystemOptions) if err != nil { return err } @@ -207,6 +208,68 @@ 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..3ecbc8fc1 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -46,6 +46,12 @@ type MknodOptions struct { DevMinor uint32 } +// MountOptions contains options to VirtualFilesystem.MountAt(). +type MountOptions struct { + // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). + GetFilesystemOptions GetFilesystemOptions +} + // OpenOptions contains options to VirtualFilesystem.OpenAt() and // FilesystemImpl.OpenAt(). type OpenOptions struct { @@ -114,6 +120,12 @@ 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/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..7a1d9e383 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" @@ -114,6 +117,12 @@ func (fs *FDTestFilesystem) UnlinkAt(ctx context.Context, rp *ResolvingPath) err 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..7262b0d0a 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 @@ -28,6 +29,11 @@ package vfs import ( "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "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. @@ -67,6 +73,11 @@ type VirtualFilesystem struct { // mountpoints is analogous to Linux's mountpoint_hashtable. mountpoints map[*Dentry]map[*Mount]struct{} + // filesystems contains all Filesystems. filesystems is protected by + // filesystemsMu. + filesystemsMu sync.Mutex + filesystems map[*Filesystem]struct{} + // fsTypes contains all FilesystemTypes that are usable in the // VirtualFilesystem. fsTypes is protected by fsTypesMu. fsTypesMu sync.RWMutex @@ -77,12 +88,379 @@ type VirtualFilesystem struct { func New() *VirtualFilesystem { vfs := &VirtualFilesystem{ mountpoints: make(map[*Dentry]map[*Mount]struct{}), + filesystems: make(map[*Filesystem]struct{}), fsTypes: make(map[string]FilesystemType), } 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 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 + } + } +} + +// 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 + } + rp, err := vfs.getResolvingPath(creds, newpop) + if err != nil { + oldVD.DecRef() + return err + } + for { + err := rp.mount.fs.impl.LinkAt(ctx, rp, oldVD) + if err == nil { + oldVD.DecRef() + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + oldVD.DecRef() + vfs.putResolvingPath(rp) + 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 { + // "Under Linux, apart from the permission bits, the S_ISVTX mode bit is + // also honored." - mkdir(2) + opts.Mode &= 0777 | linux.S_ISVTX + 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 &= 0777 | linux.S_ISUID | linux.S_ISGID | linux.S_ISVTX + + 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 + } + } +} + +// 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, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return "", err + } + 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 { + oldVD, err := vfs.GetDentryAt(ctx, creds, oldpop, &GetDentryOptions{}) + if err != nil { + return err + } + rp, err := vfs.getResolvingPath(creds, newpop) + if err != nil { + oldVD.DecRef() + return err + } + for { + err := rp.mount.fs.impl.RenameAt(ctx, rp, oldVD, *opts) + if err == nil { + oldVD.DecRef() + vfs.putResolvingPath(rp) + return nil + } + if !rp.handleError(err) { + oldVD.DecRef() + vfs.putResolvingPath(rp) + return err + } + } +} + +// RmdirAt removes the directory at the given path. +func (vfs *VirtualFilesystem) RmdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) error { + rp, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return err + } + for { + err := rp.mount.fs.impl.RmdirAt(ctx, rp) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + 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, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return err + } + 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, 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 + } + } +} + +// 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, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return linux.Statfs{}, err + } + 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 { + rp, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return err + } + for { + err := rp.mount.fs.impl.SymlinkAt(ctx, rp, target) + if err == nil { + vfs.putResolvingPath(rp) + return nil + } + 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 { + rp, err := vfs.getResolvingPath(creds, pop) + if err != nil { + return err + } + for { + err := rp.mount.fs.impl.UnlinkAt(ctx, rp) + 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). |