From a388e2d7352dd97d1621c5a84ea9277a651cfd39 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Tue, 11 Jun 2019 12:26:55 -0400 Subject: [PATCH 01/20] net-tcp_bbr: broaden app-limited rate sample detection This commit is a bug fix for the Linux TCP app-limited (application-limited) logic that is used for collecting rate (bandwidth) samples. Previously the app-limited logic only looked for "bubbles" of silence in between application writes, by checking at the start of each sendmsg. But "bubbles" of silence can also happen before retransmits: e.g. bubbles can happen between an application write and a retransmit, or between two retransmits. Retransmits are triggered by ACKs or timers. So this commit checks for bubbles of app-limited silence upon ACKs or timers. Why does this commit check for app-limited state at the start of ACKs and timer handling? Because at that point we know whether inflight was fully using the cwnd. During processing the ACK or timer event we often change the cwnd; after changing the cwnd we can't know whether inflight was fully using the old cwnd. Origin-9xx-SHA1: 3fe9b53291e018407780fb8c356adb5666722cbc Change-Id: I37221506f5166877c2b110753d39bb0757985e68 Signed-off-by: Alexandre Frade --- net/ipv4/tcp_input.c | 1 + net/ipv4/tcp_timer.c | 1 + 2 files changed, 2 insertions(+) diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index e2d3ea2e3..230c3fb9c 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -3818,6 +3818,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; rs.prior_in_flight = tcp_packets_in_flight(tp); + tcp_rate_check_app_limited(sk); /* ts_recent update must be made after we are sure that the packet * is in window. diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c index cf354c29e..7c28eadd6 100644 --- a/net/ipv4/tcp_timer.c +++ b/net/ipv4/tcp_timer.c @@ -617,6 +617,7 @@ void tcp_write_timer_handler(struct sock *sk) return; } + tcp_rate_check_app_limited(sk); tcp_mstamp_refresh(tcp_sk(sk)); event = icsk->icsk_pending; -- 2.39.3 From d91e296f20800dffa4cdfc61048ca709cd40f2e6 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Sun, 24 Jun 2018 21:55:59 -0400 Subject: [PATCH 02/20] net-tcp_bbr: v2: shrink delivered_mstamp, first_tx_mstamp to u32 to free up 8 bytes Free up some space for tracking inflight and losses for each bw sample, in upcoming commits. These timestamps are in microseconds, and are now stored in 32 bits. So they can only hold time intervals up to roughly 2^12 = 4096 seconds. But Linux TCP RTT and RTO tracking has the same 32-bit microsecond implementation approach and resulting deployment limitations. So this is not introducing a new limit. And these should not be a limitation for the foreseeable future. Effort: net-tcp_bbr Origin-9xx-SHA1: 238a7e6b5d51625fef1ce7769826a7b21b02ae55 Change-Id: I3b779603797263b52a61ad57c565eb91fe42680c Signed-off-by: Alexandre Frade --- include/net/tcp.h | 9 +++++++-- net/ipv4/tcp_rate.c | 7 ++++--- 2 files changed, 11 insertions(+), 5 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index e9c8f88f4..490c11f17 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -824,6 +824,11 @@ static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0) return max_t(s64, t1 - t0, 0); } +static inline u32 tcp_stamp32_us_delta(u32 t1, u32 t0) +{ + return max_t(s32, t1 - t0, 0); +} + static inline u32 tcp_skb_timestamp(const struct sk_buff *skb) { return tcp_ns_to_ts(skb->skb_mstamp_ns); @@ -899,9 +904,9 @@ struct tcp_skb_cb { /* pkts S/ACKed so far upon tx of skb, incl retrans: */ __u32 delivered; /* start of send pipeline phase */ - u64 first_tx_mstamp; + u32 first_tx_mstamp; /* when we reached the "delivered" count */ - u64 delivered_mstamp; + u32 delivered_mstamp; } tx; /* only used for outgoing skbs */ union { struct inet_skb_parm h4; diff --git a/net/ipv4/tcp_rate.c b/net/ipv4/tcp_rate.c index a8f6d9d06..763362159 100644 --- a/net/ipv4/tcp_rate.c +++ b/net/ipv4/tcp_rate.c @@ -101,8 +101,9 @@ void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb, /* Record send time of most recently ACKed packet: */ tp->first_tx_mstamp = tx_tstamp; /* Find the duration of the "send phase" of this window: */ - rs->interval_us = tcp_stamp_us_delta(tp->first_tx_mstamp, - scb->tx.first_tx_mstamp); + rs->interval_us = tcp_stamp32_us_delta( + tp->first_tx_mstamp, + scb->tx.first_tx_mstamp); } /* Mark off the skb delivered once it's sacked to avoid being @@ -155,7 +156,7 @@ void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost, * longer phase. */ snd_us = rs->interval_us; /* send phase */ - ack_us = tcp_stamp_us_delta(tp->tcp_mstamp, + ack_us = tcp_stamp32_us_delta(tp->tcp_mstamp, rs->prior_mstamp); /* ack phase */ rs->interval_us = max(snd_us, ack_us); -- 2.39.3 From 730dcf21588602629508e8c1f10776204f5034b1 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Sat, 5 Aug 2017 11:49:50 -0400 Subject: [PATCH 03/20] net-tcp_bbr: v2: snapshot packets in flight at transmit time and pass in rate_sample CC algorithms may want to snapshot the number of packets in flight at transmit time and pass in rate_sample, to understand the relationship between inflight and losses or ECN signals, to try to find the highest inflight value that has acceptable levels of loss/ECN marking. We split out the code to set an skb's tx.in_flight field into its own function, so that this code can be used for the TCP_REPAIR "fake send" code path that inserts skbs into the rtx queue without sending them. Effort: net-tcp_bbr Origin-9xx-SHA1: b3eb4f2d20efab4ca001f32c9294739036c493ea Origin-9xx-SHA1: e880fc907d06ea7354333f60f712748ebce9497b Origin-9xx-SHA1: 330f825a08a6fe92cef74d799cc468864c479f63 Change-Id: I7314047d0ff14dd261a04b1969a46dc658c8836a Signed-off-by: Alexandre Frade --- include/net/tcp.h | 6 ++++++ net/ipv4/tcp_output.c | 1 + net/ipv4/tcp_rate.c | 20 ++++++++++++++++++++ 3 files changed, 27 insertions(+) diff --git a/include/net/tcp.h b/include/net/tcp.h index 490c11f17..10eb4daec 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -907,6 +907,10 @@ struct tcp_skb_cb { u32 first_tx_mstamp; /* when we reached the "delivered" count */ u32 delivered_mstamp; +#define TCPCB_IN_FLIGHT_BITS 20 +#define TCPCB_IN_FLIGHT_MAX ((1U << TCPCB_IN_FLIGHT_BITS) - 1) + u32 in_flight:20, /* packets in flight at transmit */ + unused2:12; } tx; /* only used for outgoing skbs */ union { struct inet_skb_parm h4; @@ -1054,6 +1058,7 @@ struct rate_sample { u64 prior_mstamp; /* starting timestamp for interval */ u32 prior_delivered; /* tp->delivered at "prior_mstamp" */ u32 prior_delivered_ce;/* tp->delivered_ce at "prior_mstamp" */ + u32 tx_in_flight; /* packets in flight at starting timestamp */ s32 delivered; /* number of packets delivered over interval */ s32 delivered_ce; /* number of packets delivered w/ CE marks*/ long interval_us; /* time for tp->delivered to incr "delivered" */ @@ -1173,6 +1178,7 @@ static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event) void tcp_set_ca_state(struct sock *sk, const u8 ca_state); /* From tcp_rate.c */ +void tcp_set_tx_in_flight(struct sock *sk, struct sk_buff *skb); void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb); void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb, struct rate_sample *rs); diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index 26bd039f9..a7a7e9fbc 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -2635,6 +2635,7 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); tcp_init_tso_segs(skb, mss_now); + tcp_set_tx_in_flight(sk, skb); goto repair; /* Skip network transmission */ } diff --git a/net/ipv4/tcp_rate.c b/net/ipv4/tcp_rate.c index 763362159..1424911b1 100644 --- a/net/ipv4/tcp_rate.c +++ b/net/ipv4/tcp_rate.c @@ -34,6 +34,24 @@ * ready to send in the write queue. */ +void tcp_set_tx_in_flight(struct sock *sk, struct sk_buff *skb) +{ + struct tcp_sock *tp = tcp_sk(sk); + u32 in_flight; + + /* Check, sanitize, and record packets in flight after skb was sent. */ + in_flight = tcp_packets_in_flight(tp) + tcp_skb_pcount(skb); + if (WARN_ONCE(in_flight > TCPCB_IN_FLIGHT_MAX, + "insane in_flight %u cc %s mss %u " + "cwnd %u pif %u %u %u %u\n", + in_flight, inet_csk(sk)->icsk_ca_ops->name, + tp->mss_cache, tp->snd_cwnd, + tp->packets_out, tp->retrans_out, + tp->sacked_out, tp->lost_out)) + in_flight = TCPCB_IN_FLIGHT_MAX; + TCP_SKB_CB(skb)->tx.in_flight = in_flight; +} + /* Snapshot the current delivery information in the skb, to generate * a rate sample later when the skb is (s)acked in tcp_rate_skb_delivered(). */ @@ -67,6 +85,7 @@ void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb) TCP_SKB_CB(skb)->tx.delivered = tp->delivered; TCP_SKB_CB(skb)->tx.delivered_ce = tp->delivered_ce; TCP_SKB_CB(skb)->tx.is_app_limited = tp->app_limited ? 1 : 0; + tcp_set_tx_in_flight(sk, skb); } /* When an skb is sacked or acked, we fill in the rate sample with the (prior) @@ -96,6 +115,7 @@ void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb, rs->prior_mstamp = scb->tx.delivered_mstamp; rs->is_app_limited = scb->tx.is_app_limited; rs->is_retrans = scb->sacked & TCPCB_RETRANS; + rs->tx_in_flight = scb->tx.in_flight; rs->last_end_seq = scb->end_seq; /* Record send time of most recently ACKed packet: */ -- 2.39.3 From 29ff3313b86298890816f9db0ab9f75787e326fc Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Thu, 12 Oct 2017 23:44:27 -0400 Subject: [PATCH 04/20] net-tcp_bbr: v2: count packets lost over TCP rate sampling interval For understanding the relationship between inflight and packet loss signals, to try to find the highest inflight value that has acceptable levels of packet losses. Effort: net-tcp_bbr Origin-9xx-SHA1: 4527e26b2bd7756a88b5b9ef1ada3da33dd609ab Change-Id: I594c2500868d9c530770e7ddd68ffc87c57f4fd5 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 5 ++++- net/ipv4/tcp_rate.c | 3 +++ 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 10eb4daec..e17458911 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -911,6 +911,7 @@ struct tcp_skb_cb { #define TCPCB_IN_FLIGHT_MAX ((1U << TCPCB_IN_FLIGHT_BITS) - 1) u32 in_flight:20, /* packets in flight at transmit */ unused2:12; + u32 lost; /* packets lost so far upon tx of skb */ } tx; /* only used for outgoing skbs */ union { struct inet_skb_parm h4; @@ -1056,11 +1057,13 @@ struct ack_sample { */ struct rate_sample { u64 prior_mstamp; /* starting timestamp for interval */ + u32 prior_lost; /* tp->lost at "prior_mstamp" */ u32 prior_delivered; /* tp->delivered at "prior_mstamp" */ u32 prior_delivered_ce;/* tp->delivered_ce at "prior_mstamp" */ u32 tx_in_flight; /* packets in flight at starting timestamp */ + s32 lost; /* number of packets lost over interval */ s32 delivered; /* number of packets delivered over interval */ - s32 delivered_ce; /* number of packets delivered w/ CE marks*/ + s32 delivered_ce; /* packets delivered w/ CE mark over interval */ long interval_us; /* time for tp->delivered to incr "delivered" */ u32 snd_interval_us; /* snd interval for delivered packets */ u32 rcv_interval_us; /* rcv interval for delivered packets */ diff --git a/net/ipv4/tcp_rate.c b/net/ipv4/tcp_rate.c index 1424911b1..8737f2134 100644 --- a/net/ipv4/tcp_rate.c +++ b/net/ipv4/tcp_rate.c @@ -84,6 +84,7 @@ void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb) TCP_SKB_CB(skb)->tx.delivered_mstamp = tp->delivered_mstamp; TCP_SKB_CB(skb)->tx.delivered = tp->delivered; TCP_SKB_CB(skb)->tx.delivered_ce = tp->delivered_ce; + TCP_SKB_CB(skb)->tx.lost = tp->lost; TCP_SKB_CB(skb)->tx.is_app_limited = tp->app_limited ? 1 : 0; tcp_set_tx_in_flight(sk, skb); } @@ -110,6 +111,7 @@ void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb, if (!rs->prior_delivered || tcp_skb_sent_after(tx_tstamp, tp->first_tx_mstamp, scb->end_seq, rs->last_end_seq)) { + rs->prior_lost = scb->tx.lost; rs->prior_delivered_ce = scb->tx.delivered_ce; rs->prior_delivered = scb->tx.delivered; rs->prior_mstamp = scb->tx.delivered_mstamp; @@ -165,6 +167,7 @@ void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost, return; } rs->delivered = tp->delivered - rs->prior_delivered; + rs->lost = tp->lost - rs->prior_lost; rs->delivered_ce = tp->delivered_ce - rs->prior_delivered_ce; /* delivered_ce occupies less than 32 bits in the skb control block */ -- 2.39.3 From 2e5a2ca5b2fceb78eb5b4f3480a01ce0c5e5dd95 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Mon, 19 Nov 2018 13:48:36 -0500 Subject: [PATCH 05/20] net-tcp_bbr: v2: export FLAG_ECE in rate_sample.is_ece For understanding the relationship between inflight and ECN signals, to try to find the highest inflight value that has acceptable levels ECN marking. Effort: net-tcp_bbr Origin-9xx-SHA1: 3eba998f2898541406c2666781182200934965a8 Change-Id: I3a964e04cee83e11649a54507043d2dfe769a3b3 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 1 + net/ipv4/tcp_input.c | 1 + 2 files changed, 2 insertions(+) diff --git a/include/net/tcp.h b/include/net/tcp.h index e17458911..e204c0b9a 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1075,6 +1075,7 @@ struct rate_sample { bool is_app_limited; /* is sample from packet with bubble in pipe? */ bool is_retrans; /* is sample from retransmission? */ bool is_ack_delayed; /* is this (likely) a delayed ACK? */ + bool is_ece; /* did this ACK have ECN marked? */ }; struct tcp_congestion_ops { diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 230c3fb9c..d8a23112d 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -3917,6 +3917,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) delivered = tcp_newly_delivered(sk, delivered, flag); lost = tp->lost - lost; /* freshly marked lost */ rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); + rs.is_ece = !!(flag & FLAG_ECE); tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); tcp_xmit_recovery(sk, rexmit); -- 2.39.3 From 536a8e0bada4910a5a0932ecb2876f0b6263e520 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Tue, 7 Aug 2018 21:52:06 -0400 Subject: [PATCH 06/20] net-tcp_bbr: v2: introduce ca_ops->skb_marked_lost() CC module callback API For connections experiencing reordering, RACK can mark packets lost long after we receive the SACKs/ACKs hinting that the packets were actually lost. This means that CC modules cannot easily learn the volume of inflight data at which packet loss happens by looking at the current inflight or even the packets in flight when the most recently SACKed packet was sent. To learn this, CC modules need to know how many packets were in flight at the time lost packets were sent. This new callback, combined with TCP_SKB_CB(skb)->tx.in_flight, allows them to learn this. This also provides a consistent callback that is invoked whether packets are marked lost upon ACK processing, using the RACK reordering timer, or at RTO time. Effort: net-tcp_bbr Origin-9xx-SHA1: afcbebe3374e4632ac6714d39e4dc8a8455956f4 Change-Id: I54826ab53df636be537e5d3c618a46145d12d51a Signed-off-by: Alexandre Frade --- include/net/tcp.h | 3 +++ net/ipv4/tcp_input.c | 5 +++++ 2 files changed, 8 insertions(+) diff --git a/include/net/tcp.h b/include/net/tcp.h index e204c0b9a..4cb7c13b8 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1102,6 +1102,9 @@ struct tcp_congestion_ops { /* override sysctl_tcp_min_tso_segs */ u32 (*min_tso_segs)(struct sock *sk); + /* react to a specific lost skb (optional) */ + void (*skb_marked_lost)(struct sock *sk, const struct sk_buff *skb); + /* call when packets are delivered to update cwnd and pacing rate, * after all the ca_state processing. (optional) */ diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index d8a23112d..7435114e3 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -1079,7 +1079,12 @@ static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) */ static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) { + struct sock *sk = (struct sock *)tp; + const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; + tp->lost += tcp_skb_pcount(skb); + if (ca_ops->skb_marked_lost) + ca_ops->skb_marked_lost(sk, skb); } void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) -- 2.39.3 From ae99d6192d04c44ae27c7a382d20a589eed8f27f Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Wed, 1 May 2019 20:16:33 -0400 Subject: [PATCH 07/20] net-tcp_bbr: v2: adjust skb tx.in_flight upon merge in tcp_shifted_skb() When tcp_shifted_skb() updates state as adjacent SACKed skbs are coalesced, previously the tx.in_flight was not adjusted, so we could get contradictory state where the skb's recorded pcount was bigger than the tx.in_flight (the number of segments that were in_flight after sending the skb). Normally have a SACKed skb with contradictory pcount/tx.in_flight would not matter. However, with SACK reneging, the SACKed bit is removed, and an skb once again becomes eligible for retransmitting, fragmenting, SACKing, etc. Packetdrill testing verified the following sequence is possible in a kernel that does not have this commit: - skb N is SACKed - skb N+1 is SACKed and combined with skb N using tcp_shifted_skb() - tcp_shifted_skb() will increase the pcount of prev, but leave tx.in_flight as-is - so prev skb can have pcount > tx.in_flight - RTO, tcp_timeout_mark_lost(), detect reneg, remove "SACKed" bit, mark skb N as lost - find pcount of skb N is greater than its tx.in_flight I suspect this issue iw what caused the bbr2_inflight_hi_from_lost_skb(): WARN_ON_ONCE(inflight_prev < 0) to fire in production machines using bbr2. Effort: net-tcp_bbr Origin-9xx-SHA1: 1a3e997e613d2dcf32b947992882854ebe873715 Change-Id: I1b0b75c27519953430c7db51c6f358f104c7af55 Signed-off-by: Alexandre Frade --- net/ipv4/tcp_input.c | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 7435114e3..a32fc4a1a 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -1465,6 +1465,17 @@ static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); tcp_skb_pcount_add(skb, -pcount); + /* Adjust tx.in_flight as pcount is shifted from skb to prev. */ + if (WARN_ONCE(TCP_SKB_CB(skb)->tx.in_flight < pcount, + "prev in_flight: %u skb in_flight: %u pcount: %u", + TCP_SKB_CB(prev)->tx.in_flight, + TCP_SKB_CB(skb)->tx.in_flight, + pcount)) + TCP_SKB_CB(skb)->tx.in_flight = 0; + else + TCP_SKB_CB(skb)->tx.in_flight -= pcount; + TCP_SKB_CB(prev)->tx.in_flight += pcount; + /* When we're adding to gso_segs == 1, gso_size will be zero, * in theory this shouldn't be necessary but as long as DSACK * code can come after this skb later on it's better to keep -- 2.39.3 From 5a0cb01be3b8190023d6cd912a44086f528fa66b Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Wed, 1 May 2019 20:16:25 -0400 Subject: [PATCH 08/20] net-tcp_bbr: v2: adjust skb tx.in_flight upon split in tcp_fragment() When we fragment an skb that has already been sent, we need to update the tx.in_flight for the first skb in the resulting pair ("buff"). Because we were not updating the tx.in_flight, the tx.in_flight value was inconsistent with the pcount of the "buff" skb (tx.in_flight would be too high). That meant that if the "buff" skb was lost, then bbr2_inflight_hi_from_lost_skb() would calculate an inflight_hi value that is too high. This could result in longer queues and higher packet loss. Packetdrill testing verified that without this commit, when the second half of an skb is SACKed and then later the first half of that skb is marked lost, the calculated inflight_hi was incorrect. Effort: net-tcp_bbr Origin-9xx-SHA1: 385f1ddc610798fab2837f9f372857438b25f874 Origin-9xx-SHA1: a0eb099690af net-tcp_bbr: v2: fix tcp_fragment() tx.in_flight recomputation [prod feb 8 2021; use as a fixup] Origin-9xx-SHA1: 885503228153ff0c9114e net-tcp_bbr: v2: introduce tcp_skb_tx_in_flight_is_suspicious() helper for warnings Change-Id: I617f8cab4e9be7a0b8e8d30b047bf8645393354d Signed-off-by: Alexandre Frade --- include/net/tcp.h | 15 +++++++++++++++ net/ipv4/tcp_output.c | 26 +++++++++++++++++++++++++- 2 files changed, 40 insertions(+), 1 deletion(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 4cb7c13b8..04813bf38 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1198,6 +1198,21 @@ static inline bool tcp_skb_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2) return t1 > t2 || (t1 == t2 && after(seq1, seq2)); } +/* If a retransmit failed due to local qdisc congestion or other local issues, + * then we may have called tcp_set_skb_tso_segs() to increase the number of + * segments in the skb without increasing the tx.in_flight. In all other cases, + * the tx.in_flight should be at least as big as the pcount of the sk_buff. We + * do not have the state to know whether a retransmit failed due to local qdisc + * congestion or other local issues, so to avoid spurious warnings we consider + * that any skb marked lost may have suffered that fate. + */ +static inline bool tcp_skb_tx_in_flight_is_suspicious(u32 skb_pcount, + u32 skb_sacked_flags, + u32 tx_in_flight) +{ + return (skb_pcount > tx_in_flight) && !(skb_sacked_flags & TCPCB_LOST); +} + /* These functions determine how the current flow behaves in respect of SACK * handling. SACK is negotiated with the peer, and therefore it can vary * between different flows. diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index a7a7e9fbc..0818709ec 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -1533,7 +1533,7 @@ int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, { struct tcp_sock *tp = tcp_sk(sk); struct sk_buff *buff; - int nsize, old_factor; + int nsize, old_factor, inflight_prev; long limit; int nlen; u8 flags; @@ -1610,6 +1610,30 @@ int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, if (diff) tcp_adjust_pcount(sk, skb, diff); + + inflight_prev = TCP_SKB_CB(skb)->tx.in_flight - old_factor; + if (inflight_prev < 0) { + WARN_ONCE(tcp_skb_tx_in_flight_is_suspicious( + old_factor, + TCP_SKB_CB(skb)->sacked, + TCP_SKB_CB(skb)->tx.in_flight), + "inconsistent: tx.in_flight: %u " + "old_factor: %d mss: %u sacked: %u " + "1st pcount: %d 2nd pcount: %d " + "1st len: %u 2nd len: %u ", + TCP_SKB_CB(skb)->tx.in_flight, old_factor, + mss_now, TCP_SKB_CB(skb)->sacked, + tcp_skb_pcount(skb), tcp_skb_pcount(buff), + skb->len, buff->len); + inflight_prev = 0; + } + /* Set 1st tx.in_flight as if 1st were sent by itself: */ + TCP_SKB_CB(skb)->tx.in_flight = inflight_prev + + tcp_skb_pcount(skb); + /* Set 2nd tx.in_flight with new 1st and 2nd pcounts: */ + TCP_SKB_CB(buff)->tx.in_flight = inflight_prev + + tcp_skb_pcount(skb) + + tcp_skb_pcount(buff); } /* Link BUFF into the send queue. */ -- 2.39.3 From ed558c7da6dcde924cac549192603d5ced6311b4 Mon Sep 17 00:00:00 2001 From: Yousuk Seung Date: Wed, 23 May 2018 17:55:54 -0700 Subject: [PATCH 09/20] net-tcp: add new ca opts flag TCP_CONG_WANTS_CE_EVENTS Add a a new ca opts flag TCP_CONG_WANTS_CE_EVENTS that allows a congestion control module to receive CE events. Currently congestion control modules have to set the TCP_CONG_NEEDS_ECN bit in opts flag to receive CE events but this may incur changes in ECN behavior elsewhere. This patch adds a new bit TCP_CONG_WANTS_CE_EVENTS that allows congestion control modules to receive CE events independently of TCP_CONG_NEEDS_ECN. Effort: net-tcp Origin-9xx-SHA1: 9f7e14716cde760bc6c67ef8ef7e1ee48501d95b Change-Id: I2255506985242f376d910c6fd37daabaf4744f24 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 14 +++++++++++++- net/ipv4/tcp_input.c | 4 ++-- 2 files changed, 15 insertions(+), 3 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 04813bf38..8357f0e1d 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1037,7 +1037,11 @@ enum tcp_ca_ack_event_flags { #define TCP_CONG_NON_RESTRICTED 0x1 /* Requires ECN/ECT set on all packets */ #define TCP_CONG_NEEDS_ECN 0x2 -#define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | TCP_CONG_NEEDS_ECN) +/* Wants notification of CE events (CA_EVENT_ECN_IS_CE, CA_EVENT_ECN_NO_CE). */ +#define TCP_CONG_WANTS_CE_EVENTS 0x4 +#define TCP_CONG_MASK (TCP_CONG_NON_RESTRICTED | \ + TCP_CONG_NEEDS_ECN | \ + TCP_CONG_WANTS_CE_EVENTS) union tcp_cc_info; @@ -1166,6 +1170,14 @@ static inline char *tcp_ca_get_name_by_key(u32 key, char *buffer) } #endif +static inline bool tcp_ca_wants_ce_events(const struct sock *sk) +{ + const struct inet_connection_sock *icsk = inet_csk(sk); + + return icsk->icsk_ca_ops->flags & (TCP_CONG_NEEDS_ECN | + TCP_CONG_WANTS_CE_EVENTS); +} + static inline bool tcp_ca_needs_ecn(const struct sock *sk) { const struct inet_connection_sock *icsk = inet_csk(sk); diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index a32fc4a1a..a913e55d9 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -349,7 +349,7 @@ static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) tcp_enter_quickack_mode(sk, 2); break; case INET_ECN_CE: - if (tcp_ca_needs_ecn(sk)) + if (tcp_ca_wants_ce_events(sk)) tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { @@ -360,7 +360,7 @@ static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) tp->ecn_flags |= TCP_ECN_SEEN; break; default: - if (tcp_ca_needs_ecn(sk)) + if (tcp_ca_wants_ce_events(sk)) tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); tp->ecn_flags |= TCP_ECN_SEEN; break; -- 2.39.3 From ae0bf694fd7ffa62094ca72ee0cd49d474be1954 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Fri, 27 Sep 2019 17:10:26 -0400 Subject: [PATCH 10/20] net-tcp: re-generalize TSO sizing in TCP CC module API Reorganize the API for CC modules so that the CC module once again gets complete control of the TSO sizing decision. This is how the API was set up around 2016 and the initial BBRv1 upstreaming. Later Eric Dumazet simplified it. But with wider testing it now seems that to avoid CPU regressions BBR needs to have a different TSO sizing function. This is necessary to handle cases where there are many flows bottlenecked on the sender host's NIC, in which case BBR's pacing rate is much lower than CUBIC/Reno/DCTCP's. Why does this happen? Because BBR's pacing rate adapts to the low bandwidth share each flow sees. By contrast, CUBIC/Reno/DCTCP see no loss or ECN, so they grow a very large cwnd, and thus large pacing rate and large TSO burst size. Change-Id: Ic8ccfdbe4010ee8d4bf6a6334c48a2fceb2171ea Signed-off-by: Alexandre Frade --- include/net/tcp.h | 4 ++-- net/ipv4/tcp_bbr.c | 37 ++++++++++++++++++++++++++----------- net/ipv4/tcp_output.c | 11 +++++------ 3 files changed, 33 insertions(+), 19 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 8357f0e1d..76c5eb9ff 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1103,8 +1103,8 @@ struct tcp_congestion_ops { /* hook for packet ack accounting (optional) */ void (*pkts_acked)(struct sock *sk, const struct ack_sample *sample); - /* override sysctl_tcp_min_tso_segs */ - u32 (*min_tso_segs)(struct sock *sk); + /* pick target number of segments per TSO/GSO skb (optional): */ + u32 (*tso_segs)(struct sock *sk, unsigned int mss_now); /* react to a specific lost skb (optional) */ void (*skb_marked_lost)(struct sock *sk, const struct sk_buff *skb); diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c index 54eec33c6..f2c6c5d31 100644 --- a/net/ipv4/tcp_bbr.c +++ b/net/ipv4/tcp_bbr.c @@ -300,20 +300,35 @@ 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. + */ +static u32 bbr_tso_segs_generic(struct sock *sk, unsigned int mss_now, + u32 gso_max_size) +{ + u32 segs; + u64 bytes; + + /* Budget a TSO/GSO burst size allowance based on bw (pacing_rate). */ + bytes = sk->sk_pacing_rate >> sk->sk_pacing_shift; + + bytes = min_t(u32, bytes, gso_max_size - 1 - MAX_TCP_HEADER); + segs = max_t(u32, bytes / mss_now, bbr_min_tso_segs(sk)); + 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) +{ + return bbr_tso_segs_generic(sk, mss_now, sk->sk_gso_max_size); +} + +/* Like bbr_tso_segs(), using mss_cache, ignoring driver's sk_gso_max_size. */ static u32 bbr_tso_segs_goal(struct sock *sk) { struct tcp_sock *tp = tcp_sk(sk); - u32 segs, bytes; - - /* Sort of tcp_tso_autosize() but ignoring - * driver provided sk_gso_max_size. - */ - bytes = min_t(unsigned long, - sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), - GSO_LEGACY_MAX_SIZE - 1 - MAX_TCP_HEADER); - segs = max_t(u32, bytes / tp->mss_cache, bbr_min_tso_segs(sk)); - return min(segs, 0x7FU); + return bbr_tso_segs_generic(sk, tp->mss_cache, GSO_MAX_SIZE); } /* Save "last known good" cwnd so we can restore it after losses or PROBE_RTT */ @@ -1149,7 +1164,7 @@ static struct tcp_congestion_ops tcp_bbr_cong_ops __read_mostly = { .undo_cwnd = bbr_undo_cwnd, .cwnd_event = bbr_cwnd_event, .ssthresh = bbr_ssthresh, - .min_tso_segs = bbr_min_tso_segs, + .tso_segs = bbr_tso_segs, .get_info = bbr_get_info, .set_state = bbr_set_state, }; diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index 0818709ec..ed9f0e766 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -2017,13 +2017,12 @@ static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) { const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; - u32 min_tso, tso_segs; + u32 tso_segs; - min_tso = ca_ops->min_tso_segs ? - ca_ops->min_tso_segs(sk) : - READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); - - tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); + tso_segs = ca_ops->tso_segs ? + ca_ops->tso_segs(sk, mss_now) : + tcp_tso_autosize(sk, mss_now, + sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); return min_t(u32, tso_segs, sk->sk_gso_max_segs); } -- 2.39.3 From d01907b3ae9f99911ce19383dd917761364c0dfe Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Sat, 16 Nov 2019 13:16:25 -0500 Subject: [PATCH 11/20] net-tcp: add fast_ack_mode=1: skip rwin check in tcp_fast_ack_mode__tcp_ack_snd_check() Add logic for an experimental TCP connection behavior, enabled with tp->fast_ack_mode = 1, which disables checking the receive window before sending an ack in __tcp_ack_snd_check(). If this behavior is enabled, the data receiver sends an ACK if the amount of data is > RCV.MSS. Change-Id: Iaa0a0fd7108221f883137a79d5bfa724f1b096d4 Signed-off-by: Alexandre Frade --- include/linux/tcp.h | 3 ++- net/ipv4/tcp.c | 1 + net/ipv4/tcp_cong.c | 1 + net/ipv4/tcp_input.c | 5 +++-- 4 files changed, 7 insertions(+), 3 deletions(-) diff --git a/include/linux/tcp.h b/include/linux/tcp.h index 9cd289ad3..bb8fa0c9d 100644 --- a/include/linux/tcp.h +++ b/include/linux/tcp.h @@ -255,7 +255,8 @@ struct tcp_sock { u8 compressed_ack; u8 dup_ack_counter:2, tlp_retrans:1, /* TLP is a retransmission */ - unused:5; + fast_ack_mode:2, /* which fast ack mode ? */ + unused:3; u32 chrono_start; /* Start time in jiffies of a TCP chrono */ u32 chrono_stat[3]; /* Time in jiffies for chrono_stat stats */ u8 chrono_type:2, /* current chronograph type */ diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index fab25d4f3..3acfcb9a7 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -3189,6 +3189,7 @@ int tcp_disconnect(struct sock *sk, int flags) tp->rx_opt.dsack = 0; tp->rx_opt.num_sacks = 0; tp->rcv_ooopack = 0; + tp->fast_ack_mode = 0; /* Clean up fastopen related fields */ diff --git a/net/ipv4/tcp_cong.c b/net/ipv4/tcp_cong.c index d3cae4074..0f268f2ff 100644 --- a/net/ipv4/tcp_cong.c +++ b/net/ipv4/tcp_cong.c @@ -189,6 +189,7 @@ void tcp_init_congestion_control(struct sock *sk) struct inet_connection_sock *icsk = inet_csk(sk); tcp_sk(sk)->prior_ssthresh = 0; + tcp_sk(sk)->fast_ack_mode = 0; if (icsk->icsk_ca_ops->init) icsk->icsk_ca_ops->init(sk); if (tcp_ca_needs_ecn(sk)) diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index a913e55d9..519b62e9b 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -5533,13 +5533,14 @@ static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) /* More than one full frame received... */ if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && + (tp->fast_ack_mode == 1 || /* ... and right edge of window advances far enough. * (tcp_recvmsg() will send ACK otherwise). * If application uses SO_RCVLOWAT, we want send ack now if * we have not received enough bytes to satisfy the condition. */ - (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || - __tcp_select_window(sk) >= tp->rcv_wnd)) || + (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || + __tcp_select_window(sk) >= tp->rcv_wnd))) || /* We ACK each frame or... */ tcp_in_quickack_mode(sk) || /* Protocol state mandates a one-time immediate ACK */ -- 2.39.3 From d7c8ac8358203eb530b9f31c4bea9f959f2c96f9 Mon Sep 17 00:00:00 2001 From: Jianfeng Wang Date: Fri, 19 Jun 2020 17:33:45 +0000 Subject: [PATCH 12/20] net-tcp_bbr: v2: record app-limited status of TLP-repaired flight When sending a TLP retransmit, record whether the outstanding flight of data is application limited. This is important for congestion control modules that want to respond to losses repaired by TLP retransmits. This is important because the following scenarios convey very different information: (1) a packet loss with a small number of packets in flight; (2) a packet loss with the maximum amount of data in flight allowed by the CC module; Effort: net-tcp_bbr Change-Id: Ic8ae567caa4e4bfd5fd82c3d4be12a5d9171655e Signed-off-by: Alexandre Frade --- include/linux/tcp.h | 3 ++- net/ipv4/tcp_output.c | 1 + 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/include/linux/tcp.h b/include/linux/tcp.h index bb8fa0c9d..f6a913256 100644 --- a/include/linux/tcp.h +++ b/include/linux/tcp.h @@ -256,7 +256,8 @@ struct tcp_sock { u8 dup_ack_counter:2, tlp_retrans:1, /* TLP is a retransmission */ fast_ack_mode:2, /* which fast ack mode ? */ - unused:3; + tlp_orig_data_app_limited:1, /* app-limited before TLP rtx? */ + unused:2; u32 chrono_start; /* Start time in jiffies of a TCP chrono */ u32 chrono_stat[3]; /* Time in jiffies for chrono_stat stats */ u8 chrono_type:2, /* current chronograph type */ diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index ed9f0e766..96bf72970 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -2871,6 +2871,7 @@ void tcp_send_loss_probe(struct sock *sk) if (WARN_ON(!skb || !tcp_skb_pcount(skb))) goto rearm_timer; + tp->tlp_orig_data_app_limited = TCP_SKB_CB(skb)->tx.is_app_limited; if (__tcp_retransmit_skb(sk, skb, 1)) goto rearm_timer; -- 2.39.3 From 7947c5068bf452016f862bd017a438433a257141 Mon Sep 17 00:00:00 2001 From: Jianfeng Wang Date: Tue, 16 Jun 2020 17:41:19 +0000 Subject: [PATCH 13/20] net-tcp_bbr: v2: inform CC module of losses repaired by TLP probe Before this commit, when there is a packet loss that creates a sequence hole that is filled by a TLP loss probe, then tcp_process_tlp_ack() only informs the congestion control (CC) module via a back-to-back entry and exit of CWR. But some congestion control modules (e.g. BBR) do not respond to CWR events. This commit adds a new CA event with which the core TCP stack notifies the CC module when a loss is repaired by a TLP. This will allow CC modules that do not use the CWR mechanism to have a custom handler for such TLP recoveries. Effort: net-tcp_bbr Change-Id: Ieba72332b401b329bff5a641d2b2043a3fb8f632 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 1 + net/ipv4/tcp_input.c | 1 + 2 files changed, 2 insertions(+) diff --git a/include/net/tcp.h b/include/net/tcp.h index 76c5eb9ff..4cd5d49b5 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1015,6 +1015,7 @@ enum tcp_ca_event { CA_EVENT_LOSS, /* loss timeout */ CA_EVENT_ECN_NO_CE, /* ECT set, but not CE marked */ CA_EVENT_ECN_IS_CE, /* received CE marked IP packet */ + CA_EVENT_TLP_RECOVERY, /* a lost segment was repaired by TLP probe */ }; /* Information about inbound ACK, passed to cong_ops->in_ack_event() */ diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 519b62e9b..04bfbed56 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -3720,6 +3720,7 @@ static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) /* ACK advances: there was a loss, so reduce cwnd. Reset * tlp_high_seq in tcp_init_cwnd_reduction() */ + tcp_ca_event(sk, CA_EVENT_TLP_RECOVERY); tcp_init_cwnd_reduction(sk); tcp_set_ca_state(sk, TCP_CA_CWR); tcp_end_cwnd_reduction(sk); -- 2.39.3 From f9fb32814af96b3aa983566376443bacacf98147 Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Mon, 21 Sep 2020 14:46:26 -0400 Subject: [PATCH 14/20] net-tcp_bbr: v2: introduce is_acking_tlp_retrans_seq into rate_sample Introduce is_acking_tlp_retrans_seq into rate_sample. This bool will export to the CC module the knowledge of whether the current ACK matched a TLP retransmit. Note that when this bool is true, we cannot yet tell (in general) whether this ACK is for the original or the TLP retransmit. Effort: net-tcp_bbr Change-Id: I2e6494332167e75efcbdc99bd5c119034e9c39b4 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 1 + net/ipv4/tcp_input.c | 12 +++++++++--- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 4cd5d49b5..8ea074553 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -1079,6 +1079,7 @@ struct rate_sample { u32 last_end_seq; /* end_seq of most recently ACKed packet */ bool is_app_limited; /* is sample from packet with bubble in pipe? */ bool is_retrans; /* is sample from retransmission? */ + bool is_acking_tlp_retrans_seq; /* ACKed a TLP retransmit sequence? */ bool is_ack_delayed; /* is this (likely) a delayed ACK? */ bool is_ece; /* did this ACK have ECN marked? */ }; diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 04bfbed56..1aa889ccc 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -3703,7 +3703,8 @@ static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) /* This routine deals with acks during a TLP episode and ends an episode by * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack */ -static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) +static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag, + struct rate_sample *rs) { struct tcp_sock *tp = tcp_sk(sk); @@ -3731,6 +3732,11 @@ static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) FLAG_NOT_DUP | FLAG_DATA_SACKED))) { /* Pure dupack: original and TLP probe arrived; no loss */ tp->tlp_high_seq = 0; + } else { + /* This ACK matches a TLP retransmit. We cannot yet tell if + * this ACK is for the original or the TLP retransmit. + */ + rs->is_acking_tlp_retrans_seq = 1; } } @@ -3910,7 +3916,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) tcp_rack_update_reo_wnd(sk, &rs); if (tp->tlp_high_seq) - tcp_process_tlp_ack(sk, ack, flag); + tcp_process_tlp_ack(sk, ack, flag, &rs); if (tcp_ack_is_dubious(sk, flag)) { if (!(flag & (FLAG_SND_UNA_ADVANCED | @@ -3954,7 +3960,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) tcp_ack_probe(sk); if (tp->tlp_high_seq) - tcp_process_tlp_ack(sk, ack, flag); + tcp_process_tlp_ack(sk, ack, flag, &rs); return 1; old_ack: -- 2.39.3 From 8eab18f14d084ed7fbddd5dabaad9b55424a2581 Mon Sep 17 00:00:00 2001 From: David Morley Date: Fri, 14 Jul 2023 11:07:56 -0400 Subject: [PATCH 15/20] tcp: introduce per-route feature RTAX_FEATURE_ECN_LOW Define and implement a new per-route feature, RTAX_FEATURE_ECN_LOW. This feature indicates that the given destination network is a low-latency ECN environment, meaning both that ECN CE marks are applied by the network using a low-latency marking threshold and also that TCP endpoints provide precise per-data-segment ECN feedback in ACKs (where the ACK ECE flag echoes the received CE status of all newly-acknowledged data segments). This feature indication can be used by congestion control algorithms to decide how to interpret ECN signals over the given destination network. This feature is appropriate for datacenter-style ECN marking, such as the ECN marking approach expected by DCTCP or BBR congestion control modules. Signed-off-by: David Morley Signed-off-by: Neal Cardwell Signed-off-by: Yuchung Cheng Tested-by: David Morley Change-Id: I6bc06e9c6cb426fbae7243fc71c9a8c18175f5d3 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 10 ++++++++++ include/uapi/linux/rtnetlink.h | 4 +++- net/ipv4/tcp_minisocks.c | 2 ++ net/ipv4/tcp_output.c | 6 ++++-- 4 files changed, 19 insertions(+), 3 deletions(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 8ea074553..3d762a2c7 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -375,6 +375,7 @@ static inline void tcp_dec_quickack_mode(struct sock *sk, #define TCP_ECN_QUEUE_CWR 2 #define TCP_ECN_DEMAND_CWR 4 #define TCP_ECN_SEEN 8 +#define TCP_ECN_LOW 16 enum tcp_tw_status { TCP_TW_SUCCESS = 0, @@ -728,6 +729,15 @@ static inline void tcp_fast_path_check(struct sock *sk) tcp_fast_path_on(tp); } +static inline void tcp_set_ecn_low_from_dst(struct sock *sk, + const struct dst_entry *dst) +{ + struct tcp_sock *tp = tcp_sk(sk); + + if (dst_feature(dst, RTAX_FEATURE_ECN_LOW)) + tp->ecn_flags |= TCP_ECN_LOW; +} + /* Compute the actual rto_min value */ static inline u32 tcp_rto_min(struct sock *sk) { diff --git a/include/uapi/linux/rtnetlink.h b/include/uapi/linux/rtnetlink.h index 51c13cf9c..de8dcba26 100644 --- a/include/uapi/linux/rtnetlink.h +++ b/include/uapi/linux/rtnetlink.h @@ -506,9 +506,11 @@ enum { #define RTAX_FEATURE_SACK (1 << 1) #define RTAX_FEATURE_TIMESTAMP (1 << 2) #define RTAX_FEATURE_ALLFRAG (1 << 3) +#define RTAX_FEATURE_ECN_LOW (1 << 4) #define RTAX_FEATURE_MASK (RTAX_FEATURE_ECN | RTAX_FEATURE_SACK | \ - RTAX_FEATURE_TIMESTAMP | RTAX_FEATURE_ALLFRAG) + RTAX_FEATURE_TIMESTAMP | RTAX_FEATURE_ALLFRAG \ + | RTAX_FEATURE_ECN_LOW) struct rta_session { __u8 proto; diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c index 42844d20d..e153569ef 100644 --- a/net/ipv4/tcp_minisocks.c +++ b/net/ipv4/tcp_minisocks.c @@ -423,6 +423,8 @@ void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); bool ca_got_dst = false; + tcp_set_ecn_low_from_dst(sk, dst); + if (ca_key != TCP_CA_UNSPEC) { const struct tcp_congestion_ops *ca; diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index 96bf72970..e302c8e0f 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -323,10 +323,9 @@ static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || tcp_ca_needs_ecn(sk) || bpf_needs_ecn; + const struct dst_entry *dst = __sk_dst_get(sk); if (!use_ecn) { - const struct dst_entry *dst = __sk_dst_get(sk); - if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) use_ecn = true; } @@ -338,6 +337,9 @@ static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) tp->ecn_flags = TCP_ECN_OK; if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) INET_ECN_xmit(sk); + + if (dst) + tcp_set_ecn_low_from_dst(sk, dst); } } -- 2.39.3 From 1b55df7d8e4fbe95b4a2fe2d49f2294aa293f856 Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:11 +0000 Subject: [PATCH 16/20] tcp: add sysctls for TCP PLB parameters PLB (Protective Load Balancing) is a host based mechanism for load balancing across switch links. It leverages congestion signals(e.g. ECN) from transport layer to randomly change the path of the connection experiencing congestion. PLB changes the path of the connection by changing the outgoing IPv6 flow label for IPv6 connections (implemented in Linux by calling sk_rethink_txhash()). Because of this implementation mechanism, PLB can currently only work for IPv6 traffic. For more information, see the SIGCOMM 2022 paper: https://doi.org/10.1145/3544216.3544226 This commit adds new sysctl knobs and sets their default values for TCP PLB. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller Signed-off-by: Alexandre Frade --- Documentation/networking/ip-sysctl.rst | 75 ++++++++++++++++++++++++++ include/net/netns/ipv4.h | 5 ++ net/ipv4/sysctl_net_ipv4.c | 43 +++++++++++++++ net/ipv4/tcp_ipv4.c | 8 +++ 4 files changed, 131 insertions(+) diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst index 3301288a7..f485521bc 100644 --- a/Documentation/networking/ip-sysctl.rst +++ b/Documentation/networking/ip-sysctl.rst @@ -1071,6 +1071,81 @@ tcp_child_ehash_entries - INTEGER Default: 0 +tcp_plb_enabled - BOOLEAN + If set and the underlying congestion control (e.g. DCTCP) supports + and enables PLB feature, TCP PLB (Protective Load Balancing) is + enabled. PLB is described in the following paper: + https://doi.org/10.1145/3544216.3544226. Based on PLB parameters, + upon sensing sustained congestion, TCP triggers a change in + flow label field for outgoing IPv6 packets. A change in flow label + field potentially changes the path of outgoing packets for switches + that use ECMP/WCMP for routing. + + PLB changes socket txhash which results in a change in IPv6 Flow Label + field, and currently no-op for IPv4 headers. It is possible + to apply PLB for IPv4 with other network header fields (e.g. TCP + or IPv4 options) or using encapsulation where outer header is used + by switches to determine next hop. In either case, further host + and switch side changes will be needed. + + When set, PLB assumes that congestion signal (e.g. ECN) is made + available and used by congestion control module to estimate a + congestion measure (e.g. ce_ratio). PLB needs a congestion measure to + make repathing decisions. + + Default: FALSE + +tcp_plb_idle_rehash_rounds - INTEGER + Number of consecutive congested rounds (RTT) seen after which + a rehash can be performed, given there are no packets in flight. + This is referred to as M in PLB paper: + https://doi.org/10.1145/3544216.3544226. + + Possible Values: 0 - 31 + + Default: 3 + +tcp_plb_rehash_rounds - INTEGER + Number of consecutive congested rounds (RTT) seen after which + a forced rehash can be performed. Be careful when setting this + parameter, as a small value increases the risk of retransmissions. + This is referred to as N in PLB paper: + https://doi.org/10.1145/3544216.3544226. + + Possible Values: 0 - 31 + + Default: 12 + +tcp_plb_suspend_rto_sec - INTEGER + Time, in seconds, to suspend PLB in event of an RTO. In order to avoid + having PLB repath onto a connectivity "black hole", after an RTO a TCP + connection suspends PLB repathing for a random duration between 1x and + 2x of this parameter. Randomness is added to avoid concurrent rehashing + of multiple TCP connections. This should be set corresponding to the + amount of time it takes to repair a failed link. + + Possible Values: 0 - 255 + + Default: 60 + +tcp_plb_cong_thresh - INTEGER + Fraction of packets marked with congestion over a round (RTT) to + tag that round as congested. This is referred to as K in the PLB paper: + https://doi.org/10.1145/3544216.3544226. + + The 0-1 fraction range is mapped to 0-256 range to avoid floating + point operations. For example, 128 means that if at least 50% of + the packets in a round were marked as congested then the round + will be tagged as congested. + + Setting threshold to 0 means that PLB repaths every RTT regardless + of congestion. This is not intended behavior for PLB and should be + used only for experimentation purpose. + + Possible Values: 0 - 256 + + Default: 128 + UDP variables ============= diff --git a/include/net/netns/ipv4.h b/include/net/netns/ipv4.h index 1b8004679..25f90bba4 100644 --- a/include/net/netns/ipv4.h +++ b/include/net/netns/ipv4.h @@ -183,6 +183,11 @@ struct netns_ipv4 { unsigned long tfo_active_disable_stamp; u32 tcp_challenge_timestamp; u32 tcp_challenge_count; + u8 sysctl_tcp_plb_enabled; + u8 sysctl_tcp_plb_idle_rehash_rounds; + u8 sysctl_tcp_plb_rehash_rounds; + u8 sysctl_tcp_plb_suspend_rto_sec; + int sysctl_tcp_plb_cong_thresh; int sysctl_udp_wmem_min; int sysctl_udp_rmem_min; diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c index f68762ce4..cd5914250 100644 --- a/net/ipv4/sysctl_net_ipv4.c +++ b/net/ipv4/sysctl_net_ipv4.c @@ -41,6 +41,8 @@ static int one_day_secs = 24 * 3600; static u32 fib_multipath_hash_fields_all_mask __maybe_unused = FIB_MULTIPATH_HASH_FIELD_ALL_MASK; static unsigned int tcp_child_ehash_entries_max = 16 * 1024 * 1024; +static int tcp_plb_max_rounds = 31; +static int tcp_plb_max_cong_thresh = 256; /* obsolete */ static int sysctl_tcp_low_latency __read_mostly; @@ -1387,6 +1389,47 @@ static struct ctl_table ipv4_net_table[] = { .extra1 = SYSCTL_ZERO, .extra2 = SYSCTL_TWO, }, + { + .procname = "tcp_plb_enabled", + .data = &init_net.ipv4.sysctl_tcp_plb_enabled, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .procname = "tcp_plb_idle_rehash_rounds", + .data = &init_net.ipv4.sysctl_tcp_plb_idle_rehash_rounds, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra2 = &tcp_plb_max_rounds, + }, + { + .procname = "tcp_plb_rehash_rounds", + .data = &init_net.ipv4.sysctl_tcp_plb_rehash_rounds, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra2 = &tcp_plb_max_rounds, + }, + { + .procname = "tcp_plb_suspend_rto_sec", + .data = &init_net.ipv4.sysctl_tcp_plb_suspend_rto_sec, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + }, + { + .procname = "tcp_plb_cong_thresh", + .data = &init_net.ipv4.sysctl_tcp_plb_cong_thresh, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &tcp_plb_max_cong_thresh, + }, { } }; diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c index f9b8a4a1d..27fe953aa 100644 --- a/net/ipv4/tcp_ipv4.c +++ b/net/ipv4/tcp_ipv4.c @@ -3213,6 +3213,14 @@ static int __net_init tcp_sk_init(struct net *net) net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0; atomic_set(&net->ipv4.tfo_active_disable_times, 0); + /* Set default values for PLB */ + net->ipv4.sysctl_tcp_plb_enabled = 0; /* Disabled by default */ + net->ipv4.sysctl_tcp_plb_idle_rehash_rounds = 3; + net->ipv4.sysctl_tcp_plb_rehash_rounds = 12; + net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60; + /* Default congestion threshold for PLB to mark a round is 50% */ + net->ipv4.sysctl_tcp_plb_cong_thresh = 128; + /* Reno is always built in */ if (!net_eq(net, &init_net) && bpf_try_module_get(init_net.ipv4.tcp_congestion_control, -- 2.39.3 From c576f4c44355caa393a67ad8cd830f5e6905bd5c Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:12 +0000 Subject: [PATCH 17/20] tcp: add PLB functionality for TCP Congestion control algorithms track PLB state and cause the connection to trigger a path change when either of the 2 conditions is satisfied: - No packets are in flight and (# consecutive congested rounds >= sysctl_tcp_plb_idle_rehash_rounds) - (# consecutive congested rounds >= sysctl_tcp_plb_rehash_rounds) A round (RTT) is marked as congested when congestion signal (ECN ce_ratio) over an RTT is greater than sysctl_tcp_plb_cong_thresh. In the event of RTO, PLB (via tcp_write_timeout()) triggers a path change and disables congestion-triggered path changes for random time between (sysctl_tcp_plb_suspend_rto_sec, 2*sysctl_tcp_plb_suspend_rto_sec) to avoid hopping onto the "connectivity blackhole". RTO-triggered path changes can still happen during this cool-off period. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller Signed-off-by: Alexandre Frade --- include/net/tcp.h | 28 ++++++++++++ net/ipv4/Makefile | 2 +- net/ipv4/tcp_ipv4.c | 2 +- net/ipv4/tcp_plb.c | 107 ++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 137 insertions(+), 2 deletions(-) create mode 100644 net/ipv4/tcp_plb.c diff --git a/include/net/tcp.h b/include/net/tcp.h index 3d762a2c7..2b89bf7c3 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -2217,6 +2217,34 @@ extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, extern void tcp_rack_reo_timeout(struct sock *sk); extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs); +/* tcp_plb.c */ + +/* + * Scaling factor for fractions in PLB. For example, tcp_plb_update_state + * expects cong_ratio which represents fraction of traffic that experienced + * congestion over a single RTT. In order to avoid floating point operations, + * this fraction should be mapped to (1 << TCP_PLB_SCALE) and passed in. + */ +#define TCP_PLB_SCALE 8 + +/* State for PLB (Protective Load Balancing) for a single TCP connection. */ +struct tcp_plb_state { + u8 consec_cong_rounds:5, /* consecutive congested rounds */ + unused:3; + u32 pause_until; /* jiffies32 when PLB can resume rerouting */ +}; + +static inline void tcp_plb_init(const struct sock *sk, + struct tcp_plb_state *plb) +{ + plb->consec_cong_rounds = 0; + plb->pause_until = 0; +} +void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb, + const int cong_ratio); +void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb); +void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb); + /* At how many usecs into the future should the RTO fire? */ static inline s64 tcp_rto_delta_us(const struct sock *sk) { diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile index bbdd9c44f..af7d2cf49 100644 --- a/net/ipv4/Makefile +++ b/net/ipv4/Makefile @@ -10,7 +10,7 @@ obj-y := route.o inetpeer.o protocol.o \ tcp.o tcp_input.o tcp_output.o tcp_timer.o tcp_ipv4.o \ tcp_minisocks.o tcp_cong.o tcp_metrics.o tcp_fastopen.o \ tcp_rate.o tcp_recovery.o tcp_ulp.o \ - tcp_offload.o datagram.o raw.o udp.o udplite.o \ + tcp_offload.o tcp_plb.o datagram.o raw.o udp.o udplite.o \ udp_offload.o arp.o icmp.o devinet.o af_inet.o igmp.o \ fib_frontend.o fib_semantics.o fib_trie.o fib_notifier.o \ inet_fragment.o ping.o ip_tunnel_core.o gre_offload.o \ diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c index 27fe953aa..428838dd2 100644 --- a/net/ipv4/tcp_ipv4.c +++ b/net/ipv4/tcp_ipv4.c @@ -3219,7 +3219,7 @@ static int __net_init tcp_sk_init(struct net *net) net->ipv4.sysctl_tcp_plb_rehash_rounds = 12; net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60; /* Default congestion threshold for PLB to mark a round is 50% */ - net->ipv4.sysctl_tcp_plb_cong_thresh = 128; + net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2; /* Reno is always built in */ if (!net_eq(net, &init_net) && diff --git a/net/ipv4/tcp_plb.c b/net/ipv4/tcp_plb.c new file mode 100644 index 000000000..f4ced370a --- /dev/null +++ b/net/ipv4/tcp_plb.c @@ -0,0 +1,107 @@ +/* Protective Load Balancing (PLB) + * + * PLB was designed to reduce link load imbalance across datacenter + * switches. PLB is a host-based optimization; it leverages congestion + * signals from the transport layer to randomly change the path of the + * connection experiencing sustained congestion. PLB prefers to repath + * after idle periods to minimize packet reordering. It repaths by + * changing the IPv6 Flow Label on the packets of a connection, which + * datacenter switches include as part of ECMP/WCMP hashing. + * + * PLB is described in detail in: + * + * Mubashir Adnan Qureshi, Yuchung Cheng, Qianwen Yin, Qiaobin Fu, + * Gautam Kumar, Masoud Moshref, Junhua Yan, Van Jacobson, + * David Wetherall,Abdul Kabbani: + * "PLB: Congestion Signals are Simple and Effective for + * Network Load Balancing" + * In ACM SIGCOMM 2022, Amsterdam Netherlands. + * + */ + +#include + +/* Called once per round-trip to update PLB state for a connection. */ +void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb, + const int cong_ratio) +{ + struct net *net = sock_net(sk); + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + if (cong_ratio >= 0) { + if (cong_ratio < READ_ONCE(net->ipv4.sysctl_tcp_plb_cong_thresh)) + plb->consec_cong_rounds = 0; + else if (plb->consec_cong_rounds < + READ_ONCE(net->ipv4.sysctl_tcp_plb_rehash_rounds)) + plb->consec_cong_rounds++; + } +} +EXPORT_SYMBOL_GPL(tcp_plb_update_state); + +/* Check whether recent congestion has been persistent enough to warrant + * a load balancing decision that switches the connection to another path. + */ +void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb) +{ + struct net *net = sock_net(sk); + u32 max_suspend; + bool forced_rehash = false, idle_rehash = false; + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + forced_rehash = plb->consec_cong_rounds >= + READ_ONCE(net->ipv4.sysctl_tcp_plb_rehash_rounds); + /* If sender goes idle then we check whether to rehash. */ + idle_rehash = READ_ONCE(net->ipv4.sysctl_tcp_plb_idle_rehash_rounds) && + !tcp_sk(sk)->packets_out && + plb->consec_cong_rounds >= + READ_ONCE(net->ipv4.sysctl_tcp_plb_idle_rehash_rounds); + + if (!forced_rehash && !idle_rehash) + return; + + /* Note that tcp_jiffies32 can wrap; we detect wraps by checking for + * cases where the max suspension end is before the actual suspension + * end. We clear pause_until to 0 to indicate there is no recent + * RTO event that constrains PLB rehashing. + */ + max_suspend = 2 * READ_ONCE(net->ipv4.sysctl_tcp_plb_suspend_rto_sec) * HZ; + if (plb->pause_until && + (!before(tcp_jiffies32, plb->pause_until) || + before(tcp_jiffies32 + max_suspend, plb->pause_until))) + plb->pause_until = 0; + + if (plb->pause_until) + return; + + sk_rethink_txhash(sk); + plb->consec_cong_rounds = 0; +} +EXPORT_SYMBOL_GPL(tcp_plb_check_rehash); + +/* Upon RTO, disallow load balancing for a while, to avoid having load + * balancing decisions switch traffic to a black-holed path that was + * previously avoided with a sk_rethink_txhash() call at RTO time. + */ +void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb) +{ + struct net *net = sock_net(sk); + u32 pause; + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + pause = READ_ONCE(net->ipv4.sysctl_tcp_plb_suspend_rto_sec) * HZ; + pause += prandom_u32_max(pause); + plb->pause_until = tcp_jiffies32 + pause; + + /* Reset PLB state upon RTO, since an RTO causes a sk_rethink_txhash() call + * that may switch this connection to a path with completely different + * congestion characteristics. + */ + plb->consec_cong_rounds = 0; +} +EXPORT_SYMBOL_GPL(tcp_plb_update_state_upon_rto); -- 2.39.3 From 79c762e218c1f30314946622231c508ae57c4fec Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Tue, 11 Jun 2019 12:54:22 -0400 Subject: [PATCH 18/20] net-tcp_bbr: v3: update TCP "bbr" congestion control module to BBRv3 BBR v3 is an enhacement to the BBR v1 algorithm. It's designed to aim for lower queues, lower loss, and better Reno/CUBIC coexistence than BBR v1. BBR v3 maintains the core of BBR v1: an explicit model of the network path that is two-dimensional, adapting to estimate the (a) maximum available bandwidth and (b) maximum safe volume of data a flow can keep in-flight in the network. It maintains the estimated BDP as a core guide for estimating an appropriate level of in-flight data. BBR v3 makes several key enhancements: o Its bandwidth-probing time scale is adapted, within bounds, to allow improved coexistence with Reno and CUBIC. 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. o Rather than being largely agnostic to loss and ECN marks, it explicitly uses loss and (DCTCP-style) ECN signals to maintain its model. o It aims for lower losses than v1 by adjusting its model to attempt to stay within loss rate and ECN mark rate bounds (loss_thresh and ecn_thresh, respectively). o It adapts to loss/ECN signals even when the application is running out of data ("application-limited"), in case the "application-limited" flow is also "network-limited" (the bw and/or inflight available to this flow is lower than previously estimated when the flow ran out of data). o It has a three-part model: the model explicit three tracks operating points, where an operating point is a tuple: (bandwidth, inflight). The three operating points are: o latest: the latest measurement from the current round trip o upper bound: robust, optimistic, long-term upper bound o lower bound: robust, conservative, short-term lower bound These are stored in the following state variables: o latest: bw_latest, inflight_latest o lo: bw_lo, inflight_lo o hi: bw_hi[2], inflight_hi To gain intuition about the meaning of the three operating points, it may help to consider the analogs in CUBIC, which has a somewhat analogous three-part model used by its probing state machine: BBR param CUBIC param ----------- ------------- latest ~ cwnd lo ~ ssthresh hi ~ last_max_cwnd The analogy is only a loose one, though, since the BBR operating points are calculated differently, and are 2-dimensional (bw,inflight) rather than CUBIC's one-dimensional notion of operating point (inflight). o It uses the three-part model to adapt the magnitude of its bandwidth to match the estimated space available in the buffer, rather than (as in BBR v1) assuming that it was always acceptable to place 0.25*BDP in the bottleneck buffer when probing (commodity datacenter switches commonly do not have that much buffer for WAN flows). When BBR v3 estimates it hit a buffer limit during probing, its bandwidth probing then starts gently in case little space is still available in the buffer, and the accelerates, slowly at first and then rapidly if it can grow inflight without seeing congestion signals. In such cases, probing is bounded by inflight_hi + inflight_probe, where inflight_probe grows as: [0, 1, 2, 4, 8, 16,...]. This allows BBR to keep losses low and bounded if a bottleneck remains congested, while rapidly/scalably utilizing free bandwidth when it becomes available. o It has a slightly revised state machine, to achieve the goals above. BBR_BW_PROBE_UP: pushes up inflight to probe for bw/vol BBR_BW_PROBE_DOWN: drain excess inflight from the queue BBR_BW_PROBE_CRUISE: use pipe, w/ headroom in queue/pipe BBR_BW_PROBE_REFILL: try refill the pipe again to 100%, leaving queue empty o The estimated BDP: BBR v3 continues to maintain an estimate of the path's two-way propagation delay, by tracking a windowed min_rtt, and coordinating (on an as-ndeeded basis) to try to expose the two-way propagation delay by draining the bottleneck queue. BBR v3 continues to use its min_rtt and (currently-applicable) bandwidth estimate to estimate the current bandwidth-delay product. The estimated BDP still provides one important guideline for bounding inflight data. However, because any min-filtered RTT and max-filtered bw inherently tend to both overestimate, the estimated BDP is often too high; in this case loss or ECN marks can ensue, in which case BBR v3 adjusts inflight_hi and inflight_lo to adapt its sending rate and inflight down to match the available capacity of the path. o Space: Note that ICSK_CA_PRIV_SIZE increased. This is because BBR v3 requires more space. Note that much of the space is due to support for per-socket parameterization and debugging in this release for research and debugging. With that state removed, the full "struct bbr" is 140 bytes, or 144 with padding. This is an increase of 40 bytes over the existing ca_priv space. o Code: BBR v3 reuses many pieces from BBR v1. But it omits the following significant pieces: o "packet conservation" (bbr_set_cwnd_to_recover_or_restore(), bbr_can_grow_inflight()) o long-term bandwidth estimator ("policer mode") The code layout tries to keep BBR v3 code near the bottom of the file, so that v1-applicable code in the top does not accidentally refer to v3 code. o Docs: See the following docs for more details and diagrams decsribing the BBR v3 algorithm: https://datatracker.ietf.org/meeting/104/materials/slides-104-iccrg-an-update-on-bbr-00 https://datatracker.ietf.org/meeting/102/materials/slides-102-iccrg-an-update-on-bbr-work-at-google-00 o Internal notes: For this upstream rebase, Neal started from: git show fed518041ac6:net/ipv4/tcp_bbr.c > net/ipv4/tcp_bbr.c then removed dev instrumentation (dynamic get/set for parameters) and code that was only used by BBRv1 Effort: net-tcp_bbr Origin-9xx-SHA1: 2c84098e60bed6d67dde23cd7538c51dee273102 Change-Id: I125cf26ba2a7a686f2fa5e87f4c2afceb65f7a05 Signed-off-by: Alexandre Frade --- include/net/inet_connection_sock.h | 4 +- include/net/tcp.h | 2 +- include/uapi/linux/inet_diag.h | 23 + net/ipv4/Kconfig | 21 +- net/ipv4/tcp_bbr.c | 2217 +++++++++++++++++++++------- 5 files changed, 1742 insertions(+), 525 deletions(-) diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h index c2b15f7e5..a400a8408 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 2b89bf7c3..0c717d468 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 50655de04..82f8bd8f0 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 2dfb12230..2e14db3be 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 f2c6c5d31..1ce016437 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; - u64 bytes; + struct bbr *bbr = inet_csk_ca(sk); + u32 segs, r; + u32 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)); -- 2.39.3 From 67e9f4d5a29f89ead84e2daab5143521c83b1c3b Mon Sep 17 00:00:00 2001 From: Adithya Abraham Philip Date: Fri, 11 Jun 2021 21:56:10 +0000 Subject: [PATCH 19/20] net-tcp_bbr: v3: ensure ECN-enabled BBR flows set ECT on retransmits Adds a new flag TCP_ECN_ECT_PERMANENT that is used by CCAs to indicate that retransmitted packets and pure ACKs must have the ECT bit set. This is necessary for BBR, which when using ECN expects ECT to be set even on retransmitted packets and ACKs. Previous to this addition of TCP_ECN_ECT_PERMANENT, CCAs which can use ECN but don't "need" it did not have a way to indicate that ECT should be set on retransmissions/ACKs. Signed-off-by: Adithya Abraham Philip Signed-off-by: Neal Cardwell Change-Id: I8b048eaab35e136fe6501ef6cd89fd9faa15e6d2 Signed-off-by: Alexandre Frade --- include/net/tcp.h | 1 + net/ipv4/tcp_bbr.c | 3 +++ net/ipv4/tcp_output.c | 3 ++- 3 files changed, 6 insertions(+), 1 deletion(-) diff --git a/include/net/tcp.h b/include/net/tcp.h index 0c717d468..09f52c135 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -376,6 +376,7 @@ static inline void tcp_dec_quickack_mode(struct sock *sk, #define TCP_ECN_DEMAND_CWR 4 #define TCP_ECN_SEEN 8 #define TCP_ECN_LOW 16 +#define TCP_ECN_ECT_PERMANENT 32 enum tcp_tw_status { TCP_TW_SUCCESS = 0, diff --git a/net/ipv4/tcp_bbr.c b/net/ipv4/tcp_bbr.c index 1ce016437..ff3bfff15 100644 --- a/net/ipv4/tcp_bbr.c +++ b/net/ipv4/tcp_bbr.c @@ -2151,6 +2151,9 @@ static void bbr_init(struct sock *sk) bbr->plb.pause_until = 0; tp->fast_ack_mode = bbr_fast_ack_mode ? 1 : 0; + + if (bbr_can_use_ecn(sk)) + tp->ecn_flags |= TCP_ECN_ECT_PERMANENT; } /* BBR marks the current round trip as a loss round. */ diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index e302c8e0f..88d1ea03a 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -377,7 +377,8 @@ static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, th->cwr = 1; skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; } - } else if (!tcp_ca_needs_ecn(sk)) { + } else if (!(tp->ecn_flags & TCP_ECN_ECT_PERMANENT) && + !tcp_ca_needs_ecn(sk)) { /* ACK or retransmitted segment: clear ECT|CE */ INET_ECN_dontxmit(sk); } -- 2.39.3 From dcee3968edf8691383ee3495555dd527d99b176e Mon Sep 17 00:00:00 2001 From: Neal Cardwell Date: Sun, 23 Jul 2023 23:25:34 -0400 Subject: [PATCH 20/20] tcp: export TCPI_OPT_ECN_LOW in tcp_info tcpi_options field Analogous to other important ECN information, export TCPI_OPT_ECN_LOW in tcp_info tcpi_options field. Signed-off-by: Neal Cardwell Change-Id: I08d8d8c7e8780e6e37df54038ee50301ac5a0320 Signed-off-by: Alexandre Frade --- include/uapi/linux/tcp.h | 1 + net/ipv4/tcp.c | 2 ++ 2 files changed, 3 insertions(+) diff --git a/include/uapi/linux/tcp.h b/include/uapi/linux/tcp.h index 8fc09e863..ed53f2535 100644 --- a/include/uapi/linux/tcp.h +++ b/include/uapi/linux/tcp.h @@ -170,6 +170,7 @@ enum tcp_fastopen_client_fail { #define TCPI_OPT_ECN 8 /* ECN was negociated at TCP session init */ #define TCPI_OPT_ECN_SEEN 16 /* we received at least one packet with ECT */ #define TCPI_OPT_SYN_DATA 32 /* SYN-ACK acked data in SYN sent or rcvd */ +#define TCPI_OPT_ECN_LOW 64 /* Low-latency ECN configured at init */ /* * Sender's congestion state indicating normal or abnormal situations diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index 3acfcb9a7..786a165da 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -3885,6 +3885,8 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info) info->tcpi_options |= TCPI_OPT_ECN; if (tp->ecn_flags & TCP_ECN_SEEN) info->tcpi_options |= TCPI_OPT_ECN_SEEN; + if (tp->ecn_flags & TCP_ECN_LOW) + info->tcpi_options |= TCPI_OPT_ECN_LOW; if (tp->syn_data_acked) info->tcpi_options |= TCPI_OPT_SYN_DATA; -- 2.39.3