proc: invalidate task inodes when a task is destroyed

PiperOrigin-RevId: 699233555

pkg/sentry/fsimpl/cgroupfs/cgroupfs.go

@@ -511,6 +511,7 @@ type dir struct {
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 	implStatFS
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/devpts/devpts.go

@@ -229,6 +229,7 @@ type rootInode struct {
 	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 	rootInodeRefs
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/devpts/master.go

@@ -39,6 +39,7 @@ type masterInode struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
+	kernfs.InodeFSOwned
 	kernfs.InodeWatches
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/devpts/replica.go

@@ -39,6 +39,7 @@ type replicaInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/fuse/inode.go

@@ -50,6 +50,7 @@ type inode struct {
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
 	kernfs.CachedMappable
+	kernfs.InodeFSOwned
 
 	// the owning filesystem. fs is immutable.
 	fs *filesystem

pkg/sentry/fsimpl/host/host.go

@@ -100,6 +100,7 @@ type inode struct {
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid.
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go

@@ -42,6 +42,7 @@ type DynamicBytesFile struct {
 	InodeNotDirectory
 	InodeNotSymlink
 	InodeWatches
+	InodeFSOwned
 
 	locks vfs.FileLocks
 	// data can additionally implement vfs.WritableDynamicBytesSource to support

pkg/sentry/fsimpl/kernfs/filesystem.go

@@ -109,7 +109,27 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir
 	parent.dirMu.Lock()
 	defer parent.dirMu.Unlock() // may be temporarily unlocked and re-locked below
 	child := parent.children[name]
-	for child != nil {
+	for {
+		if child == nil {
+			// Dentry isn't cached; it either doesn't exist or failed revalidation.
+			// Attempt to resolve it via Lookup.
+			childInode, err := parent.inode.Lookup(ctx, name)
+			if err != nil {
+				return nil, err
+			}
+			var newChild Dentry
+			newChild.Init(fs, childInode) // childInode's ref is transferred to newChild.
+			parent.insertChildLocked(name, &newChild)
+			child = &newChild
+
+			// Drop the ref on newChild. This will cause the dentry to get pruned
+			// from the dentry tree by the end of current filesystem operation
+			// (before returning to the VFS layer) if another ref is not picked on
+			// this dentry.
+			if !childInode.Keep() {
+				fs.deferDecRef(&newChild)
+			}
+		}
 		// Cached dentry exists, revalidate.
 		if child.inode.Valid(ctx, parent, name) {
 			break
@@ -120,26 +140,7 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir
 		parent.dirMu.Lock()
 		// Check for concurrent insertion of a new cached dentry.
 		child = parent.children[name]
-	}
-	if child == nil {
-		// Dentry isn't cached; it either doesn't exist or failed revalidation.
-		// Attempt to resolve it via Lookup.
-		childInode, err := parent.inode.Lookup(ctx, name)
-		if err != nil {
-			return nil, err
-		}
-		var newChild Dentry
-		newChild.Init(fs, childInode) // childInode's ref is transferred to newChild.
-		parent.insertChildLocked(name, &newChild)
-		child = &newChild
-
-		// Drop the ref on newChild. This will cause the dentry to get pruned
-		// from the dentry tree by the end of current filesystem operation
-		// (before returning to the VFS layer) if another ref is not picked on
-		// this dentry.
-		if !childInode.Keep() {
-			fs.deferDecRef(&newChild)
-		}
+
 	}
 	return child, nil
 }

pkg/sentry/fsimpl/kernfs/inode_impl_util.go

@@ -735,6 +735,7 @@ type StaticDirectory struct {
 	InodeWatches
 	OrderedChildren
 	StaticDirectoryRefs
+	InodeFSOwned
 
 	locks  vfs.FileLocks
 	fdOpts GenericDirectoryFDOptions
@@ -845,3 +846,15 @@ type InodeNotAnonymous struct{}
 func (*InodeNotAnonymous) Anonymous() bool {
 	return false
 }
+
+// InodeFSOwned represents inodes whose lifecycle is entirely managed by the
+// filesystem.
+//
+// +stateify savable
+type InodeFSOwned struct{}
+
+// AddInvalidateCallback implements Inode.AddInvalidateCallback.
+func (*InodeFSOwned) AddInvalidateCallback(d *Dentry) {}
+
+// RemoveInvalidateCallback implements Remove.AddInvalidateCallback.
+func (*InodeFSOwned) RemoveInvalidateCallback(d *Dentry) {}

pkg/sentry/fsimpl/kernfs/kernfs.go

@@ -317,6 +317,42 @@ func (d *Dentry) decRefLocked(ctx context.Context) {
 	}
 }
 
+// Invalidate invalidates the dentry and its children.
+func (d *Dentry) Invalidate(ctx context.Context) {
+	d.fs.mu.Lock()
+	defer d.fs.mu.Unlock()
+	if d.vfsd.IsDead() {
+		return
+	}
+	parent := d.parent.Load()
+	if parent == nil {
+		return
+	}
+	parent.dirMu.Lock()
+	if parent.vfsd.IsDead() {
+		parent.dirMu.Unlock()
+		return
+	}
+	child := parent.children[d.name]
+	if child != d {
+		parent.dirMu.Unlock()
+		return
+	}
+	delete(parent.children, d.name)
+	parent.dirMu.Unlock()
+
+	children := []*Dentry{d}
+	for i := 0; i < len(children); i++ {
+		for _, c := range children[i].children {
+			children = append(children, c)
+		}
+	}
+	d.fs.invalidateRemovedChildLocked(ctx, d.fs.vfsfs.VirtualFilesystem(), child)
+	for _, c := range children {
+		c.evictLocked(ctx)
+	}
+}
+
 // cacheLocked should be called after d's reference count becomes 0. The ref
 // count check may happen before acquiring d.fs.mu so there might be a race
 // condition where the ref count is increased again by the time the caller
@@ -392,6 +428,15 @@ func (d *Dentry) cacheLocked(ctx context.Context) {
 	// back down to fs.opts.maxCachedDentries, so we don't loop.
 }
 
+// DropDentryCache cleans up the dentry cache.
+func (fs *Filesystem) DropDentryCache(ctx context.Context) {
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+	for !fs.cachedDentries.Empty() {
+		fs.cachedDentries.Back().evictLocked(ctx)
+	}
+}
+
 // Preconditions:
 //   - fs.mu must be locked for writing.
 func (fs *Filesystem) evictCachedDentryLocked(ctx context.Context) {
@@ -451,6 +496,7 @@ func (d *Dentry) destroy(ctx context.Context) {
 		panic("dentry.destroy() called with references on the dentry")
 	}
 
+	d.inode.RemoveInvalidateCallback(d)
 	d.inode.DecRef(ctx) // IncRef from Init.
 
 	refs.Unregister(d)
@@ -505,6 +551,7 @@ func (d *Dentry) Init(fs *Filesystem, inode Inode) {
 		d.flags = atomicbitops.FromUint32(d.flags.RacyLoad() | dflagsIsSymlink)
 	}
 	refs.Register(d)
+	inode.AddInvalidateCallback(d)
 }
 
 // VFSDentry returns the generic vfs dentry for this kernfs dentry.
@@ -732,6 +779,9 @@ type Inode interface {
 	// Anonymous indicates that the Inode is anonymous. It will never have
 	// a name or parent.
 	Anonymous() bool
+
+	AddInvalidateCallback(d *Dentry)
+	RemoveInvalidateCallback(d *Dentry)
 }
 
 type inodeRefs interface {

pkg/sentry/fsimpl/kernfs/kernfs_test.go

@@ -109,6 +109,7 @@ type readonlyDir struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 }
@@ -146,6 +147,7 @@ type dir struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/kernfs/symlink.go

@@ -33,6 +33,7 @@ type StaticSymlink struct {
 	InodeSymlink
 	InodeNoStatFS
 	InodeWatches
+	InodeFSOwned
 
 	target string
 }

pkg/sentry/fsimpl/kernfs/synthetic_directory.go

@@ -37,6 +37,7 @@ type syntheticDirectory struct {
 	InodeWatches
 	OrderedChildren
 	syntheticDirectoryRefs
+	InodeFSOwned
 
 	locks vfs.FileLocks
 }

pkg/sentry/fsimpl/mqfs/root.go

@@ -36,6 +36,7 @@ type rootInode struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 }

pkg/sentry/fsimpl/nsfs/nsfs.go

@@ -83,6 +83,7 @@ type Inode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 	inodeRefs
 
 	locks     vfs.FileLocks

pkg/sentry/fsimpl/pipefs/pipefs.go

@@ -95,6 +95,7 @@ type inode struct {
 	kernfs.InodeNotSymlink
 	kernfs.InodeNoopRefCount
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	locks  vfs.FileLocks
 	pipe   *pipe.VFSPipe

pkg/sentry/fsimpl/proc/BUILD

@@ -66,6 +66,7 @@ go_library(
         "fd_dir_inode_refs.go",
         "fd_info_dir_inode_refs.go",
         "filesystem.go",
+        "filesystem_state.go",
         "proc_impl.go",
         "subtasks.go",
         "subtasks_inode_refs.go",

pkg/sentry/fsimpl/proc/filesystem_state.go

@@ -0,0 +1,23 @@
+// Copyright 2024 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 proc
+
+import (
+	"gvisor.dev/gvisor/pkg/context"
+)
+
+func (fs *filesystem) beforeSave() {
+	fs.DropDentryCache(context.Background())
+}

pkg/sentry/fsimpl/proc/subtasks.go

@@ -40,6 +40,7 @@ type subtasksInode struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 	subtasksInodeRefs
 
 	locks vfs.FileLocks

pkg/sentry/fsimpl/proc/task.go

@@ -113,6 +113,16 @@ func (fs *filesystem) newTaskInode(ctx context.Context, task *kernel.Task, pidns
 	return inode, nil
 }
 
+// AddInvalidateCallback implements kernfs.Inode.AddInvalidateCallback.
+func (i *taskInode) AddInvalidateCallback(d *kernfs.Dentry) {
+	i.task.RegisterOnDestroyAction(d, d.Invalidate)
+}
+
+// RemoveInvalidateCallback implements kernfs.Inode.AddInvalidateCallback.
+func (i *taskInode) RemoveInvalidateCallback(d *kernfs.Dentry) {
+	i.task.UnregisterOnDestroyAction(d)
+}
+
 // Valid implements kernfs.Inode.Valid. This inode remains valid as long
 // as the task is still running. When it's dead, another tasks with the same
 // PID could replace it.

pkg/sentry/fsimpl/proc/task_fds.go

@@ -118,6 +118,7 @@ type fdDirInode struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 }
 
 var _ kernfs.Inode = (*fdDirInode)(nil)
@@ -202,6 +203,7 @@ type fdSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	fs   *filesystem
 	task *kernel.Task
@@ -264,6 +266,7 @@ type fdInfoDirInode struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 }
 
 var _ kernfs.Inode = (*fdInfoDirInode)(nil)