1 /* DataCenter TCP (DCTCP) congestion control. 2 * 3 * http://simula.stanford.edu/~alizade/Site/DCTCP.html 4 * 5 * This is an implementation of DCTCP over Reno, an enhancement to the 6 * TCP congestion control algorithm designed for data centers. DCTCP 7 * leverages Explicit Congestion Notification (ECN) in the network to 8 * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet 9 * the following three data center transport requirements: 10 * 11 * - High burst tolerance (incast due to partition/aggregate) 12 * - Low latency (short flows, queries) 13 * - High throughput (continuous data updates, large file transfers) 14 * with commodity shallow buffered switches 15 * 16 * The algorithm is described in detail in the following two papers: 17 * 18 * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye, 19 * Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan: 20 * "Data Center TCP (DCTCP)", Data Center Networks session 21 * Proc. ACM SIGCOMM, New Delhi, 2010. 22 * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf 23 * 24 * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar: 25 * "Analysis of DCTCP: Stability, Convergence, and Fairness" 26 * Proc. ACM SIGMETRICS, San Jose, 2011. 27 * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf 28 * 29 * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh. 30 * 31 * Authors: 32 * 33 * Daniel Borkmann <dborkman@redhat.com> 34 * Florian Westphal <fw@strlen.de> 35 * Glenn Judd <glenn.judd@morganstanley.com> 36 * 37 * This program is free software; you can redistribute it and/or modify 38 * it under the terms of the GNU General Public License as published by 39 * the Free Software Foundation; either version 2 of the License, or (at 40 * your option) any later version. 41 */ 42 43 #include <linux/module.h> 44 #include <linux/mm.h> 45 #include <net/tcp.h> 46 #include <linux/inet_diag.h> 47 #include "tcp_dctcp.h" 48 49 #define DCTCP_MAX_ALPHA 1024U 50 51 struct dctcp { 52 u32 old_delivered; 53 u32 old_delivered_ce; 54 u32 prior_rcv_nxt; 55 u32 dctcp_alpha; 56 u32 next_seq; 57 u32 ce_state; 58 u32 loss_cwnd; 59 }; 60 61 static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */ 62 module_param(dctcp_shift_g, uint, 0644); 63 MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha"); 64 65 static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA; 66 module_param(dctcp_alpha_on_init, uint, 0644); 67 MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value"); 68 69 static struct tcp_congestion_ops dctcp_reno; 70 71 static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca) 72 { 73 ca->next_seq = tp->snd_nxt; 74 75 ca->old_delivered = tp->delivered; 76 ca->old_delivered_ce = tp->delivered_ce; 77 } 78 79 static void dctcp_init(struct sock *sk) 80 { 81 const struct tcp_sock *tp = tcp_sk(sk); 82 83 if ((tp->ecn_flags & TCP_ECN_OK) || 84 (sk->sk_state == TCP_LISTEN || 85 sk->sk_state == TCP_CLOSE)) { 86 struct dctcp *ca = inet_csk_ca(sk); 87 88 ca->prior_rcv_nxt = tp->rcv_nxt; 89 90 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA); 91 92 ca->loss_cwnd = 0; 93 ca->ce_state = 0; 94 95 dctcp_reset(tp, ca); 96 return; 97 } 98 99 /* No ECN support? Fall back to Reno. Also need to clear 100 * ECT from sk since it is set during 3WHS for DCTCP. 101 */ 102 inet_csk(sk)->icsk_ca_ops = &dctcp_reno; 103 INET_ECN_dontxmit(sk); 104 } 105 106 static u32 dctcp_ssthresh(struct sock *sk) 107 { 108 struct dctcp *ca = inet_csk_ca(sk); 109 struct tcp_sock *tp = tcp_sk(sk); 110 111 ca->loss_cwnd = tp->snd_cwnd; 112 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U); 113 } 114 115 static void dctcp_update_alpha(struct sock *sk, u32 flags) 116 { 117 const struct tcp_sock *tp = tcp_sk(sk); 118 struct dctcp *ca = inet_csk_ca(sk); 119 120 /* Expired RTT */ 121 if (!before(tp->snd_una, ca->next_seq)) { 122 u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce; 123 u32 alpha = ca->dctcp_alpha; 124 125 /* alpha = (1 - g) * alpha + g * F */ 126 127 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g); 128 if (delivered_ce) { 129 u32 delivered = tp->delivered - ca->old_delivered; 130 131 /* If dctcp_shift_g == 1, a 32bit value would overflow 132 * after 8 M packets. 133 */ 134 delivered_ce <<= (10 - dctcp_shift_g); 135 delivered_ce /= max(1U, delivered); 136 137 alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA); 138 } 139 /* dctcp_alpha can be read from dctcp_get_info() without 140 * synchro, so we ask compiler to not use dctcp_alpha 141 * as a temporary variable in prior operations. 142 */ 143 WRITE_ONCE(ca->dctcp_alpha, alpha); 144 dctcp_reset(tp, ca); 145 } 146 } 147 148 static void dctcp_react_to_loss(struct sock *sk) 149 { 150 struct dctcp *ca = inet_csk_ca(sk); 151 struct tcp_sock *tp = tcp_sk(sk); 152 153 ca->loss_cwnd = tp->snd_cwnd; 154 tp->snd_ssthresh = max(tp->snd_cwnd >> 1U, 2U); 155 } 156 157 static void dctcp_state(struct sock *sk, u8 new_state) 158 { 159 if (new_state == TCP_CA_Recovery && 160 new_state != inet_csk(sk)->icsk_ca_state) 161 dctcp_react_to_loss(sk); 162 /* We handle RTO in dctcp_cwnd_event to ensure that we perform only 163 * one loss-adjustment per RTT. 164 */ 165 } 166 167 static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev) 168 { 169 struct dctcp *ca = inet_csk_ca(sk); 170 171 switch (ev) { 172 case CA_EVENT_ECN_IS_CE: 173 case CA_EVENT_ECN_NO_CE: 174 dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state); 175 break; 176 case CA_EVENT_LOSS: 177 dctcp_react_to_loss(sk); 178 break; 179 default: 180 /* Don't care for the rest. */ 181 break; 182 } 183 } 184 185 static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr, 186 union tcp_cc_info *info) 187 { 188 const struct dctcp *ca = inet_csk_ca(sk); 189 const struct tcp_sock *tp = tcp_sk(sk); 190 191 /* Fill it also in case of VEGASINFO due to req struct limits. 192 * We can still correctly retrieve it later. 193 */ 194 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) || 195 ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 196 memset(&info->dctcp, 0, sizeof(info->dctcp)); 197 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) { 198 info->dctcp.dctcp_enabled = 1; 199 info->dctcp.dctcp_ce_state = (u16) ca->ce_state; 200 info->dctcp.dctcp_alpha = ca->dctcp_alpha; 201 info->dctcp.dctcp_ab_ecn = tp->mss_cache * 202 (tp->delivered_ce - ca->old_delivered_ce); 203 info->dctcp.dctcp_ab_tot = tp->mss_cache * 204 (tp->delivered - ca->old_delivered); 205 } 206 207 *attr = INET_DIAG_DCTCPINFO; 208 return sizeof(info->dctcp); 209 } 210 return 0; 211 } 212 213 static u32 dctcp_cwnd_undo(struct sock *sk) 214 { 215 const struct dctcp *ca = inet_csk_ca(sk); 216 217 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd); 218 } 219 220 static struct tcp_congestion_ops dctcp __read_mostly = { 221 .init = dctcp_init, 222 .in_ack_event = dctcp_update_alpha, 223 .cwnd_event = dctcp_cwnd_event, 224 .ssthresh = dctcp_ssthresh, 225 .cong_avoid = tcp_reno_cong_avoid, 226 .undo_cwnd = dctcp_cwnd_undo, 227 .set_state = dctcp_state, 228 .get_info = dctcp_get_info, 229 .flags = TCP_CONG_NEEDS_ECN, 230 .owner = THIS_MODULE, 231 .name = "dctcp", 232 }; 233 234 static struct tcp_congestion_ops dctcp_reno __read_mostly = { 235 .ssthresh = tcp_reno_ssthresh, 236 .cong_avoid = tcp_reno_cong_avoid, 237 .undo_cwnd = tcp_reno_undo_cwnd, 238 .get_info = dctcp_get_info, 239 .owner = THIS_MODULE, 240 .name = "dctcp-reno", 241 }; 242 243 static int __init dctcp_register(void) 244 { 245 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE); 246 return tcp_register_congestion_control(&dctcp); 247 } 248 249 static void __exit dctcp_unregister(void) 250 { 251 tcp_unregister_congestion_control(&dctcp); 252 } 253 254 module_init(dctcp_register); 255 module_exit(dctcp_unregister); 256 257 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>"); 258 MODULE_AUTHOR("Florian Westphal <fw@strlen.de>"); 259 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>"); 260 261 MODULE_LICENSE("GPL v2"); 262 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)"); 263