Merge release-20190806.1-382-g66ebb65 (automated)

1 files changed, 293 insertions, 29 deletions

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&notifyClose != 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&notifyKeepaliveChanged != 0 {
@@ -1117,6 +1235,12 @@ func (e *endpoint) protocolMainLoop(handshake bool) *tcpip.Error {
 					}
 				}
 
+				if n&notifyTickleWorker != 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&notifyClose != 0 {
+				return nil
+			}
+			if n&notifyDrain != 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
+		}
+	}
+}