Support inotify in overlayfs.

Fixes #1479, #317.

PiperOrigin-RevId: 334258052

6 files changed, 129 insertions, 34 deletions

diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go
index 97b9165cc..94d96261b 100644
--- a/pkg/sentry/fsimpl/gofer/filesystem.go
+++ b/pkg/sentry/fsimpl/gofer/filesystem.go
@@ -1416,11 +1416,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts
 		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 		return err
 	}
-	if err := d.setStat(ctx, rp.Credentials(), &opts, rp.Mount()); err != nil {
-		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	err = d.setStat(ctx, rp.Credentials(), &opts, rp.Mount())
+	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	if err != nil {
 		return err
 	}
-	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 
 	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
 		d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
@@ -1556,11 +1556,11 @@ func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt
 		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 		return err
 	}
-	if err := d.setXattr(ctx, rp.Credentials(), &opts); err != nil {
-		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	err = d.setXattr(ctx, rp.Credentials(), &opts)
+	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	if err != nil {
 		return err
 	}
-	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 
 	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil
@@ -1575,11 +1575,11 @@ func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath,
 		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 		return err
 	}
-	if err := d.removeXattr(ctx, rp.Credentials(), name); err != nil {
-		fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	err = d.removeXattr(ctx, rp.Credentials(), name)
+	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
+	if err != nil {
 		return err
 	}
-	fs.renameMuRUnlockAndCheckCaching(ctx, &ds)
 
 	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil
diff --git a/pkg/sentry/fsimpl/overlay/directory.go b/pkg/sentry/fsimpl/overlay/directory.go
index aea599460..df4492346 100644
--- a/pkg/sentry/fsimpl/overlay/directory.go
+++ b/pkg/sentry/fsimpl/overlay/directory.go
@@ -117,10 +117,12 @@ func (fd *directoryFD) Release(ctx context.Context) {
 
 // IterDirents implements vfs.FileDescriptionImpl.IterDirents.
 func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error {
+	d := fd.dentry()
+	defer d.InotifyWithParent(ctx, linux.IN_ACCESS, 0, vfs.PathEvent)
+
 	fd.mu.Lock()
 	defer fd.mu.Unlock()
 
-	d := fd.dentry()
 	if fd.dirents == nil {
 		ds, err := d.getDirents(ctx)
 		if err != nil {
diff --git a/pkg/sentry/fsimpl/overlay/filesystem.go b/pkg/sentry/fsimpl/overlay/filesystem.go
index e9ce4bde1..bd11372d5 100644
--- a/pkg/sentry/fsimpl/overlay/filesystem.go
+++ b/pkg/sentry/fsimpl/overlay/filesystem.go
@@ -499,7 +499,13 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir
 	if err := create(parent, name, childLayer == lookupLayerUpperWhiteout); err != nil {
 		return err
 	}
+
 	parent.dirents = nil
+	ev := linux.IN_CREATE
+	if dir {
+		ev |= linux.IN_ISDIR
+	}
+	parent.watches.Notify(ctx, name, uint32(ev), 0 /* cookie */, vfs.InodeEvent, false /* unlinked */)
 	return nil
 }
 
@@ -631,6 +637,7 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.
 			}
 			return err
 		}
+		old.watches.Notify(ctx, "", linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent, false /* unlinked */)
 		return nil
 	})
 }
@@ -975,6 +982,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving
 		// just can't open it anymore for some reason.
 		return nil, err
 	}
+	parent.watches.Notify(ctx, childName, linux.IN_CREATE, 0 /* cookie */, vfs.PathEvent, false /* unlinked */)
 	return &fd.vfsfd, nil
 }
 
@@ -1236,6 +1244,7 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 		}
 	}
 
+	vfs.InotifyRename(ctx, &renamed.watches, &oldParent.watches, &newParent.watches, oldName, newName, renamed.isDir())
 	return nil
 }
 
@@ -1352,6 +1361,7 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 	delete(parent.children, name)
 	ds = appendDentry(ds, child)
 	parent.dirents = nil
+	parent.watches.Notify(ctx, name, linux.IN_DELETE|linux.IN_ISDIR, 0 /* cookie */, vfs.InodeEvent, true /* unlinked */)
 	return nil
 }
 
@@ -1359,12 +1369,25 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
 	var ds *[]*dentry
 	fs.renameMu.RLock()
-	defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
 	d, err := fs.resolveLocked(ctx, rp, &ds)
 	if err != nil {
+		fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
+		return err
+	}
+	err = d.setStatLocked(ctx, rp, opts)
+	fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
+	if err != nil {
 		return err
 	}
 
+	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
+		d.InotifyWithParent(ctx, ev, 0 /* cookie */, vfs.InodeEvent)
+	}
+	return nil
+}
+
+// Precondition: d.fs.renameMu must be held for reading.
+func (d *dentry) setStatLocked(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
 	mode := linux.FileMode(atomic.LoadUint32(&d.mode))
 	if err := vfs.CheckSetStat(ctx, rp.Credentials(), &opts, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil {
 		return err
@@ -1555,11 +1578,14 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error
 		panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout during UnlinkAt: %v", err))
 	}
 
+	var cw *vfs.Watches
 	if child != nil {
 		vfsObj.CommitDeleteDentry(ctx, &child.vfsd)
 		delete(parent.children, name)
 		ds = appendDentry(ds, child)
+		cw = &child.watches
 	}
+	vfs.InotifyRemoveChild(ctx, cw, &parent.watches, name)
 	parent.dirents = nil
 	return nil
 }
@@ -1636,13 +1662,20 @@ func (fs *filesystem) getXattr(ctx context.Context, d *dentry, creds *auth.Crede
 func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error {
 	var ds *[]*dentry
 	fs.renameMu.RLock()
-	defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
 	d, err := fs.resolveLocked(ctx, rp, &ds)
 	if err != nil {
+		fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
 		return err
 	}
 
-	return fs.setXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), &opts)
+	err = fs.setXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), &opts)
+	fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
+	if err != nil {
+		return err
+	}
+
+	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent)
+	return nil
 }
 
 // Precondition: fs.renameMu must be locked.
@@ -1673,13 +1706,20 @@ func (fs *filesystem) setXattrLocked(ctx context.Context, d *dentry, mnt *vfs.Mo
 func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error {
 	var ds *[]*dentry
 	fs.renameMu.RLock()
-	defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
 	d, err := fs.resolveLocked(ctx, rp, &ds)
 	if err != nil {
+		fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
 		return err
 	}
 
-	return fs.removeXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), name)
+	err = fs.removeXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), name)
+	fs.renameMuRUnlockAndCheckDrop(ctx, &ds)
+	if err != nil {
+		return err
+	}
+
+	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent)
+	return nil
 }
 
 // Precondition: fs.renameMu must be locked.
diff --git a/pkg/sentry/fsimpl/overlay/non_directory.go b/pkg/sentry/fsimpl/overlay/non_directory.go
index 50b4a50f1..853aee951 100644
--- a/pkg/sentry/fsimpl/overlay/non_directory.go
+++ b/pkg/sentry/fsimpl/overlay/non_directory.go
@@ -184,6 +184,9 @@ func (fd *nonDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions)
 		return err
 	}
 	d.updateAfterSetStatLocked(&opts)
+	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
+		d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
+	}
 	return nil
 }
 
diff --git a/pkg/sentry/fsimpl/overlay/overlay.go b/pkg/sentry/fsimpl/overlay/overlay.go
index b03fa9121..dfbccd05f 100644
--- a/pkg/sentry/fsimpl/overlay/overlay.go
+++ b/pkg/sentry/fsimpl/overlay/overlay.go
@@ -462,6 +462,13 @@ type dentry struct {
 	isMappable      uint32
 
 	locks vfs.FileLocks
+
+	// watches is the set of inotify watches on the file repesented by this dentry.
+	//
+	// Note that hard links to the same file will not share the same set of
+	// watches, due to the fact that we do not have inode structures in this
+	// overlay implementation.
+	watches vfs.Watches
 }
 
 // newDentry creates a new dentry. The dentry initially has no references; it
@@ -521,6 +528,14 @@ func (d *dentry) checkDropLocked(ctx context.Context) {
 	if atomic.LoadInt64(&d.refs) != 0 {
 		return
 	}
+
+	// Make sure that we do not lose watches on dentries that have not been
+	// deleted. Note that overlayfs never calls VFS.InvalidateDentry(), so
+	// d.vfsd.IsDead() indicates that d was deleted.
+	if !d.vfsd.IsDead() && d.watches.Size() > 0 {
+		return
+	}
+
 	// Refs is still zero; destroy it.
 	d.destroyLocked(ctx)
 	return
@@ -549,6 +564,8 @@ func (d *dentry) destroyLocked(ctx context.Context) {
 		lowerVD.DecRef(ctx)
 	}
 
+	d.watches.HandleDeletion(ctx)
+
 	if d.parent != nil {
 		d.parent.dirMu.Lock()
 		if !d.vfsd.IsDead() {
@@ -567,19 +584,36 @@ func (d *dentry) destroyLocked(ctx context.Context) {
 
 // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent.
 func (d *dentry) InotifyWithParent(ctx context.Context, events uint32, cookie uint32, et vfs.EventType) {
-	// TODO(gvisor.dev/issue/1479): Implement inotify.
+	if d.isDir() {
+		events |= linux.IN_ISDIR
+	}
+
+	// overlayfs never calls VFS.InvalidateDentry(), so d.vfsd.IsDead() indicates
+	// that d was deleted.
+	deleted := d.vfsd.IsDead()
+
+	d.fs.renameMu.RLock()
+	// The ordering below is important, Linux always notifies the parent first.
+	if d.parent != nil {
+		d.parent.watches.Notify(ctx, d.name, events, cookie, et, deleted)
+	}
+	d.watches.Notify(ctx, "", events, cookie, et, deleted)
+	d.fs.renameMu.RUnlock()
 }
 
 // Watches implements vfs.DentryImpl.Watches.
 func (d *dentry) Watches() *vfs.Watches {
-	// TODO(gvisor.dev/issue/1479): Implement inotify.
-	return nil
+	return &d.watches
 }
 
 // OnZeroWatches implements vfs.DentryImpl.OnZeroWatches.
-//
-// TODO(gvisor.dev/issue/1479): Implement inotify.
-func (d *dentry) OnZeroWatches(context.Context) {}
+func (d *dentry) OnZeroWatches(ctx context.Context) {
+	if atomic.LoadInt64(&d.refs) == 0 {
+		d.fs.renameMu.Lock()
+		d.checkDropLocked(ctx)
+		d.fs.renameMu.Unlock()
+	}
+}
 
 // iterLayers invokes yield on each layer comprising d, from top to bottom. If
 // any call to yield returns false, iterLayer stops iteration.
@@ -689,17 +723,33 @@ func (fd *fileDescription) GetXattr(ctx context.Context, opts vfs.GetXattrOption
 // SetXattr implements vfs.FileDescriptionImpl.SetXattr.
 func (fd *fileDescription) SetXattr(ctx context.Context, opts vfs.SetXattrOptions) error {
 	fs := fd.filesystem()
+	d := fd.dentry()
+
 	fs.renameMu.RLock()
-	defer fs.renameMu.RUnlock()
-	return fs.setXattrLocked(ctx, fd.dentry(), fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), &opts)
+	err := fs.setXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), &opts)
+	fs.renameMu.RUnlock()
+	if err != nil {
+		return err
+	}
+
+	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+	return nil
 }
 
 // RemoveXattr implements vfs.FileDescriptionImpl.RemoveXattr.
 func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error {
 	fs := fd.filesystem()
+	d := fd.dentry()
+
 	fs.renameMu.RLock()
-	defer fs.renameMu.RUnlock()
-	return fs.removeXattrLocked(ctx, fd.dentry(), fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), name)
+	err := fs.removeXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), name)
+	fs.renameMu.RUnlock()
+	if err != nil {
+		return err
+	}
+
+	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+	return nil
 }
 
 // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX.
diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go
index 1362c1602..e39cd305b 100644
--- a/pkg/sentry/fsimpl/tmpfs/filesystem.go
+++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go
@@ -673,11 +673,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts
 		fs.mu.RUnlock()
 		return err
 	}
-	if err := d.inode.setStat(ctx, rp.Credentials(), &opts); err != nil {
-		fs.mu.RUnlock()
+	err = d.inode.setStat(ctx, rp.Credentials(), &opts)
+	fs.mu.RUnlock()
+	if err != nil {
 		return err
 	}
-	fs.mu.RUnlock()
 
 	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
 		d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
@@ -822,11 +822,11 @@ func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt
 		fs.mu.RUnlock()
 		return err
 	}
-	if err := d.inode.setXattr(rp.Credentials(), &opts); err != nil {
-		fs.mu.RUnlock()
+	err = d.inode.setXattr(rp.Credentials(), &opts)
+	fs.mu.RUnlock()
+	if err != nil {
 		return err
 	}
-	fs.mu.RUnlock()
 
 	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil
@@ -840,11 +840,11 @@ func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath,
 		fs.mu.RUnlock()
 		return err
 	}
-	if err := d.inode.removeXattr(rp.Credentials(), name); err != nil {
-		fs.mu.RUnlock()
+	err = d.inode.removeXattr(rp.Credentials(), name)
+	fs.mu.RUnlock()
+	if err != nil {
 		return err
 	}
-	fs.mu.RUnlock()
 
 	d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
 	return nil