net/ipv4/tcp_dctcp.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* DataCenter TCP (DCTCP) congestion control.
   3  *
   4  * http://simula.stanford.edu/~alizade/Site/DCTCP.html
   5  *
   6  * This is an implementation of DCTCP over Reno, an enhancement to the
   7  * TCP congestion control algorithm designed for data centers. DCTCP
   8  * leverages Explicit Congestion Notification (ECN) in the network to
   9  * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet
  10  * the following three data center transport requirements:
  11  *
  12  *  - High burst tolerance (incast due to partition/aggregate)
  13  *  - Low latency (short flows, queries)
  14  *  - High throughput (continuous data updates, large file transfers)
  15  *    with commodity shallow buffered switches
  16  *
  17  * The algorithm is described in detail in the following two papers:
  18  *
  19  * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye,
  20  *    Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan:
  21  *      "Data Center TCP (DCTCP)", Data Center Networks session
  22  *      Proc. ACM SIGCOMM, New Delhi, 2010.
  23  *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
  24  *
  25  * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar:
  26  *      "Analysis of DCTCP: Stability, Convergence, and Fairness"
  27  *      Proc. ACM SIGMETRICS, San Jose, 2011.
  28  *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
  29  *
  30  * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh.
  31  *
  32  * Authors:
  33  *
  34  *      Daniel Borkmann <dborkman@redhat.com>
  35  *      Florian Westphal <fw@strlen.de>
  36  *      Glenn Judd <glenn.judd@morganstanley.com>
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/mm.h>
  41 #include <net/tcp.h>
  42 #include <linux/inet_diag.h>
  43 #include "tcp_dctcp.h"
  44
  45 #define DCTCP_MAX_ALPHA 1024U
  46
  47 struct dctcp {
  48         u32 old_delivered;
  49         u32 old_delivered_ce;
  50         u32 prior_rcv_nxt;
  51         u32 dctcp_alpha;
  52         u32 next_seq;
  53         u32 ce_state;
  54         u32 loss_cwnd;
  55 };
  56
  57 static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
  58 module_param(dctcp_shift_g, uint, 0644);
  59 MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
  60
  61 static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
  62 module_param(dctcp_alpha_on_init, uint, 0644);
  63 MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
  64
  65 static struct tcp_congestion_ops dctcp_reno;
  66
  67 static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
  68 {
  69         ca->next_seq = tp->snd_nxt;
  70
  71         ca->old_delivered = tp->delivered;
  72         ca->old_delivered_ce = tp->delivered_ce;
  73 }
  74
  75 static void dctcp_init(struct sock *sk)
  76 {
  77         const struct tcp_sock *tp = tcp_sk(sk);
  78
  79         if ((tp->ecn_flags & TCP_ECN_OK) ||
  80             (sk->sk_state == TCP_LISTEN ||
  81              sk->sk_state == TCP_CLOSE)) {
  82                 struct dctcp *ca = inet_csk_ca(sk);
  83
  84                 ca->prior_rcv_nxt = tp->rcv_nxt;
  85
  86                 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
  87
  88                 ca->loss_cwnd = 0;
  89                 ca->ce_state = 0;
  90
  91                 dctcp_reset(tp, ca);
  92                 return;
  93         }
  94
  95         /* No ECN support? Fall back to Reno. Also need to clear
  96          * ECT from sk since it is set during 3WHS for DCTCP.
  97          */
  98         inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
  99         INET_ECN_dontxmit(sk);
 100 }
 101
 102 static u32 dctcp_ssthresh(struct sock *sk)
 103 {
 104         struct dctcp *ca = inet_csk_ca(sk);
 105         struct tcp_sock *tp = tcp_sk(sk);
 106
 107         ca->loss_cwnd = tp->snd_cwnd;
 108         return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
 109 }
 110
 111 static void dctcp_update_alpha(struct sock *sk, u32 flags)
 112 {
 113         const struct tcp_sock *tp = tcp_sk(sk);
 114         struct dctcp *ca = inet_csk_ca(sk);
 115
 116         /* Expired RTT */
 117         if (!before(tp->snd_una, ca->next_seq)) {
 118                 u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
 119                 u32 alpha = ca->dctcp_alpha;
 120
 121                 /* alpha = (1 - g) * alpha + g * F */
 122
 123                 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
 124                 if (delivered_ce) {
 125                         u32 delivered = tp->delivered - ca->old_delivered;
 126
 127                         /* If dctcp_shift_g == 1, a 32bit value would overflow
 128                          * after 8 M packets.
 129                          */
 130                         delivered_ce <<= (10 - dctcp_shift_g);
 131                         delivered_ce /= max(1U, delivered);
 132
 133                         alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
 134                 }
 135                 /* dctcp_alpha can be read from dctcp_get_info() without
 136                  * synchro, so we ask compiler to not use dctcp_alpha
 137                  * as a temporary variable in prior operations.
 138                  */
 139                 WRITE_ONCE(ca->dctcp_alpha, alpha);
 140                 dctcp_reset(tp, ca);
 141         }
 142 }
 143
 144 static void dctcp_react_to_loss(struct sock *sk)
 145 {
 146         struct dctcp *ca = inet_csk_ca(sk);
 147         struct tcp_sock *tp = tcp_sk(sk);
 148
 149         ca->loss_cwnd = tp->snd_cwnd;
 150         tp->snd_ssthresh = max(tp->snd_cwnd >> 1U, 2U);
 151 }
 152
 153 static void dctcp_state(struct sock *sk, u8 new_state)
 154 {
 155         if (new_state == TCP_CA_Recovery &&
 156             new_state != inet_csk(sk)->icsk_ca_state)
 157                 dctcp_react_to_loss(sk);
 158         /* We handle RTO in dctcp_cwnd_event to ensure that we perform only
 159          * one loss-adjustment per RTT.
 160          */
 161 }
 162
 163 static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
 164 {
 165         struct dctcp *ca = inet_csk_ca(sk);
 166
 167         switch (ev) {
 168         case CA_EVENT_ECN_IS_CE:
 169         case CA_EVENT_ECN_NO_CE:
 170                 dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
 171                 break;
 172         case CA_EVENT_LOSS:
 173                 dctcp_react_to_loss(sk);
 174                 break;
 175         default:
 176                 /* Don't care for the rest. */
 177                 break;
 178         }
 179 }
 180
 181 static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
 182                              union tcp_cc_info *info)
 183 {
 184         const struct dctcp *ca = inet_csk_ca(sk);
 185         const struct tcp_sock *tp = tcp_sk(sk);
 186
 187         /* Fill it also in case of VEGASINFO due to req struct limits.
 188          * We can still correctly retrieve it later.
 189          */
 190         if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
 191             ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 192                 memset(&info->dctcp, 0, sizeof(info->dctcp));
 193                 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
 194                         info->dctcp.dctcp_enabled = 1;
 195                         info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
 196                         info->dctcp.dctcp_alpha = ca->dctcp_alpha;
 197                         info->dctcp.dctcp_ab_ecn = tp->mss_cache *
 198                                                    (tp->delivered_ce - ca->old_delivered_ce);
 199                         info->dctcp.dctcp_ab_tot = tp->mss_cache *
 200                                                    (tp->delivered - ca->old_delivered);
 201                 }
 202
 203                 *attr = INET_DIAG_DCTCPINFO;
 204                 return sizeof(info->dctcp);
 205         }
 206         return 0;
 207 }
 208
 209 static u32 dctcp_cwnd_undo(struct sock *sk)
 210 {
 211         const struct dctcp *ca = inet_csk_ca(sk);
 212
 213         return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
 214 }
 215
 216 static struct tcp_congestion_ops dctcp __read_mostly = {
 217         .init           = dctcp_init,
 218         .in_ack_event   = dctcp_update_alpha,
 219         .cwnd_event     = dctcp_cwnd_event,
 220         .ssthresh       = dctcp_ssthresh,
 221         .cong_avoid     = tcp_reno_cong_avoid,
 222         .undo_cwnd      = dctcp_cwnd_undo,
 223         .set_state      = dctcp_state,
 224         .get_info       = dctcp_get_info,
 225         .flags          = TCP_CONG_NEEDS_ECN,
 226         .owner          = THIS_MODULE,
 227         .name           = "dctcp",
 228 };
 229
 230 static struct tcp_congestion_ops dctcp_reno __read_mostly = {
 231         .ssthresh       = tcp_reno_ssthresh,
 232         .cong_avoid     = tcp_reno_cong_avoid,
 233         .undo_cwnd      = tcp_reno_undo_cwnd,
 234         .get_info       = dctcp_get_info,
 235         .owner          = THIS_MODULE,
 236         .name           = "dctcp-reno",
 237 };
 238
 239 static int __init dctcp_register(void)
 240 {
 241         BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
 242         return tcp_register_congestion_control(&dctcp);
 243 }
 244
 245 static void __exit dctcp_unregister(void)
 246 {
 247         tcp_unregister_congestion_control(&dctcp);
 248 }
 249
 250 module_init(dctcp_register);
 251 module_exit(dctcp_unregister);
 252
 253 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
 254 MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
 255 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
 256
 257 MODULE_LICENSE("GPL v2");
 258 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");