diff options
| -rw-r--r-- | pkg/sentry/socket/netstack/netstack.go | 20 | ||||
| -rw-r--r-- | pkg/tcpip/adapters/gonet/gonet_test.go | 12 | ||||
| -rw-r--r-- | pkg/tcpip/stack/stack.go | 20 | ||||
| -rw-r--r-- | pkg/tcpip/stack/transport_demuxer.go | 33 | ||||
| -rw-r--r-- | pkg/tcpip/tcpip.go | 12 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/BUILD | 2 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/accept.go | 17 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/connect.go | 322 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/endpoint.go | 101 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/endpoint_state.go | 26 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/protocol.go | 43 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/rcv.go | 167 | ||||
| -rw-r--r-- | pkg/tcpip/transport/tcp/tcp_test.go | 622 | ||||
| -rw-r--r-- | test/syscalls/BUILD | 22 | ||||
| -rw-r--r-- | test/syscalls/linux/BUILD | 1 | ||||
| -rw-r--r-- | test/syscalls/linux/socket_inet_loopback.cc | 336 | ||||
| -rw-r--r-- | test/syscalls/linux/socket_ip_tcp_generic.cc | 93 |
17 files changed, 1736 insertions, 113 deletions
diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index 27c6692c4..d92399efd 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -1173,6 +1173,18 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (interfa copy(b, v) return b, nil + case linux.TCP_LINGER2: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + var v tcpip.TCPLingerTimeoutOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + return int32(time.Duration(v) / time.Second), nil + default: emitUnimplementedEventTCP(t, name) } @@ -1556,6 +1568,14 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * } return nil + case linux.TCP_LINGER2: + if len(optVal) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + + v := usermem.ByteOrder.Uint32(optVal) + return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPLingerTimeoutOption(time.Second * time.Duration(v)))) + case linux.TCP_REPAIR_OPTIONS: t.Kernel().EmitUnimplementedEvent(t) diff --git a/pkg/tcpip/adapters/gonet/gonet_test.go b/pkg/tcpip/adapters/gonet/gonet_test.go index 8ced960bb..ee077ae83 100644 --- a/pkg/tcpip/adapters/gonet/gonet_test.go +++ b/pkg/tcpip/adapters/gonet/gonet_test.go @@ -151,10 +151,8 @@ func TestCloseReader(t *testing.T) { buf := make([]byte, 256) n, err := c.Read(buf) - got, ok := err.(*net.OpError) - want := tcpip.ErrConnectionAborted - if n != 0 || !ok || got.Err.Error() != want.String() { - t.Errorf("c.Read() = (%d, %v), want (0, OpError(%v))", n, err, want) + if n != 0 || err != io.EOF { + t.Errorf("c.Read() = (%d, %v), want (0, EOF)", n, err) } }() sender, err := connect(s, addr) @@ -203,10 +201,8 @@ func TestCloseReaderWithForwarder(t *testing.T) { buf := make([]byte, 256) n, e := c.Read(buf) - got, ok := e.(*net.OpError) - want := tcpip.ErrConnectionAborted - if n != 0 || !ok || got.Err.Error() != want.String() { - t.Errorf("c.Read() = (%d, %v), want (0, OpError(%v))", n, e, want) + if n != 0 || e != io.EOF { + t.Errorf("c.Read() = (%d, %v), want (0, EOF)", n, e) } }) s.SetTransportProtocolHandler(tcp.ProtocolNumber, fwd.HandlePacket) diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index 99809df75..2f8d8e822 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -402,11 +402,11 @@ type Stack struct { // by the stack. icmpRateLimiter *ICMPRateLimiter - // portSeed is a one-time random value initialized at stack startup + // seed is a one-time random value initialized at stack startup // and is used to seed the TCP port picking on active connections // // TODO(gvisor.dev/issue/940): S/R this field. - portSeed uint32 + seed uint32 // ndpConfigs is the default NDP configurations used by interfaces. ndpConfigs NDPConfigurations @@ -544,7 +544,7 @@ func New(opts Options) *Stack { stats: opts.Stats.FillIn(), handleLocal: opts.HandleLocal, icmpRateLimiter: NewICMPRateLimiter(), - portSeed: generateRandUint32(), + seed: generateRandUint32(), ndpConfigs: opts.NDPConfigs, autoGenIPv6LinkLocal: opts.AutoGenIPv6LinkLocal, uniqueIDGenerator: opts.UniqueID, @@ -1186,6 +1186,12 @@ func (s *Stack) CompleteTransportEndpointCleanup(ep TransportEndpoint) { s.mu.Unlock() } +// FindTransportEndpoint finds an endpoint that most closely matches the provided +// id. If no endpoint is found it returns nil. +func (s *Stack) FindTransportEndpoint(netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, id TransportEndpointID, r *Route) TransportEndpoint { + return s.demux.findTransportEndpoint(netProto, transProto, id, r) +} + // RegisterRawTransportEndpoint registers the given endpoint with the stack // transport dispatcher. Received packets that match the provided transport // protocol will be delivered to the given endpoint. @@ -1573,12 +1579,12 @@ func (s *Stack) HandleNDPRA(id tcpip.NICID, ip tcpip.Address, ra header.NDPRoute return nil } -// PortSeed returns a 32 bit value that can be used as a seed value for port -// picking. +// Seed returns a 32 bit value that can be used as a seed value for port +// picking, ISN generation etc. // // NOTE: The seed is generated once during stack initialization only. -func (s *Stack) PortSeed() uint32 { - return s.portSeed +func (s *Stack) Seed() uint32 { + return s.seed } func generateRandUint32() uint32 { diff --git a/pkg/tcpip/stack/transport_demuxer.go b/pkg/tcpip/stack/transport_demuxer.go index 594570216..cb805522b 100644 --- a/pkg/tcpip/stack/transport_demuxer.go +++ b/pkg/tcpip/stack/transport_demuxer.go @@ -103,7 +103,6 @@ func (epsByNic *endpointsByNic) handlePacket(r *Route, id TransportEndpointID, p epsByNic.mu.RUnlock() // Don't use defer for performance reasons. return } - // multiPortEndpoints are guaranteed to have at least one element. selectEndpoint(id, mpep, epsByNic.seed).HandlePacket(r, id, pkt) epsByNic.mu.RUnlock() // Don't use defer for performance reasons. @@ -507,10 +506,40 @@ func (d *transportDemuxer) findAllEndpointsLocked(eps *transportEndpoints, id Tr if ep, ok := eps.endpoints[nid]; ok { matchedEPs = append(matchedEPs, ep) } - return matchedEPs } +// findTransportEndpoint find a single endpoint that most closely matches the provided id. +func (d *transportDemuxer) findTransportEndpoint(netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, id TransportEndpointID, r *Route) TransportEndpoint { + eps, ok := d.protocol[protocolIDs{netProto, transProto}] + if !ok { + return nil + } + // Try to find the endpoint. + eps.mu.RLock() + epsByNic := d.findEndpointLocked(eps, id) + // Fail if we didn't find one. + if epsByNic == nil { + eps.mu.RUnlock() + return nil + } + + epsByNic.mu.RLock() + eps.mu.RUnlock() + + mpep, ok := epsByNic.endpoints[r.ref.nic.ID()] + if !ok { + if mpep, ok = epsByNic.endpoints[0]; !ok { + epsByNic.mu.RUnlock() // Don't use defer for performance reasons. + return nil + } + } + + ep := selectEndpoint(id, mpep, epsByNic.seed) + epsByNic.mu.RUnlock() + return ep +} + // findEndpointLocked returns the endpoint that most closely matches the given // id. func (d *transportDemuxer) findEndpointLocked(eps *transportEndpoints, id TransportEndpointID) *endpointsByNic { diff --git a/pkg/tcpip/tcpip.go b/pkg/tcpip/tcpip.go index 3edb513d4..bd5eb89ca 100644 --- a/pkg/tcpip/tcpip.go +++ b/pkg/tcpip/tcpip.go @@ -586,6 +586,16 @@ type MaxSegOption int // A zero value indicates the default. type TTLOption uint8 +// TCPLingerTimeoutOption is used by SetSockOpt/GetSockOpt to set/get the +// maximum duration for which a socket lingers in the TCP_FIN_WAIT_2 state +// before being marked closed. +type TCPLingerTimeoutOption time.Duration + +// TCPTimeWaitTimeoutOption is used by SetSockOpt/GetSockOpt to set/get the +// maximum duration for which a socket lingers in the TIME_WAIT state +// before being marked closed. +type TCPTimeWaitTimeoutOption time.Duration + // MulticastTTLOption is used by SetSockOpt/GetSockOpt to control the default // TTL value for multicast messages. The default is 1. type MulticastTTLOption uint8 @@ -1329,8 +1339,8 @@ var ( // GetDanglingEndpoints returns all dangling endpoints. func GetDanglingEndpoints() []Endpoint { - es := make([]Endpoint, 0, len(danglingEndpoints)) danglingEndpointsMu.Lock() + es := make([]Endpoint, 0, len(danglingEndpoints)) for e := range danglingEndpoints { es = append(es, e) } diff --git a/pkg/tcpip/transport/tcp/BUILD b/pkg/tcpip/transport/tcp/BUILD index f1dbc6f91..3f47b328d 100644 --- a/pkg/tcpip/transport/tcp/BUILD +++ b/pkg/tcpip/transport/tcp/BUILD @@ -71,7 +71,7 @@ filegroup( go_test( name = "tcp_test", - size = "small", + size = "medium", srcs = [ "dual_stack_test.go", "sack_scoreboard_test.go", diff --git a/pkg/tcpip/transport/tcp/accept.go b/pkg/tcpip/transport/tcp/accept.go index cb0e13ebc..0e8e0a2b4 100644 --- a/pkg/tcpip/transport/tcp/accept.go +++ b/pkg/tcpip/transport/tcp/accept.go @@ -269,8 +269,8 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *header.TCPSynOptions) (*endpoint, *tcpip.Error) { // Create new endpoint. irs := s.sequenceNumber - cookie := l.createCookie(s.id, irs, encodeMSS(opts.MSS)) - ep, err := l.createConnectingEndpoint(s, cookie, irs, opts) + isn := generateSecureISN(s.id, l.stack.Seed()) + ep, err := l.createConnectingEndpoint(s, isn, irs, opts) if err != nil { return nil, err } @@ -289,7 +289,7 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head // Perform the 3-way handshake. h := newHandshake(ep, seqnum.Size(ep.initialReceiveWindow())) - h.resetToSynRcvd(cookie, irs, opts) + h.resetToSynRcvd(isn, irs, opts) if err := h.execute(); err != nil { ep.Close() if l.listenEP != nil { @@ -361,6 +361,7 @@ func (e *endpoint) handleSynSegment(ctx *listenContext, s *segment, opts *header defer decSynRcvdCount() defer e.decSynRcvdCount() defer s.decRef() + n, err := ctx.createEndpointAndPerformHandshake(s, opts) if err != nil { e.stack.Stats().TCP.FailedConnectionAttempts.Increment() @@ -368,6 +369,11 @@ func (e *endpoint) handleSynSegment(ctx *listenContext, s *segment, opts *header return } ctx.removePendingEndpoint(n) + // Start the protocol goroutine. + wq := &waiter.Queue{} + n.startAcceptedLoop(wq) + e.stack.Stats().TCP.PassiveConnectionOpenings.Increment() + e.deliverAccepted(n) } @@ -543,6 +549,11 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { // number of goroutines as we do check before // entering here that there was at least some // space available in the backlog. + + // Start the protocol goroutine. + wq := &waiter.Queue{} + n.startAcceptedLoop(wq) + e.stack.Stats().TCP.PassiveConnectionOpenings.Increment() go e.deliverAccepted(n) } } diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go index ca982c451..a114c06c1 100644 --- a/pkg/tcpip/transport/tcp/connect.go +++ b/pkg/tcpip/transport/tcp/connect.go @@ -15,6 +15,7 @@ package tcp import ( + "encoding/binary" "sync" "time" @@ -22,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/sleep" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" + "gvisor.dev/gvisor/pkg/tcpip/hash/jenkins" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -139,7 +141,32 @@ func (h *handshake) resetState() { h.flags = header.TCPFlagSyn h.ackNum = 0 h.mss = 0 - h.iss = seqnum.Value(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24) + h.iss = generateSecureISN(h.ep.ID, h.ep.stack.Seed()) +} + +// generateSecureISN generates a secure Initial Sequence number based on the +// recommendation here https://tools.ietf.org/html/rfc6528#page-3. +func generateSecureISN(id stack.TransportEndpointID, seed uint32) seqnum.Value { + isnHasher := jenkins.Sum32(seed) + isnHasher.Write([]byte(id.LocalAddress)) + isnHasher.Write([]byte(id.RemoteAddress)) + portBuf := make([]byte, 2) + binary.LittleEndian.PutUint16(portBuf, id.LocalPort) + isnHasher.Write(portBuf) + binary.LittleEndian.PutUint16(portBuf, id.RemotePort) + isnHasher.Write(portBuf) + // The time period here is 64ns. This is similar to what linux uses + // generate a sequence number that overlaps less than one + // time per MSL (2 minutes). + // + // A 64ns clock ticks 10^9/64 = 15625000) times in a second. + // To wrap the whole 32 bit space would require + // 2^32/1562500 ~ 274 seconds. + // + // Which sort of guarantees that we won't reuse the ISN for a new + // connection for the same tuple for at least 274s. + isn := isnHasher.Sum32() + uint32(time.Now().UnixNano()>>6) + return seqnum.Value(isn) } // effectiveRcvWndScale returns the effective receive window scale to be used. @@ -809,7 +836,19 @@ func (e *endpoint) resetConnectionLocked(err *tcpip.Error) { e.state = StateError e.HardError = err if err != tcpip.ErrConnectionReset { - e.sendRaw(buffer.VectorisedView{}, header.TCPFlagAck|header.TCPFlagRst, e.snd.sndUna, e.rcv.rcvNxt, 0) + // The exact sequence number to be used for the RST is the same as the + // one used by Linux. We need to handle the case of window being shrunk + // which can cause sndNxt to be outside the acceptable window on the + // receiver. + // + // See: https://www.snellman.net/blog/archive/2016-02-01-tcp-rst/ for more + // information. + sndWndEnd := e.snd.sndUna.Add(e.snd.sndWnd) + resetSeqNum := sndWndEnd + if !sndWndEnd.LessThan(e.snd.sndNxt) || e.snd.sndNxt.Size(sndWndEnd) < (1<<e.snd.sndWndScale) { + resetSeqNum = e.snd.sndNxt + } + e.sendRaw(buffer.VectorisedView{}, header.TCPFlagAck|header.TCPFlagRst, resetSeqNum, e.rcv.rcvNxt, 0) } } @@ -823,6 +862,51 @@ func (e *endpoint) completeWorkerLocked() { } } +func (e *endpoint) handleReset(s *segment) (ok bool, err *tcpip.Error) { + if e.rcv.acceptable(s.sequenceNumber, 0) { + // RFC 793, page 37 states that "in all states + // except SYN-SENT, all reset (RST) segments are + // validated by checking their SEQ-fields." So + // we only process it if it's acceptable. + s.decRef() + e.mu.Lock() + switch e.state { + // In case of a RST in CLOSE-WAIT linux moves + // the socket to closed state with an error set + // to indicate EPIPE. + // + // Technically this seems to be at odds w/ RFC. + // As per https://tools.ietf.org/html/rfc793#section-2.7 + // page 69 the behavior for a segment arriving + // w/ RST bit set in CLOSE-WAIT is inlined below. + // + // ESTABLISHED + // FIN-WAIT-1 + // FIN-WAIT-2 + // CLOSE-WAIT + + // If the RST bit is set then, any outstanding RECEIVEs and + // SEND should receive "reset" responses. All segment queues + // should be flushed. Users should also receive an unsolicited + // general "connection reset" signal. Enter the CLOSED state, + // delete the TCB, and return. + case StateCloseWait: + e.state = StateClose + e.HardError = tcpip.ErrAborted + // We need to set this explicitly here because otherwise + // the port registrations will not be released till the + // endpoint is actively closed by the application. + e.workerCleanup = true + e.mu.Unlock() + return false, nil + default: + e.mu.Unlock() + return false, tcpip.ErrConnectionReset + } + } + return true, nil +} + // handleSegments pulls segments from the queue and processes them. It returns // no error if the protocol loop should continue, an error otherwise. func (e *endpoint) handleSegments() *tcpip.Error { @@ -840,14 +924,34 @@ func (e *endpoint) handleSegments() *tcpip.Error { } if s.flagIsSet(header.TCPFlagRst) { - if e.rcv.acceptable(s.sequenceNumber, 0) { - // RFC 793, page 37 states that "in all states - // except SYN-SENT, all reset (RST) segments are - // validated by checking their SEQ-fields." So - // we only process it if it's acceptable. - s.decRef() - return tcpip.ErrConnectionReset + if ok, err := e.handleReset(s); !ok { + return err } + } else if s.flagIsSet(header.TCPFlagSyn) { + // See: https://tools.ietf.org/html/rfc5961#section-4.1 + // 1) If the SYN bit is set, irrespective of the sequence number, TCP + // MUST send an ACK (also referred to as challenge ACK) to the remote + // peer: + // + // <SEQ=SND.NXT><ACK=RCV.NXT><CTL=ACK> + // + // After sending the acknowledgment, TCP MUST drop the unacceptable + // segment and stop processing further. + // + // By sending an ACK, the remote peer is challenged to confirm the loss + // of the previous connection and the request to start a new connection. + // A legitimate peer, after restart, would not have a TCB in the + // synchronized state. Thus, when the ACK arrives, the peer should send + // a RST segment back with the sequence number derived from the ACK + // field that caused the RST. + + // This RST will confirm that the remote peer has indeed closed the + // previous connection. Upon receipt of a valid RST, the local TCP + // endpoint MUST terminate its connection. The local TCP endpoint + // should then rely on SYN retransmission from the remote end to + // re-establish the connection. + + e.snd.sendAck() } else if s.flagIsSet(header.TCPFlagAck) { // Patch the window size in the segment according to the // send window scale. @@ -856,7 +960,15 @@ func (e *endpoint) handleSegments() *tcpip.Error { // RFC 793, page 41 states that "once in the ESTABLISHED // state all segments must carry current acknowledgment // information." - e.rcv.handleRcvdSegment(s) + drop, err := e.rcv.handleRcvdSegment(s) + if err != nil { + s.decRef() + return err + } + if drop { + s.decRef() + continue + } e.snd.handleRcvdSegment(s) } s.decRef() @@ -955,7 +1067,6 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { } e.mu.Unlock() - // When the protocol loop exits we should wake up our waiters. e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } @@ -1001,6 +1112,10 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { // RTT itself. e.rcvAutoParams.prevCopied = initialRcvWnd e.rcvListMu.Unlock() + e.stack.Stats().TCP.CurrentEstablished.Increment() + e.mu.Lock() + e.state = StateEstablished + e.mu.Unlock() } e.keepalive.timer.init(&e.keepalive.waker) @@ -1008,10 +1123,6 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { // Tell waiters that the endpoint is connected and writable. e.mu.Lock() - if e.state != StateEstablished { - e.stack.Stats().TCP.CurrentEstablished.Increment() - e.state = StateEstablished - } drained := e.drainDone != nil e.mu.Unlock() if drained { @@ -1042,7 +1153,13 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { { w: &closeWaker, f: func() *tcpip.Error { - return tcpip.ErrConnectionAborted + // This means the socket is being closed due + // to the TCP_FIN_WAIT2 timeout was hit. Just + // mark the socket as closed. + e.mu.Lock() + e.state = StateClose + e.mu.Unlock() + return nil }, }, { @@ -1085,17 +1202,18 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { e.resetConnectionLocked(tcpip.ErrConnectionAborted) e.mu.Unlock() } + if n¬ifyClose != 0 && closeTimer == nil { - // Reset the connection 3 seconds after - // the endpoint has been closed. - // - // The timer could fire in background - // when the endpoint is drained. That's - // OK as the loop here will not honor - // the firing until the undrain arrives. - closeTimer = time.AfterFunc(3*time.Second, func() { - closeWaker.Assert() - }) + e.mu.Lock() + if e.state == StateFinWait2 && e.closed { + // The socket has been closed and we are in FIN_WAIT2 + // so start the FIN_WAIT2 timer. + closeTimer = time.AfterFunc(e.tcpLingerTimeout, func() { + closeWaker.Assert() + }) + e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) + } + e.mu.Unlock() } if n¬ifyKeepaliveChanged != 0 { @@ -1117,6 +1235,12 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { } } + if n¬ifyTickleWorker != 0 { + // Just a tickle notification. No need to do + // anything. + return nil + } + return nil }, }, @@ -1143,15 +1267,16 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { } e.rcvListMu.Unlock() - e.mu.RLock() + e.mu.Lock() if e.workerCleanup { e.notifyProtocolGoroutine(notifyClose) } - e.mu.RUnlock() // Main loop. Handle segments until both send and receive ends of the // connection have completed. - for !e.rcv.closed || !e.snd.closed || e.snd.sndUna != e.snd.sndNxtList { + + for e.state != StateTimeWait && e.state != StateClose && e.state != StateError { + e.mu.Unlock() e.workMu.Unlock() v, _ := s.Fetch(true) e.workMu.Lock() @@ -1167,6 +1292,23 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { return nil } + e.mu.Lock() + } + + state := e.state + e.mu.Unlock() + var reuseTW func() + if state == StateTimeWait { + // Disable close timer as we now entering real TIME_WAIT. + if closeTimer != nil { + closeTimer.Stop() + } + // Mark the current sleeper done so as to free all associated + // wakers. + s.Done() + // Wake up any waiters before we enter TIME_WAIT. + e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) + reuseTW = e.doTimeWait() } // Mark endpoint as closed. @@ -1176,8 +1318,130 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error { e.stack.Stats().TCP.CurrentEstablished.Decrement() e.state = StateClose } + // Lock released below. epilogue() + // A new SYN was received during TIME_WAIT and we need to abort + // the timewait and redirect the segment to the listener queue + if reuseTW != nil { + reuseTW() + } + return nil } + +// handleTimeWaitSegments processes segments received during TIME_WAIT +// state. +func (e *endpoint) handleTimeWaitSegments() (extendTimeWait bool, reuseTW func()) { + checkRequeue := true + for i := 0; i < maxSegmentsPerWake; i++ { + s := e.segmentQueue.dequeue() + if s == nil { + checkRequeue = false + break + } + extTW, newSyn := e.rcv.handleTimeWaitSegment(s) + if newSyn { + info := e.EndpointInfo.TransportEndpointInfo + newID := info.ID + newID.RemoteAddress = "" + newID.RemotePort = 0 + netProtos := []tcpip.NetworkProtocolNumber{info.NetProto} + // If the local address is an IPv4 address then also + // look for IPv6 dual stack endpoints that might be + // listening on the local address. + if newID.LocalAddress.To4() != "" { + netProtos = []tcpip.NetworkProtocolNumber{header.IPv4ProtocolNumber, header.IPv6ProtocolNumber} + } + for _, netProto := range netProtos { + if listenEP := e.stack.FindTransportEndpoint(netProto, info.TransProto, newID, &s.route); listenEP != nil { + tcpEP := listenEP.(*endpoint) + if EndpointState(tcpEP.State()) == StateListen { + reuseTW = func() { + tcpEP.enqueueSegment(s) + } + // We explicitly do not decRef + // the segment as it's still + // valid and being reflected to + // a listening endpoint. + return false, reuseTW + } + } + } + } + if extTW { + extendTimeWait = true + } + s.decRef() + } + if checkRequeue && !e.segmentQueue.empty() { + e.newSegmentWaker.Assert() + } + return extendTimeWait, nil +} + +// doTimeWait is responsible for handling the TCP behaviour once a socket +// enters the TIME_WAIT state. Optionally it can return a closure that +// should be executed after releasing the endpoint registrations. This is +// done in cases where a new SYN is received during TIME_WAIT that carries +// a sequence number larger than one see on the connection. +func (e *endpoint) doTimeWait() (twReuse func()) { + // Trigger a 2 * MSL time wait state. During this period + // we will drop all incoming segments. + // NOTE: On Linux this is not configurable and is fixed at 60 seconds. + timeWaitDuration := DefaultTCPTimeWaitTimeout + + // Get the stack wide configuration. + var tcpTW tcpip.TCPTimeWaitTimeoutOption + if err := e.stack.TransportProtocolOption(ProtocolNumber, &tcpTW); err == nil { + timeWaitDuration = time.Duration(tcpTW) + } + + const newSegment = 1 + const notification = 2 + const timeWaitDone = 3 + + s := sleep.Sleeper{} + s.AddWaker(&e.newSegmentWaker, newSegment) + s.AddWaker(&e.notificationWaker, notification) + + var timeWaitWaker sleep.Waker + s.AddWaker(&timeWaitWaker, timeWaitDone) + timeWaitTimer := time.AfterFunc(timeWaitDuration, timeWaitWaker.Assert) + defer timeWaitTimer.Stop() + + for { + e.workMu.Unlock() + v, _ := s.Fetch(true) + e.workMu.Lock() + switch v { + case newSegment: + extendTimeWait, reuseTW := e.handleTimeWaitSegments() + if reuseTW != nil { + return reuseTW + } + if extendTimeWait { + timeWaitTimer.Reset(timeWaitDuration) + } + case notification: + n := e.fetchNotifications() + if n¬ifyClose != 0 { + return nil + } + if n¬ifyDrain != 0 { + for !e.segmentQueue.empty() { + // Ignore extending TIME_WAIT during a + // save. For sockets in TIME_WAIT we just + // terminate the TIME_WAIT early. + e.handleTimeWaitSegments() + } + close(e.drainDone) + <-e.undrain + return nil + } + case timeWaitDone: + return nil + } + } +} diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index 79fec6b77..04c92c04c 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -121,6 +121,11 @@ const ( notifyReset notifyKeepaliveChanged notifyMSSChanged + // notifyTickleWorker is used to tickle the protocol main loop during a + // restore after we update the endpoint state to the correct one. This + // ensures the loop terminates if the final state of the endpoint is + // say TIME_WAIT. + notifyTickleWorker ) // SACKInfo holds TCP SACK related information for a given endpoint. @@ -320,6 +325,11 @@ type endpoint struct { state EndpointState `state:".(EndpointState)"` + // origEndpointState is only used during a restore phase to save the + // endpoint state at restore time as the socket is moved to it's correct + // state. + origEndpointState EndpointState `state:"nosave"` + isPortReserved bool `state:"manual"` isRegistered bool boundNICID tcpip.NICID `state:"manual"` @@ -503,6 +513,16 @@ type endpoint struct { // TODO(b/142022063): Add ability to save and restore per endpoint stats. stats Stats `state:"nosave"` + + // tcpLingerTimeout is the maximum amount of a time a socket + // a socket stays in TIME_WAIT state before being marked + // closed. + tcpLingerTimeout time.Duration + + // closed indicates that the user has called closed on the + // endpoint and at this point the endpoint is only around + // to complete the TCP shutdown. + closed bool } // UniqueID implements stack.TransportEndpoint.UniqueID. @@ -599,6 +619,11 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue e.SetSockOptInt(tcpip.DelayOption, 1) } + var tcpLT tcpip.TCPLingerTimeoutOption + if err := s.TransportProtocolOption(ProtocolNumber, &tcpLT); err == nil { + e.tcpLingerTimeout = time.Duration(tcpLT) + } + if p := s.GetTCPProbe(); p != nil { e.probe = p } @@ -686,6 +711,13 @@ func (e *endpoint) notifyProtocolGoroutine(n uint32) { // with it. It must be called only once and with no other concurrent calls to // the endpoint. func (e *endpoint) Close() { + e.mu.Lock() + closed := e.closed + e.mu.Unlock() + if closed { + return + } + // Issue a shutdown so that the peer knows we won't send any more data // if we're connected, or stop accepting if we're listening. e.Shutdown(tcpip.ShutdownWrite | tcpip.ShutdownRead) @@ -706,6 +738,8 @@ func (e *endpoint) Close() { e.isPortReserved = false } + // Mark endpoint as closed. + e.closed = true // Either perform the local cleanup or kick the worker to make sure it // knows it needs to cleanup. tcpip.AddDanglingEndpoint(e) @@ -731,9 +765,7 @@ func (e *endpoint) closePendingAcceptableConnectionsLocked() { go func() { defer close(done) for n := range e.acceptedChan { - n.mu.Lock() - n.resetConnectionLocked(tcpip.ErrConnectionAborted) - n.mu.Unlock() + n.notifyProtocolGoroutine(notifyReset) n.Close() } }() @@ -1349,6 +1381,28 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error { e.mu.Unlock() return nil + case tcpip.TCPLingerTimeoutOption: + e.mu.Lock() + if v < 0 { + // Same as effectively disabling TCPLinger timeout. + v = 0 + } + var stkTCPLingerTimeout tcpip.TCPLingerTimeoutOption + if err := e.stack.TransportProtocolOption(header.TCPProtocolNumber, &stkTCPLingerTimeout); err != nil { + // We were unable to retrieve a stack config, just use + // the DefaultTCPLingerTimeout. + if v > tcpip.TCPLingerTimeoutOption(DefaultTCPLingerTimeout) { + stkTCPLingerTimeout = tcpip.TCPLingerTimeoutOption(DefaultTCPLingerTimeout) + } + } + // Cap it to the stack wide TCPLinger timeout. + if v > stkTCPLingerTimeout { + v = stkTCPLingerTimeout + } + e.tcpLingerTimeout = time.Duration(v) + e.mu.Unlock() + return nil + default: return nil } @@ -1562,6 +1616,12 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error { e.mu.RUnlock() return nil + case *tcpip.TCPLingerTimeoutOption: + e.mu.Lock() + *o = tcpip.TCPLingerTimeoutOption(e.tcpLingerTimeout) + e.mu.Unlock() + return nil + default: return tcpip.ErrUnknownProtocolOption } @@ -1696,7 +1756,7 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc // src IP to ensure that for a given tuple (srcIP, destIP, // destPort) the offset used as a starting point is the same to // ensure that we can cycle through the port space effectively. - h := jenkins.Sum32(e.stack.PortSeed()) + h := jenkins.Sum32(e.stack.Seed()) h.Write([]byte(e.ID.LocalAddress)) h.Write([]byte(e.ID.RemoteAddress)) portBuf := make([]byte, 2) @@ -1782,9 +1842,8 @@ func (*endpoint) ConnectEndpoint(tcpip.Endpoint) *tcpip.Error { // peer. func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { e.mu.Lock() - defer e.mu.Unlock() e.shutdownFlags |= flags - + finQueued := false switch { case e.state.connected(): // Close for read. @@ -1799,6 +1858,7 @@ func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { // the connection with a RST. if (e.shutdownFlags&tcpip.ShutdownWrite) != 0 && rcvBufUsed > 0 { e.notifyProtocolGoroutine(notifyReset) + e.mu.Unlock() return nil } } @@ -1817,14 +1877,11 @@ func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { s := newSegmentFromView(&e.route, e.ID, nil) e.sndQueue.PushBack(s) e.sndBufInQueue++ - + finQueued = true // Mark endpoint as closed. e.sndClosed = true e.sndBufMu.Unlock() - - // Tell protocol goroutine to close. - e.sndCloseWaker.Assert() } case e.state == StateListen: @@ -1832,11 +1889,20 @@ func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error { if flags&tcpip.ShutdownRead != 0 { e.notifyProtocolGoroutine(notifyClose) } - default: + e.mu.Unlock() return tcpip.ErrNotConnected } - + e.mu.Unlock() + if finQueued { + if e.workMu.TryLock() { + e.handleClose() + e.workMu.Unlock() + } else { + // Tell protocol goroutine to close. + e.sndCloseWaker.Assert() + } + } return nil } @@ -1928,12 +1994,7 @@ func (e *endpoint) Accept() (tcpip.Endpoint, *waiter.Queue, *tcpip.Error) { return nil, nil, tcpip.ErrWouldBlock } - // Start the protocol goroutine. - wq := &waiter.Queue{} - n.startAcceptedLoop(wq) - e.stack.Stats().TCP.PassiveConnectionOpenings.Increment() - - return n, wq, nil + return n, n.waiterQueue, nil } // Bind binds the endpoint to a specific local port and optionally address. @@ -2058,6 +2119,10 @@ func (e *endpoint) HandlePacket(r *stack.Route, id stack.TransportEndpointID, pk e.stack.Stats().TCP.ResetsReceived.Increment() } + e.enqueueSegment(s) +} + +func (e *endpoint) enqueueSegment(s *segment) { // Send packet to worker goroutine. if e.segmentQueue.enqueue(s) { e.newSegmentWaker.Assert() diff --git a/pkg/tcpip/transport/tcp/endpoint_state.go b/pkg/tcpip/transport/tcp/endpoint_state.go index 19f003b6b..7aa4c3f0e 100644 --- a/pkg/tcpip/transport/tcp/endpoint_state.go +++ b/pkg/tcpip/transport/tcp/endpoint_state.go @@ -78,7 +78,7 @@ func (e *endpoint) beforeSave() { } fallthrough case StateError, StateClose: - for e.state == StateError && e.workerRunning { + for (e.state == StateError || e.state == StateClose) && e.workerRunning { e.mu.Unlock() time.Sleep(100 * time.Millisecond) e.mu.Lock() @@ -165,6 +165,12 @@ func (e *endpoint) loadState(state EndpointState) { // afterLoad is invoked by stateify. func (e *endpoint) afterLoad() { + // Freeze segment queue before registering to prevent any segments + // from being delivered while it is being restored. + e.origEndpointState = e.state + // Restore the endpoint to InitialState as it will be moved to + // its origEndpointState during Resume. + e.state = StateInitial stack.StackFromEnv.RegisterRestoredEndpoint(e) } @@ -173,8 +179,8 @@ func (e *endpoint) Resume(s *stack.Stack) { e.stack = s e.segmentQueue.setLimit(MaxUnprocessedSegments) e.workMu.Init() + state := e.origEndpointState - state := e.state switch state { case StateInitial, StateBound, StateListen, StateConnecting, StateEstablished: var ss SendBufferSizeOption @@ -189,7 +195,6 @@ func (e *endpoint) Resume(s *stack.Stack) { } bind := func() { - e.state = StateInitial if len(e.BindAddr) == 0 { e.BindAddr = e.ID.LocalAddress } @@ -219,6 +224,16 @@ func (e *endpoint) Resume(s *stack.Stack) { if err := e.connect(tcpip.FullAddress{NIC: e.boundNICID, Addr: e.connectingAddress, Port: e.ID.RemotePort}, false, e.workerRunning); err != tcpip.ErrConnectStarted { panic("endpoint connecting failed: " + err.String()) } + e.mu.Lock() + e.state = e.origEndpointState + closed := e.closed + e.mu.Unlock() + e.notifyProtocolGoroutine(notifyTickleWorker) + if state == StateFinWait2 && closed { + // If the endpoint has been closed then make sure we notify so + // that the FIN_WAIT2 timer is started after a restore. + e.notifyProtocolGoroutine(notifyClose) + } connectedLoading.Done() case StateListen: tcpip.AsyncLoading.Add(1) @@ -265,8 +280,11 @@ func (e *endpoint) Resume(s *stack.Stack) { tcpip.AsyncLoading.Done() }() } - fallthrough + e.state = StateClose + e.stack.CompleteTransportEndpointCleanup(e) + tcpip.DeleteDanglingEndpoint(e) case StateError: + e.state = StateError e.stack.CompleteTransportEndpointCleanup(e) tcpip.DeleteDanglingEndpoint(e) } diff --git a/pkg/tcpip/transport/tcp/protocol.go b/pkg/tcpip/transport/tcp/protocol.go index c8e4a0d7e..89b965c23 100644 --- a/pkg/tcpip/transport/tcp/protocol.go +++ b/pkg/tcpip/transport/tcp/protocol.go @@ -23,6 +23,7 @@ package tcp import ( "strings" "sync" + "time" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" @@ -54,6 +55,14 @@ const ( // MaxUnprocessedSegments is the maximum number of unprocessed segments // that can be queued for a given endpoint. MaxUnprocessedSegments = 300 + + // DefaultTCPLingerTimeout is the amount of time that sockets linger in + // FIN_WAIT_2 state before being marked closed. + DefaultTCPLingerTimeout = 60 * time.Second + + // DefaultTCPTimeWaitTimeout is the amount of time that sockets linger + // in TIME_WAIT state before being marked closed. + DefaultTCPTimeWaitTimeout = 60 * time.Second ) // SACKEnabled option can be used to enable SACK support in the TCP @@ -93,6 +102,8 @@ type protocol struct { congestionControl string availableCongestionControl []string moderateReceiveBuffer bool + tcpLingerTimeout time.Duration + tcpTimeWaitTimeout time.Duration } // Number returns the tcp protocol number. @@ -212,6 +223,24 @@ func (p *protocol) SetOption(option interface{}) *tcpip.Error { p.mu.Unlock() return nil + case tcpip.TCPLingerTimeoutOption: + if v < 0 { + v = 0 + } + p.mu.Lock() + p.tcpLingerTimeout = time.Duration(v) + p.mu.Unlock() + return nil + + case tcpip.TCPTimeWaitTimeoutOption: + if v < 0 { + v = 0 + } + p.mu.Lock() + p.tcpTimeWaitTimeout = time.Duration(v) + p.mu.Unlock() + return nil + default: return tcpip.ErrUnknownProtocolOption } @@ -262,6 +291,18 @@ func (p *protocol) Option(option interface{}) *tcpip.Error { p.mu.Unlock() return nil + case *tcpip.TCPLingerTimeoutOption: + p.mu.Lock() + *v = tcpip.TCPLingerTimeoutOption(p.tcpLingerTimeout) + p.mu.Unlock() + return nil + + case *tcpip.TCPTimeWaitTimeoutOption: + p.mu.Lock() + *v = tcpip.TCPTimeWaitTimeoutOption(p.tcpTimeWaitTimeout) + p.mu.Unlock() + return nil + default: return tcpip.ErrUnknownProtocolOption } @@ -274,5 +315,7 @@ func NewProtocol() stack.TransportProtocol { recvBufferSize: ReceiveBufferSizeOption{MinBufferSize, DefaultReceiveBufferSize, MaxBufferSize}, congestionControl: ccReno, availableCongestionControl: []string{ccReno, ccCubic}, + tcpLingerTimeout: DefaultTCPLingerTimeout, + tcpTimeWaitTimeout: DefaultTCPTimeWaitTimeout, } } diff --git a/pkg/tcpip/transport/tcp/rcv.go b/pkg/tcpip/transport/tcp/rcv.go index e90f9a7d9..068b90fb6 100644 --- a/pkg/tcpip/transport/tcp/rcv.go +++ b/pkg/tcpip/transport/tcp/rcv.go @@ -18,6 +18,7 @@ import ( "container/heap" "time" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/seqnum" ) @@ -209,6 +210,11 @@ func (r *receiver) consumeSegment(s *segment, segSeq seqnum.Value, segLen seqnum switch r.ep.state { case StateFinWait1: r.ep.state = StateFinWait2 + // Notify protocol goroutine that we have received an + // ACK to our FIN so that it can start the FIN_WAIT2 + // timer to abort connection if the other side does + // not close within 2MSL. + r.ep.notifyProtocolGoroutine(notifyClose) case StateClosing: r.ep.state = StateTimeWait case StateLastAck: @@ -253,23 +259,105 @@ func (r *receiver) updateRTT() { r.ep.rcvListMu.Unlock() } -// handleRcvdSegment handles TCP segments directed at the connection managed by -// r as they arrive. It is called by the protocol main loop. -func (r *receiver) handleRcvdSegment(s *segment) { +func (r *receiver) handleRcvdSegmentClosing(s *segment, state EndpointState, closed bool) (drop bool, err *tcpip.Error) { + r.ep.rcvListMu.Lock() + rcvClosed := r.ep.rcvClosed || r.closed + r.ep.rcvListMu.Unlock() + + // If we are in one of the shutdown states then we need to do + // additional checks before we try and process the segment. + switch state { + case StateCloseWait, StateClosing, StateLastAck: + if !s.sequenceNumber.LessThanEq(r.rcvNxt) { + s.decRef() + // Just drop the segment as we have + // already received a FIN and this + // segment is after the sequence number + // for the FIN. + return true, nil + } + fallthrough + case StateFinWait1: + fallthrough + case StateFinWait2: + // If we are closed for reads (either due to an + // incoming FIN or the user calling shutdown(.., + // SHUT_RD) then any data past the rcvNxt should + // trigger a RST. + endDataSeq := s.sequenceNumber.Add(seqnum.Size(s.data.Size())) + if rcvClosed && r.rcvNxt.LessThan(endDataSeq) { + s.decRef() + return true, tcpip.ErrConnectionAborted + } + if state == StateFinWait1 { + break + } + + // If it's a retransmission of an old data segment + // or a pure ACK then allow it. + if s.sequenceNumber.Add(s.logicalLen()).LessThanEq(r.rcvNxt) || + s.logicalLen() == 0 { + break + } + + // In FIN-WAIT2 if the socket is fully + // closed(not owned by application on our end + // then the only acceptable segment is a + // FIN. Since FIN can technically also carry + // data we verify that the segment carrying a + // FIN ends at exactly e.rcvNxt+1. + // + // From RFC793 page 25. + // + // For sequence number purposes, the SYN is + // considered to occur before the first actual + // data octet of the segment in which it occurs, + // while the FIN is considered to occur after + // the last actual data octet in a segment in + // which it occurs. + if closed && (!s.flagIsSet(header.TCPFlagFin) || s.sequenceNumber.Add(s.logicalLen()) != r.rcvNxt+1) { + s.decRef() + return true, tcpip.ErrConnectionAborted + } + } + // We don't care about receive processing anymore if the receive side // is closed. - if r.closed { - return + // + // NOTE: We still want to permit a FIN as it's possible only our + // end has closed and the peer is yet to send a FIN. Hence we + // compare only the payload. + segEnd := s.sequenceNumber.Add(seqnum.Size(s.data.Size())) + if rcvClosed && !segEnd.LessThanEq(r.rcvNxt) { + return true, nil + } + return false, nil +} + +// handleRcvdSegment handles TCP segments directed at the connection managed by +// r as they arrive. It is called by the protocol main loop. +func (r *receiver) handleRcvdSegment(s *segment) (drop bool, err *tcpip.Error) { + r.ep.mu.RLock() + state := r.ep.state + closed := r.ep.closed + r.ep.mu.RUnlock() + + if state != StateEstablished { + drop, err := r.handleRcvdSegmentClosing(s, state, closed) + if drop || err != nil { + return drop, err + } } segLen := seqnum.Size(s.data.Size()) segSeq := s.sequenceNumber // If the sequence number range is outside the acceptable range, just - // send an ACK. This is according to RFC 793, page 37. + // send an ACK and stop further processing of the segment. + // This is according to RFC 793, page 68. if !r.acceptable(segSeq, segLen) { r.ep.snd.sendAck() - return + return true, nil } // Defer segment processing if it can't be consumed now. @@ -288,7 +376,7 @@ func (r *receiver) handleRcvdSegment(s *segment) { // have to retransmit. r.ep.snd.sendAck() } - return + return false, nil } // Since we consumed a segment update the receiver's RTT estimate @@ -315,4 +403,67 @@ func (r *receiver) handleRcvdSegment(s *segment) { r.pendingBufUsed -= s.logicalLen() s.decRef() } + return false, nil +} + +// handleTimeWaitSegment handles inbound segments received when the endpoint +// has entered the TIME_WAIT state. +func (r *receiver) handleTimeWaitSegment(s *segment) (resetTimeWait bool, newSyn bool) { + segSeq := s.sequenceNumber + segLen := seqnum.Size(s.data.Size()) + + // Just silently drop any RST packets in TIME_WAIT. We do not support + // TIME_WAIT assasination as a result we confirm w/ fix 1 as described + // in https://tools.ietf.org/html/rfc1337#section-3. + if s.flagIsSet(header.TCPFlagRst) { + return false, false + } + + // If it's a SYN and the sequence number is higher than any seen before + // for this connection then try and redirect it to a listening endpoint + // if available. + // + // RFC 1122: + // "When a connection is [...] on TIME-WAIT state [...] + // [a TCP] MAY accept a new SYN from the remote TCP to + // reopen the connection directly, if it: + + // (1) assigns its initial sequence number for the new + // connection to be larger than the largest sequence + // number it used on the previous connection incarnation, + // and + + // (2) returns to TIME-WAIT state if the SYN turns out + // to be an old duplicate". + if s.flagIsSet(header.TCPFlagSyn) && r.rcvNxt.LessThan(segSeq) { + + return false, true + } + + // Drop the segment if it does not contain an ACK. + if !s.flagIsSet(header.TCPFlagAck) { + return false, false + } + + // Update Timestamp if required. See RFC7323, section-4.3. + if r.ep.sendTSOk && s.parsedOptions.TS { + r.ep.updateRecentTimestamp(s.parsedOptions.TSVal, r.ep.snd.maxSentAck, segSeq) + } + + if segSeq.Add(1) == r.rcvNxt && s.flagIsSet(header.TCPFlagFin) { + // If it's a FIN-ACK then resetTimeWait and send an ACK, as it + // indicates our final ACK could have been lost. + r.ep.snd.sendAck() + return true, false + } + + // If the sequence number range is outside the acceptable range or + // carries data then just send an ACK. This is according to RFC 793, + // page 37. + // + // NOTE: In TIME_WAIT the only acceptable sequence number is rcvNxt. + if segSeq != r.rcvNxt || segLen != 0 { + r.ep.snd.sendAck() + } + return false, false } diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go index f4ea5f091..0c1704d74 100644 --- a/pkg/tcpip/transport/tcp/tcp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_test.go @@ -206,17 +206,18 @@ func TestTCPResetSentForACKWhenNotUsingSynCookies(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() + // Set TCPLingerTimeout to 5 seconds so that sockets are marked closed wq := &waiter.Queue{} ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) if err != nil { - t.Fatalf("NewEndpoint failed: %v", err) + t.Fatalf("NewEndpoint failed: %s", err) } if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { - t.Fatalf("Bind failed: %v", err) + t.Fatalf("Bind failed: %s", err) } if err := ep.Listen(10); err != nil { - t.Fatalf("Listen failed: %v", err) + t.Fatalf("Listen failed: %s", err) } // Send a SYN request. @@ -256,7 +257,7 @@ func TestTCPResetSentForACKWhenNotUsingSynCookies(t *testing.T) { case <-ch: c.EP, _, err = ep.Accept() if err != nil { - t.Fatalf("Accept failed: %v", err) + t.Fatalf("Accept failed: %s", err) } case <-time.After(1 * time.Second): @@ -264,6 +265,13 @@ func TestTCPResetSentForACKWhenNotUsingSynCookies(t *testing.T) { } } + // Lower stackwide TIME_WAIT timeout so that the reservations + // are released instantly on Close. + tcpTW := tcpip.TCPTimeWaitTimeoutOption(1 * time.Millisecond) + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpTW); err != nil { + t.Fatalf("e.stack.SetTransportProtocolOption(%d, %s) = %s", tcp.ProtocolNumber, tcpTW, err) + } + c.EP.Close() checker.IPv4(t, c.GetPacket(), checker.TCP( checker.SrcPort(context.StackPort), @@ -285,6 +293,11 @@ func TestTCPResetSentForACKWhenNotUsingSynCookies(t *testing.T) { // Get the ACK to the FIN we just sent. c.GetPacket() + // Since an active close was done we need to wait for a little more than + // tcpLingerTimeout for the port reservations to be released and the + // socket to move to a CLOSED state. + time.Sleep(20 * time.Millisecond) + // Now resend the same ACK, this ACK should generate a RST as there // should be no endpoint in SYN-RCVD state and we are not using // syn-cookies yet. The reason we send the same ACK is we need a valid @@ -376,6 +389,13 @@ func TestConnectResetAfterClose(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() + // Set TCPLinger to 3 seconds so that sockets are marked closed + // after 3 second in FIN_WAIT2 state. + tcpLingerTimeout := 3 * time.Second + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPLingerTimeoutOption(tcpLingerTimeout)); err != nil { + t.Fatalf("c.stack.SetTransportProtocolOption(tcp, tcpip.TCPLingerTimeoutOption(%d) failed: %s", tcpLingerTimeout, err) + } + c.CreateConnected(789, 30000, -1 /* epRcvBuf */) ep := c.EP c.EP = nil @@ -396,12 +416,24 @@ func TestConnectResetAfterClose(t *testing.T) { DstPort: c.Port, Flags: header.TCPFlagAck, SeqNum: 790, - AckNum: c.IRS.Add(1), + AckNum: c.IRS.Add(2), + RcvWnd: 30000, + }) + + // Wait for the ep to give up waiting for a FIN. + time.Sleep(tcpLingerTimeout + 1*time.Second) + + // Now send an ACK and it should trigger a RST as the endpoint should + // not exist anymore. + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: c.Port, + Flags: header.TCPFlagAck, + SeqNum: 790, + AckNum: c.IRS.Add(2), RcvWnd: 30000, }) - // Wait for the ep to give up waiting for a FIN, and send a RST. - time.Sleep(3 * time.Second) for { b := c.GetPacket() tcpHdr := header.TCP(header.IPv4(b).Payload()) @@ -413,7 +445,7 @@ func TestConnectResetAfterClose(t *testing.T) { checker.IPv4(t, b, checker.TCP( checker.DstPort(context.TestPort), - checker.SeqNum(uint32(c.IRS)+1), + checker.SeqNum(uint32(c.IRS)+2), checker.AckNum(790), checker.TCPFlags(header.TCPFlagAck|header.TCPFlagRst), ), @@ -1110,8 +1142,7 @@ func TestRstOnCloseWithUnreadDataFinConvertRst(t *testing.T) { checker.TCP( checker.DstPort(context.TestPort), checker.TCPFlags(header.TCPFlagAck|header.TCPFlagRst), - // We shouldn't consume a sequence number on RST. - checker.SeqNum(uint32(c.IRS)+1), + checker.SeqNum(uint32(c.IRS)+2), )) // The RST puts the endpoint into an error state. if got, want := tcp.EndpointState(c.EP.State()), tcp.StateError; got != want { @@ -3085,6 +3116,13 @@ func TestReadAfterClosedState(t *testing.T) { c := context.New(t, defaultMTU) defer c.Cleanup() + // Set TCPTimeWaitTimeout to 1 seconds so that sockets are marked closed + // after 1 second in TIME_WAIT state. + tcpTimeWaitTimeout := 1 * time.Second + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPTimeWaitTimeoutOption(tcpTimeWaitTimeout)); err != nil { + t.Fatalf("c.stack.SetTransportProtocolOption(tcp, tcpip.TCPTimeWaitTimeout(%d) failed: %s", tcpTimeWaitTimeout, err) + } + c.CreateConnected(789, 30000, -1 /* epRcvBuf */) we, ch := waiter.NewChannelEntry(nil) @@ -3092,12 +3130,12 @@ func TestReadAfterClosedState(t *testing.T) { defer c.WQ.EventUnregister(&we) if _, _, err := c.EP.Read(nil); err != tcpip.ErrWouldBlock { - t.Fatalf("got c.EP.Read(nil) = %v, want = %v", err, tcpip.ErrWouldBlock) + t.Fatalf("got c.EP.Read(nil) = %v, want = %s", err, tcpip.ErrWouldBlock) } // Shutdown immediately for write, check that we get a FIN. if err := c.EP.Shutdown(tcpip.ShutdownWrite); err != nil { - t.Fatalf("Shutdown failed: %v", err) + t.Fatalf("Shutdown failed: %s", err) } checker.IPv4(t, c.GetPacket(), @@ -3135,10 +3173,9 @@ func TestReadAfterClosedState(t *testing.T) { ), ) - // Give the stack the chance to transition to closed state. Note that since - // both the sender and receiver are now closed, we effectively skip the - // TIME-WAIT state. - time.Sleep(1 * time.Second) + // Give the stack the chance to transition to closed state from + // TIME_WAIT. + time.Sleep(tcpTimeWaitTimeout * 2) if got, want := tcp.EndpointState(c.EP.State()), tcp.StateClose; got != want { t.Errorf("Unexpected endpoint state: want %v, got %v", want, got) @@ -3155,7 +3192,7 @@ func TestReadAfterClosedState(t *testing.T) { peekBuf := make([]byte, 10) n, _, err := c.EP.Peek([][]byte{peekBuf}) if err != nil { - t.Fatalf("Peek failed: %v", err) + t.Fatalf("Peek failed: %s", err) } peekBuf = peekBuf[:n] @@ -3166,7 +3203,7 @@ func TestReadAfterClosedState(t *testing.T) { // Receive data. v, _, err := c.EP.Read(nil) if err != nil { - t.Fatalf("Read failed: %v", err) + t.Fatalf("Read failed: %s", err) } if !bytes.Equal(data, v) { @@ -3176,11 +3213,11 @@ func TestReadAfterClosedState(t *testing.T) { // Now that we drained the queue, check that functions fail with the // right error code. if _, _, err := c.EP.Read(nil); err != tcpip.ErrClosedForReceive { - t.Fatalf("got c.EP.Read(nil) = %v, want = %v", err, tcpip.ErrClosedForReceive) + t.Fatalf("got c.EP.Read(nil) = %v, want = %s", err, tcpip.ErrClosedForReceive) } if _, _, err := c.EP.Peek([][]byte{peekBuf}); err != tcpip.ErrClosedForReceive { - t.Fatalf("got c.EP.Peek(...) = %v, want = %v", err, tcpip.ErrClosedForReceive) + t.Fatalf("got c.EP.Peek(...) = %v, want = %s", err, tcpip.ErrClosedForReceive) } } @@ -4347,7 +4384,8 @@ func TestListenBacklogFullSynCookieInUse(t *testing.T) { // Send a SYN request. irs := seqnum.Value(789) c.SendPacket(nil, &context.Headers{ - SrcPort: context.TestPort, + // pick a different src port for new SYN. + SrcPort: context.TestPort + 1, DstPort: context.StackPort, Flags: header.TCPFlagSyn, SeqNum: irs, @@ -4893,3 +4931,545 @@ func checkDelayOption(t *testing.T, c *context.Context, wantDelayEnabled tcp.Del t.Errorf("ep.GetSockOptInt(tcpip.DelayOption) got: %d, want: %d", gotDelayOption, wantDelayOption) } } + +func TestTCPLingerTimeout(t *testing.T) { + c := context.New(t, 1500 /* mtu */) + defer c.Cleanup() + + c.CreateConnected(789, 30000, -1 /* epRcvBuf */) + + testCases := []struct { + name string + tcpLingerTimeout time.Duration + want time.Duration + }{ + {"NegativeLingerTimeout", -123123, 0}, + {"ZeroLingerTimeout", 0, 0}, + {"InRangeLingerTimeout", 10 * time.Second, 10 * time.Second}, + // Values > stack's TCPLingerTimeout are capped to the stack's + // value. Defaults to tcp.DefaultTCPLingerTimeout(60 seconds) + {"AboveMaxLingerTimeout", 65 * time.Second, 60 * time.Second}, + } + for _, tc := range testCases { + t.Run(tc.name, func(t *testing.T) { + if err := c.EP.SetSockOpt(tcpip.TCPLingerTimeoutOption(tc.tcpLingerTimeout)); err != nil { + t.Fatalf("SetSockOpt(%s) = %s", tc.tcpLingerTimeout, err) + } + var v tcpip.TCPLingerTimeoutOption + if err := c.EP.GetSockOpt(&v); err != nil { + t.Fatalf("GetSockOpt(tcpip.TCPLingerTimeoutOption) = %s", err) + } + if got, want := time.Duration(v), tc.want; got != want { + t.Fatalf("unexpected linger timeout got: %s, want: %s", got, want) + } + }) + } +} + +func TestTCPTimeWaitRSTIgnored(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + wq := &waiter.Queue{} + ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) + if err != nil { + t.Fatalf("NewEndpoint failed: %s", err) + } + if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatalf("Bind failed: %s", err) + } + + if err := ep.Listen(10); err != nil { + t.Fatalf("Listen failed: %s", err) + } + + // Send a SYN request. + iss := seqnum.Value(789) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + // Receive the SYN-ACK reply. + b := c.GetPacket() + tcpHdr := header.TCP(header.IPv4(b).Payload()) + c.IRS = seqnum.Value(tcpHdr.SequenceNumber()) + + ackHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 1, + } + + // Send ACK. + c.SendPacket(nil, ackHeaders) + + // Try to accept the connection. + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + + c.EP, _, err = ep.Accept() + if err == tcpip.ErrWouldBlock { + // Wait for connection to be established. + select { + case <-ch: + c.EP, _, err = ep.Accept() + if err != nil { + t.Fatalf("Accept failed: %s", err) + } + + case <-time.After(1 * time.Second): + t.Fatalf("Timed out waiting for accept") + } + } + + c.EP.Close() + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+1)), + checker.AckNum(uint32(iss)+1), + checker.TCPFlags(header.TCPFlagFin|header.TCPFlagAck))) + + finHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck | header.TCPFlagFin, + SeqNum: iss + 1, + AckNum: c.IRS + 2, + } + + c.SendPacket(nil, finHeaders) + + // Get the ACK to the FIN we just sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) + + // Now send a RST and this should be ignored and not + // generate an ACK. + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagRst, + SeqNum: iss + 1, + AckNum: c.IRS + 2, + }) + + c.CheckNoPacketTimeout("unexpected packet received in TIME_WAIT state", 1*time.Second) + + // Out of order ACK should generate an immediate ACK in + // TIME_WAIT. + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 3, + }) + + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) +} + +func TestTCPTimeWaitOutOfOrder(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + wq := &waiter.Queue{} + ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) + if err != nil { + t.Fatalf("NewEndpoint failed: %s", err) + } + if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatalf("Bind failed: %s", err) + } + + if err := ep.Listen(10); err != nil { + t.Fatalf("Listen failed: %s", err) + } + + // Send a SYN request. + iss := seqnum.Value(789) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + // Receive the SYN-ACK reply. + b := c.GetPacket() + tcpHdr := header.TCP(header.IPv4(b).Payload()) + c.IRS = seqnum.Value(tcpHdr.SequenceNumber()) + + ackHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 1, + } + + // Send ACK. + c.SendPacket(nil, ackHeaders) + + // Try to accept the connection. + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + + c.EP, _, err = ep.Accept() + if err == tcpip.ErrWouldBlock { + // Wait for connection to be established. + select { + case <-ch: + c.EP, _, err = ep.Accept() + if err != nil { + t.Fatalf("Accept failed: %s", err) + } + + case <-time.After(1 * time.Second): + t.Fatalf("Timed out waiting for accept") + } + } + + c.EP.Close() + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+1)), + checker.AckNum(uint32(iss)+1), + checker.TCPFlags(header.TCPFlagFin|header.TCPFlagAck))) + + finHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck | header.TCPFlagFin, + SeqNum: iss + 1, + AckNum: c.IRS + 2, + } + + c.SendPacket(nil, finHeaders) + + // Get the ACK to the FIN we just sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) + + // Out of order ACK should generate an immediate ACK in + // TIME_WAIT. + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 3, + }) + + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) +} + +func TestTCPTimeWaitNewSyn(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + wq := &waiter.Queue{} + ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) + if err != nil { + t.Fatalf("NewEndpoint failed: %s", err) + } + if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatalf("Bind failed: %s", err) + } + + if err := ep.Listen(10); err != nil { + t.Fatalf("Listen failed: %s", err) + } + + // Send a SYN request. + iss := seqnum.Value(789) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + // Receive the SYN-ACK reply. + b := c.GetPacket() + tcpHdr := header.TCP(header.IPv4(b).Payload()) + c.IRS = seqnum.Value(tcpHdr.SequenceNumber()) + + ackHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 1, + } + + // Send ACK. + c.SendPacket(nil, ackHeaders) + + // Try to accept the connection. + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + + c.EP, _, err = ep.Accept() + if err == tcpip.ErrWouldBlock { + // Wait for connection to be established. + select { + case <-ch: + c.EP, _, err = ep.Accept() + if err != nil { + t.Fatalf("Accept failed: %s", err) + } + + case <-time.After(1 * time.Second): + t.Fatalf("Timed out waiting for accept") + } + } + + c.EP.Close() + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+1)), + checker.AckNum(uint32(iss)+1), + checker.TCPFlags(header.TCPFlagFin|header.TCPFlagAck))) + + finHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck | header.TCPFlagFin, + SeqNum: iss + 1, + AckNum: c.IRS + 2, + } + + c.SendPacket(nil, finHeaders) + + // Get the ACK to the FIN we just sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) + + // Send a SYN request w/ sequence number lower than + // the highest sequence number sent. We just reuse + // the same number. + iss = seqnum.Value(789) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + c.CheckNoPacketTimeout("unexpected packet received in response to SYN", 1*time.Second) + + // Send a SYN request w/ sequence number higher than + // the highest sequence number sent. + iss = seqnum.Value(792) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + // Receive the SYN-ACK reply. + b = c.GetPacket() + tcpHdr = header.TCP(header.IPv4(b).Payload()) + c.IRS = seqnum.Value(tcpHdr.SequenceNumber()) + + ackHeaders = &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 1, + } + + // Send ACK. + c.SendPacket(nil, ackHeaders) + + // Try to accept the connection. + c.EP, _, err = ep.Accept() + if err == tcpip.ErrWouldBlock { + // Wait for connection to be established. + select { + case <-ch: + c.EP, _, err = ep.Accept() + if err != nil { + t.Fatalf("Accept failed: %s", err) + } + + case <-time.After(1 * time.Second): + t.Fatalf("Timed out waiting for accept") + } + } +} + +func TestTCPTimeWaitDuplicateFINExtendsTimeWait(t *testing.T) { + c := context.New(t, defaultMTU) + defer c.Cleanup() + + // Set TCPTimeWaitTimeout to 5 seconds so that sockets are marked closed + // after 5 seconds in TIME_WAIT state. + tcpTimeWaitTimeout := 5 * time.Second + if err := c.Stack().SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.TCPTimeWaitTimeoutOption(tcpTimeWaitTimeout)); err != nil { + t.Fatalf("c.stack.SetTransportProtocolOption(tcp, tcpip.TCPLingerTimeoutOption(%d) failed: %s", tcpTimeWaitTimeout, err) + } + + wq := &waiter.Queue{} + ep, err := c.Stack().NewEndpoint(tcp.ProtocolNumber, ipv4.ProtocolNumber, wq) + if err != nil { + t.Fatalf("NewEndpoint failed: %s", err) + } + if err := ep.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { + t.Fatalf("Bind failed: %s", err) + } + + if err := ep.Listen(10); err != nil { + t.Fatalf("Listen failed: %s", err) + } + + // Send a SYN request. + iss := seqnum.Value(789) + c.SendPacket(nil, &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagSyn, + SeqNum: iss, + }) + + // Receive the SYN-ACK reply. + b := c.GetPacket() + tcpHdr := header.TCP(header.IPv4(b).Payload()) + c.IRS = seqnum.Value(tcpHdr.SequenceNumber()) + + ackHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck, + SeqNum: iss + 1, + AckNum: c.IRS + 1, + } + + // Send ACK. + c.SendPacket(nil, ackHeaders) + + // Try to accept the connection. + we, ch := waiter.NewChannelEntry(nil) + wq.EventRegister(&we, waiter.EventIn) + defer wq.EventUnregister(&we) + + c.EP, _, err = ep.Accept() + if err == tcpip.ErrWouldBlock { + // Wait for connection to be established. + select { + case <-ch: + c.EP, _, err = ep.Accept() + if err != nil { + t.Fatalf("Accept failed: %s", err) + } + + case <-time.After(1 * time.Second): + t.Fatalf("Timed out waiting for accept") + } + } + + c.EP.Close() + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+1)), + checker.AckNum(uint32(iss)+1), + checker.TCPFlags(header.TCPFlagFin|header.TCPFlagAck))) + + finHeaders := &context.Headers{ + SrcPort: context.TestPort, + DstPort: context.StackPort, + Flags: header.TCPFlagAck | header.TCPFlagFin, + SeqNum: iss + 1, + AckNum: c.IRS + 2, + } + + c.SendPacket(nil, finHeaders) + + // Get the ACK to the FIN we just sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) + + time.Sleep(2 * time.Second) + + // Now send a duplicate FIN. This should cause the TIME_WAIT to extend + // by another 5 seconds and also send us a duplicate ACK as it should + // indicate that the final ACK was potentially lost. + c.SendPacket(nil, finHeaders) + + // Get the ACK to the FIN we just sent. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(c.IRS+2)), + checker.AckNum(uint32(iss)+2), + checker.TCPFlags(header.TCPFlagAck))) + + // Sleep for 4 seconds so at this point we are 1 second past the + // original tcpLingerTimeout of 5 seconds. + time.Sleep(4 * time.Second) + + // Send an ACK and it should not generate any packet as the socket + // should still be in TIME_WAIT for another another 5 seconds due + // to the duplicate FIN we sent earlier. + *ackHeaders = *finHeaders + ackHeaders.SeqNum = ackHeaders.SeqNum + 1 + ackHeaders.Flags = header.TCPFlagAck + c.SendPacket(nil, ackHeaders) + + c.CheckNoPacketTimeout("unexpected packet received from endpoint in TIME_WAIT", 1*time.Second) + // Now sleep for another 2 seconds so that we are past the + // extended TIME_WAIT of 7 seconds (2 + 5). + time.Sleep(2 * time.Second) + + // Resend the same ACK. + c.SendPacket(nil, ackHeaders) + + // Receive the RST that should be generated as there is no valid + // endpoint. + checker.IPv4(t, c.GetPacket(), checker.TCP( + checker.SrcPort(context.StackPort), + checker.DstPort(context.TestPort), + checker.SeqNum(uint32(ackHeaders.AckNum)), + checker.AckNum(uint32(ackHeaders.SeqNum)), + checker.TCPFlags(header.TCPFlagRst|header.TCPFlagAck))) +} diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index 3e5b6b3c3..722d14b53 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -9,7 +9,7 @@ syscall_test(test = "//test/syscalls/linux:accept_bind_stream_test") syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:accept_bind_test", ) @@ -434,7 +434,7 @@ syscall_test( syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_abstract_test", ) @@ -445,7 +445,7 @@ syscall_test( syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_domain_test", ) @@ -458,19 +458,19 @@ syscall_test( syscall_test( size = "large", add_overlay = True, - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_filesystem_test", ) syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_inet_loopback_test", ) syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_ip_tcp_generic_loopback_test", ) @@ -481,13 +481,13 @@ syscall_test( syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_ip_tcp_loopback_test", ) syscall_test( size = "medium", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_ip_tcp_udp_generic_loopback_test", ) @@ -498,7 +498,7 @@ syscall_test( syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_ip_udp_loopback_test", ) @@ -560,7 +560,7 @@ syscall_test( syscall_test( size = "large", add_overlay = True, - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_unix_pair_test", ) @@ -599,7 +599,7 @@ syscall_test( syscall_test( size = "large", - shard_count = 10, + shard_count = 50, test = "//test/syscalls/linux:socket_unix_unbound_stream_test", ) diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 93bff8299..f8b8cb724 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -2141,6 +2141,7 @@ cc_library( deps = [ ":socket_test_util", "//test/util:test_util", + "//test/util:thread_util", "@com_google_googletest//:gtest", ], alwayslink = 1, diff --git a/test/syscalls/linux/socket_inet_loopback.cc b/test/syscalls/linux/socket_inet_loopback.cc index ab375aaaf..2eeee352e 100644 --- a/test/syscalls/linux/socket_inet_loopback.cc +++ b/test/syscalls/linux/socket_inet_loopback.cc @@ -13,6 +13,7 @@ // limitations under the License. #include <arpa/inet.h> +#include <linux/tcp.h> #include <netinet/in.h> #include <poll.h> #include <string.h> @@ -31,6 +32,7 @@ #include "gtest/gtest.h" #include "absl/memory/memory.h" #include "absl/strings/str_cat.h" +#include "absl/time/clock.h" #include "absl/time/time.h" #include "test/syscalls/linux/socket_test_util.h" #include "test/util/file_descriptor.h" @@ -267,6 +269,340 @@ TEST_P(SocketInetLoopbackTest, TCPbacklog) { } } +// TCPFinWait2Test creates a pair of connected sockets then closes one end to +// trigger FIN_WAIT2 state for the closed endpoint. Then it binds the same local +// IP/port on a new socket and tries to connect. The connect should fail w/ +// an EADDRINUSE. Then we wait till the FIN_WAIT2 timeout is over and try the +// connect again with a new socket and this time it should succeed. +// +// TCP timers are not S/R today, this can cause this test to be flaky when run +// under random S/R due to timer being reset on a restore. +TEST_P(SocketInetLoopbackTest, TCPFinWait2Test_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // Lower FIN_WAIT2 state to 5 seconds for test. + constexpr int kTCPLingerTimeout = 5; + EXPECT_THAT(setsockopt(conn_fd.get(), IPPROTO_TCP, TCP_LINGER2, + &kTCPLingerTimeout, sizeof(kTCPLingerTimeout)), + SyscallSucceedsWithValue(0)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + // Now bind and connect a new socket. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // Disable cooperative saves after this point. As a save between the first + // bind/connect and the second one can cause the linger timeout timer to + // be restarted causing the final bind/connect to fail. + DisableSave ds; + + // TODO(gvisor.dev/issue/1030): Portmanager does not track all 5 tuple + // reservations which causes the bind() to succeed on gVisor but connect + // correctly fails. + if (IsRunningOnGvisor()) { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallFailsWithErrno(EADDRINUSE)); + } else { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallFailsWithErrno(EADDRINUSE)); + } + + // Sleep for a little over the linger timeout to reduce flakiness in + // save/restore tests. + absl::SleepFor(absl::Seconds(kTCPLingerTimeout + 1)); + + ds.reset(); + + if (!IsRunningOnGvisor()) { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallSucceeds()); + } + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + +// TCPLinger2TimeoutAfterClose creates a pair of connected sockets +// then closes one end to trigger FIN_WAIT2 state for the closed endpont. +// It then sleeps for the TCP_LINGER2 timeout and verifies that bind/ +// connecting the same address succeeds. +// +// TCP timers are not S/R today, this can cause this test to be flaky when run +// under random S/R due to timer being reset on a restore. +TEST_P(SocketInetLoopbackTest, TCPLinger2TimeoutAfterClose_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + constexpr int kTCPLingerTimeout = 5; + EXPECT_THAT(setsockopt(conn_fd.get(), IPPROTO_TCP, TCP_LINGER2, + &kTCPLingerTimeout, sizeof(kTCPLingerTimeout)), + SyscallSucceedsWithValue(0)); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + absl::SleepFor(absl::Seconds(kTCPLingerTimeout + 1)); + + // Now bind and connect a new socket and verify that we can immediately + // rebind the address bound by the conn_fd as it never entered TIME_WAIT. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + ASSERT_THAT(bind(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_bound_addr), conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + +// TCPResetAfterClose creates a pair of connected sockets then closes +// one end to trigger FIN_WAIT2 state for the closed endpoint verifies +// that we generate RSTs for any new data after the socket is fully +// closed. +TEST_P(SocketInetLoopbackTest, TCPResetAfterClose) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + int data = 1234; + + // Now send data which should trigger a RST as the other end should + // have timed out and closed the socket. + EXPECT_THAT(RetryEINTR(send)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceeds()); + // Sleep for a shortwhile to get a RST back. + absl::SleepFor(absl::Seconds(1)); + + // Try writing again and we should get an EPIPE back. + EXPECT_THAT(RetryEINTR(send)(accepted.get(), &data, sizeof(data), 0), + SyscallFailsWithErrno(EPIPE)); + + // Trying to read should return zero as the other end did send + // us a FIN. We do it twice to verify that the RST does not cause an + // ECONNRESET on the read after EOF has been read by applicaiton. + EXPECT_THAT(RetryEINTR(recv)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceedsWithValue(0)); + EXPECT_THAT(RetryEINTR(recv)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceedsWithValue(0)); +} + +// This test is disabled under random save as the the restore run +// results in the stack.Seed() being different which can cause +// sequence number of final connect to be one that is considered +// old and can cause the test to be flaky. +TEST_P(SocketInetLoopbackTest, TCPTimeWaitTest_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // We disable saves after this point as a S/R causes the netstack seed + // to be regenerated which changes what ports/ISN is picked for a given + // tuple (src ip,src port, dst ip, dst port). This can cause the final + // SYN to use a sequence number that looks like one from the current + // connection in TIME_WAIT and will not be accepted causing the test + // to timeout. + // + // TODO(gvisor.dev/issue/940): S/R portSeed/portHint + DisableSave ds; + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + // close the accept FD to trigger TIME_WAIT on the accepted socket which + // should cause the conn_fd to follow CLOSE_WAIT->LAST_ACK->CLOSED instead of + // TIME_WAIT. + accepted.reset(); + absl::SleepFor(absl::Seconds(1)); + conn_fd.reset(); + absl::SleepFor(absl::Seconds(1)); + + // Now bind and connect a new socket and verify that we can immediately + // rebind the address bound by the conn_fd as it never entered TIME_WAIT. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + ASSERT_THAT(bind(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_bound_addr), conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + INSTANTIATE_TEST_SUITE_P( All, SocketInetLoopbackTest, ::testing::Values( diff --git a/test/syscalls/linux/socket_ip_tcp_generic.cc b/test/syscalls/linux/socket_ip_tcp_generic.cc index 592448289..a37b49447 100644 --- a/test/syscalls/linux/socket_ip_tcp_generic.cc +++ b/test/syscalls/linux/socket_ip_tcp_generic.cc @@ -26,6 +26,7 @@ #include "gtest/gtest.h" #include "test/syscalls/linux/socket_test_util.h" #include "test/util/test_util.h" +#include "test/util/thread_util.h" namespace gvisor { namespace testing { @@ -243,6 +244,31 @@ TEST_P(TCPSocketPairTest, ShutdownRdAllowsReadOfReceivedDataBeforeEOF) { SyscallSucceedsWithValue(0)); } +// This test verifies that a shutdown(wr) by the server after sending +// data allows the client to still read() the queued data and a client +// close after sending response allows server to read the incoming +// response. +TEST_P(TCPSocketPairTest, ShutdownWrServerClientClose) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + char buf[10] = {}; + ScopedThread t([&]() { + ASSERT_THAT(RetryEINTR(read)(sockets->first_fd(), buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + ASSERT_THAT(RetryEINTR(write)(sockets->first_fd(), buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + ASSERT_THAT(close(sockets->release_first_fd()), + SyscallSucceedsWithValue(0)); + }); + ASSERT_THAT(RetryEINTR(write)(sockets->second_fd(), buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + ASSERT_THAT(RetryEINTR(shutdown)(sockets->second_fd(), SHUT_WR), + SyscallSucceedsWithValue(0)); + t.Join(); + + ASSERT_THAT(RetryEINTR(read)(sockets->second_fd(), buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); +} + TEST_P(TCPSocketPairTest, ClosedReadNonBlockingSocket) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); @@ -696,5 +722,72 @@ TEST_P(TCPSocketPairTest, SetCongestionControlFailsForUnsupported) { EXPECT_EQ(0, memcmp(got_cc, old_cc, sizeof(old_cc))); } +// Linux and Netstack both default to a 60s TCP_LINGER2 timeout. +constexpr int kDefaultTCPLingerTimeout = 60; + +TEST_P(TCPSocketPairTest, TCPLingerTimeoutDefault) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + int get = -1; + socklen_t get_len = sizeof(get); + EXPECT_THAT( + getsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kDefaultTCPLingerTimeout); +} + +TEST_P(TCPSocketPairTest, SetTCPLingerTimeoutZeroOrLess) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + constexpr int kZero = 0; + EXPECT_THAT(setsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, &kZero, + sizeof(kZero)), + SyscallSucceedsWithValue(0)); + + constexpr int kNegative = -1234; + EXPECT_THAT(setsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, + &kNegative, sizeof(kNegative)), + SyscallSucceedsWithValue(0)); +} + +TEST_P(TCPSocketPairTest, SetTCPLingerTimeoutAboveDefault) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + // Values above the net.ipv4.tcp_fin_timeout are capped to tcp_fin_timeout + // on linux (defaults to 60 seconds on linux). + constexpr int kAboveDefault = kDefaultTCPLingerTimeout + 1; + EXPECT_THAT(setsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, + &kAboveDefault, sizeof(kAboveDefault)), + SyscallSucceedsWithValue(0)); + + int get = -1; + socklen_t get_len = sizeof(get); + EXPECT_THAT( + getsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kDefaultTCPLingerTimeout); +} + +TEST_P(TCPSocketPairTest, SetTCPLingerTimeout) { + auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); + + // Values above the net.ipv4.tcp_fin_timeout are capped to tcp_fin_timeout + // on linux (defaults to 60 seconds on linux). + constexpr int kTCPLingerTimeout = kDefaultTCPLingerTimeout - 1; + EXPECT_THAT(setsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, + &kTCPLingerTimeout, sizeof(kTCPLingerTimeout)), + SyscallSucceedsWithValue(0)); + + int get = -1; + socklen_t get_len = sizeof(get); + EXPECT_THAT( + getsockopt(sockets->first_fd(), IPPROTO_TCP, TCP_LINGER2, &get, &get_len), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(get_len, sizeof(get)); + EXPECT_EQ(get, kTCPLingerTimeout); +} + } // namespace testing } // namespace gvisor |