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
51struct dctcp {
52 u32 acked_bytes_ecn;
53 u32 acked_bytes_total;
54 u32 prior_snd_una;
55 u32 prior_rcv_nxt;
56 u32 dctcp_alpha;
57 u32 next_seq;
58 u32 ce_state;
59 u32 loss_cwnd;
60};
61
62static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
63module_param(dctcp_shift_g, uint, 0644);
64MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
65
66static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
67module_param(dctcp_alpha_on_init, uint, 0644);
68MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
69
70static unsigned int dctcp_clamp_alpha_on_loss __read_mostly;
71module_param(dctcp_clamp_alpha_on_loss, uint, 0644);
72MODULE_PARM_DESC(dctcp_clamp_alpha_on_loss,
73 "parameter for clamping alpha on loss");
74
75static struct tcp_congestion_ops dctcp_reno;
76
77static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
78{
79 ca->next_seq = tp->snd_nxt;
80
81 ca->acked_bytes_ecn = 0;
82 ca->acked_bytes_total = 0;
83}
84
85static void dctcp_init(struct sock *sk)
86{
87 const struct tcp_sock *tp = tcp_sk(sk);
88
89 if ((tp->ecn_flags & TCP_ECN_OK) ||
90 (sk->sk_state == TCP_LISTEN ||
91 sk->sk_state == TCP_CLOSE)) {
92 struct dctcp *ca = inet_csk_ca(sk);
93
94 ca->prior_snd_una = tp->snd_una;
95 ca->prior_rcv_nxt = tp->rcv_nxt;
96
97 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
98
99 ca->loss_cwnd = 0;
100 ca->ce_state = 0;
101
102 dctcp_reset(tp, ca);
103 return;
104 }
105
106 /* No ECN support? Fall back to Reno. Also need to clear
107 * ECT from sk since it is set during 3WHS for DCTCP.
108 */
109 inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
110 INET_ECN_dontxmit(sk);
111}
112
113static u32 dctcp_ssthresh(struct sock *sk)
114{
115 struct dctcp *ca = inet_csk_ca(sk);
116 struct tcp_sock *tp = tcp_sk(sk);
117
118 ca->loss_cwnd = tp->snd_cwnd;
119 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
120}
121
122static void dctcp_update_alpha(struct sock *sk, u32 flags)
123{
124 const struct tcp_sock *tp = tcp_sk(sk);
125 struct dctcp *ca = inet_csk_ca(sk);
126 u32 acked_bytes = tp->snd_una - ca->prior_snd_una;
127
128 /* If ack did not advance snd_una, count dupack as MSS size.
129 * If ack did update window, do not count it at all.
130 */
131 if (acked_bytes == 0 && !(flags & CA_ACK_WIN_UPDATE))
132 acked_bytes = inet_csk(sk)->icsk_ack.rcv_mss;
133 if (acked_bytes) {
134 ca->acked_bytes_total += acked_bytes;
135 ca->prior_snd_una = tp->snd_una;
136
137 if (flags & CA_ACK_ECE)
138 ca->acked_bytes_ecn += acked_bytes;
139 }
140
141 /* Expired RTT */
142 if (!before(tp->snd_una, ca->next_seq)) {
143 u64 bytes_ecn = ca->acked_bytes_ecn;
144 u32 alpha = ca->dctcp_alpha;
145
146 /* alpha = (1 - g) * alpha + g * F */
147
148 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
149 if (bytes_ecn) {
150 /* If dctcp_shift_g == 1, a 32bit value would overflow
151 * after 8 Mbytes.
152 */
153 bytes_ecn <<= (10 - dctcp_shift_g);
154 do_div(bytes_ecn, max(1U, ca->acked_bytes_total));
155
156 alpha = min(alpha + (u32)bytes_ecn, DCTCP_MAX_ALPHA);
157 }
158 /* dctcp_alpha can be read from dctcp_get_info() without
159 * synchro, so we ask compiler to not use dctcp_alpha
160 * as a temporary variable in prior operations.
161 */
162 WRITE_ONCE(ca->dctcp_alpha, alpha);
163 dctcp_reset(tp, ca);
164 }
165}
166
167static void dctcp_state(struct sock *sk, u8 new_state)
168{
169 if (dctcp_clamp_alpha_on_loss && new_state == TCP_CA_Loss) {
170 struct dctcp *ca = inet_csk_ca(sk);
171
172 /* If this extension is enabled, we clamp dctcp_alpha to
173 * max on packet loss; the motivation is that dctcp_alpha
174 * is an indicator to the extend of congestion and packet
175 * loss is an indicator of extreme congestion; setting
176 * this in practice turned out to be beneficial, and
177 * effectively assumes total congestion which reduces the
178 * window by half.
179 */
180 ca->dctcp_alpha = DCTCP_MAX_ALPHA;
181 }
182}
183
184static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
185{
186 struct dctcp *ca = inet_csk_ca(sk);
187
188 switch (ev) {
189 case CA_EVENT_ECN_IS_CE:
190 case CA_EVENT_ECN_NO_CE:
191 dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
192 break;
193 default:
194 /* Don't care for the rest. */
195 break;
196 }
197}
198
199static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
200 union tcp_cc_info *info)
201{
202 const struct dctcp *ca = inet_csk_ca(sk);
203
204 /* Fill it also in case of VEGASINFO due to req struct limits.
205 * We can still correctly retrieve it later.
206 */
207 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
208 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
209 memset(&info->dctcp, 0, sizeof(info->dctcp));
210 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
211 info->dctcp.dctcp_enabled = 1;
212 info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
213 info->dctcp.dctcp_alpha = ca->dctcp_alpha;
214 info->dctcp.dctcp_ab_ecn = ca->acked_bytes_ecn;
215 info->dctcp.dctcp_ab_tot = ca->acked_bytes_total;
216 }
217
218 *attr = INET_DIAG_DCTCPINFO;
219 return sizeof(info->dctcp);
220 }
221 return 0;
222}
223
224static u32 dctcp_cwnd_undo(struct sock *sk)
225{
226 const struct dctcp *ca = inet_csk_ca(sk);
227
228 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
229}
230
231static struct tcp_congestion_ops dctcp __read_mostly = {
232 .init = dctcp_init,
233 .in_ack_event = dctcp_update_alpha,
234 .cwnd_event = dctcp_cwnd_event,
235 .ssthresh = dctcp_ssthresh,
236 .cong_avoid = tcp_reno_cong_avoid,
237 .undo_cwnd = dctcp_cwnd_undo,
238 .set_state = dctcp_state,
239 .get_info = dctcp_get_info,
240 .flags = TCP_CONG_NEEDS_ECN,
241 .owner = THIS_MODULE,
242 .name = "dctcp",
243};
244
245static struct tcp_congestion_ops dctcp_reno __read_mostly = {
246 .ssthresh = tcp_reno_ssthresh,
247 .cong_avoid = tcp_reno_cong_avoid,
248 .undo_cwnd = tcp_reno_undo_cwnd,
249 .get_info = dctcp_get_info,
250 .owner = THIS_MODULE,
251 .name = "dctcp-reno",
252};
253
254static int __init dctcp_register(void)
255{
256 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
257 return tcp_register_congestion_control(&dctcp);
258}
259
260static void __exit dctcp_unregister(void)
261{
262 tcp_unregister_congestion_control(&dctcp);
263}
264
265module_init(dctcp_register);
266module_exit(dctcp_unregister);
267
268MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
269MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
270MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
271
272MODULE_LICENSE("GPL v2");
273MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
274