summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/sentry')
-rw-r--r--pkg/sentry/arch/arch_amd64.go4
-rw-r--r--pkg/sentry/control/pprof.go15
-rw-r--r--pkg/sentry/control/proc.go48
-rw-r--r--pkg/sentry/control/proc_test.go10
-rw-r--r--pkg/sentry/fs/gofer/session.go6
-rw-r--r--pkg/sentry/fs/proc/task.go34
-rw-r--r--pkg/sentry/fs/tty/terminal.go4
-rw-r--r--pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go6
-rw-r--r--pkg/sentry/fsimpl/ext/ext_test.go29
-rw-r--r--pkg/sentry/fsimpl/memfs/benchmark_test.go2
-rw-r--r--pkg/sentry/fsimpl/memfs/pipe_test.go6
-rw-r--r--pkg/sentry/fsimpl/proc/filesystems.go2
-rw-r--r--pkg/sentry/fsimpl/proc/mounts.go2
-rw-r--r--pkg/sentry/kernel/kernel.go20
-rw-r--r--pkg/sentry/kernel/semaphore/semaphore.go6
-rw-r--r--pkg/sentry/kernel/syscalls.go8
-rw-r--r--pkg/sentry/kernel/task.go20
-rw-r--r--pkg/sentry/kernel/task_block.go8
-rw-r--r--pkg/sentry/kernel/task_clone.go1
-rw-r--r--pkg/sentry/kernel/task_exec.go3
-rw-r--r--pkg/sentry/kernel/task_exit.go1
-rw-r--r--pkg/sentry/kernel/task_log.go86
-rw-r--r--pkg/sentry/kernel/task_run.go14
-rw-r--r--pkg/sentry/kernel/task_start.go8
-rw-r--r--pkg/sentry/kernel/task_syscall.go8
-rw-r--r--pkg/sentry/kernel/tty.go11
-rw-r--r--pkg/sentry/loader/elf.go2
-rw-r--r--pkg/sentry/sighandling/sighandling.go75
-rw-r--r--pkg/sentry/sighandling/sighandling_unsafe.go26
-rw-r--r--pkg/sentry/socket/control/control.go63
-rw-r--r--pkg/sentry/socket/hostinet/socket.go26
-rw-r--r--pkg/sentry/socket/netstack/netstack.go2
-rw-r--r--pkg/sentry/socket/rpcinet/syscall_rpc.proto1
-rw-r--r--pkg/sentry/socket/socket.go5
-rw-r--r--pkg/sentry/strace/BUILD1
-rw-r--r--pkg/sentry/strace/linux64.go32
-rw-r--r--pkg/sentry/strace/select.go53
-rw-r--r--pkg/sentry/strace/socket.go9
-rw-r--r--pkg/sentry/strace/strace.go2
-rw-r--r--pkg/sentry/strace/syscalls.go4
-rw-r--r--pkg/sentry/syscalls/linux/linux64_amd64.go24
-rw-r--r--pkg/sentry/syscalls/linux/linux64_arm64.go21
-rw-r--r--pkg/sentry/syscalls/linux/sys_file.go70
-rw-r--r--pkg/sentry/syscalls/linux/sys_poll.go71
-rw-r--r--pkg/sentry/syscalls/linux/sys_socket.go26
-rw-r--r--pkg/sentry/vfs/BUILD1
-rw-r--r--pkg/sentry/vfs/file_description.go93
-rw-r--r--pkg/sentry/vfs/file_description_impl_util_test.go10
-rw-r--r--pkg/sentry/vfs/filesystem.go22
-rw-r--r--pkg/sentry/vfs/mount.go69
-rw-r--r--pkg/sentry/vfs/options.go12
-rw-r--r--pkg/sentry/vfs/permissions.go62
-rw-r--r--pkg/sentry/vfs/syscalls.go237
-rw-r--r--pkg/sentry/vfs/vfs.go378
-rw-r--r--pkg/sentry/watchdog/watchdog.go4
55 files changed, 1241 insertions, 522 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/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/proc/task.go b/pkg/sentry/fs/proc/task.go
index 87184ec67..0e46c5fb7 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 {
@@ -619,8 +618,11 @@ type ioData struct {
ioUsage
}
-func newIO(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
- return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t.ThreadGroup()}), msrc, fs.SpecialFile, t)
+func newIO(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode {
+ if isThreadGroup {
+ return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t.ThreadGroup()}), msrc, fs.SpecialFile, t)
+ }
+ return newProcInode(t, seqfile.NewSeqFile(t, &ioData{t}), msrc, fs.SpecialFile, t)
}
// NeedsUpdate returns whether the generation is old or not.
@@ -639,7 +641,7 @@ func (i *ioData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se
io.Accumulate(i.IOUsage())
var buf bytes.Buffer
- fmt.Fprintf(&buf, "char: %d\n", io.CharsRead)
+ fmt.Fprintf(&buf, "rchar: %d\n", io.CharsRead)
fmt.Fprintf(&buf, "wchar: %d\n", io.CharsWritten)
fmt.Fprintf(&buf, "syscr: %d\n", io.ReadSyscalls)
fmt.Fprintf(&buf, "syscw: %d\n", io.WriteSyscalls)
diff --git a/pkg/sentry/fs/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/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/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/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/filesystems.go b/pkg/sentry/fsimpl/proc/filesystems.go
index c36c4aff5..0e016bca5 100644
--- a/pkg/sentry/fsimpl/proc/filesystems.go
+++ b/pkg/sentry/fsimpl/proc/filesystems.go
@@ -19,7 +19,7 @@ package proc
// +stateify savable
type filesystemsData struct{}
-// TODO(b/138862512): Implement vfs.DynamicBytesSource.Generate for
+// TODO(gvisor.dev/issue/1195): Implement vfs.DynamicBytesSource.Generate for
// filesystemsData. We would need to retrive filesystem names from
// vfs.VirtualFilesystem. Also needs vfs replacement for
// fs.Filesystem.AllowUserList() and fs.FilesystemRequiresDev.
diff --git a/pkg/sentry/fsimpl/proc/mounts.go b/pkg/sentry/fsimpl/proc/mounts.go
index e81b1e910..8683cf677 100644
--- a/pkg/sentry/fsimpl/proc/mounts.go
+++ b/pkg/sentry/fsimpl/proc/mounts.go
@@ -16,7 +16,7 @@ package proc
import "gvisor.dev/gvisor/pkg/sentry/kernel"
-// TODO(b/138862512): Implement mountInfoFile and mountsFile.
+// TODO(gvisor.dev/issue/1195): Implement mountInfoFile and mountsFile.
// mountInfoFile implements vfs.DynamicBytesSource for /proc/[pid]/mountinfo.
//
diff --git a/pkg/sentry/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/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 8d9363aac..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
)
@@ -289,12 +289,12 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt
switch level {
case linux.SOL_IP:
switch name {
- case linux.IP_RECVTOS:
+ case linux.IP_TOS, linux.IP_RECVTOS:
optlen = sizeofInt32
}
case linux.SOL_IPV6:
switch name {
- case linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY:
+ case linux.IPV6_TCLASS, linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY:
optlen = sizeofInt32
}
case linux.SOL_SOCKET:
@@ -334,12 +334,12 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [
switch level {
case linux.SOL_IP:
switch name {
- case linux.IP_RECVTOS:
+ case linux.IP_TOS, linux.IP_RECVTOS:
optlen = sizeofInt32
}
case linux.SOL_IPV6:
switch name {
- case linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY:
+ case linux.IPV6_TCLASS, linux.IPV6_RECVTCLASS, linux.IPV6_V6ONLY:
optlen = sizeofInt32
}
case linux.SOL_SOCKET:
@@ -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..fe5a46aa3 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."),
diff --git a/pkg/sentry/socket/rpcinet/syscall_rpc.proto b/pkg/sentry/socket/rpcinet/syscall_rpc.proto
index 9586f5923..b677e9eb3 100644
--- a/pkg/sentry/socket/rpcinet/syscall_rpc.proto
+++ b/pkg/sentry/socket/rpcinet/syscall_rpc.proto
@@ -3,7 +3,6 @@ syntax = "proto3";
// package syscall_rpc is a set of networking related system calls that can be
// forwarded to a socket gofer.
//
-// TODO(b/77963526): Document individual RPCs.
package syscall_rpc;
message SendmsgRequest {
diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go
index 8c250c325..2389a9cdb 100644
--- a/pkg/sentry/socket/socket.go
+++ b/pkg/sentry/socket/socket.go
@@ -43,6 +43,11 @@ type ControlMessages struct {
IP tcpip.ControlMessages
}
+// Release releases Unix domain socket credentials and rights.
+func (c *ControlMessages) Release() {
+ c.Unix.Release()
+}
+
// Socket is the interface containing socket syscalls used by the syscall layer
// to redirect them to the appropriate implementation.
type Socket interface {
diff --git a/pkg/sentry/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..92c18083d
--- /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>>uint(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 := uint(nfds % 8)
+
+ set, err := linux.CopyInFDSet(t, addr, nBytes, nBitsInLastPartialByte)
+ if err != nil {
+ return fmt.Sprintf("%#x (error decoding fdset: %s)", addr, err)
+ }
+
+ return fmt.Sprintf("%#x %v", addr, fdsFromSet(t, set))
+}
diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go
index 94334f6d2..51f2efb39 100644
--- a/pkg/sentry/strace/socket.go
+++ b/pkg/sentry/strace/socket.go
@@ -208,6 +208,15 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64)
i += linux.SizeOfControlMessageHeader
width := t.Arch().Width()
length := int(h.Length) - linux.SizeOfControlMessageHeader
+ if length < 0 {
+ strs = append(strs, fmt.Sprintf(
+ "{level=%s, type=%s, length=%d, content too short}",
+ level,
+ typ,
+ h.Length,
+ ))
+ break
+ }
if skipData {
strs = append(strs, fmt.Sprintf("{level=%s, type=%s, length=%d}", level, typ, h.Length))
diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go
index 311389547..629c1f308 100644
--- a/pkg/sentry/strace/strace.go
+++ b/pkg/sentry/strace/strace.go
@@ -439,6 +439,8 @@ func (i *SyscallInfo) pre(t *kernel.Task, args arch.SyscallArguments, maximumBlo
output = append(output, capData(t, args[arg-1].Pointer(), args[arg].Pointer()))
case PollFDs:
output = append(output, pollFDs(t, args[arg].Pointer(), uint(args[arg+1].Uint()), false))
+ case SelectFDSet:
+ output = append(output, fdSet(t, int(args[0].Int()), args[arg].Pointer()))
case Oct:
output = append(output, "0o"+strconv.FormatUint(args[arg].Uint64(), 8))
case Hex:
diff --git a/pkg/sentry/strace/syscalls.go b/pkg/sentry/strace/syscalls.go
index 3c389d375..e5d486c4e 100644
--- a/pkg/sentry/strace/syscalls.go
+++ b/pkg/sentry/strace/syscalls.go
@@ -206,6 +206,10 @@ const (
// PollFDs is an array of struct pollfd. The number of entries in the
// array is in the next argument.
PollFDs
+
+ // SelectFDSet is an fd_set argument in select(2)/pselect(2). The number of
+ // fds represented must be the first argument.
+ SelectFDSet
)
// defaultFormat is the syscall argument format to use if the actual format is
diff --git a/pkg/sentry/syscalls/linux/linux64_amd64.go b/pkg/sentry/syscalls/linux/linux64_amd64.go
index 81e4f93a6..797542d28 100644
--- a/pkg/sentry/syscalls/linux/linux64_amd64.go
+++ b/pkg/sentry/syscalls/linux/linux64_amd64.go
@@ -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..2bc7faff5 100644
--- a/pkg/sentry/syscalls/linux/linux64_arm64.go
+++ b/pkg/sentry/syscalls/linux/linux64_arm64.go
@@ -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..631dffec6 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 int, nBitsInLastPartialByte uint) ([]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 := uint(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/vfs/BUILD b/pkg/sentry/vfs/BUILD
index 74a325309..59237c3b9 100644
--- a/pkg/sentry/vfs/BUILD
+++ b/pkg/sentry/vfs/BUILD
@@ -19,7 +19,6 @@ go_library(
"options.go",
"permissions.go",
"resolving_path.go",
- "syscalls.go",
"testutil.go",
"vfs.go",
],
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index 34007eb57..4473dfce8 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -241,3 +241,96 @@ 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)
+}
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..dfbd2372a 100644
--- a/pkg/sentry/vfs/filesystem.go
+++ b/pkg/sentry/vfs/filesystem.go
@@ -47,6 +47,9 @@ 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 +69,28 @@ 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")
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/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/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).
diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go
index ecce6c69f..5e4611333 100644
--- a/pkg/sentry/watchdog/watchdog.go
+++ b/pkg/sentry/watchdog/watchdog.go
@@ -287,7 +287,9 @@ func (w *Watchdog) runTurn() {
if !ok {
// New stuck task detected.
//
- // TODO(b/65849403): Tasks blocked doing IO may be considered stuck in kernel.
+ // Note that tasks blocked doing IO may be considered stuck in kernel,
+ // unless they are surrounded b
+ // Task.UninterruptibleSleepStart/Finish.
tc = &offender{lastUpdateTime: lastUpdateTime}
stuckTasks.Increment()
newTaskFound = true