Fix deadlock in procfs

Kernfs provides an internal mechanism to defer calls to `DecRef()` because
on the last reference `Filesystem.mu` must be held and most places that
need to call `DecRef()` are inside the lock. The same can be true for
filesystems that extend kernfs. procfs needs to look up files and `DecRef()`
them inside the `kernfs.Filesystem.mu`. If the files happen to be procfs
files, it can deadlock trying to decrement if it's the last reference.
This change extends the mechanism to external callers to defer DecRefs
to `vfs.FileDescription` and `vfs.VirtualDentries`.

PiperOrigin-RevId: 384361647

7 files changed, 84 insertions, 40 deletions

diff --git a/pkg/refsvfs2/refs.go b/pkg/refsvfs2/refs.go
index ef8beb659..fe3e4a1ca 100644
--- a/pkg/refsvfs2/refs.go
+++ b/pkg/refsvfs2/refs.go
@@ -28,6 +28,11 @@ type RefCounter interface {
 	// DecRef decrements the object's reference count. Users of refs_template.Refs
 	// may specify a destructor to be called once the reference count reaches zero.
 	DecRef(ctx context.Context)
+}
+
+// TryRefCounter is like RefCounter but allow the ref increment to be tried.
+type TryRefCounter interface {
+	RefCounter
 
 	// TryIncRef attempts to increment the reference count, but may fail if all
 	// references have already been dropped, in which case it returns false. If
diff --git a/pkg/refsvfs2/refs_template.go b/pkg/refsvfs2/refs_template.go
index 1102c8adc..55b0a60a1 100644
--- a/pkg/refsvfs2/refs_template.go
+++ b/pkg/refsvfs2/refs_template.go
@@ -101,7 +101,7 @@ func (r *Refs) IncRef() {
 	}
 }
 
-// TryIncRef implements refs.RefCounter.TryIncRef.
+// TryIncRef implements refs.TryRefCounter.TryIncRef.
 //
 // To do this safely without a loop, a speculative reference is first acquired
 // on the object. This allows multiple concurrent TryIncRef calls to distinguish
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go
index 6f699c9cd..0e2867d49 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs.go
+++ b/pkg/sentry/fsimpl/kernfs/kernfs.go
@@ -52,7 +52,7 @@
 //     vfs.VirtualFilesystem.mountMu
 //       vfs.Dentry.mu
 //   (inode implementation locks, if any)
-// kernfs.Filesystem.droppedDentriesMu
+// kernfs.Filesystem.deferredDecRefsMu
 package kernfs
 
 import (
@@ -76,12 +76,12 @@ import (
 type Filesystem struct {
 	vfsfs vfs.Filesystem
 
-	droppedDentriesMu sync.Mutex `state:"nosave"`
+	deferredDecRefsMu sync.Mutex `state:"nosave"`
 
-	// droppedDentries is a list of dentries waiting to be DecRef()ed. This is
+	// deferredDecRefs is a list of dentries waiting to be DecRef()ed. This is
 	// used to defer dentry destruction until mu can be acquired for
-	// writing. Protected by droppedDentriesMu.
-	droppedDentries []*Dentry
+	// writing. Protected by deferredDecRefsMu.
+	deferredDecRefs []refsvfs2.RefCounter
 
 	// mu synchronizes the lifetime of Dentries on this filesystem. Holding it
 	// for reading guarantees continued existence of any resolved dentries, but
@@ -131,25 +131,49 @@ type Filesystem struct {
 // deferDecRef defers dropping a dentry ref until the next call to
 // processDeferredDecRefs{,Locked}. See comment on Filesystem.mu.
 // This may be called while Filesystem.mu or Dentry.dirMu is locked.
-func (fs *Filesystem) deferDecRef(d *Dentry) {
-	fs.droppedDentriesMu.Lock()
-	fs.droppedDentries = append(fs.droppedDentries, d)
-	fs.droppedDentriesMu.Unlock()
+func (fs *Filesystem) deferDecRef(d refsvfs2.RefCounter) {
+	fs.deferredDecRefsMu.Lock()
+	fs.deferredDecRefs = append(fs.deferredDecRefs, d)
+	fs.deferredDecRefsMu.Unlock()
+}
+
+// SafeDecRefFD safely DecRef the FileDescription making sure DecRef is deferred
+// in case Filesystem.mu is held. See comment on Filesystem.mu.
+func (fs *Filesystem) SafeDecRefFD(ctx context.Context, fd *vfs.FileDescription) {
+	if d, ok := fd.Dentry().Impl().(*Dentry); ok && d.fs == fs {
+		// Only defer if dentry belongs to this filesystem, since locks cannot cross
+		// filesystems.
+		fs.deferDecRef(fd)
+		return
+	}
+	fd.DecRef(ctx)
+}
+
+// SafeDecRef safely DecRef the virtual dentry making sure DecRef is deferred
+// in case Filesystem.mu is held. See comment on Filesystem.mu.
+func (fs *Filesystem) SafeDecRef(ctx context.Context, vd vfs.VirtualDentry) {
+	if d, ok := vd.Dentry().Impl().(*Dentry); ok && d.fs == fs {
+		// Only defer if dentry belongs to this filesystem, since locks cannot cross
+		// filesystems.
+		fs.deferDecRef(&vd)
+		return
+	}
+	vd.DecRef(ctx)
 }
 
 // processDeferredDecRefs calls vfs.Dentry.DecRef on all dentries in the
-// droppedDentries list. See comment on Filesystem.mu.
+// deferredDecRefs list. See comment on Filesystem.mu.
 //
 // Precondition: Filesystem.mu or Dentry.dirMu must NOT be locked.
 func (fs *Filesystem) processDeferredDecRefs(ctx context.Context) {
-	fs.droppedDentriesMu.Lock()
-	for _, d := range fs.droppedDentries {
-		// Defer the DecRef call so that we are not holding droppedDentriesMu
+	fs.deferredDecRefsMu.Lock()
+	for _, d := range fs.deferredDecRefs {
+		// Defer the DecRef call so that we are not holding deferredDecRefsMu
 		// when DecRef is called.
 		defer d.DecRef(ctx)
 	}
-	fs.droppedDentries = fs.droppedDentries[:0] // Keep slice memory for reuse.
-	fs.droppedDentriesMu.Unlock()
+	fs.deferredDecRefs = fs.deferredDecRefs[:0] // Keep slice memory for reuse.
+	fs.deferredDecRefsMu.Unlock()
 }
 
 // VFSFilesystem returns the generic vfs filesystem object.
diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go
index 2717e1359..cbbc0935a 100644
--- a/pkg/sentry/fsimpl/proc/task.go
+++ b/pkg/sentry/fsimpl/proc/task.go
@@ -65,8 +65,8 @@ func (fs *filesystem) newTaskInode(ctx context.Context, task *kernel.Task, pidns
 		"io":        fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0400, newIO(task, isThreadGroup)),
 		"maps":      fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &mapsData{task: task}),
 		"mem":       fs.newMemInode(ctx, task, fs.NextIno(), 0400),
-		"mountinfo": fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &mountInfoData{task: task}),
-		"mounts":    fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &mountsData{task: task}),
+		"mountinfo": fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &mountInfoData{fs: fs, task: task}),
+		"mounts":    fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &mountsData{fs: fs, task: task}),
 		"net":       fs.newTaskNetDir(ctx, task),
 		"ns": fs.newTaskOwnedDir(ctx, task, fs.NextIno(), 0511, map[string]kernfs.Inode{
 			"net":  fs.newNamespaceSymlink(ctx, task, fs.NextIno(), "net"),
diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go
index 4718fac7a..dfc0a924e 100644
--- a/pkg/sentry/fsimpl/proc/task_fds.go
+++ b/pkg/sentry/fsimpl/proc/task_fds.go
@@ -42,12 +42,12 @@ func getTaskFD(t *kernel.Task, fd int32) (*vfs.FileDescription, kernel.FDFlags)
 	return file, flags
 }
 
-func taskFDExists(ctx context.Context, t *kernel.Task, fd int32) bool {
+func taskFDExists(ctx context.Context, fs *filesystem, t *kernel.Task, fd int32) bool {
 	file, _ := getTaskFD(t, fd)
 	if file == nil {
 		return false
 	}
-	file.DecRef(ctx)
+	fs.SafeDecRefFD(ctx, file)
 	return true
 }
 
@@ -145,7 +145,7 @@ func (i *fdDirInode) Lookup(ctx context.Context, name string) (kernfs.Inode, err
 		return nil, syserror.ENOENT
 	}
 	fd := int32(fdInt)
-	if !taskFDExists(ctx, i.task, fd) {
+	if !taskFDExists(ctx, i.fs, i.task, fd) {
 		return nil, syserror.ENOENT
 	}
 	return i.fs.newFDSymlink(ctx, i.task, fd, i.fs.NextIno()), nil
@@ -198,6 +198,7 @@ type fdSymlink struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
 
+	fs   *filesystem
 	task *kernel.Task
 	fd   int32
 }
@@ -206,6 +207,7 @@ var _ kernfs.Inode = (*fdSymlink)(nil)
 
 func (fs *filesystem) newFDSymlink(ctx context.Context, task *kernel.Task, fd int32, ino uint64) kernfs.Inode {
 	inode := &fdSymlink{
+		fs:   fs,
 		task: task,
 		fd:   fd,
 	}
@@ -218,9 +220,9 @@ func (s *fdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error)
 	if file == nil {
 		return "", syserror.ENOENT
 	}
-	defer file.DecRef(ctx)
+	defer s.fs.SafeDecRefFD(ctx, file)
 	root := vfs.RootFromContext(ctx)
-	defer root.DecRef(ctx)
+	defer s.fs.SafeDecRef(ctx, root)
 
 	// Note: it's safe to reenter kernfs from Readlink if needed to resolve path.
 	return s.task.Kernel().VFS().PathnameWithDeleted(ctx, root, file.VirtualDentry())
@@ -231,7 +233,7 @@ func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDen
 	if file == nil {
 		return vfs.VirtualDentry{}, "", syserror.ENOENT
 	}
-	defer file.DecRef(ctx)
+	defer s.fs.SafeDecRefFD(ctx, file)
 	vd := file.VirtualDentry()
 	vd.IncRef()
 	return vd, "", nil
@@ -239,7 +241,7 @@ func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDen
 
 // Valid implements kernfs.Inode.Valid.
 func (s *fdSymlink) Valid(ctx context.Context) bool {
-	return taskFDExists(ctx, s.task, s.fd)
+	return taskFDExists(ctx, s.fs, s.task, s.fd)
 }
 
 // fdInfoDirInode represents the inode for /proc/[pid]/fdinfo directory.
@@ -279,10 +281,11 @@ func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (kernfs.Inode,
 		return nil, syserror.ENOENT
 	}
 	fd := int32(fdInt)
-	if !taskFDExists(ctx, i.task, fd) {
+	if !taskFDExists(ctx, i.fs, i.task, fd) {
 		return nil, syserror.ENOENT
 	}
 	data := &fdInfoData{
+		fs:   i.fs,
 		task: i.task,
 		fd:   fd,
 	}
@@ -316,6 +319,7 @@ func (i *fdInfoDirInode) DecRef(ctx context.Context) {
 type fdInfoData struct {
 	kernfs.DynamicBytesFile
 
+	fs   *filesystem
 	task *kernel.Task
 	fd   int32
 }
@@ -328,7 +332,7 @@ func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 	if file == nil {
 		return syserror.ENOENT
 	}
-	defer file.DecRef(ctx)
+	defer d.fs.SafeDecRefFD(ctx, file)
 	// TODO(b/121266871): Include pos, locks, and other data. For now we only
 	// have flags.
 	// See https://www.kernel.org/doc/Documentation/filesystems/proc.txt
@@ -339,5 +343,5 @@ func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 
 // Valid implements kernfs.Inode.Valid.
 func (d *fdInfoData) Valid(ctx context.Context) bool {
-	return taskFDExists(ctx, d.task, d.fd)
+	return taskFDExists(ctx, d.fs, d.task, d.fd)
 }
diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go
index 5526cac1e..5bb6bc372 100644
--- a/pkg/sentry/fsimpl/proc/task_files.go
+++ b/pkg/sentry/fsimpl/proc/task_files.go
@@ -803,13 +803,17 @@ type exeSymlink struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
 
+	fs   *filesystem
 	task *kernel.Task
 }
 
 var _ kernfs.Inode = (*exeSymlink)(nil)
 
 func (fs *filesystem) newExeSymlink(ctx context.Context, task *kernel.Task, ino uint64) kernfs.Inode {
-	inode := &exeSymlink{task: task}
+	inode := &exeSymlink{
+		fs:   fs,
+		task: task,
+	}
 	inode.Init(ctx, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777)
 	return inode
 }
@@ -820,14 +824,14 @@ func (s *exeSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error)
 	if err != nil {
 		return "", err
 	}
-	defer exec.DecRef(ctx)
+	defer s.fs.SafeDecRef(ctx, exec)
 
 	root := vfs.RootFromContext(ctx)
 	if !root.Ok() {
 		// It could have raced with process deletion.
 		return "", linuxerr.ESRCH
 	}
-	defer root.DecRef(ctx)
+	defer s.fs.SafeDecRef(ctx, root)
 
 	vfsObj := exec.Mount().Filesystem().VirtualFilesystem()
 	name, _ := vfsObj.PathnameWithDeleted(ctx, root, exec)
@@ -879,13 +883,17 @@ type cwdSymlink struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
 
+	fs   *filesystem
 	task *kernel.Task
 }
 
 var _ kernfs.Inode = (*cwdSymlink)(nil)
 
 func (fs *filesystem) newCwdSymlink(ctx context.Context, task *kernel.Task, ino uint64) kernfs.Inode {
-	inode := &cwdSymlink{task: task}
+	inode := &cwdSymlink{
+		fs:   fs,
+		task: task,
+	}
 	inode.Init(ctx, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777)
 	return inode
 }
@@ -896,14 +904,14 @@ func (s *cwdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error)
 	if err != nil {
 		return "", err
 	}
-	defer cwd.DecRef(ctx)
+	defer s.fs.SafeDecRef(ctx, cwd)
 
 	root := vfs.RootFromContext(ctx)
 	if !root.Ok() {
 		// It could have raced with process deletion.
 		return "", linuxerr.ESRCH
 	}
-	defer root.DecRef(ctx)
+	defer s.fs.SafeDecRef(ctx, root)
 
 	vfsObj := cwd.Mount().Filesystem().VirtualFilesystem()
 	name, _ := vfsObj.PathnameWithDeleted(ctx, root, cwd)
@@ -923,6 +931,7 @@ func (s *cwdSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDent
 		// It could have raced with process deletion.
 		return vfs.VirtualDentry{}, "", linuxerr.ESRCH
 	}
+	// The reference is transferred to the caller.
 	return cwd, "", nil
 }
 
@@ -932,6 +941,7 @@ func (s *cwdSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDent
 type mountInfoData struct {
 	kernfs.DynamicBytesFile
 
+	fs   *filesystem
 	task *kernel.Task
 }
 
@@ -952,7 +962,7 @@ func (i *mountInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 		// Root has been destroyed. Don't try to read mounts.
 		return nil
 	}
-	defer rootDir.DecRef(ctx)
+	defer i.fs.SafeDecRef(ctx, rootDir)
 	i.task.Kernel().VFS().GenerateProcMountInfo(ctx, rootDir, buf)
 	return nil
 }
@@ -963,6 +973,7 @@ func (i *mountInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 type mountsData struct {
 	kernfs.DynamicBytesFile
 
+	fs   *filesystem
 	task *kernel.Task
 }
 
@@ -983,7 +994,7 @@ func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 		// Root has been destroyed. Don't try to read mounts.
 		return nil
 	}
-	defer rootDir.DecRef(ctx)
+	defer i.fs.SafeDecRef(ctx, rootDir)
 	i.task.Kernel().VFS().GenerateProcMounts(ctx, rootDir, buf)
 	return nil
 }
diff --git a/pkg/sentry/kernel/abstract_socket_namespace.go b/pkg/sentry/kernel/abstract_socket_namespace.go
index d100e58d7..5d86a04f3 100644
--- a/pkg/sentry/kernel/abstract_socket_namespace.go
+++ b/pkg/sentry/kernel/abstract_socket_namespace.go
@@ -27,7 +27,7 @@ import (
 // +stateify savable
 type abstractEndpoint struct {
 	ep     transport.BoundEndpoint
-	socket refsvfs2.RefCounter
+	socket refsvfs2.TryRefCounter
 	name   string
 	ns     *AbstractSocketNamespace
 }
@@ -57,7 +57,7 @@ func NewAbstractSocketNamespace() *AbstractSocketNamespace {
 // its backing socket.
 type boundEndpoint struct {
 	transport.BoundEndpoint
-	socket refsvfs2.RefCounter
+	socket refsvfs2.TryRefCounter
 }
 
 // Release implements transport.BoundEndpoint.Release.
@@ -89,7 +89,7 @@ func (a *AbstractSocketNamespace) BoundEndpoint(name string) transport.BoundEndp
 //
 // When the last reference managed by socket is dropped, ep may be removed from the
 // namespace.
-func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep transport.BoundEndpoint, socket refsvfs2.RefCounter) error {
+func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep transport.BoundEndpoint, socket refsvfs2.TryRefCounter) error {
 	a.mu.Lock()
 	defer a.mu.Unlock()
 
@@ -109,7 +109,7 @@ func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep tran
 
 // Remove removes the specified socket at name from the abstract socket
 // namespace, if it has not yet been replaced.
-func (a *AbstractSocketNamespace) Remove(name string, socket refsvfs2.RefCounter) {
+func (a *AbstractSocketNamespace) Remove(name string, socket refsvfs2.TryRefCounter) {
 	a.mu.Lock()
 	defer a.mu.Unlock()