net/ipv4/tcp_cubic.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
0004  * Home page:
0005  *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
0006  * This is from the implementation of CUBIC TCP in
0007  * Sangtae Ha, Injong Rhee and Lisong Xu,
0008  *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant"
0009  *  in ACM SIGOPS Operating System Review, July 2008.
0010  * Available from:
0011  *  http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
0012  *
0013  * CUBIC integrates a new slow start algorithm, called HyStart.
0014  * The details of HyStart are presented in
0015  *  Sangtae Ha and Injong Rhee,
0016  *  "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008.
0017  * Available from:
0018  *  http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
0019  *
0020  * All testing results are available from:
0021  * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
0022  *
0023  * Unless CUBIC is enabled and congestion window is large
0024  * this behaves the same as the original Reno.
0025  */
0026
0027 #include <linux/mm.h>
0028 #include <linux/btf.h>
0029 #include <linux/btf_ids.h>
0030 #include <linux/module.h>
0031 #include <linux/math64.h>
0032 #include <net/tcp.h>
0033
0034 #define BICTCP_BETA_SCALE    1024   /* Scale factor beta calculation
0035                      * max_cwnd = snd_cwnd * beta
0036                      */
0037 #define BICTCP_HZ       10  /* BIC HZ 2^10 = 1024 */
0038
0039 /* Two methods of hybrid slow start */
0040 #define HYSTART_ACK_TRAIN   0x1
0041 #define HYSTART_DELAY       0x2
0042
0043 /* Number of delay samples for detecting the increase of delay */
0044 #define HYSTART_MIN_SAMPLES 8
0045 #define HYSTART_DELAY_MIN   (4000U) /* 4 ms */
0046 #define HYSTART_DELAY_MAX   (16000U)    /* 16 ms */
0047 #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
0048
0049 static int fast_convergence __read_mostly = 1;
0050 static int beta __read_mostly = 717;    /* = 717/1024 (BICTCP_BETA_SCALE) */
0051 static int initial_ssthresh __read_mostly;
0052 static int bic_scale __read_mostly = 41;
0053 static int tcp_friendliness __read_mostly = 1;
0054
0055 static int hystart __read_mostly = 1;
0056 static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
0057 static int hystart_low_window __read_mostly = 16;
0058 static int hystart_ack_delta_us __read_mostly = 2000;
0059
0060 static u32 cube_rtt_scale __read_mostly;
0061 static u32 beta_scale __read_mostly;
0062 static u64 cube_factor __read_mostly;
0063
0064 /* Note parameters that are used for precomputing scale factors are read-only */
0065 module_param(fast_convergence, int, 0644);
0066 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
0067 module_param(beta, int, 0644);
0068 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
0069 module_param(initial_ssthresh, int, 0644);
0070 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
0071 module_param(bic_scale, int, 0444);
0072 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
0073 module_param(tcp_friendliness, int, 0644);
0074 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
0075 module_param(hystart, int, 0644);
0076 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
0077 module_param(hystart_detect, int, 0644);
0078 MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms"
0079          " 1: packet-train 2: delay 3: both packet-train and delay");
0080 module_param(hystart_low_window, int, 0644);
0081 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
0082 module_param(hystart_ack_delta_us, int, 0644);
0083 MODULE_PARM_DESC(hystart_ack_delta_us, "spacing between ack's indicating train (usecs)");
0084
0085 /* BIC TCP Parameters */
0086 struct bictcp {
0087     u32 cnt;        /* increase cwnd by 1 after ACKs */
0088     u32 last_max_cwnd;  /* last maximum snd_cwnd */
0089     u32 last_cwnd;  /* the last snd_cwnd */
0090     u32 last_time;  /* time when updated last_cwnd */
0091     u32 bic_origin_point;/* origin point of bic function */
0092     u32 bic_K;      /* time to origin point
0093                    from the beginning of the current epoch */
0094     u32 delay_min;  /* min delay (usec) */
0095     u32 epoch_start;    /* beginning of an epoch */
0096     u32 ack_cnt;    /* number of acks */
0097     u32 tcp_cwnd;   /* estimated tcp cwnd */
0098     u16 unused;
0099     u8  sample_cnt; /* number of samples to decide curr_rtt */
0100     u8  found;      /* the exit point is found? */
0101     u32 round_start;    /* beginning of each round */
0102     u32 end_seq;    /* end_seq of the round */
0103     u32 last_ack;   /* last time when the ACK spacing is close */
0104     u32 curr_rtt;   /* the minimum rtt of current round */
0105 };
0106
0107 static inline void bictcp_reset(struct bictcp *ca)
0108 {
0109     memset(ca, 0, offsetof(struct bictcp, unused));
0110     ca->found = 0;
0111 }
0112
0113 static inline u32 bictcp_clock_us(const struct sock *sk)
0114 {
0115     return tcp_sk(sk)->tcp_mstamp;
0116 }
0117
0118 static inline void bictcp_hystart_reset(struct sock *sk)
0119 {
0120     struct tcp_sock *tp = tcp_sk(sk);
0121     struct bictcp *ca = inet_csk_ca(sk);
0122
0123     ca->round_start = ca->last_ack = bictcp_clock_us(sk);
0124     ca->end_seq = tp->snd_nxt;
0125     ca->curr_rtt = ~0U;
0126     ca->sample_cnt = 0;
0127 }
0128
0129 static void cubictcp_init(struct sock *sk)
0130 {
0131     struct bictcp *ca = inet_csk_ca(sk);
0132
0133     bictcp_reset(ca);
0134
0135     if (hystart)
0136         bictcp_hystart_reset(sk);
0137
0138     if (!hystart && initial_ssthresh)
0139         tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
0140 }
0141
0142 static void cubictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
0143 {
0144     if (event == CA_EVENT_TX_START) {
0145         struct bictcp *ca = inet_csk_ca(sk);
0146         u32 now = tcp_jiffies32;
0147         s32 delta;
0148
0149         delta = now - tcp_sk(sk)->lsndtime;
0150
0151         /* We were application limited (idle) for a while.
0152          * Shift epoch_start to keep cwnd growth to cubic curve.
0153          */
0154         if (ca->epoch_start && delta > 0) {
0155             ca->epoch_start += delta;
0156             if (after(ca->epoch_start, now))
0157                 ca->epoch_start = now;
0158         }
0159         return;
0160     }
0161 }
0162
0163 /* calculate the cubic root of x using a table lookup followed by one
0164  * Newton-Raphson iteration.
0165  * Avg err ~= 0.195%
0166  */
0167 static u32 cubic_root(u64 a)
0168 {
0169     u32 x, b, shift;
0170     /*
0171      * cbrt(x) MSB values for x MSB values in [0..63].
0172      * Precomputed then refined by hand - Willy Tarreau
0173      *
0174      * For x in [0..63],
0175      *   v = cbrt(x << 18) - 1
0176      *   cbrt(x) = (v[x] + 10) >> 6
0177      */
0178     static const u8 v[] = {
0179         /* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
0180         /* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
0181         /* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
0182         /* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
0183         /* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
0184         /* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
0185         /* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
0186         /* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
0187     };
0188
0189     b = fls64(a);
0190     if (b < 7) {
0191         /* a in [0..63] */
0192         return ((u32)v[(u32)a] + 35) >> 6;
0193     }
0194
0195     b = ((b * 84) >> 8) - 1;
0196     shift = (a >> (b * 3));
0197
0198     x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
0199
0200     /*
0201      * Newton-Raphson iteration
0202      *                         2
0203      * x    = ( 2 * x  +  a / x  ) / 3
0204      *  k+1          k         k
0205      */
0206     x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
0207     x = ((x * 341) >> 10);
0208     return x;
0209 }
0210
0211 /*
0212  * Compute congestion window to use.
0213  */
0214 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
0215 {
0216     u32 delta, bic_target, max_cnt;
0217     u64 offs, t;
0218
0219     ca->ack_cnt += acked;   /* count the number of ACKed packets */
0220
0221     if (ca->last_cwnd == cwnd &&
0222         (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
0223         return;
0224
0225     /* The CUBIC function can update ca->cnt at most once per jiffy.
0226      * On all cwnd reduction events, ca->epoch_start is set to 0,
0227      * which will force a recalculation of ca->cnt.
0228      */
0229     if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
0230         goto tcp_friendliness;
0231
0232     ca->last_cwnd = cwnd;
0233     ca->last_time = tcp_jiffies32;
0234
0235     if (ca->epoch_start == 0) {
0236         ca->epoch_start = tcp_jiffies32;    /* record beginning */
0237         ca->ack_cnt = acked;            /* start counting */
0238         ca->tcp_cwnd = cwnd;            /* syn with cubic */
0239
0240         if (ca->last_max_cwnd <= cwnd) {
0241             ca->bic_K = 0;
0242             ca->bic_origin_point = cwnd;
0243         } else {
0244             /* Compute new K based on
0245              * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
0246              */
0247             ca->bic_K = cubic_root(cube_factor
0248                            * (ca->last_max_cwnd - cwnd));
0249             ca->bic_origin_point = ca->last_max_cwnd;
0250         }
0251     }
0252
0253     /* cubic function - calc*/
0254     /* calculate c * time^3 / rtt,
0255      *  while considering overflow in calculation of time^3
0256      * (so time^3 is done by using 64 bit)
0257      * and without the support of division of 64bit numbers
0258      * (so all divisions are done by using 32 bit)
0259      *  also NOTE the unit of those veriables
0260      *    time  = (t - K) / 2^bictcp_HZ
0261      *    c = bic_scale >> 10
0262      * rtt  = (srtt >> 3) / HZ
0263      * !!! The following code does not have overflow problems,
0264      * if the cwnd < 1 million packets !!!
0265      */
0266
0267     t = (s32)(tcp_jiffies32 - ca->epoch_start);
0268     t += usecs_to_jiffies(ca->delay_min);
0269     /* change the unit from HZ to bictcp_HZ */
0270     t <<= BICTCP_HZ;
0271     do_div(t, HZ);
0272
0273     if (t < ca->bic_K)      /* t - K */
0274         offs = ca->bic_K - t;
0275     else
0276         offs = t - ca->bic_K;
0277
0278     /* c/rtt * (t-K)^3 */
0279     delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
0280     if (t < ca->bic_K)                            /* below origin*/
0281         bic_target = ca->bic_origin_point - delta;
0282     else                                          /* above origin*/
0283         bic_target = ca->bic_origin_point + delta;
0284
0285     /* cubic function - calc bictcp_cnt*/
0286     if (bic_target > cwnd) {
0287         ca->cnt = cwnd / (bic_target - cwnd);
0288     } else {
0289         ca->cnt = 100 * cwnd;              /* very small increment*/
0290     }
0291
0292     /*
0293      * The initial growth of cubic function may be too conservative
0294      * when the available bandwidth is still unknown.
0295      */
0296     if (ca->last_max_cwnd == 0 && ca->cnt > 20)
0297         ca->cnt = 20;   /* increase cwnd 5% per RTT */
0298
0299 tcp_friendliness:
0300     /* TCP Friendly */
0301     if (tcp_friendliness) {
0302         u32 scale = beta_scale;
0303
0304         delta = (cwnd * scale) >> 3;
0305         while (ca->ack_cnt > delta) {       /* update tcp cwnd */
0306             ca->ack_cnt -= delta;
0307             ca->tcp_cwnd++;
0308         }
0309
0310         if (ca->tcp_cwnd > cwnd) {  /* if bic is slower than tcp */
0311             delta = ca->tcp_cwnd - cwnd;
0312             max_cnt = cwnd / delta;
0313             if (ca->cnt > max_cnt)
0314                 ca->cnt = max_cnt;
0315         }
0316     }
0317
0318     /* The maximum rate of cwnd increase CUBIC allows is 1 packet per
0319      * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
0320      */
0321     ca->cnt = max(ca->cnt, 2U);
0322 }
0323
0324 static void cubictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
0325 {
0326     struct tcp_sock *tp = tcp_sk(sk);
0327     struct bictcp *ca = inet_csk_ca(sk);
0328
0329     if (!tcp_is_cwnd_limited(sk))
0330         return;
0331
0332     if (tcp_in_slow_start(tp)) {
0333         acked = tcp_slow_start(tp, acked);
0334         if (!acked)
0335             return;
0336     }
0337     bictcp_update(ca, tcp_snd_cwnd(tp), acked);
0338     tcp_cong_avoid_ai(tp, ca->cnt, acked);
0339 }
0340
0341 static u32 cubictcp_recalc_ssthresh(struct sock *sk)
0342 {
0343     const struct tcp_sock *tp = tcp_sk(sk);
0344     struct bictcp *ca = inet_csk_ca(sk);
0345
0346     ca->epoch_start = 0;    /* end of epoch */
0347
0348     /* Wmax and fast convergence */
0349     if (tcp_snd_cwnd(tp) < ca->last_max_cwnd && fast_convergence)
0350         ca->last_max_cwnd = (tcp_snd_cwnd(tp) * (BICTCP_BETA_SCALE + beta))
0351             / (2 * BICTCP_BETA_SCALE);
0352     else
0353         ca->last_max_cwnd = tcp_snd_cwnd(tp);
0354
0355     return max((tcp_snd_cwnd(tp) * beta) / BICTCP_BETA_SCALE, 2U);
0356 }
0357
0358 static void cubictcp_state(struct sock *sk, u8 new_state)
0359 {
0360     if (new_state == TCP_CA_Loss) {
0361         bictcp_reset(inet_csk_ca(sk));
0362         bictcp_hystart_reset(sk);
0363     }
0364 }
0365
0366 /* Account for TSO/GRO delays.
0367  * Otherwise short RTT flows could get too small ssthresh, since during
0368  * slow start we begin with small TSO packets and ca->delay_min would
0369  * not account for long aggregation delay when TSO packets get bigger.
0370  * Ideally even with a very small RTT we would like to have at least one
0371  * TSO packet being sent and received by GRO, and another one in qdisc layer.
0372  * We apply another 100% factor because @rate is doubled at this point.
0373  * We cap the cushion to 1ms.
0374  */
0375 static u32 hystart_ack_delay(const struct sock *sk)
0376 {
0377     unsigned long rate;
0378
0379     rate = READ_ONCE(sk->sk_pacing_rate);
0380     if (!rate)
0381         return 0;
0382     return min_t(u64, USEC_PER_MSEC,
0383              div64_ul((u64)sk->sk_gso_max_size * 4 * USEC_PER_SEC, rate));
0384 }
0385
0386 static void hystart_update(struct sock *sk, u32 delay)
0387 {
0388     struct tcp_sock *tp = tcp_sk(sk);
0389     struct bictcp *ca = inet_csk_ca(sk);
0390     u32 threshold;
0391
0392     if (after(tp->snd_una, ca->end_seq))
0393         bictcp_hystart_reset(sk);
0394
0395     if (hystart_detect & HYSTART_ACK_TRAIN) {
0396         u32 now = bictcp_clock_us(sk);
0397
0398         /* first detection parameter - ack-train detection */
0399         if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) {
0400             ca->last_ack = now;
0401
0402             threshold = ca->delay_min + hystart_ack_delay(sk);
0403
0404             /* Hystart ack train triggers if we get ack past
0405              * ca->delay_min/2.
0406              * Pacing might have delayed packets up to RTT/2
0407              * during slow start.
0408              */
0409             if (sk->sk_pacing_status == SK_PACING_NONE)
0410                 threshold >>= 1;
0411
0412             if ((s32)(now - ca->round_start) > threshold) {
0413                 ca->found = 1;
0414                 pr_debug("hystart_ack_train (%u > %u) delay_min %u (+ ack_delay %u) cwnd %u\n",
0415                      now - ca->round_start, threshold,
0416                      ca->delay_min, hystart_ack_delay(sk), tcp_snd_cwnd(tp));
0417                 NET_INC_STATS(sock_net(sk),
0418                           LINUX_MIB_TCPHYSTARTTRAINDETECT);
0419                 NET_ADD_STATS(sock_net(sk),
0420                           LINUX_MIB_TCPHYSTARTTRAINCWND,
0421                           tcp_snd_cwnd(tp));
0422                 tp->snd_ssthresh = tcp_snd_cwnd(tp);
0423             }
0424         }
0425     }
0426
0427     if (hystart_detect & HYSTART_DELAY) {
0428         /* obtain the minimum delay of more than sampling packets */
0429         if (ca->curr_rtt > delay)
0430             ca->curr_rtt = delay;
0431         if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
0432             ca->sample_cnt++;
0433         } else {
0434             if (ca->curr_rtt > ca->delay_min +
0435                 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
0436                 ca->found = 1;
0437                 NET_INC_STATS(sock_net(sk),
0438                           LINUX_MIB_TCPHYSTARTDELAYDETECT);
0439                 NET_ADD_STATS(sock_net(sk),
0440                           LINUX_MIB_TCPHYSTARTDELAYCWND,
0441                           tcp_snd_cwnd(tp));
0442                 tp->snd_ssthresh = tcp_snd_cwnd(tp);
0443             }
0444         }
0445     }
0446 }
0447
0448 static void cubictcp_acked(struct sock *sk, const struct ack_sample *sample)
0449 {
0450     const struct tcp_sock *tp = tcp_sk(sk);
0451     struct bictcp *ca = inet_csk_ca(sk);
0452     u32 delay;
0453
0454     /* Some calls are for duplicates without timetamps */
0455     if (sample->rtt_us < 0)
0456         return;
0457
0458     /* Discard delay samples right after fast recovery */
0459     if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
0460         return;
0461
0462     delay = sample->rtt_us;
0463     if (delay == 0)
0464         delay = 1;
0465
0466     /* first time call or link delay decreases */
0467     if (ca->delay_min == 0 || ca->delay_min > delay)
0468         ca->delay_min = delay;
0469
0470     /* hystart triggers when cwnd is larger than some threshold */
0471     if (!ca->found && tcp_in_slow_start(tp) && hystart &&
0472         tcp_snd_cwnd(tp) >= hystart_low_window)
0473         hystart_update(sk, delay);
0474 }
0475
0476 static struct tcp_congestion_ops cubictcp __read_mostly = {
0477     .init       = cubictcp_init,
0478     .ssthresh   = cubictcp_recalc_ssthresh,
0479     .cong_avoid = cubictcp_cong_avoid,
0480     .set_state  = cubictcp_state,
0481     .undo_cwnd  = tcp_reno_undo_cwnd,
0482     .cwnd_event = cubictcp_cwnd_event,
0483     .pkts_acked     = cubictcp_acked,
0484     .owner      = THIS_MODULE,
0485     .name       = "cubic",
0486 };
0487
0488 BTF_SET8_START(tcp_cubic_check_kfunc_ids)
0489 #ifdef CONFIG_X86
0490 #ifdef CONFIG_DYNAMIC_FTRACE
0491 BTF_ID_FLAGS(func, cubictcp_init)
0492 BTF_ID_FLAGS(func, cubictcp_recalc_ssthresh)
0493 BTF_ID_FLAGS(func, cubictcp_cong_avoid)
0494 BTF_ID_FLAGS(func, cubictcp_state)
0495 BTF_ID_FLAGS(func, cubictcp_cwnd_event)
0496 BTF_ID_FLAGS(func, cubictcp_acked)
0497 #endif
0498 #endif
0499 BTF_SET8_END(tcp_cubic_check_kfunc_ids)
0500
0501 static const struct btf_kfunc_id_set tcp_cubic_kfunc_set = {
0502     .owner = THIS_MODULE,
0503     .set   = &tcp_cubic_check_kfunc_ids,
0504 };
0505
0506 static int __init cubictcp_register(void)
0507 {
0508     int ret;
0509
0510     BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
0511
0512     /* Precompute a bunch of the scaling factors that are used per-packet
0513      * based on SRTT of 100ms
0514      */
0515
0516     beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
0517         / (BICTCP_BETA_SCALE - beta);
0518
0519     cube_rtt_scale = (bic_scale * 10);  /* 1024*c/rtt */
0520
0521     /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
0522      *  so K = cubic_root( (wmax-cwnd)*rtt/c )
0523      * the unit of K is bictcp_HZ=2^10, not HZ
0524      *
0525      *  c = bic_scale >> 10
0526      *  rtt = 100ms
0527      *
0528      * the following code has been designed and tested for
0529      * cwnd < 1 million packets
0530      * RTT < 100 seconds
0531      * HZ < 1,000,00  (corresponding to 10 nano-second)
0532      */
0533
0534     /* 1/c * 2^2*bictcp_HZ * srtt */
0535     cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
0536
0537     /* divide by bic_scale and by constant Srtt (100ms) */
0538     do_div(cube_factor, bic_scale * 10);
0539
0540     ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_cubic_kfunc_set);
0541     if (ret < 0)
0542         return ret;
0543     return tcp_register_congestion_control(&cubictcp);
0544 }
0545
0546 static void __exit cubictcp_unregister(void)
0547 {
0548     tcp_unregister_congestion_control(&cubictcp);
0549 }
0550
0551 module_init(cubictcp_register);
0552 module_exit(cubictcp_unregister);
0553
0554 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
0555 MODULE_LICENSE("GPL");
0556 MODULE_DESCRIPTION("CUBIC TCP");
0557 MODULE_VERSION("2.3");