5 files changed, 114 insertions, 12 deletions
diff --git a/pkg/tcpip/network/ipv4/ipv4.go b/pkg/tcpip/network/ipv4/ipv4.go
index 7e9f16c90..b1776e5ee 100644
--- a/pkg/tcpip/network/ipv4/ipv4.go
+++ b/pkg/tcpip/network/ipv4/ipv4.go
@@ -225,12 +225,10 @@ func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, mtu int,
 func (e *endpoint) addIPHeader(r *stack.Route, hdr *buffer.Prependable, payloadSize int, params stack.NetworkHeaderParams) header.IPv4 {
 	ip := header.IPv4(hdr.Prepend(header.IPv4MinimumSize))
 	length := uint16(hdr.UsedLength() + payloadSize)
-	id := uint32(0)
-	if length > header.IPv4MaximumHeaderSize+8 {
-		// Packets of 68 bytes or less are required by RFC 791 to not be
-		// fragmented, so we only assign ids to larger packets.
-		id = atomic.AddUint32(&e.protocol.ids[hashRoute(r, params.Protocol, e.protocol.hashIV)%buckets], 1)
-	}
+	// 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.
+	id := atomic.AddUint32(&e.protocol.ids[hashRoute(r, params.Protocol, e.protocol.hashIV)%buckets], 1)
 	ip.Encode(&header.IPv4Fields{
 		IHL:         header.IPv4MinimumSize,
 		TotalLength: length,
@@ -376,13 +374,12 @@ func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, pkt *stack.PacketBu
 
 	// Set the packet ID when zero.
 	if ip.ID() == 0 {
-		id := uint32(0)
-		if pkt.Data.Size() > header.IPv4MaximumHeaderSize+8 {
-			// Packets of 68 bytes or less are required by RFC 791 to not be
-			// fragmented, so we only assign ids to larger packets.
-			id = atomic.AddUint32(&e.protocol.ids[hashRoute(r, 0 /* protocol */, e.protocol.hashIV)%buckets], 1)
+		// RFC 6864 section 4.3 mandates uniqueness of ID values for
+		// non-atomic datagrams, so assign an ID to all such datagrams
+		// according to the definition given in RFC 6864 section 4.
+		if ip.Flags()&header.IPv4FlagDontFragment == 0 || ip.Flags()&header.IPv4FlagMoreFragments != 0 || ip.FragmentOffset() > 0 {
+			ip.SetID(uint16(atomic.AddUint32(&e.protocol.ids[hashRoute(r, 0 /* protocol */, e.protocol.hashIV)%buckets], 1)))
 		}
-		ip.SetID(uint16(id))
 	}
 
 	// Always set the checksum.
diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go
index 2be1c107a..2f9872dc6 100644
--- a/pkg/tcpip/tcpip.go
+++ b/pkg/tcpip/tcpip.go
@@ -673,6 +673,13 @@ const (
 	// TCP_MAXSEG option.
 	MaxSegOption
 
+	// MTUDiscoverOption is used to set/get the path MTU discovery setting.
+	//
+	// NOTE: Setting this option to any other value than PMTUDiscoveryDont
+	// is not supported and will fail as such, and getting this option will
+	// always return PMTUDiscoveryDont.
+	MTUDiscoverOption
+
 	// MulticastTTLOption is used by SetSockOptInt/GetSockOptInt to control
 	// the default TTL value for multicast messages. The default is 1.
 	MulticastTTLOption
@@ -714,6 +721,24 @@ const (
 	TCPWindowClampOption
 )
 
+const (
+	// PMTUDiscoveryWant is a setting of the MTUDiscoverOption to use
+	// per-route settings.
+	PMTUDiscoveryWant int = iota
+
+	// PMTUDiscoveryDont is a setting of the MTUDiscoverOption to disable
+	// path MTU discovery.
+	PMTUDiscoveryDont
+
+	// PMTUDiscoveryDo is a setting of the MTUDiscoverOption to always do
+	// path MTU discovery.
+	PMTUDiscoveryDo
+
+	// PMTUDiscoveryProbe is a setting of the MTUDiscoverOption to set DF
+	// but ignore path MTU.
+	PMTUDiscoveryProbe
+)
+
 // ErrorOption is used in GetSockOpt to specify that the last error reported by
 // the endpoint should be cleared and returned.
 type ErrorOption struct{}
diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go
index bd3ec5a8d..caac6ef57 100644
--- a/pkg/tcpip/transport/tcp/endpoint.go
+++ b/pkg/tcpip/transport/tcp/endpoint.go
@@ -1589,6 +1589,13 @@ func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error {
 		e.UnlockUser()
 		e.notifyProtocolGoroutine(notifyMSSChanged)
 
+	case tcpip.MTUDiscoverOption:
+		// Return not supported if attempting to set this option to
+		// anything other than path MTU discovery disabled.
+		if v != tcpip.PMTUDiscoveryDont {
+			return tcpip.ErrNotSupported
+		}
+
 	case tcpip.ReceiveBufferSizeOption:
 		// Make sure the receive buffer size is within the min and max
 		// allowed.
@@ -1896,6 +1903,11 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) {
 		v := header.TCPDefaultMSS
 		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
+
 	case tcpip.ReceiveQueueSizeOption:
 		return e.readyReceiveSize()
 
diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go
index 169adb16b..e67ec42b1 100644
--- a/pkg/tcpip/transport/tcp/tcp_test.go
+++ b/pkg/tcpip/transport/tcp/tcp_test.go
@@ -3095,6 +3095,63 @@ func TestMaxRTO(t *testing.T) {
 	}
 }
 
+// TestRetransmitIPv4IDUniqueness tests that the IPv4 Identification field is
+// unique on retransmits.
+func TestRetransmitIPv4IDUniqueness(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		size int
+	}{
+		{"1Byte", 1},
+		{"512Bytes", 512},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			c := context.New(t, defaultMTU)
+			defer c.Cleanup()
+
+			c.CreateConnected(789 /* iss */, 30000 /* rcvWnd */, -1 /* epRcvBuf */)
+
+			// Disabling PMTU discovery causes all packets sent from this socket to
+			// 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 {
+				t.Fatalf("disabling PMTU discovery via sockopt to force DF=0 failed: %s", err)
+			}
+
+			if _, _, err := c.EP.Write(tcpip.SlicePayload(buffer.NewView(tc.size)), tcpip.WriteOptions{}); err != nil {
+				t.Fatalf("Write failed: %s", err)
+			}
+			pkt := c.GetPacket()
+			checker.IPv4(t, pkt,
+				checker.FragmentFlags(0),
+				checker.TCP(
+					checker.DstPort(context.TestPort),
+					checker.TCPFlagsMatch(header.TCPFlagAck, ^uint8(header.TCPFlagPsh)),
+				),
+			)
+			idSet := map[uint16]struct{}{header.IPv4(pkt).ID(): struct{}{}}
+			// Expect two retransmitted packets, and that all packets received have
+			// unique IPv4 ID values.
+			for i := 0; i <= 2; i++ {
+				pkt := c.GetPacket()
+				checker.IPv4(t, pkt,
+					checker.FragmentFlags(0),
+					checker.TCP(
+						checker.DstPort(context.TestPort),
+						checker.TCPFlagsMatch(header.TCPFlagAck, ^uint8(header.TCPFlagPsh)),
+					),
+				)
+				id := header.IPv4(pkt).ID()
+				if _, exists := idSet[id]; exists {
+					t.Fatalf("duplicate IPv4 ID=%d found in retransmitted packet", id)
+				}
+				idSet[id] = struct{}{}
+			}
+		})
+	}
+}
+
 func TestFinImmediately(t *testing.T) {
 	c := context.New(t, defaultMTU)
 	defer c.Cleanup()
diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go
index cae29fbff..0584ec8dc 100644
--- a/pkg/tcpip/transport/udp/endpoint.go
+++ b/pkg/tcpip/transport/udp/endpoint.go
@@ -612,6 +612,13 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error {
 // SetSockOptInt implements tcpip.Endpoint.SetSockOptInt.
 func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error {
 	switch opt {
+	case tcpip.MTUDiscoverOption:
+		// Return not supported if the value is not disabling path
+		// MTU discovery.
+		if v != tcpip.PMTUDiscoveryDont {
+			return tcpip.ErrNotSupported
+		}
+
 	case tcpip.MulticastTTLOption:
 		e.mu.Lock()
 		e.multicastTTL = uint8(v)
@@ -906,6 +913,10 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) {
 		e.mu.RUnlock()
 		return v, nil
 
+	case tcpip.MTUDiscoverOption:
+		// The only supported setting is path MTU discovery disabled.
+		return tcpip.PMTUDiscoveryDont, nil
+
 	case tcpip.MulticastTTLOption:
 		e.mu.Lock()
 		v := int(e.multicastTTL)