make connect(2) fail when dest is unreachable

Previously, ICMP destination unreachable datagrams were ignored by TCP
endpoints. This caused connect to hang when an intermediate router
couldn't find a route to the host.

This manifested as a Kokoro error when Docker IPv6 was enabled. The Ruby
image test would try to install the sinatra gem and hang indefinitely
attempting to use an IPv6 address.

Fixes #3079.

13 files changed, 262 insertions, 14 deletions

diff --git a/pkg/tcpip/header/icmpv4.go b/pkg/tcpip/header/icmpv4.go
index 7908c5744..1a631b31a 100644
--- a/pkg/tcpip/header/icmpv4.go
+++ b/pkg/tcpip/header/icmpv4.go
@@ -72,6 +72,7 @@ const (
 // Values for ICMP code as defined in RFC 792.
 const (
 	ICMPv4TTLExceeded         = 0
+	ICMPv4HostUnreachable     = 1
 	ICMPv4PortUnreachable     = 3
 	ICMPv4FragmentationNeeded = 4
 )
diff --git a/pkg/tcpip/header/icmpv6.go b/pkg/tcpip/header/icmpv6.go
index c7ee2de57..a13b4b809 100644
--- a/pkg/tcpip/header/icmpv6.go
+++ b/pkg/tcpip/header/icmpv6.go
@@ -110,9 +110,16 @@ const (
 	ICMPv6RedirectMsg     ICMPv6Type = 137
 )
 
-// Values for ICMP code as defined in RFC 4443.
+// Values for ICMP destination unreachable code as defined in RFC 4443 section
+// 3.1.
 const (
-	ICMPv6PortUnreachable = 4
+	ICMPv6NetworkUnreachable = 0
+	ICMPv6Prohibited         = 1
+	ICMPv6BeyondScope        = 2
+	ICMPv6AddressUnreachable = 3
+	ICMPv6PortUnreachable    = 4
+	ICMPv6Policy             = 5
+	ICMPv6RejectRoute        = 6
 )
 
 // Type is the ICMP type field.
diff --git a/pkg/tcpip/network/ipv4/icmp.go b/pkg/tcpip/network/ipv4/icmp.go
index 1b67aa066..83e71cb8c 100644
--- a/pkg/tcpip/network/ipv4/icmp.go
+++ b/pkg/tcpip/network/ipv4/icmp.go
@@ -129,6 +129,9 @@ func (e *endpoint) handleICMP(r *stack.Route, pkt *stack.PacketBuffer) {
 
 		pkt.Data.TrimFront(header.ICMPv4MinimumSize)
 		switch h.Code() {
+		case header.ICMPv4HostUnreachable:
+			e.handleControl(stack.ControlNoRoute, 0, pkt)
+
 		case header.ICMPv4PortUnreachable:
 			e.handleControl(stack.ControlPortUnreachable, 0, pkt)
 
diff --git a/pkg/tcpip/network/ipv6/icmp.go b/pkg/tcpip/network/ipv6/icmp.go
index 3b79749b5..ff1cb53dd 100644
--- a/pkg/tcpip/network/ipv6/icmp.go
+++ b/pkg/tcpip/network/ipv6/icmp.go
@@ -128,6 +128,8 @@ func (e *endpoint) handleICMP(r *stack.Route, pkt *stack.PacketBuffer, hasFragme
 		}
 		pkt.Data.TrimFront(header.ICMPv6DstUnreachableMinimumSize)
 		switch header.ICMPv6(hdr).Code() {
+		case header.ICMPv6NetworkUnreachable:
+			e.handleControl(stack.ControlNetworkUnreachable, 0, pkt)
 		case header.ICMPv6PortUnreachable:
 			e.handleControl(stack.ControlPortUnreachable, 0, pkt)
 		}
diff --git a/pkg/tcpip/stack/registration.go b/pkg/tcpip/stack/registration.go
index cd4b7a449..9e1b2d25f 100644
--- a/pkg/tcpip/stack/registration.go
+++ b/pkg/tcpip/stack/registration.go
@@ -52,8 +52,11 @@ type TransportEndpointID struct {
 type ControlType int
 
 // The following are the allowed values for ControlType values.
+// TODO(http://gvisor.dev/issue/3210): Support time exceeded messages.
 const (
-	ControlPacketTooBig ControlType = iota
+	ControlNetworkUnreachable ControlType = iota
+	ControlNoRoute
+	ControlPacketTooBig
 	ControlPortUnreachable
 	ControlUnknown
 )
diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go
index 81b740115..1798510bc 100644
--- a/pkg/tcpip/transport/tcp/connect.go
+++ b/pkg/tcpip/transport/tcp/connect.go
@@ -490,6 +490,9 @@ func (h *handshake) resolveRoute() *tcpip.Error {
 				<-h.ep.undrain
 				h.ep.mu.Lock()
 			}
+			if n&notifyError != 0 {
+				return h.ep.takeLastError()
+			}
 		}
 
 		// Wait for notification.
@@ -616,6 +619,9 @@ func (h *handshake) execute() *tcpip.Error {
 				<-h.ep.undrain
 				h.ep.mu.Lock()
 			}
+			if n&notifyError != 0 {
+				return h.ep.takeLastError()
+			}
 
 		case wakerForNewSegment:
 			if err := h.processSegments(); err != nil {
diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go
index 83dc10ed0..0f7487963 100644
--- a/pkg/tcpip/transport/tcp/endpoint.go
+++ b/pkg/tcpip/transport/tcp/endpoint.go
@@ -1209,6 +1209,14 @@ func (e *endpoint) SetOwner(owner tcpip.PacketOwner) {
 	e.owner = owner
 }
 
+func (e *endpoint) takeLastError() *tcpip.Error {
+	e.lastErrorMu.Lock()
+	defer e.lastErrorMu.Unlock()
+	err := e.lastError
+	e.lastError = nil
+	return err
+}
+
 // Read reads data from the endpoint.
 func (e *endpoint) Read(*tcpip.FullAddress) (buffer.View, tcpip.ControlMessages, *tcpip.Error) {
 	e.LockUser()
@@ -1956,11 +1964,7 @@ func (e *endpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) {
 func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error {
 	switch o := opt.(type) {
 	case tcpip.ErrorOption:
-		e.lastErrorMu.Lock()
-		err := e.lastError
-		e.lastError = nil
-		e.lastErrorMu.Unlock()
-		return err
+		return e.takeLastError()
 
 	case *tcpip.BindToDeviceOption:
 		e.LockUser()
@@ -2546,6 +2550,18 @@ func (e *endpoint) HandleControlPacket(id stack.TransportEndpointID, typ stack.C
 		e.sndBufMu.Unlock()
 
 		e.notifyProtocolGoroutine(notifyMTUChanged)
+
+	case stack.ControlNoRoute:
+		e.lastErrorMu.Lock()
+		e.lastError = tcpip.ErrNoRoute
+		e.lastErrorMu.Unlock()
+		e.notifyProtocolGoroutine(notifyError)
+
+	case stack.ControlNetworkUnreachable:
+		e.lastErrorMu.Lock()
+		e.lastError = tcpip.ErrNetworkUnreachable
+		e.lastErrorMu.Unlock()
+		e.notifyProtocolGoroutine(notifyError)
 	}
 }
 
diff --git a/pkg/test/dockerutil/exec.go b/pkg/test/dockerutil/exec.go
index 921d1da9e..4c739c9e9 100644
--- a/pkg/test/dockerutil/exec.go
+++ b/pkg/test/dockerutil/exec.go
@@ -87,7 +87,6 @@ func (c *Container) doExec(ctx context.Context, r ExecOpts, args []string) (Proc
 		execid:    resp.ID,
 		conn:      hijack,
 	}, nil
-
 }
 
 func (c *Container) execConfig(r ExecOpts, cmd []string) types.ExecConfig {
diff --git a/test/packetimpact/runner/packetimpact_test.go b/test/packetimpact/runner/packetimpact_test.go
index ff5f5c7f1..1a0221893 100644
--- a/test/packetimpact/runner/packetimpact_test.go
+++ b/test/packetimpact/runner/packetimpact_test.go
@@ -280,11 +280,13 @@ func TestOne(t *testing.T) {
 	}
 
 	// Because the Linux kernel receives the SYN-ACK but didn't send the SYN it
-	// will issue a RST. To prevent this IPtables can be used to filter out all
+	// will issue an RST. To prevent this IPtables can be used to filter out all
 	// incoming packets. The raw socket that packetimpact tests use will still see
 	// everything.
-	if logs, err := testbench.Exec(ctx, dockerutil.ExecOpts{}, "iptables", "-A", "INPUT", "-i", testNetDev, "-j", "DROP"); err != nil {
-		t.Fatalf("unable to Exec iptables on container %s: %s, logs from testbench:\n%s", testbench.Name, err, logs)
+	for _, bin := range []string{"iptables", "ip6tables"} {
+		if logs, err := testbench.Exec(ctx, dockerutil.ExecOpts{}, bin, "-A", "INPUT", "-i", testNetDev, "-p", "tcp", "-j", "DROP"); err != nil {
+			t.Fatalf("unable to Exec %s on container %s: %s, logs from testbench:\n%s", bin, testbench.Name, err, logs)
+		}
 	}
 
 	// FIXME(b/156449515): Some piece of the system has a race. The old
diff --git a/test/packetimpact/testbench/connections.go b/test/packetimpact/testbench/connections.go
index 5d9cec73e..87ce58c24 100644
--- a/test/packetimpact/testbench/connections.go
+++ b/test/packetimpact/testbench/connections.go
@@ -41,7 +41,8 @@ func portFromSockaddr(sa unix.Sockaddr) (uint16, error) {
 	return 0, fmt.Errorf("sockaddr type %T does not contain port", sa)
 }
 
-// pickPort makes a new socket and returns the socket FD and port. The domain should be AF_INET or AF_INET6. The caller must close the FD when done with
+// pickPort makes a new socket and returns the socket FD and port. The domain
+// should be AF_INET or AF_INET6. The caller must close the FD when done with
 // the port if there is no error.
 func pickPort(domain, typ int) (fd int, port uint16, err error) {
 	fd, err = unix.Socket(domain, typ, 0)
@@ -1061,3 +1062,58 @@ func (conn *UDPIPv6) Close() {
 func (conn *UDPIPv6) Drain() {
 	conn.sniffer.Drain()
 }
+
+// TCPIPv6 maintains the state for all the layers in a TCP/IPv6 connection.
+type TCPIPv6 Connection
+
+// NewTCPIPv6 creates a new TCPIPv6 connection with reasonable defaults.
+func NewTCPIPv6(t *testing.T, outgoingTCP, incomingTCP TCP) TCPIPv6 {
+	etherState, err := newEtherState(Ether{}, Ether{})
+	if err != nil {
+		t.Fatalf("can't make etherState: %s", err)
+	}
+	ipv6State, err := newIPv6State(IPv6{}, IPv6{})
+	if err != nil {
+		t.Fatalf("can't make ipv6State: %s", err)
+	}
+	tcpState, err := newTCPState(unix.AF_INET6, outgoingTCP, incomingTCP)
+	if err != nil {
+		t.Fatalf("can't make tcpState: %s", err)
+	}
+	injector, err := NewInjector(t)
+	if err != nil {
+		t.Fatalf("can't make injector: %s", err)
+	}
+	sniffer, err := NewSniffer(t)
+	if err != nil {
+		t.Fatalf("can't make sniffer: %s", err)
+	}
+
+	return TCPIPv6{
+		layerStates: []layerState{etherState, ipv6State, tcpState},
+		injector:    injector,
+		sniffer:     sniffer,
+		t:           t,
+	}
+}
+
+func (conn *TCPIPv6) SrcPort() uint16 {
+	state := conn.layerStates[2].(*tcpState)
+	return *state.out.SrcPort
+}
+
+// ExpectData is a convenient method that expects a Layer and the Layer after
+// it. If it doens't arrive in time, it returns nil.
+func (conn *TCPIPv6) ExpectData(tcp *TCP, payload *Payload, timeout time.Duration) (Layers, error) {
+	expected := make([]Layer, len(conn.layerStates))
+	expected[len(expected)-1] = tcp
+	if payload != nil {
+		expected = append(expected, payload)
+	}
+	return (*Connection)(conn).ExpectFrame(expected, timeout)
+}
+
+// Close frees associated resources held by the TCPIPv6 connection.
+func (conn *TCPIPv6) Close() {
+	(*Connection)(conn).Close()
+}
diff --git a/test/packetimpact/testbench/layers.go b/test/packetimpact/testbench/layers.go
index 645f6c1a9..24aa46cce 100644
--- a/test/packetimpact/testbench/layers.go
+++ b/test/packetimpact/testbench/layers.go
@@ -805,7 +805,11 @@ func (l *ICMPv6) ToBytes() ([]byte, error) {
 		// We need to search forward to find the IPv6 header.
 		for prev := l.Prev(); prev != nil; prev = prev.Prev() {
 			if ipv6, ok := prev.(*IPv6); ok {
-				h.SetChecksum(header.ICMPv6Checksum(h, *ipv6.SrcAddr, *ipv6.DstAddr, buffer.VectorisedView{}))
+				payload, err := payload(l)
+				if err != nil {
+					return nil, err
+				}
+				h.SetChecksum(header.ICMPv6Checksum(h, *ipv6.SrcAddr, *ipv6.DstAddr, payload))
 				break
 			}
 		}
diff --git a/test/packetimpact/tests/BUILD b/test/packetimpact/tests/BUILD
index 6a07889be..27905dcff 100644
--- a/test/packetimpact/tests/BUILD
+++ b/test/packetimpact/tests/BUILD
@@ -220,6 +220,16 @@ packetimpact_go_test(
 )
 
 packetimpact_go_test(
+    name = "tcp_network_unreachable",
+    srcs = ["tcp_network_unreachable_test.go"],
+    deps = [
+        "//pkg/tcpip/header",
+        "//test/packetimpact/testbench",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+packetimpact_go_test(
     name = "tcp_cork_mss",
     srcs = ["tcp_cork_mss_test.go"],
     deps = [
diff --git a/test/packetimpact/tests/tcp_network_unreachable_test.go b/test/packetimpact/tests/tcp_network_unreachable_test.go
new file mode 100644
index 000000000..868a08da8
--- /dev/null
+++ b/test/packetimpact/tests/tcp_network_unreachable_test.go
@@ -0,0 +1,139 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tcp_synsent_reset_test
+
+import (
+	"context"
+	"flag"
+	"net"
+	"syscall"
+	"testing"
+	"time"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/header"
+	"gvisor.dev/gvisor/test/packetimpact/testbench"
+)
+
+func init() {
+	testbench.RegisterFlags(flag.CommandLine)
+}
+
+// TestTCPSynSentUnreachable verifies that TCP connections fail immediately when
+// an ICMP destination unreachable message is sent in response to the inital
+// SYN.
+func TestTCPSynSentUnreachable(t *testing.T) {
+	// Create the DUT and connection.
+	dut := testbench.NewDUT(t)
+	defer dut.TearDown()
+	clientFD, clientPort := dut.CreateBoundSocket(unix.SOCK_STREAM|unix.SOCK_NONBLOCK, unix.IPPROTO_TCP, net.ParseIP(testbench.RemoteIPv4))
+	port := uint16(9001)
+	conn := testbench.NewTCPIPv4(t, testbench.TCP{SrcPort: &port, DstPort: &clientPort}, testbench.TCP{SrcPort: &clientPort, DstPort: &port})
+	defer conn.Close()
+
+	// Bring the DUT to SYN-SENT state with a non-blocking connect.
+	ctx, cancel := context.WithTimeout(context.Background(), testbench.RPCTimeout)
+	defer cancel()
+	sa := unix.SockaddrInet4{Port: int(port)}
+	copy(sa.Addr[:], net.IP(net.ParseIP(testbench.LocalIPv4)).To4())
+	if _, err := dut.ConnectWithErrno(ctx, clientFD, &sa); err != syscall.Errno(unix.EINPROGRESS) {
+		t.Errorf("expected connect to fail with EINPROGRESS, but got %v", err)
+	}
+
+	// Get the SYN.
+	tcpLayers, err := conn.ExpectData(&testbench.TCP{Flags: testbench.Uint8(header.TCPFlagSyn)}, nil, time.Second)
+	if err != nil {
+		t.Fatalf("expected SYN: %s", err)
+	}
+
+	// Send a host unreachable message.
+	rawConn := (*testbench.Connection)(&conn)
+	layers := rawConn.CreateFrame(nil)
+	layers = layers[:len(layers)-1]
+	const ipLayer = 1
+	const tcpLayer = ipLayer + 1
+	ip, ok := tcpLayers[ipLayer].(*testbench.IPv4)
+	if !ok {
+		t.Fatalf("expected %s to be IPv4", tcpLayers[ipLayer])
+	}
+	tcp, ok := tcpLayers[tcpLayer].(*testbench.TCP)
+	if !ok {
+		t.Fatalf("expected %s to be TCP", tcpLayers[tcpLayer])
+	}
+	var icmpv4 testbench.ICMPv4 = testbench.ICMPv4{Type: testbench.ICMPv4Type(header.ICMPv4DstUnreachable), Code: testbench.Uint8(header.ICMPv4HostUnreachable)}
+	layers = append(layers, &icmpv4, ip, tcp)
+	rawConn.SendFrameStateless(layers)
+
+	if _, err = dut.ConnectWithErrno(ctx, clientFD, &sa); err != syscall.Errno(unix.EHOSTUNREACH) {
+		t.Errorf("expected connect to fail with EHOSTUNREACH, but got %v", err)
+	}
+}
+
+// TestTCPSynSentUnreachable6 verifies that TCP connections fail immediately when
+// an ICMP destination unreachable message is sent in response to the inital
+// SYN.
+func TestTCPSynSentUnreachable6(t *testing.T) {
+	// Create the DUT and connection.
+	dut := testbench.NewDUT(t)
+	defer dut.TearDown()
+	clientFD, clientPort := dut.CreateBoundSocket(unix.SOCK_STREAM|unix.SOCK_NONBLOCK, unix.IPPROTO_TCP, net.ParseIP(testbench.RemoteIPv6))
+	conn := testbench.NewTCPIPv6(t, testbench.TCP{DstPort: &clientPort}, testbench.TCP{SrcPort: &clientPort})
+	defer conn.Close()
+
+	// Bring the DUT to SYN-SENT state with a non-blocking connect.
+	ctx, cancel := context.WithTimeout(context.Background(), testbench.RPCTimeout)
+	defer cancel()
+	sa := unix.SockaddrInet6{
+		Port:   int(conn.SrcPort()),
+		ZoneId: uint32(testbench.RemoteInterfaceID),
+	}
+	copy(sa.Addr[:], net.IP(net.ParseIP(testbench.LocalIPv6)).To16())
+	if _, err := dut.ConnectWithErrno(ctx, clientFD, &sa); err != syscall.Errno(unix.EINPROGRESS) {
+		t.Errorf("expected connect to fail with EINPROGRESS, but got %v", err)
+	}
+
+	// Get the SYN.
+	tcpLayers, err := conn.ExpectData(&testbench.TCP{Flags: testbench.Uint8(header.TCPFlagSyn)}, nil, time.Second)
+	if err != nil {
+		t.Fatalf("expected SYN: %s", err)
+	}
+
+	// Send a host unreachable message.
+	rawConn := (*testbench.Connection)(&conn)
+	layers := rawConn.CreateFrame(nil)
+	layers = layers[:len(layers)-1]
+	const ipLayer = 1
+	const tcpLayer = ipLayer + 1
+	ip, ok := tcpLayers[ipLayer].(*testbench.IPv6)
+	if !ok {
+		t.Fatalf("expected %s to be IPv6", tcpLayers[ipLayer])
+	}
+	tcp, ok := tcpLayers[tcpLayer].(*testbench.TCP)
+	if !ok {
+		t.Fatalf("expected %s to be TCP", tcpLayers[tcpLayer])
+	}
+	var icmpv6 testbench.ICMPv6 = testbench.ICMPv6{
+		Type: testbench.ICMPv6Type(header.ICMPv6DstUnreachable),
+		Code: testbench.Uint8(header.ICMPv6NetworkUnreachable),
+		// Per RFC 4443 3.1, the payload contains 4 zeroed bytes.
+		Payload: []byte{0, 0, 0, 0},
+	}
+	layers = append(layers, &icmpv6, ip, tcp)
+	rawConn.SendFrameStateless(layers)
+
+	if _, err = dut.ConnectWithErrno(ctx, clientFD, &sa); err != syscall.Errno(unix.ENETUNREACH) {
+		t.Errorf("expected connect to fail with ENETUNREACH, but got %v", err)
+	}
+}