diff --git a/pkg/sentry/fsimpl/cgroupfs/cgroupfs.go b/pkg/sentry/fsimpl/cgroupfs/cgroupfs.go
index 3bce26e90..c291a369f 100644
--- a/pkg/sentry/fsimpl/cgroupfs/cgroupfs.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go
index ab6427e06..dc5b416fc 100644
--- a/pkg/sentry/fsimpl/devpts/devpts.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/devpts/master.go b/pkg/sentry/fsimpl/devpts/master.go
index 5e0b8a8c7..498026fa6 100644
--- a/pkg/sentry/fsimpl/devpts/master.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/devpts/replica.go b/pkg/sentry/fsimpl/devpts/replica.go
index fedfc23ff..e7af3512a 100644
--- a/pkg/sentry/fsimpl/devpts/replica.go
+++ b/pkg/sentry/fsimpl/devpts/replica.go
@@ -39,6 +39,7 @@ type replicaInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 
diff --git a/pkg/sentry/fsimpl/fuse/inode.go b/pkg/sentry/fsimpl/fuse/inode.go
index 09c74a5b6..9f026372d 100644
--- a/pkg/sentry/fsimpl/fuse/inode.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go
index 1b2cfd5e5..2aa0a7491 100644
--- a/pkg/sentry/fsimpl/host/host.go
+++ b/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
 
diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go
index cae0a81f4..49770b554 100644
--- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go
index 7caa2025d..b3b045bf0 100644
--- a/pkg/sentry/fsimpl/kernfs/filesystem.go
+++ b/pkg/sentry/fsimpl/kernfs/filesystem.go
@@ -155,6 +155,14 @@ func (fs *Filesystem) invalidateRemovedChildLocked(ctx context.Context, vfsObj *
 		d := toInvalidate[len(toInvalidate)-1]
 		toInvalidate = toInvalidate[:len(toInvalidate)-1]
 
+		if d.cached {
+			// The dentry is removed from the cache when its
+			// reference counter drops to 0. It can't be removed
+			// from the cache here, because fs.mu isn't locked for
+			// write.
+			d.IncRef()
+			fs.deferDecRef(d)
+		}
 		if d.inode.Keep() {
 			fs.deferDecRef(d)
 		}
diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go
index 3ec38cb82..07c5aac31 100644
--- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go
+++ b/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{}
+
+// RegisterDentry implements Inode.RegisterDentry.
+func (*InodeFSOwned) RegisterDentry(d *Dentry) {}
+
+// UnregisterDentry implements Remove.UnregisterDentry.
+func (*InodeFSOwned) UnregisterDentry(d *Dentry) {}
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go
index e197f6433..d7d600626 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs.go
+++ b/pkg/sentry/fsimpl/kernfs/kernfs.go
@@ -317,6 +317,27 @@ func (d *Dentry) decRefLocked(ctx context.Context) {
 	}
 }
 
+// Invalidate invalidates the dentry and its children.
+func (d *Dentry) Invalidate(ctx context.Context) {
+	d.fs.mu.RLock()
+	defer d.fs.processDeferredDecRefs(ctx)
+	defer d.fs.mu.RUnlock()
+	parent := d.parent.Load()
+	if parent == nil {
+		return
+	}
+	parent.dirMu.Lock()
+	child := parent.children[d.name]
+	if child != d {
+		parent.dirMu.Unlock()
+		return
+	}
+	delete(parent.children, d.name)
+	parent.dirMu.Unlock()
+
+	d.fs.invalidateRemovedChildLocked(ctx, d.fs.vfsfs.VirtualFilesystem(), child)
+}
+
 // 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
@@ -451,6 +472,7 @@ func (d *Dentry) destroy(ctx context.Context) {
 		panic("dentry.destroy() called with references on the dentry")
 	}
 
+	d.inode.UnregisterDentry(d)
 	d.inode.DecRef(ctx) // IncRef from Init.
 
 	refs.Unregister(d)
@@ -505,6 +527,7 @@ func (d *Dentry) Init(fs *Filesystem, inode Inode) {
 		d.flags = atomicbitops.FromUint32(d.flags.RacyLoad() | dflagsIsSymlink)
 	}
 	refs.Register(d)
+	inode.RegisterDentry(d)
 }
 
 // VFSDentry returns the generic vfs dentry for this kernfs dentry.
@@ -732,6 +755,13 @@ type Inode interface {
 	// Anonymous indicates that the Inode is anonymous. It will never have
 	// a name or parent.
 	Anonymous() bool
+
+	// RegisterDentry is called when a new dentry representing the inode is
+	// created.
+	RegisterDentry(d *Dentry)
+
+	// UnregisterDentry is called when the specified dentry is destroyed.
+	UnregisterDentry(d *Dentry)
 }
 
 type inodeRefs interface {
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go
index e67537bda..6e487e3dd 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go
+++ b/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
 
diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go
index 0c9a7b3a6..933da99d5 100644
--- a/pkg/sentry/fsimpl/kernfs/symlink.go
+++ b/pkg/sentry/fsimpl/kernfs/symlink.go
@@ -33,6 +33,7 @@ type StaticSymlink struct {
 	InodeSymlink
 	InodeNoStatFS
 	InodeWatches
+	InodeFSOwned
 
 	target string
 }
diff --git a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go
index 4b28181d5..41700de43 100644
--- a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go
+++ b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go
@@ -37,6 +37,7 @@ type syntheticDirectory struct {
 	InodeWatches
 	OrderedChildren
 	syntheticDirectoryRefs
+	InodeFSOwned
 
 	locks vfs.FileLocks
 }
diff --git a/pkg/sentry/fsimpl/mqfs/root.go b/pkg/sentry/fsimpl/mqfs/root.go
index 7c26e2219..ecf01d500 100644
--- a/pkg/sentry/fsimpl/mqfs/root.go
+++ b/pkg/sentry/fsimpl/mqfs/root.go
@@ -36,6 +36,7 @@ type rootInode struct {
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
 	kernfs.OrderedChildren
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 }
diff --git a/pkg/sentry/fsimpl/nsfs/nsfs.go b/pkg/sentry/fsimpl/nsfs/nsfs.go
index 0d5dc2531..b4cb86820 100644
--- a/pkg/sentry/fsimpl/nsfs/nsfs.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/pipefs/pipefs.go b/pkg/sentry/fsimpl/pipefs/pipefs.go
index 73e85654f..b332fdbfb 100644
--- a/pkg/sentry/fsimpl/pipefs/pipefs.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD
index 794779d92..d4ca3b08e 100644
--- a/pkg/sentry/fsimpl/proc/BUILD
+++ b/pkg/sentry/fsimpl/proc/BUILD
@@ -1,3 +1,4 @@
+load("//pkg/sync/locking:locking.bzl", "declare_rwmutex")
 load("//tools:defs.bzl", "go_library", "go_test")
 load("//tools/go_generics:defs.bzl", "go_template_instance")
 
@@ -60,9 +61,17 @@ go_template_instance(
     },
 )
 
+declare_rwmutex(
+    name = "dentries_mutex",
+    out = "dentries_mutex.go",
+    package = "proc",
+    prefix = "dentries",
+)
+
 go_library(
     name = "proc",
     srcs = [
+        "dentries_mutex.go",
         "fd_dir_inode_refs.go",
         "fd_info_dir_inode_refs.go",
         "filesystem.go",
@@ -106,6 +115,7 @@ go_library(
         "//pkg/sentry/usage",
         "//pkg/sentry/vfs",
         "//pkg/sync",
+        "//pkg/sync/locking",
         "//pkg/tcpip/header",
         "//pkg/tcpip/network/ipv4",
         "//pkg/usermem",
diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go
index 60e4dcb07..dab64758a 100644
--- a/pkg/sentry/fsimpl/proc/subtasks.go
+++ b/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
diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go
index d41a3c001..606553239 100644
--- a/pkg/sentry/fsimpl/proc/task.go
+++ b/pkg/sentry/fsimpl/proc/task.go
@@ -45,6 +45,10 @@ type taskInode struct {
 	locks vfs.FileLocks
 
 	task *kernel.Task
+
+	dentriesMu dentriesRWMutex `state:"nosave"`
+	// dentries is a list of dentries to be invalidated when the task is destroyed.
+	dentries map[*kernfs.Dentry]struct{}
 }
 
 var _ kernfs.Inode = (*taskInode)(nil)
@@ -99,7 +103,10 @@ func (fs *filesystem) newTaskInode(ctx context.Context, task *kernel.Task, pidns
 		contents["cgroup"] = fs.newTaskOwnedInode(ctx, task, fs.NextIno(), 0444, &taskCgroupData{task: task})
 	}
 
-	taskInode := &taskInode{task: task}
+	taskInode := &taskInode{
+		task:     task,
+		dentries: make(map[*kernfs.Dentry]struct{}),
+	}
 	// Note: credentials are overridden by taskOwnedInode.
 	taskInode.InodeAttrs.Init(ctx, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
 	taskInode.InitRefs()
@@ -113,6 +120,44 @@ func (fs *filesystem) newTaskInode(ctx context.Context, task *kernel.Task, pidns
 	return inode, nil
 }
 
+func (i *taskInode) TaskDestroyAction(ctx context.Context) {
+	i.dentriesMu.Lock()
+	dentries := i.dentries
+	i.dentries = nil
+	i.dentriesMu.Unlock()
+
+	for d := range dentries {
+		d.Invalidate(ctx)
+	}
+}
+
+// RegisterDentry implements kernfs.Inode.RegisterDentry.
+func (i *taskInode) RegisterDentry(d *kernfs.Dentry) {
+	i.dentriesMu.Lock()
+	defer i.dentriesMu.Unlock()
+
+	if i.dentries == nil {
+		return
+	}
+	if len(i.dentries) == 0 && !i.task.RegisterOnDestroyAction(i) {
+		// The task has been destroyed.
+		i.dentries = nil
+		return
+	}
+	i.dentries[d] = struct{}{}
+}
+
+// UnregisterDentry implements kernfs.Inode.RegisterDentry.
+func (i *taskInode) UnregisterDentry(d *kernfs.Dentry) {
+	i.dentriesMu.Lock()
+	defer i.dentriesMu.Unlock()
+
+	delete(i.dentries, d)
+	if len(i.dentries) == 0 {
+		i.task.UnregisterOnDestroyAction(i)
+	}
+}
+
 // 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.
@@ -141,6 +186,17 @@ func (i *taskInode) DecRef(ctx context.Context) {
 	i.taskInodeRefs.DecRef(func() { i.Destroy(ctx) })
 }
 
+func (i *taskInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) {
+	i.dentriesMu.RLock()
+	if i.dentries == nil {
+		// The task has been destroyed and the inode invalidate callback has been executed.
+		i.dentriesMu.RUnlock()
+		return nil, linuxerr.ESRCH
+	}
+	i.dentriesMu.RUnlock()
+	return i.OrderedChildren.Lookup(ctx, name)
+}
+
 // taskOwnedInode implements kernfs.Inode and overrides inode owner with task
 // effective user and group.
 //
diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go
index 83996a446..14ad5d093 100644
--- a/pkg/sentry/fsimpl/proc/task_fds.go
+++ b/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)
diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go
index c0d818b47..366d9476a 100644
--- a/pkg/sentry/fsimpl/proc/task_files.go
+++ b/pkg/sentry/fsimpl/proc/task_files.go
@@ -422,6 +422,7 @@ type memInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	task  *kernel.Task
 	locks vfs.FileLocks
@@ -743,6 +744,7 @@ type statusInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	task  *kernel.Task
 	pidns *kernel.PIDNamespace
@@ -981,6 +983,7 @@ type exeSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	fs   *filesystem
 	task *kernel.Task
@@ -1054,6 +1057,7 @@ type cwdSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	fs   *filesystem
 	task *kernel.Task
@@ -1116,6 +1120,7 @@ type rootSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	fs   *filesystem
 	task *kernel.Task
@@ -1360,6 +1365,7 @@ type namespaceInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	locks vfs.FileLocks
 }
diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go
index e4948791c..68dac3160 100644
--- a/pkg/sentry/fsimpl/proc/tasks.go
+++ b/pkg/sentry/fsimpl/proc/tasks.go
@@ -46,6 +46,7 @@ type tasksInode 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
 	tasksInodeRefs
 
 	locks vfs.FileLocks
diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go
index 97eef2917..1febbbc5b 100644
--- a/pkg/sentry/fsimpl/proc/tasks_files.go
+++ b/pkg/sentry/fsimpl/proc/tasks_files.go
@@ -40,6 +40,7 @@ type selfSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	pidns *kernel.PIDNamespace
 }
@@ -83,6 +84,7 @@ type threadSelfSymlink struct {
 	kernfs.InodeNotAnonymous
 	kernfs.InodeSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 
 	pidns *kernel.PIDNamespace
 }
diff --git a/pkg/sentry/fsimpl/sockfs/sockfs.go b/pkg/sentry/fsimpl/sockfs/sockfs.go
index 455481447..3e3f01cf2 100644
--- a/pkg/sentry/fsimpl/sockfs/sockfs.go
+++ b/pkg/sentry/fsimpl/sockfs/sockfs.go
@@ -100,6 +100,7 @@ type inode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 }
 
 // Open implements kernfs.Inode.Open.
diff --git a/pkg/sentry/fsimpl/sys/kcov.go b/pkg/sentry/fsimpl/sys/kcov.go
index f71313543..1d6afcd95 100644
--- a/pkg/sentry/fsimpl/sys/kcov.go
+++ b/pkg/sentry/fsimpl/sys/kcov.go
@@ -43,6 +43,7 @@ type kcovInode struct {
 	kernfs.InodeNotDirectory
 	kernfs.InodeNotSymlink
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 	implStatFS
 }
 
diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go
index 2c3efd0af..cffb0bfaf 100644
--- a/pkg/sentry/fsimpl/sys/sys.go
+++ b/pkg/sentry/fsimpl/sys/sys.go
@@ -383,6 +383,7 @@ type dir struct {
 	kernfs.InodeNotSymlink
 	kernfs.InodeTemporary
 	kernfs.InodeWatches
+	kernfs.InodeFSOwned
 	kernfs.OrderedChildren
 
 	locks vfs.FileLocks
diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go
index 0c65b1828..7d62227ff 100644
--- a/pkg/sentry/kernel/task.go
+++ b/pkg/sentry/kernel/task.go
@@ -630,6 +630,10 @@ type Task struct {
 
 	// Origin is the origin of the task.
 	Origin TaskOrigin
+
+	// onDestroyAction is a set of callbacks that are executed when the
+	// task is destroyed.
+	onDestroyAction map[TaskDestroyAction]struct{}
 }
 
 // Task related metrics
diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go
index 0e8ac8067..38cc377cb 100644
--- a/pkg/sentry/kernel/task_exit.go
+++ b/pkg/sentry/kernel/task_exit.go
@@ -30,6 +30,7 @@ import (
 	"strconv"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/errors/linuxerr"
 	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -756,9 +757,58 @@ func (t *Task) exitNotifyLocked(fromPtraceDetach bool) {
 			// Do not clear t.parent. It may be still be needed after the task has exited
 			// (for example, to perform ptrace access checks on /proc/[pid] files).
 		}
+		t.execOnDestroyActions()
 	}
 }
 
+// TaskDestroyAction defines an action to be executed when a task is destroyed.
+type TaskDestroyAction interface {
+	TaskDestroyAction(ctx context.Context)
+}
+
+// RegisterOnDestroyAction registers an action to be executed when the task
+// is destroyed.
+//
+// It returns true if the action was successfully registered.
+// If the task is already terminated, it returns false.
+func (t *Task) RegisterOnDestroyAction(act TaskDestroyAction) bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if t.onDestroyAction == nil {
+		return false
+	}
+	t.onDestroyAction[act] = struct{}{}
+	return true
+}
+
+// UnregisterOnDestroyAction unregisters an action previously registered with
+// RegisterOnDestroyAction.
+func (t *Task) UnregisterOnDestroyAction(key TaskDestroyAction) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	delete(t.onDestroyAction, key)
+}
+
+func (t *Task) execOnDestroyActions() {
+	t.mu.Lock()
+	actions := t.onDestroyAction
+	t.onDestroyAction = nil
+	t.mu.Unlock()
+
+	if len(actions) == 0 {
+		return
+	}
+	// Block S/R until all actions is executed.
+	t.k.tasks.aioGoroutines.Add(1)
+	// Run in another goroutine to avoid extra lock dependencies.
+	go func() {
+		defer t.k.tasks.aioGoroutines.Done()
+		for act := range actions {
+			act.TaskDestroyAction(t)
+		}
+	}()
+}
+
 // Preconditions: The TaskSet mutex must be locked.
 func (t *Task) exitNotificationSignal(sig linux.Signal, receiver *Task) *linux.SignalInfo {
 	info := &linux.SignalInfo{
diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go
index 84260ba1e..c7616ee9b 100644
--- a/pkg/sentry/kernel/task_start.go
+++ b/pkg/sentry/kernel/task_start.go
@@ -151,30 +151,31 @@ func (ts *TaskSet) newTask(ctx context.Context, cfg *TaskConfig) (*Task, error)
 			parent:   cfg.Parent,
 			children: make(map[*Task]struct{}),
 		},
-		runState:       (*runApp)(nil),
-		interruptChan:  make(chan struct{}, 1),
-		signalMask:     atomicbitops.FromUint64(uint64(cfg.SignalMask)),
-		signalStack:    linux.SignalStack{Flags: linux.SS_DISABLE},
-		image:          *image,
-		fsContext:      cfg.FSContext,
-		fdTable:        cfg.FDTable,
-		k:              cfg.Kernel,
-		ptraceTracees:  make(map[*Task]struct{}),
-		allowedCPUMask: cfg.AllowedCPUMask.Copy(),
-		ioUsage:        &usage.IO{},
-		niceness:       cfg.Niceness,
-		utsns:          cfg.UTSNamespace,
-		ipcns:          cfg.IPCNamespace,
-		mountNamespace: cfg.MountNamespace,
-		rseqCPU:        -1,
-		rseqAddr:       cfg.RSeqAddr,
-		rseqSignature:  cfg.RSeqSignature,
-		futexWaiter:    futex.NewWaiter(),
-		containerID:    cfg.ContainerID,
-		cgroups:        make(map[Cgroup]struct{}),
-		userCounters:   cfg.UserCounters,
-		sessionKeyring: cfg.SessionKeyring,
-		Origin:         cfg.Origin,
+		runState:        (*runApp)(nil),
+		interruptChan:   make(chan struct{}, 1),
+		signalMask:      atomicbitops.FromUint64(uint64(cfg.SignalMask)),
+		signalStack:     linux.SignalStack{Flags: linux.SS_DISABLE},
+		image:           *image,
+		fsContext:       cfg.FSContext,
+		fdTable:         cfg.FDTable,
+		k:               cfg.Kernel,
+		ptraceTracees:   make(map[*Task]struct{}),
+		allowedCPUMask:  cfg.AllowedCPUMask.Copy(),
+		ioUsage:         &usage.IO{},
+		niceness:        cfg.Niceness,
+		utsns:           cfg.UTSNamespace,
+		ipcns:           cfg.IPCNamespace,
+		mountNamespace:  cfg.MountNamespace,
+		rseqCPU:         -1,
+		rseqAddr:        cfg.RSeqAddr,
+		rseqSignature:   cfg.RSeqSignature,
+		futexWaiter:     futex.NewWaiter(),
+		containerID:     cfg.ContainerID,
+		cgroups:         make(map[Cgroup]struct{}),
+		userCounters:    cfg.UserCounters,
+		sessionKeyring:  cfg.SessionKeyring,
+		Origin:          cfg.Origin,
+		onDestroyAction: make(map[TaskDestroyAction]struct{}),
 	}
 	t.netns = cfg.NetworkNamespace
 	t.creds.Store(cfg.Credentials)