net/ipv4/tcp_bic.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Binary Increase Congestion control for TCP
0004  * Home page:
0005  *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
0006  * This is from the implementation of BICTCP in
0007  * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
0008  *  "Binary Increase Congestion Control for Fast, Long Distance
0009  *  Networks" in InfoComm 2004
0010  * Available from:
0011  *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
0012  *
0013  * Unless BIC is enabled and congestion window is large
0014  * this behaves the same as the original Reno.
0015  */
0016
0017 #include <linux/mm.h>
0018 #include <linux/module.h>
0019 #include <net/tcp.h>
0020
0021 #define BICTCP_BETA_SCALE    1024   /* Scale factor beta calculation
0022                      * max_cwnd = snd_cwnd * beta
0023                      */
0024 #define BICTCP_B        4    /*
0025                       * In binary search,
0026                       * go to point (max+min)/N
0027                       */
0028
0029 static int fast_convergence = 1;
0030 static int max_increment = 16;
0031 static int low_window = 14;
0032 static int beta = 819;      /* = 819/1024 (BICTCP_BETA_SCALE) */
0033 static int initial_ssthresh;
0034 static int smooth_part = 20;
0035
0036 module_param(fast_convergence, int, 0644);
0037 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
0038 module_param(max_increment, int, 0644);
0039 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
0040 module_param(low_window, int, 0644);
0041 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
0042 module_param(beta, int, 0644);
0043 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
0044 module_param(initial_ssthresh, int, 0644);
0045 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
0046 module_param(smooth_part, int, 0644);
0047 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
0048
0049 /* BIC TCP Parameters */
0050 struct bictcp {
0051     u32 cnt;        /* increase cwnd by 1 after ACKs */
0052     u32 last_max_cwnd;  /* last maximum snd_cwnd */
0053     u32 last_cwnd;  /* the last snd_cwnd */
0054     u32 last_time;  /* time when updated last_cwnd */
0055     u32 epoch_start;    /* beginning of an epoch */
0056 #define ACK_RATIO_SHIFT 4
0057     u32 delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
0058 };
0059
0060 static inline void bictcp_reset(struct bictcp *ca)
0061 {
0062     ca->cnt = 0;
0063     ca->last_max_cwnd = 0;
0064     ca->last_cwnd = 0;
0065     ca->last_time = 0;
0066     ca->epoch_start = 0;
0067     ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
0068 }
0069
0070 static void bictcp_init(struct sock *sk)
0071 {
0072     struct bictcp *ca = inet_csk_ca(sk);
0073
0074     bictcp_reset(ca);
0075
0076     if (initial_ssthresh)
0077         tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
0078 }
0079
0080 /*
0081  * Compute congestion window to use.
0082  */
0083 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
0084 {
0085     if (ca->last_cwnd == cwnd &&
0086         (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
0087         return;
0088
0089     ca->last_cwnd = cwnd;
0090     ca->last_time = tcp_jiffies32;
0091
0092     if (ca->epoch_start == 0) /* record the beginning of an epoch */
0093         ca->epoch_start = tcp_jiffies32;
0094
0095     /* start off normal */
0096     if (cwnd <= low_window) {
0097         ca->cnt = cwnd;
0098         return;
0099     }
0100
0101     /* binary increase */
0102     if (cwnd < ca->last_max_cwnd) {
0103         __u32   dist = (ca->last_max_cwnd - cwnd)
0104             / BICTCP_B;
0105
0106         if (dist > max_increment)
0107             /* linear increase */
0108             ca->cnt = cwnd / max_increment;
0109         else if (dist <= 1U)
0110             /* binary search increase */
0111             ca->cnt = (cwnd * smooth_part) / BICTCP_B;
0112         else
0113             /* binary search increase */
0114             ca->cnt = cwnd / dist;
0115     } else {
0116         /* slow start AMD linear increase */
0117         if (cwnd < ca->last_max_cwnd + BICTCP_B)
0118             /* slow start */
0119             ca->cnt = (cwnd * smooth_part) / BICTCP_B;
0120         else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
0121             /* slow start */
0122             ca->cnt = (cwnd * (BICTCP_B-1))
0123                 / (cwnd - ca->last_max_cwnd);
0124         else
0125             /* linear increase */
0126             ca->cnt = cwnd / max_increment;
0127     }
0128
0129     /* if in slow start or link utilization is very low */
0130     if (ca->last_max_cwnd == 0) {
0131         if (ca->cnt > 20) /* increase cwnd 5% per RTT */
0132             ca->cnt = 20;
0133     }
0134
0135     ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
0136     if (ca->cnt == 0)           /* cannot be zero */
0137         ca->cnt = 1;
0138 }
0139
0140 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
0141 {
0142     struct tcp_sock *tp = tcp_sk(sk);
0143     struct bictcp *ca = inet_csk_ca(sk);
0144
0145     if (!tcp_is_cwnd_limited(sk))
0146         return;
0147
0148     if (tcp_in_slow_start(tp)) {
0149         acked = tcp_slow_start(tp, acked);
0150         if (!acked)
0151             return;
0152     }
0153     bictcp_update(ca, tcp_snd_cwnd(tp));
0154     tcp_cong_avoid_ai(tp, ca->cnt, acked);
0155 }
0156
0157 /*
0158  *  behave like Reno until low_window is reached,
0159  *  then increase congestion window slowly
0160  */
0161 static u32 bictcp_recalc_ssthresh(struct sock *sk)
0162 {
0163     const struct tcp_sock *tp = tcp_sk(sk);
0164     struct bictcp *ca = inet_csk_ca(sk);
0165
0166     ca->epoch_start = 0;    /* end of epoch */
0167
0168     /* Wmax and fast convergence */
0169     if (tcp_snd_cwnd(tp) < ca->last_max_cwnd && fast_convergence)
0170         ca->last_max_cwnd = (tcp_snd_cwnd(tp) * (BICTCP_BETA_SCALE + beta))
0171             / (2 * BICTCP_BETA_SCALE);
0172     else
0173         ca->last_max_cwnd = tcp_snd_cwnd(tp);
0174
0175     if (tcp_snd_cwnd(tp) <= low_window)
0176         return max(tcp_snd_cwnd(tp) >> 1U, 2U);
0177     else
0178         return max((tcp_snd_cwnd(tp) * beta) / BICTCP_BETA_SCALE, 2U);
0179 }
0180
0181 static void bictcp_state(struct sock *sk, u8 new_state)
0182 {
0183     if (new_state == TCP_CA_Loss)
0184         bictcp_reset(inet_csk_ca(sk));
0185 }
0186
0187 /* Track delayed acknowledgment ratio using sliding window
0188  * ratio = (15*ratio + sample) / 16
0189  */
0190 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
0191 {
0192     const struct inet_connection_sock *icsk = inet_csk(sk);
0193
0194     if (icsk->icsk_ca_state == TCP_CA_Open) {
0195         struct bictcp *ca = inet_csk_ca(sk);
0196
0197         ca->delayed_ack += sample->pkts_acked -
0198             (ca->delayed_ack >> ACK_RATIO_SHIFT);
0199     }
0200 }
0201
0202 static struct tcp_congestion_ops bictcp __read_mostly = {
0203     .init       = bictcp_init,
0204     .ssthresh   = bictcp_recalc_ssthresh,
0205     .cong_avoid = bictcp_cong_avoid,
0206     .set_state  = bictcp_state,
0207     .undo_cwnd  = tcp_reno_undo_cwnd,
0208     .pkts_acked     = bictcp_acked,
0209     .owner      = THIS_MODULE,
0210     .name       = "bic",
0211 };
0212
0213 static int __init bictcp_register(void)
0214 {
0215     BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
0216     return tcp_register_congestion_control(&bictcp);
0217 }
0218
0219 static void __exit bictcp_unregister(void)
0220 {
0221     tcp_unregister_congestion_control(&bictcp);
0222 }
0223
0224 module_init(bictcp_register);
0225 module_exit(bictcp_unregister);
0226
0227 MODULE_AUTHOR("Stephen Hemminger");
0228 MODULE_LICENSE("GPL");
0229 MODULE_DESCRIPTION("BIC TCP");