tcp_bbr: add BBR congestion control

This commit implements a new TCP congestion control algorithm: BBR
(Bottleneck Bandwidth and RTT). A detailed description of BBR will be
published in ACM Queue, Vol. 14 No. 5, September-October 2016, as
"BBR: Congestion-Based Congestion Control".

BBR has significantly increased throughput and reduced latency for
connections on Google's internal backbone networks and google.com and
YouTube Web servers.

BBR requires only changes on the sender side, not in the network or
the receiver side. Thus it can be incrementally deployed on today's
Internet, or in datacenters.

The Internet has predominantly used loss-based congestion control
(largely Reno or CUBIC) since the 1980s, relying on packet loss as the
signal to slow down. While this worked well for many years, loss-based
congestion control is unfortunately out-dated in today's networks. On
today's Internet, loss-based congestion control causes the infamous
bufferbloat problem, often causing seconds of needless queuing delay,
since it fills the bloated buffers in many last-mile links. On today's
high-speed long-haul links using commodity switches with shallow
buffers, loss-based congestion control has abysmal throughput because
it over-reacts to losses caused by transient traffic bursts.

In 1981 Kleinrock and Gale showed that the optimal operating point for
a network maximizes delivered bandwidth while minimizing delay and
loss, not only for single connections but for the network as a
whole. Finding that optimal operating point has been elusive, since
any single network measurement is ambiguous: network measurements are
the result of both bandwidth and propagation delay, and those two
cannot be measured simultaneously.

While it is impossible to disambiguate any single bandwidth or RTT
measurement, a connection's behavior over time tells a clearer
story. BBR uses a measurement strategy designed to resolve this
ambiguity. It combines these measurements with a robust servo loop
using recent control systems advances to implement a distributed
congestion control algorithm that reacts to actual congestion, not
packet loss or transient queue delay, and is designed to converge with
high probability to a point near the optimal operating point.

In a nutshell, BBR creates an explicit model of the network pipe by
sequentially probing the bottleneck bandwidth and RTT. On the arrival
of each ACK, BBR derives the current delivery rate of the last round
trip, and feeds it through a windowed max-filter to estimate the
bottleneck bandwidth. Conversely it uses a windowed min-filter to
estimate the round trip propagation delay. The max-filtered bandwidth
and min-filtered RTT estimates form BBR's model of the network pipe.

Using its model, BBR sets control parameters to govern sending
behavior. The primary control is the pacing rate: BBR applies a gain
multiplier to transmit faster or slower than the observed bottleneck
bandwidth. The conventional congestion window (cwnd) is now the
secondary control; the cwnd is set to a small multiple of the
estimated BDP (bandwidth-delay product) in order to allow full
utilization and bandwidth probing while bounding the potential amount
of queue at the bottleneck.

When a BBR connection starts, it enters STARTUP mode and applies a
high gain to perform an exponential search to quickly probe the
bottleneck bandwidth (doubling its sending rate each round trip, like
slow start). However, instead of continuing until it fills up the
buffer (i.e. a loss), or until delay or ACK spacing reaches some
threshold (like Hystart), it uses its model of the pipe to estimate
when that pipe is full: it estimates the pipe is full when it notices
the estimated bandwidth has stopped growing. At that point it exits
STARTUP and enters DRAIN mode, where it reduces its pacing rate to
drain the queue it estimates it has created.

Then BBR enters steady state. In steady state, PROBE_BW mode cycles
between first pacing faster to probe for more bandwidth, then pacing
slower to drain any queue that created if no more bandwidth was
available, and then cruising at the estimated bandwidth to utilize the
pipe without creating excess queue. Occasionally, on an as-needed
basis, it sends significantly slower to probe for RTT (PROBE_RTT
mode).

BBR has been fully deployed on Google's wide-area backbone networks
and we're experimenting with BBR on Google.com and YouTube on a global
scale.  Replacing CUBIC with BBR has resulted in significant
improvements in network latency and application (RPC, browser, and
video) metrics. For more details please refer to our upcoming ACM
Queue publication.

Example performance results, to illustrate the difference between BBR
and CUBIC:

Resilience to random loss (e.g. from shallow buffers):
  Consider a netperf TCP_STREAM test lasting 30 secs on an emulated
  path with a 10Gbps bottleneck, 100ms RTT, and 1% packet loss
  rate. CUBIC gets 3.27 Mbps, and BBR gets 9150 Mbps (2798x higher).

Low latency with the bloated buffers common in today's last-mile links:
  Consider a netperf TCP_STREAM test lasting 120 secs on an emulated
  path with a 10Mbps bottleneck, 40ms RTT, and 1000-packet bottleneck
  buffer. Both fully utilize the bottleneck bandwidth, but BBR
  achieves this with a median RTT 25x lower (43 ms instead of 1.09
  secs).

Our long-term goal is to improve the congestion control algorithms
used on the Internet. We are hopeful that BBR can help advance the
efforts toward this goal, and motivate the community to do further
research.

Test results, performance evaluations, feedback, and BBR-related
discussions are very welcome in the public e-mail list for BBR:

  https://groups.google.com/forum/#!forum/bbr-dev

NOTE: BBR *must* be used with the fq qdisc ("man tc-fq") with pacing
enabled, since pacing is integral to the BBR design and
implementation. BBR without pacing would not function properly, and
may incur unnecessary high packet loss rates.

Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 896 insertions, 0 deletions

diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c
new file mode 100644
index 000000000000..0ea66c2c9344
--- /dev/null
+++ b/net/ipv4/tcp_bbr.c
@@ -0,0 +1,896 @@
+/* Bottleneck Bandwidth and RTT (BBR) congestion control
+ *
+ * BBR congestion control computes the sending rate based on the delivery
+ * rate (throughput) estimated from ACKs. In a nutshell:
+ *
+ *   On each ACK, update our model of the network path:
+ *      bottleneck_bandwidth = windowed_max(delivered / elapsed, 10 round trips)
+ *      min_rtt = windowed_min(rtt, 10 seconds)
+ *   pacing_rate = pacing_gain * bottleneck_bandwidth
+ *   cwnd = max(cwnd_gain * bottleneck_bandwidth * min_rtt, 4)
+ *
+ * The core algorithm does not react directly to packet losses or delays,
+ * although BBR may adjust the size of next send per ACK when loss is
+ * observed, or adjust the sending rate if it estimates there is a
+ * traffic policer, in order to keep the drop rate reasonable.
+ *
+ * BBR is described in detail in:
+ *   "BBR: Congestion-Based Congestion Control",
+ *   Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh,
+ *   Van Jacobson. ACM Queue, Vol. 14 No. 5, September-October 2016.
+ *
+ * There is a public e-mail list for discussing BBR development and testing:
+ *   https://groups.google.com/forum/#!forum/bbr-dev
+ *
+ * NOTE: BBR *must* be used with the fq qdisc ("man tc-fq") with pacing enabled,
+ * since pacing is integral to the BBR design and implementation.
+ * BBR without pacing would not function properly, and may incur unnecessary
+ * high packet loss rates.
+ */
+#include <linux/module.h>
+#include <net/tcp.h>
+#include <linux/inet_diag.h>
+#include <linux/inet.h>
+#include <linux/random.h>
+#include <linux/win_minmax.h>
+
+/* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth
+ * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps.
+ * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32.
+ * Since the minimum window is >=4 packets, the lower bound isn't
+ * an issue. The upper bound isn't an issue with existing technologies.
+ */
+#define BW_SCALE 24
+#define BW_UNIT (1 << BW_SCALE)
+
+#define BBR_SCALE 8	/* scaling factor for fractions in BBR (e.g. gains) */
+#define BBR_UNIT (1 << BBR_SCALE)
+
+/* BBR has the following modes for deciding how fast to send: */
+enum bbr_mode {
+	BBR_STARTUP,	/* ramp up sending rate rapidly to fill pipe */
+	BBR_DRAIN,	/* drain any queue created during startup */
+	BBR_PROBE_BW,	/* discover, share bw: pace around estimated bw */
+	BBR_PROBE_RTT,	/* cut cwnd to min to probe min_rtt */
+};
+
+/* BBR congestion control block */
+struct bbr {
+	u32	min_rtt_us;	        /* min RTT in min_rtt_win_sec window */
+	u32	min_rtt_stamp;	        /* timestamp of min_rtt_us */
+	u32	probe_rtt_done_stamp;   /* end time for BBR_PROBE_RTT mode */
+	struct minmax bw;	/* Max recent delivery rate in pkts/uS << 24 */
+	u32	rtt_cnt;	    /* count of packet-timed rounds elapsed */
+	u32     next_rtt_delivered; /* scb->tx.delivered at end of round */
+	struct skb_mstamp cycle_mstamp;  /* time of this cycle phase start */
+	u32     mode:3,		     /* current bbr_mode in state machine */
+		prev_ca_state:3,     /* CA state on previous ACK */
+		packet_conservation:1,  /* use packet conservation? */
+		restore_cwnd:1,	     /* decided to revert cwnd to old value */
+		round_start:1,	     /* start of packet-timed tx->ack round? */
+		tso_segs_goal:7,     /* segments we want in each skb we send */
+		idle_restart:1,	     /* restarting after idle? */
+		probe_rtt_round_done:1,  /* a BBR_PROBE_RTT round at 4 pkts? */
+		unused:5,
+		lt_is_sampling:1,    /* taking long-term ("LT") samples now? */
+		lt_rtt_cnt:7,	     /* round trips in long-term interval */
+		lt_use_bw:1;	     /* use lt_bw as our bw estimate? */
+	u32	lt_bw;		     /* LT est delivery rate in pkts/uS << 24 */
+	u32	lt_last_delivered;   /* LT intvl start: tp->delivered */
+	u32	lt_last_stamp;	     /* LT intvl start: tp->delivered_mstamp */
+	u32	lt_last_lost;	     /* LT intvl start: tp->lost */
+	u32	pacing_gain:10,	/* current gain for setting pacing rate */
+		cwnd_gain:10,	/* current gain for setting cwnd */
+		full_bw_cnt:3,	/* number of rounds without large bw gains */
+		cycle_idx:3,	/* current index in pacing_gain cycle array */
+		unused_b:6;
+	u32	prior_cwnd;	/* prior cwnd upon entering loss recovery */
+	u32	full_bw;	/* recent bw, to estimate if pipe is full */
+};
+
+#define CYCLE_LEN	8	/* number of phases in a pacing gain cycle */
+
+/* Window length of bw filter (in rounds): */
+static const int bbr_bw_rtts = CYCLE_LEN + 2;
+/* Window length of min_rtt filter (in sec): */
+static const u32 bbr_min_rtt_win_sec = 10;
+/* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
+static const u32 bbr_probe_rtt_mode_ms = 200;
+/* Skip TSO below the following bandwidth (bits/sec): */
+static const int bbr_min_tso_rate = 1200000;
+
+/* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
+ * that will allow a smoothly increasing pacing rate that will double each RTT
+ * and send the same number of packets per RTT that an un-paced, slow-starting
+ * Reno or CUBIC flow would:
+ */
+static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
+/* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
+ * the queue created in BBR_STARTUP in a single round:
+ */
+static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
+/* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
+static const int bbr_cwnd_gain  = BBR_UNIT * 2;
+/* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
+static const int bbr_pacing_gain[] = {
+	BBR_UNIT * 5 / 4,	/* probe for more available bw */
+	BBR_UNIT * 3 / 4,	/* drain queue and/or yield bw to other flows */
+	BBR_UNIT, BBR_UNIT, BBR_UNIT,	/* cruise at 1.0*bw to utilize pipe, */
+	BBR_UNIT, BBR_UNIT, BBR_UNIT	/* without creating excess queue... */
+};
+/* Randomize the starting gain cycling phase over N phases: */
+static const u32 bbr_cycle_rand = 7;
+
+/* Try to keep at least this many packets in flight, if things go smoothly. For
+ * smooth functioning, a sliding window protocol ACKing every other packet
+ * needs at least 4 packets in flight:
+ */
+static const u32 bbr_cwnd_min_target = 4;
+
+/* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
+/* If bw has increased significantly (1.25x), there may be more bw available: */
+static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
+/* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
+static const u32 bbr_full_bw_cnt = 3;
+
+/* "long-term" ("LT") bandwidth estimator parameters... */
+/* The minimum number of rounds in an LT bw sampling interval: */
+static const u32 bbr_lt_intvl_min_rtts = 4;
+/* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
+static const u32 bbr_lt_loss_thresh = 50;
+/* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
+static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
+/* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
+static const u32 bbr_lt_bw_diff = 4000 / 8;
+/* If we estimate we're policed, use lt_bw for this many round trips: */
+static const u32 bbr_lt_bw_max_rtts = 48;
+
+/* Do we estimate that STARTUP filled the pipe? */
+static bool bbr_full_bw_reached(const struct sock *sk)
+{
+	const struct bbr *bbr = inet_csk_ca(sk);
+
+	return bbr->full_bw_cnt >= bbr_full_bw_cnt;
+}
+
+/* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
+static u32 bbr_max_bw(const struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	return minmax_get(&bbr->bw);
+}
+
+/* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
+static u32 bbr_bw(const struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
+}
+
+/* Return rate in bytes per second, optionally with a gain.
+ * The order here is chosen carefully to avoid overflow of u64. This should
+ * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
+ */
+static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
+{
+	rate *= tcp_mss_to_mtu(sk, tcp_sk(sk)->mss_cache);
+	rate *= gain;
+	rate >>= BBR_SCALE;
+	rate *= USEC_PER_SEC;
+	return rate >> BW_SCALE;
+}
+
+/* Pace using current bw estimate and a gain factor. In order to help drive the
+ * network toward lower queues while maintaining high utilization and low
+ * latency, the average pacing rate aims to be slightly (~1%) lower than the
+ * estimated bandwidth. This is an important aspect of the design. In this
+ * implementation this slightly lower pacing rate is achieved implicitly by not
+ * including link-layer headers in the packet size used for the pacing rate.
+ */
+static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+	u64 rate = bw;
+
+	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
+	rate = min_t(u64, rate, sk->sk_max_pacing_rate);
+	if (bbr->mode != BBR_STARTUP || rate > sk->sk_pacing_rate)
+		sk->sk_pacing_rate = rate;
+}
+
+/* Return count of segments we want in the skbs we send, or 0 for default. */
+static u32 bbr_tso_segs_goal(struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	return bbr->tso_segs_goal;
+}
+
+static void bbr_set_tso_segs_goal(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 min_segs;
+
+	min_segs = sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2;
+	bbr->tso_segs_goal = min(tcp_tso_autosize(sk, tp->mss_cache, min_segs),
+				 0x7FU);
+}
+
+/* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
+static void bbr_save_cwnd(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
+		bbr->prior_cwnd = tp->snd_cwnd;  /* this cwnd is good enough */
+	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
+		bbr->prior_cwnd = max(bbr->prior_cwnd, tp->snd_cwnd);
+}
+
+static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	if (event == CA_EVENT_TX_START && tp->app_limited) {
+		bbr->idle_restart = 1;
+		/* Avoid pointless buffer overflows: pace at est. bw if we don't
+		 * need more speed (we're restarting from idle and app-limited).
+		 */
+		if (bbr->mode == BBR_PROBE_BW)
+			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
+	}
+}
+
+/* Find target cwnd. Right-size the cwnd based on min RTT and the
+ * estimated bottleneck bandwidth:
+ *
+ * cwnd = bw * min_rtt * gain = BDP * gain
+ *
+ * The key factor, gain, controls the amount of queue. While a small gain
+ * builds a smaller queue, it becomes more vulnerable to noise in RTT
+ * measurements (e.g., delayed ACKs or other ACK compression effects). This
+ * noise may cause BBR to under-estimate the rate.
+ *
+ * To achieve full performance in high-speed paths, we budget enough cwnd to
+ * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
+ *   - one skb in sending host Qdisc,
+ *   - one skb in sending host TSO/GSO engine
+ *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
+ * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
+ * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
+ * which allows 2 outstanding 2-packet sequences, to try to keep pipe
+ * full even with ACK-every-other-packet delayed ACKs.
+ */
+static u32 bbr_target_cwnd(struct sock *sk, u32 bw, int gain)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 cwnd;
+	u64 w;
+
+	/* If we've never had a valid RTT sample, cap cwnd at the initial
+	 * default. This should only happen when the connection is not using TCP
+	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
+	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
+	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
+	 */
+	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
+		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
+
+	w = (u64)bw * bbr->min_rtt_us;
+
+	/* Apply a gain to the given value, then remove the BW_SCALE shift. */
+	cwnd = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
+
+	/* Allow enough full-sized skbs in flight to utilize end systems. */
+	cwnd += 3 * bbr->tso_segs_goal;
+
+	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
+	cwnd = (cwnd + 1) & ~1U;
+
+	return cwnd;
+}
+
+/* An optimization in BBR to reduce losses: On the first round of recovery, we
+ * follow the packet conservation principle: send P packets per P packets acked.
+ * After that, we slow-start and send at most 2*P packets per P packets acked.
+ * After recovery finishes, or upon undo, we restore the cwnd we had when
+ * recovery started (capped by the target cwnd based on estimated BDP).
+ *
+ * TODO(ycheng/ncardwell): implement a rate-based approach.
+ */
+static bool bbr_set_cwnd_to_recover_or_restore(
+	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
+	u32 cwnd = tp->snd_cwnd;
+
+	/* An ACK for P pkts should release at most 2*P packets. We do this
+	 * in two steps. First, here we deduct the number of lost packets.
+	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
+	 */
+	if (rs->losses > 0)
+		cwnd = max_t(s32, cwnd - rs->losses, 1);
+
+	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
+		/* Starting 1st round of Recovery, so do packet conservation. */
+		bbr->packet_conservation = 1;
+		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
+		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
+		cwnd = tcp_packets_in_flight(tp) + acked;
+	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
+		/* Exiting loss recovery; restore cwnd saved before recovery. */
+		bbr->restore_cwnd = 1;
+		bbr->packet_conservation = 0;
+	}
+	bbr->prev_ca_state = state;
+
+	if (bbr->restore_cwnd) {
+		/* Restore cwnd after exiting loss recovery or PROBE_RTT. */
+		cwnd = max(cwnd, bbr->prior_cwnd);
+		bbr->restore_cwnd = 0;
+	}
+
+	if (bbr->packet_conservation) {
+		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
+		return true;	/* yes, using packet conservation */
+	}
+	*new_cwnd = cwnd;
+	return false;
+}
+
+/* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
+ * has drawn us down below target), or snap down to target if we're above it.
+ */
+static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
+			 u32 acked, u32 bw, int gain)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 cwnd = 0, target_cwnd = 0;
+
+	if (!acked)
+		return;
+
+	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
+		goto done;
+
+	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
+	target_cwnd = bbr_target_cwnd(sk, bw, gain);
+	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
+		cwnd = min(cwnd + acked, target_cwnd);
+	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
+		cwnd = cwnd + acked;
+	cwnd = max(cwnd, bbr_cwnd_min_target);
+
+done:
+	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);	/* apply global cap */
+	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
+		tp->snd_cwnd = min(tp->snd_cwnd, bbr_cwnd_min_target);
+}
+
+/* End cycle phase if it's time and/or we hit the phase's in-flight target. */
+static bool bbr_is_next_cycle_phase(struct sock *sk,
+				    const struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	bool is_full_length =
+		skb_mstamp_us_delta(&tp->delivered_mstamp, &bbr->cycle_mstamp) >
+		bbr->min_rtt_us;
+	u32 inflight, bw;
+
+	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
+	 * use the pipe without increasing the queue.
+	 */
+	if (bbr->pacing_gain == BBR_UNIT)
+		return is_full_length;		/* just use wall clock time */
+
+	inflight = rs->prior_in_flight;  /* what was in-flight before ACK? */
+	bw = bbr_max_bw(sk);
+
+	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
+	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
+	 * small (e.g. on a LAN). We do not persist if packets are lost, since
+	 * a path with small buffers may not hold that much.
+	 */
+	if (bbr->pacing_gain > BBR_UNIT)
+		return is_full_length &&
+			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
+			 inflight >= bbr_target_cwnd(sk, bw, bbr->pacing_gain));
+
+	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
+	 * probing didn't find more bw. If inflight falls to match BDP then we
+	 * estimate queue is drained; persisting would underutilize the pipe.
+	 */
+	return is_full_length ||
+		inflight <= bbr_target_cwnd(sk, bw, BBR_UNIT);
+}
+
+static void bbr_advance_cycle_phase(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
+	bbr->cycle_mstamp = tp->delivered_mstamp;
+	bbr->pacing_gain = bbr_pacing_gain[bbr->cycle_idx];
+}
+
+/* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
+static void bbr_update_cycle_phase(struct sock *sk,
+				   const struct rate_sample *rs)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	if ((bbr->mode == BBR_PROBE_BW) && !bbr->lt_use_bw &&
+	    bbr_is_next_cycle_phase(sk, rs))
+		bbr_advance_cycle_phase(sk);
+}
+
+static void bbr_reset_startup_mode(struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	bbr->mode = BBR_STARTUP;
+	bbr->pacing_gain = bbr_high_gain;
+	bbr->cwnd_gain	 = bbr_high_gain;
+}
+
+static void bbr_reset_probe_bw_mode(struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	bbr->mode = BBR_PROBE_BW;
+	bbr->pacing_gain = BBR_UNIT;
+	bbr->cwnd_gain = bbr_cwnd_gain;
+	bbr->cycle_idx = CYCLE_LEN - 1 - prandom_u32_max(bbr_cycle_rand);
+	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
+}
+
+static void bbr_reset_mode(struct sock *sk)
+{
+	if (!bbr_full_bw_reached(sk))
+		bbr_reset_startup_mode(sk);
+	else
+		bbr_reset_probe_bw_mode(sk);
+}
+
+/* Start a new long-term sampling interval. */
+static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	bbr->lt_last_stamp = tp->delivered_mstamp.stamp_jiffies;
+	bbr->lt_last_delivered = tp->delivered;
+	bbr->lt_last_lost = tp->lost;
+	bbr->lt_rtt_cnt = 0;
+}
+
+/* Completely reset long-term bandwidth sampling. */
+static void bbr_reset_lt_bw_sampling(struct sock *sk)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	bbr->lt_bw = 0;
+	bbr->lt_use_bw = 0;
+	bbr->lt_is_sampling = false;
+	bbr_reset_lt_bw_sampling_interval(sk);
+}
+
+/* Long-term bw sampling interval is done. Estimate whether we're policed. */
+static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 diff;
+
+	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
+		/* Is new bw close to the lt_bw from the previous interval? */
+		diff = abs(bw - bbr->lt_bw);
+		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
+		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
+		     bbr_lt_bw_diff)) {
+			/* All criteria are met; estimate we're policed. */
+			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
+			bbr->lt_use_bw = 1;
+			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
+			bbr->lt_rtt_cnt = 0;
+			return;
+		}
+	}
+	bbr->lt_bw = bw;
+	bbr_reset_lt_bw_sampling_interval(sk);
+}
+
+/* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
+ * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
+ * explicitly models their policed rate, to reduce unnecessary losses. We
+ * estimate that we're policed if we see 2 consecutive sampling intervals with
+ * consistent throughput and high packet loss. If we think we're being policed,
+ * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
+ */
+static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 lost, delivered;
+	u64 bw;
+	s32 t;
+
+	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
+		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
+		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
+			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
+			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
+		}
+		return;
+	}
+
+	/* Wait for the first loss before sampling, to let the policer exhaust
+	 * its tokens and estimate the steady-state rate allowed by the policer.
+	 * Starting samples earlier includes bursts that over-estimate the bw.
+	 */
+	if (!bbr->lt_is_sampling) {
+		if (!rs->losses)
+			return;
+		bbr_reset_lt_bw_sampling_interval(sk);
+		bbr->lt_is_sampling = true;
+	}
+
+	/* To avoid underestimates, reset sampling if we run out of data. */
+	if (rs->is_app_limited) {
+		bbr_reset_lt_bw_sampling(sk);
+		return;
+	}
+
+	if (bbr->round_start)
+		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
+	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
+		return;		/* sampling interval needs to be longer */
+	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
+		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
+		return;
+	}
+
+	/* End sampling interval when a packet is lost, so we estimate the
+	 * policer tokens were exhausted. Stopping the sampling before the
+	 * tokens are exhausted under-estimates the policed rate.
+	 */
+	if (!rs->losses)
+		return;
+
+	/* Calculate packets lost and delivered in sampling interval. */
+	lost = tp->lost - bbr->lt_last_lost;
+	delivered = tp->delivered - bbr->lt_last_delivered;
+	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
+	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
+		return;
+
+	/* Find average delivery rate in this sampling interval. */
+	t = (s32)(tp->delivered_mstamp.stamp_jiffies - bbr->lt_last_stamp);
+	if (t < 1)
+		return;		/* interval is less than one jiffy, so wait */
+	t = jiffies_to_usecs(t);
+	/* Interval long enough for jiffies_to_usecs() to return a bogus 0? */
+	if (t < 1) {
+		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
+		return;
+	}
+	bw = (u64)delivered * BW_UNIT;
+	do_div(bw, t);
+	bbr_lt_bw_interval_done(sk, bw);
+}
+
+/* Estimate the bandwidth based on how fast packets are delivered */
+static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u64 bw;
+
+	bbr->round_start = 0;
+	if (rs->delivered < 0 || rs->interval_us <= 0)
+		return; /* Not a valid observation */
+
+	/* See if we've reached the next RTT */
+	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
+		bbr->next_rtt_delivered = tp->delivered;
+		bbr->rtt_cnt++;
+		bbr->round_start = 1;
+		bbr->packet_conservation = 0;
+	}
+
+	bbr_lt_bw_sampling(sk, rs);
+
+	/* Divide delivered by the interval to find a (lower bound) bottleneck
+	 * bandwidth sample. Delivered is in packets and interval_us in uS and
+	 * ratio will be <<1 for most connections. So delivered is first scaled.
+	 */
+	bw = (u64)rs->delivered * BW_UNIT;
+	do_div(bw, rs->interval_us);
+
+	/* If this sample is application-limited, it is likely to have a very
+	 * low delivered count that represents application behavior rather than
+	 * the available network rate. Such a sample could drag down estimated
+	 * bw, causing needless slow-down. Thus, to continue to send at the
+	 * last measured network rate, we filter out app-limited samples unless
+	 * they describe the path bw at least as well as our bw model.
+	 *
+	 * So the goal during app-limited phase is to proceed with the best
+	 * network rate no matter how long. We automatically leave this
+	 * phase when app writes faster than the network can deliver :)
+	 */
+	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
+		/* Incorporate new sample into our max bw filter. */
+		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
+	}
+}
+
+/* Estimate when the pipe is full, using the change in delivery rate: BBR
+ * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
+ * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
+ * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
+ * higher rwin, 3: we get higher delivery rate samples. Or transient
+ * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
+ * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
+ */
+static void bbr_check_full_bw_reached(struct sock *sk,
+				      const struct rate_sample *rs)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 bw_thresh;
+
+	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
+		return;
+
+	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
+	if (bbr_max_bw(sk) >= bw_thresh) {
+		bbr->full_bw = bbr_max_bw(sk);
+		bbr->full_bw_cnt = 0;
+		return;
+	}
+	++bbr->full_bw_cnt;
+}
+
+/* If pipe is probably full, drain the queue and then enter steady-state. */
+static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
+		bbr->mode = BBR_DRAIN;	/* drain queue we created */
+		bbr->pacing_gain = bbr_drain_gain;	/* pace slow to drain */
+		bbr->cwnd_gain = bbr_high_gain;	/* maintain cwnd */
+	}	/* fall through to check if in-flight is already small: */
+	if (bbr->mode == BBR_DRAIN &&
+	    tcp_packets_in_flight(tcp_sk(sk)) <=
+	    bbr_target_cwnd(sk, bbr_max_bw(sk), BBR_UNIT))
+		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
+}
+
+/* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
+ * periodically drain the bottleneck queue, to converge to measure the true
+ * min_rtt (unloaded propagation delay). This allows the flows to keep queues
+ * small (reducing queuing delay and packet loss) and achieve fairness among
+ * BBR flows.
+ *
+ * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
+ * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
+ * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
+ * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
+ * re-enter the previous mode. BBR uses 200ms to approximately bound the
+ * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
+ *
+ * Note that flows need only pay 2% if they are busy sending over the last 10
+ * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
+ * natural silences or low-rate periods within 10 seconds where the rate is low
+ * enough for long enough to drain its queue in the bottleneck. We pick up
+ * these min RTT measurements opportunistically with our min_rtt filter. :-)
+ */
+static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	bool filter_expired;
+
+	/* Track min RTT seen in the min_rtt_win_sec filter window: */
+	filter_expired = after(tcp_time_stamp,
+			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
+	if (rs->rtt_us >= 0 &&
+	    (rs->rtt_us <= bbr->min_rtt_us || filter_expired)) {
+		bbr->min_rtt_us = rs->rtt_us;
+		bbr->min_rtt_stamp = tcp_time_stamp;
+	}
+
+	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
+	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
+		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
+		bbr->pacing_gain = BBR_UNIT;
+		bbr->cwnd_gain = BBR_UNIT;
+		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
+		bbr->probe_rtt_done_stamp = 0;
+	}
+
+	if (bbr->mode == BBR_PROBE_RTT) {
+		/* Ignore low rate samples during this mode. */
+		tp->app_limited =
+			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
+		/* Maintain min packets in flight for max(200 ms, 1 round). */
+		if (!bbr->probe_rtt_done_stamp &&
+		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
+			bbr->probe_rtt_done_stamp = tcp_time_stamp +
+				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
+			bbr->probe_rtt_round_done = 0;
+			bbr->next_rtt_delivered = tp->delivered;
+		} else if (bbr->probe_rtt_done_stamp) {
+			if (bbr->round_start)
+				bbr->probe_rtt_round_done = 1;
+			if (bbr->probe_rtt_round_done &&
+			    after(tcp_time_stamp, bbr->probe_rtt_done_stamp)) {
+				bbr->min_rtt_stamp = tcp_time_stamp;
+				bbr->restore_cwnd = 1;  /* snap to prior_cwnd */
+				bbr_reset_mode(sk);
+			}
+		}
+	}
+	bbr->idle_restart = 0;
+}
+
+static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
+{
+	bbr_update_bw(sk, rs);
+	bbr_update_cycle_phase(sk, rs);
+	bbr_check_full_bw_reached(sk, rs);
+	bbr_check_drain(sk, rs);
+	bbr_update_min_rtt(sk, rs);
+}
+
+static void bbr_main(struct sock *sk, const struct rate_sample *rs)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+	u32 bw;
+
+	bbr_update_model(sk, rs);
+
+	bw = bbr_bw(sk);
+	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
+	bbr_set_tso_segs_goal(sk);
+	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
+}
+
+static void bbr_init(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct bbr *bbr = inet_csk_ca(sk);
+	u64 bw;
+
+	bbr->prior_cwnd = 0;
+	bbr->tso_segs_goal = 0;	 /* default segs per skb until first ACK */
+	bbr->rtt_cnt = 0;
+	bbr->next_rtt_delivered = 0;
+	bbr->prev_ca_state = TCP_CA_Open;
+	bbr->packet_conservation = 0;
+
+	bbr->probe_rtt_done_stamp = 0;
+	bbr->probe_rtt_round_done = 0;
+	bbr->min_rtt_us = tcp_min_rtt(tp);
+	bbr->min_rtt_stamp = tcp_time_stamp;
+
+	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
+
+	/* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
+	bw = (u64)tp->snd_cwnd * BW_UNIT;
+	do_div(bw, (tp->srtt_us >> 3) ? : USEC_PER_MSEC);
+	sk->sk_pacing_rate = 0;		/* force an update of sk_pacing_rate */
+	bbr_set_pacing_rate(sk, bw, bbr_high_gain);
+
+	bbr->restore_cwnd = 0;
+	bbr->round_start = 0;
+	bbr->idle_restart = 0;
+	bbr->full_bw = 0;
+	bbr->full_bw_cnt = 0;
+	bbr->cycle_mstamp.v64 = 0;
+	bbr->cycle_idx = 0;
+	bbr_reset_lt_bw_sampling(sk);
+	bbr_reset_startup_mode(sk);
+}
+
+static u32 bbr_sndbuf_expand(struct sock *sk)
+{
+	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
+	return 3;
+}
+
+/* In theory BBR does not need to undo the cwnd since it does not
+ * always reduce cwnd on losses (see bbr_main()). Keep it for now.
+ */
+static u32 bbr_undo_cwnd(struct sock *sk)
+{
+	return tcp_sk(sk)->snd_cwnd;
+}
+
+/* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
+static u32 bbr_ssthresh(struct sock *sk)
+{
+	bbr_save_cwnd(sk);
+	return TCP_INFINITE_SSTHRESH;	 /* BBR does not use ssthresh */
+}
+
+static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
+			   union tcp_cc_info *info)
+{
+	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
+	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
+		struct tcp_sock *tp = tcp_sk(sk);
+		struct bbr *bbr = inet_csk_ca(sk);
+		u64 bw = bbr_bw(sk);
+
+		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
+		memset(&info->bbr, 0, sizeof(info->bbr));
+		info->bbr.bbr_bw_lo		= (u32)bw;
+		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
+		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
+		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
+		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
+		*attr = INET_DIAG_BBRINFO;
+		return sizeof(info->bbr);
+	}
+	return 0;
+}
+
+static void bbr_set_state(struct sock *sk, u8 new_state)
+{
+	struct bbr *bbr = inet_csk_ca(sk);
+
+	if (new_state == TCP_CA_Loss) {
+		struct rate_sample rs = { .losses = 1 };
+
+		bbr->prev_ca_state = TCP_CA_Loss;
+		bbr->full_bw = 0;
+		bbr->round_start = 1;	/* treat RTO like end of a round */
+		bbr_lt_bw_sampling(sk, &rs);
+	}
+}
+
+static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
+	.flags		= TCP_CONG_NON_RESTRICTED,
+	.name		= "bbr",
+	.owner		= THIS_MODULE,
+	.init		= bbr_init,
+	.cong_control	= bbr_main,
+	.sndbuf_expand	= bbr_sndbuf_expand,
+	.undo_cwnd	= bbr_undo_cwnd,
+	.cwnd_event	= bbr_cwnd_event,
+	.ssthresh	= bbr_ssthresh,
+	.tso_segs_goal	= bbr_tso_segs_goal,
+	.get_info	= bbr_get_info,
+	.set_state	= bbr_set_state,
+};
+
+static int __init bbr_register(void)
+{
+	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
+	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
+}
+
+static void __exit bbr_unregister(void)
+{
+	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
+}
+
+module_init(bbr_register);
+module_exit(bbr_unregister);
+
+MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
+MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
+MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
+MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");