Support inotify in vfs2 gofer fs.

Because there is no inode structure stored in the sandbox, inotify watches
must be held on the dentry. This would be an issue in the presence of hard
links, where multiple dentries would need to share the same set of watches,
but in VFS2, we do not support the internal creation of hard links on gofer
fs. As a result, we make the assumption that every dentry corresponds to a
unique inode.

Furthermore, dentries can be cached and then evicted, even if the underlying
file has not be deleted. We must prevent this from occurring if there are any
watches that would be lost. Note that if the dentry was deleted or invalidated
(d.vfsd.IsDead()), we should still destroy it along with its watches.

Additionally, when a dentry’s last watch is removed, we cache it if it also
has zero references. This way, the dentry can eventually be evicted from
memory if it is no longer needed. This is accomplished with a new dentry
method, OnZeroWatches(), which is called by Inotify.RmWatch and
Inotify.Release. Note that it must be called after all inotify locks are
released to avoid violating lock order. Stress tests are added to make sure
that inotify operations don't deadlock with gofer.OnZeroWatches.

Updates #1479.

PiperOrigin-RevId: 317958034

3 files changed, 20 insertions, 15 deletions

diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go
index 913b8a6c5..b172abc15 100644
--- a/pkg/sentry/fsimpl/tmpfs/directory.go
+++ b/pkg/sentry/fsimpl/tmpfs/directory.go
@@ -79,7 +79,6 @@ func (dir *directory) removeChildLocked(child *dentry) {
 	dir.iterMu.Lock()
 	dir.childList.Remove(child)
 	dir.iterMu.Unlock()
-	child.unlinked = true
 }
 
 type directoryFD struct {
@@ -107,13 +106,14 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba
 	fs := fd.filesystem()
 	dir := fd.inode().impl.(*directory)
 
+	defer fd.dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent)
+
 	// fs.mu is required to read d.parent and dentry.name.
 	fs.mu.RLock()
 	defer fs.mu.RUnlock()
 	dir.iterMu.Lock()
 	defer dir.iterMu.Unlock()
 
-	fd.dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent)
 	fd.inode().touchAtime(fd.vfsfd.Mount())
 
 	if fd.off == 0 {
diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go
index 637d84e04..85d8e37b2 100644
--- a/pkg/sentry/fsimpl/tmpfs/filesystem.go
+++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go
@@ -638,14 +638,16 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 // SetStatAt implements vfs.FilesystemImpl.SetStatAt.
 func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
 	fs.mu.RLock()
-	defer fs.mu.RUnlock()
 	d, err := resolveLocked(rp)
 	if err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
 	if err := d.inode.setStat(ctx, rp.Credentials(), &opts.Stat); err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
+	fs.mu.RUnlock()
 
 	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
 		d.InotifyWithParent(ev, 0, vfs.InodeEvent)
@@ -788,14 +790,16 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt
 // SetxattrAt implements vfs.FilesystemImpl.SetxattrAt.
 func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error {
 	fs.mu.RLock()
-	defer fs.mu.RUnlock()
 	d, err := resolveLocked(rp)
 	if err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
 	if err := d.inode.setxattr(rp.Credentials(), &opts); err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
+	fs.mu.RUnlock()
 
 	d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil
@@ -804,14 +808,16 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt
 // RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt.
 func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error {
 	fs.mu.RLock()
-	defer fs.mu.RUnlock()
 	d, err := resolveLocked(rp)
 	if err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
 	if err := d.inode.removexattr(rp.Credentials(), name); err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
+	fs.mu.RUnlock()
 
 	d.InotifyWithParent(linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
index a94333ee0..a85bfc968 100644
--- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
@@ -215,11 +215,6 @@ type dentry struct {
 	// filesystem.mu.
 	name string
 
-	// unlinked indicates whether this dentry has been unlinked from its parent.
-	// It is only set to true on an unlink operation, and never set from true to
-	// false. unlinked is protected by filesystem.mu.
-	unlinked bool
-
 	// dentryEntry (ugh) links dentries into their parent directory.childList.
 	dentryEntry
 
@@ -259,18 +254,20 @@ func (d *dentry) DecRef() {
 }
 
 // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent.
-func (d *dentry) InotifyWithParent(events uint32, cookie uint32, et vfs.EventType) {
+func (d *dentry) InotifyWithParent(events, cookie uint32, et vfs.EventType) {
 	if d.inode.isDir() {
 		events |= linux.IN_ISDIR
 	}
 
+	d.inode.fs.mu.RLock()
 	// The ordering below is important, Linux always notifies the parent first.
 	if d.parent != nil {
-		// Note that d.parent or d.name may be stale if there is a concurrent
-		// rename operation. Inotify does not provide consistency guarantees.
-		d.parent.inode.watches.NotifyWithExclusions(d.name, events, cookie, et, d.unlinked)
+		// tmpfs never calls VFS.InvalidateDentry(), so d.vfsd.IsDead() indicates
+		// that d was deleted.
+		d.parent.inode.watches.NotifyWithExclusions(d.name, events, cookie, et, d.vfsd.IsDead())
 	}
 	d.inode.watches.Notify("", events, cookie, et)
+	d.inode.fs.mu.RUnlock()
 }
 
 // Watches implements vfs.DentryImpl.Watches.
@@ -278,6 +275,9 @@ func (d *dentry) Watches() *vfs.Watches {
 	return &d.inode.watches
 }
 
+// OnZeroWatches implements vfs.Dentry.OnZeroWatches.
+func (d *dentry) OnZeroWatches() {}
+
 // inode represents a filesystem object.
 type inode struct {
 	// fs is the owning filesystem. fs is immutable.
@@ -336,7 +336,6 @@ func (i *inode) init(impl interface{}, fs *filesystem, kuid auth.KUID, kgid auth
 	i.ctime = now
 	i.mtime = now
 	// i.nlink initialized by caller
-	i.watches = vfs.Watches{}
 	i.impl = impl
 }