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
48#define DCTCP_MAX_ALPHA 1024U
49
50struct dctcp {
51 u32 acked_bytes_ecn;
52 u32 acked_bytes_total;
53 u32 prior_snd_una;
54 u32 prior_rcv_nxt;
55 u32 dctcp_alpha;
56 u32 next_seq;
57 u32 ce_state;
58 u32 loss_cwnd;
59};
60
61static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
62module_param(dctcp_shift_g, uint, 0644);
63MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
64
65static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
66module_param(dctcp_alpha_on_init, uint, 0644);
67MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
68
69static unsigned int dctcp_clamp_alpha_on_loss __read_mostly;
70module_param(dctcp_clamp_alpha_on_loss, uint, 0644);
71MODULE_PARM_DESC(dctcp_clamp_alpha_on_loss,
72 "parameter for clamping alpha on loss");
73
74static struct tcp_congestion_ops dctcp_reno;
75
76static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
77{
78 ca->next_seq = tp->snd_nxt;
79
80 ca->acked_bytes_ecn = 0;
81 ca->acked_bytes_total = 0;
82}
83
84static void dctcp_init(struct sock *sk)
85{
86 const struct tcp_sock *tp = tcp_sk(sk);
87
88 if ((tp->ecn_flags & TCP_ECN_OK) ||
89 (sk->sk_state == TCP_LISTEN ||
90 sk->sk_state == TCP_CLOSE)) {
91 struct dctcp *ca = inet_csk_ca(sk);
92
93 ca->prior_snd_una = tp->snd_una;
94 ca->prior_rcv_nxt = tp->rcv_nxt;
95
96 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
97
98 ca->loss_cwnd = 0;
99 ca->ce_state = 0;
100
101 dctcp_reset(tp, ca);
102 return;
103 }
104
105 /* No ECN support? Fall back to Reno. Also need to clear
106 * ECT from sk since it is set during 3WHS for DCTCP.
107 */
108 inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
109 INET_ECN_dontxmit(sk);
110}
111
112static u32 dctcp_ssthresh(struct sock *sk)
113{
114 struct dctcp *ca = inet_csk_ca(sk);
115 struct tcp_sock *tp = tcp_sk(sk);
116
117 ca->loss_cwnd = tp->snd_cwnd;
118 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
119}
120
121/* Minimal DCTP CE state machine:
122 *
123 * S: 0 <- last pkt was non-CE
124 * 1 <- last pkt was CE
125 */
126
127static void dctcp_ce_state_0_to_1(struct sock *sk)
128{
129 struct dctcp *ca = inet_csk_ca(sk);
130 struct tcp_sock *tp = tcp_sk(sk);
131
132 if (!ca->ce_state) {
133 /* State has changed from CE=0 to CE=1, force an immediate
134 * ACK to reflect the new CE state. If an ACK was delayed,
135 * send that first to reflect the prior CE state.
136 */
137 if (inet_csk(sk)->icsk_ack.pending & ICSK_ACK_TIMER)
138 __tcp_send_ack(sk, ca->prior_rcv_nxt);
139 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
140 }
141
142 ca->prior_rcv_nxt = tp->rcv_nxt;
143 ca->ce_state = 1;
144
145 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
146}
147
148static void dctcp_ce_state_1_to_0(struct sock *sk)
149{
150 struct dctcp *ca = inet_csk_ca(sk);
151 struct tcp_sock *tp = tcp_sk(sk);
152
153 if (ca->ce_state) {
154 /* State has changed from CE=1 to CE=0, force an immediate
155 * ACK to reflect the new CE state. If an ACK was delayed,
156 * send that first to reflect the prior CE state.
157 */
158 if (inet_csk(sk)->icsk_ack.pending & ICSK_ACK_TIMER)
159 __tcp_send_ack(sk, ca->prior_rcv_nxt);
160 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
161 }
162
163 ca->prior_rcv_nxt = tp->rcv_nxt;
164 ca->ce_state = 0;
165
166 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
167}
168
169static void dctcp_update_alpha(struct sock *sk, u32 flags)
170{
171 const struct tcp_sock *tp = tcp_sk(sk);
172 struct dctcp *ca = inet_csk_ca(sk);
173 u32 acked_bytes = tp->snd_una - ca->prior_snd_una;
174
175 /* If ack did not advance snd_una, count dupack as MSS size.
176 * If ack did update window, do not count it at all.
177 */
178 if (acked_bytes == 0 && !(flags & CA_ACK_WIN_UPDATE))
179 acked_bytes = inet_csk(sk)->icsk_ack.rcv_mss;
180 if (acked_bytes) {
181 ca->acked_bytes_total += acked_bytes;
182 ca->prior_snd_una = tp->snd_una;
183
184 if (flags & CA_ACK_ECE)
185 ca->acked_bytes_ecn += acked_bytes;
186 }
187
188 /* Expired RTT */
189 if (!before(tp->snd_una, ca->next_seq)) {
190 u64 bytes_ecn = ca->acked_bytes_ecn;
191 u32 alpha = ca->dctcp_alpha;
192
193 /* alpha = (1 - g) * alpha + g * F */
194
195 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
196 if (bytes_ecn) {
197 /* If dctcp_shift_g == 1, a 32bit value would overflow
198 * after 8 Mbytes.
199 */
200 bytes_ecn <<= (10 - dctcp_shift_g);
201 do_div(bytes_ecn, max(1U, ca->acked_bytes_total));
202
203 alpha = min(alpha + (u32)bytes_ecn, DCTCP_MAX_ALPHA);
204 }
205 /* dctcp_alpha can be read from dctcp_get_info() without
206 * synchro, so we ask compiler to not use dctcp_alpha
207 * as a temporary variable in prior operations.
208 */
209 WRITE_ONCE(ca->dctcp_alpha, alpha);
210 dctcp_reset(tp, ca);
211 }
212}
213
214static void dctcp_state(struct sock *sk, u8 new_state)
215{
216 if (dctcp_clamp_alpha_on_loss && new_state == TCP_CA_Loss) {
217 struct dctcp *ca = inet_csk_ca(sk);
218
219 /* If this extension is enabled, we clamp dctcp_alpha to
220 * max on packet loss; the motivation is that dctcp_alpha
221 * is an indicator to the extend of congestion and packet
222 * loss is an indicator of extreme congestion; setting
223 * this in practice turned out to be beneficial, and
224 * effectively assumes total congestion which reduces the
225 * window by half.
226 */
227 ca->dctcp_alpha = DCTCP_MAX_ALPHA;
228 }
229}
230
231static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
232{
233 switch (ev) {
234 case CA_EVENT_ECN_IS_CE:
235 dctcp_ce_state_0_to_1(sk);
236 break;
237 case CA_EVENT_ECN_NO_CE:
238 dctcp_ce_state_1_to_0(sk);
239 break;
240 default:
241 /* Don't care for the rest. */
242 break;
243 }
244}
245
246static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
247 union tcp_cc_info *info)
248{
249 const struct dctcp *ca = inet_csk_ca(sk);
250
251 /* Fill it also in case of VEGASINFO due to req struct limits.
252 * We can still correctly retrieve it later.
253 */
254 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
255 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
256 memset(&info->dctcp, 0, sizeof(info->dctcp));
257 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
258 info->dctcp.dctcp_enabled = 1;
259 info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
260 info->dctcp.dctcp_alpha = ca->dctcp_alpha;
261 info->dctcp.dctcp_ab_ecn = ca->acked_bytes_ecn;
262 info->dctcp.dctcp_ab_tot = ca->acked_bytes_total;
263 }
264
265 *attr = INET_DIAG_DCTCPINFO;
266 return sizeof(info->dctcp);
267 }
268 return 0;
269}
270
271static u32 dctcp_cwnd_undo(struct sock *sk)
272{
273 const struct dctcp *ca = inet_csk_ca(sk);
274
275 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
276}
277
278static struct tcp_congestion_ops dctcp __read_mostly = {
279 .init = dctcp_init,
280 .in_ack_event = dctcp_update_alpha,
281 .cwnd_event = dctcp_cwnd_event,
282 .ssthresh = dctcp_ssthresh,
283 .cong_avoid = tcp_reno_cong_avoid,
284 .undo_cwnd = dctcp_cwnd_undo,
285 .set_state = dctcp_state,
286 .get_info = dctcp_get_info,
287 .flags = TCP_CONG_NEEDS_ECN,
288 .owner = THIS_MODULE,
289 .name = "dctcp",
290};
291
292static struct tcp_congestion_ops dctcp_reno __read_mostly = {
293 .ssthresh = tcp_reno_ssthresh,
294 .cong_avoid = tcp_reno_cong_avoid,
295 .undo_cwnd = tcp_reno_undo_cwnd,
296 .get_info = dctcp_get_info,
297 .owner = THIS_MODULE,
298 .name = "dctcp-reno",
299};
300
301static int __init dctcp_register(void)
302{
303 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
304 return tcp_register_congestion_control(&dctcp);
305}
306
307static void __exit dctcp_unregister(void)
308{
309 tcp_unregister_congestion_control(&dctcp);
310}
311
312module_init(dctcp_register);
313module_exit(dctcp_unregister);
314
315MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
316MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
317MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
318
319MODULE_LICENSE("GPL v2");
320MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
321