Merge pull request #3875 from btw616:fix/issue-3874

PiperOrigin-RevId: 334428344

8 files changed, 134 insertions, 46 deletions

diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD
index a43c549f1..083071b5e 100644
--- a/pkg/sentry/kernel/BUILD
+++ b/pkg/sentry/kernel/BUILD
@@ -69,8 +69,8 @@ go_template_instance(
     prefix = "socket",
     template = "//pkg/ilist:generic_list",
     types = {
-        "Element": "*SocketEntry",
-        "Linker": "*SocketEntry",
+        "Element": "*SocketRecordVFS1",
+        "Linker": "*SocketRecordVFS1",
     },
 )
 
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index 08bb5bd12..d6c21adb7 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -220,13 +220,18 @@ type Kernel struct {
 	// danglingEndpoints is used to save / restore tcpip.DanglingEndpoints.
 	danglingEndpoints struct{} `state:".([]tcpip.Endpoint)"`
 
-	// sockets is the list of all network sockets the system. Protected by
-	// extMu.
+	// sockets is the list of all network sockets in the system.
+	// Protected by extMu.
+	// TODO(gvisor.dev/issue/1624): Only used by VFS1.
 	sockets socketList
 
-	// nextSocketEntry is the next entry number to use in sockets. Protected
+	// socketsVFS2 records all network sockets in the system. Protected by
+	// extMu.
+	socketsVFS2 map[*vfs.FileDescription]*SocketRecord
+
+	// nextSocketRecord is the next entry number to use in sockets. Protected
 	// by extMu.
-	nextSocketEntry uint64
+	nextSocketRecord uint64
 
 	// deviceRegistry is used to save/restore device.SimpleDevices.
 	deviceRegistry struct{} `state:".(*device.Registry)"`
@@ -414,6 +419,8 @@ func (k *Kernel) Init(args InitKernelArgs) error {
 			return fmt.Errorf("failed to create sockfs mount: %v", err)
 		}
 		k.socketMount = socketMount
+
+		k.socketsVFS2 = make(map[*vfs.FileDescription]*SocketRecord)
 	}
 
 	return nil
@@ -1512,20 +1519,27 @@ func (k *Kernel) SupervisorContext() context.Context {
 	}
 }
 
-// SocketEntry represents a socket recorded in Kernel.sockets. It implements
+// SocketRecord represents a socket recorded in Kernel.socketsVFS2.
+//
+// +stateify savable
+type SocketRecord struct {
+	k        *Kernel
+	Sock     *refs.WeakRef        // TODO(gvisor.dev/issue/1624): Only used by VFS1.
+	SockVFS2 *vfs.FileDescription // Only used by VFS2.
+	ID       uint64               // Socket table entry number.
+}
+
+// SocketRecordVFS1 represents a socket recorded in Kernel.sockets. It implements
 // refs.WeakRefUser for sockets stored in the socket table.
 //
 // +stateify savable
-type SocketEntry struct {
+type SocketRecordVFS1 struct {
 	socketEntry
-	k        *Kernel
-	Sock     *refs.WeakRef
-	SockVFS2 *vfs.FileDescription
-	ID       uint64 // Socket table entry number.
+	SocketRecord
 }
 
 // WeakRefGone implements refs.WeakRefUser.WeakRefGone.
-func (s *SocketEntry) WeakRefGone(context.Context) {
+func (s *SocketRecordVFS1) WeakRefGone(context.Context) {
 	s.k.extMu.Lock()
 	s.k.sockets.Remove(s)
 	s.k.extMu.Unlock()
@@ -1536,9 +1550,14 @@ func (s *SocketEntry) WeakRefGone(context.Context) {
 // Precondition: Caller must hold a reference to sock.
 func (k *Kernel) RecordSocket(sock *fs.File) {
 	k.extMu.Lock()
-	id := k.nextSocketEntry
-	k.nextSocketEntry++
-	s := &SocketEntry{k: k, ID: id}
+	id := k.nextSocketRecord
+	k.nextSocketRecord++
+	s := &SocketRecordVFS1{
+		SocketRecord: SocketRecord{
+			k:  k,
+			ID: id,
+		},
+	}
 	s.Sock = refs.NewWeakRef(sock, s)
 	k.sockets.PushBack(s)
 	k.extMu.Unlock()
@@ -1550,29 +1569,45 @@ func (k *Kernel) RecordSocket(sock *fs.File) {
 // Precondition: Caller must hold a reference to sock.
 //
 // Note that the socket table will not hold a reference on the
-// vfs.FileDescription, because we do not support weak refs on VFS2 files.
+// vfs.FileDescription.
 func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) {
 	k.extMu.Lock()
-	id := k.nextSocketEntry
-	k.nextSocketEntry++
-	s := &SocketEntry{
+	if _, ok := k.socketsVFS2[sock]; ok {
+		panic(fmt.Sprintf("Socket %p added twice", sock))
+	}
+	id := k.nextSocketRecord
+	k.nextSocketRecord++
+	s := &SocketRecord{
 		k:        k,
 		ID:       id,
 		SockVFS2: sock,
 	}
-	k.sockets.PushBack(s)
+	k.socketsVFS2[sock] = s
+	k.extMu.Unlock()
+}
+
+// DeleteSocketVFS2 removes a VFS2 socket from the system-wide socket table.
+func (k *Kernel) DeleteSocketVFS2(sock *vfs.FileDescription) {
+	k.extMu.Lock()
+	delete(k.socketsVFS2, sock)
 	k.extMu.Unlock()
 }
 
 // ListSockets returns a snapshot of all sockets.
 //
-// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef()
+// Callers of ListSockets() in VFS2 should use SocketRecord.SockVFS2.TryIncRef()
 // to get a reference on a socket in the table.
-func (k *Kernel) ListSockets() []*SocketEntry {
+func (k *Kernel) ListSockets() []*SocketRecord {
 	k.extMu.Lock()
-	var socks []*SocketEntry
-	for s := k.sockets.Front(); s != nil; s = s.Next() {
-		socks = append(socks, s)
+	var socks []*SocketRecord
+	if VFS2Enabled {
+		for _, s := range k.socketsVFS2 {
+			socks = append(socks, s)
+		}
+	} else {
+		for s := k.sockets.Front(); s != nil; s = s.Next() {
+			socks = append(socks, &s.SocketRecord)
+		}
 	}
 	k.extMu.Unlock()
 	return socks
diff --git a/pkg/sentry/socket/hostinet/socket_vfs2.go b/pkg/sentry/socket/hostinet/socket_vfs2.go
index 87b077e68..163af329b 100644
--- a/pkg/sentry/socket/hostinet/socket_vfs2.go
+++ b/pkg/sentry/socket/hostinet/socket_vfs2.go
@@ -78,6 +78,13 @@ func newVFS2Socket(t *kernel.Task, family int, stype linux.SockType, protocol in
 	return vfsfd, nil
 }
 
+// Release implements vfs.FileDescriptionImpl.Release.
+func (s *socketVFS2) Release(ctx context.Context) {
+	t := kernel.TaskFromContext(ctx)
+	t.Kernel().DeleteSocketVFS2(&s.vfsfd)
+	s.socketOpsCommon.Release(ctx)
+}
+
 // Readiness implements waiter.Waitable.Readiness.
 func (s *socketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return s.socketOpsCommon.Readiness(mask)
diff --git a/pkg/sentry/socket/netlink/socket_vfs2.go b/pkg/sentry/socket/netlink/socket_vfs2.go
index a38d25da9..c83b23242 100644
--- a/pkg/sentry/socket/netlink/socket_vfs2.go
+++ b/pkg/sentry/socket/netlink/socket_vfs2.go
@@ -82,6 +82,13 @@ func NewVFS2(t *kernel.Task, skType linux.SockType, protocol Protocol) (*SocketV
 	return fd, nil
 }
 
+// Release implements vfs.FileDescriptionImpl.Release.
+func (s *SocketVFS2) Release(ctx context.Context) {
+	t := kernel.TaskFromContext(ctx)
+	t.Kernel().DeleteSocketVFS2(&s.vfsfd)
+	s.socketOpsCommon.Release(ctx)
+}
+
 // Readiness implements waiter.Waitable.Readiness.
 func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return s.socketOpsCommon.Readiness(mask)
diff --git a/pkg/sentry/socket/netstack/netstack_vfs2.go b/pkg/sentry/socket/netstack/netstack_vfs2.go
index c0212ad76..4c6791fff 100644
--- a/pkg/sentry/socket/netstack/netstack_vfs2.go
+++ b/pkg/sentry/socket/netstack/netstack_vfs2.go
@@ -79,6 +79,13 @@ func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, qu
 	return vfsfd, nil
 }
 
+// Release implements vfs.FileDescriptionImpl.Release.
+func (s *SocketVFS2) Release(ctx context.Context) {
+	t := kernel.TaskFromContext(ctx)
+	t.Kernel().DeleteSocketVFS2(&s.vfsfd)
+	s.socketOpsCommon.Release(ctx)
+}
+
 // Readiness implements waiter.Waitable.Readiness.
 func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return s.socketOpsCommon.Readiness(mask)
diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD
index a89583dad..cc7408698 100644
--- a/pkg/sentry/socket/unix/BUILD
+++ b/pkg/sentry/socket/unix/BUILD
@@ -7,10 +7,21 @@ go_template_instance(
     name = "socket_refs",
     out = "socket_refs.go",
     package = "unix",
-    prefix = "socketOpsCommon",
+    prefix = "socketOperations",
     template = "//pkg/refs_vfs2:refs_template",
     types = {
-        "T": "socketOpsCommon",
+        "T": "SocketOperations",
+    },
+)
+
+go_template_instance(
+    name = "socket_vfs2_refs",
+    out = "socket_vfs2_refs.go",
+    package = "unix",
+    prefix = "socketVFS2",
+    template = "//pkg/refs_vfs2:refs_template",
+    types = {
+        "T": "SocketVFS2",
     },
 )
 
@@ -20,6 +31,7 @@ go_library(
         "device.go",
         "io.go",
         "socket_refs.go",
+        "socket_vfs2_refs.go",
         "unix.go",
         "unix_vfs2.go",
     ],
diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go
index 917055cea..f80011ce4 100644
--- a/pkg/sentry/socket/unix/unix.go
+++ b/pkg/sentry/socket/unix/unix.go
@@ -55,6 +55,7 @@ type SocketOperations struct {
 	fsutil.FileNoopFlush            `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
 
+	socketOperationsRefs
 	socketOpsCommon
 }
 
@@ -84,11 +85,27 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty
 	return fs.NewFile(ctx, d, flags, &s)
 }
 
+// DecRef implements RefCounter.DecRef.
+func (s *SocketOperations) DecRef(ctx context.Context) {
+	s.socketOperationsRefs.DecRef(func() {
+		s.ep.Close(ctx)
+		if s.abstractNamespace != nil {
+			s.abstractNamespace.Remove(s.abstractName, s)
+		}
+	})
+}
+
+// Release implemements fs.FileOperations.Release.
+func (s *SocketOperations) Release(ctx context.Context) {
+	// Release only decrements a reference on s because s may be referenced in
+	// the abstract socket namespace.
+	s.DecRef(ctx)
+}
+
 // socketOpsCommon contains the socket operations common to VFS1 and VFS2.
 //
 // +stateify savable
 type socketOpsCommon struct {
-	socketOpsCommonRefs
 	socket.SendReceiveTimeout
 
 	ep    transport.Endpoint
@@ -101,23 +118,6 @@ type socketOpsCommon struct {
 	abstractNamespace *kernel.AbstractSocketNamespace
 }
 
-// DecRef implements RefCounter.DecRef.
-func (s *socketOpsCommon) DecRef(ctx context.Context) {
-	s.socketOpsCommonRefs.DecRef(func() {
-		s.ep.Close(ctx)
-		if s.abstractNamespace != nil {
-			s.abstractNamespace.Remove(s.abstractName, s)
-		}
-	})
-}
-
-// Release implemements fs.FileOperations.Release.
-func (s *socketOpsCommon) Release(ctx context.Context) {
-	// Release only decrements a reference on s because s may be referenced in
-	// the abstract socket namespace.
-	s.DecRef(ctx)
-}
-
 func (s *socketOpsCommon) isPacket() bool {
 	switch s.stype {
 	case linux.SOCK_DGRAM, linux.SOCK_SEQPACKET:
diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go
index 8b1abd922..3345124cc 100644
--- a/pkg/sentry/socket/unix/unix_vfs2.go
+++ b/pkg/sentry/socket/unix/unix_vfs2.go
@@ -45,6 +45,7 @@ type SocketVFS2 struct {
 	vfs.DentryMetadataFileDescriptionImpl
 	vfs.LockFD
 
+	socketVFS2Refs
 	socketOpsCommon
 }
 
@@ -91,6 +92,25 @@ func NewFileDescription(ep transport.Endpoint, stype linux.SockType, flags uint3
 	return vfsfd, nil
 }
 
+// DecRef implements RefCounter.DecRef.
+func (s *SocketVFS2) DecRef(ctx context.Context) {
+	s.socketVFS2Refs.DecRef(func() {
+		t := kernel.TaskFromContext(ctx)
+		t.Kernel().DeleteSocketVFS2(&s.vfsfd)
+		s.ep.Close(ctx)
+		if s.abstractNamespace != nil {
+			s.abstractNamespace.Remove(s.abstractName, s)
+		}
+	})
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (s *SocketVFS2) Release(ctx context.Context) {
+	// Release only decrements a reference on s because s may be referenced in
+	// the abstract socket namespace.
+	s.DecRef(ctx)
+}
+
 // GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by
 // a transport.Endpoint.
 func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) {