Add inotify support to kernfs and anonfs.

This is consistent with Linux. Some applications recursively add
files to inotify FD. They fail if they run into kernfs files
because inotify_add_watch(2) returns EPERM in that case.

PiperOrigin-RevId: 462642204

pkg/sentry/fsimpl/cgroupfs/cgroupfs.go

@@ -493,6 +493,7 @@ type dir struct {
 	kernfs.InodeAttrs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	implStatFS
 

pkg/sentry/fsimpl/devpts/devpts.go

@@ -152,6 +152,7 @@ type rootInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	rootInodeRefs
 

pkg/sentry/fsimpl/devpts/master.go

@@ -38,6 +38,7 @@ type masterInode struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/devpts/replica.go

@@ -37,6 +37,7 @@ type replicaInode struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/fuse/fusefs.go

@@ -305,6 +305,7 @@ type inode struct {
 	kernfs.InodeAttrs
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 
 	// the owning filesystem. fs is immutable.

pkg/sentry/fsimpl/host/host.go

@@ -99,6 +99,7 @@ type inode struct {
 	kernfs.InodeNotSymlink
 	kernfs.CachedMappable
 	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
+	kernfs.InodeWatches
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go

@@ -40,6 +40,7 @@ type DynamicBytesFile struct {
 	InodeNoopRefCount
 	InodeNotDirectory
 	InodeNotSymlink
+	InodeWatches
 
 	locks vfs.FileLocks
 	// data can additionally implement vfs.WritableDynamicBytesSource to support

pkg/sentry/fsimpl/kernfs/filesystem.go

@@ -395,6 +395,8 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.
 	if err != nil {
 		return err
 	}
+	parent.inode.Watches().Notify(ctx, pc, linux.IN_CREATE, 0, vfs.InodeEvent, false /* unlinked */)
+	d.inode.Watches().Notify(ctx, "", linux.IN_ATTRIB, 0, vfs.InodeEvent, false /* unlinked */)
 	var child Dentry
 	child.Init(fs, childI)
 	parent.insertChildLocked(pc, &child)
@@ -433,6 +435,7 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v
 	}
 	var child Dentry
 	child.Init(fs, childI)
+	parent.inode.Watches().Notify(ctx, pc, linux.IN_CREATE|linux.IN_ISDIR, 0, vfs.InodeEvent, false /* unlinked */)
 	parent.insertChildLocked(pc, &child)
 	return nil
 }
@@ -467,6 +470,7 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v
 	if err != nil {
 		return err
 	}
+	parent.inode.Watches().Notify(ctx, pc, linux.IN_CREATE, 0, vfs.InodeEvent, false /* unlinked */)
 	var newD Dentry
 	newD.Init(fs, newI)
 	parent.insertChildLocked(pc, &newD)
@@ -559,6 +563,9 @@ afterTrailingSymlink:
 	if len(pc) > linux.NAME_MAX {
 		return nil, linuxerr.ENAMETOOLONG
 	}
+	if parent.VFSDentry().IsDead() {
+		return nil, linuxerr.ENOENT
+	}
 	// Determine whether or not we need to create a file.
 	child, err := fs.stepExistingLocked(ctx, rp, parent, false /* mayFollowSymlinks */)
 	if linuxerr.Equals(linuxerr.ENOENT, err) {
@@ -582,6 +589,7 @@ afterTrailingSymlink:
 		// its destruction while fs.mu is unlocked.
 		child.IncRef()
 		unlock()
+		parent.inode.Watches().Notify(ctx, pc, linux.IN_CREATE, 0, vfs.PathEvent, false /* unlinked */)
 		fd, err := child.inode.Open(ctx, rp, &child, opts)
 		child.DecRef(ctx)
 		return fd, err
@@ -765,7 +773,9 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 		// deferDecRef so that fs.mu and dstDir.mu are unlocked by then.
 		fs.deferDecRef(replaced)
 		replaceVFSD = replaced.VFSDentry()
+		replaced.setDeleted()
 	}
+	vfs.InotifyRename(ctx, src.inode.Watches(), srcDir.inode.Watches(), dstDir.inode.Watches(), oldName, newName, src.isDir())
 	virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaceVFSD) // +checklocksforce: to may be nil, that's okay.
 	return nil
 }
@@ -810,9 +820,11 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 		return err
 	}
 	delete(parentDentry.children, d.name)
+	parentDentry.inode.Watches().Notify(ctx, d.name, linux.IN_DELETE|linux.IN_ISDIR, 0, vfs.InodeEvent, true /* unlinked */)
 	// Defer decref so that fs.mu and parentDentry.dirMu are unlocked by then.
 	fs.deferDecRef(d)
 	virtfs.CommitDeleteDentry(ctx, vfsd)
+	d.setDeleted()
 	return nil
 }
 
@@ -820,15 +832,24 @@ 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 {
 	fs.mu.RLock()
 	defer fs.processDeferredDecRefs(ctx)
-	defer fs.mu.RUnlock()
 	d, err := fs.walkExistingLocked(ctx, rp)
 	if err != nil {
+		fs.mu.RUnlock()
 		return err
 	}
 	if opts.Stat.Mask == 0 {
+		fs.mu.RUnlock()
 		return nil
 	}
-	return d.inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts)
+	err = d.inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts)
+	fs.mu.RUnlock()
+	if err != nil {
+		return err
+	}
+	if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
+		d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
+	}
+	return nil
 }
 
 // StatAt implements vfs.FilesystemImpl.StatAt.
@@ -885,6 +906,7 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ
 	if err != nil {
 		return err
 	}
+	parent.inode.Watches().Notify(ctx, pc, linux.IN_CREATE, 0, vfs.InodeEvent, false /* unlinked */)
 	var child Dentry
 	child.Init(fs, childI)
 	parent.insertChildLocked(pc, &child)
@@ -926,9 +948,11 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error
 		return err
 	}
 	delete(parentDentry.children, d.name)
+	vfs.InotifyRemoveChild(ctx, d.inode.Watches(), parentDentry.inode.Watches(), d.name)
 	// Defer decref so that fs.mu and parentDentry.dirMu are unlocked by then.
 	fs.deferDecRef(d)
 	virtfs.CommitDeleteDentry(ctx, vfsd)
+	d.setDeleted()
 	return nil
 }
 

pkg/sentry/fsimpl/kernfs/inode_impl_util.go

@@ -716,6 +716,7 @@ type StaticDirectory struct {
 	InodeNoStatFS
 	InodeNotSymlink
 	InodeTemporary
+	InodeWatches
 	OrderedChildren
 	StaticDirectoryRefs
 
@@ -796,3 +797,15 @@ type InodeNoStatFS struct{}
 func (*InodeNoStatFS) StatFS(context.Context, *vfs.Filesystem) (linux.Statfs, error) {
 	return linux.Statfs{}, linuxerr.ENOSYS
 }
+
+// InodeWatches partially implements Inode.
+//
+// +stateify savable
+type InodeWatches struct {
+	watches vfs.Watches
+}
+
+// Watches implements Inode.Watches.
+func (i *InodeWatches) Watches() *vfs.Watches {
+	return &i.watches
+}

pkg/sentry/fsimpl/kernfs/kernfs.go

@@ -244,6 +244,10 @@ type Dentry struct {
 	children map[string]*Dentry
 
 	inode Inode
+
+	// If deleted is non-zero, the file represented by this dentry has been
+	// deleted. deleted is accessed using atomic memory operations.
+	deleted atomicbitops.Uint32
 }
 
 // IncRef implements vfs.DentryImpl.IncRef.
@@ -339,6 +343,9 @@ func (d *Dentry) cacheLocked(ctx context.Context) {
 			d.fs.cachedDentriesLen--
 			d.cached = false
 		}
+		if d.isDeleted() {
+			d.inode.Watches().HandleDeletion(ctx)
+		}
 		d.destroyLocked(ctx)
 		return
 	}
@@ -420,7 +427,6 @@ func (d *Dentry) destroyLocked(ctx context.Context) {
 	}
 
 	d.inode.DecRef(ctx) // IncRef from Init.
-	d.inode = nil
 
 	if d.parent != nil {
 		d.parent.decRefLocked(ctx)
@@ -485,6 +491,14 @@ func (d *Dentry) VFSDentry() *vfs.Dentry {
 	return &d.vfsd
 }
 
+func (d *Dentry) isDeleted() bool {
+	return d.deleted.Load() != 0
+}
+
+func (d *Dentry) setDeleted() {
+	d.deleted.Store(1)
+}
+
 // isDir checks whether the dentry points to a directory inode.
 func (d *Dentry) isDir() bool {
 	return d.flags.Load()&dflagsIsDir != 0
@@ -496,15 +510,23 @@ func (d *Dentry) isSymlink() bool {
 }
 
 // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent.
-//
-// Although Linux technically supports inotify on pseudo filesystems (inotify
-// is implemented at the vfs layer), it is not particularly useful. It is left
-// unimplemented until someone actually needs it.
-func (d *Dentry) InotifyWithParent(ctx context.Context, events, cookie uint32, et vfs.EventType) {}
+func (d *Dentry) InotifyWithParent(ctx context.Context, events, cookie uint32, et vfs.EventType) {
+	if d.isDir() {
+		events |= linux.IN_ISDIR
+	}
+
+	d.fs.mu.RLock()
+	defer d.fs.mu.RUnlock()
+	// The ordering below is important, Linux always notifies the parent first.
+	if d.parent != nil {
+		d.parent.inode.Watches().Notify(ctx, d.name, events, cookie, et, d.isDeleted())
+	}
+	d.inode.Watches().Notify(ctx, "", events, cookie, et, d.isDeleted())
+}
 
 // Watches implements vfs.DentryImpl.Watches.
 func (d *Dentry) Watches() *vfs.Watches {
-	return nil
+	return d.inode.Watches()
 }
 
 // OnZeroWatches implements vfs.Dentry.OnZeroWatches.
@@ -680,6 +702,9 @@ type Inode interface {
 	// Valid should return true if this inode is still valid, or needs to
 	// be resolved again by a call to Lookup.
 	Valid(ctx context.Context) bool
+
+	// Watches returns the set of inotify watches associated with this inode.
+	Watches() *vfs.Watches
 }
 
 type inodeRefs interface {

pkg/sentry/fsimpl/kernfs/kernfs_test.go

@@ -106,6 +106,7 @@ type readonlyDir struct {
 	kernfs.InodeNoStatFS
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 
 	locks vfs.FileLocks
@@ -141,6 +142,7 @@ type dir struct {
 	kernfs.InodeNotSymlink
 	kernfs.InodeNoStatFS
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/kernfs/symlink.go

@@ -31,6 +31,7 @@ type StaticSymlink struct {
 	InodeNoopRefCount
 	InodeSymlink
 	InodeNoStatFS
+	InodeWatches
 
 	target string
 }

pkg/sentry/fsimpl/kernfs/synthetic_directory.go

@@ -33,6 +33,7 @@ type syntheticDirectory struct {
 	InodeAttrs
 	InodeNoStatFS
 	InodeNotSymlink
+	InodeWatches
 	OrderedChildren
 	syntheticDirectoryRefs
 

pkg/sentry/fsimpl/mqfs/root.go

@@ -33,6 +33,7 @@ type rootInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/pipefs/pipefs.go

@@ -92,6 +92,7 @@ type inode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeNoopRefCount
+	kernfs.InodeWatches
 
 	locks vfs.FileLocks
 	pipe  *pipe.VFSPipe

pkg/sentry/fsimpl/proc/subtasks.go

@@ -37,6 +37,7 @@ type subtasksInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	subtasksInodeRefs
 

pkg/sentry/fsimpl/proc/task.go

@@ -37,6 +37,7 @@ type taskInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	taskInodeRefs
 

pkg/sentry/fsimpl/proc/task_fds.go

@@ -114,6 +114,7 @@ type fdDirInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 }
 
@@ -197,6 +198,7 @@ type fdSymlink struct {
 	kernfs.InodeAttrs
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
+	kernfs.InodeWatches
 
 	fs   *filesystem
 	task *kernel.Task
@@ -256,6 +258,7 @@ type fdInfoDirInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 }
 

pkg/sentry/fsimpl/proc/task_files.go

@@ -394,6 +394,7 @@ type memInode struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 
 	task  *kernel.Task
 	locks vfs.FileLocks
@@ -675,6 +676,7 @@ type statusInode struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 
 	task  *kernel.Task
 	pidns *kernel.PIDNamespace
@@ -906,6 +908,7 @@ type exeSymlink struct {
 	kernfs.InodeAttrs
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
+	kernfs.InodeWatches
 
 	fs   *filesystem
 	task *kernel.Task
@@ -978,6 +981,7 @@ type cwdSymlink struct {
 	kernfs.InodeAttrs
 	kernfs.InodeNoopRefCount
 	kernfs.InodeSymlink
+	kernfs.InodeWatches
 
 	fs   *filesystem
 	task *kernel.Task
@@ -1152,6 +1156,7 @@ type namespaceInode struct {
 	kernfs.InodeNoopRefCount
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeWatches
 
 	locks vfs.FileLocks
 }

pkg/sentry/fsimpl/proc/tasks.go

@@ -43,6 +43,7 @@ type tasksInode struct {
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
+	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	tasksInodeRefs