diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h index c2b15f7e55161..a400a84088d38 100644 --- a/include/net/inet_connection_sock.h +++ b/include/net/inet_connection_sock.h @@ -135,8 +135,8 @@ struct inet_connection_sock { u32 icsk_probes_tstamp; u32 icsk_user_timeout; - u64 icsk_ca_priv[104 / sizeof(u64)]; -#define ICSK_CA_PRIV_SIZE sizeof_field(struct inet_connection_sock, icsk_ca_priv) +#define ICSK_CA_PRIV_SIZE (144) + u64 icsk_ca_priv[ICSK_CA_PRIV_SIZE / sizeof(u64)]; }; #define ICSK_TIME_RETRANS 1 /* Retransmit timer */ diff --git a/include/net/tcp.h b/include/net/tcp.h index 2b89bf7c3317c..0c717d4687fc9 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -2232,7 +2232,7 @@ struct tcp_plb_state { u8 consec_cong_rounds:5, /* consecutive congested rounds */ unused:3; u32 pause_until; /* jiffies32 when PLB can resume rerouting */ -}; +} __attribute__ ((__packed__)); static inline void tcp_plb_init(const struct sock *sk, struct tcp_plb_state *plb) diff --git a/include/uapi/linux/inet_diag.h b/include/uapi/linux/inet_diag.h index 50655de04c9b6..82f8bd8f0d161 100644 --- a/include/uapi/linux/inet_diag.h +++ b/include/uapi/linux/inet_diag.h @@ -229,6 +229,29 @@ struct tcp_bbr_info { __u32 bbr_min_rtt; /* min-filtered RTT in uSec */ __u32 bbr_pacing_gain; /* pacing gain shifted left 8 bits */ __u32 bbr_cwnd_gain; /* cwnd gain shifted left 8 bits */ + __u32 bbr_bw_hi_lsb; /* lower 32 bits of bw_hi */ + __u32 bbr_bw_hi_msb; /* upper 32 bits of bw_hi */ + __u32 bbr_bw_lo_lsb; /* lower 32 bits of bw_lo */ + __u32 bbr_bw_lo_msb; /* upper 32 bits of bw_lo */ + __u8 bbr_mode; /* current bbr_mode in state machine */ + __u8 bbr_phase; /* current state machine phase */ + __u8 unused1; /* alignment padding; not used yet */ + __u8 bbr_version; /* BBR algorithm version */ + __u32 bbr_inflight_lo; /* lower short-term data volume bound */ + __u32 bbr_inflight_hi; /* higher long-term data volume bound */ + __u32 bbr_extra_acked; /* max excess packets ACKed in epoch */ +}; + +/* TCP BBR congestion control bbr_phase as reported in netlink/ss stats. */ +enum tcp_bbr_phase { + BBR_PHASE_INVALID = 0, + BBR_PHASE_STARTUP = 1, + BBR_PHASE_DRAIN = 2, + BBR_PHASE_PROBE_RTT = 3, + BBR_PHASE_PROBE_BW_UP = 4, + BBR_PHASE_PROBE_BW_DOWN = 5, + BBR_PHASE_PROBE_BW_CRUISE = 6, + BBR_PHASE_PROBE_BW_REFILL = 7, }; union tcp_cc_info { diff --git a/net/ipv4/Kconfig b/net/ipv4/Kconfig index 2dfb12230f089..2e14db3bee704 100644 --- a/net/ipv4/Kconfig +++ b/net/ipv4/Kconfig @@ -668,15 +668,18 @@ config TCP_CONG_BBR default n help - BBR (Bottleneck Bandwidth and RTT) TCP congestion control aims to - maximize network utilization and minimize queues. It builds an explicit - model of the bottleneck delivery rate and path round-trip propagation - delay. It tolerates packet loss and delay unrelated to congestion. It - can operate over LAN, WAN, cellular, wifi, or cable modem links. It can - coexist with flows that use loss-based congestion control, and can - operate with shallow buffers, deep buffers, bufferbloat, policers, or - AQM schemes that do not provide a delay signal. It requires the fq - ("Fair Queue") pacing packet scheduler. + BBR (Bottleneck Bandwidth and RTT) TCP congestion control is a + model-based congestion control algorithm that aims to maximize + network utilization, keep queues and retransmit rates low, and to be + able to coexist with Reno/CUBIC in common scenarios. It builds an + explicit model of the network path. It tolerates a targeted degree + of random packet loss and delay. It can operate over LAN, WAN, + cellular, wifi, or cable modem links, and can use shallow-threshold + ECN signals. It can coexist to some degree with flows that use + loss-based congestion control, and can operate with shallow buffers, + deep buffers, bufferbloat, policers, or AQM schemes that do not + provide a delay signal. It requires pacing, using either TCP internal + pacing or the fq ("Fair Queue") pacing packet scheduler. choice prompt "Default TCP congestion control" diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c index f2c6c5d310686..1ce01643709c5 100644 --- a/net/ipv4/tcp_bbr.c +++ b/net/ipv4/tcp_bbr.c @@ -1,18 +1,19 @@ -/* Bottleneck Bandwidth and RTT (BBR) congestion control +/* BBR (Bottleneck Bandwidth and RTT) congestion control * - * BBR congestion control computes the sending rate based on the delivery - * rate (throughput) estimated from ACKs. In a nutshell: + * BBR is a model-based congestion control algorithm that aims for low queues, + * low loss, and (bounded) Reno/CUBIC coexistence. To maintain a model of the + * network path, it uses measurements of bandwidth and RTT, as well as (if they + * occur) packet loss and/or shallow-threshold ECN signals. Note that although + * it can use ECN or loss signals explicitly, it does not require either; it + * can bound its in-flight data based on its estimate of the BDP. * - * 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. + * The model has both higher and lower bounds for the operating range: + * lo: bw_lo, inflight_lo: conservative short-term lower bound + * hi: bw_hi, inflight_hi: robust long-term upper bound + * The bandwidth-probing time scale is (a) extended dynamically based on + * estimated BDP to improve coexistence with Reno/CUBIC; (b) bounded by + * an interactive wall-clock time-scale to be more scalable and responsive + * than Reno and CUBIC. * * Here is a state transition diagram for BBR: * @@ -65,6 +66,13 @@ #include #include +#include +#include "tcp_dctcp.h" + +#define BBR_VERSION 3 + +#define bbr_param(sk,name) (bbr_ ## name) + /* 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. @@ -85,36 +93,41 @@ enum bbr_mode { BBR_PROBE_RTT, /* cut inflight to min to probe min_rtt */ }; +/* How does the incoming ACK stream relate to our bandwidth probing? */ +enum bbr_ack_phase { + BBR_ACKS_INIT, /* not probing; not getting probe feedback */ + BBR_ACKS_REFILLING, /* sending at est. bw to fill pipe */ + BBR_ACKS_PROBE_STARTING, /* inflight rising to probe bw */ + BBR_ACKS_PROBE_FEEDBACK, /* getting feedback from bw probing */ + BBR_ACKS_PROBE_STOPPING, /* stopped probing; still getting feedback */ +}; + /* 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 probe_rtt_min_us; /* min RTT in probe_rtt_win_ms win */ + u32 probe_rtt_min_stamp; /* timestamp of probe_rtt_min_us*/ u32 next_rtt_delivered; /* scb->tx.delivered at end of round */ u64 cycle_mstamp; /* time of this cycle phase start */ - u32 mode:3, /* current bbr_mode in state machine */ + u32 mode:2, /* current bbr_mode in state machine */ prev_ca_state:3, /* CA state on previous ACK */ - packet_conservation:1, /* use packet conservation? */ round_start:1, /* start of packet-timed tx->ack round? */ + ce_state:1, /* If most recent data has CE bit set */ + bw_probe_up_rounds:5, /* cwnd-limited rounds in PROBE_UP */ + try_fast_path:1, /* can we take fast path? */ idle_restart:1, /* restarting after idle? */ probe_rtt_round_done:1, /* a BBR_PROBE_RTT round at 4 pkts? */ - unused:13, - 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 */ + init_cwnd:7, /* initial cwnd */ + unused_1:10; u32 pacing_gain:10, /* current gain for setting pacing rate */ cwnd_gain:10, /* current gain for setting cwnd */ full_bw_reached:1, /* reached full bw in Startup? */ full_bw_cnt:2, /* number of rounds without large bw gains */ - cycle_idx:3, /* current index in pacing_gain cycle array */ + cycle_idx:2, /* current index in pacing_gain cycle array */ has_seen_rtt:1, /* have we seen an RTT sample yet? */ - unused_b:5; + unused_2:6; u32 prior_cwnd; /* prior cwnd upon entering loss recovery */ u32 full_bw; /* recent bw, to estimate if pipe is full */ @@ -124,19 +137,67 @@ struct bbr { u32 ack_epoch_acked:20, /* packets (S)ACKed in sampling epoch */ extra_acked_win_rtts:5, /* age of extra_acked, in round trips */ extra_acked_win_idx:1, /* current index in extra_acked array */ - unused_c:6; + /* BBR v3 state: */ + full_bw_now:1, /* recently reached full bw plateau? */ + startup_ecn_rounds:2, /* consecutive hi ECN STARTUP rounds */ + loss_in_cycle:1, /* packet loss in this cycle? */ + ecn_in_cycle:1, /* ECN in this cycle? */ + unused_3:1; + u32 loss_round_delivered; /* scb->tx.delivered ending loss round */ + u32 undo_bw_lo; /* bw_lo before latest losses */ + u32 undo_inflight_lo; /* inflight_lo before latest losses */ + u32 undo_inflight_hi; /* inflight_hi before latest losses */ + u32 bw_latest; /* max delivered bw in last round trip */ + u32 bw_lo; /* lower bound on sending bandwidth */ + u32 bw_hi[2]; /* max recent measured bw sample */ + u32 inflight_latest; /* max delivered data in last round trip */ + u32 inflight_lo; /* lower bound of inflight data range */ + u32 inflight_hi; /* upper bound of inflight data range */ + u32 bw_probe_up_cnt; /* packets delivered per inflight_hi incr */ + u32 bw_probe_up_acks; /* packets (S)ACKed since inflight_hi incr */ + u32 probe_wait_us; /* PROBE_DOWN until next clock-driven probe */ + u32 prior_rcv_nxt; /* tp->rcv_nxt when CE state last changed */ + u32 ecn_eligible:1, /* sender can use ECN (RTT, handshake)? */ + ecn_alpha:9, /* EWMA delivered_ce/delivered; 0..256 */ + bw_probe_samples:1, /* rate samples reflect bw probing? */ + prev_probe_too_high:1, /* did last PROBE_UP go too high? */ + stopped_risky_probe:1, /* last PROBE_UP stopped due to risk? */ + rounds_since_probe:8, /* packet-timed rounds since probed bw */ + loss_round_start:1, /* loss_round_delivered round trip? */ + loss_in_round:1, /* loss marked in this round trip? */ + ecn_in_round:1, /* ECN marked in this round trip? */ + ack_phase:3, /* bbr_ack_phase: meaning of ACKs */ + loss_events_in_round:4,/* losses in STARTUP round */ + initialized:1; /* has bbr_init() been called? */ + u32 alpha_last_delivered; /* tp->delivered at alpha update */ + u32 alpha_last_delivered_ce; /* tp->delivered_ce at alpha update */ + + u8 unused_4; /* to preserve alignment */ + struct tcp_plb_state plb; }; -#define CYCLE_LEN 8 /* number of phases in a pacing gain cycle */ +struct bbr_context { + u32 sample_bw; +}; -/* 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; +/* Window length of probe_rtt_min_us filter (in ms), and consequently the + * typical interval between PROBE_RTT mode entries. The default is 5000ms. + * Note that bbr_probe_rtt_win_ms must be <= bbr_min_rtt_win_sec * MSEC_PER_SEC + */ +static const u32 bbr_probe_rtt_win_ms = 5000; +/* Proportion of cwnd to estimated BDP in PROBE_RTT, in units of BBR_UNIT: */ +static const u32 bbr_probe_rtt_cwnd_gain = BBR_UNIT * 1 / 2; + +/* Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting + * in bigger TSO bursts. We cut the RTT-based allowance in half + * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance + * is below 1500 bytes after 6 * ~500 usec = 3ms. + */ +static const u32 bbr_tso_rtt_shift = 9; /* Pace at ~1% below estimated bw, on average, to reduce queue at bottleneck. * In order to help drive the network toward lower queues and low latency while @@ -146,13 +207,15 @@ static const int bbr_min_tso_rate = 1200000; */ static const int bbr_pacing_margin_percent = 1; -/* We use a high_gain value of 2/ln(2) because it's the smallest pacing gain +/* We use a startup_pacing_gain of 4*ln(2) because it's the smallest value * 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 +static const int bbr_startup_pacing_gain = BBR_UNIT * 277 / 100 + 1; +/* The gain for deriving startup cwnd: */ +static const int bbr_startup_cwnd_gain = BBR_UNIT * 2; +/* The pacing 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; @@ -160,13 +223,17 @@ static const int bbr_drain_gain = BBR_UNIT * 1000 / 2885; 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... */ + BBR_UNIT * 5 / 4, /* UP: probe for more available bw */ + BBR_UNIT * 91 / 100, /* DOWN: drain queue and/or yield bw */ + BBR_UNIT, /* CRUISE: try to use pipe w/ some headroom */ + BBR_UNIT, /* REFILL: refill pipe to estimated 100% */ +}; +enum bbr_pacing_gain_phase { + BBR_BW_PROBE_UP = 0, /* push up inflight to probe for bw/vol */ + BBR_BW_PROBE_DOWN = 1, /* drain excess inflight from the queue */ + BBR_BW_PROBE_CRUISE = 2, /* use pipe, w/ headroom in queue/pipe */ + BBR_BW_PROBE_REFILL = 3, /* v2: refill the pipe again to 100% */ }; -/* 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 @@ -174,24 +241,12 @@ static const u32 bbr_cycle_rand = 7; */ static const u32 bbr_cwnd_min_target = 4; -/* To estimate if BBR_STARTUP mode (i.e. high_gain) has filled pipe... */ +/* To estimate if BBR_STARTUP or BBR_BW_PROBE_UP 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; - /* Gain factor for adding extra_acked to target cwnd: */ static const int bbr_extra_acked_gain = BBR_UNIT; /* Window length of extra_acked window. */ @@ -201,8 +256,121 @@ static const u32 bbr_ack_epoch_acked_reset_thresh = 1U << 20; /* Time period for clamping cwnd increment due to ack aggregation */ static const u32 bbr_extra_acked_max_us = 100 * 1000; +/* Flags to control BBR ECN-related behavior... */ + +/* Ensure ACKs only ACK packets with consistent ECN CE status? */ +static const bool bbr_precise_ece_ack = true; + +/* Max RTT (in usec) at which to use sender-side ECN logic. + * Disabled when 0 (ECN allowed at any RTT). + */ +static const u32 bbr_ecn_max_rtt_us = 5000; + +/* On losses, scale down inflight and pacing rate by beta scaled by BBR_SCALE. + * No loss response when 0. + */ +static const u32 bbr_beta = BBR_UNIT * 30 / 100; + +/* Gain factor for ECN mark ratio samples, scaled by BBR_SCALE (1/16 = 6.25%) */ +static const u32 bbr_ecn_alpha_gain = BBR_UNIT * 1 / 16; + +/* The initial value for ecn_alpha; 1.0 allows a flow to respond quickly + * to congestion if the bottleneck is congested when the flow starts up. + */ +static const u32 bbr_ecn_alpha_init = BBR_UNIT; + +/* On ECN, cut inflight_lo to (1 - ecn_factor * ecn_alpha) scaled by BBR_SCALE. + * No ECN based bounding when 0. + */ +static const u32 bbr_ecn_factor = BBR_UNIT * 1 / 3; /* 1/3 = 33% */ + +/* Estimate bw probing has gone too far if CE ratio exceeds this threshold. + * Scaled by BBR_SCALE. Disabled when 0. + */ +static const u32 bbr_ecn_thresh = BBR_UNIT * 1 / 2; /* 1/2 = 50% */ + +/* If non-zero, if in a cycle with no losses but some ECN marks, after ECN + * clears then make the first round's increment to inflight_hi the following + * fraction of inflight_hi. + */ +static const u32 bbr_ecn_reprobe_gain = BBR_UNIT * 1 / 2; + +/* Estimate bw probing has gone too far if loss rate exceeds this level. */ +static const u32 bbr_loss_thresh = BBR_UNIT * 2 / 100; /* 2% loss */ + +/* Slow down for a packet loss recovered by TLP? */ +static const bool bbr_loss_probe_recovery = true; + +/* Exit STARTUP if number of loss marking events in a Recovery round is >= N, + * and loss rate is higher than bbr_loss_thresh. + * Disabled if 0. + */ +static const u32 bbr_full_loss_cnt = 6; + +/* Exit STARTUP if number of round trips with ECN mark rate above ecn_thresh + * meets this count. + */ +static const u32 bbr_full_ecn_cnt = 2; + +/* Fraction of unutilized headroom to try to leave in path upon high loss. */ +static const u32 bbr_inflight_headroom = BBR_UNIT * 15 / 100; + +/* How much do we increase cwnd_gain when probing for bandwidth in + * BBR_BW_PROBE_UP? This specifies the increment in units of + * BBR_UNIT/4. The default is 1, meaning 0.25. + * The min value is 0 (meaning 0.0); max is 3 (meaning 0.75). + */ +static const u32 bbr_bw_probe_cwnd_gain = 1; + +/* Max number of packet-timed rounds to wait before probing for bandwidth. If + * we want to tolerate 1% random loss per round, and not have this cut our + * inflight too much, we must probe for bw periodically on roughly this scale. + * If low, limits Reno/CUBIC coexistence; if high, limits loss tolerance. + * We aim to be fair with Reno/CUBIC up to a BDP of at least: + * BDP = 25Mbps * .030sec /(1514bytes) = 61.9 packets + */ +static const u32 bbr_bw_probe_max_rounds = 63; + +/* Max amount of randomness to inject in round counting for Reno-coexistence. + */ +static const u32 bbr_bw_probe_rand_rounds = 2; + +/* Use BBR-native probe time scale starting at this many usec. + * We aim to be fair with Reno/CUBIC up to an inter-loss time epoch of at least: + * BDP*RTT = 25Mbps * .030sec /(1514bytes) * 0.030sec = 1.9 secs + */ +static const u32 bbr_bw_probe_base_us = 2 * USEC_PER_SEC; /* 2 secs */ + +/* Use BBR-native probes spread over this many usec: */ +static const u32 bbr_bw_probe_rand_us = 1 * USEC_PER_SEC; /* 1 secs */ + +/* Use fast path if app-limited, no loss/ECN, and target cwnd was reached? */ +static const bool bbr_fast_path = true; + +/* Use fast ack mode? */ +static const bool bbr_fast_ack_mode = true; + +static u32 bbr_max_bw(const struct sock *sk); +static u32 bbr_bw(const struct sock *sk); +static void bbr_exit_probe_rtt(struct sock *sk); +static void bbr_reset_congestion_signals(struct sock *sk); +static void bbr_run_loss_probe_recovery(struct sock *sk); + static void bbr_check_probe_rtt_done(struct sock *sk); +/* This connection can use ECN if both endpoints have signaled ECN support in + * the handshake and the per-route settings indicated this is a + * shallow-threshold ECN environment, meaning both: + * (a) ECN CE marks indicate low-latency/shallow-threshold congestion, and + * (b) TCP endpoints provide precise ACKs that only ACK data segments + * with consistent ECN CE status + */ +static bool bbr_can_use_ecn(const struct sock *sk) +{ + return (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) && + (tcp_sk(sk)->ecn_flags & TCP_ECN_LOW); +} + /* Do we estimate that STARTUP filled the pipe? */ static bool bbr_full_bw_reached(const struct sock *sk) { @@ -214,17 +382,17 @@ static bool bbr_full_bw_reached(const struct sock *sk) /* 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); + const struct bbr *bbr = inet_csk_ca(sk); - return minmax_get(&bbr->bw); + return max(bbr->bw_hi[0], bbr->bw_hi[1]); } /* 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); + const struct bbr *bbr = inet_csk_ca(sk); - return bbr->lt_use_bw ? bbr->lt_bw : bbr_max_bw(sk); + return min(bbr_max_bw(sk), bbr->bw_lo); } /* Return maximum extra acked in past k-2k round trips, @@ -241,15 +409,23 @@ static u16 bbr_extra_acked(const struct sock *sk) * 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) +static u64 bbr_rate_bytes_per_sec(struct sock *sk, u64 rate, int gain, + int margin) { unsigned int mss = tcp_sk(sk)->mss_cache; rate *= mss; rate *= gain; rate >>= BBR_SCALE; - rate *= USEC_PER_SEC / 100 * (100 - bbr_pacing_margin_percent); - return rate >> BW_SCALE; + rate *= USEC_PER_SEC / 100 * (100 - margin); + rate >>= BW_SCALE; + rate = max(rate, 1ULL); + return rate; +} + +static u64 bbr_bw_bytes_per_sec(struct sock *sk, u64 rate) +{ + return bbr_rate_bytes_per_sec(sk, rate, BBR_UNIT, 0); } /* Convert a BBR bw and gain factor to a pacing rate in bytes per second. */ @@ -257,12 +433,13 @@ static unsigned long bbr_bw_to_pacing_rate(struct sock *sk, u32 bw, int gain) { u64 rate = bw; - rate = bbr_rate_bytes_per_sec(sk, rate, gain); + rate = bbr_rate_bytes_per_sec(sk, rate, gain, + bbr_pacing_margin_percent); rate = min_t(u64, rate, sk->sk_max_pacing_rate); return rate; } -/* Initialize pacing rate to: high_gain * init_cwnd / RTT. */ +/* Initialize pacing rate to: startup_pacing_gain * init_cwnd / RTT. */ static void bbr_init_pacing_rate_from_rtt(struct sock *sk) { struct tcp_sock *tp = tcp_sk(sk); @@ -278,7 +455,8 @@ static void bbr_init_pacing_rate_from_rtt(struct sock *sk) } bw = (u64)tcp_snd_cwnd(tp) * BW_UNIT; do_div(bw, rtt_us); - sk->sk_pacing_rate = bbr_bw_to_pacing_rate(sk, bw, bbr_high_gain); + sk->sk_pacing_rate = + bbr_bw_to_pacing_rate(sk, bw, bbr_param(sk, startup_pacing_gain)); } /* Pace using current bw estimate and a gain factor. */ @@ -294,31 +472,38 @@ static void bbr_set_pacing_rate(struct sock *sk, u32 bw, int gain) sk->sk_pacing_rate = rate; } -/* override sysctl_tcp_min_tso_segs */ -static u32 bbr_min_tso_segs(struct sock *sk) -{ - return sk->sk_pacing_rate < (bbr_min_tso_rate >> 3) ? 1 : 2; -} - -/* Return the number of segments BBR would like in a TSO/GSO skb, given - * a particular max gso size as a constraint. +/* Return the number of segments BBR would like in a TSO/GSO skb, given a + * particular max gso size as a constraint. TODO: make this simpler and more + * consistent by switching bbr to just call tcp_tso_autosize(). */ static u32 bbr_tso_segs_generic(struct sock *sk, unsigned int mss_now, u32 gso_max_size) { - u32 segs; + struct bbr *bbr = inet_csk_ca(sk); + u32 segs, r; u64 bytes; /* Budget a TSO/GSO burst size allowance based on bw (pacing_rate). */ bytes = sk->sk_pacing_rate >> sk->sk_pacing_shift; + /* Budget a TSO/GSO burst size allowance based on min_rtt. For every + * K = 2^tso_rtt_shift microseconds of min_rtt, halve the burst. + * The min_rtt-based burst allowance is: 64 KBytes / 2^(min_rtt/K) + */ + if (bbr_param(sk, tso_rtt_shift)) { + r = bbr->min_rtt_us >> bbr_param(sk, tso_rtt_shift); + if (r < BITS_PER_TYPE(u32)) /* prevent undefined behavior */ + bytes += GSO_LEGACY_MAX_SIZE >> r; + } + bytes = min_t(u32, bytes, gso_max_size - 1 - MAX_TCP_HEADER); - segs = max_t(u32, bytes / mss_now, bbr_min_tso_segs(sk)); + segs = max_t(u32, bytes / mss_now, + sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); return segs; } /* Custom tcp_tso_autosize() for BBR, used at transmit time to cap skb size. */ -static u32 bbr_tso_segs(struct sock *sk, unsigned int mss_now) +static u32 bbr_tso_segs(struct sock *sk, unsigned int mss_now) { return bbr_tso_segs_generic(sk, mss_now, sk->sk_gso_max_size); } @@ -328,7 +513,7 @@ static u32 bbr_tso_segs_goal(struct sock *sk) { struct tcp_sock *tp = tcp_sk(sk); - return bbr_tso_segs_generic(sk, tp->mss_cache, GSO_MAX_SIZE); + return bbr_tso_segs_generic(sk, tp->mss_cache, GSO_LEGACY_MAX_SIZE); } /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ @@ -348,7 +533,9 @@ 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) { + if (event == CA_EVENT_TX_START) { + if (!tp->app_limited) + return; bbr->idle_restart = 1; bbr->ack_epoch_mstamp = tp->tcp_mstamp; bbr->ack_epoch_acked = 0; @@ -359,6 +546,16 @@ static void bbr_cwnd_event(struct sock *sk, enum tcp_ca_event event) bbr_set_pacing_rate(sk, bbr_bw(sk), BBR_UNIT); else if (bbr->mode == BBR_PROBE_RTT) bbr_check_probe_rtt_done(sk); + } else if ((event == CA_EVENT_ECN_IS_CE || + event == CA_EVENT_ECN_NO_CE) && + bbr_can_use_ecn(sk) && + bbr_param(sk, precise_ece_ack)) { + u32 state = bbr->ce_state; + dctcp_ece_ack_update(sk, event, &bbr->prior_rcv_nxt, &state); + bbr->ce_state = state; + } else if (event == CA_EVENT_TLP_RECOVERY && + bbr_param(sk, loss_probe_recovery)) { + bbr_run_loss_probe_recovery(sk); } } @@ -381,10 +578,10 @@ static u32 bbr_bdp(struct sock *sk, u32 bw, int gain) * 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. + * case we need to slow-start up toward something safe: initial cwnd. */ if (unlikely(bbr->min_rtt_us == ~0U)) /* no valid RTT samples yet? */ - return TCP_INIT_CWND; /* be safe: cap at default initial cwnd*/ + return bbr->init_cwnd; /* be safe: cap at initial cwnd */ w = (u64)bw * bbr->min_rtt_us; @@ -401,23 +598,23 @@ static u32 bbr_bdp(struct sock *sk, u32 bw, int gain) * - 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, + * Don't worry, at low rates this won't bloat cwnd because + * in such cases tso_segs_goal is small. 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_quantization_budget(struct sock *sk, u32 cwnd) { struct bbr *bbr = inet_csk_ca(sk); + u32 tso_segs_goal; - /* Allow enough full-sized skbs in flight to utilize end systems. */ - cwnd += 3 * bbr_tso_segs_goal(sk); - - /* Reduce delayed ACKs by rounding up cwnd to the next even number. */ - cwnd = (cwnd + 1) & ~1U; + tso_segs_goal = 3 * bbr_tso_segs_goal(sk); + /* Allow enough full-sized skbs in flight to utilize end systems. */ + cwnd = max_t(u32, cwnd, tso_segs_goal); + cwnd = max_t(u32, cwnd, bbr_param(sk, cwnd_min_target)); /* Ensure gain cycling gets inflight above BDP even for small BDPs. */ - if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == 0) + if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == BBR_BW_PROBE_UP) cwnd += 2; return cwnd; @@ -472,10 +669,10 @@ static u32 bbr_ack_aggregation_cwnd(struct sock *sk) { u32 max_aggr_cwnd, aggr_cwnd = 0; - if (bbr_extra_acked_gain && bbr_full_bw_reached(sk)) { + if (bbr_param(sk, extra_acked_gain)) { max_aggr_cwnd = ((u64)bbr_bw(sk) * bbr_extra_acked_max_us) / BW_UNIT; - aggr_cwnd = (bbr_extra_acked_gain * bbr_extra_acked(sk)) + aggr_cwnd = (bbr_param(sk, extra_acked_gain) * bbr_extra_acked(sk)) >> BBR_SCALE; aggr_cwnd = min(aggr_cwnd, max_aggr_cwnd); } @@ -483,66 +680,27 @@ static u32 bbr_ack_aggregation_cwnd(struct sock *sk) return aggr_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) +/* Returns the cwnd for PROBE_RTT mode. */ +static u32 bbr_probe_rtt_cwnd(struct sock *sk) { - 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 = tcp_snd_cwnd(tp); - - /* 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. */ - cwnd = max(cwnd, bbr->prior_cwnd); - bbr->packet_conservation = 0; - } - bbr->prev_ca_state = state; - - 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; + return max_t(u32, bbr_param(sk, cwnd_min_target), + bbr_bdp(sk, bbr_bw(sk), bbr_param(sk, probe_rtt_cwnd_gain))); } /* 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) + u32 acked, u32 bw, int gain, u32 cwnd, + struct bbr_context *ctx) { struct tcp_sock *tp = tcp_sk(sk); struct bbr *bbr = inet_csk_ca(sk); - u32 cwnd = tcp_snd_cwnd(tp), target_cwnd = 0; + u32 target_cwnd = 0; if (!acked) goto done; /* no packet fully ACKed; just apply caps */ - if (bbr_set_cwnd_to_recover_or_restore(sk, rs, acked, &cwnd)) - goto done; - target_cwnd = bbr_bdp(sk, bw, gain); /* Increment the cwnd to account for excess ACKed data that seems @@ -551,74 +709,26 @@ static void bbr_set_cwnd(struct sock *sk, const struct rate_sample *rs, target_cwnd += bbr_ack_aggregation_cwnd(sk); target_cwnd = bbr_quantization_budget(sk, target_cwnd); - /* If we're below target cwnd, slow start cwnd toward target cwnd. */ - 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); + /* Update cwnd and enable fast path if cwnd reaches target_cwnd. */ + bbr->try_fast_path = 0; + if (bbr_full_bw_reached(sk)) { /* only cut cwnd if we filled the pipe */ + cwnd += acked; + if (cwnd >= target_cwnd) { + cwnd = target_cwnd; + bbr->try_fast_path = 1; + } + } else if (cwnd < target_cwnd || cwnd < 2 * bbr->init_cwnd) { + cwnd += acked; + } else { + bbr->try_fast_path = 1; + } + cwnd = max_t(u32, cwnd, bbr_param(sk, cwnd_min_target)); done: - tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp)); /* apply global cap */ + tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp)); /* global cap */ if (bbr->mode == BBR_PROBE_RTT) /* drain queue, refresh min_rtt */ - tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), 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 = - tcp_stamp_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 = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight); - 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_inflight(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_inflight(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; -} - -/* 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_is_next_cycle_phase(sk, rs)) - bbr_advance_cycle_phase(sk); + tcp_snd_cwnd_set(tp, min_t(u32, tcp_snd_cwnd(tp), + bbr_probe_rtt_cwnd(sk))); } static void bbr_reset_startup_mode(struct sock *sk) @@ -628,191 +738,49 @@ static void bbr_reset_startup_mode(struct sock *sk) bbr->mode = BBR_STARTUP; } -static void bbr_reset_probe_bw_mode(struct sock *sk) -{ - struct bbr *bbr = inet_csk_ca(sk); - - bbr->mode = BBR_PROBE_BW; - 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 = div_u64(tp->delivered_mstamp, USEC_PER_MSEC); - 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. +/* See if we have reached next round trip. Upon start of the new round, + * returns packets delivered since previous round start plus this ACK. */ -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; - u32 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 = div_u64(tp->delivered_mstamp, USEC_PER_MSEC) - bbr->lt_last_stamp; - if ((s32)t < 1) - return; /* interval is less than one ms, so wait */ - /* Check if can multiply without overflow */ - if (t >= ~0U / USEC_PER_MSEC) { - bbr_reset_lt_bw_sampling(sk); /* interval too long; reset */ - return; - } - t *= USEC_PER_MSEC; - 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) +static u32 bbr_update_round_start(struct sock *sk, + const struct rate_sample *rs, struct bbr_context *ctx) { struct tcp_sock *tp = tcp_sk(sk); struct bbr *bbr = inet_csk_ca(sk); - u64 bw; + u32 round_delivered = 0; 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)) { + if (rs->interval_us > 0 && + !before(rs->prior_delivered, bbr->next_rtt_delivered)) { + round_delivered = tp->delivered - bbr->next_rtt_delivered; bbr->next_rtt_delivered = tp->delivered; - bbr->rtt_cnt++; bbr->round_start = 1; - bbr->packet_conservation = 0; } + return round_delivered; +} - bbr_lt_bw_sampling(sk, rs); +/* Calculate the bandwidth based on how fast packets are delivered */ +static void bbr_calculate_bw_sample(struct sock *sk, + const struct rate_sample *rs, struct bbr_context *ctx) +{ + u64 bw = 0; /* 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. + * Round up to allow growth at low rates, even with integer division. */ - bw = div64_long((u64)rs->delivered * BW_UNIT, 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); + if (rs->interval_us > 0) { + if (WARN_ONCE(rs->delivered < 0, + "negative delivered: %d interval_us: %ld\n", + rs->delivered, rs->interval_us)) + return; + + bw = DIV_ROUND_UP_ULL((u64)rs->delivered * BW_UNIT, rs->interval_us); } + + ctx->sample_bw = bw; } /* Estimates the windowed max degree of ack aggregation. @@ -826,7 +794,7 @@ static void bbr_update_bw(struct sock *sk, const struct rate_sample *rs) * * Max extra_acked is clamped by cwnd and bw * bbr_extra_acked_max_us (100 ms). * Max filter is an approximate sliding window of 5-10 (packet timed) round - * trips. + * trips for non-startup phase, and 1-2 round trips for startup. */ static void bbr_update_ack_aggregation(struct sock *sk, const struct rate_sample *rs) @@ -834,15 +802,19 @@ static void bbr_update_ack_aggregation(struct sock *sk, u32 epoch_us, expected_acked, extra_acked; struct bbr *bbr = inet_csk_ca(sk); struct tcp_sock *tp = tcp_sk(sk); + u32 extra_acked_win_rtts_thresh = bbr_param(sk, extra_acked_win_rtts); - if (!bbr_extra_acked_gain || rs->acked_sacked <= 0 || + if (!bbr_param(sk, extra_acked_gain) || rs->acked_sacked <= 0 || rs->delivered < 0 || rs->interval_us <= 0) return; if (bbr->round_start) { bbr->extra_acked_win_rtts = min(0x1F, bbr->extra_acked_win_rtts + 1); - if (bbr->extra_acked_win_rtts >= bbr_extra_acked_win_rtts) { + if (!bbr_full_bw_reached(sk)) + extra_acked_win_rtts_thresh = 1; + if (bbr->extra_acked_win_rtts >= + extra_acked_win_rtts_thresh) { bbr->extra_acked_win_rtts = 0; bbr->extra_acked_win_idx = bbr->extra_acked_win_idx ? 0 : 1; @@ -876,49 +848,6 @@ static void bbr_update_ack_aggregation(struct sock *sk, bbr->extra_acked[bbr->extra_acked_win_idx] = extra_acked; } -/* 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; - bbr->full_bw_reached = bbr->full_bw_cnt >= 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 */ - tcp_sk(sk)->snd_ssthresh = - bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT); - } /* fall through to check if in-flight is already small: */ - if (bbr->mode == BBR_DRAIN && - bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <= - bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT)) - bbr_reset_probe_bw_mode(sk); /* we estimate queue is drained */ -} - static void bbr_check_probe_rtt_done(struct sock *sk) { struct tcp_sock *tp = tcp_sk(sk); @@ -928,9 +857,9 @@ static void bbr_check_probe_rtt_done(struct sock *sk) after(tcp_jiffies32, bbr->probe_rtt_done_stamp))) return; - bbr->min_rtt_stamp = tcp_jiffies32; /* wait a while until PROBE_RTT */ + bbr->probe_rtt_min_stamp = tcp_jiffies32; /* schedule next PROBE_RTT */ tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp), bbr->prior_cwnd)); - bbr_reset_mode(sk); + bbr_exit_probe_rtt(sk); } /* The goal of PROBE_RTT mode is to have BBR flows cooperatively and @@ -956,23 +885,35 @@ 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; + bool probe_rtt_expired, min_rtt_expired; + u32 expire; - /* Track min RTT seen in the min_rtt_win_sec filter window: */ - filter_expired = after(tcp_jiffies32, - bbr->min_rtt_stamp + bbr_min_rtt_win_sec * HZ); + /* Track min RTT in probe_rtt_win_ms to time next PROBE_RTT state. */ + expire = bbr->probe_rtt_min_stamp + + msecs_to_jiffies(bbr_param(sk, probe_rtt_win_ms)); + probe_rtt_expired = after(tcp_jiffies32, expire); if (rs->rtt_us >= 0 && - (rs->rtt_us < bbr->min_rtt_us || - (filter_expired && !rs->is_ack_delayed))) { - bbr->min_rtt_us = rs->rtt_us; - bbr->min_rtt_stamp = tcp_jiffies32; + (rs->rtt_us < bbr->probe_rtt_min_us || + (probe_rtt_expired && !rs->is_ack_delayed))) { + bbr->probe_rtt_min_us = rs->rtt_us; + bbr->probe_rtt_min_stamp = tcp_jiffies32; + } + /* Track min RTT seen in the min_rtt_win_sec filter window: */ + expire = bbr->min_rtt_stamp + bbr_param(sk, min_rtt_win_sec) * HZ; + min_rtt_expired = after(tcp_jiffies32, expire); + if (bbr->probe_rtt_min_us <= bbr->min_rtt_us || + min_rtt_expired) { + bbr->min_rtt_us = bbr->probe_rtt_min_us; + bbr->min_rtt_stamp = bbr->probe_rtt_min_stamp; } - if (bbr_probe_rtt_mode_ms > 0 && filter_expired && + if (bbr_param(sk, probe_rtt_mode_ms) > 0 && probe_rtt_expired && !bbr->idle_restart && bbr->mode != BBR_PROBE_RTT) { bbr->mode = BBR_PROBE_RTT; /* dip, drain queue */ bbr_save_cwnd(sk); /* note cwnd so we can restore it */ bbr->probe_rtt_done_stamp = 0; + bbr->ack_phase = BBR_ACKS_PROBE_STOPPING; + bbr->next_rtt_delivered = tp->delivered; } if (bbr->mode == BBR_PROBE_RTT) { @@ -981,9 +922,9 @@ static void bbr_update_min_rtt(struct sock *sk, const struct rate_sample *rs) (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) { + tcp_packets_in_flight(tp) <= bbr_probe_rtt_cwnd(sk)) { bbr->probe_rtt_done_stamp = tcp_jiffies32 + - msecs_to_jiffies(bbr_probe_rtt_mode_ms); + msecs_to_jiffies(bbr_param(sk, probe_rtt_mode_ms)); bbr->probe_rtt_round_done = 0; bbr->next_rtt_delivered = tp->delivered; } else if (bbr->probe_rtt_done_stamp) { @@ -1004,18 +945,20 @@ static void bbr_update_gains(struct sock *sk) switch (bbr->mode) { case BBR_STARTUP: - bbr->pacing_gain = bbr_high_gain; - bbr->cwnd_gain = bbr_high_gain; + bbr->pacing_gain = bbr_param(sk, startup_pacing_gain); + bbr->cwnd_gain = bbr_param(sk, startup_cwnd_gain); break; case BBR_DRAIN: - bbr->pacing_gain = bbr_drain_gain; /* slow, to drain */ - bbr->cwnd_gain = bbr_high_gain; /* keep cwnd */ + bbr->pacing_gain = bbr_param(sk, drain_gain); /* slow, to drain */ + bbr->cwnd_gain = bbr_param(sk, startup_cwnd_gain); /* keep cwnd */ break; case BBR_PROBE_BW: - bbr->pacing_gain = (bbr->lt_use_bw ? - BBR_UNIT : - bbr_pacing_gain[bbr->cycle_idx]); - bbr->cwnd_gain = bbr_cwnd_gain; + bbr->pacing_gain = bbr_pacing_gain[bbr->cycle_idx]; + bbr->cwnd_gain = bbr_param(sk, cwnd_gain); + if (bbr_param(sk, bw_probe_cwnd_gain) && + bbr->cycle_idx == BBR_BW_PROBE_UP) + bbr->cwnd_gain += + BBR_UNIT * bbr_param(sk, bw_probe_cwnd_gain) / 4; break; case BBR_PROBE_RTT: bbr->pacing_gain = BBR_UNIT; @@ -1027,140 +970,1380 @@ static void bbr_update_gains(struct sock *sk) } } -static void bbr_update_model(struct sock *sk, const struct rate_sample *rs) +static u32 bbr_sndbuf_expand(struct sock *sk) { - bbr_update_bw(sk, rs); - bbr_update_ack_aggregation(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); - bbr_update_gains(sk); + /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ + return 3; } -static void bbr_main(struct sock *sk, const struct rate_sample *rs) +/* Incorporate a new bw sample into the current window of our max filter. */ +static void bbr_take_max_bw_sample(struct sock *sk, u32 bw) { 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_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain); + bbr->bw_hi[1] = max(bw, bbr->bw_hi[1]); } -static void bbr_init(struct sock *sk) +/* Keep max of last 1-2 cycles. Each PROBE_BW cycle, flip filter window. */ +static void bbr_advance_max_bw_filter(struct sock *sk) { - struct tcp_sock *tp = tcp_sk(sk); struct bbr *bbr = inet_csk_ca(sk); - bbr->prior_cwnd = 0; - tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; - bbr->rtt_cnt = 0; - bbr->next_rtt_delivered = tp->delivered; - 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_jiffies32; - - minmax_reset(&bbr->bw, bbr->rtt_cnt, 0); /* init max bw to 0 */ + if (!bbr->bw_hi[1]) + return; /* no samples in this window; remember old window */ + bbr->bw_hi[0] = bbr->bw_hi[1]; + bbr->bw_hi[1] = 0; +} - bbr->has_seen_rtt = 0; - bbr_init_pacing_rate_from_rtt(sk); +/* Reset the estimator for reaching full bandwidth based on bw plateau. */ +static void bbr_reset_full_bw(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); - bbr->round_start = 0; - bbr->idle_restart = 0; - bbr->full_bw_reached = 0; bbr->full_bw = 0; bbr->full_bw_cnt = 0; - bbr->cycle_mstamp = 0; - bbr->cycle_idx = 0; - bbr_reset_lt_bw_sampling(sk); - bbr_reset_startup_mode(sk); - - bbr->ack_epoch_mstamp = tp->tcp_mstamp; - bbr->ack_epoch_acked = 0; - bbr->extra_acked_win_rtts = 0; - bbr->extra_acked_win_idx = 0; - bbr->extra_acked[0] = 0; - bbr->extra_acked[1] = 0; - - cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED); + bbr->full_bw_now = 0; } -static u32 bbr_sndbuf_expand(struct sock *sk) +/* How much do we want in flight? Our BDP, unless congestion cut cwnd. */ +static u32 bbr_target_inflight(struct sock *sk) { - /* Provision 3 * cwnd since BBR may slow-start even during recovery. */ - return 3; + u32 bdp = bbr_inflight(sk, bbr_bw(sk), BBR_UNIT); + + return min(bdp, tcp_sk(sk)->snd_cwnd); } -/* 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) +static bool bbr_is_probing_bandwidth(struct sock *sk) { struct bbr *bbr = inet_csk_ca(sk); - bbr->full_bw = 0; /* spurious slow-down; reset full pipe detection */ - bbr->full_bw_cnt = 0; - bbr_reset_lt_bw_sampling(sk); - return tcp_snd_cwnd(tcp_sk(sk)); + return (bbr->mode == BBR_STARTUP) || + (bbr->mode == BBR_PROBE_BW && + (bbr->cycle_idx == BBR_BW_PROBE_REFILL || + bbr->cycle_idx == BBR_BW_PROBE_UP)); } -/* Entering loss recovery, so save cwnd for when we exit or undo recovery. */ +/* Has the given amount of time elapsed since we marked the phase start? */ +static bool bbr_has_elapsed_in_phase(const struct sock *sk, u32 interval_us) +{ + const struct tcp_sock *tp = tcp_sk(sk); + const struct bbr *bbr = inet_csk_ca(sk); + + return tcp_stamp_us_delta(tp->tcp_mstamp, + bbr->cycle_mstamp + interval_us) > 0; +} + +static void bbr_handle_queue_too_high_in_startup(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 bdp; /* estimated BDP in packets, with quantization budget */ + + bbr->full_bw_reached = 1; + + bdp = bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT); + bbr->inflight_hi = max(bdp, bbr->inflight_latest); +} + +/* Exit STARTUP upon N consecutive rounds with ECN mark rate > ecn_thresh. */ +static void bbr_check_ecn_too_high_in_startup(struct sock *sk, u32 ce_ratio) +{ + struct bbr *bbr = inet_csk_ca(sk); + + if (bbr_full_bw_reached(sk) || !bbr->ecn_eligible || + !bbr_param(sk, full_ecn_cnt) || !bbr_param(sk, ecn_thresh)) + return; + + if (ce_ratio >= bbr_param(sk, ecn_thresh)) + bbr->startup_ecn_rounds++; + else + bbr->startup_ecn_rounds = 0; + + if (bbr->startup_ecn_rounds >= bbr_param(sk, full_ecn_cnt)) { + bbr_handle_queue_too_high_in_startup(sk); + return; + } +} + +/* Updates ecn_alpha and returns ce_ratio. -1 if not available. */ +static int bbr_update_ecn_alpha(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct net *net = sock_net(sk); + struct bbr *bbr = inet_csk_ca(sk); + s32 delivered, delivered_ce; + u64 alpha, ce_ratio; + u32 gain; + bool want_ecn_alpha; + + /* See if we should use ECN sender logic for this connection. */ + if (!bbr->ecn_eligible && bbr_can_use_ecn(sk) && + bbr_param(sk, ecn_factor) && + (bbr->min_rtt_us <= bbr_ecn_max_rtt_us || + !bbr_ecn_max_rtt_us)) + bbr->ecn_eligible = 1; + + /* Skip updating alpha only if not ECN-eligible and PLB is disabled. */ + want_ecn_alpha = (bbr->ecn_eligible || + (bbr_can_use_ecn(sk) && + READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled))); + if (!want_ecn_alpha) + return -1; + + delivered = tp->delivered - bbr->alpha_last_delivered; + delivered_ce = tp->delivered_ce - bbr->alpha_last_delivered_ce; + + if (delivered == 0 || /* avoid divide by zero */ + WARN_ON_ONCE(delivered < 0 || delivered_ce < 0)) /* backwards? */ + return -1; + + BUILD_BUG_ON(BBR_SCALE != TCP_PLB_SCALE); + ce_ratio = (u64)delivered_ce << BBR_SCALE; + do_div(ce_ratio, delivered); + + gain = bbr_param(sk, ecn_alpha_gain); + alpha = ((BBR_UNIT - gain) * bbr->ecn_alpha) >> BBR_SCALE; + alpha += (gain * ce_ratio) >> BBR_SCALE; + bbr->ecn_alpha = min_t(u32, alpha, BBR_UNIT); + + bbr->alpha_last_delivered = tp->delivered; + bbr->alpha_last_delivered_ce = tp->delivered_ce; + + bbr_check_ecn_too_high_in_startup(sk, ce_ratio); + return (int)ce_ratio; +} + +/* Protective Load Balancing (PLB). PLB rehashes outgoing data (to a new IPv6 + * flow label) if it encounters sustained congestion in the form of ECN marks. + */ +static void bbr_plb(struct sock *sk, const struct rate_sample *rs, int ce_ratio) +{ + struct bbr *bbr = inet_csk_ca(sk); + + if (bbr->round_start && ce_ratio >= 0) + tcp_plb_update_state(sk, &bbr->plb, ce_ratio); + + tcp_plb_check_rehash(sk, &bbr->plb); +} + +/* Each round trip of BBR_BW_PROBE_UP, double volume of probing data. */ +static void bbr_raise_inflight_hi_slope(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + u32 growth_this_round, cnt; + + /* Calculate "slope": packets S/Acked per inflight_hi increment. */ + growth_this_round = 1 << bbr->bw_probe_up_rounds; + bbr->bw_probe_up_rounds = min(bbr->bw_probe_up_rounds + 1, 30); + cnt = tcp_snd_cwnd(tp) / growth_this_round; + cnt = max(cnt, 1U); + bbr->bw_probe_up_cnt = cnt; +} + +/* In BBR_BW_PROBE_UP, not seeing high loss/ECN/queue, so raise inflight_hi. */ +static void bbr_probe_inflight_hi_upward(struct sock *sk, + const struct rate_sample *rs) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + u32 delta; + + if (!tp->is_cwnd_limited || tcp_snd_cwnd(tp) < bbr->inflight_hi) + return; /* not fully using inflight_hi, so don't grow it */ + + /* For each bw_probe_up_cnt packets ACKed, increase inflight_hi by 1. */ + bbr->bw_probe_up_acks += rs->acked_sacked; + if (bbr->bw_probe_up_acks >= bbr->bw_probe_up_cnt) { + delta = bbr->bw_probe_up_acks / bbr->bw_probe_up_cnt; + bbr->bw_probe_up_acks -= delta * bbr->bw_probe_up_cnt; + bbr->inflight_hi += delta; + bbr->try_fast_path = 0; /* Need to update cwnd */ + } + + if (bbr->round_start) + bbr_raise_inflight_hi_slope(sk); +} + +/* Does loss/ECN rate for this sample say inflight is "too high"? + * This is used by both the bbr_check_loss_too_high_in_startup() function, + * which can be used in either v1 or v2, and the PROBE_UP phase of v2, which + * uses it to notice when loss/ECN rates suggest inflight is too high. + */ +static bool bbr_is_inflight_too_high(const struct sock *sk, + const struct rate_sample *rs) +{ + const struct bbr *bbr = inet_csk_ca(sk); + u32 loss_thresh, ecn_thresh; + + if (rs->lost > 0 && rs->tx_in_flight) { + loss_thresh = (u64)rs->tx_in_flight * bbr_param(sk, loss_thresh) >> + BBR_SCALE; + if (rs->lost > loss_thresh) { + return true; + } + } + + if (rs->delivered_ce > 0 && rs->delivered > 0 && + bbr->ecn_eligible && bbr_param(sk, ecn_thresh)) { + ecn_thresh = (u64)rs->delivered * bbr_param(sk, ecn_thresh) >> + BBR_SCALE; + if (rs->delivered_ce > ecn_thresh) { + return true; + } + } + + return false; +} + +/* Calculate the tx_in_flight level that corresponded to excessive loss. + * We find "lost_prefix" segs of the skb where loss rate went too high, + * by solving for "lost_prefix" in the following equation: + * lost / inflight >= loss_thresh + * (lost_prev + lost_prefix) / (inflight_prev + lost_prefix) >= loss_thresh + * Then we take that equation, convert it to fixed point, and + * round up to the nearest packet. + */ +static u32 bbr_inflight_hi_from_lost_skb(const struct sock *sk, + const struct rate_sample *rs, + const struct sk_buff *skb) +{ + const struct tcp_sock *tp = tcp_sk(sk); + u32 loss_thresh = bbr_param(sk, loss_thresh); + u32 pcount, divisor, inflight_hi; + s32 inflight_prev, lost_prev; + u64 loss_budget, lost_prefix; + + pcount = tcp_skb_pcount(skb); + + /* How much data was in flight before this skb? */ + inflight_prev = rs->tx_in_flight - pcount; + if (inflight_prev < 0) { + WARN_ONCE(tcp_skb_tx_in_flight_is_suspicious( + pcount, + TCP_SKB_CB(skb)->sacked, + rs->tx_in_flight), + "tx_in_flight: %u pcount: %u reneg: %u", + rs->tx_in_flight, pcount, tcp_sk(sk)->is_sack_reneg); + return ~0U; + } + + /* How much inflight data was marked lost before this skb? */ + lost_prev = rs->lost - pcount; + if (WARN_ONCE(lost_prev < 0, + "cwnd: %u ca: %d out: %u lost: %u pif: %u " + "tx_in_flight: %u tx.lost: %u tp->lost: %u rs->lost: %d " + "lost_prev: %d pcount: %d seq: %u end_seq: %u reneg: %u", + tcp_snd_cwnd(tp), inet_csk(sk)->icsk_ca_state, + tp->packets_out, tp->lost_out, tcp_packets_in_flight(tp), + rs->tx_in_flight, TCP_SKB_CB(skb)->tx.lost, tp->lost, + rs->lost, lost_prev, pcount, + TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, + tp->is_sack_reneg)) + return ~0U; + + /* At what prefix of this lost skb did losss rate exceed loss_thresh? */ + loss_budget = (u64)inflight_prev * loss_thresh + BBR_UNIT - 1; + loss_budget >>= BBR_SCALE; + if (lost_prev >= loss_budget) { + lost_prefix = 0; /* previous losses crossed loss_thresh */ + } else { + lost_prefix = loss_budget - lost_prev; + lost_prefix <<= BBR_SCALE; + divisor = BBR_UNIT - loss_thresh; + if (WARN_ON_ONCE(!divisor)) /* loss_thresh is 8 bits */ + return ~0U; + do_div(lost_prefix, divisor); + } + + inflight_hi = inflight_prev + lost_prefix; + return inflight_hi; +} + +/* If loss/ECN rates during probing indicated we may have overfilled a + * buffer, return an operating point that tries to leave unutilized headroom in + * the path for other flows, for fairness convergence and lower RTTs and loss. + */ +static u32 bbr_inflight_with_headroom(const struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 headroom, headroom_fraction; + + if (bbr->inflight_hi == ~0U) + return ~0U; + + headroom_fraction = bbr_param(sk, inflight_headroom); + headroom = ((u64)bbr->inflight_hi * headroom_fraction) >> BBR_SCALE; + headroom = max(headroom, 1U); + return max_t(s32, bbr->inflight_hi - headroom, + bbr_param(sk, cwnd_min_target)); +} + +/* Bound cwnd to a sensible level, based on our current probing state + * machine phase and model of a good inflight level (inflight_lo, inflight_hi). + */ +static void bbr_bound_cwnd_for_inflight_model(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + u32 cap; + + /* tcp_rcv_synsent_state_process() currently calls tcp_ack() + * and thus cong_control() without first initializing us(!). + */ + if (!bbr->initialized) + return; + + cap = ~0U; + if (bbr->mode == BBR_PROBE_BW && + bbr->cycle_idx != BBR_BW_PROBE_CRUISE) { + /* Probe to see if more packets fit in the path. */ + cap = bbr->inflight_hi; + } else { + if (bbr->mode == BBR_PROBE_RTT || + (bbr->mode == BBR_PROBE_BW && + bbr->cycle_idx == BBR_BW_PROBE_CRUISE)) + cap = bbr_inflight_with_headroom(sk); + } + /* Adapt to any loss/ECN since our last bw probe. */ + cap = min(cap, bbr->inflight_lo); + + cap = max_t(u32, cap, bbr_param(sk, cwnd_min_target)); + tcp_snd_cwnd_set(tp, min(cap, tcp_snd_cwnd(tp))); +} + +/* How should we multiplicatively cut bw or inflight limits based on ECN? */ +u32 bbr_ecn_cut(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + return BBR_UNIT - + ((bbr->ecn_alpha * bbr_param(sk, ecn_factor)) >> BBR_SCALE); +} + +/* Init lower bounds if have not inited yet. */ +static void bbr_init_lower_bounds(struct sock *sk, bool init_bw) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + if (init_bw && bbr->bw_lo == ~0U) + bbr->bw_lo = bbr_max_bw(sk); + if (bbr->inflight_lo == ~0U) + bbr->inflight_lo = tcp_snd_cwnd(tp); +} + +/* Reduce bw and inflight to (1 - beta). */ +static void bbr_loss_lower_bounds(struct sock *sk, u32 *bw, u32 *inflight) +{ + struct bbr* bbr = inet_csk_ca(sk); + u32 loss_cut = BBR_UNIT - bbr_param(sk, beta); + + *bw = max_t(u32, bbr->bw_latest, + (u64)bbr->bw_lo * loss_cut >> BBR_SCALE); + *inflight = max_t(u32, bbr->inflight_latest, + (u64)bbr->inflight_lo * loss_cut >> BBR_SCALE); +} + +/* Reduce inflight to (1 - alpha*ecn_factor). */ +static void bbr_ecn_lower_bounds(struct sock *sk, u32 *inflight) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 ecn_cut = bbr_ecn_cut(sk); + + *inflight = (u64)bbr->inflight_lo * ecn_cut >> BBR_SCALE; +} + +/* Estimate a short-term lower bound on the capacity available now, based + * on measurements of the current delivery process and recent history. When we + * are seeing loss/ECN at times when we are not probing bw, then conservatively + * move toward flow balance by multiplicatively cutting our short-term + * estimated safe rate and volume of data (bw_lo and inflight_lo). We use a + * multiplicative decrease in order to converge to a lower capacity in time + * logarithmic in the magnitude of the decrease. + * + * However, we do not cut our short-term estimates lower than the current rate + * and volume of delivered data from this round trip, since from the current + * delivery process we can estimate the measured capacity available now. + * + * Anything faster than that approach would knowingly risk high loss, which can + * cause low bw for Reno/CUBIC and high loss recovery latency for + * request/response flows using any congestion control. + */ +static void bbr_adapt_lower_bounds(struct sock *sk, + const struct rate_sample *rs) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 ecn_inflight_lo = ~0U; + + /* We only use lower-bound estimates when not probing bw. + * When probing we need to push inflight higher to probe bw. + */ + if (bbr_is_probing_bandwidth(sk)) + return; + + /* ECN response. */ + if (bbr->ecn_in_round && bbr_param(sk, ecn_factor)) { + bbr_init_lower_bounds(sk, false); + bbr_ecn_lower_bounds(sk, &ecn_inflight_lo); + } + + /* Loss response. */ + if (bbr->loss_in_round) { + bbr_init_lower_bounds(sk, true); + bbr_loss_lower_bounds(sk, &bbr->bw_lo, &bbr->inflight_lo); + } + + /* Adjust to the lower of the levels implied by loss/ECN. */ + bbr->inflight_lo = min(bbr->inflight_lo, ecn_inflight_lo); + bbr->bw_lo = max(1U, bbr->bw_lo); +} + +/* Reset any short-term lower-bound adaptation to congestion, so that we can + * push our inflight up. + */ +static void bbr_reset_lower_bounds(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + bbr->bw_lo = ~0U; + bbr->inflight_lo = ~0U; +} + +/* After bw probing (STARTUP/PROBE_UP), reset signals before entering a state + * machine phase where we adapt our lower bound based on congestion signals. + */ +static void bbr_reset_congestion_signals(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + bbr->loss_in_round = 0; + bbr->ecn_in_round = 0; + bbr->loss_in_cycle = 0; + bbr->ecn_in_cycle = 0; + bbr->bw_latest = 0; + bbr->inflight_latest = 0; +} + +static void bbr_exit_loss_recovery(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + tcp_snd_cwnd_set(tp, max(tcp_snd_cwnd(tp), bbr->prior_cwnd)); + bbr->try_fast_path = 0; /* bound cwnd using latest model */ +} + +/* Update rate and volume of delivered data from latest round trip. */ +static void bbr_update_latest_delivery_signals( + struct sock *sk, const struct rate_sample *rs, struct bbr_context *ctx) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + bbr->loss_round_start = 0; + if (rs->interval_us <= 0 || !rs->acked_sacked) + return; /* Not a valid observation */ + + bbr->bw_latest = max_t(u32, bbr->bw_latest, ctx->sample_bw); + bbr->inflight_latest = max_t(u32, bbr->inflight_latest, rs->delivered); + + if (!before(rs->prior_delivered, bbr->loss_round_delivered)) { + bbr->loss_round_delivered = tp->delivered; + bbr->loss_round_start = 1; /* mark start of new round trip */ + } +} + +/* Once per round, reset filter for latest rate and volume of delivered data. */ +static void bbr_advance_latest_delivery_signals( + struct sock *sk, const struct rate_sample *rs, struct bbr_context *ctx) +{ + struct bbr *bbr = inet_csk_ca(sk); + + /* If ACK matches a TLP retransmit, persist the filter. If we detect + * that a TLP retransmit plugged a tail loss, we'll want to remember + * how much data the path delivered before the tail loss. + */ + if (bbr->loss_round_start && !rs->is_acking_tlp_retrans_seq) { + bbr->bw_latest = ctx->sample_bw; + bbr->inflight_latest = rs->delivered; + } +} + +/* Update (most of) our congestion signals: track the recent rate and volume of + * delivered data, presence of loss, and EWMA degree of ECN marking. + */ +static void bbr_update_congestion_signals( + struct sock *sk, const struct rate_sample *rs, struct bbr_context *ctx) +{ + struct bbr *bbr = inet_csk_ca(sk); + u64 bw; + + if (rs->interval_us <= 0 || !rs->acked_sacked) + return; /* Not a valid observation */ + bw = ctx->sample_bw; + + if (!rs->is_app_limited || bw >= bbr_max_bw(sk)) + bbr_take_max_bw_sample(sk, bw); + + bbr->loss_in_round |= (rs->losses > 0); + + if (!bbr->loss_round_start) + return; /* skip the per-round-trip updates */ + /* Now do per-round-trip updates. */ + bbr_adapt_lower_bounds(sk, rs); + + bbr->loss_in_round = 0; + bbr->ecn_in_round = 0; +} + +/* Bandwidth probing can cause loss. To help coexistence with loss-based + * congestion control we spread out our probing in a Reno-conscious way. Due to + * the shape of the Reno sawtooth, the time required between loss epochs for an + * idealized Reno flow is a number of round trips that is the BDP of that + * flow. We count packet-timed round trips directly, since measured RTT can + * vary widely, and Reno is driven by packet-timed round trips. + */ +static bool bbr_is_reno_coexistence_probe_time(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 rounds; + + /* Random loss can shave some small percentage off of our inflight + * in each round. To survive this, flows need robust periodic probes. + */ + rounds = min_t(u32, bbr_param(sk, bw_probe_max_rounds), bbr_target_inflight(sk)); + return bbr->rounds_since_probe >= rounds; +} + +/* How long do we want to wait before probing for bandwidth (and risking + * loss)? We randomize the wait, for better mixing and fairness convergence. + * + * We bound the Reno-coexistence inter-bw-probe time to be 62-63 round trips. + * This is calculated to allow fairness with a 25Mbps, 30ms Reno flow, + * (eg 4K video to a broadband user): + * BDP = 25Mbps * .030sec /(1514bytes) = 61.9 packets + * + * We bound the BBR-native inter-bw-probe wall clock time to be: + * (a) higher than 2 sec: to try to avoid causing loss for a long enough time + * to allow Reno at 30ms to get 4K video bw, the inter-bw-probe time must + * be at least: 25Mbps * .030sec / (1514bytes) * 0.030sec = 1.9secs + * (b) lower than 3 sec: to ensure flows can start probing in a reasonable + * amount of time to discover unutilized bw on human-scale interactive + * time-scales (e.g. perhaps traffic from a web page download that we + * were competing with is now complete). + */ +static void bbr_pick_probe_wait(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + /* Decide the random round-trip bound for wait until probe: */ + bbr->rounds_since_probe = + get_random_u32_below(bbr_param(sk, bw_probe_rand_rounds)); + /* Decide the random wall clock bound for wait until probe: */ + bbr->probe_wait_us = bbr_param(sk, bw_probe_base_us) + + get_random_u32_below(bbr_param(sk, bw_probe_rand_us)); +} + +static void bbr_set_cycle_idx(struct sock *sk, int cycle_idx) +{ + struct bbr *bbr = inet_csk_ca(sk); + + bbr->cycle_idx = cycle_idx; + /* New phase, so need to update cwnd and pacing rate. */ + bbr->try_fast_path = 0; +} + +/* Send at estimated bw to fill the pipe, but not queue. We need this phase + * before PROBE_UP, because as soon as we send faster than the available bw + * we will start building a queue, and if the buffer is shallow we can cause + * loss. If we do not fill the pipe before we cause this loss, our bw_hi and + * inflight_hi estimates will underestimate. + */ +static void bbr_start_bw_probe_refill(struct sock *sk, u32 bw_probe_up_rounds) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + bbr_reset_lower_bounds(sk); + bbr->bw_probe_up_rounds = bw_probe_up_rounds; + bbr->bw_probe_up_acks = 0; + bbr->stopped_risky_probe = 0; + bbr->ack_phase = BBR_ACKS_REFILLING; + bbr->next_rtt_delivered = tp->delivered; + bbr_set_cycle_idx(sk, BBR_BW_PROBE_REFILL); +} + +/* Now probe max deliverable data rate and volume. */ +static void bbr_start_bw_probe_up(struct sock *sk, struct bbr_context *ctx) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + bbr->ack_phase = BBR_ACKS_PROBE_STARTING; + bbr->next_rtt_delivered = tp->delivered; + bbr->cycle_mstamp = tp->tcp_mstamp; + bbr_reset_full_bw(sk); + bbr->full_bw = ctx->sample_bw; + bbr_set_cycle_idx(sk, BBR_BW_PROBE_UP); + bbr_raise_inflight_hi_slope(sk); +} + +/* Start a new PROBE_BW probing cycle of some wall clock length. Pick a wall + * clock time at which to probe beyond an inflight that we think to be + * safe. This will knowingly risk packet loss, so we want to do this rarely, to + * keep packet loss rates low. Also start a round-trip counter, to probe faster + * if we estimate a Reno flow at our BDP would probe faster. + */ +static void bbr_start_bw_probe_down(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + bbr_reset_congestion_signals(sk); + bbr->bw_probe_up_cnt = ~0U; /* not growing inflight_hi any more */ + bbr_pick_probe_wait(sk); + bbr->cycle_mstamp = tp->tcp_mstamp; /* start wall clock */ + bbr->ack_phase = BBR_ACKS_PROBE_STOPPING; + bbr->next_rtt_delivered = tp->delivered; + bbr_set_cycle_idx(sk, BBR_BW_PROBE_DOWN); +} + +/* Cruise: maintain what we estimate to be a neutral, conservative + * operating point, without attempting to probe up for bandwidth or down for + * RTT, and only reducing inflight in response to loss/ECN signals. + */ +static void bbr_start_bw_probe_cruise(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + if (bbr->inflight_lo != ~0U) + bbr->inflight_lo = min(bbr->inflight_lo, bbr->inflight_hi); + + bbr_set_cycle_idx(sk, BBR_BW_PROBE_CRUISE); +} + +/* Loss and/or ECN rate is too high while probing. + * Adapt (once per bw probe) by cutting inflight_hi and then restarting cycle. + */ +static void bbr_handle_inflight_too_high(struct sock *sk, + const struct rate_sample *rs) +{ + struct bbr *bbr = inet_csk_ca(sk); + const u32 beta = bbr_param(sk, beta); + + bbr->prev_probe_too_high = 1; + bbr->bw_probe_samples = 0; /* only react once per probe */ + /* If we are app-limited then we are not robustly + * probing the max volume of inflight data we think + * might be safe (analogous to how app-limited bw + * samples are not known to be robustly probing bw). + */ + if (!rs->is_app_limited) { + bbr->inflight_hi = max_t(u32, rs->tx_in_flight, + (u64)bbr_target_inflight(sk) * + (BBR_UNIT - beta) >> BBR_SCALE); + } + if (bbr->mode == BBR_PROBE_BW && bbr->cycle_idx == BBR_BW_PROBE_UP) + bbr_start_bw_probe_down(sk); +} + +/* If we're seeing bw and loss samples reflecting our bw probing, adapt + * using the signals we see. If loss or ECN mark rate gets too high, then adapt + * inflight_hi downward. If we're able to push inflight higher without such + * signals, push higher: adapt inflight_hi upward. + */ +static bool bbr_adapt_upper_bounds(struct sock *sk, + const struct rate_sample *rs, + struct bbr_context *ctx) +{ + struct bbr *bbr = inet_csk_ca(sk); + + /* Track when we'll see bw/loss samples resulting from our bw probes. */ + if (bbr->ack_phase == BBR_ACKS_PROBE_STARTING && bbr->round_start) + bbr->ack_phase = BBR_ACKS_PROBE_FEEDBACK; + if (bbr->ack_phase == BBR_ACKS_PROBE_STOPPING && bbr->round_start) { + /* End of samples from bw probing phase. */ + bbr->bw_probe_samples = 0; + bbr->ack_phase = BBR_ACKS_INIT; + /* At this point in the cycle, our current bw sample is also + * our best recent chance at finding the highest available bw + * for this flow. So now is the best time to forget the bw + * samples from the previous cycle, by advancing the window. + */ + if (bbr->mode == BBR_PROBE_BW && !rs->is_app_limited) + bbr_advance_max_bw_filter(sk); + /* If we had an inflight_hi, then probed and pushed inflight all + * the way up to hit that inflight_hi without seeing any + * high loss/ECN in all the resulting ACKs from that probing, + * then probe up again, this time letting inflight persist at + * inflight_hi for a round trip, then accelerating beyond. + */ + if (bbr->mode == BBR_PROBE_BW && + bbr->stopped_risky_probe && !bbr->prev_probe_too_high) { + bbr_start_bw_probe_refill(sk, 0); + return true; /* yes, decided state transition */ + } + } + if (bbr_is_inflight_too_high(sk, rs)) { + if (bbr->bw_probe_samples) /* sample is from bw probing? */ + bbr_handle_inflight_too_high(sk, rs); + } else { + /* Loss/ECN rate is declared safe. Adjust upper bound upward. */ + + if (bbr->inflight_hi == ~0U) + return false; /* no excess queue signals yet */ + + /* To be resilient to random loss, we must raise bw/inflight_hi + * if we observe in any phase that a higher level is safe. + */ + if (rs->tx_in_flight > bbr->inflight_hi) { + bbr->inflight_hi = rs->tx_in_flight; + } + + if (bbr->mode == BBR_PROBE_BW && + bbr->cycle_idx == BBR_BW_PROBE_UP) + bbr_probe_inflight_hi_upward(sk, rs); + } + + return false; +} + +/* Check if it's time to probe for bandwidth now, and if so, kick it off. */ +static bool bbr_check_time_to_probe_bw(struct sock *sk, + const struct rate_sample *rs) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 n; + + /* If we seem to be at an operating point where we are not seeing loss + * but we are seeing ECN marks, then when the ECN marks cease we reprobe + * quickly (in case cross-traffic has ceased and freed up bw). + */ + if (bbr_param(sk, ecn_reprobe_gain) && bbr->ecn_eligible && + bbr->ecn_in_cycle && !bbr->loss_in_cycle && + inet_csk(sk)->icsk_ca_state == TCP_CA_Open) { + /* Calculate n so that when bbr_raise_inflight_hi_slope() + * computes growth_this_round as 2^n it will be roughly the + * desired volume of data (inflight_hi*ecn_reprobe_gain). + */ + n = ilog2((((u64)bbr->inflight_hi * + bbr_param(sk, ecn_reprobe_gain)) >> BBR_SCALE)); + bbr_start_bw_probe_refill(sk, n); + return true; + } + + if (bbr_has_elapsed_in_phase(sk, bbr->probe_wait_us) || + bbr_is_reno_coexistence_probe_time(sk)) { + bbr_start_bw_probe_refill(sk, 0); + return true; + } + return false; +} + +/* Is it time to transition from PROBE_DOWN to PROBE_CRUISE? */ +static bool bbr_check_time_to_cruise(struct sock *sk, u32 inflight, u32 bw) +{ + /* Always need to pull inflight down to leave headroom in queue. */ + if (inflight > bbr_inflight_with_headroom(sk)) + return false; + + return inflight <= bbr_inflight(sk, bw, BBR_UNIT); +} + +/* PROBE_BW state machine: cruise, refill, probe for bw, or drain? */ +static void bbr_update_cycle_phase(struct sock *sk, + const struct rate_sample *rs, + struct bbr_context *ctx) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + bool is_bw_probe_done = false; + u32 inflight, bw; + + if (!bbr_full_bw_reached(sk)) + return; + + /* In DRAIN, PROBE_BW, or PROBE_RTT, adjust upper bounds. */ + if (bbr_adapt_upper_bounds(sk, rs, ctx)) + return; /* already decided state transition */ + + if (bbr->mode != BBR_PROBE_BW) + return; + + inflight = bbr_packets_in_net_at_edt(sk, rs->prior_in_flight); + bw = bbr_max_bw(sk); + + switch (bbr->cycle_idx) { + /* First we spend most of our time cruising with a pacing_gain of 1.0, + * which paces at the estimated bw, to try to fully use the pipe + * without building queue. If we encounter loss/ECN marks, we adapt + * by slowing down. + */ + case BBR_BW_PROBE_CRUISE: + if (bbr_check_time_to_probe_bw(sk, rs)) + return; /* already decided state transition */ + break; + + /* After cruising, when it's time to probe, we first "refill": we send + * at the estimated bw to fill the pipe, before probing higher and + * knowingly risking overflowing the bottleneck buffer (causing loss). + */ + case BBR_BW_PROBE_REFILL: + if (bbr->round_start) { + /* After one full round trip of sending in REFILL, we + * start to see bw samples reflecting our REFILL, which + * may be putting too much data in flight. + */ + bbr->bw_probe_samples = 1; + bbr_start_bw_probe_up(sk, ctx); + } + break; + + /* After we refill the pipe, we probe by using a pacing_gain > 1.0, to + * probe for bw. If we have not seen loss/ECN, we try to raise inflight + * to at least pacing_gain*BDP; note that this may take more than + * min_rtt if min_rtt is small (e.g. on a LAN). + * + * We terminate PROBE_UP bandwidth probing upon any of the following: + * + * (1) We've pushed inflight up to hit the inflight_hi target set in the + * most recent previous bw probe phase. Thus we want to start + * draining the queue immediately because it's very likely the most + * recently sent packets will fill the queue and cause drops. + * (2) If inflight_hi has not limited bandwidth growth recently, and + * yet delivered bandwidth has not increased much recently + * (bbr->full_bw_now). + * (3) Loss filter says loss rate is "too high". + * (4) ECN filter says ECN mark rate is "too high". + * + * (1) (2) checked here, (3) (4) checked in bbr_is_inflight_too_high() + */ + case BBR_BW_PROBE_UP: + if (bbr->prev_probe_too_high && + inflight >= bbr->inflight_hi) { + bbr->stopped_risky_probe = 1; + is_bw_probe_done = true; + } else { + if (tp->is_cwnd_limited && + tcp_snd_cwnd(tp) >= bbr->inflight_hi) { + /* inflight_hi is limiting bw growth */ + bbr_reset_full_bw(sk); + bbr->full_bw = ctx->sample_bw; + } else if (bbr->full_bw_now) { + /* Plateau in estimated bw. Pipe looks full. */ + is_bw_probe_done = true; + } + } + if (is_bw_probe_done) { + bbr->prev_probe_too_high = 0; /* no loss/ECN (yet) */ + bbr_start_bw_probe_down(sk); /* restart w/ down */ + } + break; + + /* After probing in PROBE_UP, we have usually accumulated some data in + * the bottleneck buffer (if bw probing didn't find more bw). We next + * enter PROBE_DOWN to try to drain any excess data from the queue. To + * do this, we use a pacing_gain < 1.0. We hold this pacing gain until + * our inflight is less then that target cruising point, which is the + * minimum of (a) the amount needed to leave headroom, and (b) the + * estimated BDP. Once inflight falls to match the target, we estimate + * the queue is drained; persisting would underutilize the pipe. + */ + case BBR_BW_PROBE_DOWN: + if (bbr_check_time_to_probe_bw(sk, rs)) + return; /* already decided state transition */ + if (bbr_check_time_to_cruise(sk, inflight, bw)) + bbr_start_bw_probe_cruise(sk); + break; + + default: + WARN_ONCE(1, "BBR invalid cycle index %u\n", bbr->cycle_idx); + } +} + +/* Exiting PROBE_RTT, so return to bandwidth probing in STARTUP or PROBE_BW. */ +static void bbr_exit_probe_rtt(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + bbr_reset_lower_bounds(sk); + if (bbr_full_bw_reached(sk)) { + bbr->mode = BBR_PROBE_BW; + /* Raising inflight after PROBE_RTT may cause loss, so reset + * the PROBE_BW clock and schedule the next bandwidth probe for + * a friendly and randomized future point in time. + */ + bbr_start_bw_probe_down(sk); + /* Since we are exiting PROBE_RTT, we know inflight is + * below our estimated BDP, so it is reasonable to cruise. + */ + bbr_start_bw_probe_cruise(sk); + } else { + bbr->mode = BBR_STARTUP; + } +} + +/* Exit STARTUP based on loss rate > 1% and loss gaps in round >= N. Wait until + * the end of the round in recovery to get a good estimate of how many packets + * have been lost, and how many we need to drain with a low pacing rate. + */ +static void bbr_check_loss_too_high_in_startup(struct sock *sk, + const struct rate_sample *rs) +{ + struct bbr *bbr = inet_csk_ca(sk); + + if (bbr_full_bw_reached(sk)) + return; + + /* For STARTUP exit, check the loss rate at the end of each round trip + * of Recovery episodes in STARTUP. We check the loss rate at the end + * of the round trip to filter out noisy/low loss and have a better + * sense of inflight (extent of loss), so we can drain more accurately. + */ + if (rs->losses && bbr->loss_events_in_round < 0xf) + bbr->loss_events_in_round++; /* update saturating counter */ + if (bbr_param(sk, full_loss_cnt) && bbr->loss_round_start && + inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery && + bbr->loss_events_in_round >= bbr_param(sk, full_loss_cnt) && + bbr_is_inflight_too_high(sk, rs)) { + bbr_handle_queue_too_high_in_startup(sk); + return; + } + if (bbr->loss_round_start) + bbr->loss_events_in_round = 0; +} + +/* Estimate when the pipe is full, using the change in delivery rate: BBR + * estimates bw probing 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_context *ctx) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 bw_thresh, full_cnt, thresh; + + if (bbr->full_bw_now || rs->is_app_limited) + return; + + thresh = bbr_param(sk, full_bw_thresh); + full_cnt = bbr_param(sk, full_bw_cnt); + bw_thresh = (u64)bbr->full_bw * thresh >> BBR_SCALE; + if (ctx->sample_bw >= bw_thresh) { + bbr_reset_full_bw(sk); + bbr->full_bw = ctx->sample_bw; + return; + } + if (!bbr->round_start) + return; + ++bbr->full_bw_cnt; + bbr->full_bw_now = bbr->full_bw_cnt >= full_cnt; + bbr->full_bw_reached |= bbr->full_bw_now; +} + +/* 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_context *ctx) +{ + 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 */ + /* Set ssthresh to export purely for monitoring, to signal + * completion of initial STARTUP by setting to a non- + * TCP_INFINITE_SSTHRESH value (ssthresh is not used by BBR). + */ + tcp_sk(sk)->snd_ssthresh = + bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT); + bbr_reset_congestion_signals(sk); + } /* fall through to check if in-flight is already small: */ + if (bbr->mode == BBR_DRAIN && + bbr_packets_in_net_at_edt(sk, tcp_packets_in_flight(tcp_sk(sk))) <= + bbr_inflight(sk, bbr_max_bw(sk), BBR_UNIT)) { + bbr->mode = BBR_PROBE_BW; + bbr_start_bw_probe_down(sk); + } +} + +static void bbr_update_model(struct sock *sk, const struct rate_sample *rs, + struct bbr_context *ctx) +{ + bbr_update_congestion_signals(sk, rs, ctx); + bbr_update_ack_aggregation(sk, rs); + bbr_check_loss_too_high_in_startup(sk, rs); + bbr_check_full_bw_reached(sk, rs, ctx); + bbr_check_drain(sk, rs, ctx); + bbr_update_cycle_phase(sk, rs, ctx); + bbr_update_min_rtt(sk, rs); +} + +/* Fast path for app-limited case. + * + * On each ack, we execute bbr state machine, which primarily consists of: + * 1) update model based on new rate sample, and + * 2) update control based on updated model or state change. + * + * There are certain workload/scenarios, e.g. app-limited case, where + * either we can skip updating model or we can skip update of both model + * as well as control. This provides signifcant softirq cpu savings for + * processing incoming acks. + * + * In case of app-limited, if there is no congestion (loss/ecn) and + * if observed bw sample is less than current estimated bw, then we can + * skip some of the computation in bbr state processing: + * + * - if there is no rtt/mode/phase change: In this case, since all the + * parameters of the network model are constant, we can skip model + * as well control update. + * + * - else we can skip rest of the model update. But we still need to + * update the control to account for the new rtt/mode/phase. + * + * Returns whether we can take fast path or not. + */ +static bool bbr_run_fast_path(struct sock *sk, bool *update_model, + const struct rate_sample *rs, struct bbr_context *ctx) +{ + struct bbr *bbr = inet_csk_ca(sk); + u32 prev_min_rtt_us, prev_mode; + + if (bbr_param(sk, fast_path) && bbr->try_fast_path && + rs->is_app_limited && ctx->sample_bw < bbr_max_bw(sk) && + !bbr->loss_in_round && !bbr->ecn_in_round ) { + prev_mode = bbr->mode; + prev_min_rtt_us = bbr->min_rtt_us; + bbr_check_drain(sk, rs, ctx); + bbr_update_cycle_phase(sk, rs, ctx); + bbr_update_min_rtt(sk, rs); + + if (bbr->mode == prev_mode && + bbr->min_rtt_us == prev_min_rtt_us && + bbr->try_fast_path) { + return true; + } + + /* Skip model update, but control still needs to be updated */ + *update_model = false; + } + return false; +} + +void bbr_main(struct sock *sk, const struct rate_sample *rs) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + struct bbr_context ctx = { 0 }; + bool update_model = true; + u32 bw, round_delivered; + int ce_ratio = -1; + + round_delivered = bbr_update_round_start(sk, rs, &ctx); + if (bbr->round_start) { + bbr->rounds_since_probe = + min_t(s32, bbr->rounds_since_probe + 1, 0xFF); + ce_ratio = bbr_update_ecn_alpha(sk); + } + bbr_plb(sk, rs, ce_ratio); + + bbr->ecn_in_round |= (bbr->ecn_eligible && rs->is_ece); + bbr_calculate_bw_sample(sk, rs, &ctx); + bbr_update_latest_delivery_signals(sk, rs, &ctx); + + if (bbr_run_fast_path(sk, &update_model, rs, &ctx)) + goto out; + + if (update_model) + bbr_update_model(sk, rs, &ctx); + + bbr_update_gains(sk); + bw = bbr_bw(sk); + bbr_set_pacing_rate(sk, bw, bbr->pacing_gain); + bbr_set_cwnd(sk, rs, rs->acked_sacked, bw, bbr->cwnd_gain, + tcp_snd_cwnd(tp), &ctx); + bbr_bound_cwnd_for_inflight_model(sk); + +out: + bbr_advance_latest_delivery_signals(sk, rs, &ctx); + bbr->prev_ca_state = inet_csk(sk)->icsk_ca_state; + bbr->loss_in_cycle |= rs->lost > 0; + bbr->ecn_in_cycle |= rs->delivered_ce > 0; +} + +static void bbr_init(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + bbr->initialized = 1; + + bbr->init_cwnd = min(0x7FU, tcp_snd_cwnd(tp)); + bbr->prior_cwnd = tp->prior_cwnd; + tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; + bbr->next_rtt_delivered = tp->delivered; + bbr->prev_ca_state = TCP_CA_Open; + + bbr->probe_rtt_done_stamp = 0; + bbr->probe_rtt_round_done = 0; + bbr->probe_rtt_min_us = tcp_min_rtt(tp); + bbr->probe_rtt_min_stamp = tcp_jiffies32; + bbr->min_rtt_us = tcp_min_rtt(tp); + bbr->min_rtt_stamp = tcp_jiffies32; + + bbr->has_seen_rtt = 0; + bbr_init_pacing_rate_from_rtt(sk); + + bbr->round_start = 0; + bbr->idle_restart = 0; + bbr->full_bw_reached = 0; + bbr->full_bw = 0; + bbr->full_bw_cnt = 0; + bbr->cycle_mstamp = 0; + bbr->cycle_idx = 0; + + bbr_reset_startup_mode(sk); + + bbr->ack_epoch_mstamp = tp->tcp_mstamp; + bbr->ack_epoch_acked = 0; + bbr->extra_acked_win_rtts = 0; + bbr->extra_acked_win_idx = 0; + bbr->extra_acked[0] = 0; + bbr->extra_acked[1] = 0; + + bbr->ce_state = 0; + bbr->prior_rcv_nxt = tp->rcv_nxt; + bbr->try_fast_path = 0; + + cmpxchg(&sk->sk_pacing_status, SK_PACING_NONE, SK_PACING_NEEDED); + + /* Start sampling ECN mark rate after first full flight is ACKed: */ + bbr->loss_round_delivered = tp->delivered + 1; + bbr->loss_round_start = 0; + bbr->undo_bw_lo = 0; + bbr->undo_inflight_lo = 0; + bbr->undo_inflight_hi = 0; + bbr->loss_events_in_round = 0; + bbr->startup_ecn_rounds = 0; + bbr_reset_congestion_signals(sk); + bbr->bw_lo = ~0U; + bbr->bw_hi[0] = 0; + bbr->bw_hi[1] = 0; + bbr->inflight_lo = ~0U; + bbr->inflight_hi = ~0U; + bbr_reset_full_bw(sk); + bbr->bw_probe_up_cnt = ~0U; + bbr->bw_probe_up_acks = 0; + bbr->bw_probe_up_rounds = 0; + bbr->probe_wait_us = 0; + bbr->stopped_risky_probe = 0; + bbr->ack_phase = BBR_ACKS_INIT; + bbr->rounds_since_probe = 0; + bbr->bw_probe_samples = 0; + bbr->prev_probe_too_high = 0; + bbr->ecn_eligible = 0; + bbr->ecn_alpha = bbr_param(sk, ecn_alpha_init); + bbr->alpha_last_delivered = 0; + bbr->alpha_last_delivered_ce = 0; + bbr->plb.pause_until = 0; + + tp->fast_ack_mode = bbr_fast_ack_mode ? 1 : 0; +} + +/* BBR marks the current round trip as a loss round. */ +static void bbr_note_loss(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + + /* Capture "current" data over the full round trip of loss, to + * have a better chance of observing the full capacity of the path. + */ + if (!bbr->loss_in_round) /* first loss in this round trip? */ + bbr->loss_round_delivered = tp->delivered; /* set round trip */ + bbr->loss_in_round = 1; + bbr->loss_in_cycle = 1; +} + +/* Core TCP stack informs us that the given skb was just marked lost. */ +static void bbr_skb_marked_lost(struct sock *sk, + const struct sk_buff *skb) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + struct tcp_skb_cb *scb = TCP_SKB_CB(skb); + struct rate_sample rs = {}; + + bbr_note_loss(sk); + + if (!bbr->bw_probe_samples) + return; /* not an skb sent while probing for bandwidth */ + if (unlikely(!scb->tx.delivered_mstamp)) + return; /* skb was SACKed, reneged, marked lost; ignore it */ + /* We are probing for bandwidth. Construct a rate sample that + * estimates what happened in the flight leading up to this lost skb, + * then see if the loss rate went too high, and if so at which packet. + */ + rs.tx_in_flight = scb->tx.in_flight; + rs.lost = tp->lost - scb->tx.lost; + rs.is_app_limited = scb->tx.is_app_limited; + if (bbr_is_inflight_too_high(sk, &rs)) { + rs.tx_in_flight = bbr_inflight_hi_from_lost_skb(sk, &rs, skb); + bbr_handle_inflight_too_high(sk, &rs); + } +} + +static void bbr_run_loss_probe_recovery(struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct bbr *bbr = inet_csk_ca(sk); + struct rate_sample rs = {0}; + + bbr_note_loss(sk); + + if (!bbr->bw_probe_samples) + return; /* not sent while probing for bandwidth */ + /* We are probing for bandwidth. Construct a rate sample that + * estimates what happened in the flight leading up to this + * loss, then see if the loss rate went too high. + */ + rs.lost = 1; /* TLP probe repaired loss of a single segment */ + rs.tx_in_flight = bbr->inflight_latest + rs.lost; + rs.is_app_limited = tp->tlp_orig_data_app_limited; + if (bbr_is_inflight_too_high(sk, &rs)) + bbr_handle_inflight_too_high(sk, &rs); +} + +/* Revert short-term model if current loss recovery event was spurious. */ +static u32 bbr_undo_cwnd(struct sock *sk) +{ + struct bbr *bbr = inet_csk_ca(sk); + + bbr_reset_full_bw(sk); /* spurious slow-down; reset full bw detector */ + bbr->loss_in_round = 0; + + /* Revert to cwnd and other state saved before loss episode. */ + bbr->bw_lo = max(bbr->bw_lo, bbr->undo_bw_lo); + bbr->inflight_lo = max(bbr->inflight_lo, bbr->undo_inflight_lo); + bbr->inflight_hi = max(bbr->inflight_hi, bbr->undo_inflight_hi); + bbr->try_fast_path = 0; /* take slow path to set proper cwnd, pacing */ + return bbr->prior_cwnd; +} + +/* Entering loss recovery, so save state for when we undo recovery. */ static u32 bbr_ssthresh(struct sock *sk) { + struct bbr *bbr = inet_csk_ca(sk); + bbr_save_cwnd(sk); + /* For undo, save state that adapts based on loss signal. */ + bbr->undo_bw_lo = bbr->bw_lo; + bbr->undo_inflight_lo = bbr->inflight_lo; + bbr->undo_inflight_hi = bbr->inflight_hi; return tcp_sk(sk)->snd_ssthresh; } +static enum tcp_bbr_phase bbr_get_phase(struct bbr *bbr) +{ + switch (bbr->mode) { + case BBR_STARTUP: + return BBR_PHASE_STARTUP; + case BBR_DRAIN: + return BBR_PHASE_DRAIN; + case BBR_PROBE_BW: + break; + case BBR_PROBE_RTT: + return BBR_PHASE_PROBE_RTT; + default: + return BBR_PHASE_INVALID; + } + switch (bbr->cycle_idx) { + case BBR_BW_PROBE_UP: + return BBR_PHASE_PROBE_BW_UP; + case BBR_BW_PROBE_DOWN: + return BBR_PHASE_PROBE_BW_DOWN; + case BBR_BW_PROBE_CRUISE: + return BBR_PHASE_PROBE_BW_CRUISE; + case BBR_BW_PROBE_REFILL: + return BBR_PHASE_PROBE_BW_REFILL; + default: + return BBR_PHASE_INVALID; + } +} + static size_t bbr_get_info(struct sock *sk, u32 ext, int *attr, - union tcp_cc_info *info) + 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; + u64 bw = bbr_bw_bytes_per_sec(sk, bbr_bw(sk)); + u64 bw_hi = bbr_bw_bytes_per_sec(sk, bbr_max_bw(sk)); + u64 bw_lo = bbr->bw_lo == ~0U ? + ~0ULL : bbr_bw_bytes_per_sec(sk, bbr->bw_lo); + struct tcp_bbr_info *bbr_info = &info->bbr; + + memset(bbr_info, 0, sizeof(*bbr_info)); + bbr_info->bbr_bw_lo = (u32)bw; + bbr_info->bbr_bw_hi = (u32)(bw >> 32); + bbr_info->bbr_min_rtt = bbr->min_rtt_us; + bbr_info->bbr_pacing_gain = bbr->pacing_gain; + bbr_info->bbr_cwnd_gain = bbr->cwnd_gain; + bbr_info->bbr_bw_hi_lsb = (u32)bw_hi; + bbr_info->bbr_bw_hi_msb = (u32)(bw_hi >> 32); + bbr_info->bbr_bw_lo_lsb = (u32)bw_lo; + bbr_info->bbr_bw_lo_msb = (u32)(bw_lo >> 32); + bbr_info->bbr_mode = bbr->mode; + bbr_info->bbr_phase = (__u8)bbr_get_phase(bbr); + bbr_info->bbr_version = (__u8)BBR_VERSION; + bbr_info->bbr_inflight_lo = bbr->inflight_lo; + bbr_info->bbr_inflight_hi = bbr->inflight_hi; + bbr_info->bbr_extra_acked = bbr_extra_acked(sk); *attr = INET_DIAG_BBRINFO; - return sizeof(info->bbr); + return sizeof(*bbr_info); } return 0; } static void bbr_set_state(struct sock *sk, u8 new_state) { + struct tcp_sock *tp = tcp_sk(sk); 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); + tcp_plb_update_state_upon_rto(sk, &bbr->plb); + /* The tcp_write_timeout() call to sk_rethink_txhash() likely + * repathed this flow, so re-learn the min network RTT on the + * new path: + */ + bbr_reset_full_bw(sk); + if (!bbr_is_probing_bandwidth(sk) && bbr->inflight_lo == ~0U) { + /* bbr_adapt_lower_bounds() needs cwnd before + * we suffered an RTO, to update inflight_lo: + */ + bbr->inflight_lo = + max(tcp_snd_cwnd(tp), bbr->prior_cwnd); + } + } else if (bbr->prev_ca_state == TCP_CA_Loss && + new_state != TCP_CA_Loss) { + bbr_exit_loss_recovery(sk); } } + static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { - .flags = TCP_CONG_NON_RESTRICTED, + .flags = TCP_CONG_NON_RESTRICTED | TCP_CONG_WANTS_CE_EVENTS, .name = "bbr", .owner = THIS_MODULE, .init = bbr_init, .cong_control = bbr_main, .sndbuf_expand = bbr_sndbuf_expand, + .skb_marked_lost = bbr_skb_marked_lost, .undo_cwnd = bbr_undo_cwnd, .cwnd_event = bbr_cwnd_event, .ssthresh = bbr_ssthresh, @@ -1175,10 +2358,11 @@ BTF_SET8_START(tcp_bbr_check_kfunc_ids) BTF_ID_FLAGS(func, bbr_init) BTF_ID_FLAGS(func, bbr_main) BTF_ID_FLAGS(func, bbr_sndbuf_expand) +BTF_ID_FLAGS(func, bbr_skb_marked_lost) BTF_ID_FLAGS(func, bbr_undo_cwnd) BTF_ID_FLAGS(func, bbr_cwnd_event) BTF_ID_FLAGS(func, bbr_ssthresh) -BTF_ID_FLAGS(func, bbr_min_tso_segs) +BTF_ID_FLAGS(func, bbr_tso_segs) BTF_ID_FLAGS(func, bbr_set_state) #endif #endif @@ -1213,5 +2397,12 @@ MODULE_AUTHOR("Van Jacobson "); MODULE_AUTHOR("Neal Cardwell "); MODULE_AUTHOR("Yuchung Cheng "); MODULE_AUTHOR("Soheil Hassas Yeganeh "); +MODULE_AUTHOR("Priyaranjan Jha "); +MODULE_AUTHOR("Yousuk Seung "); +MODULE_AUTHOR("Kevin Yang "); +MODULE_AUTHOR("Arjun Roy "); +MODULE_AUTHOR("David Morley "); + MODULE_LICENSE("Dual BSD/GPL"); MODULE_DESCRIPTION("TCP BBR (Bottleneck Bandwidth and RTT)"); +MODULE_VERSION(__stringify(BBR_VERSION));