2 files changed, 68 insertions, 4 deletions

diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go
index 044123185..355719beb 100644
--- a/pkg/tcpip/transport/tcp/endpoint.go
+++ b/pkg/tcpip/transport/tcp/endpoint.go
@@ -1717,6 +1717,27 @@ func (e *endpoint) OnSetReceiveBufferSize(rcvBufSz, oldSz int64) (newSz int64) {
 	return rcvBufSz
 }
 
+// OnSetSendBufferSize implements tcpip.SocketOptionsHandler.OnSetSendBufferSize.
+func (e *endpoint) OnSetSendBufferSize(sz int64) int64 {
+	atomic.StoreUint32(&e.sndQueueInfo.TCPSndBufState.AutoTuneSndBufDisabled, 1)
+	return sz
+}
+
+// WakeupWriters implements tcpip.SocketOptionsHandler.WakeupWriters.
+func (e *endpoint) WakeupWriters() {
+	e.LockUser()
+	defer e.UnlockUser()
+
+	sendBufferSize := e.getSendBufferSize()
+	e.sndQueueInfo.sndQueueMu.Lock()
+	notify := (sendBufferSize - e.sndQueueInfo.SndBufUsed) >= e.sndQueueInfo.SndBufUsed>>1
+	e.sndQueueInfo.sndQueueMu.Unlock()
+
+	if notify {
+		e.waiterQueue.Notify(waiter.WritableEvents)
+	}
+}
+
 // SetSockOptInt sets a socket option.
 func (e *endpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) tcpip.Error {
 	// Lower 2 bits represents ECN bits. RFC 3168, section 23.1
@@ -2329,6 +2350,9 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) tcp
 		e.segmentQueue.mu.Unlock()
 		e.snd.updateMaxPayloadSize(int(e.route.MTU()), 0)
 		e.setEndpointState(StateEstablished)
+		// Set the new auto tuned send buffer size after entering
+		// established state.
+		e.ops.SetSendBufferSize(e.computeTCPSendBufferSize(), false /* notify */)
 	}
 
 	if run {
@@ -2763,13 +2787,20 @@ func (e *endpoint) updateSndBufferUsage(v int) {
 	e.sndQueueInfo.sndQueueMu.Lock()
 	notify := e.sndQueueInfo.SndBufUsed >= sendBufferSize>>1
 	e.sndQueueInfo.SndBufUsed -= v
+
+	// Get the new send buffer size with auto tuning, but do not set it
+	// unless we decide to notify the writers.
+	newSndBufSz := e.computeTCPSendBufferSize()
+
 	// We only notify when there is half the sendBufferSize available after
 	// a full buffer event occurs. This ensures that we don't wake up
 	// writers to queue just 1-2 segments and go back to sleep.
-	notify = notify && e.sndQueueInfo.SndBufUsed < sendBufferSize>>1
+	notify = notify && e.sndQueueInfo.SndBufUsed < int(newSndBufSz)>>1
 	e.sndQueueInfo.sndQueueMu.Unlock()
 
 	if notify {
+		// Set the new send buffer size calculated from auto tuning.
+		e.ops.SetSendBufferSize(newSndBufSz, false /* notify */)
 		e.waiterQueue.Notify(waiter.WritableEvents)
 	}
 }
@@ -3091,3 +3122,36 @@ func GetTCPReceiveBufferLimits(s tcpip.StackHandler) tcpip.ReceiveBufferSizeOpti
 		Max:     ss.Max,
 	}
 }
+
+// computeTCPSendBufferSize implements auto tuning of send buffer size and
+// returns the new send buffer size.
+func (e *endpoint) computeTCPSendBufferSize() int64 {
+	curSndBufSz := int64(e.getSendBufferSize())
+
+	// Auto tuning is disabled when the user explicitly sets the send
+	// buffer size with SO_SNDBUF option.
+	if disabled := atomic.LoadUint32(&e.sndQueueInfo.TCPSndBufState.AutoTuneSndBufDisabled); disabled == 1 {
+		return curSndBufSz
+	}
+
+	const packetOverheadFactor = 2
+	curMSS := e.snd.MaxPayloadSize
+	numSeg := InitialCwnd
+	if numSeg < e.snd.SndCwnd {
+		numSeg = e.snd.SndCwnd
+	}
+
+	// SndCwnd indicates the number of segments that can be sent. This means
+	// that the sender can send upto #SndCwnd segments and the send buffer
+	// size should be set to SndCwnd*MSS to accommodate sending of all the
+	// segments.
+	newSndBufSz := int64(numSeg * curMSS * packetOverheadFactor)
+	if newSndBufSz < curSndBufSz {
+		return curSndBufSz
+	}
+	if ss := GetTCPSendBufferLimits(e.stack); int64(ss.Max) < newSndBufSz {
+		newSndBufSz = int64(ss.Max)
+	}
+
+	return newSndBufSz
+}
diff --git a/pkg/tcpip/transport/tcp/snd.go b/pkg/tcpip/transport/tcp/snd.go
index 92a66f17e..64302f576 100644
--- a/pkg/tcpip/transport/tcp/snd.go
+++ b/pkg/tcpip/transport/tcp/snd.go
@@ -1415,9 +1415,6 @@ func (s *sender) handleRcvdSegment(rcvdSeg *segment) {
 			ackLeft -= datalen
 		}
 
-		// Update the send buffer usage and notify potential waiters.
-		s.ep.updateSndBufferUsage(int(acked))
-
 		// Clear SACK information for all acked data.
 		s.ep.scoreboard.Delete(s.SndUna)
 
@@ -1437,6 +1434,9 @@ func (s *sender) handleRcvdSegment(rcvdSeg *segment) {
 			}
 		}
 
+		// Update the send buffer usage and notify potential waiters.
+		s.ep.updateSndBufferUsage(int(acked))
+
 		// It is possible for s.outstanding to drop below zero if we get
 		// a retransmit timeout, reset outstanding to zero but later
 		// get an ack that cover previously sent data.