xref: /linux/net/ipv4/tcp_bbr.c (revision 621cde16e49b3ecf7d59a8106a20aaebfb4a59a9)
10f8782eaSNeal Cardwell /* Bottleneck Bandwidth and RTT (BBR) congestion control
20f8782eaSNeal Cardwell  *
30f8782eaSNeal Cardwell  * BBR congestion control computes the sending rate based on the delivery
40f8782eaSNeal Cardwell  * rate (throughput) estimated from ACKs. In a nutshell:
50f8782eaSNeal Cardwell  *
60f8782eaSNeal Cardwell  *   On each ACK, update our model of the network path:
70f8782eaSNeal Cardwell  *      bottleneck_bandwidth = windowed_max(delivered / elapsed, 10 round trips)
80f8782eaSNeal Cardwell  *      min_rtt = windowed_min(rtt, 10 seconds)
90f8782eaSNeal Cardwell  *   pacing_rate = pacing_gain * bottleneck_bandwidth
100f8782eaSNeal Cardwell  *   cwnd = max(cwnd_gain * bottleneck_bandwidth * min_rtt, 4)
110f8782eaSNeal Cardwell  *
120f8782eaSNeal Cardwell  * The core algorithm does not react directly to packet losses or delays,
130f8782eaSNeal Cardwell  * although BBR may adjust the size of next send per ACK when loss is
140f8782eaSNeal Cardwell  * observed, or adjust the sending rate if it estimates there is a
150f8782eaSNeal Cardwell  * traffic policer, in order to keep the drop rate reasonable.
160f8782eaSNeal Cardwell  *
179b9375b5SNeal Cardwell  * Here is a state transition diagram for BBR:
189b9375b5SNeal Cardwell  *
199b9375b5SNeal Cardwell  *             |
209b9375b5SNeal Cardwell  *             V
219b9375b5SNeal Cardwell  *    +---> STARTUP  ----+
229b9375b5SNeal Cardwell  *    |        |         |
239b9375b5SNeal Cardwell  *    |        V         |
249b9375b5SNeal Cardwell  *    |      DRAIN   ----+
259b9375b5SNeal Cardwell  *    |        |         |
269b9375b5SNeal Cardwell  *    |        V         |
279b9375b5SNeal Cardwell  *    +---> PROBE_BW ----+
289b9375b5SNeal Cardwell  *    |      ^    |      |
299b9375b5SNeal Cardwell  *    |      |    |      |
309b9375b5SNeal Cardwell  *    |      +----+      |
319b9375b5SNeal Cardwell  *    |                  |
329b9375b5SNeal Cardwell  *    +---- PROBE_RTT <--+
339b9375b5SNeal Cardwell  *
349b9375b5SNeal Cardwell  * A BBR flow starts in STARTUP, and ramps up its sending rate quickly.
359b9375b5SNeal Cardwell  * When it estimates the pipe is full, it enters DRAIN to drain the queue.
369b9375b5SNeal Cardwell  * In steady state a BBR flow only uses PROBE_BW and PROBE_RTT.
379b9375b5SNeal Cardwell  * A long-lived BBR flow spends the vast majority of its time remaining
389b9375b5SNeal Cardwell  * (repeatedly) in PROBE_BW, fully probing and utilizing the pipe's bandwidth
399b9375b5SNeal Cardwell  * in a fair manner, with a small, bounded queue. *If* a flow has been
409b9375b5SNeal Cardwell  * continuously sending for the entire min_rtt window, and hasn't seen an RTT
419b9375b5SNeal Cardwell  * sample that matches or decreases its min_rtt estimate for 10 seconds, then
429b9375b5SNeal Cardwell  * it briefly enters PROBE_RTT to cut inflight to a minimum value to re-probe
439b9375b5SNeal Cardwell  * the path's two-way propagation delay (min_rtt). When exiting PROBE_RTT, if
449b9375b5SNeal Cardwell  * we estimated that we reached the full bw of the pipe then we enter PROBE_BW;
459b9375b5SNeal Cardwell  * otherwise we enter STARTUP to try to fill the pipe.
469b9375b5SNeal Cardwell  *
470f8782eaSNeal Cardwell  * BBR is described in detail in:
480f8782eaSNeal Cardwell  *   "BBR: Congestion-Based Congestion Control",
490f8782eaSNeal Cardwell  *   Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh,
500f8782eaSNeal Cardwell  *   Van Jacobson. ACM Queue, Vol. 14 No. 5, September-October 2016.
510f8782eaSNeal Cardwell  *
520f8782eaSNeal Cardwell  * There is a public e-mail list for discussing BBR development and testing:
530f8782eaSNeal Cardwell  *   https://groups.google.com/forum/#!forum/bbr-dev
540f8782eaSNeal Cardwell  *
55218af599SEric Dumazet  * NOTE: BBR might be used with the fq qdisc ("man tc-fq") with pacing enabled,
56218af599SEric Dumazet  * otherwise TCP stack falls back to an internal pacing using one high
57218af599SEric Dumazet  * resolution timer per TCP socket and may use more resources.
580f8782eaSNeal Cardwell  */
590e32dfc8SKumar Kartikeya Dwivedi #include <linux/btf.h>
600e32dfc8SKumar Kartikeya Dwivedi #include <linux/btf_ids.h>
610f8782eaSNeal Cardwell #include <linux/module.h>
620f8782eaSNeal Cardwell #include <net/tcp.h>
630f8782eaSNeal Cardwell #include <linux/inet_diag.h>
640f8782eaSNeal Cardwell #include <linux/inet.h>
650f8782eaSNeal Cardwell #include <linux/random.h>
660f8782eaSNeal Cardwell #include <linux/win_minmax.h>
670f8782eaSNeal Cardwell 
680f8782eaSNeal Cardwell /* Scale factor for rate in pkt/uSec unit to avoid truncation in bandwidth
690f8782eaSNeal Cardwell  * estimation. The rate unit ~= (1500 bytes / 1 usec / 2^24) ~= 715 bps.
700f8782eaSNeal Cardwell  * This handles bandwidths from 0.06pps (715bps) to 256Mpps (3Tbps) in a u32.
710f8782eaSNeal Cardwell  * Since the minimum window is >=4 packets, the lower bound isn't
720f8782eaSNeal Cardwell  * an issue. The upper bound isn't an issue with existing technologies.
730f8782eaSNeal Cardwell  */
740f8782eaSNeal Cardwell #define BW_SCALE 24
750f8782eaSNeal Cardwell #define BW_UNIT (1 << BW_SCALE)
760f8782eaSNeal Cardwell 
770f8782eaSNeal Cardwell #define BBR_SCALE 8	/* scaling factor for fractions in BBR (e.g. gains) */
780f8782eaSNeal Cardwell #define BBR_UNIT (1 << BBR_SCALE)
790f8782eaSNeal Cardwell 
800f8782eaSNeal Cardwell /* BBR has the following modes for deciding how fast to send: */
810f8782eaSNeal Cardwell enum bbr_mode {
820f8782eaSNeal Cardwell 	BBR_STARTUP,	/* ramp up sending rate rapidly to fill pipe */
830f8782eaSNeal Cardwell 	BBR_DRAIN,	/* drain any queue created during startup */
840f8782eaSNeal Cardwell 	BBR_PROBE_BW,	/* discover, share bw: pace around estimated bw */
859b9375b5SNeal Cardwell 	BBR_PROBE_RTT,	/* cut inflight to min to probe min_rtt */
860f8782eaSNeal Cardwell };
870f8782eaSNeal Cardwell 
880f8782eaSNeal Cardwell /* BBR congestion control block */
890f8782eaSNeal Cardwell struct bbr {
900f8782eaSNeal Cardwell 	u32	min_rtt_us;	        /* min RTT in min_rtt_win_sec window */
910f8782eaSNeal Cardwell 	u32	min_rtt_stamp;	        /* timestamp of min_rtt_us */
920f8782eaSNeal Cardwell 	u32	probe_rtt_done_stamp;   /* end time for BBR_PROBE_RTT mode */
930f8782eaSNeal Cardwell 	struct minmax bw;	/* Max recent delivery rate in pkts/uS << 24 */
940f8782eaSNeal Cardwell 	u32	rtt_cnt;	    /* count of packet-timed rounds elapsed */
950f8782eaSNeal Cardwell 	u32     next_rtt_delivered; /* scb->tx.delivered at end of round */
969a568de4SEric Dumazet 	u64	cycle_mstamp;	     /* time of this cycle phase start */
970f8782eaSNeal Cardwell 	u32     mode:3,		     /* current bbr_mode in state machine */
980f8782eaSNeal Cardwell 		prev_ca_state:3,     /* CA state on previous ACK */
990f8782eaSNeal Cardwell 		packet_conservation:1,  /* use packet conservation? */
1000f8782eaSNeal Cardwell 		round_start:1,	     /* start of packet-timed tx->ack round? */
1010f8782eaSNeal Cardwell 		idle_restart:1,	     /* restarting after idle? */
1020f8782eaSNeal Cardwell 		probe_rtt_round_done:1,  /* a BBR_PROBE_RTT round at 4 pkts? */
103fb998862SKevin Yang 		unused:13,
1040f8782eaSNeal Cardwell 		lt_is_sampling:1,    /* taking long-term ("LT") samples now? */
1050f8782eaSNeal Cardwell 		lt_rtt_cnt:7,	     /* round trips in long-term interval */
1060f8782eaSNeal Cardwell 		lt_use_bw:1;	     /* use lt_bw as our bw estimate? */
1070f8782eaSNeal Cardwell 	u32	lt_bw;		     /* LT est delivery rate in pkts/uS << 24 */
1080f8782eaSNeal Cardwell 	u32	lt_last_delivered;   /* LT intvl start: tp->delivered */
1090f8782eaSNeal Cardwell 	u32	lt_last_stamp;	     /* LT intvl start: tp->delivered_mstamp */
1100f8782eaSNeal Cardwell 	u32	lt_last_lost;	     /* LT intvl start: tp->lost */
1110f8782eaSNeal Cardwell 	u32	pacing_gain:10,	/* current gain for setting pacing rate */
1120f8782eaSNeal Cardwell 		cwnd_gain:10,	/* current gain for setting cwnd */
113c589e69bSNeal Cardwell 		full_bw_reached:1,   /* reached full bw in Startup? */
114c589e69bSNeal Cardwell 		full_bw_cnt:2,	/* number of rounds without large bw gains */
1150f8782eaSNeal Cardwell 		cycle_idx:3,	/* current index in pacing_gain cycle array */
11632984565SNeal Cardwell 		has_seen_rtt:1, /* have we seen an RTT sample yet? */
11732984565SNeal Cardwell 		unused_b:5;
1180f8782eaSNeal Cardwell 	u32	prior_cwnd;	/* prior cwnd upon entering loss recovery */
1190f8782eaSNeal Cardwell 	u32	full_bw;	/* recent bw, to estimate if pipe is full */
12078dc70ebSPriyaranjan Jha 
12178dc70ebSPriyaranjan Jha 	/* For tracking ACK aggregation: */
12278dc70ebSPriyaranjan Jha 	u64	ack_epoch_mstamp;	/* start of ACK sampling epoch */
12378dc70ebSPriyaranjan Jha 	u16	extra_acked[2];		/* max excess data ACKed in epoch */
12478dc70ebSPriyaranjan Jha 	u32	ack_epoch_acked:20,	/* packets (S)ACKed in sampling epoch */
12578dc70ebSPriyaranjan Jha 		extra_acked_win_rtts:5,	/* age of extra_acked, in round trips */
12678dc70ebSPriyaranjan Jha 		extra_acked_win_idx:1,	/* current index in extra_acked array */
12778dc70ebSPriyaranjan Jha 		unused_c:6;
1280f8782eaSNeal Cardwell };
1290f8782eaSNeal Cardwell 
1300f8782eaSNeal Cardwell #define CYCLE_LEN	8	/* number of phases in a pacing gain cycle */
1310f8782eaSNeal Cardwell 
1320f8782eaSNeal Cardwell /* Window length of bw filter (in rounds): */
1330f8782eaSNeal Cardwell static const int bbr_bw_rtts = CYCLE_LEN + 2;
1340f8782eaSNeal Cardwell /* Window length of min_rtt filter (in sec): */
1350f8782eaSNeal Cardwell static const u32 bbr_min_rtt_win_sec = 10;
1360f8782eaSNeal Cardwell /* Minimum time (in ms) spent at bbr_cwnd_min_target in BBR_PROBE_RTT mode: */
1370f8782eaSNeal Cardwell static const u32 bbr_probe_rtt_mode_ms = 200;
1380f8782eaSNeal Cardwell /* Skip TSO below the following bandwidth (bits/sec): */
1390f8782eaSNeal Cardwell static const int bbr_min_tso_rate = 1200000;
1400f8782eaSNeal Cardwell 
1411106a5adSNeal Cardwell /* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck.
1421106a5adSNeal Cardwell  * In order to help drive the network toward lower queues and low latency while
1431106a5adSNeal Cardwell  * maintaining high utilization, the average pacing rate aims to be slightly
1441106a5adSNeal Cardwell  * lower than the estimated bandwidth. This is an important aspect of the
1451106a5adSNeal Cardwell  * design.
1461106a5adSNeal Cardwell  */
14797ec3eb3SNeal Cardwell static const int bbr_pacing_margin_percent = 1;
148ab408b6dSEric Dumazet 
1490f8782eaSNeal Cardwell /* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain
1500f8782eaSNeal Cardwell  * that will allow a smoothly increasing pacing rate that will double each RTT
1510f8782eaSNeal Cardwell  * and send the same number of packets per RTT that an un-paced, slow-starting
1520f8782eaSNeal Cardwell  * Reno or CUBIC flow would:
1530f8782eaSNeal Cardwell  */
1540f8782eaSNeal Cardwell static const int bbr_high_gain  = BBR_UNIT * 2885 / 1000 + 1;
1550f8782eaSNeal Cardwell /* The pacing gain of 1/high_gain in BBR_DRAIN is calculated to typically drain
1560f8782eaSNeal Cardwell  * the queue created in BBR_STARTUP in a single round:
1570f8782eaSNeal Cardwell  */
1580f8782eaSNeal Cardwell static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885;
1590f8782eaSNeal Cardwell /* The gain for deriving steady-state cwnd tolerates delayed/stretched ACKs: */
1600f8782eaSNeal Cardwell static const int bbr_cwnd_gain  = BBR_UNIT * 2;
1610f8782eaSNeal Cardwell /* The pacing_gain values for the PROBE_BW gain cycle, to discover/share bw: */
1620f8782eaSNeal Cardwell static const int bbr_pacing_gain[] = {
1630f8782eaSNeal Cardwell 	BBR_UNIT * 5 / 4,	/* probe for more available bw */
1640f8782eaSNeal Cardwell 	BBR_UNIT * 3 / 4,	/* drain queue and/or yield bw to other flows */
1650f8782eaSNeal Cardwell 	BBR_UNIT, BBR_UNIT, BBR_UNIT,	/* cruise at 1.0*bw to utilize pipe, */
1660f8782eaSNeal Cardwell 	BBR_UNIT, BBR_UNIT, BBR_UNIT	/* without creating excess queue... */
1670f8782eaSNeal Cardwell };
1680f8782eaSNeal Cardwell /* Randomize the starting gain cycling phase over N phases: */
1690f8782eaSNeal Cardwell static const u32 bbr_cycle_rand = 7;
1700f8782eaSNeal Cardwell 
1710f8782eaSNeal Cardwell /* Try to keep at least this many packets in flight, if things go smoothly. For
1720f8782eaSNeal Cardwell  * smooth functioning, a sliding window protocol ACKing every other packet
1730f8782eaSNeal Cardwell  * needs at least 4 packets in flight:
1740f8782eaSNeal Cardwell  */
1750f8782eaSNeal Cardwell static const u32 bbr_cwnd_min_target = 4;
1760f8782eaSNeal Cardwell 
1770f8782eaSNeal Cardwell /* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */
1780f8782eaSNeal Cardwell /* If bw has increased significantly (1.25x), there may be more bw available: */
1790f8782eaSNeal Cardwell static const u32 bbr_full_bw_thresh = BBR_UNIT * 5 / 4;
1800f8782eaSNeal Cardwell /* But after 3 rounds w/o significant bw growth, estimate pipe is full: */
1810f8782eaSNeal Cardwell static const u32 bbr_full_bw_cnt = 3;
1820f8782eaSNeal Cardwell 
1830f8782eaSNeal Cardwell /* "long-term" ("LT") bandwidth estimator parameters... */
1840f8782eaSNeal Cardwell /* The minimum number of rounds in an LT bw sampling interval: */
1850f8782eaSNeal Cardwell static const u32 bbr_lt_intvl_min_rtts = 4;
1860f8782eaSNeal Cardwell /* If lost/delivered ratio > 20%, interval is "lossy" and we may be policed: */
1870f8782eaSNeal Cardwell static const u32 bbr_lt_loss_thresh = 50;
1880f8782eaSNeal Cardwell /* If 2 intervals have a bw ratio <= 1/8, their bw is "consistent": */
1890f8782eaSNeal Cardwell static const u32 bbr_lt_bw_ratio = BBR_UNIT / 8;
1900f8782eaSNeal Cardwell /* If 2 intervals have a bw diff <= 4 Kbit/sec their bw is "consistent": */
1910f8782eaSNeal Cardwell static const u32 bbr_lt_bw_diff = 4000 / 8;
1920f8782eaSNeal Cardwell /* If we estimate we're policed, use lt_bw for this many round trips: */
1930f8782eaSNeal Cardwell static const u32 bbr_lt_bw_max_rtts = 48;
1940f8782eaSNeal Cardwell 
19578dc70ebSPriyaranjan Jha /* Gain factor for adding extra_acked to target cwnd: */
19678dc70ebSPriyaranjan Jha static const int bbr_extra_acked_gain = BBR_UNIT;
19778dc70ebSPriyaranjan Jha /* Window length of extra_acked window. */
19878dc70ebSPriyaranjan Jha static const u32 bbr_extra_acked_win_rtts = 5;
19978dc70ebSPriyaranjan Jha /* Max allowed val for ack_epoch_acked, after which sampling epoch is reset */
20078dc70ebSPriyaranjan Jha static const u32 bbr_ack_epoch_acked_reset_thresh = 1U << 20;
20178dc70ebSPriyaranjan Jha /* Time period for clamping cwnd increment due to ack aggregation */
20278dc70ebSPriyaranjan Jha static const u32 bbr_extra_acked_max_us = 100 * 1000;
20378dc70ebSPriyaranjan Jha 
2045490b32dSKevin Yang static void bbr_check_probe_rtt_done(struct sock *sk);
2055490b32dSKevin Yang 
2060f8782eaSNeal Cardwell /* Do we estimate that STARTUP filled the pipe? */
bbr_full_bw_reached(const struct sock * sk)2070f8782eaSNeal Cardwell static bool bbr_full_bw_reached(const struct sock *sk)
2080f8782eaSNeal Cardwell {
2090f8782eaSNeal Cardwell 	const struct bbr *bbr = inet_csk_ca(sk);
2100f8782eaSNeal Cardwell 
211c589e69bSNeal Cardwell 	return bbr->full_bw_reached;
2120f8782eaSNeal Cardwell }
2130f8782eaSNeal Cardwell 
2140f8782eaSNeal Cardwell /* Return the windowed max recent bandwidth sample, in pkts/uS << BW_SCALE. */
bbr_max_bw(const struct sock * sk)2150f8782eaSNeal Cardwell static u32 bbr_max_bw(const struct sock *sk)
2160f8782eaSNeal Cardwell {
2170f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
2180f8782eaSNeal Cardwell 
2190f8782eaSNeal Cardwell 	return minmax_get(&bbr->bw);
2200f8782eaSNeal Cardwell }
2210f8782eaSNeal Cardwell 
2220f8782eaSNeal Cardwell /* Return the estimated bandwidth of the path, in pkts/uS << BW_SCALE. */
bbr_bw(const struct sock * sk)2230f8782eaSNeal Cardwell static u32 bbr_bw(const struct sock *sk)
2240f8782eaSNeal Cardwell {
2250f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
2260f8782eaSNeal Cardwell 
2270f8782eaSNeal Cardwell 	return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk);
2280f8782eaSNeal Cardwell }
2290f8782eaSNeal Cardwell 
23078dc70ebSPriyaranjan Jha /* Return maximum extra acked in past k-2k round trips,
23178dc70ebSPriyaranjan Jha  * where k = bbr_extra_acked_win_rtts.
23278dc70ebSPriyaranjan Jha  */
bbr_extra_acked(const struct sock * sk)23378dc70ebSPriyaranjan Jha static u16 bbr_extra_acked(const struct sock *sk)
23478dc70ebSPriyaranjan Jha {
23578dc70ebSPriyaranjan Jha 	struct bbr *bbr = inet_csk_ca(sk);
23678dc70ebSPriyaranjan Jha 
23778dc70ebSPriyaranjan Jha 	return max(bbr->extra_acked[0], bbr->extra_acked[1]);
23878dc70ebSPriyaranjan Jha }
23978dc70ebSPriyaranjan Jha 
2400f8782eaSNeal Cardwell /* Return rate in bytes per second, optionally with a gain.
2410f8782eaSNeal Cardwell  * The order here is chosen carefully to avoid overflow of u64. This should
2420f8782eaSNeal Cardwell  * work for input rates of up to 2.9Tbit/sec and gain of 2.89x.
2430f8782eaSNeal Cardwell  */
bbr_rate_bytes_per_sec(struct sock * sk,u64 rate,int gain)2440f8782eaSNeal Cardwell static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain)
2450f8782eaSNeal Cardwell {
246cadefe5fSEric Dumazet 	unsigned int mss = tcp_sk(sk)->mss_cache;
247cadefe5fSEric Dumazet 
248cadefe5fSEric Dumazet 	rate *= mss;
2490f8782eaSNeal Cardwell 	rate *= gain;
2500f8782eaSNeal Cardwell 	rate >>= BBR_SCALE;
25197ec3eb3SNeal Cardwell 	rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent);
2520f8782eaSNeal Cardwell 	return rate >> BW_SCALE;
2530f8782eaSNeal Cardwell }
2540f8782eaSNeal Cardwell 
255f19fd62dSNeal Cardwell /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */
bbr_bw_to_pacing_rate(struct sock * sk,u32 bw,int gain)25676a9ebe8SEric Dumazet static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain)
257f19fd62dSNeal Cardwell {
258f19fd62dSNeal Cardwell 	u64 rate = bw;
259f19fd62dSNeal Cardwell 
260f19fd62dSNeal Cardwell 	rate = bbr_rate_bytes_per_sec(sk, rate, gain);
26128b24f90SEric Dumazet 	rate = min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate));
262f19fd62dSNeal Cardwell 	return rate;
263f19fd62dSNeal Cardwell }
264f19fd62dSNeal Cardwell 
26579135b89SNeal Cardwell /* Initialize pacing rate to: high_gain * init_cwnd / RTT. */
bbr_init_pacing_rate_from_rtt(struct sock * sk)26679135b89SNeal Cardwell static void bbr_init_pacing_rate_from_rtt(struct sock *sk)
26779135b89SNeal Cardwell {
26879135b89SNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
26932984565SNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
27079135b89SNeal Cardwell 	u64 bw;
27179135b89SNeal Cardwell 	u32 rtt_us;
27279135b89SNeal Cardwell 
27379135b89SNeal Cardwell 	if (tp->srtt_us) {		/* any RTT sample yet? */
27479135b89SNeal Cardwell 		rtt_us = max(tp->srtt_us >> 3, 1U);
27532984565SNeal Cardwell 		bbr->has_seen_rtt = 1;
27679135b89SNeal Cardwell 	} else {			 /* no RTT sample yet */
27779135b89SNeal Cardwell 		rtt_us = USEC_PER_MSEC;	 /* use nominal default RTT */
27879135b89SNeal Cardwell 	}
27940570375SEric Dumazet 	bw = (u64)tcp_snd_cwnd(tp) * BW_UNIT;
28079135b89SNeal Cardwell 	do_div(bw, rtt_us);
28128b24f90SEric Dumazet 	WRITE_ONCE(sk->sk_pacing_rate,
28228b24f90SEric Dumazet 		   bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain));
28379135b89SNeal Cardwell }
28479135b89SNeal Cardwell 
2851106a5adSNeal Cardwell /* Pace using current bw estimate and a gain factor. */
bbr_set_pacing_rate(struct sock * sk,u32 bw,int gain)2860f8782eaSNeal Cardwell static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain)
2870f8782eaSNeal Cardwell {
28832984565SNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
28932984565SNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
29076a9ebe8SEric Dumazet 	unsigned long rate = bbr_bw_to_pacing_rate(sk, bw, gain);
2910f8782eaSNeal Cardwell 
29232984565SNeal Cardwell 	if (unlikely(!bbr->has_seen_rtt && tp->srtt_us))
29332984565SNeal Cardwell 		bbr_init_pacing_rate_from_rtt(sk);
29428b24f90SEric Dumazet 	if (bbr_full_bw_reached(sk) || rate > READ_ONCE(sk->sk_pacing_rate))
29528b24f90SEric Dumazet 		WRITE_ONCE(sk->sk_pacing_rate, rate);
2960f8782eaSNeal Cardwell }
2970f8782eaSNeal Cardwell 
298dcb8c9b4SEric Dumazet /* override sysctl_tcp_min_tso_segs */
bbr_min_tso_segs(struct sock * sk)299400031e0SDavid Vernet __bpf_kfunc static u32 bbr_min_tso_segs(struct sock *sk)
3000f8782eaSNeal Cardwell {
30128b24f90SEric Dumazet 	return READ_ONCE(sk->sk_pacing_rate) < (bbr_min_tso_rate >> 3) ? 1 : 2;
3020f8782eaSNeal Cardwell }
3030f8782eaSNeal Cardwell 
bbr_tso_segs_goal(struct sock * sk)30471abf467SEric Dumazet static u32 bbr_tso_segs_goal(struct sock *sk)
3050f8782eaSNeal Cardwell {
3060f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
307dcb8c9b4SEric Dumazet 	u32 segs, bytes;
3080f8782eaSNeal Cardwell 
309dcb8c9b4SEric Dumazet 	/* Sort of tcp_tso_autosize() but ignoring
310dcb8c9b4SEric Dumazet 	 * driver provided sk_gso_max_size.
311dcb8c9b4SEric Dumazet 	 */
3127c68fa2bSEric Dumazet 	bytes = min_t(unsigned long,
31328b24f90SEric Dumazet 		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift),
3147c4e983cSAlexander Duyck 		      GSO_LEGACY_MAX_SIZE - 1 - MAX_TCP_HEADER);
315dcb8c9b4SEric Dumazet 	segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk));
316dcb8c9b4SEric Dumazet 
31771abf467SEric Dumazet 	return min(segs, 0x7FU);
3180f8782eaSNeal Cardwell }
3190f8782eaSNeal Cardwell 
3200f8782eaSNeal Cardwell /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */
bbr_save_cwnd(struct sock * sk)3210f8782eaSNeal Cardwell static void bbr_save_cwnd(struct sock *sk)
3220f8782eaSNeal Cardwell {
3230f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
3240f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
3250f8782eaSNeal Cardwell 
3260f8782eaSNeal Cardwell 	if (bbr->prev_ca_state < TCP_CA_Recovery && bbr->mode != BBR_PROBE_RTT)
32740570375SEric Dumazet 		bbr->prior_cwnd = tcp_snd_cwnd(tp);  /* this cwnd is good enough */
3280f8782eaSNeal Cardwell 	else  /* loss recovery or BBR_PROBE_RTT have temporarily cut cwnd */
32940570375SEric Dumazet 		bbr->prior_cwnd = max(bbr->prior_cwnd, tcp_snd_cwnd(tp));
3300f8782eaSNeal Cardwell }
3310f8782eaSNeal Cardwell 
bbr_cwnd_event(struct sock * sk,enum tcp_ca_event event)332400031e0SDavid Vernet __bpf_kfunc static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event)
3330f8782eaSNeal Cardwell {
3340f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
3350f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
3360f8782eaSNeal Cardwell 
3370f8782eaSNeal Cardwell 	if (event == CA_EVENT_TX_START && tp->app_limited) {
3380f8782eaSNeal Cardwell 		bbr->idle_restart = 1;
33978dc70ebSPriyaranjan Jha 		bbr->ack_epoch_mstamp = tp->tcp_mstamp;
34078dc70ebSPriyaranjan Jha 		bbr->ack_epoch_acked = 0;
3410f8782eaSNeal Cardwell 		/* Avoid pointless buffer overflows: pace at est. bw if we don't
3420f8782eaSNeal Cardwell 		 * need more speed (we're restarting from idle and app-limited).
3430f8782eaSNeal Cardwell 		 */
3440f8782eaSNeal Cardwell 		if (bbr->mode == BBR_PROBE_BW)
3450f8782eaSNeal Cardwell 			bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT);
3465490b32dSKevin Yang 		else if (bbr->mode == BBR_PROBE_RTT)
3475490b32dSKevin Yang 			bbr_check_probe_rtt_done(sk);
3480f8782eaSNeal Cardwell 	}
3490f8782eaSNeal Cardwell }
3500f8782eaSNeal Cardwell 
351232aa8ecSPriyaranjan Jha /* Calculate bdp based on min RTT and the estimated bottleneck bandwidth:
3520f8782eaSNeal Cardwell  *
353de8e1bebSLuke Hsiao  * bdp = ceil(bw * min_rtt * gain)
3540f8782eaSNeal Cardwell  *
3550f8782eaSNeal Cardwell  * The key factor, gain, controls the amount of queue. While a small gain
3560f8782eaSNeal Cardwell  * builds a smaller queue, it becomes more vulnerable to noise in RTT
3570f8782eaSNeal Cardwell  * measurements (e.g., delayed ACKs or other ACK compression effects). This
3580f8782eaSNeal Cardwell  * noise may cause BBR to under-estimate the rate.
3590f8782eaSNeal Cardwell  */
bbr_bdp(struct sock * sk,u32 bw,int gain)360232aa8ecSPriyaranjan Jha static u32 bbr_bdp(struct sock *sk, u32 bw, int gain)
3610f8782eaSNeal Cardwell {
3620f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
363232aa8ecSPriyaranjan Jha 	u32 bdp;
3640f8782eaSNeal Cardwell 	u64 w;
3650f8782eaSNeal Cardwell 
3660f8782eaSNeal Cardwell 	/* If we've never had a valid RTT sample, cap cwnd at the initial
3670f8782eaSNeal Cardwell 	 * default. This should only happen when the connection is not using TCP
3680f8782eaSNeal Cardwell 	 * timestamps and has retransmitted all of the SYN/SYNACK/data packets
3690f8782eaSNeal Cardwell 	 * ACKed so far. In this case, an RTO can cut cwnd to 1, in which
3700f8782eaSNeal Cardwell 	 * case we need to slow-start up toward something safe: TCP_INIT_CWND.
3710f8782eaSNeal Cardwell 	 */
3720f8782eaSNeal Cardwell 	if (unlikely(bbr->min_rtt_us == ~0U))	 /* no valid RTT samples yet? */
3730f8782eaSNeal Cardwell 		return TCP_INIT_CWND;  /* be safe: cap at default initial cwnd*/
3740f8782eaSNeal Cardwell 
3750f8782eaSNeal Cardwell 	w = (u64)bw * bbr->min_rtt_us;
3760f8782eaSNeal Cardwell 
377de8e1bebSLuke Hsiao 	/* Apply a gain to the given value, remove the BW_SCALE shift, and
378de8e1bebSLuke Hsiao 	 * round the value up to avoid a negative feedback loop.
379de8e1bebSLuke Hsiao 	 */
380232aa8ecSPriyaranjan Jha 	bdp = (((w * gain) >> BBR_SCALE) + BW_UNIT - 1) / BW_UNIT;
381232aa8ecSPriyaranjan Jha 
382232aa8ecSPriyaranjan Jha 	return bdp;
383232aa8ecSPriyaranjan Jha }
384232aa8ecSPriyaranjan Jha 
385232aa8ecSPriyaranjan Jha /* To achieve full performance in high-speed paths, we budget enough cwnd to
386232aa8ecSPriyaranjan Jha  * fit full-sized skbs in-flight on both end hosts to fully utilize the path:
387232aa8ecSPriyaranjan Jha  *   - one skb in sending host Qdisc,
388232aa8ecSPriyaranjan Jha  *   - one skb in sending host TSO/GSO engine
389232aa8ecSPriyaranjan Jha  *   - one skb being received by receiver host LRO/GRO/delayed-ACK engine
390232aa8ecSPriyaranjan Jha  * Don't worry, at low rates (bbr_min_tso_rate) this won't bloat cwnd because
391232aa8ecSPriyaranjan Jha  * in such cases tso_segs_goal is 1. The minimum cwnd is 4 packets,
392232aa8ecSPriyaranjan Jha  * which allows 2 outstanding 2-packet sequences, to try to keep pipe
393232aa8ecSPriyaranjan Jha  * full even with ACK-every-other-packet delayed ACKs.
394232aa8ecSPriyaranjan Jha  */
bbr_quantization_budget(struct sock * sk,u32 cwnd)3956b3656a6SKevin(Yudong) Yang static u32 bbr_quantization_budget(struct sock *sk, u32 cwnd)
396232aa8ecSPriyaranjan Jha {
397232aa8ecSPriyaranjan Jha 	struct bbr *bbr = inet_csk_ca(sk);
3980f8782eaSNeal Cardwell 
3990f8782eaSNeal Cardwell 	/* Allow enough full-sized skbs in flight to utilize end systems. */
40071abf467SEric Dumazet 	cwnd += 3 * bbr_tso_segs_goal(sk);
4010f8782eaSNeal Cardwell 
4020f8782eaSNeal Cardwell 	/* Reduce delayed ACKs by rounding up cwnd to the next even number. */
4030f8782eaSNeal Cardwell 	cwnd = (cwnd + 1) & ~1U;
4040f8782eaSNeal Cardwell 
405383d4709SNeal Cardwell 	/* Ensure gain cycling gets inflight above BDP even for small BDPs. */
4066b3656a6SKevin(Yudong) Yang 	if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == 0)
407383d4709SNeal Cardwell 		cwnd += 2;
408383d4709SNeal Cardwell 
4090f8782eaSNeal Cardwell 	return cwnd;
4100f8782eaSNeal Cardwell }
4110f8782eaSNeal Cardwell 
412232aa8ecSPriyaranjan Jha /* Find inflight based on min RTT and the estimated bottleneck bandwidth. */
bbr_inflight(struct sock * sk,u32 bw,int gain)413232aa8ecSPriyaranjan Jha static u32 bbr_inflight(struct sock *sk, u32 bw, int gain)
414232aa8ecSPriyaranjan Jha {
415232aa8ecSPriyaranjan Jha 	u32 inflight;
416232aa8ecSPriyaranjan Jha 
417232aa8ecSPriyaranjan Jha 	inflight = bbr_bdp(sk, bw, gain);
4186b3656a6SKevin(Yudong) Yang 	inflight = bbr_quantization_budget(sk, inflight);
419232aa8ecSPriyaranjan Jha 
420232aa8ecSPriyaranjan Jha 	return inflight;
421232aa8ecSPriyaranjan Jha }
422232aa8ecSPriyaranjan Jha 
423a87c83d5SNeal Cardwell /* With pacing at lower layers, there's often less data "in the network" than
424a87c83d5SNeal Cardwell  * "in flight". With TSQ and departure time pacing at lower layers (e.g. fq),
425a87c83d5SNeal Cardwell  * we often have several skbs queued in the pacing layer with a pre-scheduled
426a87c83d5SNeal Cardwell  * earliest departure time (EDT). BBR adapts its pacing rate based on the
427a87c83d5SNeal Cardwell  * inflight level that it estimates has already been "baked in" by previous
428a87c83d5SNeal Cardwell  * departure time decisions. We calculate a rough estimate of the number of our
429a87c83d5SNeal Cardwell  * packets that might be in the network at the earliest departure time for the
430a87c83d5SNeal Cardwell  * next skb scheduled:
431a87c83d5SNeal Cardwell  *   in_network_at_edt = inflight_at_edt - (EDT - now) * bw
432a87c83d5SNeal Cardwell  * If we're increasing inflight, then we want to know if the transmit of the
433a87c83d5SNeal Cardwell  * EDT skb will push inflight above the target, so inflight_at_edt includes
434a87c83d5SNeal Cardwell  * bbr_tso_segs_goal() from the skb departing at EDT. If decreasing inflight,
435a87c83d5SNeal Cardwell  * then estimate if inflight will sink too low just before the EDT transmit.
436a87c83d5SNeal Cardwell  */
bbr_packets_in_net_at_edt(struct sock * sk,u32 inflight_now)437a87c83d5SNeal Cardwell static u32 bbr_packets_in_net_at_edt(struct sock *sk, u32 inflight_now)
438a87c83d5SNeal Cardwell {
439a87c83d5SNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
440a87c83d5SNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
441a87c83d5SNeal Cardwell 	u64 now_ns, edt_ns, interval_us;
442a87c83d5SNeal Cardwell 	u32 interval_delivered, inflight_at_edt;
443a87c83d5SNeal Cardwell 
444a87c83d5SNeal Cardwell 	now_ns = tp->tcp_clock_cache;
445a87c83d5SNeal Cardwell 	edt_ns = max(tp->tcp_wstamp_ns, now_ns);
446a87c83d5SNeal Cardwell 	interval_us = div_u64(edt_ns - now_ns, NSEC_PER_USEC);
447a87c83d5SNeal Cardwell 	interval_delivered = (u64)bbr_bw(sk) * interval_us >> BW_SCALE;
448a87c83d5SNeal Cardwell 	inflight_at_edt = inflight_now;
449a87c83d5SNeal Cardwell 	if (bbr->pacing_gain > BBR_UNIT)              /* increasing inflight */
450a87c83d5SNeal Cardwell 		inflight_at_edt += bbr_tso_segs_goal(sk);  /* include EDT skb */
451a87c83d5SNeal Cardwell 	if (interval_delivered >= inflight_at_edt)
452a87c83d5SNeal Cardwell 		return 0;
453a87c83d5SNeal Cardwell 	return inflight_at_edt - interval_delivered;
454a87c83d5SNeal Cardwell }
455a87c83d5SNeal Cardwell 
45678dc70ebSPriyaranjan Jha /* Find the cwnd increment based on estimate of ack aggregation */
bbr_ack_aggregation_cwnd(struct sock * sk)45778dc70ebSPriyaranjan Jha static u32 bbr_ack_aggregation_cwnd(struct sock *sk)
45878dc70ebSPriyaranjan Jha {
45978dc70ebSPriyaranjan Jha 	u32 max_aggr_cwnd, aggr_cwnd = 0;
46078dc70ebSPriyaranjan Jha 
46178dc70ebSPriyaranjan Jha 	if (bbr_extra_acked_gain && bbr_full_bw_reached(sk)) {
46278dc70ebSPriyaranjan Jha 		max_aggr_cwnd = ((u64)bbr_bw(sk) * bbr_extra_acked_max_us)
46378dc70ebSPriyaranjan Jha 				/ BW_UNIT;
46478dc70ebSPriyaranjan Jha 		aggr_cwnd = (bbr_extra_acked_gain * bbr_extra_acked(sk))
46578dc70ebSPriyaranjan Jha 			     >> BBR_SCALE;
46678dc70ebSPriyaranjan Jha 		aggr_cwnd = min(aggr_cwnd, max_aggr_cwnd);
46778dc70ebSPriyaranjan Jha 	}
46878dc70ebSPriyaranjan Jha 
46978dc70ebSPriyaranjan Jha 	return aggr_cwnd;
47078dc70ebSPriyaranjan Jha }
47178dc70ebSPriyaranjan Jha 
4720f8782eaSNeal Cardwell /* An optimization in BBR to reduce losses: On the first round of recovery, we
4730f8782eaSNeal Cardwell  * follow the packet conservation principle: send P packets per P packets acked.
4740f8782eaSNeal Cardwell  * After that, we slow-start and send at most 2*P packets per P packets acked.
4750f8782eaSNeal Cardwell  * After recovery finishes, or upon undo, we restore the cwnd we had when
4760f8782eaSNeal Cardwell  * recovery started (capped by the target cwnd based on estimated BDP).
4770f8782eaSNeal Cardwell  *
4780f8782eaSNeal Cardwell  * TODO(ycheng/ncardwell): implement a rate-based approach.
4790f8782eaSNeal Cardwell  */
bbr_set_cwnd_to_recover_or_restore(struct sock * sk,const struct rate_sample * rs,u32 acked,u32 * new_cwnd)4800f8782eaSNeal Cardwell static bool bbr_set_cwnd_to_recover_or_restore(
4810f8782eaSNeal Cardwell 	struct sock *sk, const struct rate_sample *rs, u32 acked, u32 *new_cwnd)
4820f8782eaSNeal Cardwell {
4830f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
4840f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
4850f8782eaSNeal Cardwell 	u8 prev_state = bbr->prev_ca_state, state = inet_csk(sk)->icsk_ca_state;
48640570375SEric Dumazet 	u32 cwnd = tcp_snd_cwnd(tp);
4870f8782eaSNeal Cardwell 
4880f8782eaSNeal Cardwell 	/* An ACK for P pkts should release at most 2*P packets. We do this
4890f8782eaSNeal Cardwell 	 * in two steps. First, here we deduct the number of lost packets.
4900f8782eaSNeal Cardwell 	 * Then, in bbr_set_cwnd() we slow start up toward the target cwnd.
4910f8782eaSNeal Cardwell 	 */
4920f8782eaSNeal Cardwell 	if (rs->losses > 0)
4930f8782eaSNeal Cardwell 		cwnd = max_t(s32, cwnd - rs->losses, 1);
4940f8782eaSNeal Cardwell 
4950f8782eaSNeal Cardwell 	if (state == TCP_CA_Recovery && prev_state != TCP_CA_Recovery) {
4960f8782eaSNeal Cardwell 		/* Starting 1st round of Recovery, so do packet conservation. */
4970f8782eaSNeal Cardwell 		bbr->packet_conservation = 1;
4980f8782eaSNeal Cardwell 		bbr->next_rtt_delivered = tp->delivered;  /* start round now */
4990f8782eaSNeal Cardwell 		/* Cut unused cwnd from app behavior, TSQ, or TSO deferral: */
5000f8782eaSNeal Cardwell 		cwnd = tcp_packets_in_flight(tp) + acked;
5010f8782eaSNeal Cardwell 	} else if (prev_state >= TCP_CA_Recovery && state < TCP_CA_Recovery) {
5020f8782eaSNeal Cardwell 		/* Exiting loss recovery; restore cwnd saved before recovery. */
503fb998862SKevin Yang 		cwnd = max(cwnd, bbr->prior_cwnd);
5040f8782eaSNeal Cardwell 		bbr->packet_conservation = 0;
5050f8782eaSNeal Cardwell 	}
5060f8782eaSNeal Cardwell 	bbr->prev_ca_state = state;
5070f8782eaSNeal Cardwell 
5080f8782eaSNeal Cardwell 	if (bbr->packet_conservation) {
5090f8782eaSNeal Cardwell 		*new_cwnd = max(cwnd, tcp_packets_in_flight(tp) + acked);
5100f8782eaSNeal Cardwell 		return true;	/* yes, using packet conservation */
5110f8782eaSNeal Cardwell 	}
5120f8782eaSNeal Cardwell 	*new_cwnd = cwnd;
5130f8782eaSNeal Cardwell 	return false;
5140f8782eaSNeal Cardwell }
5150f8782eaSNeal Cardwell 
5160f8782eaSNeal Cardwell /* Slow-start up toward target cwnd (if bw estimate is growing, or packet loss
5170f8782eaSNeal Cardwell  * has drawn us down below target), or snap down to target if we're above it.
5180f8782eaSNeal Cardwell  */
bbr_set_cwnd(struct sock * sk,const struct rate_sample * rs,u32 acked,u32 bw,int gain)5190f8782eaSNeal Cardwell static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs,
5200f8782eaSNeal Cardwell 			 u32 acked, u32 bw, int gain)
5210f8782eaSNeal Cardwell {
5220f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
5230f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
52440570375SEric Dumazet 	u32 cwnd = tcp_snd_cwnd(tp), target_cwnd = 0;
5250f8782eaSNeal Cardwell 
5260f8782eaSNeal Cardwell 	if (!acked)
5278e995bf1SKevin Yang 		goto done;  /* no packet fully ACKed; just apply caps */
5280f8782eaSNeal Cardwell 
5290f8782eaSNeal Cardwell 	if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd))
5300f8782eaSNeal Cardwell 		goto done;
5310f8782eaSNeal Cardwell 
532232aa8ecSPriyaranjan Jha 	target_cwnd = bbr_bdp(sk, bw, gain);
53378dc70ebSPriyaranjan Jha 
53478dc70ebSPriyaranjan Jha 	/* Increment the cwnd to account for excess ACKed data that seems
53578dc70ebSPriyaranjan Jha 	 * due to aggregation (of data and/or ACKs) visible in the ACK stream.
53678dc70ebSPriyaranjan Jha 	 */
53778dc70ebSPriyaranjan Jha 	target_cwnd += bbr_ack_aggregation_cwnd(sk);
5386b3656a6SKevin(Yudong) Yang 	target_cwnd = bbr_quantization_budget(sk, target_cwnd);
53978dc70ebSPriyaranjan Jha 
54078dc70ebSPriyaranjan Jha 	/* If we're below target cwnd, slow start cwnd toward target cwnd. */
5410f8782eaSNeal Cardwell 	if (bbr_full_bw_reached(sk))  /* only cut cwnd if we filled the pipe */
5420f8782eaSNeal Cardwell 		cwnd = min(cwnd + acked, target_cwnd);
5430f8782eaSNeal Cardwell 	else if (cwnd < target_cwnd || tp->delivered < TCP_INIT_CWND)
5440f8782eaSNeal Cardwell 		cwnd = cwnd + acked;
5450f8782eaSNeal Cardwell 	cwnd = max(cwnd, bbr_cwnd_min_target);
5460f8782eaSNeal Cardwell 
5470f8782eaSNeal Cardwell done:
54840570375SEric Dumazet 	tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));	/* apply global cap */
5490f8782eaSNeal Cardwell 	if (bbr->mode == BBR_PROBE_RTT)  /* drain queue, refresh min_rtt */
55040570375SEric Dumazet 		tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), bbr_cwnd_min_target));
5510f8782eaSNeal Cardwell }
5520f8782eaSNeal Cardwell 
5530f8782eaSNeal Cardwell /* End cycle phase if it's time and/or we hit the phase's in-flight target. */
bbr_is_next_cycle_phase(struct sock * sk,const struct rate_sample * rs)5540f8782eaSNeal Cardwell static bool bbr_is_next_cycle_phase(struct sock *sk,
5550f8782eaSNeal Cardwell 				    const struct rate_sample *rs)
5560f8782eaSNeal Cardwell {
5570f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
5580f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
5590f8782eaSNeal Cardwell 	bool is_full_length =
5609a568de4SEric Dumazet 		tcp_stamp_us_delta(tp->delivered_mstamp, bbr->cycle_mstamp) >
5610f8782eaSNeal Cardwell 		bbr->min_rtt_us;
5620f8782eaSNeal Cardwell 	u32 inflight, bw;
5630f8782eaSNeal Cardwell 
5640f8782eaSNeal Cardwell 	/* The pacing_gain of 1.0 paces at the estimated bw to try to fully
5650f8782eaSNeal Cardwell 	 * use the pipe without increasing the queue.
5660f8782eaSNeal Cardwell 	 */
5670f8782eaSNeal Cardwell 	if (bbr->pacing_gain == BBR_UNIT)
5680f8782eaSNeal Cardwell 		return is_full_length;		/* just use wall clock time */
5690f8782eaSNeal Cardwell 
570a87c83d5SNeal Cardwell 	inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight);
5710f8782eaSNeal Cardwell 	bw = bbr_max_bw(sk);
5720f8782eaSNeal Cardwell 
5730f8782eaSNeal Cardwell 	/* A pacing_gain > 1.0 probes for bw by trying to raise inflight to at
5740f8782eaSNeal Cardwell 	 * least pacing_gain*BDP; this may take more than min_rtt if min_rtt is
5750f8782eaSNeal Cardwell 	 * small (e.g. on a LAN). We do not persist if packets are lost, since
5760f8782eaSNeal Cardwell 	 * a path with small buffers may not hold that much.
5770f8782eaSNeal Cardwell 	 */
5780f8782eaSNeal Cardwell 	if (bbr->pacing_gain > BBR_UNIT)
5790f8782eaSNeal Cardwell 		return is_full_length &&
5800f8782eaSNeal Cardwell 			(rs->losses ||  /* perhaps pacing_gain*BDP won't fit */
581232aa8ecSPriyaranjan Jha 			 inflight >= bbr_inflight(sk, bw, bbr->pacing_gain));
5820f8782eaSNeal Cardwell 
5830f8782eaSNeal Cardwell 	/* A pacing_gain < 1.0 tries to drain extra queue we added if bw
5840f8782eaSNeal Cardwell 	 * probing didn't find more bw. If inflight falls to match BDP then we
5850f8782eaSNeal Cardwell 	 * estimate queue is drained; persisting would underutilize the pipe.
5860f8782eaSNeal Cardwell 	 */
5870f8782eaSNeal Cardwell 	return is_full_length ||
588232aa8ecSPriyaranjan Jha 		inflight <= bbr_inflight(sk, bw, BBR_UNIT);
5890f8782eaSNeal Cardwell }
5900f8782eaSNeal Cardwell 
bbr_advance_cycle_phase(struct sock * sk)5910f8782eaSNeal Cardwell static void bbr_advance_cycle_phase(struct sock *sk)
5920f8782eaSNeal Cardwell {
5930f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
5940f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
5950f8782eaSNeal Cardwell 
5960f8782eaSNeal Cardwell 	bbr->cycle_idx = (bbr->cycle_idx + 1) & (CYCLE_LEN - 1);
5970f8782eaSNeal Cardwell 	bbr->cycle_mstamp = tp->delivered_mstamp;
5980f8782eaSNeal Cardwell }
5990f8782eaSNeal Cardwell 
6000f8782eaSNeal Cardwell /* Gain cycling: cycle pacing gain to converge to fair share of available bw. */
bbr_update_cycle_phase(struct sock * sk,const struct rate_sample * rs)6010f8782eaSNeal Cardwell static void bbr_update_cycle_phase(struct sock *sk,
6020f8782eaSNeal Cardwell 				   const struct rate_sample *rs)
6030f8782eaSNeal Cardwell {
6040f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6050f8782eaSNeal Cardwell 
6063aff3b4bSNeal Cardwell 	if (bbr->mode == BBR_PROBE_BW && bbr_is_next_cycle_phase(sk, rs))
6070f8782eaSNeal Cardwell 		bbr_advance_cycle_phase(sk);
6080f8782eaSNeal Cardwell }
6090f8782eaSNeal Cardwell 
bbr_reset_startup_mode(struct sock * sk)6100f8782eaSNeal Cardwell static void bbr_reset_startup_mode(struct sock *sk)
6110f8782eaSNeal Cardwell {
6120f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6130f8782eaSNeal Cardwell 
6140f8782eaSNeal Cardwell 	bbr->mode = BBR_STARTUP;
6150f8782eaSNeal Cardwell }
6160f8782eaSNeal Cardwell 
bbr_reset_probe_bw_mode(struct sock * sk)6170f8782eaSNeal Cardwell static void bbr_reset_probe_bw_mode(struct sock *sk)
6180f8782eaSNeal Cardwell {
6190f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6200f8782eaSNeal Cardwell 
6210f8782eaSNeal Cardwell 	bbr->mode = BBR_PROBE_BW;
6228032bf12SJason A. Donenfeld 	bbr->cycle_idx = CYCLE_LEN - 1 - get_random_u32_below(bbr_cycle_rand);
6230f8782eaSNeal Cardwell 	bbr_advance_cycle_phase(sk);	/* flip to next phase of gain cycle */
6240f8782eaSNeal Cardwell }
6250f8782eaSNeal Cardwell 
bbr_reset_mode(struct sock * sk)6260f8782eaSNeal Cardwell static void bbr_reset_mode(struct sock *sk)
6270f8782eaSNeal Cardwell {
6280f8782eaSNeal Cardwell 	if (!bbr_full_bw_reached(sk))
6290f8782eaSNeal Cardwell 		bbr_reset_startup_mode(sk);
6300f8782eaSNeal Cardwell 	else
6310f8782eaSNeal Cardwell 		bbr_reset_probe_bw_mode(sk);
6320f8782eaSNeal Cardwell }
6330f8782eaSNeal Cardwell 
6340f8782eaSNeal Cardwell /* Start a new long-term sampling interval. */
bbr_reset_lt_bw_sampling_interval(struct sock * sk)6350f8782eaSNeal Cardwell static void bbr_reset_lt_bw_sampling_interval(struct sock *sk)
6360f8782eaSNeal Cardwell {
6370f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
6380f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6390f8782eaSNeal Cardwell 
6409a568de4SEric Dumazet 	bbr->lt_last_stamp = div_u64(tp->delivered_mstamp, USEC_PER_MSEC);
6410f8782eaSNeal Cardwell 	bbr->lt_last_delivered = tp->delivered;
6420f8782eaSNeal Cardwell 	bbr->lt_last_lost = tp->lost;
6430f8782eaSNeal Cardwell 	bbr->lt_rtt_cnt = 0;
6440f8782eaSNeal Cardwell }
6450f8782eaSNeal Cardwell 
6460f8782eaSNeal Cardwell /* Completely reset long-term bandwidth sampling. */
bbr_reset_lt_bw_sampling(struct sock * sk)6470f8782eaSNeal Cardwell static void bbr_reset_lt_bw_sampling(struct sock *sk)
6480f8782eaSNeal Cardwell {
6490f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6500f8782eaSNeal Cardwell 
6510f8782eaSNeal Cardwell 	bbr->lt_bw = 0;
6520f8782eaSNeal Cardwell 	bbr->lt_use_bw = 0;
6530f8782eaSNeal Cardwell 	bbr->lt_is_sampling = false;
6540f8782eaSNeal Cardwell 	bbr_reset_lt_bw_sampling_interval(sk);
6550f8782eaSNeal Cardwell }
6560f8782eaSNeal Cardwell 
6570f8782eaSNeal Cardwell /* Long-term bw sampling interval is done. Estimate whether we're policed. */
bbr_lt_bw_interval_done(struct sock * sk,u32 bw)6580f8782eaSNeal Cardwell static void bbr_lt_bw_interval_done(struct sock *sk, u32 bw)
6590f8782eaSNeal Cardwell {
6600f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6610f8782eaSNeal Cardwell 	u32 diff;
6620f8782eaSNeal Cardwell 
6630f8782eaSNeal Cardwell 	if (bbr->lt_bw) {  /* do we have bw from a previous interval? */
6640f8782eaSNeal Cardwell 		/* Is new bw close to the lt_bw from the previous interval? */
6650f8782eaSNeal Cardwell 		diff = abs(bw - bbr->lt_bw);
6660f8782eaSNeal Cardwell 		if ((diff * BBR_UNIT <= bbr_lt_bw_ratio * bbr->lt_bw) ||
6670f8782eaSNeal Cardwell 		    (bbr_rate_bytes_per_sec(sk, diff, BBR_UNIT) <=
6680f8782eaSNeal Cardwell 		     bbr_lt_bw_diff)) {
6690f8782eaSNeal Cardwell 			/* All criteria are met; estimate we're policed. */
6700f8782eaSNeal Cardwell 			bbr->lt_bw = (bw + bbr->lt_bw) >> 1;  /* avg 2 intvls */
6710f8782eaSNeal Cardwell 			bbr->lt_use_bw = 1;
6720f8782eaSNeal Cardwell 			bbr->pacing_gain = BBR_UNIT;  /* try to avoid drops */
6730f8782eaSNeal Cardwell 			bbr->lt_rtt_cnt = 0;
6740f8782eaSNeal Cardwell 			return;
6750f8782eaSNeal Cardwell 		}
6760f8782eaSNeal Cardwell 	}
6770f8782eaSNeal Cardwell 	bbr->lt_bw = bw;
6780f8782eaSNeal Cardwell 	bbr_reset_lt_bw_sampling_interval(sk);
6790f8782eaSNeal Cardwell }
6800f8782eaSNeal Cardwell 
6810f8782eaSNeal Cardwell /* Token-bucket traffic policers are common (see "An Internet-Wide Analysis of
6820f8782eaSNeal Cardwell  * Traffic Policing", SIGCOMM 2016). BBR detects token-bucket policers and
6830f8782eaSNeal Cardwell  * explicitly models their policed rate, to reduce unnecessary losses. We
6840f8782eaSNeal Cardwell  * estimate that we're policed if we see 2 consecutive sampling intervals with
6850f8782eaSNeal Cardwell  * consistent throughput and high packet loss. If we think we're being policed,
6860f8782eaSNeal Cardwell  * set lt_bw to the "long-term" average delivery rate from those 2 intervals.
6870f8782eaSNeal Cardwell  */
bbr_lt_bw_sampling(struct sock * sk,const struct rate_sample * rs)6880f8782eaSNeal Cardwell static void bbr_lt_bw_sampling(struct sock *sk, const struct rate_sample *rs)
6890f8782eaSNeal Cardwell {
6900f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
6910f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
6920f8782eaSNeal Cardwell 	u32 lost, delivered;
6930f8782eaSNeal Cardwell 	u64 bw;
6949a568de4SEric Dumazet 	u32 t;
6950f8782eaSNeal Cardwell 
6960f8782eaSNeal Cardwell 	if (bbr->lt_use_bw) {	/* already using long-term rate, lt_bw? */
6970f8782eaSNeal Cardwell 		if (bbr->mode == BBR_PROBE_BW && bbr->round_start &&
6980f8782eaSNeal Cardwell 		    ++bbr->lt_rtt_cnt >= bbr_lt_bw_max_rtts) {
6990f8782eaSNeal Cardwell 			bbr_reset_lt_bw_sampling(sk);    /* stop using lt_bw */
7000f8782eaSNeal Cardwell 			bbr_reset_probe_bw_mode(sk);  /* restart gain cycling */
7010f8782eaSNeal Cardwell 		}
7020f8782eaSNeal Cardwell 		return;
7030f8782eaSNeal Cardwell 	}
7040f8782eaSNeal Cardwell 
7050f8782eaSNeal Cardwell 	/* Wait for the first loss before sampling, to let the policer exhaust
7060f8782eaSNeal Cardwell 	 * its tokens and estimate the steady-state rate allowed by the policer.
7070f8782eaSNeal Cardwell 	 * Starting samples earlier includes bursts that over-estimate the bw.
7080f8782eaSNeal Cardwell 	 */
7090f8782eaSNeal Cardwell 	if (!bbr->lt_is_sampling) {
7100f8782eaSNeal Cardwell 		if (!rs->losses)
7110f8782eaSNeal Cardwell 			return;
7120f8782eaSNeal Cardwell 		bbr_reset_lt_bw_sampling_interval(sk);
7130f8782eaSNeal Cardwell 		bbr->lt_is_sampling = true;
7140f8782eaSNeal Cardwell 	}
7150f8782eaSNeal Cardwell 
7160f8782eaSNeal Cardwell 	/* To avoid underestimates, reset sampling if we run out of data. */
7170f8782eaSNeal Cardwell 	if (rs->is_app_limited) {
7180f8782eaSNeal Cardwell 		bbr_reset_lt_bw_sampling(sk);
7190f8782eaSNeal Cardwell 		return;
7200f8782eaSNeal Cardwell 	}
7210f8782eaSNeal Cardwell 
7220f8782eaSNeal Cardwell 	if (bbr->round_start)
7230f8782eaSNeal Cardwell 		bbr->lt_rtt_cnt++;	/* count round trips in this interval */
7240f8782eaSNeal Cardwell 	if (bbr->lt_rtt_cnt < bbr_lt_intvl_min_rtts)
7250f8782eaSNeal Cardwell 		return;		/* sampling interval needs to be longer */
7260f8782eaSNeal Cardwell 	if (bbr->lt_rtt_cnt > 4 * bbr_lt_intvl_min_rtts) {
7270f8782eaSNeal Cardwell 		bbr_reset_lt_bw_sampling(sk);  /* interval is too long */
7280f8782eaSNeal Cardwell 		return;
7290f8782eaSNeal Cardwell 	}
7300f8782eaSNeal Cardwell 
7310f8782eaSNeal Cardwell 	/* End sampling interval when a packet is lost, so we estimate the
7320f8782eaSNeal Cardwell 	 * policer tokens were exhausted. Stopping the sampling before the
7330f8782eaSNeal Cardwell 	 * tokens are exhausted under-estimates the policed rate.
7340f8782eaSNeal Cardwell 	 */
7350f8782eaSNeal Cardwell 	if (!rs->losses)
7360f8782eaSNeal Cardwell 		return;
7370f8782eaSNeal Cardwell 
7380f8782eaSNeal Cardwell 	/* Calculate packets lost and delivered in sampling interval. */
7390f8782eaSNeal Cardwell 	lost = tp->lost - bbr->lt_last_lost;
7400f8782eaSNeal Cardwell 	delivered = tp->delivered - bbr->lt_last_delivered;
7410f8782eaSNeal Cardwell 	/* Is loss rate (lost/delivered) >= lt_loss_thresh? If not, wait. */
7420f8782eaSNeal Cardwell 	if (!delivered || (lost << BBR_SCALE) < bbr_lt_loss_thresh * delivered)
7430f8782eaSNeal Cardwell 		return;
7440f8782eaSNeal Cardwell 
7450f8782eaSNeal Cardwell 	/* Find average delivery rate in this sampling interval. */
7469a568de4SEric Dumazet 	t = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp;
7479a568de4SEric Dumazet 	if ((s32)t < 1)
7489a568de4SEric Dumazet 		return;		/* interval is less than one ms, so wait */
7499a568de4SEric Dumazet 	/* Check if can multiply without overflow */
7509a568de4SEric Dumazet 	if (t >= ~0U / USEC_PER_MSEC) {
7510f8782eaSNeal Cardwell 		bbr_reset_lt_bw_sampling(sk);  /* interval too long; reset */
7520f8782eaSNeal Cardwell 		return;
7530f8782eaSNeal Cardwell 	}
7549a568de4SEric Dumazet 	t *= USEC_PER_MSEC;
7550f8782eaSNeal Cardwell 	bw = (u64)delivered * BW_UNIT;
7560f8782eaSNeal Cardwell 	do_div(bw, t);
7570f8782eaSNeal Cardwell 	bbr_lt_bw_interval_done(sk, bw);
7580f8782eaSNeal Cardwell }
7590f8782eaSNeal Cardwell 
7600f8782eaSNeal Cardwell /* Estimate the bandwidth based on how fast packets are delivered */
bbr_update_bw(struct sock * sk,const struct rate_sample * rs)7610f8782eaSNeal Cardwell static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs)
7620f8782eaSNeal Cardwell {
7630f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
7640f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
7650f8782eaSNeal Cardwell 	u64 bw;
7660f8782eaSNeal Cardwell 
7670f8782eaSNeal Cardwell 	bbr->round_start = 0;
7680f8782eaSNeal Cardwell 	if (rs->delivered < 0 || rs->interval_us <= 0)
7690f8782eaSNeal Cardwell 		return; /* Not a valid observation */
7700f8782eaSNeal Cardwell 
7710f8782eaSNeal Cardwell 	/* See if we've reached the next RTT */
7720f8782eaSNeal Cardwell 	if (!before(rs->prior_delivered, bbr->next_rtt_delivered)) {
7730f8782eaSNeal Cardwell 		bbr->next_rtt_delivered = tp->delivered;
7740f8782eaSNeal Cardwell 		bbr->rtt_cnt++;
7750f8782eaSNeal Cardwell 		bbr->round_start = 1;
7760f8782eaSNeal Cardwell 		bbr->packet_conservation = 0;
7770f8782eaSNeal Cardwell 	}
7780f8782eaSNeal Cardwell 
7790f8782eaSNeal Cardwell 	bbr_lt_bw_sampling(sk, rs);
7800f8782eaSNeal Cardwell 
7810f8782eaSNeal Cardwell 	/* Divide delivered by the interval to find a (lower bound) bottleneck
7820f8782eaSNeal Cardwell 	 * bandwidth sample. Delivered is in packets and interval_us in uS and
7830f8782eaSNeal Cardwell 	 * ratio will be <<1 for most connections. So delivered is first scaled.
7840f8782eaSNeal Cardwell 	 */
7855b2f1f30SWen Yang 	bw = div64_long((u64)rs->delivered * BW_UNIT, rs->interval_us);
7860f8782eaSNeal Cardwell 
7870f8782eaSNeal Cardwell 	/* If this sample is application-limited, it is likely to have a very
7880f8782eaSNeal Cardwell 	 * low delivered count that represents application behavior rather than
7890f8782eaSNeal Cardwell 	 * the available network rate. Such a sample could drag down estimated
7900f8782eaSNeal Cardwell 	 * bw, causing needless slow-down. Thus, to continue to send at the
7910f8782eaSNeal Cardwell 	 * last measured network rate, we filter out app-limited samples unless
7920f8782eaSNeal Cardwell 	 * they describe the path bw at least as well as our bw model.
7930f8782eaSNeal Cardwell 	 *
7940f8782eaSNeal Cardwell 	 * So the goal during app-limited phase is to proceed with the best
7950f8782eaSNeal Cardwell 	 * network rate no matter how long. We automatically leave this
7960f8782eaSNeal Cardwell 	 * phase when app writes faster than the network can deliver :)
7970f8782eaSNeal Cardwell 	 */
7980f8782eaSNeal Cardwell 	if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) {
7990f8782eaSNeal Cardwell 		/* Incorporate new sample into our max bw filter. */
8000f8782eaSNeal Cardwell 		minmax_running_max(&bbr->bw, bbr_bw_rtts, bbr->rtt_cnt, bw);
8010f8782eaSNeal Cardwell 	}
8020f8782eaSNeal Cardwell }
8030f8782eaSNeal Cardwell 
80478dc70ebSPriyaranjan Jha /* Estimates the windowed max degree of ack aggregation.
80578dc70ebSPriyaranjan Jha  * This is used to provision extra in-flight data to keep sending during
80678dc70ebSPriyaranjan Jha  * inter-ACK silences.
80778dc70ebSPriyaranjan Jha  *
80878dc70ebSPriyaranjan Jha  * Degree of ack aggregation is estimated as extra data acked beyond expected.
80978dc70ebSPriyaranjan Jha  *
81078dc70ebSPriyaranjan Jha  * max_extra_acked = "maximum recent excess data ACKed beyond max_bw * interval"
81178dc70ebSPriyaranjan Jha  * cwnd += max_extra_acked
81278dc70ebSPriyaranjan Jha  *
81378dc70ebSPriyaranjan Jha  * Max extra_acked is clamped by cwnd and bw * bbr_extra_acked_max_us (100 ms).
81478dc70ebSPriyaranjan Jha  * Max filter is an approximate sliding window of 5-10 (packet timed) round
81578dc70ebSPriyaranjan Jha  * trips.
81678dc70ebSPriyaranjan Jha  */
bbr_update_ack_aggregation(struct sock * sk,const struct rate_sample * rs)81778dc70ebSPriyaranjan Jha static void bbr_update_ack_aggregation(struct sock *sk,
81878dc70ebSPriyaranjan Jha 				       const struct rate_sample *rs)
81978dc70ebSPriyaranjan Jha {
82078dc70ebSPriyaranjan Jha 	u32 epoch_us, expected_acked, extra_acked;
82178dc70ebSPriyaranjan Jha 	struct bbr *bbr = inet_csk_ca(sk);
82278dc70ebSPriyaranjan Jha 	struct tcp_sock *tp = tcp_sk(sk);
82378dc70ebSPriyaranjan Jha 
82478dc70ebSPriyaranjan Jha 	if (!bbr_extra_acked_gain || rs->acked_sacked <= 0 ||
82578dc70ebSPriyaranjan Jha 	    rs->delivered < 0 || rs->interval_us <= 0)
82678dc70ebSPriyaranjan Jha 		return;
82778dc70ebSPriyaranjan Jha 
82878dc70ebSPriyaranjan Jha 	if (bbr->round_start) {
82978dc70ebSPriyaranjan Jha 		bbr->extra_acked_win_rtts = min(0x1F,
83078dc70ebSPriyaranjan Jha 						bbr->extra_acked_win_rtts + 1);
83178dc70ebSPriyaranjan Jha 		if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) {
83278dc70ebSPriyaranjan Jha 			bbr->extra_acked_win_rtts = 0;
83378dc70ebSPriyaranjan Jha 			bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ?
83478dc70ebSPriyaranjan Jha 						   0 : 1;
83578dc70ebSPriyaranjan Jha 			bbr->extra_acked[bbr->extra_acked_win_idx] = 0;
83678dc70ebSPriyaranjan Jha 		}
83778dc70ebSPriyaranjan Jha 	}
83878dc70ebSPriyaranjan Jha 
83978dc70ebSPriyaranjan Jha 	/* Compute how many packets we expected to be delivered over epoch. */
84078dc70ebSPriyaranjan Jha 	epoch_us = tcp_stamp_us_delta(tp->delivered_mstamp,
84178dc70ebSPriyaranjan Jha 				      bbr->ack_epoch_mstamp);
84278dc70ebSPriyaranjan Jha 	expected_acked = ((u64)bbr_bw(sk) * epoch_us) / BW_UNIT;
84378dc70ebSPriyaranjan Jha 
84478dc70ebSPriyaranjan Jha 	/* Reset the aggregation epoch if ACK rate is below expected rate or
84578dc70ebSPriyaranjan Jha 	 * significantly large no. of ack received since epoch (potentially
84678dc70ebSPriyaranjan Jha 	 * quite old epoch).
84778dc70ebSPriyaranjan Jha 	 */
84878dc70ebSPriyaranjan Jha 	if (bbr->ack_epoch_acked <= expected_acked ||
84978dc70ebSPriyaranjan Jha 	    (bbr->ack_epoch_acked + rs->acked_sacked >=
85078dc70ebSPriyaranjan Jha 	     bbr_ack_epoch_acked_reset_thresh)) {
85178dc70ebSPriyaranjan Jha 		bbr->ack_epoch_acked = 0;
85278dc70ebSPriyaranjan Jha 		bbr->ack_epoch_mstamp = tp->delivered_mstamp;
85378dc70ebSPriyaranjan Jha 		expected_acked = 0;
85478dc70ebSPriyaranjan Jha 	}
85578dc70ebSPriyaranjan Jha 
85678dc70ebSPriyaranjan Jha 	/* Compute excess data delivered, beyond what was expected. */
85778dc70ebSPriyaranjan Jha 	bbr->ack_epoch_acked = min_t(u32, 0xFFFFF,
85878dc70ebSPriyaranjan Jha 				     bbr->ack_epoch_acked + rs->acked_sacked);
85978dc70ebSPriyaranjan Jha 	extra_acked = bbr->ack_epoch_acked - expected_acked;
86040570375SEric Dumazet 	extra_acked = min(extra_acked, tcp_snd_cwnd(tp));
86178dc70ebSPriyaranjan Jha 	if (extra_acked > bbr->extra_acked[bbr->extra_acked_win_idx])
86278dc70ebSPriyaranjan Jha 		bbr->extra_acked[bbr->extra_acked_win_idx] = extra_acked;
86378dc70ebSPriyaranjan Jha }
86478dc70ebSPriyaranjan Jha 
8650f8782eaSNeal Cardwell /* Estimate when the pipe is full, using the change in delivery rate: BBR
8660f8782eaSNeal Cardwell  * estimates that STARTUP filled the pipe if the estimated bw hasn't changed by
8670f8782eaSNeal Cardwell  * at least bbr_full_bw_thresh (25%) after bbr_full_bw_cnt (3) non-app-limited
8680f8782eaSNeal Cardwell  * rounds. Why 3 rounds: 1: rwin autotuning grows the rwin, 2: we fill the
8690f8782eaSNeal Cardwell  * higher rwin, 3: we get higher delivery rate samples. Or transient
8700f8782eaSNeal Cardwell  * cross-traffic or radio noise can go away. CUBIC Hystart shares a similar
8710f8782eaSNeal Cardwell  * design goal, but uses delay and inter-ACK spacing instead of bandwidth.
8720f8782eaSNeal Cardwell  */
bbr_check_full_bw_reached(struct sock * sk,const struct rate_sample * rs)8730f8782eaSNeal Cardwell static void bbr_check_full_bw_reached(struct sock *sk,
8740f8782eaSNeal Cardwell 				      const struct rate_sample *rs)
8750f8782eaSNeal Cardwell {
8760f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
8770f8782eaSNeal Cardwell 	u32 bw_thresh;
8780f8782eaSNeal Cardwell 
8790f8782eaSNeal Cardwell 	if (bbr_full_bw_reached(sk) || !bbr->round_start || rs->is_app_limited)
8800f8782eaSNeal Cardwell 		return;
8810f8782eaSNeal Cardwell 
8820f8782eaSNeal Cardwell 	bw_thresh = (u64)bbr->full_bw * bbr_full_bw_thresh >> BBR_SCALE;
8830f8782eaSNeal Cardwell 	if (bbr_max_bw(sk) >= bw_thresh) {
8840f8782eaSNeal Cardwell 		bbr->full_bw = bbr_max_bw(sk);
8850f8782eaSNeal Cardwell 		bbr->full_bw_cnt = 0;
8860f8782eaSNeal Cardwell 		return;
8870f8782eaSNeal Cardwell 	}
8880f8782eaSNeal Cardwell 	++bbr->full_bw_cnt;
889c589e69bSNeal Cardwell 	bbr->full_bw_reached = bbr->full_bw_cnt >= bbr_full_bw_cnt;
8900f8782eaSNeal Cardwell }
8910f8782eaSNeal Cardwell 
8920f8782eaSNeal Cardwell /* If pipe is probably full, drain the queue and then enter steady-state. */
bbr_check_drain(struct sock * sk,const struct rate_sample * rs)8930f8782eaSNeal Cardwell static void bbr_check_drain(struct sock *sk, const struct rate_sample *rs)
8940f8782eaSNeal Cardwell {
8950f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
8960f8782eaSNeal Cardwell 
8970f8782eaSNeal Cardwell 	if (bbr->mode == BBR_STARTUP && bbr_full_bw_reached(sk)) {
8980f8782eaSNeal Cardwell 		bbr->mode = BBR_DRAIN;	/* drain queue we created */
89953794570SYousuk Seung 		tcp_sk(sk)->snd_ssthresh =
900232aa8ecSPriyaranjan Jha 				bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT);
9010f8782eaSNeal Cardwell 	}	/* fall through to check if in-flight is already small: */
9020f8782eaSNeal Cardwell 	if (bbr->mode == BBR_DRAIN &&
903a87c83d5SNeal Cardwell 	    bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <=
904232aa8ecSPriyaranjan Jha 	    bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT))
9050f8782eaSNeal Cardwell 		bbr_reset_probe_bw_mode(sk);  /* we estimate queue is drained */
9060f8782eaSNeal Cardwell }
9070f8782eaSNeal Cardwell 
bbr_check_probe_rtt_done(struct sock * sk)908fb998862SKevin Yang static void bbr_check_probe_rtt_done(struct sock *sk)
909fb998862SKevin Yang {
910fb998862SKevin Yang 	struct tcp_sock *tp = tcp_sk(sk);
911fb998862SKevin Yang 	struct bbr *bbr = inet_csk_ca(sk);
912fb998862SKevin Yang 
913fb998862SKevin Yang 	if (!(bbr->probe_rtt_done_stamp &&
914fb998862SKevin Yang 	      after(tcp_jiffies32, bbr->probe_rtt_done_stamp)))
915fb998862SKevin Yang 		return;
916fb998862SKevin Yang 
917fb998862SKevin Yang 	bbr->min_rtt_stamp = tcp_jiffies32;  /* wait a while until PROBE_RTT */
91840570375SEric Dumazet 	tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp), bbr->prior_cwnd));
919fb998862SKevin Yang 	bbr_reset_mode(sk);
920fb998862SKevin Yang }
921fb998862SKevin Yang 
9220f8782eaSNeal Cardwell /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and
9230f8782eaSNeal Cardwell  * periodically drain the bottleneck queue, to converge to measure the true
9240f8782eaSNeal Cardwell  * min_rtt (unloaded propagation delay). This allows the flows to keep queues
9250f8782eaSNeal Cardwell  * small (reducing queuing delay and packet loss) and achieve fairness among
9260f8782eaSNeal Cardwell  * BBR flows.
9270f8782eaSNeal Cardwell  *
9280f8782eaSNeal Cardwell  * The min_rtt filter window is 10 seconds. When the min_rtt estimate expires,
9290f8782eaSNeal Cardwell  * we enter PROBE_RTT mode and cap the cwnd at bbr_cwnd_min_target=4 packets.
9300f8782eaSNeal Cardwell  * After at least bbr_probe_rtt_mode_ms=200ms and at least one packet-timed
9310f8782eaSNeal Cardwell  * round trip elapsed with that flight size <= 4, we leave PROBE_RTT mode and
9320f8782eaSNeal Cardwell  * re-enter the previous mode. BBR uses 200ms to approximately bound the
9330f8782eaSNeal Cardwell  * performance penalty of PROBE_RTT's cwnd capping to roughly 2% (200ms/10s).
9340f8782eaSNeal Cardwell  *
9350f8782eaSNeal Cardwell  * Note that flows need only pay 2% if they are busy sending over the last 10
9360f8782eaSNeal Cardwell  * seconds. Interactive applications (e.g., Web, RPCs, video chunks) often have
9370f8782eaSNeal Cardwell  * natural silences or low-rate periods within 10 seconds where the rate is low
9380f8782eaSNeal Cardwell  * enough for long enough to drain its queue in the bottleneck. We pick up
9390f8782eaSNeal Cardwell  * these min RTT measurements opportunistically with our min_rtt filter. :-)
9400f8782eaSNeal Cardwell  */
bbr_update_min_rtt(struct sock * sk,const struct rate_sample * rs)9410f8782eaSNeal Cardwell static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs)
9420f8782eaSNeal Cardwell {
9430f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
9440f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
9450f8782eaSNeal Cardwell 	bool filter_expired;
9460f8782eaSNeal Cardwell 
9470f8782eaSNeal Cardwell 	/* Track min RTT seen in the min_rtt_win_sec filter window: */
9482660bfa8SEric Dumazet 	filter_expired = after(tcp_jiffies32,
9490f8782eaSNeal Cardwell 			       bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ);
9500f8782eaSNeal Cardwell 	if (rs->rtt_us >= 0 &&
9511b9e2a8cSRyan Sharpelletti 	    (rs->rtt_us < bbr->min_rtt_us ||
952e4286603SYuchung Cheng 	     (filter_expired && !rs->is_ack_delayed))) {
9530f8782eaSNeal Cardwell 		bbr->min_rtt_us = rs->rtt_us;
9542660bfa8SEric Dumazet 		bbr->min_rtt_stamp = tcp_jiffies32;
9550f8782eaSNeal Cardwell 	}
9560f8782eaSNeal Cardwell 
9570f8782eaSNeal Cardwell 	if (bbr_probe_rtt_mode_ms > 0 && filter_expired &&
9580f8782eaSNeal Cardwell 	    !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) {
9590f8782eaSNeal Cardwell 		bbr->mode = BBR_PROBE_RTT;  /* dip, drain queue */
9600f8782eaSNeal Cardwell 		bbr_save_cwnd(sk);  /* note cwnd so we can restore it */
9610f8782eaSNeal Cardwell 		bbr->probe_rtt_done_stamp = 0;
9620f8782eaSNeal Cardwell 	}
9630f8782eaSNeal Cardwell 
9640f8782eaSNeal Cardwell 	if (bbr->mode == BBR_PROBE_RTT) {
9650f8782eaSNeal Cardwell 		/* Ignore low rate samples during this mode. */
9660f8782eaSNeal Cardwell 		tp->app_limited =
9670f8782eaSNeal Cardwell 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
9680f8782eaSNeal Cardwell 		/* Maintain min packets in flight for max(200 ms, 1 round). */
9690f8782eaSNeal Cardwell 		if (!bbr->probe_rtt_done_stamp &&
9700f8782eaSNeal Cardwell 		    tcp_packets_in_flight(tp) <= bbr_cwnd_min_target) {
9712660bfa8SEric Dumazet 			bbr->probe_rtt_done_stamp = tcp_jiffies32 +
9720f8782eaSNeal Cardwell 				msecs_to_jiffies(bbr_probe_rtt_mode_ms);
9730f8782eaSNeal Cardwell 			bbr->probe_rtt_round_done = 0;
9740f8782eaSNeal Cardwell 			bbr->next_rtt_delivered = tp->delivered;
9750f8782eaSNeal Cardwell 		} else if (bbr->probe_rtt_done_stamp) {
9760f8782eaSNeal Cardwell 			if (bbr->round_start)
9770f8782eaSNeal Cardwell 				bbr->probe_rtt_round_done = 1;
978fb998862SKevin Yang 			if (bbr->probe_rtt_round_done)
979fb998862SKevin Yang 				bbr_check_probe_rtt_done(sk);
9800f8782eaSNeal Cardwell 		}
9810f8782eaSNeal Cardwell 	}
982e6e6a278SNeal Cardwell 	/* Restart after idle ends only once we process a new S/ACK for data */
983e6e6a278SNeal Cardwell 	if (rs->delivered > 0)
9840f8782eaSNeal Cardwell 		bbr->idle_restart = 0;
9850f8782eaSNeal Cardwell }
9860f8782eaSNeal Cardwell 
bbr_update_gains(struct sock * sk)987cf33e25cSNeal Cardwell static void bbr_update_gains(struct sock *sk)
988cf33e25cSNeal Cardwell {
989cf33e25cSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
990cf33e25cSNeal Cardwell 
991cf33e25cSNeal Cardwell 	switch (bbr->mode) {
992cf33e25cSNeal Cardwell 	case BBR_STARTUP:
993cf33e25cSNeal Cardwell 		bbr->pacing_gain = bbr_high_gain;
994cf33e25cSNeal Cardwell 		bbr->cwnd_gain	 = bbr_high_gain;
995cf33e25cSNeal Cardwell 		break;
996cf33e25cSNeal Cardwell 	case BBR_DRAIN:
997cf33e25cSNeal Cardwell 		bbr->pacing_gain = bbr_drain_gain;	/* slow, to drain */
998cf33e25cSNeal Cardwell 		bbr->cwnd_gain	 = bbr_high_gain;	/* keep cwnd */
999cf33e25cSNeal Cardwell 		break;
1000cf33e25cSNeal Cardwell 	case BBR_PROBE_BW:
1001cf33e25cSNeal Cardwell 		bbr->pacing_gain = (bbr->lt_use_bw ?
1002cf33e25cSNeal Cardwell 				    BBR_UNIT :
1003cf33e25cSNeal Cardwell 				    bbr_pacing_gain[bbr->cycle_idx]);
1004cf33e25cSNeal Cardwell 		bbr->cwnd_gain	 = bbr_cwnd_gain;
1005cf33e25cSNeal Cardwell 		break;
1006cf33e25cSNeal Cardwell 	case BBR_PROBE_RTT:
1007cf33e25cSNeal Cardwell 		bbr->pacing_gain = BBR_UNIT;
1008cf33e25cSNeal Cardwell 		bbr->cwnd_gain	 = BBR_UNIT;
1009cf33e25cSNeal Cardwell 		break;
1010cf33e25cSNeal Cardwell 	default:
1011cf33e25cSNeal Cardwell 		WARN_ONCE(1, "BBR bad mode: %u\n", bbr->mode);
1012cf33e25cSNeal Cardwell 		break;
1013cf33e25cSNeal Cardwell 	}
1014cf33e25cSNeal Cardwell }
1015cf33e25cSNeal Cardwell 
bbr_update_model(struct sock * sk,const struct rate_sample * rs)10160f8782eaSNeal Cardwell static void bbr_update_model(struct sock *sk, const struct rate_sample *rs)
10170f8782eaSNeal Cardwell {
10180f8782eaSNeal Cardwell 	bbr_update_bw(sk, rs);
101978dc70ebSPriyaranjan Jha 	bbr_update_ack_aggregation(sk, rs);
10200f8782eaSNeal Cardwell 	bbr_update_cycle_phase(sk, rs);
10210f8782eaSNeal Cardwell 	bbr_check_full_bw_reached(sk, rs);
10220f8782eaSNeal Cardwell 	bbr_check_drain(sk, rs);
10230f8782eaSNeal Cardwell 	bbr_update_min_rtt(sk, rs);
1024cf33e25cSNeal Cardwell 	bbr_update_gains(sk);
10250f8782eaSNeal Cardwell }
10260f8782eaSNeal Cardwell 
bbr_main(struct sock * sk,u32 ack,int flag,const struct rate_sample * rs)1027*57bfc760SMiao Xu __bpf_kfunc static void bbr_main(struct sock *sk, u32 ack, int flag, const struct rate_sample *rs)
10280f8782eaSNeal Cardwell {
10290f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
10300f8782eaSNeal Cardwell 	u32 bw;
10310f8782eaSNeal Cardwell 
10320f8782eaSNeal Cardwell 	bbr_update_model(sk, rs);
10330f8782eaSNeal Cardwell 
10340f8782eaSNeal Cardwell 	bw = bbr_bw(sk);
10350f8782eaSNeal Cardwell 	bbr_set_pacing_rate(sk, bw, bbr->pacing_gain);
10360f8782eaSNeal Cardwell 	bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain);
10370f8782eaSNeal Cardwell }
10380f8782eaSNeal Cardwell 
bbr_init(struct sock * sk)1039400031e0SDavid Vernet __bpf_kfunc static void bbr_init(struct sock *sk)
10400f8782eaSNeal Cardwell {
10410f8782eaSNeal Cardwell 	struct tcp_sock *tp = tcp_sk(sk);
10420f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
10430f8782eaSNeal Cardwell 
10440f8782eaSNeal Cardwell 	bbr->prior_cwnd = 0;
104553794570SYousuk Seung 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
10460f8782eaSNeal Cardwell 	bbr->rtt_cnt = 0;
10476de035feSNeal Cardwell 	bbr->next_rtt_delivered = tp->delivered;
10480f8782eaSNeal Cardwell 	bbr->prev_ca_state = TCP_CA_Open;
10490f8782eaSNeal Cardwell 	bbr->packet_conservation = 0;
10500f8782eaSNeal Cardwell 
10510f8782eaSNeal Cardwell 	bbr->probe_rtt_done_stamp = 0;
10520f8782eaSNeal Cardwell 	bbr->probe_rtt_round_done = 0;
10530f8782eaSNeal Cardwell 	bbr->min_rtt_us = tcp_min_rtt(tp);
10542660bfa8SEric Dumazet 	bbr->min_rtt_stamp = tcp_jiffies32;
10550f8782eaSNeal Cardwell 
10560f8782eaSNeal Cardwell 	minmax_reset(&bbr->bw, bbr->rtt_cnt, 0);  /* init max bw to 0 */
10570f8782eaSNeal Cardwell 
105832984565SNeal Cardwell 	bbr->has_seen_rtt = 0;
105979135b89SNeal Cardwell 	bbr_init_pacing_rate_from_rtt(sk);
10600f8782eaSNeal Cardwell 
10610f8782eaSNeal Cardwell 	bbr->round_start = 0;
10620f8782eaSNeal Cardwell 	bbr->idle_restart = 0;
1063c589e69bSNeal Cardwell 	bbr->full_bw_reached = 0;
10640f8782eaSNeal Cardwell 	bbr->full_bw = 0;
10650f8782eaSNeal Cardwell 	bbr->full_bw_cnt = 0;
10669a568de4SEric Dumazet 	bbr->cycle_mstamp = 0;
10670f8782eaSNeal Cardwell 	bbr->cycle_idx = 0;
10680f8782eaSNeal Cardwell 	bbr_reset_lt_bw_sampling(sk);
10690f8782eaSNeal Cardwell 	bbr_reset_startup_mode(sk);
1070218af599SEric Dumazet 
107178dc70ebSPriyaranjan Jha 	bbr->ack_epoch_mstamp = tp->tcp_mstamp;
107278dc70ebSPriyaranjan Jha 	bbr->ack_epoch_acked = 0;
107378dc70ebSPriyaranjan Jha 	bbr->extra_acked_win_rtts = 0;
107478dc70ebSPriyaranjan Jha 	bbr->extra_acked_win_idx = 0;
107578dc70ebSPriyaranjan Jha 	bbr->extra_acked[0] = 0;
107678dc70ebSPriyaranjan Jha 	bbr->extra_acked[1] = 0;
107778dc70ebSPriyaranjan Jha 
1078218af599SEric Dumazet 	cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED);
10790f8782eaSNeal Cardwell }
10800f8782eaSNeal Cardwell 
bbr_sndbuf_expand(struct sock * sk)1081400031e0SDavid Vernet __bpf_kfunc static u32 bbr_sndbuf_expand(struct sock *sk)
10820f8782eaSNeal Cardwell {
10830f8782eaSNeal Cardwell 	/* Provision 3 * cwnd since BBR may slow-start even during recovery. */
10840f8782eaSNeal Cardwell 	return 3;
10850f8782eaSNeal Cardwell }
10860f8782eaSNeal Cardwell 
10870f8782eaSNeal Cardwell /* In theory BBR does not need to undo the cwnd since it does not
10880f8782eaSNeal Cardwell  * always reduce cwnd on losses (see bbr_main()). Keep it for now.
10890f8782eaSNeal Cardwell  */
bbr_undo_cwnd(struct sock * sk)1090400031e0SDavid Vernet __bpf_kfunc static u32 bbr_undo_cwnd(struct sock *sk)
10910f8782eaSNeal Cardwell {
10922f6c498eSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
10932f6c498eSNeal Cardwell 
10942f6c498eSNeal Cardwell 	bbr->full_bw = 0;   /* spurious slow-down; reset full pipe detection */
10952f6c498eSNeal Cardwell 	bbr->full_bw_cnt = 0;
1096600647d4SNeal Cardwell 	bbr_reset_lt_bw_sampling(sk);
109740570375SEric Dumazet 	return tcp_snd_cwnd(tcp_sk(sk));
10980f8782eaSNeal Cardwell }
10990f8782eaSNeal Cardwell 
11000f8782eaSNeal Cardwell /* Entering loss recovery, so save cwnd for when we exit or undo recovery. */
bbr_ssthresh(struct sock * sk)1101400031e0SDavid Vernet __bpf_kfunc static u32 bbr_ssthresh(struct sock *sk)
11020f8782eaSNeal Cardwell {
11030f8782eaSNeal Cardwell 	bbr_save_cwnd(sk);
110453794570SYousuk Seung 	return tcp_sk(sk)->snd_ssthresh;
11050f8782eaSNeal Cardwell }
11060f8782eaSNeal Cardwell 
bbr_get_info(struct sock * sk,u32 ext,int * attr,union tcp_cc_info * info)11070f8782eaSNeal Cardwell static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr,
11080f8782eaSNeal Cardwell 			   union tcp_cc_info *info)
11090f8782eaSNeal Cardwell {
11100f8782eaSNeal Cardwell 	if (ext & (1 << (INET_DIAG_BBRINFO - 1)) ||
11110f8782eaSNeal Cardwell 	    ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
11120f8782eaSNeal Cardwell 		struct tcp_sock *tp = tcp_sk(sk);
11130f8782eaSNeal Cardwell 		struct bbr *bbr = inet_csk_ca(sk);
11140f8782eaSNeal Cardwell 		u64 bw = bbr_bw(sk);
11150f8782eaSNeal Cardwell 
11160f8782eaSNeal Cardwell 		bw = bw * tp->mss_cache * USEC_PER_SEC >> BW_SCALE;
11170f8782eaSNeal Cardwell 		memset(&info->bbr, 0, sizeof(info->bbr));
11180f8782eaSNeal Cardwell 		info->bbr.bbr_bw_lo		= (u32)bw;
11190f8782eaSNeal Cardwell 		info->bbr.bbr_bw_hi		= (u32)(bw >> 32);
11200f8782eaSNeal Cardwell 		info->bbr.bbr_min_rtt		= bbr->min_rtt_us;
11210f8782eaSNeal Cardwell 		info->bbr.bbr_pacing_gain	= bbr->pacing_gain;
11220f8782eaSNeal Cardwell 		info->bbr.bbr_cwnd_gain		= bbr->cwnd_gain;
11230f8782eaSNeal Cardwell 		*attr = INET_DIAG_BBRINFO;
11240f8782eaSNeal Cardwell 		return sizeof(info->bbr);
11250f8782eaSNeal Cardwell 	}
11260f8782eaSNeal Cardwell 	return 0;
11270f8782eaSNeal Cardwell }
11280f8782eaSNeal Cardwell 
bbr_set_state(struct sock * sk,u8 new_state)1129400031e0SDavid Vernet __bpf_kfunc static void bbr_set_state(struct sock *sk, u8 new_state)
11300f8782eaSNeal Cardwell {
11310f8782eaSNeal Cardwell 	struct bbr *bbr = inet_csk_ca(sk);
11320f8782eaSNeal Cardwell 
11330f8782eaSNeal Cardwell 	if (new_state == TCP_CA_Loss) {
11340f8782eaSNeal Cardwell 		struct rate_sample rs = { .losses = 1 };
11350f8782eaSNeal Cardwell 
11360f8782eaSNeal Cardwell 		bbr->prev_ca_state = TCP_CA_Loss;
11370f8782eaSNeal Cardwell 		bbr->full_bw = 0;
11380f8782eaSNeal Cardwell 		bbr->round_start = 1;	/* treat RTO like end of a round */
11390f8782eaSNeal Cardwell 		bbr_lt_bw_sampling(sk, &rs);
11400f8782eaSNeal Cardwell 	}
11410f8782eaSNeal Cardwell }
11420f8782eaSNeal Cardwell 
11430f8782eaSNeal Cardwell static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = {
11440f8782eaSNeal Cardwell 	.flags		= TCP_CONG_NON_RESTRICTED,
11450f8782eaSNeal Cardwell 	.name		= "bbr",
11460f8782eaSNeal Cardwell 	.owner		= THIS_MODULE,
11470f8782eaSNeal Cardwell 	.init		= bbr_init,
11480f8782eaSNeal Cardwell 	.cong_control	= bbr_main,
11490f8782eaSNeal Cardwell 	.sndbuf_expand	= bbr_sndbuf_expand,
11500f8782eaSNeal Cardwell 	.undo_cwnd	= bbr_undo_cwnd,
11510f8782eaSNeal Cardwell 	.cwnd_event	= bbr_cwnd_event,
11520f8782eaSNeal Cardwell 	.ssthresh	= bbr_ssthresh,
1153dcb8c9b4SEric Dumazet 	.min_tso_segs	= bbr_min_tso_segs,
11540f8782eaSNeal Cardwell 	.get_info	= bbr_get_info,
11550f8782eaSNeal Cardwell 	.set_state	= bbr_set_state,
11560f8782eaSNeal Cardwell };
11570f8782eaSNeal Cardwell 
11586f3189f3SDaniel Xu BTF_KFUNCS_START(tcp_bbr_check_kfunc_ids)
1159a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_init)
1160a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_main)
1161a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_sndbuf_expand)
1162a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_undo_cwnd)
1163a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_cwnd_event)
1164a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_ssthresh)
1165a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_min_tso_segs)
1166a4703e31SKumar Kartikeya Dwivedi BTF_ID_FLAGS(func, bbr_set_state)
11676f3189f3SDaniel Xu BTF_KFUNCS_END(tcp_bbr_check_kfunc_ids)
11680e32dfc8SKumar Kartikeya Dwivedi 
1169b202d844SKumar Kartikeya Dwivedi static const struct btf_kfunc_id_set tcp_bbr_kfunc_set = {
1170b202d844SKumar Kartikeya Dwivedi 	.owner = THIS_MODULE,
1171a4703e31SKumar Kartikeya Dwivedi 	.set   = &tcp_bbr_check_kfunc_ids,
1172b202d844SKumar Kartikeya Dwivedi };
11730e32dfc8SKumar Kartikeya Dwivedi 
bbr_register(void)11740f8782eaSNeal Cardwell static int __init bbr_register(void)
11750f8782eaSNeal Cardwell {
11760e32dfc8SKumar Kartikeya Dwivedi 	int ret;
11770e32dfc8SKumar Kartikeya Dwivedi 
11780f8782eaSNeal Cardwell 	BUILD_BUG_ON(sizeof(struct bbr) > ICSK_CA_PRIV_SIZE);
1179b202d844SKumar Kartikeya Dwivedi 
1180b202d844SKumar Kartikeya Dwivedi 	ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_bbr_kfunc_set);
1181b202d844SKumar Kartikeya Dwivedi 	if (ret < 0)
11820e32dfc8SKumar Kartikeya Dwivedi 		return ret;
1183b202d844SKumar Kartikeya Dwivedi 	return tcp_register_congestion_control(&tcp_bbr_cong_ops);
11840f8782eaSNeal Cardwell }
11850f8782eaSNeal Cardwell 
bbr_unregister(void)11860f8782eaSNeal Cardwell static void __exit bbr_unregister(void)
11870f8782eaSNeal Cardwell {
11880f8782eaSNeal Cardwell 	tcp_unregister_congestion_control(&tcp_bbr_cong_ops);
11890f8782eaSNeal Cardwell }
11900f8782eaSNeal Cardwell 
11910f8782eaSNeal Cardwell module_init(bbr_register);
11920f8782eaSNeal Cardwell module_exit(bbr_unregister);
11930f8782eaSNeal Cardwell 
11940f8782eaSNeal Cardwell MODULE_AUTHOR("Van Jacobson <vanj@google.com>");
11950f8782eaSNeal Cardwell MODULE_AUTHOR("Neal Cardwell <ncardwell@google.com>");
11960f8782eaSNeal Cardwell MODULE_AUTHOR("Yuchung Cheng <ycheng@google.com>");
11970f8782eaSNeal Cardwell MODULE_AUTHOR("Soheil Hassas Yeganeh <soheil@google.com>");
11980f8782eaSNeal Cardwell MODULE_LICENSE("Dual BSD/GPL");
11990f8782eaSNeal Cardwell MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)");
1200