make PMTUD on by default and settable via sockopt

We've supported PMTUD for a long time and just never turned it on.

Addresses #10344.

PiperOrigin-RevId: 632186215

pkg/sentry/socket/hostinet/sockopt.go

@@ -63,6 +63,7 @@ var SockOpts = []SockOpt{
 	{linux.SOL_IP, linux.IP_MULTICAST_IF, 0 /* kernel allows multiple structures to be passed */, true, true},
 	{linux.SOL_IP, linux.IP_MULTICAST_LOOP, 0 /* can be 32-bit int or 8-bit uint */, true, true},
 	{linux.SOL_IP, linux.IP_MULTICAST_TTL, 0 /* can be 32-bit int or 8-bit uint */, true, true},
+	{linux.SOL_IP, linux.IP_MTU_DISCOVER, 0 /* can be 32-bit int or 8-bit uint */, true, true},
 	{linux.SOL_IP, linux.IP_PKTINFO, sizeofInt32, true, true},
 	{linux.SOL_IP, linux.IP_RECVERR, sizeofInt32, true, true},
 	{linux.SOL_IP, linux.IP_RECVORIGDSTADDR, sizeofInt32, true, true},

pkg/sentry/socket/netstack/netstack.go

@@ -1762,6 +1762,30 @@ func getSockOptIP(t *kernel.Task, s socket.Socket, ep commonEndpoint, name int,
 			return nil, err
 		}
 		return &ret, nil
+
+	case linux.IP_MTU_DISCOVER:
+		if outLen < sizeOfInt32 {
+			return nil, syserr.ErrInvalidArgument
+		}
+
+		v, err := ep.GetSockOptInt(tcpip.MTUDiscoverOption)
+		if err != nil {
+			return nil, syserr.TranslateNetstackError(err)
+		}
+		switch tcpip.PMTUDStrategy(v) {
+		case tcpip.PMTUDiscoveryWant:
+			v = linux.IP_PMTUDISC_WANT
+		case tcpip.PMTUDiscoveryDont:
+			v = linux.IP_PMTUDISC_DONT
+		case tcpip.PMTUDiscoveryDo:
+			v = linux.IP_PMTUDISC_DO
+		case tcpip.PMTUDiscoveryProbe:
+			v = linux.IP_PMTUDISC_PROBE
+		default:
+			panic(fmt.Errorf("unknown PMTUD option: %d", v))
+		}
+		vP := primitive.Int32(v)
+		return &vP, nil
 	}
 	return nil, syserr.ErrProtocolNotAvailable
 }
@@ -2576,6 +2600,28 @@ func setSockOptIP(t *kernel.Task, s socket.Socket, ep commonEndpoint, name int,
 		log.Infof("IPT_SO_SET_ADD_COUNTERS is not supported")
 		return nil
 
+	case linux.IP_MTU_DISCOVER:
+		if len(optVal) == 0 {
+			return nil
+		}
+		v, err := parseIntOrChar(optVal)
+		if err != nil {
+			return err
+		}
+		switch v {
+		case linux.IP_PMTUDISC_DONT:
+			v = int32(tcpip.PMTUDiscoveryDont)
+		case linux.IP_PMTUDISC_WANT:
+			v = int32(tcpip.PMTUDiscoveryWant)
+		case linux.IP_PMTUDISC_DO:
+			v = int32(tcpip.PMTUDiscoveryDo)
+		case linux.IP_PMTUDISC_PROBE:
+			v = int32(tcpip.PMTUDiscoveryProbe)
+		default:
+			return syserr.ErrNotSupported
+		}
+		return syserr.TranslateNetstackError(ep.SetSockOptInt(tcpip.MTUDiscoverOption, int(v)))
+
 	case linux.IP_ADD_SOURCE_MEMBERSHIP,
 		linux.IP_BIND_ADDRESS_NO_PORT,
 		linux.IP_BLOCK_SOURCE,
@@ -2585,7 +2631,6 @@ func setSockOptIP(t *kernel.Task, s socket.Socket, ep commonEndpoint, name int,
 		linux.IP_IPSEC_POLICY,
 		linux.IP_MINTTL,
 		linux.IP_MSFILTER,
-		linux.IP_MTU_DISCOVER,
 		linux.IP_MULTICAST_ALL,
 		linux.IP_NODEFRAG,
 		linux.IP_OPTIONS,

pkg/tcpip/network/ipv4/ipv4.go

@@ -462,19 +462,26 @@ func (e *endpoint) addIPHeader(srcAddr, dstAddr tcpip.Address, pkt *stack.Packet
 	if length > math.MaxUint16 {
 		return &tcpip.ErrMessageTooLong{}
 	}
-	// RFC 6864 section 4.3 mandates uniqueness of ID values for non-atomic
-	// datagrams. Since the DF bit is never being set here, all datagrams
-	// are non-atomic and need an ID.
-	ipH.Encode(&header.IPv4Fields{
+
+	fields := header.IPv4Fields{
 		TotalLength: uint16(length),
-		ID:          e.getID(),
 		TTL:         params.TTL,
 		TOS:         params.TOS,
 		Protocol:    uint8(params.Protocol),
 		SrcAddr:     srcAddr,
 		DstAddr:     dstAddr,
 		Options:     options,
-	})
+	}
+	if params.DF {
+		// Treat want and do the same.
+		fields.Flags = header.IPv4FlagDontFragment
+	} else {
+		// RFC 6864 section 4.3 mandates uniqueness of ID values for
+		// non-atomic datagrams.
+		fields.ID = e.getID()
+	}
+	ipH.Encode(&fields)
+
 	ipH.SetChecksum(^ipH.CalculateChecksum())
 	pkt.NetworkProtocolNumber = ProtocolNumber
 	return nil

pkg/tcpip/stack/registration.go

@@ -319,6 +319,9 @@ type NetworkHeaderParams struct {
 
 	// TOS refers to TypeOfService or TrafficClass field of the IP-header.
 	TOS uint8
+
+	// DF indicates whether the DF bit should be set.
+	DF bool
 }
 
 // GroupAddressableEndpoint is an endpoint that supports group addressing.

pkg/tcpip/tcpip.go

@@ -995,10 +995,13 @@ const (
 	UseDefaultIPv6HopLimit = -1
 )
 
+// PMTUDStrategy is the kind of PMTUD to perform.
+type PMTUDStrategy int
+
 const (
 	// PMTUDiscoveryWant is a setting of the MTUDiscoverOption to use
 	// per-route settings.
-	PMTUDiscoveryWant int = iota
+	PMTUDiscoveryWant PMTUDStrategy = iota
 
 	// PMTUDiscoveryDont is a setting of the MTUDiscoverOption to disable
 	// path MTU discovery.

pkg/tcpip/transport/internal/network/endpoint.go

@@ -824,7 +824,7 @@ func (e *Endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) tcpip.Error {
 	case tcpip.MTUDiscoverOption:
 		// Return not supported if the value is not disabling path
 		// MTU discovery.
-		if v != tcpip.PMTUDiscoveryDont {
+		if tcpip.PMTUDStrategy(v) != tcpip.PMTUDiscoveryDont {
 			return &tcpip.ErrNotSupported{}
 		}
 
@@ -862,7 +862,7 @@ func (e *Endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, tcpip.Error) {
 	switch opt {
 	case tcpip.MTUDiscoverOption:
 		// The only supported setting is path MTU discovery disabled.
-		return tcpip.PMTUDiscoveryDont, nil
+		return int(tcpip.PMTUDiscoveryDont), nil
 
 	case tcpip.MulticastTTLOption:
 		e.mu.Lock()

pkg/tcpip/transport/tcp/connect.go

@@ -794,6 +794,7 @@ type tcpFields struct {
 	rcvWnd seqnum.Size
 	opts   []byte
 	txHash uint32
+	df     bool
 }
 
 func (e *Endpoint) sendSynTCP(r *stack.Route, tf tcpFields, opts header.TCPSynOptions) tcpip.Error {
@@ -881,7 +882,7 @@ func sendTCPBatch(r *stack.Route, tf tcpFields, pkt *stack.PacketBuffer, gso sta
 		buildTCPHdr(r, tf, pkt, gso)
 		tf.seq = tf.seq.Add(seqnum.Size(packetSize))
 		pkt.GSOOptions = gso
-		if err := r.WritePacket(stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos}, pkt); err != nil {
+		if err := r.WritePacket(stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos, DF: tf.df}, pkt); err != nil {
 			r.Stats().TCP.SegmentSendErrors.Increment()
 			if shouldSplitPacket {
 				pkt.DecRef()
@@ -913,7 +914,7 @@ func sendTCP(r *stack.Route, tf tcpFields, pkt *stack.PacketBuffer, gso stack.GS
 	pkt.Owner = owner
 	buildTCPHdr(r, tf, pkt, gso)
 
-	if err := r.WritePacket(stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos}, pkt); err != nil {
+	if err := r.WritePacket(stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos, DF: tf.df}, pkt); err != nil {
 		r.Stats().TCP.SegmentSendErrors.Increment()
 		return err
 	}
@@ -964,6 +965,9 @@ func (e *Endpoint) makeOptions(sackBlocks []header.SACKBlock) []byte {
 }
 
 // sendEmptyRaw sends a TCP segment with no payload to the endpoint's peer.
+//
+// +checklocks:e.mu
+// +checklocksalias:e.snd.ep.mu=e.mu
 func (e *Endpoint) sendEmptyRaw(flags header.TCPFlags, seq, ack seqnum.Value, rcvWnd seqnum.Size) tcpip.Error {
 	pkt := stack.NewPacketBuffer(stack.PacketBufferOptions{})
 	defer pkt.DecRef()
@@ -972,6 +976,9 @@ func (e *Endpoint) sendEmptyRaw(flags header.TCPFlags, seq, ack seqnum.Value, rc
 
 // sendRaw sends a TCP segment to the endpoint's peer. This method takes
 // ownership of pkt. pkt must not have any headers set.
+//
+// +checklocks:e.mu
+// +checklocksalias:e.snd.ep.mu=e.mu
 func (e *Endpoint) sendRaw(pkt *stack.PacketBuffer, flags header.TCPFlags, seq, ack seqnum.Value, rcvWnd seqnum.Size) tcpip.Error {
 	var sackBlocks []header.SACKBlock
 	if e.EndpointState() == StateEstablished && e.rcv.pendingRcvdSegments.Len() > 0 && (flags&header.TCPFlagAck != 0) {
@@ -989,6 +996,7 @@ func (e *Endpoint) sendRaw(pkt *stack.PacketBuffer, flags header.TCPFlags, seq,
 		ack:    ack,
 		rcvWnd: rcvWnd,
 		opts:   options,
+		df:     e.pmtud == tcpip.PMTUDiscoveryWant || e.pmtud == tcpip.PMTUDiscoveryDo,
 	}, pkt, e.gso)
 }
 

pkg/tcpip/transport/tcp/endpoint.go

@@ -600,6 +600,11 @@ type Endpoint struct {
 	//
 	// +checklocks:mu
 	limRdr *io.LimitedReader `state:"nosave"`
+
+	// pmtud is the PMTUD strategy to use.
+	//
+	// +checklocks:mu
+	pmtud tcpip.PMTUDStrategy
 }
 
 // UniqueID implements stack.TransportEndpoint.UniqueID.
@@ -1890,9 +1895,16 @@ func (e *Endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) tcpip.Error {
 		e.UnlockUser()
 
 	case tcpip.MTUDiscoverOption:
-		// Return not supported if attempting to set this option to
-		// anything other than path MTU discovery disabled.
-		if v != tcpip.PMTUDiscoveryDont {
+		switch v := tcpip.PMTUDStrategy(v); v {
+		case tcpip.PMTUDiscoveryWant, tcpip.PMTUDiscoveryDont, tcpip.PMTUDiscoveryDo:
+			e.LockUser()
+			e.pmtud = v
+			e.UnlockUser()
+		case tcpip.PMTUDiscoveryProbe:
+			// We don't support a way to ignore MTU updates; it's
+			// either on or it's off.
+			return &tcpip.ErrNotSupported{}
+		default:
 			return &tcpip.ErrNotSupported{}
 		}
 
@@ -2089,9 +2101,10 @@ func (e *Endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, tcpip.Error) {
 		return v, nil
 
 	case tcpip.MTUDiscoverOption:
-		// Always return the path MTU discovery disabled setting since
-		// it's the only one supported.
-		return tcpip.PMTUDiscoveryDont, nil
+		e.LockUser()
+		v := e.pmtud
+		e.UnlockUser()
+		return int(v), nil
 
 	case tcpip.ReceiveQueueSizeOption:
 		return e.readyReceiveSize()

pkg/tcpip/transport/tcp/test/e2e/tcp_test.go

@@ -4245,7 +4245,7 @@ func TestRetransmitIPv4IDUniqueness(t *testing.T) {
 			// have DF=0. This needs to be done because the IPv4 ID uniqueness
 			// applies only to non-atomic IPv4 datagrams as defined in RFC 6864
 			// Section 4, and datagrams with DF=0 are non-atomic.
-			if err := c.EP.SetSockOptInt(tcpip.MTUDiscoverOption, tcpip.PMTUDiscoveryDont); err != nil {
+			if err := c.EP.SetSockOptInt(tcpip.MTUDiscoverOption, int(tcpip.PMTUDiscoveryDont)); err != nil {
 				t.Fatalf("disabling PMTU discovery via sockopt to force DF=0 failed: %s", err)
 			}
 

test/syscalls/linux/tcp_socket.cc

@@ -18,6 +18,7 @@
 #include <linux/filter.h>
 #include <sys/epoll.h>
 #endif  // __linux__
+#include <errno.h>
 #include <netinet/in.h>
 #include <netinet/tcp.h>
 #include <poll.h>
@@ -2327,6 +2328,49 @@ TEST_P(TcpSocketTest, GetSocketAcceptConnNonListener) {
   EXPECT_EQ(got, 0);
 }
 
+TEST_P(TcpSocketTest, SetPMTUD) {
+  // IP_PMTUDISC_WANT should be default.
+  int got = -1;
+  socklen_t length = sizeof(got);
+  ASSERT_THAT(
+      getsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &got, &length),
+      SyscallSucceeds());
+  EXPECT_EQ(got, IP_PMTUDISC_WANT);
+
+  int set = IP_PMTUDISC_DO;
+  ASSERT_THAT(
+      setsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &set, length),
+      SyscallSucceeds());
+  ASSERT_THAT(
+      getsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &got, &length),
+      SyscallSucceeds());
+  EXPECT_EQ(got, IP_PMTUDISC_DO);
+  set = IP_PMTUDISC_DONT;
+  ASSERT_THAT(
+      setsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &set, length),
+      SyscallSucceeds());
+  ASSERT_THAT(
+      getsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &got, &length),
+      SyscallSucceeds());
+  EXPECT_EQ(got, IP_PMTUDISC_DONT);
+
+  // IP_PMTUDISC_PROBE is not supported by gVisor.
+  set = IP_PMTUDISC_PROBE;
+  if (IsRunningOnGvisor() && !IsRunningWithHostinet()) {
+    ASSERT_THAT(
+        setsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &set, length),
+        SyscallFailsWithErrno(ENOTSUP));
+  } else {
+    ASSERT_THAT(
+        setsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &set, length),
+        SyscallSucceeds());
+    ASSERT_THAT(
+        getsockopt(accepted_.get(), SOL_IP, IP_MTU_DISCOVER, &got, &length),
+        SyscallSucceeds());
+    EXPECT_EQ(got, IP_PMTUDISC_PROBE);
+  }
+}
+
 TEST_P(SimpleTcpSocketTest, GetSocketAcceptConnWithShutdown) {
   // TODO(b/171345701): Fix the TCP state for listening socket on shutdown.
   SKIP_IF(IsRunningOnGvisor());