Re: [PATCH net-next v3 09/15] quic: add congestion control
Xin Long <[email protected]> Tue, 23 Sep 2025 15:37:35 -0400
| Newsgroups | dev.linux.lists.quic,dev.linux.lists.kernel-tls-handshake,org.kernel.vger.linux-cifs,org.kernel.vger.netdev |
|---|---|
| Message-ID | <CADvbK_dxa0G-drauyggde2_46PdRfd9qzzmfuhj5NcMsOkGHRw@mail.gmail.com> |
On Tue, Sep 23, 2025 at 9:55 AM Paolo Abeni <[email protected]> wrote: > > On 9/19/25 12:34 AM, Xin Long wrote: > > This patch introduces 'quic_cong' for RTT measurement and congestion > > control. The 'quic_cong_ops' is added to define the congestion > > control algorithm. > > > > It implements a congestion control state machine with slow start, > > congestion avoidance, and recovery phases, and introduces the New > > Reno and CUBIC algorithms. > > To moderate the initial submission size, you could initially introduce > just one of the above. That sounds like a good idea, and I will keep the one specified in rfc9002: the New Reno, and remove the CUBIC. > > > The implementation updates RTT estimates when packets are acknowledged, > > reacts to loss and ECN signals, and adjusts the congestion window > > accordingly during packet transmission and acknowledgment processing. > > > > - quic_cong_rtt_update(): Performs RTT measurement, invoked when a > > packet is acknowledged by the largest number in the ACK frame. > > > > - quic_cong_on_packet_acked(): Invoked when a packet is acknowledged. > > > > - quic_cong_on_packet_lost(): Invoked when a packet is marked as lost. > > > > - quic_cong_on_process_ecn(): Invoked when an ACK_ECN frame is received. > > > > - quic_cong_on_packet_sent(): Invoked when a packet is transmitted. > > > > - quic_cong_on_ack_recv(): Invoked when an ACK frame is received. > > > > Signed-off-by: Xin Long <[email protected]> > > --- > > net/quic/Makefile | 3 +- > > net/quic/cong.c | 700 ++++++++++++++++++++++++++++++++++++++++++++++ > > net/quic/cong.h | 120 ++++++++ > > net/quic/socket.c | 1 + > > net/quic/socket.h | 7 + > > 5 files changed, 830 insertions(+), 1 deletion(-) > > create mode 100644 net/quic/cong.c > > create mode 100644 net/quic/cong.h > > > > diff --git a/net/quic/Makefile b/net/quic/Makefile > > index 1565fb5cef9d..4d4a42c6d565 100644 > > --- a/net/quic/Makefile > > +++ b/net/quic/Makefile > > @@ -5,4 +5,5 @@ > > > > obj-$(CONFIG_IP_QUIC) += quic.o > > > > -quic-y := common.o family.o protocol.o socket.o stream.o connid.o path.o > > +quic-y := common.o family.o protocol.o socket.o stream.o connid.o path.o \ > > + cong.o > > diff --git a/net/quic/cong.c b/net/quic/cong.c > > new file mode 100644 > > index 000000000000..d598cc14b15e > > --- /dev/null > > +++ b/net/quic/cong.c > > @@ -0,0 +1,700 @@ > > +// SPDX-License-Identifier: GPL-2.0-or-later > > +/* QUIC kernel implementation > > + * (C) Copyright Red Hat Corp. 2023 > > + * > > + * This file is part of the QUIC kernel implementation > > + * > > + * Initialization/cleanup for QUIC protocol support. > > + * > > + * Written or modified by: > > + * Xin Long <[email protected]> > > + */ > > + > > +#include <linux/jiffies.h> > > +#include <linux/quic.h> > > +#include <net/sock.h> > > + > > +#include "common.h" > > +#include "cong.h" > > + > > +/* CUBIC APIs */ > > +struct quic_cubic { > > + /* Variables of Interest in rfc9438#section-4.1.2 */ > > + u32 pending_w_add; /* Accumulate fractional increments to W_est */ > > + u32 origin_point; /* W_max */ > > + u32 epoch_start; /* t_epoch */ > > + u32 pending_add; /* Accumulates fractional additions to W_cubic */ > > + u32 w_last_max; /* last W_max */ > > + u32 w_tcp; /* W_est */ > > + u64 k; /* K */ > > + > > + /* HyStart++ variables in rfc9406#section-4.2 */ > > + u32 current_round_min_rtt; /* currentRoundMinRTT */ > > + u32 css_baseline_min_rtt; /* cssBaselineMinRtt */ > > + u32 last_round_min_rtt; /* lastRoundMinRTT */ > > + u16 rtt_sample_count; /* rttSampleCount */ > > + u16 css_rounds; /* Counter for consecutive rounds showing RTT increase */ > > + s64 window_end; /* End of current CSS round (packet number) */ > > +}; > > + > > +/* HyStart++ constants in rfc9406#section-4.3 */ > > +#define QUIC_HS_MIN_SSTHRESH 16 > > +#define QUIC_HS_N_RTT_SAMPLE 8 > > +#define QUIC_HS_MIN_ETA 4000 > > +#define QUIC_HS_MAX_ETA 16000 > > +#define QUIC_HS_MIN_RTT_DIVISOR 8 > > +#define QUIC_HS_CSS_GROWTH_DIVISOR 4 > > +#define QUIC_HS_CSS_ROUNDS 5 > > + > > +static u64 cubic_root(u64 n) > > +{ > > + u64 a, d; > > + > > + if (!n) > > + return 0; > > + > > + d = (64 - __builtin_clzll(n)) / 3; > > + a = BIT_ULL(d + 1); > > + > > + for (; a * a * a > n;) { > > + d = div64_ul(n, a * a); > > + a = div64_ul(2 * a + d, 3); > > + } > > + return a; > > +} > > tcp_cubic() has already an helper to compute the square root. You could > re-use that one. > > > + > > +/* rfc9406#section-4: HyStart++ Algorithm */ > > +static void cubic_slow_start(struct quic_cong *cong, u32 bytes, s64 number) > > +{ > > + struct quic_cubic *cubic = quic_cong_priv(cong); > > + u32 eta; > > + > > + if (cubic->window_end <= number) > > + cubic->window_end = -1; > > + > > + /* cwnd = cwnd + (min(N, L * SMSS) / CSS_GROWTH_DIVISOR) */ > > + if (cubic->css_baseline_min_rtt != U32_MAX) > > + bytes = bytes / QUIC_HS_CSS_GROWTH_DIVISOR; > > + cong->window = min_t(u32, cong->window + bytes, cong->max_window); > > + > > + if (cubic->css_baseline_min_rtt != U32_MAX) { > > + /* If CSS_ROUNDS rounds are complete, enter congestion avoidance. */ > > + if (++cubic->css_rounds > QUIC_HS_CSS_ROUNDS) { > > + cubic->css_baseline_min_rtt = U32_MAX; > > + cubic->w_last_max = cong->window; > > + cong->ssthresh = cong->window; > > + cubic->css_rounds = 0; > > + } > > + return; > > + } > > + > > + /* if ((rttSampleCount >= N_RTT_SAMPLE) AND > > + * (currentRoundMinRTT != infinity) AND > > + * (lastRoundMinRTT != infinity)) > > + * RttThresh = max(MIN_RTT_THRESH, > > + * min(lastRoundMinRTT / MIN_RTT_DIVISOR, MAX_RTT_THRESH)) > > + * if (currentRoundMinRTT >= (lastRoundMinRTT + RttThresh)) > > + * cssBaselineMinRtt = currentRoundMinRTT > > + * exit slow start and enter CSS > > + */ > > + if (cubic->last_round_min_rtt != U32_MAX && > > + cubic->current_round_min_rtt != U32_MAX && > > + cong->window >= QUIC_HS_MIN_SSTHRESH * cong->mss && > > + cubic->rtt_sample_count >= QUIC_HS_N_RTT_SAMPLE) { > > + eta = cubic->last_round_min_rtt / QUIC_HS_MIN_RTT_DIVISOR; > > + if (eta < QUIC_HS_MIN_ETA) > > + eta = QUIC_HS_MIN_ETA; > > + else if (eta > QUIC_HS_MAX_ETA) > > + eta = QUIC_HS_MAX_ETA; > > + > > + pr_debug("%s: current_round_min_rtt: %u, last_round_min_rtt: %u, eta: %u\n", > > + __func__, cubic->current_round_min_rtt, cubic->last_round_min_rtt, eta); > > + > > + /* Delay increase triggers slow start exit and enter CSS. */ > > + if (cubic->current_round_min_rtt >= cubic->last_round_min_rtt + eta) > > + cubic->css_baseline_min_rtt = cubic->current_round_min_rtt; > > + } > > +} > > + > > +/* rfc9438#section-4: CUBIC Congestion Control */ > > +static void cubic_cong_avoid(struct quic_cong *cong, u32 bytes) > > +{ > > + struct quic_cubic *cubic = quic_cong_priv(cong); > > + u64 tx, kx, time_delta, delta, t; > > + u64 target_add, tcp_add = 0; > > + u64 target, cwnd_thres, m; > > + > > + if (cubic->epoch_start == U32_MAX) { > > + cubic->epoch_start = cong->time; > > + if (cong->window < cubic->w_last_max) { > > + /* > > + * ┌────────────────┐ > > + * 3 │W - cwnd > > + * ╲ │ max epoch > > + * K = ╲ │──────────────── > > + * ╲│ C > > + */ > > + cubic->k = cubic->w_last_max - cong->window; > > + cubic->k = cubic_root(div64_ul(cubic->k * 10, (u64)cong->mss * 4)); > > Can `mss` be 0 at this point? Why? There is QUIC_PATH_MIN_PMTU (1200) defined. quic_flow_route() must be done and mss is calculated to set cong->mss via quic_cong_set_mss() based on pmtu before coming to this place. It should be ensured on the 2nd patchset, I will double check it. > > > + cubic->origin_point = cubic->w_last_max; > > + } else { > > + cubic->k = 0; > > + cubic->origin_point = cong->window; > > + } > > + cubic->w_tcp = cong->window; > > + cubic->pending_add = 0; > > + cubic->pending_w_add = 0; > > + } > > + > > + /* > > + * t = t - t > > + * current epoch > > + */ > > + t = cong->time - cubic->epoch_start; > > + tx = div64_ul(t << 10, USEC_PER_SEC); > > + kx = (cubic->k << 10); > > + if (tx > kx) > > + time_delta = tx - kx; > > + else > > + time_delta = kx - tx; > > + /* > > + * 3 > > + * W (t) = C * (t - K) + W > > + * cubic max > > + */ > > + delta = cong->mss * ((((time_delta * time_delta) >> 10) * time_delta) >> 10); > > + delta = div64_ul(delta * 4, 10) >> 10; > > + if (tx > kx) > > + target = cubic->origin_point + delta; > > + else > > + target = cubic->origin_point - delta; > > + > > + /* > > + * W (t + RTT) > > + * cubic > > + */ > > + cwnd_thres = (div64_ul((t + cong->smoothed_rtt) << 10, USEC_PER_SEC) * target) >> 10; > > + pr_debug("%s: tgt: %llu, thres: %llu, delta: %llu, t: %llu, srtt: %u, tx: %llu, kx: %llu\n", > > + __func__, target, cwnd_thres, delta, t, cong->smoothed_rtt, tx, kx); > > + /* > > + * ⎧ > > + * ⎪cwnd if W (t + RTT) < cwnd > > + * ⎪ cubic > > + * ⎨1.5 * cwnd if W (t + RTT) > 1.5 * cwnd > > + * target = ⎪ cubic > > + * ⎪W (t + RTT) otherwise > > + * ⎩ cubic > > + */ > > + if (cwnd_thres < cong->window) > > + target = cong->window; > > + else if (cwnd_thres * 2 > (u64)cong->window * 3) > > + target = cong->window * 3 / 2; > > + else > > + target = cwnd_thres; > > + > > + /* > > + * target - cwnd > > + * ───────────── > > + * cwnd > > + */ > > + if (target > cong->window) { > > + target_add = cubic->pending_add + cong->mss * (target - cong->window); > > + cubic->pending_add = do_div(target_add, cong->window); > > + } else { > > + target_add = cubic->pending_add + cong->mss; > > + cubic->pending_add = do_div(target_add, 100 * cong->window); > > + } > > Can `window` be 0 here? why? It should not. When changing cong->window, there's always a check to cong->min_window, which is set via quic_cong_set_mss(). > > > + > > + pr_debug("%s: target: %llu, window: %u, target_add: %llu\n", > > + __func__, target, cong->window, target_add); > > + > > + /* > > + * segments_acked > > + * W = W + α * ────────────── > > + * est est cubic cwnd > > + */ > > + m = cubic->pending_w_add + cong->mss * bytes; > > + cubic->pending_w_add = do_div(m, cong->window); > > + cubic->w_tcp += m; > > + > > + if (cubic->w_tcp > cong->window) > > + tcp_add = div64_ul((u64)cong->mss * (cubic->w_tcp - cong->window), cong->window); > > + > > + pr_debug("%s: w_tcp: %u, window: %u, tcp_add: %llu\n", > > + __func__, cubic->w_tcp, cong->window, tcp_add); > > + > > + /* W_cubic(_t_) or _W_est_, whichever is bigger. */ > > + cong->window += max(tcp_add, target_add); > > +} > > + > > +static void cubic_recovery(struct quic_cong *cong) > > +{ > > + struct quic_cubic *cubic = quic_cong_priv(cong); > > + > > + cong->recovery_time = cong->time; > > + cubic->epoch_start = U32_MAX; > > + > > + /* rfc9438#section-3.4: > > + * CUBIC sets the multiplicative window decrease factor (β__cubic_) to 0.7, > > + * whereas Reno uses 0.5. > > + * > > + * rfc9438#section-4.6: > > + * ssthresh = flight_size * β new ssthresh > > + * > > + * Some implementations of CUBIC currently use _cwnd_ instead of _flight_size_ when > > + * calculating a new _ssthresh_. > > + * > > + * rfc9438#section-4.7: > > + * > > + * ⎧ 1 + β > > + * ⎪ cubic > > + * ⎪cwnd * ────────── if cwnd < W_max and fast convergence > > + * W = ⎨ 2 > > + * max ⎪ enabled, further reduce W_max > > + * ⎪ > > + * ⎩cwnd otherwise, remember cwnd before reduction > > + */ > > + if (cong->window < cubic->w_last_max) > > + cubic->w_last_max = cong->window * 17 / 10 / 2; > > + else > > + cubic->w_last_max = cong->window; > > + > > + cong->ssthresh = cong->window * 7 / 10; > > There are quite a bit of magic numbers that should be replaced by macros > and/or associated with explainatory comments. Yes, better to use macros. I was expecting people to understand if from the comment I put in above. > > > + > > +/* rfc9002#section-5: Estimating the Round-Trip Time */ > > +void quic_cong_rtt_update(struct quic_cong *cong, u32 time, u32 ack_delay) > > +{ > > + u32 adjusted_rtt, rttvar_sample; > > + > > + /* Ignore RTT sample if ACK delay is suspiciously large. */ > > + if (ack_delay > cong->max_ack_delay * 2) > > + return; > > + > > + /* rfc9002#section-5.1: latest_rtt = ack_time - send_time_of_largest_acked */ > > + cong->latest_rtt = cong->time - time; > > + > > + /* rfc9002#section-5.2: Estimating min_rtt */ > > + if (!cong->min_rtt_valid) { > > + cong->min_rtt = cong->latest_rtt; > > + cong->min_rtt_valid = 1; > > + } > > + if (cong->min_rtt > cong->latest_rtt) > > + cong->min_rtt = cong->latest_rtt; > > + > > + if (!cong->is_rtt_set) { > > + /* rfc9002#section-5.3: > > + * smoothed_rtt = latest_rtt > > + * rttvar = latest_rtt / 2 > > + */ > > + cong->smoothed_rtt = cong->latest_rtt; > > + cong->rttvar = cong->smoothed_rtt / 2; > > + quic_cong_pto_update(cong); > > + cong->is_rtt_set = 1; > > + return; > > + } > > + > > + /* rfc9002#section-5.3: > > + * adjusted_rtt = latest_rtt > > + * if (latest_rtt >= min_rtt + ack_delay): > > + * adjusted_rtt = latest_rtt - ack_delay > > + * smoothed_rtt = 7/8 * smoothed_rtt + 1/8 * adjusted_rtt > > + * rttvar_sample = abs(smoothed_rtt - adjusted_rtt) > > + * rttvar = 3/4 * rttvar + 1/4 * rttvar_sample > > + */ > > + adjusted_rtt = cong->latest_rtt; > > + if (cong->latest_rtt >= cong->min_rtt + ack_delay) > > + adjusted_rtt = cong->latest_rtt - ack_delay; > > + > > + cong->smoothed_rtt = (cong->smoothed_rtt * 7 + adjusted_rtt) / 8; > > + if (cong->smoothed_rtt >= adjusted_rtt) > > + rttvar_sample = cong->smoothed_rtt - adjusted_rtt; > > + else > > + rttvar_sample = adjusted_rtt - cong->smoothed_rtt; > > Here in a few other place before you could use abs_diff() Sure. Thanks.