Merge release-20200810.0-39-g3bd066d50 (automated)

4 files changed, 166 insertions, 9 deletions

diff --git a/pkg/sentry/socket/unix/socket_refs.go b/pkg/sentry/socket/unix/socket_refs.go
new file mode 100644
index 000000000..8eb0c9327
--- /dev/null
+++ b/pkg/sentry/socket/unix/socket_refs.go
@@ -0,0 +1,111 @@
+package unix
+
+import (
+	"gvisor.dev/gvisor/pkg/log"
+	refs_vfs1 "gvisor.dev/gvisor/pkg/refs"
+	"runtime"
+	"sync/atomic"
+)
+
+// 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 socketOpsCommonownerType *socketOpsCommon
+
+// 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.
+//
+// +stateify savable
+type socketOpsCommonRefs 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 *socketOpsCommonRefs) 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, socketOpsCommonownerType, n)
+	}
+}
+
+// EnableLeakCheck checks for reference leaks when Refs gets garbage collected.
+func (r *socketOpsCommonRefs) EnableLeakCheck() {
+	if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking {
+		runtime.SetFinalizer(r, (*socketOpsCommonRefs).finalize)
+	}
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *socketOpsCommonRefs) ReadRefs() int64 {
+
+	return atomic.LoadInt64(&r.refCount) + 1
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *socketOpsCommonRefs) IncRef() {
+	if v := atomic.AddInt64(&r.refCount, 1); v <= 0 {
+		panic("Incrementing non-positive ref count")
+	}
+}
+
+// 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 *socketOpsCommonRefs) 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 *socketOpsCommonRefs) DecRef(destroy func()) {
+	switch v := atomic.AddInt64(&r.refCount, -1); {
+	case v < -1:
+		panic("Decrementing non-positive ref count")
+
+	case v == -1:
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go
index 2b8454edb..b7e8e4325 100644
--- a/pkg/sentry/socket/unix/unix.go
+++ b/pkg/sentry/socket/unix/unix.go
@@ -24,7 +24,6 @@ import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/fspath"
-	"gvisor.dev/gvisor/pkg/refs"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/fs/fsutil"
@@ -80,7 +79,7 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty
 			stype: stype,
 		},
 	}
-	s.EnableLeakCheck("unix.SocketOperations")
+	s.EnableLeakCheck()
 
 	return fs.NewFile(ctx, d, flags, &s)
 }
@@ -89,17 +88,26 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty
 //
 // +stateify savable
 type socketOpsCommon struct {
-	refs.AtomicRefCount
+	socketOpsCommonRefs
 	socket.SendReceiveTimeout
 
 	ep    transport.Endpoint
 	stype linux.SockType
+
+	// abstractName and abstractNamespace indicate the name and namespace of the
+	// socket if it is bound to an abstract socket namespace. Once the socket is
+	// bound, they cannot be modified.
+	abstractName      string
+	abstractNamespace *kernel.AbstractSocketNamespace
 }
 
 // DecRef implements RefCounter.DecRef.
 func (s *socketOpsCommon) DecRef(ctx context.Context) {
-	s.DecRefWithDestructor(ctx, func(context.Context) {
+	s.socketOpsCommonRefs.DecRef(func() {
 		s.ep.Close(ctx)
+		if s.abstractNamespace != nil {
+			s.abstractNamespace.Remove(s.abstractName, s)
+		}
 	})
 }
 
@@ -284,10 +292,14 @@ func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
 			if t.IsNetworkNamespaced() {
 				return syserr.ErrInvalidEndpointState
 			}
-			if err := t.AbstractSockets().Bind(t, p[1:], bep, s); err != nil {
+			asn := t.AbstractSockets()
+			name := p[1:]
+			if err := asn.Bind(t, name, bep, s); err != nil {
 				// syserr.ErrPortInUse corresponds to EADDRINUSE.
 				return syserr.ErrPortInUse
 			}
+			s.abstractName = name
+			s.abstractNamespace = asn
 		} else {
 			// The parent and name.
 			var d *fs.Dirent
diff --git a/pkg/sentry/socket/unix/unix_state_autogen.go b/pkg/sentry/socket/unix/unix_state_autogen.go
index 4a3bbc11b..6966529c6 100644
--- a/pkg/sentry/socket/unix/unix_state_autogen.go
+++ b/pkg/sentry/socket/unix/unix_state_autogen.go
@@ -6,6 +6,29 @@ import (
 	"gvisor.dev/gvisor/pkg/state"
 )
 
+func (x *socketOpsCommonRefs) StateTypeName() string {
+	return "pkg/sentry/socket/unix.socketOpsCommonRefs"
+}
+
+func (x *socketOpsCommonRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (x *socketOpsCommonRefs) beforeSave() {}
+
+func (x *socketOpsCommonRefs) StateSave(m state.Sink) {
+	x.beforeSave()
+	m.Save(0, &x.refCount)
+}
+
+func (x *socketOpsCommonRefs) afterLoad() {}
+
+func (x *socketOpsCommonRefs) StateLoad(m state.Source) {
+	m.Load(0, &x.refCount)
+}
+
 func (x *SocketOperations) StateTypeName() string {
 	return "pkg/sentry/socket/unix.SocketOperations"
 }
@@ -35,10 +58,12 @@ func (x *socketOpsCommon) StateTypeName() string {
 
 func (x *socketOpsCommon) StateFields() []string {
 	return []string{
-		"AtomicRefCount",
+		"socketOpsCommonRefs",
 		"SendReceiveTimeout",
 		"ep",
 		"stype",
+		"abstractName",
+		"abstractNamespace",
 	}
 }
 
@@ -46,22 +71,27 @@ func (x *socketOpsCommon) beforeSave() {}
 
 func (x *socketOpsCommon) StateSave(m state.Sink) {
 	x.beforeSave()
-	m.Save(0, &x.AtomicRefCount)
+	m.Save(0, &x.socketOpsCommonRefs)
 	m.Save(1, &x.SendReceiveTimeout)
 	m.Save(2, &x.ep)
 	m.Save(3, &x.stype)
+	m.Save(4, &x.abstractName)
+	m.Save(5, &x.abstractNamespace)
 }
 
 func (x *socketOpsCommon) afterLoad() {}
 
 func (x *socketOpsCommon) StateLoad(m state.Source) {
-	m.Load(0, &x.AtomicRefCount)
+	m.Load(0, &x.socketOpsCommonRefs)
 	m.Load(1, &x.SendReceiveTimeout)
 	m.Load(2, &x.ep)
 	m.Load(3, &x.stype)
+	m.Load(4, &x.abstractName)
+	m.Load(5, &x.abstractNamespace)
 }
 
 func init() {
+	state.Register((*socketOpsCommonRefs)(nil))
 	state.Register((*SocketOperations)(nil))
 	state.Register((*socketOpsCommon)(nil))
 }
diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go
index dfa25241a..d066ef8ab 100644
--- a/pkg/sentry/socket/unix/unix_vfs2.go
+++ b/pkg/sentry/socket/unix/unix_vfs2.go
@@ -183,10 +183,14 @@ func (s *SocketVFS2) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
 			if t.IsNetworkNamespaced() {
 				return syserr.ErrInvalidEndpointState
 			}
-			if err := t.AbstractSockets().Bind(t, p[1:], bep, s); err != nil {
+			asn := t.AbstractSockets()
+			name := p[1:]
+			if err := asn.Bind(t, name, bep, s); err != nil {
 				// syserr.ErrPortInUse corresponds to EADDRINUSE.
 				return syserr.ErrPortInUse
 			}
+			s.abstractName = name
+			s.abstractNamespace = asn
 		} else {
 			path := fspath.Parse(p)
 			root := t.FSContext().RootDirectoryVFS2()