net/ipv4/tcp_cdg.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * CAIA Delay-Gradient (CDG) congestion control
0004  *
0005  * This implementation is based on the paper:
0006  *   D.A. Hayes and G. Armitage. "Revisiting TCP congestion control using
0007  *   delay gradients." In IFIP Networking, pages 328-341. Springer, 2011.
0008  *
0009  * Scavenger traffic (Less-than-Best-Effort) should disable coexistence
0010  * heuristics using parameters use_shadow=0 and use_ineff=0.
0011  *
0012  * Parameters window, backoff_beta, and backoff_factor are crucial for
0013  * throughput and delay. Future work is needed to determine better defaults,
0014  * and to provide guidelines for use in different environments/contexts.
0015  *
0016  * Except for window, knobs are configured via /sys/module/tcp_cdg/parameters/.
0017  * Parameter window is only configurable when loading tcp_cdg as a module.
0018  *
0019  * Notable differences from paper/FreeBSD:
0020  *   o Using Hybrid Slow start and Proportional Rate Reduction.
0021  *   o Add toggle for shadow window mechanism. Suggested by David Hayes.
0022  *   o Add toggle for non-congestion loss tolerance.
0023  *   o Scaling parameter G is changed to a backoff factor;
0024  *     conversion is given by: backoff_factor = 1000/(G * window).
0025  *   o Limit shadow window to 2 * cwnd, or to cwnd when application limited.
0026  *   o More accurate e^-x.
0027  */
0028 #include <linux/kernel.h>
0029 #include <linux/random.h>
0030 #include <linux/module.h>
0031 #include <linux/sched/clock.h>
0032
0033 #include <net/tcp.h>
0034
0035 #define HYSTART_ACK_TRAIN   1
0036 #define HYSTART_DELAY       2
0037
0038 static int window __read_mostly = 8;
0039 static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
0040 static unsigned int backoff_factor __read_mostly = 42;
0041 static unsigned int hystart_detect __read_mostly = 3;
0042 static unsigned int use_ineff __read_mostly = 5;
0043 static bool use_shadow __read_mostly = true;
0044 static bool use_tolerance __read_mostly;
0045
0046 module_param(window, int, 0444);
0047 MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
0048 module_param(backoff_beta, uint, 0644);
0049 MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
0050 module_param(backoff_factor, uint, 0644);
0051 MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
0052 module_param(hystart_detect, uint, 0644);
0053 MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
0054          "(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
0055 module_param(use_ineff, uint, 0644);
0056 MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
0057 module_param(use_shadow, bool, 0644);
0058 MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
0059 module_param(use_tolerance, bool, 0644);
0060 MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
0061
0062 struct cdg_minmax {
0063     union {
0064         struct {
0065             s32 min;
0066             s32 max;
0067         };
0068         u64 v64;
0069     };
0070 };
0071
0072 enum cdg_state {
0073     CDG_UNKNOWN = 0,
0074     CDG_NONFULL = 1,
0075     CDG_FULL    = 2,
0076     CDG_BACKOFF = 3,
0077 };
0078
0079 struct cdg {
0080     struct cdg_minmax rtt;
0081     struct cdg_minmax rtt_prev;
0082     struct cdg_minmax *gradients;
0083     struct cdg_minmax gsum;
0084     bool gfilled;
0085     u8  tail;
0086     u8  state;
0087     u8  delack;
0088     u32 rtt_seq;
0089     u32 shadow_wnd;
0090     u16 backoff_cnt;
0091     u16 sample_cnt;
0092     s32 delay_min;
0093     u32 last_ack;
0094     u32 round_start;
0095 };
0096
0097 /**
0098  * nexp_u32 - negative base-e exponential
0099  * @ux: x in units of micro
0100  *
0101  * Returns exp(ux * -1e-6) * U32_MAX.
0102  */
0103 static u32 __pure nexp_u32(u32 ux)
0104 {
0105     static const u16 v[] = {
0106         /* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
0107         65535,
0108         65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
0109         61378, 57484, 50423, 38795, 22965, 8047,  987,   14,
0110     };
0111     u32 msb = ux >> 8;
0112     u32 res;
0113     int i;
0114
0115     /* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
0116     if (msb > U16_MAX)
0117         return 0;
0118
0119     /* Scale first eight bits linearly: */
0120     res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
0121
0122     /* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
0123     for (i = 1; msb; i++, msb >>= 1) {
0124         u32 y = v[i & -(msb & 1)] + U32_C(1);
0125
0126         res = ((u64)res * y) >> 16;
0127     }
0128
0129     return res;
0130 }
0131
0132 /* Based on the HyStart algorithm (by Ha et al.) that is implemented in
0133  * tcp_cubic. Differences/experimental changes:
0134  *   o Using Hayes' delayed ACK filter.
0135  *   o Using a usec clock for the ACK train.
0136  *   o Reset ACK train when application limited.
0137  *   o Invoked at any cwnd (i.e. also when cwnd < 16).
0138  *   o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
0139  */
0140 static void tcp_cdg_hystart_update(struct sock *sk)
0141 {
0142     struct cdg *ca = inet_csk_ca(sk);
0143     struct tcp_sock *tp = tcp_sk(sk);
0144
0145     ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
0146     if (ca->delay_min == 0)
0147         return;
0148
0149     if (hystart_detect & HYSTART_ACK_TRAIN) {
0150         u32 now_us = tp->tcp_mstamp;
0151
0152         if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk)) {
0153             ca->last_ack = now_us;
0154             ca->round_start = now_us;
0155         } else if (before(now_us, ca->last_ack + 3000)) {
0156             u32 base_owd = max(ca->delay_min / 2U, 125U);
0157
0158             ca->last_ack = now_us;
0159             if (after(now_us, ca->round_start + base_owd)) {
0160                 NET_INC_STATS(sock_net(sk),
0161                           LINUX_MIB_TCPHYSTARTTRAINDETECT);
0162                 NET_ADD_STATS(sock_net(sk),
0163                           LINUX_MIB_TCPHYSTARTTRAINCWND,
0164                           tcp_snd_cwnd(tp));
0165                 tp->snd_ssthresh = tcp_snd_cwnd(tp);
0166                 return;
0167             }
0168         }
0169     }
0170
0171     if (hystart_detect & HYSTART_DELAY) {
0172         if (ca->sample_cnt < 8) {
0173             ca->sample_cnt++;
0174         } else {
0175             s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
0176                      125U);
0177
0178             if (ca->rtt.min > thresh) {
0179                 NET_INC_STATS(sock_net(sk),
0180                           LINUX_MIB_TCPHYSTARTDELAYDETECT);
0181                 NET_ADD_STATS(sock_net(sk),
0182                           LINUX_MIB_TCPHYSTARTDELAYCWND,
0183                           tcp_snd_cwnd(tp));
0184                 tp->snd_ssthresh = tcp_snd_cwnd(tp);
0185             }
0186         }
0187     }
0188 }
0189
0190 static s32 tcp_cdg_grad(struct cdg *ca)
0191 {
0192     s32 gmin = ca->rtt.min - ca->rtt_prev.min;
0193     s32 gmax = ca->rtt.max - ca->rtt_prev.max;
0194     s32 grad;
0195
0196     if (ca->gradients) {
0197         ca->gsum.min += gmin - ca->gradients[ca->tail].min;
0198         ca->gsum.max += gmax - ca->gradients[ca->tail].max;
0199         ca->gradients[ca->tail].min = gmin;
0200         ca->gradients[ca->tail].max = gmax;
0201         ca->tail = (ca->tail + 1) & (window - 1);
0202         gmin = ca->gsum.min;
0203         gmax = ca->gsum.max;
0204     }
0205
0206     /* We keep sums to ignore gradients during cwnd reductions;
0207      * the paper's smoothed gradients otherwise simplify to:
0208      * (rtt_latest - rtt_oldest) / window.
0209      *
0210      * We also drop division by window here.
0211      */
0212     grad = gmin > 0 ? gmin : gmax;
0213
0214     /* Extrapolate missing values in gradient window: */
0215     if (!ca->gfilled) {
0216         if (!ca->gradients && window > 1)
0217             grad *= window; /* Memory allocation failed. */
0218         else if (ca->tail == 0)
0219             ca->gfilled = true;
0220         else
0221             grad = (grad * window) / (int)ca->tail;
0222     }
0223
0224     /* Backoff was effectual: */
0225     if (gmin <= -32 || gmax <= -32)
0226         ca->backoff_cnt = 0;
0227
0228     if (use_tolerance) {
0229         /* Reduce small variations to zero: */
0230         gmin = DIV_ROUND_CLOSEST(gmin, 64);
0231         gmax = DIV_ROUND_CLOSEST(gmax, 64);
0232
0233         if (gmin > 0 && gmax <= 0)
0234             ca->state = CDG_FULL;
0235         else if ((gmin > 0 && gmax > 0) || gmax < 0)
0236             ca->state = CDG_NONFULL;
0237     }
0238     return grad;
0239 }
0240
0241 static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
0242 {
0243     struct cdg *ca = inet_csk_ca(sk);
0244     struct tcp_sock *tp = tcp_sk(sk);
0245
0246     if (prandom_u32() <= nexp_u32(grad * backoff_factor))
0247         return false;
0248
0249     if (use_ineff) {
0250         ca->backoff_cnt++;
0251         if (ca->backoff_cnt > use_ineff)
0252             return false;
0253     }
0254
0255     ca->shadow_wnd = max(ca->shadow_wnd, tcp_snd_cwnd(tp));
0256     ca->state = CDG_BACKOFF;
0257     tcp_enter_cwr(sk);
0258     return true;
0259 }
0260
0261 /* Not called in CWR or Recovery state. */
0262 static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
0263 {
0264     struct cdg *ca = inet_csk_ca(sk);
0265     struct tcp_sock *tp = tcp_sk(sk);
0266     u32 prior_snd_cwnd;
0267     u32 incr;
0268
0269     if (tcp_in_slow_start(tp) && hystart_detect)
0270         tcp_cdg_hystart_update(sk);
0271
0272     if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
0273         s32 grad = 0;
0274
0275         if (ca->rtt_prev.v64)
0276             grad = tcp_cdg_grad(ca);
0277         ca->rtt_seq = tp->snd_nxt;
0278         ca->rtt_prev = ca->rtt;
0279         ca->rtt.v64 = 0;
0280         ca->last_ack = 0;
0281         ca->sample_cnt = 0;
0282
0283         if (grad > 0 && tcp_cdg_backoff(sk, grad))
0284             return;
0285     }
0286
0287     if (!tcp_is_cwnd_limited(sk)) {
0288         ca->shadow_wnd = min(ca->shadow_wnd, tcp_snd_cwnd(tp));
0289         return;
0290     }
0291
0292     prior_snd_cwnd = tcp_snd_cwnd(tp);
0293     tcp_reno_cong_avoid(sk, ack, acked);
0294
0295     incr = tcp_snd_cwnd(tp) - prior_snd_cwnd;
0296     ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
0297 }
0298
0299 static void tcp_cdg_acked(struct sock *sk, const struct ack_sample *sample)
0300 {
0301     struct cdg *ca = inet_csk_ca(sk);
0302     struct tcp_sock *tp = tcp_sk(sk);
0303
0304     if (sample->rtt_us <= 0)
0305         return;
0306
0307     /* A heuristic for filtering delayed ACKs, adapted from:
0308      * D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
0309      * delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
0310      */
0311     if (tp->sacked_out == 0) {
0312         if (sample->pkts_acked == 1 && ca->delack) {
0313             /* A delayed ACK is only used for the minimum if it is
0314              * provenly lower than an existing non-zero minimum.
0315              */
0316             ca->rtt.min = min(ca->rtt.min, sample->rtt_us);
0317             ca->delack--;
0318             return;
0319         } else if (sample->pkts_acked > 1 && ca->delack < 5) {
0320             ca->delack++;
0321         }
0322     }
0323
0324     ca->rtt.min = min_not_zero(ca->rtt.min, sample->rtt_us);
0325     ca->rtt.max = max(ca->rtt.max, sample->rtt_us);
0326 }
0327
0328 static u32 tcp_cdg_ssthresh(struct sock *sk)
0329 {
0330     struct cdg *ca = inet_csk_ca(sk);
0331     struct tcp_sock *tp = tcp_sk(sk);
0332
0333     if (ca->state == CDG_BACKOFF)
0334         return max(2U, (tcp_snd_cwnd(tp) * min(1024U, backoff_beta)) >> 10);
0335
0336     if (ca->state == CDG_NONFULL && use_tolerance)
0337         return tcp_snd_cwnd(tp);
0338
0339     ca->shadow_wnd = min(ca->shadow_wnd >> 1, tcp_snd_cwnd(tp));
0340     if (use_shadow)
0341         return max3(2U, ca->shadow_wnd, tcp_snd_cwnd(tp) >> 1);
0342     return max(2U, tcp_snd_cwnd(tp) >> 1);
0343 }
0344
0345 static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
0346 {
0347     struct cdg *ca = inet_csk_ca(sk);
0348     struct tcp_sock *tp = tcp_sk(sk);
0349     struct cdg_minmax *gradients;
0350
0351     switch (ev) {
0352     case CA_EVENT_CWND_RESTART:
0353         gradients = ca->gradients;
0354         if (gradients)
0355             memset(gradients, 0, window * sizeof(gradients[0]));
0356         memset(ca, 0, sizeof(*ca));
0357
0358         ca->gradients = gradients;
0359         ca->rtt_seq = tp->snd_nxt;
0360         ca->shadow_wnd = tcp_snd_cwnd(tp);
0361         break;
0362     case CA_EVENT_COMPLETE_CWR:
0363         ca->state = CDG_UNKNOWN;
0364         ca->rtt_seq = tp->snd_nxt;
0365         ca->rtt_prev = ca->rtt;
0366         ca->rtt.v64 = 0;
0367         break;
0368     default:
0369         break;
0370     }
0371 }
0372
0373 static void tcp_cdg_init(struct sock *sk)
0374 {
0375     struct cdg *ca = inet_csk_ca(sk);
0376     struct tcp_sock *tp = tcp_sk(sk);
0377
0378     /* We silently fall back to window = 1 if allocation fails. */
0379     if (window > 1)
0380         ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
0381                     GFP_NOWAIT | __GFP_NOWARN);
0382     ca->rtt_seq = tp->snd_nxt;
0383     ca->shadow_wnd = tcp_snd_cwnd(tp);
0384 }
0385
0386 static void tcp_cdg_release(struct sock *sk)
0387 {
0388     struct cdg *ca = inet_csk_ca(sk);
0389
0390     kfree(ca->gradients);
0391 }
0392
0393 static struct tcp_congestion_ops tcp_cdg __read_mostly = {
0394     .cong_avoid = tcp_cdg_cong_avoid,
0395     .cwnd_event = tcp_cdg_cwnd_event,
0396     .pkts_acked = tcp_cdg_acked,
0397     .undo_cwnd = tcp_reno_undo_cwnd,
0398     .ssthresh = tcp_cdg_ssthresh,
0399     .release = tcp_cdg_release,
0400     .init = tcp_cdg_init,
0401     .owner = THIS_MODULE,
0402     .name = "cdg",
0403 };
0404
0405 static int __init tcp_cdg_register(void)
0406 {
0407     if (backoff_beta > 1024 || window < 1 || window > 256)
0408         return -ERANGE;
0409     if (!is_power_of_2(window))
0410         return -EINVAL;
0411
0412     BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
0413     tcp_register_congestion_control(&tcp_cdg);
0414     return 0;
0415 }
0416
0417 static void __exit tcp_cdg_unregister(void)
0418 {
0419     tcp_unregister_congestion_control(&tcp_cdg);
0420 }
0421
0422 module_init(tcp_cdg_register);
0423 module_exit(tcp_cdg_unregister);
0424 MODULE_AUTHOR("Kenneth Klette Jonassen");
0425 MODULE_LICENSE("GPL");
0426 MODULE_DESCRIPTION("TCP CDG");