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/btf.h>
40 #include <linux/btf_ids.h>
41 #include <linux/module.h>
42 #include <linux/mm.h>
43 #include <net/tcp.h>
44 #include <linux/inet_diag.h>
45 #include "tcp_dctcp.h"
46
47 #define DCTCP_MAX_ALPHA 1024U
48
49 struct dctcp {
50 u32 old_delivered;
51 u32 old_delivered_ce;
52 u32 prior_rcv_nxt;
53 u32 dctcp_alpha;
54 u32 next_seq;
55 u32 ce_state;
56 u32 loss_cwnd;
57 };
58
59 static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
60 module_param(dctcp_shift_g, uint, 0644);
61 MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
62
63 static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
64 module_param(dctcp_alpha_on_init, uint, 0644);
65 MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
66
67 static struct tcp_congestion_ops dctcp_reno;
68
dctcp_reset(const struct tcp_sock * tp,struct dctcp * ca)69 static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
70 {
71 ca->next_seq = tp->snd_nxt;
72
73 ca->old_delivered = tp->delivered;
74 ca->old_delivered_ce = tp->delivered_ce;
75 }
76
dctcp_init(struct sock * sk)77 static void dctcp_init(struct sock *sk)
78 {
79 const struct tcp_sock *tp = tcp_sk(sk);
80
81 if ((tp->ecn_flags & TCP_ECN_OK) ||
82 (sk->sk_state == TCP_LISTEN ||
83 sk->sk_state == TCP_CLOSE)) {
84 struct dctcp *ca = inet_csk_ca(sk);
85
86 ca->prior_rcv_nxt = tp->rcv_nxt;
87
88 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
89
90 ca->loss_cwnd = 0;
91 ca->ce_state = 0;
92
93 dctcp_reset(tp, ca);
94 return;
95 }
96
97 /* No ECN support? Fall back to Reno. Also need to clear
98 * ECT from sk since it is set during 3WHS for DCTCP.
99 */
100 inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
101 INET_ECN_dontxmit(sk);
102 }
103
dctcp_ssthresh(struct sock * sk)104 static u32 dctcp_ssthresh(struct sock *sk)
105 {
106 struct dctcp *ca = inet_csk_ca(sk);
107 struct tcp_sock *tp = tcp_sk(sk);
108
109 ca->loss_cwnd = tcp_snd_cwnd(tp);
110 return max(tcp_snd_cwnd(tp) - ((tcp_snd_cwnd(tp) * ca->dctcp_alpha) >> 11U), 2U);
111 }
112
dctcp_update_alpha(struct sock * sk,u32 flags)113 static void dctcp_update_alpha(struct sock *sk, u32 flags)
114 {
115 const struct tcp_sock *tp = tcp_sk(sk);
116 struct dctcp *ca = inet_csk_ca(sk);
117
118 /* Expired RTT */
119 if (!before(tp->snd_una, ca->next_seq)) {
120 u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
121 u32 alpha = ca->dctcp_alpha;
122
123 /* alpha = (1 - g) * alpha + g * F */
124
125 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
126 if (delivered_ce) {
127 u32 delivered = tp->delivered - ca->old_delivered;
128
129 /* If dctcp_shift_g == 1, a 32bit value would overflow
130 * after 8 M packets.
131 */
132 delivered_ce <<= (10 - dctcp_shift_g);
133 delivered_ce /= max(1U, delivered);
134
135 alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
136 }
137 /* dctcp_alpha can be read from dctcp_get_info() without
138 * synchro, so we ask compiler to not use dctcp_alpha
139 * as a temporary variable in prior operations.
140 */
141 WRITE_ONCE(ca->dctcp_alpha, alpha);
142 dctcp_reset(tp, ca);
143 }
144 }
145
dctcp_react_to_loss(struct sock * sk)146 static void dctcp_react_to_loss(struct sock *sk)
147 {
148 struct dctcp *ca = inet_csk_ca(sk);
149 struct tcp_sock *tp = tcp_sk(sk);
150
151 ca->loss_cwnd = tcp_snd_cwnd(tp);
152 tp->snd_ssthresh = max(tcp_snd_cwnd(tp) >> 1U, 2U);
153 }
154
dctcp_state(struct sock * sk,u8 new_state)155 static void dctcp_state(struct sock *sk, u8 new_state)
156 {
157 if (new_state == TCP_CA_Recovery &&
158 new_state != inet_csk(sk)->icsk_ca_state)
159 dctcp_react_to_loss(sk);
160 /* We handle RTO in dctcp_cwnd_event to ensure that we perform only
161 * one loss-adjustment per RTT.
162 */
163 }
164
dctcp_cwnd_event(struct sock * sk,enum tcp_ca_event ev)165 static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
166 {
167 struct dctcp *ca = inet_csk_ca(sk);
168
169 switch (ev) {
170 case CA_EVENT_ECN_IS_CE:
171 case CA_EVENT_ECN_NO_CE:
172 dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
173 break;
174 case CA_EVENT_LOSS:
175 dctcp_react_to_loss(sk);
176 break;
177 default:
178 /* Don't care for the rest. */
179 break;
180 }
181 }
182
dctcp_get_info(struct sock * sk,u32 ext,int * attr,union tcp_cc_info * info)183 static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
184 union tcp_cc_info *info)
185 {
186 const struct dctcp *ca = inet_csk_ca(sk);
187 const struct tcp_sock *tp = tcp_sk(sk);
188
189 /* Fill it also in case of VEGASINFO due to req struct limits.
190 * We can still correctly retrieve it later.
191 */
192 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
193 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
194 memset(&info->dctcp, 0, sizeof(info->dctcp));
195 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
196 info->dctcp.dctcp_enabled = 1;
197 info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
198 info->dctcp.dctcp_alpha = ca->dctcp_alpha;
199 info->dctcp.dctcp_ab_ecn = tp->mss_cache *
200 (tp->delivered_ce - ca->old_delivered_ce);
201 info->dctcp.dctcp_ab_tot = tp->mss_cache *
202 (tp->delivered - ca->old_delivered);
203 }
204
205 *attr = INET_DIAG_DCTCPINFO;
206 return sizeof(info->dctcp);
207 }
208 return 0;
209 }
210
dctcp_cwnd_undo(struct sock * sk)211 static u32 dctcp_cwnd_undo(struct sock *sk)
212 {
213 const struct dctcp *ca = inet_csk_ca(sk);
214 struct tcp_sock *tp = tcp_sk(sk);
215
216 return max(tcp_snd_cwnd(tp), ca->loss_cwnd);
217 }
218
219 static struct tcp_congestion_ops dctcp __read_mostly = {
220 .init = dctcp_init,
221 .in_ack_event = dctcp_update_alpha,
222 .cwnd_event = dctcp_cwnd_event,
223 .ssthresh = dctcp_ssthresh,
224 .cong_avoid = tcp_reno_cong_avoid,
225 .undo_cwnd = dctcp_cwnd_undo,
226 .set_state = dctcp_state,
227 .get_info = dctcp_get_info,
228 .flags = TCP_CONG_NEEDS_ECN,
229 .owner = THIS_MODULE,
230 .name = "dctcp",
231 };
232
233 static struct tcp_congestion_ops dctcp_reno __read_mostly = {
234 .ssthresh = tcp_reno_ssthresh,
235 .cong_avoid = tcp_reno_cong_avoid,
236 .undo_cwnd = tcp_reno_undo_cwnd,
237 .get_info = dctcp_get_info,
238 .owner = THIS_MODULE,
239 .name = "dctcp-reno",
240 };
241
242 BTF_SET8_START(tcp_dctcp_check_kfunc_ids)
243 #ifdef CONFIG_X86
244 #ifdef CONFIG_DYNAMIC_FTRACE
245 BTF_ID_FLAGS(func, dctcp_init)
246 BTF_ID_FLAGS(func, dctcp_update_alpha)
247 BTF_ID_FLAGS(func, dctcp_cwnd_event)
248 BTF_ID_FLAGS(func, dctcp_ssthresh)
249 BTF_ID_FLAGS(func, dctcp_cwnd_undo)
250 BTF_ID_FLAGS(func, dctcp_state)
251 #endif
252 #endif
253 BTF_SET8_END(tcp_dctcp_check_kfunc_ids)
254
255 static const struct btf_kfunc_id_set tcp_dctcp_kfunc_set = {
256 .owner = THIS_MODULE,
257 .set = &tcp_dctcp_check_kfunc_ids,
258 };
259
dctcp_register(void)260 static int __init dctcp_register(void)
261 {
262 int ret;
263
264 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
265
266 ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_dctcp_kfunc_set);
267 if (ret < 0)
268 return ret;
269 return tcp_register_congestion_control(&dctcp);
270 }
271
dctcp_unregister(void)272 static void __exit dctcp_unregister(void)
273 {
274 tcp_unregister_congestion_control(&dctcp);
275 }
276
277 module_init(dctcp_register);
278 module_exit(dctcp_unregister);
279
280 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
281 MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
282 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
283
284 MODULE_LICENSE("GPL v2");
285 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
286
