// Copyright (c) 2012 The Chromium Authors. All rights reserved. // Use of this source code is governed by a BSD-style license that can be // found in the LICENSE file. #include "net/quic/congestion_control/tcp_cubic_sender.h" #include #include "base/metrics/histogram.h" #include "net/quic/congestion_control/prr_sender.h" #include "net/quic/congestion_control/rtt_stats.h" #include "net/quic/crypto/crypto_protocol.h" using std::max; using std::min; namespace net { namespace { // Constants based on TCP defaults. // The minimum cwnd based on RFC 3782 (TCP NewReno) for cwnd reductions on a // fast retransmission. The cwnd after a timeout is still 1. const QuicPacketCount kMinimumCongestionWindow = 2; const QuicByteCount kMaxSegmentSize = kDefaultTCPMSS; const int kMaxBurstLength = 3; const float kRenoBeta = 0.7f; // Reno backoff factor. const uint32 kDefaultNumConnections = 2; // N-connection emulation. } // namespace TcpCubicSender::TcpCubicSender(const QuicClock* clock, const RttStats* rtt_stats, bool reno, QuicPacketCount initial_tcp_congestion_window, QuicPacketCount max_tcp_congestion_window, QuicConnectionStats* stats) : hybrid_slow_start_(clock), cubic_(clock, stats), rtt_stats_(rtt_stats), stats_(stats), reno_(reno), num_connections_(kDefaultNumConnections), congestion_window_count_(0), largest_sent_sequence_number_(0), largest_acked_sequence_number_(0), largest_sent_at_last_cutback_(0), congestion_window_(initial_tcp_congestion_window), previous_congestion_window_(0), slowstart_threshold_(max_tcp_congestion_window), previous_slowstart_threshold_(0), last_cutback_exited_slowstart_(false), max_tcp_congestion_window_(max_tcp_congestion_window), clock_(clock) {} TcpCubicSender::~TcpCubicSender() { UMA_HISTOGRAM_COUNTS("Net.QuicSession.FinalTcpCwnd", congestion_window_); } void TcpCubicSender::SetFromConfig(const QuicConfig& config, bool is_server, bool using_pacing) { if (is_server) { if (config.HasReceivedConnectionOptions() && ContainsQuicTag(config.ReceivedConnectionOptions(), kIW10)) { // Initial window experiment. congestion_window_ = 10; } if (using_pacing) { // Disable the ack train mode in hystart when pacing is enabled, since it // may be falsely triggered. hybrid_slow_start_.set_ack_train_detection(false); } } } bool TcpCubicSender::ResumeConnectionState( const CachedNetworkParameters& cached_network_params) { // If the previous bandwidth estimate is less than an hour old, store in // preparation for doing bandwidth resumption. int64 seconds_since_estimate = clock_->WallNow().ToUNIXSeconds() - cached_network_params.timestamp(); if (seconds_since_estimate > kNumSecondsPerHour) { return false; } QuicBandwidth bandwidth = QuicBandwidth::FromBytesPerSecond( cached_network_params.bandwidth_estimate_bytes_per_second()); QuicTime::Delta rtt_ms = QuicTime::Delta::FromMilliseconds(cached_network_params.min_rtt_ms()); // Make sure CWND is in appropriate range (in case of bad data). QuicPacketCount new_congestion_window = bandwidth.ToBytesPerPeriod(rtt_ms) / kMaxPacketSize; congestion_window_ = max(min(new_congestion_window, kMaxTcpCongestionWindow), kMinCongestionWindowForBandwidthResumption); // TODO(rjshade): Set appropriate CWND when previous connection was in slow // start at time of estimate. return true; } void TcpCubicSender::SetNumEmulatedConnections(int num_connections) { num_connections_ = max(1, num_connections); cubic_.SetNumConnections(num_connections_); } float TcpCubicSender::RenoBeta() const { // kNConnectionBeta is the backoff factor after loss for our N-connection // emulation, which emulates the effective backoff of an ensemble of N // TCP-Reno connections on a single loss event. The effective multiplier is // computed as: return (num_connections_ - 1 + kRenoBeta) / num_connections_; } void TcpCubicSender::OnCongestionEvent( bool rtt_updated, QuicByteCount bytes_in_flight, const CongestionVector& acked_packets, const CongestionVector& lost_packets) { if (rtt_updated && InSlowStart() && hybrid_slow_start_.ShouldExitSlowStart(rtt_stats_->latest_rtt(), rtt_stats_->min_rtt(), congestion_window_)) { slowstart_threshold_ = congestion_window_; } for (CongestionVector::const_iterator it = lost_packets.begin(); it != lost_packets.end(); ++it) { OnPacketLost(it->first, bytes_in_flight); } for (CongestionVector::const_iterator it = acked_packets.begin(); it != acked_packets.end(); ++it) { OnPacketAcked(it->first, it->second.bytes_sent, bytes_in_flight); } } void TcpCubicSender::OnPacketAcked( QuicPacketSequenceNumber acked_sequence_number, QuicByteCount acked_bytes, QuicByteCount bytes_in_flight) { largest_acked_sequence_number_ = max(acked_sequence_number, largest_acked_sequence_number_); // As soon as a packet is acked, ensure we're no longer in RTO mode. previous_congestion_window_ = 0; if (InRecovery()) { // PRR is used when in recovery. prr_.OnPacketAcked(acked_bytes); return; } MaybeIncreaseCwnd(acked_sequence_number, bytes_in_flight); // TODO(ianswett): Should this even be called when not in slow start? hybrid_slow_start_.OnPacketAcked(acked_sequence_number, InSlowStart()); } void TcpCubicSender::OnPacketLost(QuicPacketSequenceNumber sequence_number, QuicByteCount bytes_in_flight) { // TCP NewReno (RFC6582) says that once a loss occurs, any losses in packets // already sent should be treated as a single loss event, since it's expected. if (sequence_number <= largest_sent_at_last_cutback_) { if (last_cutback_exited_slowstart_) { ++stats_->slowstart_packets_lost; } DVLOG(1) << "Ignoring loss for largest_missing:" << sequence_number << " because it was sent prior to the last CWND cutback."; return; } ++stats_->tcp_loss_events; last_cutback_exited_slowstart_ = InSlowStart(); if (InSlowStart()) { ++stats_->slowstart_packets_lost; } prr_.OnPacketLost(bytes_in_flight); if (reno_) { congestion_window_ = congestion_window_ * RenoBeta(); } else { congestion_window_ = cubic_.CongestionWindowAfterPacketLoss(congestion_window_); } slowstart_threshold_ = congestion_window_; // Enforce TCP's minimum congestion window of 2*MSS. if (congestion_window_ < kMinimumCongestionWindow) { congestion_window_ = kMinimumCongestionWindow; } largest_sent_at_last_cutback_ = largest_sent_sequence_number_; // reset packet count from congestion avoidance mode. We start // counting again when we're out of recovery. congestion_window_count_ = 0; DVLOG(1) << "Incoming loss; congestion window: " << congestion_window_ << " slowstart threshold: " << slowstart_threshold_; } bool TcpCubicSender::OnPacketSent(QuicTime /*sent_time*/, QuicByteCount /*bytes_in_flight*/, QuicPacketSequenceNumber sequence_number, QuicByteCount bytes, HasRetransmittableData is_retransmittable) { // Only update bytes_in_flight_ for data packets. if (is_retransmittable != HAS_RETRANSMITTABLE_DATA) { return false; } if (InRecovery()) { // PRR is used when in recovery. prr_.OnPacketSent(bytes); } DCHECK_LT(largest_sent_sequence_number_, sequence_number); largest_sent_sequence_number_ = sequence_number; hybrid_slow_start_.OnPacketSent(sequence_number); return true; } QuicTime::Delta TcpCubicSender::TimeUntilSend( QuicTime /* now */, QuicByteCount bytes_in_flight, HasRetransmittableData has_retransmittable_data) const { if (has_retransmittable_data == NO_RETRANSMITTABLE_DATA) { // For TCP we can always send an ACK immediately. return QuicTime::Delta::Zero(); } if (InRecovery()) { // PRR is used when in recovery. return prr_.TimeUntilSend(GetCongestionWindow(), bytes_in_flight, slowstart_threshold_); } if (GetCongestionWindow() > bytes_in_flight) { return QuicTime::Delta::Zero(); } return QuicTime::Delta::Infinite(); } QuicBandwidth TcpCubicSender::PacingRate() const { // We pace at twice the rate of the underlying sender's bandwidth estimate // during slow start and 1.25x during congestion avoidance to ensure pacing // doesn't prevent us from filling the window. QuicTime::Delta srtt = rtt_stats_->smoothed_rtt(); if (srtt.IsZero()) { srtt = QuicTime::Delta::FromMicroseconds(rtt_stats_->initial_rtt_us()); } const QuicBandwidth bandwidth = QuicBandwidth::FromBytesAndTimeDelta(GetCongestionWindow(), srtt); return bandwidth.Scale(InSlowStart() ? 2 : 1.25); } QuicBandwidth TcpCubicSender::BandwidthEstimate() const { QuicTime::Delta srtt = rtt_stats_->smoothed_rtt(); if (srtt.IsZero()) { // If we haven't measured an rtt, the bandwidth estimate is unknown. return QuicBandwidth::Zero(); } return QuicBandwidth::FromBytesAndTimeDelta(GetCongestionWindow(), srtt); } bool TcpCubicSender::HasReliableBandwidthEstimate() const { return !InSlowStart() && !InRecovery() && !rtt_stats_->smoothed_rtt().IsZero();; } QuicTime::Delta TcpCubicSender::RetransmissionDelay() const { if (rtt_stats_->smoothed_rtt().IsZero()) { return QuicTime::Delta::Zero(); } return rtt_stats_->smoothed_rtt().Add( rtt_stats_->mean_deviation().Multiply(4)); } QuicByteCount TcpCubicSender::GetCongestionWindow() const { return congestion_window_ * kMaxSegmentSize; } bool TcpCubicSender::InSlowStart() const { return congestion_window_ < slowstart_threshold_; } QuicByteCount TcpCubicSender::GetSlowStartThreshold() const { return slowstart_threshold_ * kMaxSegmentSize; } bool TcpCubicSender::IsCwndLimited(QuicByteCount bytes_in_flight) const { const QuicByteCount congestion_window_bytes = congestion_window_ * kMaxSegmentSize; if (bytes_in_flight >= congestion_window_bytes) { return true; } const QuicByteCount max_burst = kMaxBurstLength * kMaxSegmentSize; const QuicByteCount available_bytes = congestion_window_bytes - bytes_in_flight; const bool slow_start_limited = InSlowStart() && bytes_in_flight > congestion_window_bytes / 2; return slow_start_limited || available_bytes <= max_burst; } bool TcpCubicSender::InRecovery() const { return largest_acked_sequence_number_ <= largest_sent_at_last_cutback_ && largest_acked_sequence_number_ != 0; } // Called when we receive an ack. Normal TCP tracks how many packets one ack // represents, but quic has a separate ack for each packet. void TcpCubicSender::MaybeIncreaseCwnd( QuicPacketSequenceNumber acked_sequence_number, QuicByteCount bytes_in_flight) { LOG_IF(DFATAL, InRecovery()) << "Never increase the CWND during recovery."; if (!IsCwndLimited(bytes_in_flight)) { // We don't update the congestion window unless we are close to using the // window we have available. return; } if (InSlowStart()) { // congestion_window_cnt is the number of acks since last change of snd_cwnd if (congestion_window_ < max_tcp_congestion_window_) { // TCP slow start, exponential growth, increase by one for each ACK. ++congestion_window_; } DVLOG(1) << "Slow start; congestion window: " << congestion_window_ << " slowstart threshold: " << slowstart_threshold_; return; } if (congestion_window_ >= max_tcp_congestion_window_) { return; } // Congestion avoidance if (reno_) { // Classic Reno congestion avoidance. ++congestion_window_count_; // Divide by num_connections to smoothly increase the CWND at a faster // rate than conventional Reno. if (congestion_window_count_ * num_connections_ >= congestion_window_) { ++congestion_window_; congestion_window_count_ = 0; } DVLOG(1) << "Reno; congestion window: " << congestion_window_ << " slowstart threshold: " << slowstart_threshold_ << " congestion window count: " << congestion_window_count_; } else { congestion_window_ = min(max_tcp_congestion_window_, cubic_.CongestionWindowAfterAck( congestion_window_, rtt_stats_->min_rtt())); DVLOG(1) << "Cubic; congestion window: " << congestion_window_ << " slowstart threshold: " << slowstart_threshold_; } } void TcpCubicSender::OnRetransmissionTimeout(bool packets_retransmitted) { largest_sent_at_last_cutback_ = 0; if (!packets_retransmitted) { return; } cubic_.Reset(); hybrid_slow_start_.Restart(); // Only reduce ssthresh once over multiple retransmissions. if (previous_congestion_window_ != 0) { return; } previous_slowstart_threshold_ = slowstart_threshold_; slowstart_threshold_ = congestion_window_ / 2; previous_congestion_window_ = congestion_window_; congestion_window_ = kMinimumCongestionWindow; } void TcpCubicSender::RevertRetransmissionTimeout() { if (previous_congestion_window_ == 0) { LOG(DFATAL) << "No previous congestion window to revert to."; return; } congestion_window_ = previous_congestion_window_; slowstart_threshold_ = previous_slowstart_threshold_; previous_congestion_window_ = 0; } CongestionControlType TcpCubicSender::GetCongestionControlType() const { return reno_ ? kReno : kCubic; } } // namespace net