Merge release-20200818.0-56-gdf3c105f4 (automated)

64 files changed, 3521 insertions, 372 deletions

diff --git a/pkg/abi/linux/linux_abi_autogen_unsafe.go b/pkg/abi/linux/linux_abi_autogen_unsafe.go
index 7d829e731..77a17f267 100644
--- a/pkg/abi/linux/linux_abi_autogen_unsafe.go
+++ b/pkg/abi/linux/linux_abi_autogen_unsafe.go
@@ -154,7 +154,7 @@ func (s *Statx) UnmarshalBytes(src []byte) {
 // Packed implements marshal.Marshallable.Packed.
 //go:nosplit
 func (s *Statx) Packed() bool {
-    return s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() && s.Mtime.Packed()
+    return s.Mtime.Packed() && s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed()
 }
 
 // MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
@@ -210,7 +210,7 @@ func (s *Statx) CopyOut(task marshal.Task, addr usermem.Addr) (int, error) {
 // CopyIn implements marshal.Marshallable.CopyIn.
 //go:nosplit
 func (s *Statx) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) {
-    if !s.Atime.Packed() && s.Btime.Packed() && s.Ctime.Packed() && s.Mtime.Packed() {
+    if !s.Ctime.Packed() && s.Mtime.Packed() && s.Atime.Packed() && s.Btime.Packed() {
         // Type Statx doesn't have a packed layout in memory, fall back to UnmarshalBytes.
         buf := task.CopyScratchBuffer(s.SizeBytes()) // escapes: okay.
         length, err := task.CopyInBytes(addr, buf) // escapes: okay.
@@ -644,7 +644,7 @@ func (f *FUSEHeaderIn) MarshalUnsafe(dst []byte) {
 
 // UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe.
 func (f *FUSEHeaderIn) UnmarshalUnsafe(src []byte) {
-    if f.Unique.Packed() && f.Opcode.Packed() {
+    if f.Opcode.Packed() && f.Unique.Packed() {
         safecopy.CopyOut(unsafe.Pointer(f), src)
     } else {
         // Type FUSEHeaderIn doesn't have a packed layout in memory, fallback to UnmarshalBytes.
@@ -2104,7 +2104,7 @@ func (i *IPTEntry) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) {
 
 // WriteTo implements io.WriterTo.WriteTo.
 func (i *IPTEntry) WriteTo(writer io.Writer) (int64, error) {
-    if !i.Counters.Packed() && i.IP.Packed() {
+    if !i.IP.Packed() && i.Counters.Packed() {
         // Type IPTEntry doesn't have a packed layout in memory, fall back to MarshalBytes.
         buf := make([]byte, i.SizeBytes())
         i.MarshalBytes(buf)
@@ -2215,7 +2215,7 @@ func (i *IPTIP) Packed() bool {
 
 // MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
 func (i *IPTIP) MarshalUnsafe(dst []byte) {
-    if i.SrcMask.Packed() && i.DstMask.Packed() && i.Src.Packed() && i.Dst.Packed() {
+    if i.Src.Packed() && i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed() {
         safecopy.CopyIn(dst, unsafe.Pointer(i))
     } else {
         // Type IPTIP doesn't have a packed layout in memory, fallback to MarshalBytes.
@@ -2266,7 +2266,7 @@ func (i *IPTIP) CopyOut(task marshal.Task, addr usermem.Addr) (int, error) {
 // CopyIn implements marshal.Marshallable.CopyIn.
 //go:nosplit
 func (i *IPTIP) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) {
-    if !i.Src.Packed() && i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed() {
+    if !i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed() && i.Src.Packed() {
         // Type IPTIP doesn't have a packed layout in memory, fall back to UnmarshalBytes.
         buf := task.CopyScratchBuffer(i.SizeBytes()) // escapes: okay.
         length, err := task.CopyInBytes(addr, buf) // escapes: okay.
@@ -2292,7 +2292,7 @@ func (i *IPTIP) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) {
 
 // WriteTo implements io.WriterTo.WriteTo.
 func (i *IPTIP) WriteTo(writer io.Writer) (int64, error) {
-    if !i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed() && i.Src.Packed() {
+    if !i.Src.Packed() && i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed() {
         // Type IPTIP doesn't have a packed layout in memory, fall back to MarshalBytes.
         buf := make([]byte, i.SizeBytes())
         i.MarshalBytes(buf)
@@ -3198,7 +3198,7 @@ func (i *IP6TIP) UnmarshalBytes(src []byte) {
 // Packed implements marshal.Marshallable.Packed.
 //go:nosplit
 func (i *IP6TIP) Packed() bool {
-    return i.Src.Packed() && i.Dst.Packed() && i.SrcMask.Packed() && i.DstMask.Packed()
+    return i.SrcMask.Packed() && i.DstMask.Packed() && i.Src.Packed() && i.Dst.Packed()
 }
 
 // MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
diff --git a/pkg/abi/linux/linux_arm64_abi_autogen_unsafe.go b/pkg/abi/linux/linux_arm64_abi_autogen_unsafe.go
index 33fb22105..79630e8d7 100644
--- a/pkg/abi/linux/linux_arm64_abi_autogen_unsafe.go
+++ b/pkg/abi/linux/linux_arm64_abi_autogen_unsafe.go
@@ -295,7 +295,7 @@ func (s *Stat) UnmarshalBytes(src []byte) {
 // Packed implements marshal.Marshallable.Packed.
 //go:nosplit
 func (s *Stat) Packed() bool {
-    return s.ATime.Packed() && s.MTime.Packed() && s.CTime.Packed()
+    return s.MTime.Packed() && s.CTime.Packed() && s.ATime.Packed()
 }
 
 // MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
@@ -351,7 +351,7 @@ func (s *Stat) CopyOut(task marshal.Task, addr usermem.Addr) (int, error) {
 // CopyIn implements marshal.Marshallable.CopyIn.
 //go:nosplit
 func (s *Stat) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) {
-    if !s.ATime.Packed() && s.MTime.Packed() && s.CTime.Packed() {
+    if !s.CTime.Packed() && s.ATime.Packed() && s.MTime.Packed() {
         // Type Stat doesn't have a packed layout in memory, fall back to UnmarshalBytes.
         buf := task.CopyScratchBuffer(s.SizeBytes()) // escapes: okay.
         length, err := task.CopyInBytes(addr, buf) // escapes: okay.
diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go
index 3f3a099bd..0eaff9087 100644
--- a/pkg/sentry/fsimpl/devpts/devpts.go
+++ b/pkg/sentry/fsimpl/devpts/devpts.go
@@ -83,6 +83,7 @@ func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds
 	}
 	root.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 1, linux.ModeDirectory|0555)
 	root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
+	root.EnableLeakCheck()
 	root.dentry.Init(root)
 
 	// Construct the pts master inode and dentry. Linux always uses inode
@@ -110,6 +111,7 @@ func (fs *filesystem) Release(ctx context.Context) {
 
 // rootInode is the root directory inode for the devpts mounts.
 type rootInode struct {
+	rootInodeRefs
 	kernfs.AlwaysValid
 	kernfs.InodeAttrs
 	kernfs.InodeDirectoryNoNewChildren
@@ -233,3 +235,8 @@ func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback,
 	}
 	return offset, nil
 }
+
+// DecRef implements kernfs.Inode.
+func (i *rootInode) DecRef(context.Context) {
+	i.rootInodeRefs.DecRef(i.Destroy)
+}
diff --git a/pkg/sentry/fsimpl/devpts/devpts_state_autogen.go b/pkg/sentry/fsimpl/devpts/devpts_state_autogen.go
index 5942f7bac..adc184d1b 100644
--- a/pkg/sentry/fsimpl/devpts/devpts_state_autogen.go
+++ b/pkg/sentry/fsimpl/devpts/devpts_state_autogen.go
@@ -120,6 +120,29 @@ func (x *queue) StateLoad(m state.Source) {
 	m.Load(4, &x.transformer)
 }
 
+func (x *rootInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/devpts.rootInodeRefs"
+}
+
+func (x *rootInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *rootInodeRefs) beforeSave() {}
+
+func (x *rootInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *rootInodeRefs) afterLoad() {}
+
+func (x *rootInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *Terminal) StateTypeName() string {
 	return "pkg/sentry/fsimpl/devpts.Terminal"
 }
@@ -157,5 +180,6 @@ func init() {
 	state.Register((*outputQueueTransformer)(nil))
 	state.Register((*inputQueueTransformer)(nil))
 	state.Register((*queue)(nil))
+	state.Register((*rootInodeRefs)(nil))
 	state.Register((*Terminal)(nil))
 }
diff --git a/pkg/sentry/fsimpl/devpts/root_inode_refs.go b/pkg/sentry/fsimpl/devpts/root_inode_refs.go
new file mode 100644
index 000000000..051801202
--- /dev/null
+++ b/pkg/sentry/fsimpl/devpts/root_inode_refs.go
@@ -0,0 +1,118 @@
+package devpts
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var rootInodeownerType *rootInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type rootInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *rootInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, rootInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *rootInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*rootInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *rootInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *rootInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, rootInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *rootInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *rootInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, rootInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/fuse/fuse_state_autogen.go b/pkg/sentry/fsimpl/fuse/fuse_state_autogen.go
index e4ce04322..f72fe342e 100644
--- a/pkg/sentry/fsimpl/fuse/fuse_state_autogen.go
+++ b/pkg/sentry/fsimpl/fuse/fuse_state_autogen.go
@@ -99,6 +99,29 @@ func (x *futureResponse) StateLoad(m state.Source) {
 	m.Load(3, &x.data)
 }
 
+func (x *inodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/fuse.inodeRefs"
+}
+
+func (x *inodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *inodeRefs) beforeSave() {}
+
+func (x *inodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *inodeRefs) afterLoad() {}
+
+func (x *inodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *requestList) StateTypeName() string {
 	return "pkg/sentry/fsimpl/fuse.requestList"
 }
@@ -155,6 +178,7 @@ func init() {
 	state.Register((*Request)(nil))
 	state.Register((*Response)(nil))
 	state.Register((*futureResponse)(nil))
+	state.Register((*inodeRefs)(nil))
 	state.Register((*requestList)(nil))
 	state.Register((*requestEntry)(nil))
 }
diff --git a/pkg/sentry/fsimpl/fuse/fusefs.go b/pkg/sentry/fsimpl/fuse/fusefs.go
index 44021ee4b..9717c0e15 100644
--- a/pkg/sentry/fsimpl/fuse/fusefs.go
+++ b/pkg/sentry/fsimpl/fuse/fusefs.go
@@ -198,6 +198,7 @@ func (fs *filesystem) Release(ctx context.Context) {
 
 // inode implements kernfs.Inode.
 type inode struct {
+	inodeRefs
 	kernfs.InodeAttrs
 	kernfs.InodeNoDynamicLookup
 	kernfs.InodeNotSymlink
@@ -213,6 +214,7 @@ func (fs *filesystem) newInode(creds *auth.Credentials, mode linux.FileMode) *ke
 	i := &inode{}
 	i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755)
 	i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
+	i.EnableLeakCheck()
 	i.dentry.Init(i)
 
 	return &i.dentry
@@ -324,3 +326,8 @@ func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptio
 
 	return statFromFUSEAttr(out.Attr, opts.Mask, fusefs.devMinor), nil
 }
+
+// DecRef implements kernfs.Inode.
+func (i *inode) DecRef(context.Context) {
+	i.inodeRefs.DecRef(i.Destroy)
+}
diff --git a/pkg/sentry/fsimpl/fuse/inode_refs.go b/pkg/sentry/fsimpl/fuse/inode_refs.go
new file mode 100644
index 000000000..6b9456e1d
--- /dev/null
+++ b/pkg/sentry/fsimpl/fuse/inode_refs.go
@@ -0,0 +1,118 @@
+package fuse
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var inodeownerType *inode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type inodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *inodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, inodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *inodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*inodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *inodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *inodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, inodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *inodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *inodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, inodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/host/connected_endpoint_refs.go b/pkg/sentry/fsimpl/host/connected_endpoint_refs.go
index 3b7bf599e..babb3f664 100644
--- a/pkg/sentry/fsimpl/host/connected_endpoint_refs.go
+++ b/pkg/sentry/fsimpl/host/connected_endpoint_refs.go
@@ -1,6 +1,7 @@
 package host
 
 import (
+	"fmt"
 	"runtime"
 	"sync/atomic"
 
@@ -18,6 +19,11 @@ var ConnectedEndpointownerType *ConnectedEndpoint
 // Note that the number of references is actually refCount + 1 so that a default
 // zero-value Refs object contains one reference.
 //
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
 // +stateify savable
 type ConnectedEndpointRefs struct {
 	// refCount is composed of two fields:
@@ -62,7 +68,7 @@ func (r *ConnectedEndpointRefs) ReadRefs() int64 {
 //go:nosplit
 func (r *ConnectedEndpointRefs) IncRef() {
 	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
-		panic("Incrementing non-positive ref count")
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, ConnectedEndpointownerType))
 	}
 }
 
@@ -101,7 +107,7 @@ func (r *ConnectedEndpointRefs) TryIncRef() bool {
 func (r *ConnectedEndpointRefs) DecRef(destroy func()) {
 	switch v := atomic.AddInt64(&r.refCount, -1); {
 	case v < -1:
-		panic("Decrementing non-positive ref count")
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, ConnectedEndpointownerType))
 
 	case v == -1:
 
diff --git a/pkg/sentry/fsimpl/host/inode_refs.go b/pkg/sentry/fsimpl/host/inode_refs.go
index 55c0fb3a9..17f90ce4a 100644
--- a/pkg/sentry/fsimpl/host/inode_refs.go
+++ b/pkg/sentry/fsimpl/host/inode_refs.go
@@ -1,6 +1,7 @@
 package host
 
 import (
+	"fmt"
 	"runtime"
 	"sync/atomic"
 
@@ -18,6 +19,11 @@ var inodeownerType *inode
 // Note that the number of references is actually refCount + 1 so that a default
 // zero-value Refs object contains one reference.
 //
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
 // +stateify savable
 type inodeRefs struct {
 	// refCount is composed of two fields:
@@ -62,7 +68,7 @@ func (r *inodeRefs) ReadRefs() int64 {
 //go:nosplit
 func (r *inodeRefs) IncRef() {
 	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
-		panic("Incrementing non-positive ref count")
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, inodeownerType))
 	}
 }
 
@@ -101,7 +107,7 @@ func (r *inodeRefs) TryIncRef() bool {
 func (r *inodeRefs) DecRef(destroy func()) {
 	switch v := atomic.AddInt64(&r.refCount, -1); {
 	case v < -1:
-		panic("Decrementing non-positive ref count")
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, inodeownerType))
 
 	case v == -1:
 
diff --git a/pkg/sentry/fsimpl/kernfs/dentry_refs.go b/pkg/sentry/fsimpl/kernfs/dentry_refs.go
new file mode 100644
index 000000000..79863b3bc
--- /dev/null
+++ b/pkg/sentry/fsimpl/kernfs/dentry_refs.go
@@ -0,0 +1,118 @@
+package kernfs
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var DentryownerType *Dentry
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type DentryRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *DentryRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, DentryownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *DentryRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*DentryRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *DentryRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *DentryRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, DentryownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *DentryRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *DentryRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, DentryownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go
index 885856868..f442a5606 100644
--- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go
+++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go
@@ -20,7 +20,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
@@ -344,8 +343,6 @@ type OrderedChildrenOptions struct {
 //
 // Must be initialize with Init before first use.
 type OrderedChildren struct {
-	refs.AtomicRefCount
-
 	// Can children be modified by user syscalls? It set to false, interface
 	// methods that would modify the children return EPERM. Immutable.
 	writable bool
@@ -361,14 +358,14 @@ func (o *OrderedChildren) Init(opts OrderedChildrenOptions) {
 	o.set = make(map[string]*slot)
 }
 
-// DecRef implements Inode.DecRef.
-func (o *OrderedChildren) DecRef(ctx context.Context) {
-	o.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) {
-		o.mu.Lock()
-		defer o.mu.Unlock()
-		o.order.Reset()
-		o.set = nil
-	})
+// Destroy clears the children stored in o. It should be called by structs
+// embedding OrderedChildren upon destruction, i.e. when their reference count
+// reaches zero.
+func (o *OrderedChildren) Destroy() {
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	o.order.Reset()
+	o.set = nil
 }
 
 // Populate inserts children into this OrderedChildren, and d's dentry
@@ -549,6 +546,7 @@ func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D
 //
 // +stateify savable
 type StaticDirectory struct {
+	StaticDirectoryRefs
 	InodeNotSymlink
 	InodeDirectoryNoNewChildren
 	InodeAttrs
@@ -594,11 +592,16 @@ func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd
 	return fd.VFSFileDescription(), nil
 }
 
-// SetStat implements Inode.SetStat not allowing inode attributes to be changed.
+// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed.
 func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error {
 	return syserror.EPERM
 }
 
+// DecRef implements kernfs.Inode.
+func (s *StaticDirectory) DecRef(context.Context) {
+	s.StaticDirectoryRefs.DecRef(s.Destroy)
+}
+
 // AlwaysValid partially implements kernfs.inodeDynamicLookup.
 type AlwaysValid struct{}
 
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go
index 51dbc050c..ca3685800 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs.go
+++ b/pkg/sentry/fsimpl/kernfs/kernfs.go
@@ -57,7 +57,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
@@ -161,9 +160,9 @@ const (
 //
 // Must be initialized by Init prior to first use.
 type Dentry struct {
-	vfsd vfs.Dentry
+	DentryRefs
 
-	refs.AtomicRefCount
+	vfsd vfs.Dentry
 
 	// flags caches useful information about the dentry from the inode. See the
 	// dflags* consts above. Must be accessed by atomic ops.
@@ -194,6 +193,7 @@ func (d *Dentry) Init(inode Inode) {
 	if ftype == linux.ModeSymlink {
 		d.flags |= dflagsIsSymlink
 	}
+	d.EnableLeakCheck()
 }
 
 // VFSDentry returns the generic vfs dentry for this kernfs dentry.
@@ -213,16 +213,14 @@ func (d *Dentry) isSymlink() bool {
 
 // DecRef implements vfs.DentryImpl.DecRef.
 func (d *Dentry) DecRef(ctx context.Context) {
-	d.AtomicRefCount.DecRefWithDestructor(ctx, d.destroy)
-}
-
-// Precondition: Dentry must be removed from VFS' dentry cache.
-func (d *Dentry) destroy(ctx context.Context) {
-	d.inode.DecRef(ctx) // IncRef from Init.
-	d.inode = nil
-	if d.parent != nil {
-		d.parent.DecRef(ctx) // IncRef from Dentry.InsertChild.
-	}
+	// Before the destructor is called, Dentry must be removed from VFS' dentry cache.
+	d.DentryRefs.DecRef(func() {
+		d.inode.DecRef(ctx) // IncRef from Init.
+		d.inode = nil
+		if d.parent != nil {
+			d.parent.DecRef(ctx) // IncRef from Dentry.InsertChild.
+		}
+	})
 }
 
 // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent.
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_state_autogen.go b/pkg/sentry/fsimpl/kernfs/kernfs_state_autogen.go
index f396affb0..12aaf797f 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs_state_autogen.go
+++ b/pkg/sentry/fsimpl/kernfs/kernfs_state_autogen.go
@@ -6,6 +6,29 @@ import (
 	"gvisor.dev/gvisor/pkg/state"
 )
 
+func (x *DentryRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/kernfs.DentryRefs"
+}
+
+func (x *DentryRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *DentryRefs) beforeSave() {}
+
+func (x *DentryRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *DentryRefs) afterLoad() {}
+
+func (x *DentryRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *DynamicBytesFile) StateTypeName() string {
 	return "pkg/sentry/fsimpl/kernfs.DynamicBytesFile"
 }
@@ -85,6 +108,7 @@ func (x *StaticDirectory) StateTypeName() string {
 
 func (x *StaticDirectory) StateFields() []string {
 	return []string{
+		"StaticDirectoryRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeAttrs",
@@ -99,25 +123,27 @@ func (x *StaticDirectory) beforeSave() {}
 
 func (x *StaticDirectory) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeAttrs)
-	m.Save(3, &x.InodeNoDynamicLookup)
-	m.Save(4, &x.OrderedChildren)
-	m.Save(5, &x.locks)
-	m.Save(6, &x.fdOpts)
+	m.Save(0, &x.StaticDirectoryRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeAttrs)
+	m.Save(4, &x.InodeNoDynamicLookup)
+	m.Save(5, &x.OrderedChildren)
+	m.Save(6, &x.locks)
+	m.Save(7, &x.fdOpts)
 }
 
 func (x *StaticDirectory) afterLoad() {}
 
 func (x *StaticDirectory) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeAttrs)
-	m.Load(3, &x.InodeNoDynamicLookup)
-	m.Load(4, &x.OrderedChildren)
-	m.Load(5, &x.locks)
-	m.Load(6, &x.fdOpts)
+	m.Load(0, &x.StaticDirectoryRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeAttrs)
+	m.Load(4, &x.InodeNoDynamicLookup)
+	m.Load(5, &x.OrderedChildren)
+	m.Load(6, &x.locks)
+	m.Load(7, &x.fdOpts)
 }
 
 func (x *slotList) StateTypeName() string {
@@ -172,10 +198,35 @@ func (x *slotEntry) StateLoad(m state.Source) {
 	m.Load(1, &x.prev)
 }
 
+func (x *StaticDirectoryRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/kernfs.StaticDirectoryRefs"
+}
+
+func (x *StaticDirectoryRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *StaticDirectoryRefs) beforeSave() {}
+
+func (x *StaticDirectoryRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *StaticDirectoryRefs) afterLoad() {}
+
+func (x *StaticDirectoryRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func init() {
+	state.Register((*DentryRefs)(nil))
 	state.Register((*DynamicBytesFile)(nil))
 	state.Register((*DynamicBytesFD)(nil))
 	state.Register((*StaticDirectory)(nil))
 	state.Register((*slotList)(nil))
 	state.Register((*slotEntry)(nil))
+	state.Register((*StaticDirectoryRefs)(nil))
 }
diff --git a/pkg/sentry/fsimpl/kernfs/static_directory_refs.go b/pkg/sentry/fsimpl/kernfs/static_directory_refs.go
new file mode 100644
index 000000000..478b04bdd
--- /dev/null
+++ b/pkg/sentry/fsimpl/kernfs/static_directory_refs.go
@@ -0,0 +1,118 @@
+package kernfs
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var StaticDirectoryownerType *StaticDirectory
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type StaticDirectoryRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *StaticDirectoryRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, StaticDirectoryownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *StaticDirectoryRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*StaticDirectoryRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *StaticDirectoryRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *StaticDirectoryRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, StaticDirectoryownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *StaticDirectoryRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *StaticDirectoryRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, StaticDirectoryownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/proc/fd_dir_inode_refs.go b/pkg/sentry/fsimpl/proc/fd_dir_inode_refs.go
new file mode 100644
index 000000000..9431c1506
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/fd_dir_inode_refs.go
@@ -0,0 +1,118 @@
+package proc
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var fdDirInodeownerType *fdDirInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type fdDirInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *fdDirInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, fdDirInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *fdDirInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*fdDirInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *fdDirInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *fdDirInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, fdDirInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *fdDirInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *fdDirInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, fdDirInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/proc/fd_info_dir_inode_refs.go b/pkg/sentry/fsimpl/proc/fd_info_dir_inode_refs.go
new file mode 100644
index 000000000..872b20eb0
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/fd_info_dir_inode_refs.go
@@ -0,0 +1,118 @@
+package proc
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var fdInfoDirInodeownerType *fdInfoDirInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type fdInfoDirInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *fdInfoDirInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, fdInfoDirInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *fdInfoDirInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*fdInfoDirInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *fdInfoDirInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *fdInfoDirInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, fdInfoDirInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *fdInfoDirInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *fdInfoDirInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, fdInfoDirInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/proc/proc_state_autogen.go b/pkg/sentry/fsimpl/proc/proc_state_autogen.go
index a7e43b7a7..907ef38e0 100644
--- a/pkg/sentry/fsimpl/proc/proc_state_autogen.go
+++ b/pkg/sentry/fsimpl/proc/proc_state_autogen.go
@@ -6,6 +6,52 @@ import (
 	"gvisor.dev/gvisor/pkg/state"
 )
 
+func (x *fdDirInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.fdDirInodeRefs"
+}
+
+func (x *fdDirInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *fdDirInodeRefs) beforeSave() {}
+
+func (x *fdDirInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *fdDirInodeRefs) afterLoad() {}
+
+func (x *fdDirInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
+func (x *fdInfoDirInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.fdInfoDirInodeRefs"
+}
+
+func (x *fdInfoDirInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *fdInfoDirInodeRefs) beforeSave() {}
+
+func (x *fdInfoDirInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *fdInfoDirInodeRefs) afterLoad() {}
+
+func (x *fdInfoDirInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *FilesystemType) StateTypeName() string {
 	return "pkg/sentry/fsimpl/proc.FilesystemType"
 }
@@ -31,6 +77,7 @@ func (x *subtasksInode) StateTypeName() string {
 
 func (x *subtasksInode) StateFields() []string {
 	return []string{
+		"subtasksInodeRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeAttrs",
@@ -48,31 +95,56 @@ func (x *subtasksInode) beforeSave() {}
 
 func (x *subtasksInode) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeAttrs)
-	m.Save(3, &x.OrderedChildren)
-	m.Save(4, &x.AlwaysValid)
-	m.Save(5, &x.locks)
-	m.Save(6, &x.fs)
-	m.Save(7, &x.task)
-	m.Save(8, &x.pidns)
-	m.Save(9, &x.cgroupControllers)
+	m.Save(0, &x.subtasksInodeRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeAttrs)
+	m.Save(4, &x.OrderedChildren)
+	m.Save(5, &x.AlwaysValid)
+	m.Save(6, &x.locks)
+	m.Save(7, &x.fs)
+	m.Save(8, &x.task)
+	m.Save(9, &x.pidns)
+	m.Save(10, &x.cgroupControllers)
 }
 
 func (x *subtasksInode) afterLoad() {}
 
 func (x *subtasksInode) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeAttrs)
-	m.Load(3, &x.OrderedChildren)
-	m.Load(4, &x.AlwaysValid)
-	m.Load(5, &x.locks)
-	m.Load(6, &x.fs)
-	m.Load(7, &x.task)
-	m.Load(8, &x.pidns)
-	m.Load(9, &x.cgroupControllers)
+	m.Load(0, &x.subtasksInodeRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeAttrs)
+	m.Load(4, &x.OrderedChildren)
+	m.Load(5, &x.AlwaysValid)
+	m.Load(6, &x.locks)
+	m.Load(7, &x.fs)
+	m.Load(8, &x.task)
+	m.Load(9, &x.pidns)
+	m.Load(10, &x.cgroupControllers)
+}
+
+func (x *subtasksInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.subtasksInodeRefs"
+}
+
+func (x *subtasksInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *subtasksInodeRefs) beforeSave() {}
+
+func (x *subtasksInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *subtasksInodeRefs) afterLoad() {}
+
+func (x *subtasksInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
 }
 
 func (x *taskInode) StateTypeName() string {
@@ -81,6 +153,7 @@ func (x *taskInode) StateTypeName() string {
 
 func (x *taskInode) StateFields() []string {
 	return []string{
+		"taskInodeRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeNoDynamicLookup",
@@ -95,25 +168,27 @@ func (x *taskInode) beforeSave() {}
 
 func (x *taskInode) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeNoDynamicLookup)
-	m.Save(3, &x.InodeAttrs)
-	m.Save(4, &x.OrderedChildren)
-	m.Save(5, &x.locks)
-	m.Save(6, &x.task)
+	m.Save(0, &x.taskInodeRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeNoDynamicLookup)
+	m.Save(4, &x.InodeAttrs)
+	m.Save(5, &x.OrderedChildren)
+	m.Save(6, &x.locks)
+	m.Save(7, &x.task)
 }
 
 func (x *taskInode) afterLoad() {}
 
 func (x *taskInode) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeNoDynamicLookup)
-	m.Load(3, &x.InodeAttrs)
-	m.Load(4, &x.OrderedChildren)
-	m.Load(5, &x.locks)
-	m.Load(6, &x.task)
+	m.Load(0, &x.taskInodeRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeNoDynamicLookup)
+	m.Load(4, &x.InodeAttrs)
+	m.Load(5, &x.OrderedChildren)
+	m.Load(6, &x.locks)
+	m.Load(7, &x.task)
 }
 
 func (x *fdDirInode) StateTypeName() string {
@@ -122,6 +197,7 @@ func (x *fdDirInode) StateTypeName() string {
 
 func (x *fdDirInode) StateFields() []string {
 	return []string{
+		"fdDirInodeRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeAttrs",
@@ -135,23 +211,25 @@ func (x *fdDirInode) beforeSave() {}
 
 func (x *fdDirInode) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeAttrs)
-	m.Save(3, &x.OrderedChildren)
-	m.Save(4, &x.AlwaysValid)
-	m.Save(5, &x.fdDir)
+	m.Save(0, &x.fdDirInodeRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeAttrs)
+	m.Save(4, &x.OrderedChildren)
+	m.Save(5, &x.AlwaysValid)
+	m.Save(6, &x.fdDir)
 }
 
 func (x *fdDirInode) afterLoad() {}
 
 func (x *fdDirInode) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeAttrs)
-	m.Load(3, &x.OrderedChildren)
-	m.Load(4, &x.AlwaysValid)
-	m.Load(5, &x.fdDir)
+	m.Load(0, &x.fdDirInodeRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeAttrs)
+	m.Load(4, &x.OrderedChildren)
+	m.Load(5, &x.AlwaysValid)
+	m.Load(6, &x.fdDir)
 }
 
 func (x *fdSymlink) StateTypeName() string {
@@ -195,6 +273,7 @@ func (x *fdInfoDirInode) StateTypeName() string {
 
 func (x *fdInfoDirInode) StateFields() []string {
 	return []string{
+		"fdInfoDirInodeRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeAttrs",
@@ -208,23 +287,25 @@ func (x *fdInfoDirInode) beforeSave() {}
 
 func (x *fdInfoDirInode) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeAttrs)
-	m.Save(3, &x.OrderedChildren)
-	m.Save(4, &x.AlwaysValid)
-	m.Save(5, &x.fdDir)
+	m.Save(0, &x.fdInfoDirInodeRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeAttrs)
+	m.Save(4, &x.OrderedChildren)
+	m.Save(5, &x.AlwaysValid)
+	m.Save(6, &x.fdDir)
 }
 
 func (x *fdInfoDirInode) afterLoad() {}
 
 func (x *fdInfoDirInode) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeAttrs)
-	m.Load(3, &x.OrderedChildren)
-	m.Load(4, &x.AlwaysValid)
-	m.Load(5, &x.fdDir)
+	m.Load(0, &x.fdInfoDirInodeRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeAttrs)
+	m.Load(4, &x.OrderedChildren)
+	m.Load(5, &x.AlwaysValid)
+	m.Load(6, &x.fdDir)
 }
 
 func (x *fdInfoData) StateTypeName() string {
@@ -234,7 +315,6 @@ func (x *fdInfoData) StateTypeName() string {
 func (x *fdInfoData) StateFields() []string {
 	return []string{
 		"DynamicBytesFile",
-		"AtomicRefCount",
 		"task",
 		"fd",
 	}
@@ -245,18 +325,16 @@ func (x *fdInfoData) beforeSave() {}
 func (x *fdInfoData) StateSave(m state.Sink) {
 	x.beforeSave()
 	m.Save(0, &x.DynamicBytesFile)
-	m.Save(1, &x.AtomicRefCount)
-	m.Save(2, &x.task)
-	m.Save(3, &x.fd)
+	m.Save(1, &x.task)
+	m.Save(2, &x.fd)
 }
 
 func (x *fdInfoData) afterLoad() {}
 
 func (x *fdInfoData) StateLoad(m state.Source) {
 	m.Load(0, &x.DynamicBytesFile)
-	m.Load(1, &x.AtomicRefCount)
-	m.Load(2, &x.task)
-	m.Load(3, &x.fd)
+	m.Load(1, &x.task)
+	m.Load(2, &x.fd)
 }
 
 func (x *auxvData) StateTypeName() string {
@@ -670,6 +748,29 @@ func (x *mountsData) StateLoad(m state.Source) {
 	m.Load(1, &x.task)
 }
 
+func (x *taskInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.taskInodeRefs"
+}
+
+func (x *taskInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *taskInodeRefs) beforeSave() {}
+
+func (x *taskInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *taskInodeRefs) afterLoad() {}
+
+func (x *taskInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *ifinet6) StateTypeName() string {
 	return "pkg/sentry/fsimpl/proc.ifinet6"
 }
@@ -910,6 +1011,7 @@ func (x *tasksInode) StateTypeName() string {
 
 func (x *tasksInode) StateFields() []string {
 	return []string{
+		"tasksInodeRefs",
 		"InodeNotSymlink",
 		"InodeDirectoryNoNewChildren",
 		"InodeAttrs",
@@ -928,33 +1030,35 @@ func (x *tasksInode) beforeSave() {}
 
 func (x *tasksInode) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.InodeNotSymlink)
-	m.Save(1, &x.InodeDirectoryNoNewChildren)
-	m.Save(2, &x.InodeAttrs)
-	m.Save(3, &x.OrderedChildren)
-	m.Save(4, &x.AlwaysValid)
-	m.Save(5, &x.locks)
-	m.Save(6, &x.fs)
-	m.Save(7, &x.pidns)
-	m.Save(8, &x.selfSymlink)
-	m.Save(9, &x.threadSelfSymlink)
-	m.Save(10, &x.cgroupControllers)
+	m.Save(0, &x.tasksInodeRefs)
+	m.Save(1, &x.InodeNotSymlink)
+	m.Save(2, &x.InodeDirectoryNoNewChildren)
+	m.Save(3, &x.InodeAttrs)
+	m.Save(4, &x.OrderedChildren)
+	m.Save(5, &x.AlwaysValid)
+	m.Save(6, &x.locks)
+	m.Save(7, &x.fs)
+	m.Save(8, &x.pidns)
+	m.Save(9, &x.selfSymlink)
+	m.Save(10, &x.threadSelfSymlink)
+	m.Save(11, &x.cgroupControllers)
 }
 
 func (x *tasksInode) afterLoad() {}
 
 func (x *tasksInode) StateLoad(m state.Source) {
-	m.Load(0, &x.InodeNotSymlink)
-	m.Load(1, &x.InodeDirectoryNoNewChildren)
-	m.Load(2, &x.InodeAttrs)
-	m.Load(3, &x.OrderedChildren)
-	m.Load(4, &x.AlwaysValid)
-	m.Load(5, &x.locks)
-	m.Load(6, &x.fs)
-	m.Load(7, &x.pidns)
-	m.Load(8, &x.selfSymlink)
-	m.Load(9, &x.threadSelfSymlink)
-	m.Load(10, &x.cgroupControllers)
+	m.Load(0, &x.tasksInodeRefs)
+	m.Load(1, &x.InodeNotSymlink)
+	m.Load(2, &x.InodeDirectoryNoNewChildren)
+	m.Load(3, &x.InodeAttrs)
+	m.Load(4, &x.OrderedChildren)
+	m.Load(5, &x.AlwaysValid)
+	m.Load(6, &x.locks)
+	m.Load(7, &x.fs)
+	m.Load(8, &x.pidns)
+	m.Load(9, &x.selfSymlink)
+	m.Load(10, &x.threadSelfSymlink)
+	m.Load(11, &x.cgroupControllers)
 }
 
 func (x *statData) StateTypeName() string {
@@ -1095,6 +1199,29 @@ func (x *filesystemsData) StateLoad(m state.Source) {
 	m.Load(0, &x.DynamicBytesFile)
 }
 
+func (x *tasksInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.tasksInodeRefs"
+}
+
+func (x *tasksInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *tasksInodeRefs) beforeSave() {}
+
+func (x *tasksInodeRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *tasksInodeRefs) afterLoad() {}
+
+func (x *tasksInodeRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *mmapMinAddrData) StateTypeName() string {
 	return "pkg/sentry/fsimpl/proc.mmapMinAddrData"
 }
@@ -1229,8 +1356,11 @@ func (x *tcpMemData) StateLoad(m state.Source) {
 }
 
 func init() {
+	state.Register((*fdDirInodeRefs)(nil))
+	state.Register((*fdInfoDirInodeRefs)(nil))
 	state.Register((*FilesystemType)(nil))
 	state.Register((*subtasksInode)(nil))
+	state.Register((*subtasksInodeRefs)(nil))
 	state.Register((*taskInode)(nil))
 	state.Register((*fdDirInode)(nil))
 	state.Register((*fdSymlink)(nil))
@@ -1251,6 +1381,7 @@ func init() {
 	state.Register((*exeSymlink)(nil))
 	state.Register((*mountInfoData)(nil))
 	state.Register((*mountsData)(nil))
+	state.Register((*taskInodeRefs)(nil))
 	state.Register((*ifinet6)(nil))
 	state.Register((*netDevData)(nil))
 	state.Register((*netUnixData)(nil))
@@ -1267,6 +1398,7 @@ func init() {
 	state.Register((*uptimeData)(nil))
 	state.Register((*versionData)(nil))
 	state.Register((*filesystemsData)(nil))
+	state.Register((*tasksInodeRefs)(nil))
 	state.Register((*mmapMinAddrData)(nil))
 	state.Register((*hostnameData)(nil))
 	state.Register((*tcpSackData)(nil))
diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go
index f25747da3..01c0efb3a 100644
--- a/pkg/sentry/fsimpl/proc/subtasks.go
+++ b/pkg/sentry/fsimpl/proc/subtasks.go
@@ -31,6 +31,7 @@ import (
 //
 // +stateify savable
 type subtasksInode struct {
+	subtasksInodeRefs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeAttrs
@@ -57,6 +58,7 @@ func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace,
 	// Note: credentials are overridden by taskOwnedInode.
 	subInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
 	subInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
+	subInode.EnableLeakCheck()
 
 	inode := &taskOwnedInode{Inode: subInode, owner: task}
 	dentry := &kernfs.Dentry{}
@@ -182,3 +184,8 @@ func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs
 func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error {
 	return syserror.EPERM
 }
+
+// DecRef implements kernfs.Inode.
+func (i *subtasksInode) DecRef(context.Context) {
+	i.subtasksInodeRefs.DecRef(i.Destroy)
+}
diff --git a/pkg/sentry/fsimpl/proc/subtasks_inode_refs.go b/pkg/sentry/fsimpl/proc/subtasks_inode_refs.go
new file mode 100644
index 000000000..c6d9b3522
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/subtasks_inode_refs.go
@@ -0,0 +1,118 @@
+package proc
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var subtasksInodeownerType *subtasksInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type subtasksInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *subtasksInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, subtasksInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *subtasksInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*subtasksInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *subtasksInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *subtasksInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, subtasksInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *subtasksInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *subtasksInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, subtasksInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go
index 109b31b4c..66b557abd 100644
--- a/pkg/sentry/fsimpl/proc/task.go
+++ b/pkg/sentry/fsimpl/proc/task.go
@@ -32,6 +32,7 @@ import (
 //
 // +stateify savable
 type taskInode struct {
+	taskInodeRefs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeNoDynamicLookup
@@ -84,6 +85,7 @@ func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace
 	taskInode := &taskInode{task: task}
 	// Note: credentials are overridden by taskOwnedInode.
 	taskInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
+	taskInode.EnableLeakCheck()
 
 	inode := &taskOwnedInode{Inode: taskInode, owner: task}
 	dentry := &kernfs.Dentry{}
@@ -119,6 +121,11 @@ func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, v
 	return syserror.EPERM
 }
 
+// DecRef implements kernfs.Inode.
+func (i *taskInode) DecRef(context.Context) {
+	i.taskInodeRefs.DecRef(i.Destroy)
+}
+
 // taskOwnedInode implements kernfs.Inode and overrides inode owner with task
 // effective user and group.
 type taskOwnedInode struct {
@@ -147,6 +154,7 @@ func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.
 	dir.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, perm, kernfs.GenericDirectoryFDOptions{
 		SeekEnd: kernfs.SeekEndZero,
 	})
+	dir.EnableLeakCheck()
 
 	inode := &taskOwnedInode{Inode: dir, owner: task}
 	d := &kernfs.Dentry{}
diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go
index e8fcb9aa1..0527b2de8 100644
--- a/pkg/sentry/fsimpl/proc/task_fds.go
+++ b/pkg/sentry/fsimpl/proc/task_fds.go
@@ -22,7 +22,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -101,6 +100,7 @@ func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, off
 //
 // +stateify savable
 type fdDirInode struct {
+	fdDirInodeRefs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeAttrs
@@ -120,6 +120,7 @@ func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry {
 		},
 	}
 	inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
+	inode.EnableLeakCheck()
 
 	dentry := &kernfs.Dentry{}
 	dentry.Init(inode)
@@ -175,6 +176,11 @@ func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentia
 	return err
 }
 
+// DecRef implements kernfs.Inode.
+func (i *fdDirInode) DecRef(context.Context) {
+	i.fdDirInodeRefs.DecRef(i.Destroy)
+}
+
 // fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file.
 //
 // +stateify savable
@@ -227,6 +233,7 @@ func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDen
 //
 // +stateify savable
 type fdInfoDirInode struct {
+	fdInfoDirInodeRefs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeAttrs
@@ -245,6 +252,7 @@ func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) *kernfs.Dentry {
 		},
 	}
 	inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
+	inode.EnableLeakCheck()
 
 	dentry := &kernfs.Dentry{}
 	dentry.Init(inode)
@@ -282,12 +290,16 @@ func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *
 	return fd.VFSFileDescription(), nil
 }
 
+// DecRef implements kernfs.Inode.
+func (i *fdInfoDirInode) DecRef(context.Context) {
+	i.fdInfoDirInodeRefs.DecRef(i.Destroy)
+}
+
 // fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd].
 //
 // +stateify savable
 type fdInfoData struct {
 	kernfs.DynamicBytesFile
-	refs.AtomicRefCount
 
 	task *kernel.Task
 	fd   int32
diff --git a/pkg/sentry/fsimpl/proc/task_inode_refs.go b/pkg/sentry/fsimpl/proc/task_inode_refs.go
new file mode 100644
index 000000000..714488450
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/task_inode_refs.go
@@ -0,0 +1,118 @@
+package proc
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var taskInodeownerType *taskInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type taskInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *taskInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, taskInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *taskInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*taskInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *taskInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *taskInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, taskInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *taskInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *taskInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, taskInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/proc/task_net.go b/pkg/sentry/fsimpl/proc/task_net.go
index a4c884bf9..4e69782c7 100644
--- a/pkg/sentry/fsimpl/proc/task_net.go
+++ b/pkg/sentry/fsimpl/proc/task_net.go
@@ -262,7 +262,7 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 		// For now, we always redact this pointer.
 		fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %8d",
 			(*unix.SocketOperations)(nil), // Num, pointer to kernel socket struct.
-			s.Refs()-1,                    // RefCount, don't count our own ref.
+			s.ReadRefs()-1,                // RefCount, don't count our own ref.
 			0,                             // Protocol, always 0 for UDS.
 			sockFlags,                     // Flags.
 			sops.Endpoint().Type(),        // Type.
@@ -430,7 +430,7 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel,
 
 		// Field: refcount. Don't count the ref we obtain while deferencing
 		// the weakref to this socket.
-		fmt.Fprintf(buf, "%d ", s.Refs()-1)
+		fmt.Fprintf(buf, "%d ", s.ReadRefs()-1)
 
 		// Field: Socket struct address. Redacted due to the same reason as
 		// the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData.
@@ -589,7 +589,7 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error {
 
 		// Field: ref; reference count on the socket inode. Don't count the ref
 		// we obtain while deferencing the weakref to this socket.
-		fmt.Fprintf(buf, "%d ", s.Refs()-1)
+		fmt.Fprintf(buf, "%d ", s.ReadRefs()-1)
 
 		// Field: Socket struct address. Redacted due to the same reason as
 		// the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData.
diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go
index 1391992b7..863c4467e 100644
--- a/pkg/sentry/fsimpl/proc/tasks.go
+++ b/pkg/sentry/fsimpl/proc/tasks.go
@@ -37,6 +37,7 @@ const (
 //
 // +stateify savable
 type tasksInode struct {
+	tasksInodeRefs
 	kernfs.InodeNotSymlink
 	kernfs.InodeDirectoryNoNewChildren
 	kernfs.InodeAttrs
@@ -84,6 +85,7 @@ func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace
 		cgroupControllers: cgroupControllers,
 	}
 	inode.InodeAttrs.Init(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555)
+	inode.EnableLeakCheck()
 
 	dentry := &kernfs.Dentry{}
 	dentry.Init(inode)
@@ -226,6 +228,11 @@ func (i *tasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.St
 	return stat, nil
 }
 
+// DecRef implements kernfs.Inode.
+func (i *tasksInode) DecRef(context.Context) {
+	i.tasksInodeRefs.DecRef(i.Destroy)
+}
+
 // staticFileSetStat implements a special static file that allows inode
 // attributes to be set. This is to support /proc files that are readonly, but
 // allow attributes to be set.
diff --git a/pkg/sentry/fsimpl/proc/tasks_inode_refs.go b/pkg/sentry/fsimpl/proc/tasks_inode_refs.go
new file mode 100644
index 000000000..22d9cc488
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/tasks_inode_refs.go
@@ -0,0 +1,118 @@
+package proc
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var tasksInodeownerType *tasksInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type tasksInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *tasksInodeRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, tasksInodeownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *tasksInodeRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*tasksInodeRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *tasksInodeRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *tasksInodeRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, tasksInodeownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *tasksInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *tasksInodeRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, tasksInodeownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/sys/dir_refs.go b/pkg/sentry/fsimpl/sys/dir_refs.go
new file mode 100644
index 000000000..89609b198
--- /dev/null
+++ b/pkg/sentry/fsimpl/sys/dir_refs.go
@@ -0,0 +1,118 @@
+package sys
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var dirownerType *dir
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type dirRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *dirRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, dirownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *dirRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*dirRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *dirRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *dirRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, dirownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *dirRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *dirRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, dirownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go
index 1f042d9f7..ea30a4ec2 100644
--- a/pkg/sentry/fsimpl/sys/sys.go
+++ b/pkg/sentry/fsimpl/sys/sys.go
@@ -118,6 +118,7 @@ func (fs *filesystem) Release(ctx context.Context) {
 
 // dir implements kernfs.Inode.
 type dir struct {
+	dirRefs
 	kernfs.InodeAttrs
 	kernfs.InodeNoDynamicLookup
 	kernfs.InodeNotSymlink
@@ -133,6 +134,7 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte
 	d := &dir{}
 	d.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755)
 	d.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
+	d.EnableLeakCheck()
 	d.dentry.Init(d)
 
 	d.IncLinks(d.OrderedChildren.Populate(&d.dentry, contents))
@@ -140,7 +142,7 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte
 	return &d.dentry
 }
 
-// SetStat implements Inode.SetStat not allowing inode attributes to be changed.
+// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed.
 func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error {
 	return syserror.EPERM
 }
@@ -156,6 +158,11 @@ func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry,
 	return fd.VFSFileDescription(), nil
 }
 
+// DecRef implements kernfs.Inode.DecRef.
+func (d *dir) DecRef(context.Context) {
+	d.dirRefs.DecRef(d.Destroy)
+}
+
 // cpuFile implements kernfs.Inode.
 type cpuFile struct {
 	kernfs.DynamicBytesFile
diff --git a/pkg/sentry/fsimpl/sys/sys_state_autogen.go b/pkg/sentry/fsimpl/sys/sys_state_autogen.go
index 8866ddff9..347a46318 100644
--- a/pkg/sentry/fsimpl/sys/sys_state_autogen.go
+++ b/pkg/sentry/fsimpl/sys/sys_state_autogen.go
@@ -1,3 +1,34 @@
 // automatically generated by stateify.
 
 package sys
+
+import (
+	"gvisor.dev/gvisor/pkg/state"
+)
+
+func (x *dirRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/sys.dirRefs"
+}
+
+func (x *dirRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *dirRefs) beforeSave() {}
+
+func (x *dirRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *dirRefs) afterLoad() {}
+
+func (x *dirRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
+func init() {
+	state.Register((*dirRefs)(nil))
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/inode_refs.go b/pkg/sentry/fsimpl/tmpfs/inode_refs.go
index 8b7ff185f..dbf0b2766 100644
--- a/pkg/sentry/fsimpl/tmpfs/inode_refs.go
+++ b/pkg/sentry/fsimpl/tmpfs/inode_refs.go
@@ -1,6 +1,7 @@
 package tmpfs
 
 import (
+	"fmt"
 	"runtime"
 	"sync/atomic"
 
@@ -18,6 +19,11 @@ var inodeownerType *inode
 // Note that the number of references is actually refCount + 1 so that a default
 // zero-value Refs object contains one reference.
 //
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
 // +stateify savable
 type inodeRefs struct {
 	// refCount is composed of two fields:
@@ -62,7 +68,7 @@ func (r *inodeRefs) ReadRefs() int64 {
 //go:nosplit
 func (r *inodeRefs) IncRef() {
 	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
-		panic("Incrementing non-positive ref count")
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, inodeownerType))
 	}
 }
 
@@ -101,7 +107,7 @@ func (r *inodeRefs) TryIncRef() bool {
 func (r *inodeRefs) DecRef(destroy func()) {
 	switch v := atomic.AddInt64(&r.refCount, -1); {
 	case v < -1:
-		panic("Decrementing non-positive ref count")
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, inodeownerType))
 
 	case v == -1:
 
diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go
index ce53af69b..5773244ac 100644
--- a/pkg/sentry/kernel/fd_table.go
+++ b/pkg/sentry/kernel/fd_table.go
@@ -23,7 +23,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/fs/lock"
 	"gvisor.dev/gvisor/pkg/sentry/limits"
@@ -78,7 +77,8 @@ type descriptor struct {
 //
 // +stateify savable
 type FDTable struct {
-	refs.AtomicRefCount
+	FDTableRefs
+
 	k *Kernel
 
 	// mu protects below.
@@ -176,16 +176,15 @@ func (k *Kernel) NewFDTable() *FDTable {
 	return f
 }
 
-// destroy removes all of the file descriptors from the map.
-func (f *FDTable) destroy(ctx context.Context) {
-	f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool {
-		return true
-	})
-}
-
-// DecRef implements RefCounter.DecRef with destructor f.destroy.
+// DecRef implements RefCounter.DecRef.
+//
+// If f reaches zero references, all of its file descriptors are removed.
 func (f *FDTable) DecRef(ctx context.Context) {
-	f.DecRefWithDestructor(ctx, f.destroy)
+	f.FDTableRefs.DecRef(func() {
+		f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool {
+			return true
+		})
+	})
 }
 
 // Size returns the number of file descriptor slots currently allocated.
diff --git a/pkg/sentry/kernel/fd_table_refs.go b/pkg/sentry/kernel/fd_table_refs.go
new file mode 100644
index 000000000..ecba138ac
--- /dev/null
+++ b/pkg/sentry/kernel/fd_table_refs.go
@@ -0,0 +1,118 @@
+package kernel
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var FDTableownerType *FDTable
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type FDTableRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *FDTableRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, FDTableownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *FDTableRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*FDTableRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *FDTableRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *FDTableRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, FDTableownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *FDTableRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *FDTableRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, FDTableownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/fd_table_unsafe.go b/pkg/sentry/kernel/fd_table_unsafe.go
index 7fd97dc53..6b8feb107 100644
--- a/pkg/sentry/kernel/fd_table_unsafe.go
+++ b/pkg/sentry/kernel/fd_table_unsafe.go
@@ -31,6 +31,8 @@ type descriptorTable struct {
 }
 
 // init initializes the table.
+//
+// TODO(gvisor.dev/1486): Enable leak check for FDTable.
 func (f *FDTable) init() {
 	var slice []unsafe.Pointer // Empty slice.
 	atomic.StorePointer(&f.slice, unsafe.Pointer(&slice))
diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go
index 8f2d36d5a..d46d1e1c1 100644
--- a/pkg/sentry/kernel/fs_context.go
+++ b/pkg/sentry/kernel/fs_context.go
@@ -18,7 +18,6 @@ import (
 	"fmt"
 
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
@@ -30,7 +29,7 @@ import (
 //
 // +stateify savable
 type FSContext struct {
-	refs.AtomicRefCount
+	FSContextRefs
 
 	// mu protects below.
 	mu sync.Mutex `state:"nosave"`
@@ -64,7 +63,7 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext {
 		cwd:   cwd,
 		umask: umask,
 	}
-	f.EnableLeakCheck("kernel.FSContext")
+	f.EnableLeakCheck()
 	return &f
 }
 
@@ -77,54 +76,56 @@ func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext {
 		cwdVFS2:  cwd,
 		umask:    umask,
 	}
-	f.EnableLeakCheck("kernel.FSContext")
+	f.EnableLeakCheck()
 	return &f
 }
 
-// destroy is the destructor for an FSContext.
+// DecRef implements RefCounter.DecRef.
 //
-// This will call DecRef on both root and cwd Dirents.  If either call to
-// DecRef returns an error, then it will be propagated.  If both calls to
-// DecRef return an error, then the one from root.DecRef will be propagated.
+// When f reaches zero references, DecRef will be called on both root and cwd
+// Dirents.
 //
 // Note that there may still be calls to WorkingDirectory() or RootDirectory()
 // (that return nil).  This is because valid references may still be held via
 // proc files or other mechanisms.
-func (f *FSContext) destroy(ctx context.Context) {
-	// Hold f.mu so that we don't race with RootDirectory() and
-	// WorkingDirectory().
-	f.mu.Lock()
-	defer f.mu.Unlock()
-
-	if VFS2Enabled {
-		f.rootVFS2.DecRef(ctx)
-		f.rootVFS2 = vfs.VirtualDentry{}
-		f.cwdVFS2.DecRef(ctx)
-		f.cwdVFS2 = vfs.VirtualDentry{}
-	} else {
-		f.root.DecRef(ctx)
-		f.root = nil
-		f.cwd.DecRef(ctx)
-		f.cwd = nil
-	}
-}
-
-// DecRef implements RefCounter.DecRef with destructor f.destroy.
 func (f *FSContext) DecRef(ctx context.Context) {
-	f.DecRefWithDestructor(ctx, f.destroy)
+	f.FSContextRefs.DecRef(func() {
+		// Hold f.mu so that we don't race with RootDirectory() and
+		// WorkingDirectory().
+		f.mu.Lock()
+		defer f.mu.Unlock()
+
+		if VFS2Enabled {
+			f.rootVFS2.DecRef(ctx)
+			f.rootVFS2 = vfs.VirtualDentry{}
+			f.cwdVFS2.DecRef(ctx)
+			f.cwdVFS2 = vfs.VirtualDentry{}
+		} else {
+			f.root.DecRef(ctx)
+			f.root = nil
+			f.cwd.DecRef(ctx)
+			f.cwd = nil
+		}
+	})
 }
 
 // Fork forks this FSContext.
 //
-// This is not a valid call after destroy.
+// This is not a valid call after f is destroyed.
 func (f *FSContext) Fork() *FSContext {
 	f.mu.Lock()
 	defer f.mu.Unlock()
 
 	if VFS2Enabled {
+		if !f.cwdVFS2.Ok() {
+			panic("FSContext.Fork() called after destroy")
+		}
 		f.cwdVFS2.IncRef()
 		f.rootVFS2.IncRef()
 	} else {
+		if f.cwd == nil {
+			panic("FSContext.Fork() called after destroy")
+		}
 		f.cwd.IncRef()
 		f.root.IncRef()
 	}
@@ -140,8 +141,8 @@ func (f *FSContext) Fork() *FSContext {
 
 // WorkingDirectory returns the current working directory.
 //
-// This will return nil if called after destroy(), otherwise it will return a
-// Dirent with a reference taken.
+// This will return nil if called after f is destroyed, otherwise it will return
+// a Dirent with a reference taken.
 func (f *FSContext) WorkingDirectory() *fs.Dirent {
 	f.mu.Lock()
 	defer f.mu.Unlock()
@@ -152,8 +153,8 @@ func (f *FSContext) WorkingDirectory() *fs.Dirent {
 
 // WorkingDirectoryVFS2 returns the current working directory.
 //
-// This will return nil if called after destroy(), otherwise it will return a
-// Dirent with a reference taken.
+// This will return nil if called after f is destroyed, otherwise it will return
+// a Dirent with a reference taken.
 func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry {
 	f.mu.Lock()
 	defer f.mu.Unlock()
@@ -165,7 +166,7 @@ func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry {
 // SetWorkingDirectory sets the current working directory.
 // This will take an extra reference on the Dirent.
 //
-// This is not a valid call after destroy.
+// This is not a valid call after f is destroyed.
 func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) {
 	if d == nil {
 		panic("FSContext.SetWorkingDirectory called with nil dirent")
@@ -187,11 +188,15 @@ func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) {
 // SetWorkingDirectoryVFS2 sets the current working directory.
 // This will take an extra reference on the VirtualDentry.
 //
-// This is not a valid call after destroy.
+// This is not a valid call after f is destroyed.
 func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDentry) {
 	f.mu.Lock()
 	defer f.mu.Unlock()
 
+	if !f.cwdVFS2.Ok() {
+		panic(fmt.Sprintf("FSContext.SetWorkingDirectoryVFS2(%v)) called after destroy", d))
+	}
+
 	old := f.cwdVFS2
 	f.cwdVFS2 = d
 	d.IncRef()
@@ -200,8 +205,8 @@ func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDe
 
 // RootDirectory returns the current filesystem root.
 //
-// This will return nil if called after destroy(), otherwise it will return a
-// Dirent with a reference taken.
+// This will return nil if called after f is destroyed, otherwise it will return
+// a Dirent with a reference taken.
 func (f *FSContext) RootDirectory() *fs.Dirent {
 	f.mu.Lock()
 	defer f.mu.Unlock()
@@ -213,8 +218,8 @@ func (f *FSContext) RootDirectory() *fs.Dirent {
 
 // RootDirectoryVFS2 returns the current filesystem root.
 //
-// This will return nil if called after destroy(), otherwise it will return a
-// Dirent with a reference taken.
+// This will return nil if called after f is destroyed, otherwise it will return
+// a Dirent with a reference taken.
 func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry {
 	f.mu.Lock()
 	defer f.mu.Unlock()
@@ -226,7 +231,7 @@ func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry {
 // SetRootDirectory sets the root directory.
 // This will take an extra reference on the Dirent.
 //
-// This is not a valid call after free.
+// This is not a valid call after f is destroyed.
 func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) {
 	if d == nil {
 		panic("FSContext.SetRootDirectory called with nil dirent")
@@ -247,7 +252,7 @@ func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) {
 
 // SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd.
 //
-// This is not a valid call after free.
+// This is not a valid call after f is destroyed.
 func (f *FSContext) SetRootDirectoryVFS2(ctx context.Context, vd vfs.VirtualDentry) {
 	if !vd.Ok() {
 		panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry")
diff --git a/pkg/sentry/kernel/fs_context_refs.go b/pkg/sentry/kernel/fs_context_refs.go
new file mode 100644
index 000000000..fb2fde971
--- /dev/null
+++ b/pkg/sentry/kernel/fs_context_refs.go
@@ -0,0 +1,118 @@
+package kernel
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var FSContextownerType *FSContext
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type FSContextRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *FSContextRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, FSContextownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *FSContextRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*FSContextRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *FSContextRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *FSContextRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, FSContextownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *FSContextRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *FSContextRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, FSContextownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/kernel_state_autogen.go b/pkg/sentry/kernel/kernel_state_autogen.go
index e514b07d6..d0ff135d7 100644
--- a/pkg/sentry/kernel/kernel_state_autogen.go
+++ b/pkg/sentry/kernel/kernel_state_autogen.go
@@ -122,7 +122,7 @@ func (x *FDTable) StateTypeName() string {
 
 func (x *FDTable) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"FDTableRefs",
 		"k",
 		"next",
 		"used",
@@ -136,7 +136,7 @@ func (x *FDTable) StateSave(m state.Sink) {
 	x.beforeSave()
 	var descriptorTable map[int32]descriptor = x.saveDescriptorTable()
 	m.SaveValue(4, descriptorTable)
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.FDTableRefs)
 	m.Save(1, &x.k)
 	m.Save(2, &x.next)
 	m.Save(3, &x.used)
@@ -145,20 +145,43 @@ func (x *FDTable) StateSave(m state.Sink) {
 func (x *FDTable) afterLoad() {}
 
 func (x *FDTable) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.FDTableRefs)
 	m.Load(1, &x.k)
 	m.Load(2, &x.next)
 	m.Load(3, &x.used)
 	m.LoadValue(4, new(map[int32]descriptor), func(y interface{}) { x.loadDescriptorTable(y.(map[int32]descriptor)) })
 }
 
+func (x *FDTableRefs) StateTypeName() string {
+	return "pkg/sentry/kernel.FDTableRefs"
+}
+
+func (x *FDTableRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *FDTableRefs) beforeSave() {}
+
+func (x *FDTableRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *FDTableRefs) afterLoad() {}
+
+func (x *FDTableRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *FSContext) StateTypeName() string {
 	return "pkg/sentry/kernel.FSContext"
 }
 
 func (x *FSContext) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"FSContextRefs",
 		"root",
 		"rootVFS2",
 		"cwd",
@@ -171,7 +194,7 @@ func (x *FSContext) beforeSave() {}
 
 func (x *FSContext) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.FSContextRefs)
 	m.Save(1, &x.root)
 	m.Save(2, &x.rootVFS2)
 	m.Save(3, &x.cwd)
@@ -182,7 +205,7 @@ func (x *FSContext) StateSave(m state.Sink) {
 func (x *FSContext) afterLoad() {}
 
 func (x *FSContext) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.FSContextRefs)
 	m.Load(1, &x.root)
 	m.Load(2, &x.rootVFS2)
 	m.Load(3, &x.cwd)
@@ -190,6 +213,29 @@ func (x *FSContext) StateLoad(m state.Source) {
 	m.Load(5, &x.umask)
 }
 
+func (x *FSContextRefs) StateTypeName() string {
+	return "pkg/sentry/kernel.FSContextRefs"
+}
+
+func (x *FSContextRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *FSContextRefs) beforeSave() {}
+
+func (x *FSContextRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *FSContextRefs) afterLoad() {}
+
+func (x *FSContextRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *IPCNamespace) StateTypeName() string {
 	return "pkg/sentry/kernel.IPCNamespace"
 }
@@ -643,6 +689,29 @@ func (x *processGroupEntry) StateLoad(m state.Source) {
 	m.Load(1, &x.prev)
 }
 
+func (x *ProcessGroupRefs) StateTypeName() string {
+	return "pkg/sentry/kernel.ProcessGroupRefs"
+}
+
+func (x *ProcessGroupRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *ProcessGroupRefs) beforeSave() {}
+
+func (x *ProcessGroupRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *ProcessGroupRefs) afterLoad() {}
+
+func (x *ProcessGroupRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *ptraceOptions) StateTypeName() string {
 	return "pkg/sentry/kernel.ptraceOptions"
 }
@@ -794,13 +863,36 @@ func (x *sessionEntry) StateLoad(m state.Source) {
 	m.Load(1, &x.prev)
 }
 
+func (x *SessionRefs) StateTypeName() string {
+	return "pkg/sentry/kernel.SessionRefs"
+}
+
+func (x *SessionRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *SessionRefs) beforeSave() {}
+
+func (x *SessionRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *SessionRefs) afterLoad() {}
+
+func (x *SessionRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *Session) StateTypeName() string {
 	return "pkg/sentry/kernel.Session"
 }
 
 func (x *Session) StateFields() []string {
 	return []string{
-		"refs",
+		"SessionRefs",
 		"leader",
 		"id",
 		"foreground",
@@ -813,7 +905,7 @@ func (x *Session) beforeSave() {}
 
 func (x *Session) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.refs)
+	m.Save(0, &x.SessionRefs)
 	m.Save(1, &x.leader)
 	m.Save(2, &x.id)
 	m.Save(3, &x.foreground)
@@ -824,7 +916,7 @@ func (x *Session) StateSave(m state.Sink) {
 func (x *Session) afterLoad() {}
 
 func (x *Session) StateLoad(m state.Source) {
-	m.Load(0, &x.refs)
+	m.Load(0, &x.SessionRefs)
 	m.Load(1, &x.leader)
 	m.Load(2, &x.id)
 	m.Load(3, &x.foreground)
@@ -2167,7 +2259,9 @@ func init() {
 	state.Register((*FDFlags)(nil))
 	state.Register((*descriptor)(nil))
 	state.Register((*FDTable)(nil))
+	state.Register((*FDTableRefs)(nil))
 	state.Register((*FSContext)(nil))
+	state.Register((*FSContextRefs)(nil))
 	state.Register((*IPCNamespace)(nil))
 	state.Register((*Kernel)(nil))
 	state.Register((*SocketEntry)(nil))
@@ -2180,11 +2274,13 @@ func init() {
 	state.Register((*IntervalTimer)(nil))
 	state.Register((*processGroupList)(nil))
 	state.Register((*processGroupEntry)(nil))
+	state.Register((*ProcessGroupRefs)(nil))
 	state.Register((*ptraceOptions)(nil))
 	state.Register((*ptraceStop)(nil))
 	state.Register((*OldRSeqCriticalRegion)(nil))
 	state.Register((*sessionList)(nil))
 	state.Register((*sessionEntry)(nil))
+	state.Register((*SessionRefs)(nil))
 	state.Register((*Session)(nil))
 	state.Register((*ProcessGroup)(nil))
 	state.Register((*SignalHandlers)(nil))
diff --git a/pkg/sentry/kernel/process_group_refs.go b/pkg/sentry/kernel/process_group_refs.go
new file mode 100644
index 000000000..4ed6e6458
--- /dev/null
+++ b/pkg/sentry/kernel/process_group_refs.go
@@ -0,0 +1,118 @@
+package kernel
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var ProcessGroupownerType *ProcessGroup
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type ProcessGroupRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *ProcessGroupRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, ProcessGroupownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *ProcessGroupRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*ProcessGroupRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *ProcessGroupRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *ProcessGroupRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, ProcessGroupownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *ProcessGroupRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *ProcessGroupRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, ProcessGroupownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/session_refs.go b/pkg/sentry/kernel/session_refs.go
new file mode 100644
index 000000000..f2e1bb797
--- /dev/null
+++ b/pkg/sentry/kernel/session_refs.go
@@ -0,0 +1,118 @@
+package kernel
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var SessionownerType *Session
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type SessionRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *SessionRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, SessionownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *SessionRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*SessionRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *SessionRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *SessionRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, SessionownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *SessionRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *SessionRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, SessionownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go
index 5c4c622c2..df5c8421b 100644
--- a/pkg/sentry/kernel/sessions.go
+++ b/pkg/sentry/kernel/sessions.go
@@ -16,8 +16,6 @@ package kernel
 
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
-	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/syserror"
 )
@@ -32,7 +30,7 @@ type ProcessGroupID ThreadID
 //
 // +stateify savable
 type Session struct {
-	refs refs.AtomicRefCount
+	SessionRefs
 
 	// leader is the originator of the Session.
 	//
@@ -62,16 +60,11 @@ type Session struct {
 	sessionEntry
 }
 
-// incRef grabs a reference.
-func (s *Session) incRef() {
-	s.refs.IncRef()
-}
-
-// decRef drops a reference.
+// DecRef drops a reference.
 //
 // Precondition: callers must hold TaskSet.mu for writing.
-func (s *Session) decRef() {
-	s.refs.DecRefWithDestructor(nil, func(context.Context) {
+func (s *Session) DecRef() {
+	s.SessionRefs.DecRef(func() {
 		// Remove translations from the leader.
 		for ns := s.leader.pidns; ns != nil; ns = ns.parent {
 			id := ns.sids[s]
@@ -88,7 +81,7 @@ func (s *Session) decRef() {
 //
 // +stateify savable
 type ProcessGroup struct {
-	refs refs.AtomicRefCount // not exported.
+	refs ProcessGroupRefs
 
 	// originator is the originator of the group.
 	//
@@ -163,7 +156,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) {
 	}
 
 	alive := true
-	pg.refs.DecRefWithDestructor(nil, func(context.Context) {
+	pg.refs.DecRef(func() {
 		alive = false // don't bother with handleOrphan.
 
 		// Remove translations from the originator.
@@ -175,7 +168,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) {
 
 		// Remove the list of process groups.
 		pg.session.processGroups.Remove(pg)
-		pg.session.decRef()
+		pg.session.DecRef()
 	})
 	if alive {
 		pg.handleOrphan()
@@ -302,7 +295,7 @@ func (tg *ThreadGroup) createSession() error {
 		id:     SessionID(id),
 		leader: tg,
 	}
-	s.refs.EnableLeakCheck("kernel.Session")
+	s.EnableLeakCheck()
 
 	// Create a new ProcessGroup, belonging to that Session.
 	// This also has a single reference (assigned below).
@@ -316,7 +309,7 @@ func (tg *ThreadGroup) createSession() error {
 		session:    s,
 		ancestors:  0,
 	}
-	pg.refs.EnableLeakCheck("kernel.ProcessGroup")
+	pg.refs.EnableLeakCheck()
 
 	// Tie them and return the result.
 	s.processGroups.PushBack(pg)
@@ -396,13 +389,13 @@ func (tg *ThreadGroup) CreateProcessGroup() error {
 	//
 	// We manually adjust the ancestors if the parent is in the same
 	// session.
-	tg.processGroup.session.incRef()
+	tg.processGroup.session.IncRef()
 	pg := ProcessGroup{
 		id:         ProcessGroupID(id),
 		originator: tg,
 		session:    tg.processGroup.session,
 	}
-	pg.refs.EnableLeakCheck("kernel.ProcessGroup")
+	pg.refs.EnableLeakCheck()
 
 	if tg.leader.parent != nil && tg.leader.parent.tg.processGroup.session == pg.session {
 		pg.ancestors++
diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go
index 13ec7afe0..00c03585e 100644
--- a/pkg/sentry/kernel/shm/shm.go
+++ b/pkg/sentry/kernel/shm/shm.go
@@ -39,7 +39,6 @@ import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/log"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time"
@@ -252,7 +251,7 @@ func (r *Registry) newShm(ctx context.Context, pid int32, key Key, creator fs.Fi
 		creatorPID:    pid,
 		changeTime:    ktime.NowFromContext(ctx),
 	}
-	shm.EnableLeakCheck("kernel.Shm")
+	shm.EnableLeakCheck()
 
 	// Find the next available ID.
 	for id := r.lastIDUsed + 1; id != r.lastIDUsed; id++ {
@@ -337,14 +336,14 @@ func (r *Registry) remove(s *Shm) {
 //
 // +stateify savable
 type Shm struct {
-	// AtomicRefCount tracks the number of references to this segment.
+	// ShmRefs tracks the number of references to this segment.
 	//
 	// A segment holds a reference to itself until it is marked for
 	// destruction.
 	//
 	// In addition to direct users, the MemoryManager will hold references
 	// via MappingIdentity.
-	refs.AtomicRefCount
+	ShmRefs
 
 	mfp pgalloc.MemoryFileProvider
 
@@ -428,11 +427,14 @@ func (s *Shm) InodeID() uint64 {
 	return uint64(s.ID)
 }
 
-// DecRef overrides refs.RefCount.DecRef with a destructor.
+// DecRef drops a reference on s.
 //
 // Precondition: Caller must not hold s.mu.
 func (s *Shm) DecRef(ctx context.Context) {
-	s.DecRefWithDestructor(ctx, s.destroy)
+	s.ShmRefs.DecRef(func() {
+		s.mfp.MemoryFile().DecRef(s.fr)
+		s.registry.remove(s)
+	})
 }
 
 // Msync implements memmap.MappingIdentity.Msync. Msync is a no-op for shm
@@ -642,11 +644,6 @@ func (s *Shm) Set(ctx context.Context, ds *linux.ShmidDS) error {
 	return nil
 }
 
-func (s *Shm) destroy(context.Context) {
-	s.mfp.MemoryFile().DecRef(s.fr)
-	s.registry.remove(s)
-}
-
 // MarkDestroyed marks a segment for destruction. The segment is actually
 // destroyed once it has no references. MarkDestroyed may be called multiple
 // times, and is safe to call after a segment has already been destroyed. See
diff --git a/pkg/sentry/kernel/shm/shm_refs.go b/pkg/sentry/kernel/shm/shm_refs.go
new file mode 100644
index 000000000..51e07d0b3
--- /dev/null
+++ b/pkg/sentry/kernel/shm/shm_refs.go
@@ -0,0 +1,118 @@
+package shm
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var ShmownerType *Shm
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type ShmRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *ShmRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, ShmownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *ShmRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*ShmRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *ShmRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *ShmRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, ShmownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *ShmRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *ShmRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, ShmownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/shm/shm_state_autogen.go b/pkg/sentry/kernel/shm/shm_state_autogen.go
index c98632087..fa80353b6 100644
--- a/pkg/sentry/kernel/shm/shm_state_autogen.go
+++ b/pkg/sentry/kernel/shm/shm_state_autogen.go
@@ -47,7 +47,7 @@ func (x *Shm) StateTypeName() string {
 
 func (x *Shm) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"ShmRefs",
 		"mfp",
 		"registry",
 		"ID",
@@ -71,7 +71,7 @@ func (x *Shm) beforeSave() {}
 
 func (x *Shm) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.ShmRefs)
 	m.Save(1, &x.mfp)
 	m.Save(2, &x.registry)
 	m.Save(3, &x.ID)
@@ -93,7 +93,7 @@ func (x *Shm) StateSave(m state.Sink) {
 func (x *Shm) afterLoad() {}
 
 func (x *Shm) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.ShmRefs)
 	m.Load(1, &x.mfp)
 	m.Load(2, &x.registry)
 	m.Load(3, &x.ID)
@@ -112,7 +112,31 @@ func (x *Shm) StateLoad(m state.Source) {
 	m.Load(16, &x.pendingDestruction)
 }
 
+func (x *ShmRefs) StateTypeName() string {
+	return "pkg/sentry/kernel/shm.ShmRefs"
+}
+
+func (x *ShmRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *ShmRefs) beforeSave() {}
+
+func (x *ShmRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *ShmRefs) afterLoad() {}
+
+func (x *ShmRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func init() {
 	state.Register((*Registry)(nil))
 	state.Register((*Shm)(nil))
+	state.Register((*ShmRefs)(nil))
 }
diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go
index 16fea53c4..7bf48cb2c 100644
--- a/pkg/sentry/mm/aio_context.go
+++ b/pkg/sentry/mm/aio_context.go
@@ -17,7 +17,6 @@ package mm
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/memmap"
 	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
 	"gvisor.dev/gvisor/pkg/sentry/usage"
@@ -239,7 +238,7 @@ func (ctx *AIOContext) Drain() {
 //
 // +stateify savable
 type aioMappable struct {
-	refs.AtomicRefCount
+	aioMappableRefs
 
 	mfp pgalloc.MemoryFileProvider
 	fr  memmap.FileRange
@@ -253,13 +252,13 @@ func newAIOMappable(mfp pgalloc.MemoryFileProvider) (*aioMappable, error) {
 		return nil, err
 	}
 	m := aioMappable{mfp: mfp, fr: fr}
-	m.EnableLeakCheck("mm.aioMappable")
+	m.EnableLeakCheck()
 	return &m, nil
 }
 
 // DecRef implements refs.RefCounter.DecRef.
 func (m *aioMappable) DecRef(ctx context.Context) {
-	m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) {
+	m.aioMappableRefs.DecRef(func() {
 		m.mfp.MemoryFile().DecRef(m.fr)
 	})
 }
diff --git a/pkg/sentry/mm/aio_mappable_refs.go b/pkg/sentry/mm/aio_mappable_refs.go
new file mode 100644
index 000000000..b99909f07
--- /dev/null
+++ b/pkg/sentry/mm/aio_mappable_refs.go
@@ -0,0 +1,118 @@
+package mm
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var aioMappableownerType *aioMappable
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type aioMappableRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *aioMappableRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, aioMappableownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *aioMappableRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*aioMappableRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *aioMappableRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *aioMappableRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, aioMappableownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *aioMappableRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *aioMappableRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, aioMappableownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/mm/mm_state_autogen.go b/pkg/sentry/mm/mm_state_autogen.go
index 2f81a8240..8ab51450c 100644
--- a/pkg/sentry/mm/mm_state_autogen.go
+++ b/pkg/sentry/mm/mm_state_autogen.go
@@ -92,7 +92,7 @@ func (x *aioMappable) StateTypeName() string {
 
 func (x *aioMappable) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"aioMappableRefs",
 		"mfp",
 		"fr",
 	}
@@ -102,7 +102,7 @@ func (x *aioMappable) beforeSave() {}
 
 func (x *aioMappable) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.aioMappableRefs)
 	m.Save(1, &x.mfp)
 	m.Save(2, &x.fr)
 }
@@ -110,11 +110,34 @@ func (x *aioMappable) StateSave(m state.Sink) {
 func (x *aioMappable) afterLoad() {}
 
 func (x *aioMappable) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.aioMappableRefs)
 	m.Load(1, &x.mfp)
 	m.Load(2, &x.fr)
 }
 
+func (x *aioMappableRefs) StateTypeName() string {
+	return "pkg/sentry/mm.aioMappableRefs"
+}
+
+func (x *aioMappableRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *aioMappableRefs) beforeSave() {}
+
+func (x *aioMappableRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *aioMappableRefs) afterLoad() {}
+
+func (x *aioMappableRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *fileRefcountSet) StateTypeName() string {
 	return "pkg/sentry/mm.fileRefcountSet"
 }
@@ -565,7 +588,7 @@ func (x *SpecialMappable) StateTypeName() string {
 
 func (x *SpecialMappable) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"SpecialMappableRefs",
 		"mfp",
 		"fr",
 		"name",
@@ -576,7 +599,7 @@ func (x *SpecialMappable) beforeSave() {}
 
 func (x *SpecialMappable) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.SpecialMappableRefs)
 	m.Save(1, &x.mfp)
 	m.Save(2, &x.fr)
 	m.Save(3, &x.name)
@@ -585,12 +608,35 @@ func (x *SpecialMappable) StateSave(m state.Sink) {
 func (x *SpecialMappable) afterLoad() {}
 
 func (x *SpecialMappable) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.SpecialMappableRefs)
 	m.Load(1, &x.mfp)
 	m.Load(2, &x.fr)
 	m.Load(3, &x.name)
 }
 
+func (x *SpecialMappableRefs) StateTypeName() string {
+	return "pkg/sentry/mm.SpecialMappableRefs"
+}
+
+func (x *SpecialMappableRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *SpecialMappableRefs) beforeSave() {}
+
+func (x *SpecialMappableRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *SpecialMappableRefs) afterLoad() {}
+
+func (x *SpecialMappableRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *vmaSet) StateTypeName() string {
 	return "pkg/sentry/mm.vmaSet"
 }
@@ -693,6 +739,7 @@ func init() {
 	state.Register((*ioResult)(nil))
 	state.Register((*AIOContext)(nil))
 	state.Register((*aioMappable)(nil))
+	state.Register((*aioMappableRefs)(nil))
 	state.Register((*fileRefcountSet)(nil))
 	state.Register((*fileRefcountnode)(nil))
 	state.Register((*fileRefcountSegmentDataSlices)(nil))
@@ -706,6 +753,7 @@ func init() {
 	state.Register((*pmanode)(nil))
 	state.Register((*pmaSegmentDataSlices)(nil))
 	state.Register((*SpecialMappable)(nil))
+	state.Register((*SpecialMappableRefs)(nil))
 	state.Register((*vmaSet)(nil))
 	state.Register((*vmanode)(nil))
 	state.Register((*vmaSegmentDataSlices)(nil))
diff --git a/pkg/sentry/mm/special_mappable.go b/pkg/sentry/mm/special_mappable.go
index 4cdb52eb6..f4c93baeb 100644
--- a/pkg/sentry/mm/special_mappable.go
+++ b/pkg/sentry/mm/special_mappable.go
@@ -16,7 +16,6 @@ package mm
 
 import (
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/memmap"
 	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
 	"gvisor.dev/gvisor/pkg/sentry/usage"
@@ -31,7 +30,7 @@ import (
 //
 // +stateify savable
 type SpecialMappable struct {
-	refs.AtomicRefCount
+	SpecialMappableRefs
 
 	mfp  pgalloc.MemoryFileProvider
 	fr   memmap.FileRange
@@ -45,13 +44,13 @@ type SpecialMappable struct {
 // Preconditions: fr.Length() != 0.
 func NewSpecialMappable(name string, mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *SpecialMappable {
 	m := SpecialMappable{mfp: mfp, fr: fr, name: name}
-	m.EnableLeakCheck("mm.SpecialMappable")
+	m.EnableLeakCheck()
 	return &m
 }
 
 // DecRef implements refs.RefCounter.DecRef.
 func (m *SpecialMappable) DecRef(ctx context.Context) {
-	m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) {
+	m.SpecialMappableRefs.DecRef(func() {
 		m.mfp.MemoryFile().DecRef(m.fr)
 	})
 }
diff --git a/pkg/sentry/mm/special_mappable_refs.go b/pkg/sentry/mm/special_mappable_refs.go
new file mode 100644
index 000000000..035bbe690
--- /dev/null
+++ b/pkg/sentry/mm/special_mappable_refs.go
@@ -0,0 +1,118 @@
+package mm
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var SpecialMappableownerType *SpecialMappable
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type SpecialMappableRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *SpecialMappableRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, SpecialMappableownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *SpecialMappableRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*SpecialMappableRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *SpecialMappableRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *SpecialMappableRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, SpecialMappableownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *SpecialMappableRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *SpecialMappableRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, SpecialMappableownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/defs_impl_arm64.go b/pkg/sentry/platform/ring0/defs_impl_arm64.go
index 2dac9ad14..424b66f76 100644
--- a/pkg/sentry/platform/ring0/defs_impl_arm64.go
+++ b/pkg/sentry/platform/ring0/defs_impl_arm64.go
@@ -3,11 +3,11 @@ package ring0
 import (
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables"
-	"reflect"
 
 	"fmt"
 	"gvisor.dev/gvisor/pkg/usermem"
 	"io"
+	"reflect"
 )
 
 // Useful bits.
diff --git a/pkg/sentry/socket/unix/socket_refs.go b/pkg/sentry/socket/unix/socket_refs.go
index 4c6ec186b..dababb85f 100644
--- a/pkg/sentry/socket/unix/socket_refs.go
+++ b/pkg/sentry/socket/unix/socket_refs.go
@@ -1,6 +1,7 @@
 package unix
 
 import (
+	"fmt"
 	"runtime"
 	"sync/atomic"
 
@@ -18,6 +19,11 @@ var socketOpsCommonownerType *socketOpsCommon
 // Note that the number of references is actually refCount + 1 so that a default
 // zero-value Refs object contains one reference.
 //
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
 // +stateify savable
 type socketOpsCommonRefs struct {
 	// refCount is composed of two fields:
@@ -62,7 +68,7 @@ func (r *socketOpsCommonRefs) ReadRefs() int64 {
 //go:nosplit
 func (r *socketOpsCommonRefs) IncRef() {
 	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
-		panic("Incrementing non-positive ref count")
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, socketOpsCommonownerType))
 	}
 }
 
@@ -101,7 +107,7 @@ func (r *socketOpsCommonRefs) TryIncRef() bool {
 func (r *socketOpsCommonRefs) DecRef(destroy func()) {
 	switch v := atomic.AddInt64(&r.refCount, -1); {
 	case v < -1:
-		panic("Decrementing non-positive ref count")
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, socketOpsCommonownerType))
 
 	case v == -1:
 
diff --git a/pkg/sentry/socket/unix/transport/connectioned.go b/pkg/sentry/socket/unix/transport/connectioned.go
index c67b602f0..e3a75b519 100644
--- a/pkg/sentry/socket/unix/transport/connectioned.go
+++ b/pkg/sentry/socket/unix/transport/connectioned.go
@@ -142,9 +142,9 @@ func NewPair(ctx context.Context, stype linux.SockType, uid UniqueIDProvider) (E
 	}
 
 	q1 := &queue{ReaderQueue: a.Queue, WriterQueue: b.Queue, limit: initialLimit}
-	q1.EnableLeakCheck("transport.queue")
+	q1.EnableLeakCheck()
 	q2 := &queue{ReaderQueue: b.Queue, WriterQueue: a.Queue, limit: initialLimit}
-	q2.EnableLeakCheck("transport.queue")
+	q2.EnableLeakCheck()
 
 	if stype == linux.SOCK_STREAM {
 		a.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{q1}}
@@ -300,14 +300,14 @@ func (e *connectionedEndpoint) BidirectionalConnect(ctx context.Context, ce Conn
 	}
 
 	readQueue := &queue{ReaderQueue: ce.WaiterQueue(), WriterQueue: ne.Queue, limit: initialLimit}
-	readQueue.EnableLeakCheck("transport.queue")
+	readQueue.EnableLeakCheck()
 	ne.connected = &connectedEndpoint{
 		endpoint:   ce,
 		writeQueue: readQueue,
 	}
 
 	writeQueue := &queue{ReaderQueue: ne.Queue, WriterQueue: ce.WaiterQueue(), limit: initialLimit}
-	writeQueue.EnableLeakCheck("transport.queue")
+	writeQueue.EnableLeakCheck()
 	if e.stype == linux.SOCK_STREAM {
 		ne.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{readQueue: writeQueue}}
 	} else {
diff --git a/pkg/sentry/socket/unix/transport/connectionless.go b/pkg/sentry/socket/unix/transport/connectionless.go
index 70ee8f9b8..4751b2fd8 100644
--- a/pkg/sentry/socket/unix/transport/connectionless.go
+++ b/pkg/sentry/socket/unix/transport/connectionless.go
@@ -42,7 +42,7 @@ var (
 func NewConnectionless(ctx context.Context) Endpoint {
 	ep := &connectionlessEndpoint{baseEndpoint{Queue: &waiter.Queue{}}}
 	q := queue{ReaderQueue: ep.Queue, WriterQueue: &waiter.Queue{}, limit: initialLimit}
-	q.EnableLeakCheck("transport.queue")
+	q.EnableLeakCheck()
 	ep.receiver = &queueReceiver{readQueue: &q}
 	return ep
 }
diff --git a/pkg/sentry/socket/unix/transport/queue.go b/pkg/sentry/socket/unix/transport/queue.go
index ef6043e19..342def28f 100644
--- a/pkg/sentry/socket/unix/transport/queue.go
+++ b/pkg/sentry/socket/unix/transport/queue.go
@@ -16,7 +16,6 @@ package transport
 
 import (
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserr"
 	"gvisor.dev/gvisor/pkg/tcpip"
@@ -28,7 +27,7 @@ import (
 //
 // +stateify savable
 type queue struct {
-	refs.AtomicRefCount
+	queueRefs
 
 	ReaderQueue *waiter.Queue
 	WriterQueue *waiter.Queue
@@ -68,11 +67,13 @@ func (q *queue) Reset(ctx context.Context) {
 	q.mu.Unlock()
 }
 
-// DecRef implements RefCounter.DecRef with destructor q.Reset.
+// DecRef implements RefCounter.DecRef.
 func (q *queue) DecRef(ctx context.Context) {
-	q.DecRefWithDestructor(ctx, q.Reset)
-	// We don't need to notify after resetting because no one cares about
-	// this queue after all references have been dropped.
+	q.queueRefs.DecRef(func() {
+		// We don't need to notify after resetting because no one cares about
+		// this queue after all references have been dropped.
+		q.Reset(ctx)
+	})
 }
 
 // IsReadable determines if q is currently readable.
diff --git a/pkg/sentry/socket/unix/transport/queue_refs.go b/pkg/sentry/socket/unix/transport/queue_refs.go
new file mode 100644
index 000000000..0d4e34988
--- /dev/null
+++ b/pkg/sentry/socket/unix/transport/queue_refs.go
@@ -0,0 +1,118 @@
+package transport
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var queueownerType *queue
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type queueRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *queueRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, queueownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *queueRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*queueRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *queueRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *queueRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, queueownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *queueRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *queueRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, queueownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/socket/unix/transport/transport_state_autogen.go b/pkg/sentry/socket/unix/transport/transport_state_autogen.go
index 4b88ea3ae..91e632833 100644
--- a/pkg/sentry/socket/unix/transport/transport_state_autogen.go
+++ b/pkg/sentry/socket/unix/transport/transport_state_autogen.go
@@ -71,7 +71,7 @@ func (x *queue) StateTypeName() string {
 
 func (x *queue) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"queueRefs",
 		"ReaderQueue",
 		"WriterQueue",
 		"closed",
@@ -86,7 +86,7 @@ func (x *queue) beforeSave() {}
 
 func (x *queue) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.queueRefs)
 	m.Save(1, &x.ReaderQueue)
 	m.Save(2, &x.WriterQueue)
 	m.Save(3, &x.closed)
@@ -99,7 +99,7 @@ func (x *queue) StateSave(m state.Sink) {
 func (x *queue) afterLoad() {}
 
 func (x *queue) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.queueRefs)
 	m.Load(1, &x.ReaderQueue)
 	m.Load(2, &x.WriterQueue)
 	m.Load(3, &x.closed)
@@ -109,6 +109,29 @@ func (x *queue) StateLoad(m state.Source) {
 	m.Load(7, &x.dataList)
 }
 
+func (x *queueRefs) StateTypeName() string {
+	return "pkg/sentry/socket/unix/transport.queueRefs"
+}
+
+func (x *queueRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *queueRefs) beforeSave() {}
+
+func (x *queueRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *queueRefs) afterLoad() {}
+
+func (x *queueRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *messageList) StateTypeName() string {
 	return "pkg/sentry/socket/unix/transport.messageList"
 }
@@ -339,6 +362,7 @@ func init() {
 	state.Register((*connectionedEndpoint)(nil))
 	state.Register((*connectionlessEndpoint)(nil))
 	state.Register((*queue)(nil))
+	state.Register((*queueRefs)(nil))
 	state.Register((*messageList)(nil))
 	state.Register((*messageEntry)(nil))
 	state.Register((*ControlMessages)(nil))
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index 3219a9e13..22a54fa48 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -38,9 +38,7 @@ import (
 //
 // FileDescription is analogous to Linux's struct file.
 type FileDescription struct {
-	// refs is the reference count. refs is accessed using atomic memory
-	// operations.
-	refs int64
+	FileDescriptionRefs
 
 	// flagsMu protects statusFlags and asyncHandler below.
 	flagsMu sync.Mutex
@@ -131,7 +129,7 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou
 		}
 	}
 
-	fd.refs = 1
+	fd.EnableLeakCheck()
 
 	// Remove "file creation flags" to mirror the behavior from file.f_flags in
 	// fs/open.c:do_dentry_open.
@@ -149,30 +147,9 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou
 	return nil
 }
 
-// IncRef increments fd's reference count.
-func (fd *FileDescription) IncRef() {
-	atomic.AddInt64(&fd.refs, 1)
-}
-
-// TryIncRef increments fd's reference count and returns true. If fd's
-// reference count is already zero, TryIncRef does nothing and returns false.
-//
-// TryIncRef does not require that a reference is held on fd.
-func (fd *FileDescription) TryIncRef() bool {
-	for {
-		refs := atomic.LoadInt64(&fd.refs)
-		if refs <= 0 {
-			return false
-		}
-		if atomic.CompareAndSwapInt64(&fd.refs, refs, refs+1) {
-			return true
-		}
-	}
-}
-
 // DecRef decrements fd's reference count.
 func (fd *FileDescription) DecRef(ctx context.Context) {
-	if refs := atomic.AddInt64(&fd.refs, -1); refs == 0 {
+	fd.FileDescriptionRefs.DecRef(func() {
 		// Unregister fd from all epoll instances.
 		fd.epollMu.Lock()
 		epolls := fd.epolls
@@ -208,15 +185,7 @@ func (fd *FileDescription) DecRef(ctx context.Context) {
 		}
 		fd.asyncHandler = nil
 		fd.flagsMu.Unlock()
-	} else if refs < 0 {
-		panic("FileDescription.DecRef() called without holding a reference")
-	}
-}
-
-// Refs returns the current number of references. The returned count
-// is inherently racy and is unsafe to use without external synchronization.
-func (fd *FileDescription) Refs() int64 {
-	return atomic.LoadInt64(&fd.refs)
+	})
 }
 
 // Mount returns the mount on which fd was opened. It does not take a reference
diff --git a/pkg/sentry/vfs/file_description_refs.go b/pkg/sentry/vfs/file_description_refs.go
new file mode 100644
index 000000000..bdd7e6554
--- /dev/null
+++ b/pkg/sentry/vfs/file_description_refs.go
@@ -0,0 +1,118 @@
+package vfs
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var FileDescriptionownerType *FileDescription
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type FileDescriptionRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *FileDescriptionRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, FileDescriptionownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *FileDescriptionRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*FileDescriptionRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *FileDescriptionRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *FileDescriptionRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, FileDescriptionownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *FileDescriptionRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *FileDescriptionRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, FileDescriptionownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go
index 2c60cfab2..46851f638 100644
--- a/pkg/sentry/vfs/filesystem.go
+++ b/pkg/sentry/vfs/filesystem.go
@@ -15,8 +15,6 @@
 package vfs
 
 import (
-	"sync/atomic"
-
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/fspath"
@@ -34,9 +32,7 @@ import (
 //
 // +stateify savable
 type Filesystem struct {
-	// refs is the reference count. refs is accessed using atomic memory
-	// operations.
-	refs int64
+	FilesystemRefs
 
 	// vfs is the VirtualFilesystem that uses this Filesystem. vfs is
 	// immutable.
@@ -52,7 +48,7 @@ type Filesystem struct {
 
 // Init must be called before first use of fs.
 func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, fsType FilesystemType, impl FilesystemImpl) {
-	fs.refs = 1
+	fs.EnableLeakCheck()
 	fs.vfs = vfsObj
 	fs.fsType = fsType
 	fs.impl = impl
@@ -76,39 +72,14 @@ func (fs *Filesystem) Impl() FilesystemImpl {
 	return fs.impl
 }
 
-// IncRef increments fs' reference count.
-func (fs *Filesystem) IncRef() {
-	if atomic.AddInt64(&fs.refs, 1) <= 1 {
-		panic("Filesystem.IncRef() called without holding a reference")
-	}
-}
-
-// TryIncRef increments fs' reference count and returns true. If fs' reference
-// count is zero, TryIncRef does nothing and returns false.
-//
-// TryIncRef does not require that a reference is held on fs.
-func (fs *Filesystem) TryIncRef() bool {
-	for {
-		refs := atomic.LoadInt64(&fs.refs)
-		if refs <= 0 {
-			return false
-		}
-		if atomic.CompareAndSwapInt64(&fs.refs, refs, refs+1) {
-			return true
-		}
-	}
-}
-
 // DecRef decrements fs' reference count.
 func (fs *Filesystem) DecRef(ctx context.Context) {
-	if refs := atomic.AddInt64(&fs.refs, -1); refs == 0 {
+	fs.FilesystemRefs.DecRef(func() {
 		fs.vfs.filesystemsMu.Lock()
 		delete(fs.vfs.filesystems, fs)
 		fs.vfs.filesystemsMu.Unlock()
 		fs.impl.Release(ctx)
-	} else if refs < 0 {
-		panic("Filesystem.decRef() called without holding a reference")
-	}
+	})
 }
 
 // FilesystemImpl contains implementation details for a Filesystem.
diff --git a/pkg/sentry/vfs/filesystem_refs.go b/pkg/sentry/vfs/filesystem_refs.go
new file mode 100644
index 000000000..38a9a986f
--- /dev/null
+++ b/pkg/sentry/vfs/filesystem_refs.go
@@ -0,0 +1,118 @@
+package vfs
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var FilesystemownerType *Filesystem
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type FilesystemRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *FilesystemRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, FilesystemownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *FilesystemRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*FilesystemRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *FilesystemRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *FilesystemRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, FilesystemownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *FilesystemRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *FilesystemRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, FilesystemownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go
index cd5456eef..db5fb3bb1 100644
--- a/pkg/sentry/vfs/mount.go
+++ b/pkg/sentry/vfs/mount.go
@@ -128,16 +128,14 @@ func (mnt *Mount) Options() MountOptions {
 //
 // +stateify savable
 type MountNamespace struct {
+	MountNamespaceRefs
+
 	// Owner is the usernamespace that owns this mount namespace.
 	Owner *auth.UserNamespace
 
 	// root is the MountNamespace's root mount. root is immutable.
 	root *Mount
 
-	// refs is the reference count. refs is accessed using atomic memory
-	// operations.
-	refs int64
-
 	// mountpoints maps all Dentries which are mount points in this namespace
 	// to the number of Mounts for which they are mount points. mountpoints is
 	// protected by VirtualFilesystem.mountMu.
@@ -168,9 +166,9 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth
 	}
 	mntns := &MountNamespace{
 		Owner:       creds.UserNamespace,
-		refs:        1,
 		mountpoints: make(map[*Dentry]uint32),
 	}
+	mntns.EnableLeakCheck()
 	mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{})
 	return mntns, nil
 }
@@ -509,17 +507,10 @@ func (mnt *Mount) DecRef(ctx context.Context) {
 	}
 }
 
-// IncRef increments mntns' reference count.
-func (mntns *MountNamespace) IncRef() {
-	if atomic.AddInt64(&mntns.refs, 1) <= 1 {
-		panic("MountNamespace.IncRef() called without holding a reference")
-	}
-}
-
 // DecRef decrements mntns' reference count.
 func (mntns *MountNamespace) DecRef(ctx context.Context) {
 	vfs := mntns.root.fs.VirtualFilesystem()
-	if refs := atomic.AddInt64(&mntns.refs, -1); refs == 0 {
+	mntns.MountNamespaceRefs.DecRef(func() {
 		vfs.mountMu.Lock()
 		vfs.mounts.seq.BeginWrite()
 		vdsToDecRef, mountsToDecRef := vfs.umountRecursiveLocked(mntns.root, &umountRecursiveOptions{
@@ -533,9 +524,7 @@ func (mntns *MountNamespace) DecRef(ctx context.Context) {
 		for _, mnt := range mountsToDecRef {
 			mnt.DecRef(ctx)
 		}
-	} else if refs < 0 {
-		panic("MountNamespace.DecRef() called without holding a reference")
-	}
+	})
 }
 
 // getMountAt returns the last Mount in the stack mounted at (mnt, d). It takes
diff --git a/pkg/sentry/vfs/mount_namespace_refs.go b/pkg/sentry/vfs/mount_namespace_refs.go
new file mode 100644
index 000000000..63285fb8e
--- /dev/null
+++ b/pkg/sentry/vfs/mount_namespace_refs.go
@@ -0,0 +1,118 @@
+package vfs
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var MountNamespaceownerType *MountNamespace
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type MountNamespaceRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *MountNamespaceRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, MountNamespaceownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *MountNamespaceRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*MountNamespaceRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *MountNamespaceRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *MountNamespaceRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, MountNamespaceownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *MountNamespaceRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *MountNamespaceRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, MountNamespaceownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/vfs/vfs_state_autogen.go b/pkg/sentry/vfs/vfs_state_autogen.go
index 7bd988336..be649f60a 100644
--- a/pkg/sentry/vfs/vfs_state_autogen.go
+++ b/pkg/sentry/vfs/vfs_state_autogen.go
@@ -188,13 +188,36 @@ func (x *eventEntry) StateLoad(m state.Source) {
 	m.Load(1, &x.prev)
 }
 
+func (x *FileDescriptionRefs) StateTypeName() string {
+	return "pkg/sentry/vfs.FileDescriptionRefs"
+}
+
+func (x *FileDescriptionRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *FileDescriptionRefs) beforeSave() {}
+
+func (x *FileDescriptionRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *FileDescriptionRefs) afterLoad() {}
+
+func (x *FileDescriptionRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *Filesystem) StateTypeName() string {
 	return "pkg/sentry/vfs.Filesystem"
 }
 
 func (x *Filesystem) StateFields() []string {
 	return []string{
-		"refs",
+		"FilesystemRefs",
 		"vfs",
 		"fsType",
 		"impl",
@@ -205,7 +228,7 @@ func (x *Filesystem) beforeSave() {}
 
 func (x *Filesystem) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.refs)
+	m.Save(0, &x.FilesystemRefs)
 	m.Save(1, &x.vfs)
 	m.Save(2, &x.fsType)
 	m.Save(3, &x.impl)
@@ -214,12 +237,35 @@ func (x *Filesystem) StateSave(m state.Sink) {
 func (x *Filesystem) afterLoad() {}
 
 func (x *Filesystem) StateLoad(m state.Source) {
-	m.Load(0, &x.refs)
+	m.Load(0, &x.FilesystemRefs)
 	m.Load(1, &x.vfs)
 	m.Load(2, &x.fsType)
 	m.Load(3, &x.impl)
 }
 
+func (x *FilesystemRefs) StateTypeName() string {
+	return "pkg/sentry/vfs.FilesystemRefs"
+}
+
+func (x *FilesystemRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *FilesystemRefs) beforeSave() {}
+
+func (x *FilesystemRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *FilesystemRefs) afterLoad() {}
+
+func (x *FilesystemRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *registeredFilesystemType) StateTypeName() string {
 	return "pkg/sentry/vfs.registeredFilesystemType"
 }
@@ -448,9 +494,9 @@ func (x *MountNamespace) StateTypeName() string {
 
 func (x *MountNamespace) StateFields() []string {
 	return []string{
+		"MountNamespaceRefs",
 		"Owner",
 		"root",
-		"refs",
 		"mountpoints",
 	}
 }
@@ -459,21 +505,44 @@ func (x *MountNamespace) beforeSave() {}
 
 func (x *MountNamespace) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.Owner)
-	m.Save(1, &x.root)
-	m.Save(2, &x.refs)
+	m.Save(0, &x.MountNamespaceRefs)
+	m.Save(1, &x.Owner)
+	m.Save(2, &x.root)
 	m.Save(3, &x.mountpoints)
 }
 
 func (x *MountNamespace) afterLoad() {}
 
 func (x *MountNamespace) StateLoad(m state.Source) {
-	m.Load(0, &x.Owner)
-	m.Load(1, &x.root)
-	m.Load(2, &x.refs)
+	m.Load(0, &x.MountNamespaceRefs)
+	m.Load(1, &x.Owner)
+	m.Load(2, &x.root)
 	m.Load(3, &x.mountpoints)
 }
 
+func (x *MountNamespaceRefs) StateTypeName() string {
+	return "pkg/sentry/vfs.MountNamespaceRefs"
+}
+
+func (x *MountNamespaceRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *MountNamespaceRefs) beforeSave() {}
+
+func (x *MountNamespaceRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *MountNamespaceRefs) afterLoad() {}
+
+func (x *MountNamespaceRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *VirtualFilesystem) StateTypeName() string {
 	return "pkg/sentry/vfs.VirtualFilesystem"
 }
@@ -555,7 +624,9 @@ func init() {
 	state.Register((*epollInterestEntry)(nil))
 	state.Register((*eventList)(nil))
 	state.Register((*eventEntry)(nil))
+	state.Register((*FileDescriptionRefs)(nil))
 	state.Register((*Filesystem)(nil))
+	state.Register((*FilesystemRefs)(nil))
 	state.Register((*registeredFilesystemType)(nil))
 	state.Register((*Inotify)(nil))
 	state.Register((*Watches)(nil))
@@ -563,6 +634,7 @@ func init() {
 	state.Register((*Event)(nil))
 	state.Register((*Mount)(nil))
 	state.Register((*MountNamespace)(nil))
+	state.Register((*MountNamespaceRefs)(nil))
 	state.Register((*VirtualFilesystem)(nil))
 	state.Register((*VirtualDentry)(nil))
 }
diff --git a/pkg/tcpip/link/tun/device.go b/pkg/tcpip/link/tun/device.go
index 3b1510a33..b6ddbe81e 100644
--- a/pkg/tcpip/link/tun/device.go
+++ b/pkg/tcpip/link/tun/device.go
@@ -19,7 +19,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/tcpip"
@@ -135,6 +134,7 @@ func attachOrCreateNIC(s *stack.Stack, name, prefix string, linkCaps stack.LinkE
 
 		// 2. Creating a new NIC.
 		id := tcpip.NICID(s.UniqueID())
+		// TODO(gvisor.dev/1486): enable leak check for tunEndpoint.
 		endpoint := &tunEndpoint{
 			Endpoint: channel.New(defaultDevOutQueueLen, defaultDevMtu, ""),
 			stack:    s,
@@ -331,19 +331,18 @@ func (d *Device) WriteNotify() {
 // It is ref-counted as multiple opening files can attach to the same NIC.
 // The last owner is responsible for deleting the NIC.
 type tunEndpoint struct {
+	tunEndpointRefs
 	*channel.Endpoint
 
-	refs.AtomicRefCount
-
 	stack *stack.Stack
 	nicID tcpip.NICID
 	name  string
 	isTap bool
 }
 
-// DecRef decrements refcount of e, removes NIC if refcount goes to 0.
+// DecRef decrements refcount of e, removing NIC if it reaches 0.
 func (e *tunEndpoint) DecRef(ctx context.Context) {
-	e.DecRefWithDestructor(ctx, func(context.Context) {
+	e.tunEndpointRefs.DecRef(func() {
 		e.stack.RemoveNIC(e.nicID)
 	})
 }
diff --git a/pkg/tcpip/link/tun/tun_endpoint_refs.go b/pkg/tcpip/link/tun/tun_endpoint_refs.go
new file mode 100644
index 000000000..e0595429c
--- /dev/null
+++ b/pkg/tcpip/link/tun/tun_endpoint_refs.go
@@ -0,0 +1,118 @@
+package tun
+
+import (
+	"fmt"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+)
+
+// ownerType is used to customize logging. Note that we use a pointer to T so
+// that we do not copy the entire object when passed as a format parameter.
+var tunEndpointownerType *tunEndpoint
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// Note that the number of references is actually refCount + 1 so that a default
+// zero-value Refs object contains one reference.
+//
+// TODO(gvisor.dev/issue/1486): Store stack traces when leak check is enabled in
+// a map with 16-bit hashes, and store the hash in the top 16 bits of refCount.
+// This will allow us to add stack trace information to the leak messages
+// without growing the size of Refs.
+//
+// +stateify savable
+type tunEndpointRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+func (r *tunEndpointRefs) finalize() {
+	var note string
+	switch refs_vfs1.GetLeakMode() {
+	case refs_vfs1.NoLeakChecking:
+		return
+	case refs_vfs1.UninitializedLeakChecking:
+		note = "(Leak checker uninitialized): "
+	}
+	if n := r.ReadRefs(); n != 0 {
+		log.Warningf("%sRefs %p owned by %T garbage collected with ref count of %d (want 0)", note, r, tunEndpointownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *tunEndpointRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*tunEndpointRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *tunEndpointRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *tunEndpointRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic(fmt.Sprintf("Incrementing non-positive ref count %p owned by %T", r, tunEndpointownerType))
+	}
+}
+
+// TryIncRef implements refs.RefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *tunEndpointRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	v := atomic.AddInt64(&r.refCount, speculativeRef)
+	if int32(v) < 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *tunEndpointRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %T", r, tunEndpointownerType))
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/tcpip/link/tun/tun_state_autogen.go b/pkg/tcpip/link/tun/tun_state_autogen.go
index 7f75aef19..0b3041e84 100644
--- a/pkg/tcpip/link/tun/tun_state_autogen.go
+++ b/pkg/tcpip/link/tun/tun_state_autogen.go
@@ -36,6 +36,30 @@ func (x *Device) StateLoad(m state.Source) {
 	m.Load(3, &x.flags)
 }
 
+func (x *tunEndpointRefs) StateTypeName() string {
+	return "pkg/tcpip/link/tun.tunEndpointRefs"
+}
+
+func (x *tunEndpointRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *tunEndpointRefs) beforeSave() {}
+
+func (x *tunEndpointRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *tunEndpointRefs) afterLoad() {}
+
+func (x *tunEndpointRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func init() {
 	state.Register((*Device)(nil))
+	state.Register((*tunEndpointRefs)(nil))
 }