Loading...
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
47struct 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
57static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
58module_param(dctcp_shift_g, uint, 0644);
59MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
60
61static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
62module_param(dctcp_alpha_on_init, uint, 0644);
63MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
64
65static struct tcp_congestion_ops dctcp_reno;
66
67static 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
75static 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
102static 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
111static 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
144static 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
153static 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
163static 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
181static 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
209static 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
216static 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
230static 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
239static 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
245static void __exit dctcp_unregister(void)
246{
247 tcp_unregister_congestion_control(&dctcp);
248}
249
250module_init(dctcp_register);
251module_exit(dctcp_unregister);
252
253MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
254MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
255MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
256
257MODULE_LICENSE("GPL v2");
258MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
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 delayed_ack_reserved;
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->delayed_ack_reserved = 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 const struct dctcp *ca = inet_csk_ca(sk);
115 struct tcp_sock *tp = tcp_sk(sk);
116
117 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
118}
119
120/* Minimal DCTP CE state machine:
121 *
122 * S: 0 <- last pkt was non-CE
123 * 1 <- last pkt was CE
124 */
125
126static void dctcp_ce_state_0_to_1(struct sock *sk)
127{
128 struct dctcp *ca = inet_csk_ca(sk);
129 struct tcp_sock *tp = tcp_sk(sk);
130
131 /* State has changed from CE=0 to CE=1 and delayed
132 * ACK has not sent yet.
133 */
134 if (!ca->ce_state && ca->delayed_ack_reserved) {
135 u32 tmp_rcv_nxt;
136
137 /* Save current rcv_nxt. */
138 tmp_rcv_nxt = tp->rcv_nxt;
139
140 /* Generate previous ack with CE=0. */
141 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
142 tp->rcv_nxt = ca->prior_rcv_nxt;
143
144 tcp_send_ack(sk);
145
146 /* Recover current rcv_nxt. */
147 tp->rcv_nxt = tmp_rcv_nxt;
148 }
149
150 ca->prior_rcv_nxt = tp->rcv_nxt;
151 ca->ce_state = 1;
152
153 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
154}
155
156static void dctcp_ce_state_1_to_0(struct sock *sk)
157{
158 struct dctcp *ca = inet_csk_ca(sk);
159 struct tcp_sock *tp = tcp_sk(sk);
160
161 /* State has changed from CE=1 to CE=0 and delayed
162 * ACK has not sent yet.
163 */
164 if (ca->ce_state && ca->delayed_ack_reserved) {
165 u32 tmp_rcv_nxt;
166
167 /* Save current rcv_nxt. */
168 tmp_rcv_nxt = tp->rcv_nxt;
169
170 /* Generate previous ack with CE=1. */
171 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
172 tp->rcv_nxt = ca->prior_rcv_nxt;
173
174 tcp_send_ack(sk);
175
176 /* Recover current rcv_nxt. */
177 tp->rcv_nxt = tmp_rcv_nxt;
178 }
179
180 ca->prior_rcv_nxt = tp->rcv_nxt;
181 ca->ce_state = 0;
182
183 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
184}
185
186static void dctcp_update_alpha(struct sock *sk, u32 flags)
187{
188 const struct tcp_sock *tp = tcp_sk(sk);
189 struct dctcp *ca = inet_csk_ca(sk);
190 u32 acked_bytes = tp->snd_una - ca->prior_snd_una;
191
192 /* If ack did not advance snd_una, count dupack as MSS size.
193 * If ack did update window, do not count it at all.
194 */
195 if (acked_bytes == 0 && !(flags & CA_ACK_WIN_UPDATE))
196 acked_bytes = inet_csk(sk)->icsk_ack.rcv_mss;
197 if (acked_bytes) {
198 ca->acked_bytes_total += acked_bytes;
199 ca->prior_snd_una = tp->snd_una;
200
201 if (flags & CA_ACK_ECE)
202 ca->acked_bytes_ecn += acked_bytes;
203 }
204
205 /* Expired RTT */
206 if (!before(tp->snd_una, ca->next_seq)) {
207 u64 bytes_ecn = ca->acked_bytes_ecn;
208 u32 alpha = ca->dctcp_alpha;
209
210 /* alpha = (1 - g) * alpha + g * F */
211
212 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
213 if (bytes_ecn) {
214 /* If dctcp_shift_g == 1, a 32bit value would overflow
215 * after 8 Mbytes.
216 */
217 bytes_ecn <<= (10 - dctcp_shift_g);
218 do_div(bytes_ecn, max(1U, ca->acked_bytes_total));
219
220 alpha = min(alpha + (u32)bytes_ecn, DCTCP_MAX_ALPHA);
221 }
222 /* dctcp_alpha can be read from dctcp_get_info() without
223 * synchro, so we ask compiler to not use dctcp_alpha
224 * as a temporary variable in prior operations.
225 */
226 WRITE_ONCE(ca->dctcp_alpha, alpha);
227 dctcp_reset(tp, ca);
228 }
229}
230
231static void dctcp_state(struct sock *sk, u8 new_state)
232{
233 if (dctcp_clamp_alpha_on_loss && new_state == TCP_CA_Loss) {
234 struct dctcp *ca = inet_csk_ca(sk);
235
236 /* If this extension is enabled, we clamp dctcp_alpha to
237 * max on packet loss; the motivation is that dctcp_alpha
238 * is an indicator to the extend of congestion and packet
239 * loss is an indicator of extreme congestion; setting
240 * this in practice turned out to be beneficial, and
241 * effectively assumes total congestion which reduces the
242 * window by half.
243 */
244 ca->dctcp_alpha = DCTCP_MAX_ALPHA;
245 }
246}
247
248static void dctcp_update_ack_reserved(struct sock *sk, enum tcp_ca_event ev)
249{
250 struct dctcp *ca = inet_csk_ca(sk);
251
252 switch (ev) {
253 case CA_EVENT_DELAYED_ACK:
254 if (!ca->delayed_ack_reserved)
255 ca->delayed_ack_reserved = 1;
256 break;
257 case CA_EVENT_NON_DELAYED_ACK:
258 if (ca->delayed_ack_reserved)
259 ca->delayed_ack_reserved = 0;
260 break;
261 default:
262 /* Don't care for the rest. */
263 break;
264 }
265}
266
267static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
268{
269 switch (ev) {
270 case CA_EVENT_ECN_IS_CE:
271 dctcp_ce_state_0_to_1(sk);
272 break;
273 case CA_EVENT_ECN_NO_CE:
274 dctcp_ce_state_1_to_0(sk);
275 break;
276 case CA_EVENT_DELAYED_ACK:
277 case CA_EVENT_NON_DELAYED_ACK:
278 dctcp_update_ack_reserved(sk, ev);
279 break;
280 default:
281 /* Don't care for the rest. */
282 break;
283 }
284}
285
286static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
287 union tcp_cc_info *info)
288{
289 const struct dctcp *ca = inet_csk_ca(sk);
290
291 /* Fill it also in case of VEGASINFO due to req struct limits.
292 * We can still correctly retrieve it later.
293 */
294 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
295 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
296 memset(info, 0, sizeof(struct tcp_dctcp_info));
297 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
298 info->dctcp.dctcp_enabled = 1;
299 info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
300 info->dctcp.dctcp_alpha = ca->dctcp_alpha;
301 info->dctcp.dctcp_ab_ecn = ca->acked_bytes_ecn;
302 info->dctcp.dctcp_ab_tot = ca->acked_bytes_total;
303 }
304
305 *attr = INET_DIAG_DCTCPINFO;
306 return sizeof(*info);
307 }
308 return 0;
309}
310
311static struct tcp_congestion_ops dctcp __read_mostly = {
312 .init = dctcp_init,
313 .in_ack_event = dctcp_update_alpha,
314 .cwnd_event = dctcp_cwnd_event,
315 .ssthresh = dctcp_ssthresh,
316 .cong_avoid = tcp_reno_cong_avoid,
317 .set_state = dctcp_state,
318 .get_info = dctcp_get_info,
319 .flags = TCP_CONG_NEEDS_ECN,
320 .owner = THIS_MODULE,
321 .name = "dctcp",
322};
323
324static struct tcp_congestion_ops dctcp_reno __read_mostly = {
325 .ssthresh = tcp_reno_ssthresh,
326 .cong_avoid = tcp_reno_cong_avoid,
327 .get_info = dctcp_get_info,
328 .owner = THIS_MODULE,
329 .name = "dctcp-reno",
330};
331
332static int __init dctcp_register(void)
333{
334 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
335 return tcp_register_congestion_control(&dctcp);
336}
337
338static void __exit dctcp_unregister(void)
339{
340 tcp_unregister_congestion_control(&dctcp);
341}
342
343module_init(dctcp_register);
344module_exit(dctcp_unregister);
345
346MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
347MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
348MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
349
350MODULE_LICENSE("GPL v2");
351MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");