1 // SPDX-License-Identifier: GPL-2.0-or-later 1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* DataCenter TCP (DCTCP) congestion control. 2 /* DataCenter TCP (DCTCP) congestion control.
3 * 3 *
4 * http://simula.stanford.edu/~alizade/Site/DC 4 * http://simula.stanford.edu/~alizade/Site/DCTCP.html
5 * 5 *
6 * This is an implementation of DCTCP over Ren 6 * This is an implementation of DCTCP over Reno, an enhancement to the
7 * TCP congestion control algorithm designed f 7 * TCP congestion control algorithm designed for data centers. DCTCP
8 * leverages Explicit Congestion Notification 8 * leverages Explicit Congestion Notification (ECN) in the network to
9 * provide multi-bit feedback to the end hosts 9 * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet
10 * the following three data center transport r 10 * the following three data center transport requirements:
11 * 11 *
12 * - High burst tolerance (incast due to part 12 * - High burst tolerance (incast due to partition/aggregate)
13 * - Low latency (short flows, queries) 13 * - Low latency (short flows, queries)
14 * - High throughput (continuous data updates 14 * - High throughput (continuous data updates, large file transfers)
15 * with commodity shallow buffered switches 15 * with commodity shallow buffered switches
16 * 16 *
17 * The algorithm is described in detail in the 17 * The algorithm is described in detail in the following two papers:
18 * 18 *
19 * 1) Mohammad Alizadeh, Albert Greenberg, Dav 19 * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye,
20 * Parveen Patel, Balaji Prabhakar, Sudipta 20 * Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan:
21 * "Data Center TCP (DCTCP)", Data Center 21 * "Data Center TCP (DCTCP)", Data Center Networks session
22 * Proc. ACM SIGCOMM, New Delhi, 2010. 22 * Proc. ACM SIGCOMM, New Delhi, 2010.
23 * http://simula.stanford.edu/~alizade/Site/ 23 * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
24 * 24 *
25 * 2) Mohammad Alizadeh, Adel Javanmard, and B 25 * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar:
26 * "Analysis of DCTCP: Stability, Converg 26 * "Analysis of DCTCP: Stability, Convergence, and Fairness"
27 * Proc. ACM SIGMETRICS, San Jose, 2011. 27 * Proc. ACM SIGMETRICS, San Jose, 2011.
28 * http://simula.stanford.edu/~alizade/Site/ 28 * http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
29 * 29 *
30 * Initial prototype from Abdul Kabbani, Masat 30 * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh.
31 * 31 *
32 * Authors: 32 * Authors:
33 * 33 *
34 * Daniel Borkmann <dborkman@redhat.com> 34 * Daniel Borkmann <dborkman@redhat.com>
35 * Florian Westphal <fw@strlen.de> 35 * Florian Westphal <fw@strlen.de>
36 * Glenn Judd <glenn.judd@morganstanley.c 36 * Glenn Judd <glenn.judd@morganstanley.com>
37 */ 37 */
38 38
39 #include <linux/btf.h> <<
40 #include <linux/btf_ids.h> <<
41 #include <linux/module.h> 39 #include <linux/module.h>
42 #include <linux/mm.h> 40 #include <linux/mm.h>
43 #include <net/tcp.h> 41 #include <net/tcp.h>
44 #include <linux/inet_diag.h> 42 #include <linux/inet_diag.h>
45 #include "tcp_dctcp.h" 43 #include "tcp_dctcp.h"
46 44
47 #define DCTCP_MAX_ALPHA 1024U 45 #define DCTCP_MAX_ALPHA 1024U
48 46
49 struct dctcp { 47 struct dctcp {
50 u32 old_delivered; 48 u32 old_delivered;
51 u32 old_delivered_ce; 49 u32 old_delivered_ce;
52 u32 prior_rcv_nxt; 50 u32 prior_rcv_nxt;
53 u32 dctcp_alpha; 51 u32 dctcp_alpha;
54 u32 next_seq; 52 u32 next_seq;
55 u32 ce_state; 53 u32 ce_state;
56 u32 loss_cwnd; 54 u32 loss_cwnd;
57 struct tcp_plb_state plb; <<
58 }; 55 };
59 56
60 static unsigned int dctcp_shift_g __read_mostl 57 static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
61 !! 58 module_param(dctcp_shift_g, uint, 0644);
62 static int dctcp_shift_g_set(const char *val, <<
63 { <<
64 return param_set_uint_minmax(val, kp, <<
65 } <<
66 <<
67 static const struct kernel_param_ops dctcp_shi <<
68 .set = dctcp_shift_g_set, <<
69 .get = param_get_uint, <<
70 }; <<
71 <<
72 module_param_cb(dctcp_shift_g, &dctcp_shift_g_ <<
73 MODULE_PARM_DESC(dctcp_shift_g, "parameter g f 59 MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
74 60
75 static unsigned int dctcp_alpha_on_init __read 61 static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
76 module_param(dctcp_alpha_on_init, uint, 0644); 62 module_param(dctcp_alpha_on_init, uint, 0644);
77 MODULE_PARM_DESC(dctcp_alpha_on_init, "paramet 63 MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
78 64
79 static struct tcp_congestion_ops dctcp_reno; 65 static struct tcp_congestion_ops dctcp_reno;
80 66
81 static void dctcp_reset(const struct tcp_sock 67 static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
82 { 68 {
83 ca->next_seq = tp->snd_nxt; 69 ca->next_seq = tp->snd_nxt;
84 70
85 ca->old_delivered = tp->delivered; 71 ca->old_delivered = tp->delivered;
86 ca->old_delivered_ce = tp->delivered_c 72 ca->old_delivered_ce = tp->delivered_ce;
87 } 73 }
88 74
89 __bpf_kfunc static void dctcp_init(struct sock !! 75 static void dctcp_init(struct sock *sk)
90 { 76 {
91 const struct tcp_sock *tp = tcp_sk(sk) 77 const struct tcp_sock *tp = tcp_sk(sk);
92 78
93 if ((tp->ecn_flags & TCP_ECN_OK) || 79 if ((tp->ecn_flags & TCP_ECN_OK) ||
94 (sk->sk_state == TCP_LISTEN || 80 (sk->sk_state == TCP_LISTEN ||
95 sk->sk_state == TCP_CLOSE)) { 81 sk->sk_state == TCP_CLOSE)) {
96 struct dctcp *ca = inet_csk_ca 82 struct dctcp *ca = inet_csk_ca(sk);
97 83
98 ca->prior_rcv_nxt = tp->rcv_nx 84 ca->prior_rcv_nxt = tp->rcv_nxt;
99 85
100 ca->dctcp_alpha = min(dctcp_al 86 ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
101 87
102 ca->loss_cwnd = 0; 88 ca->loss_cwnd = 0;
103 ca->ce_state = 0; 89 ca->ce_state = 0;
104 90
105 dctcp_reset(tp, ca); 91 dctcp_reset(tp, ca);
106 tcp_plb_init(sk, &ca->plb); <<
107 <<
108 return; 92 return;
109 } 93 }
110 94
111 /* No ECN support? Fall back to Reno. 95 /* No ECN support? Fall back to Reno. Also need to clear
112 * ECT from sk since it is set during 96 * ECT from sk since it is set during 3WHS for DCTCP.
113 */ 97 */
114 inet_csk(sk)->icsk_ca_ops = &dctcp_ren 98 inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
115 INET_ECN_dontxmit(sk); 99 INET_ECN_dontxmit(sk);
116 } 100 }
117 101
118 __bpf_kfunc static u32 dctcp_ssthresh(struct s !! 102 static u32 dctcp_ssthresh(struct sock *sk)
119 { 103 {
120 struct dctcp *ca = inet_csk_ca(sk); 104 struct dctcp *ca = inet_csk_ca(sk);
121 struct tcp_sock *tp = tcp_sk(sk); 105 struct tcp_sock *tp = tcp_sk(sk);
122 106
123 ca->loss_cwnd = tcp_snd_cwnd(tp); !! 107 ca->loss_cwnd = tp->snd_cwnd;
124 return max(tcp_snd_cwnd(tp) - ((tcp_sn !! 108 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U);
125 } 109 }
126 110
127 __bpf_kfunc static void dctcp_update_alpha(str !! 111 static void dctcp_update_alpha(struct sock *sk, u32 flags)
128 { 112 {
129 const struct tcp_sock *tp = tcp_sk(sk) 113 const struct tcp_sock *tp = tcp_sk(sk);
130 struct dctcp *ca = inet_csk_ca(sk); 114 struct dctcp *ca = inet_csk_ca(sk);
131 115
132 /* Expired RTT */ 116 /* Expired RTT */
133 if (!before(tp->snd_una, ca->next_seq) 117 if (!before(tp->snd_una, ca->next_seq)) {
134 u32 delivered = tp->delivered <<
135 u32 delivered_ce = tp->deliver 118 u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
136 u32 alpha = ca->dctcp_alpha; 119 u32 alpha = ca->dctcp_alpha;
137 u32 ce_ratio = 0; <<
138 <<
139 if (delivered > 0) { <<
140 /* dctcp_alpha keeps E <<
141 * packets. Because of <<
142 * be slow so we use c <<
143 * measure of congesti <<
144 * ECN marked packets <<
145 */ <<
146 if (delivered_ce > 0) <<
147 ce_ratio = (de <<
148 tcp_plb_update_state(s <<
149 tcp_plb_check_rehash(s <<
150 } <<
151 120
152 /* alpha = (1 - g) * alpha + g 121 /* alpha = (1 - g) * alpha + g * F */
153 122
154 alpha -= min_not_zero(alpha, a 123 alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
155 if (delivered_ce) { 124 if (delivered_ce) {
>> 125 u32 delivered = tp->delivered - ca->old_delivered;
156 126
157 /* If dctcp_shift_g == 127 /* If dctcp_shift_g == 1, a 32bit value would overflow
158 * after 8 M packets. 128 * after 8 M packets.
159 */ 129 */
160 delivered_ce <<= (10 - 130 delivered_ce <<= (10 - dctcp_shift_g);
161 delivered_ce /= max(1U 131 delivered_ce /= max(1U, delivered);
162 132
163 alpha = min(alpha + de 133 alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
164 } 134 }
165 /* dctcp_alpha can be read fro 135 /* dctcp_alpha can be read from dctcp_get_info() without
166 * synchro, so we ask compiler 136 * synchro, so we ask compiler to not use dctcp_alpha
167 * as a temporary variable in 137 * as a temporary variable in prior operations.
168 */ 138 */
169 WRITE_ONCE(ca->dctcp_alpha, al 139 WRITE_ONCE(ca->dctcp_alpha, alpha);
170 dctcp_reset(tp, ca); 140 dctcp_reset(tp, ca);
171 } 141 }
172 } 142 }
173 143
174 static void dctcp_react_to_loss(struct sock *s 144 static void dctcp_react_to_loss(struct sock *sk)
175 { 145 {
176 struct dctcp *ca = inet_csk_ca(sk); 146 struct dctcp *ca = inet_csk_ca(sk);
177 struct tcp_sock *tp = tcp_sk(sk); 147 struct tcp_sock *tp = tcp_sk(sk);
178 148
179 ca->loss_cwnd = tcp_snd_cwnd(tp); !! 149 ca->loss_cwnd = tp->snd_cwnd;
180 tp->snd_ssthresh = max(tcp_snd_cwnd(tp !! 150 tp->snd_ssthresh = max(tp->snd_cwnd >> 1U, 2U);
181 } 151 }
182 152
183 __bpf_kfunc static void dctcp_state(struct soc !! 153 static void dctcp_state(struct sock *sk, u8 new_state)
184 { 154 {
185 if (new_state == TCP_CA_Recovery && 155 if (new_state == TCP_CA_Recovery &&
186 new_state != inet_csk(sk)->icsk_ca 156 new_state != inet_csk(sk)->icsk_ca_state)
187 dctcp_react_to_loss(sk); 157 dctcp_react_to_loss(sk);
188 /* We handle RTO in dctcp_cwnd_event t 158 /* We handle RTO in dctcp_cwnd_event to ensure that we perform only
189 * one loss-adjustment per RTT. 159 * one loss-adjustment per RTT.
190 */ 160 */
191 } 161 }
192 162
193 __bpf_kfunc static void dctcp_cwnd_event(struc !! 163 static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
194 { 164 {
195 struct dctcp *ca = inet_csk_ca(sk); 165 struct dctcp *ca = inet_csk_ca(sk);
196 166
197 switch (ev) { 167 switch (ev) {
198 case CA_EVENT_ECN_IS_CE: 168 case CA_EVENT_ECN_IS_CE:
199 case CA_EVENT_ECN_NO_CE: 169 case CA_EVENT_ECN_NO_CE:
200 dctcp_ece_ack_update(sk, ev, & 170 dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
201 break; 171 break;
202 case CA_EVENT_LOSS: 172 case CA_EVENT_LOSS:
203 tcp_plb_update_state_upon_rto( <<
204 dctcp_react_to_loss(sk); 173 dctcp_react_to_loss(sk);
205 break; 174 break;
206 case CA_EVENT_TX_START: <<
207 tcp_plb_check_rehash(sk, &ca-> <<
208 break; <<
209 default: 175 default:
210 /* Don't care for the rest. */ 176 /* Don't care for the rest. */
211 break; 177 break;
212 } 178 }
213 } 179 }
214 180
215 static size_t dctcp_get_info(struct sock *sk, 181 static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
216 union tcp_cc_info 182 union tcp_cc_info *info)
217 { 183 {
218 const struct dctcp *ca = inet_csk_ca(s 184 const struct dctcp *ca = inet_csk_ca(sk);
219 const struct tcp_sock *tp = tcp_sk(sk) 185 const struct tcp_sock *tp = tcp_sk(sk);
220 186
221 /* Fill it also in case of VEGASINFO d 187 /* Fill it also in case of VEGASINFO due to req struct limits.
222 * We can still correctly retrieve it 188 * We can still correctly retrieve it later.
223 */ 189 */
224 if (ext & (1 << (INET_DIAG_DCTCPINFO - 190 if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
225 ext & (1 << (INET_DIAG_VEGASINFO - 191 ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
226 memset(&info->dctcp, 0, sizeof 192 memset(&info->dctcp, 0, sizeof(info->dctcp));
227 if (inet_csk(sk)->icsk_ca_ops 193 if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
228 info->dctcp.dctcp_enab 194 info->dctcp.dctcp_enabled = 1;
229 info->dctcp.dctcp_ce_s 195 info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
230 info->dctcp.dctcp_alph 196 info->dctcp.dctcp_alpha = ca->dctcp_alpha;
231 info->dctcp.dctcp_ab_e 197 info->dctcp.dctcp_ab_ecn = tp->mss_cache *
232 198 (tp->delivered_ce - ca->old_delivered_ce);
233 info->dctcp.dctcp_ab_t 199 info->dctcp.dctcp_ab_tot = tp->mss_cache *
234 200 (tp->delivered - ca->old_delivered);
235 } 201 }
236 202
237 *attr = INET_DIAG_DCTCPINFO; 203 *attr = INET_DIAG_DCTCPINFO;
238 return sizeof(info->dctcp); 204 return sizeof(info->dctcp);
239 } 205 }
240 return 0; 206 return 0;
241 } 207 }
242 208
243 __bpf_kfunc static u32 dctcp_cwnd_undo(struct !! 209 static u32 dctcp_cwnd_undo(struct sock *sk)
244 { 210 {
245 const struct dctcp *ca = inet_csk_ca(s 211 const struct dctcp *ca = inet_csk_ca(sk);
246 struct tcp_sock *tp = tcp_sk(sk); <<
247 212
248 return max(tcp_snd_cwnd(tp), ca->loss_ !! 213 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
249 } 214 }
250 215
251 static struct tcp_congestion_ops dctcp __read_ 216 static struct tcp_congestion_ops dctcp __read_mostly = {
252 .init = dctcp_init, 217 .init = dctcp_init,
253 .in_ack_event = dctcp_update_alpha, 218 .in_ack_event = dctcp_update_alpha,
254 .cwnd_event = dctcp_cwnd_event, 219 .cwnd_event = dctcp_cwnd_event,
255 .ssthresh = dctcp_ssthresh, 220 .ssthresh = dctcp_ssthresh,
256 .cong_avoid = tcp_reno_cong_avoid, 221 .cong_avoid = tcp_reno_cong_avoid,
257 .undo_cwnd = dctcp_cwnd_undo, 222 .undo_cwnd = dctcp_cwnd_undo,
258 .set_state = dctcp_state, 223 .set_state = dctcp_state,
259 .get_info = dctcp_get_info, 224 .get_info = dctcp_get_info,
260 .flags = TCP_CONG_NEEDS_ECN, 225 .flags = TCP_CONG_NEEDS_ECN,
261 .owner = THIS_MODULE, 226 .owner = THIS_MODULE,
262 .name = "dctcp", 227 .name = "dctcp",
263 }; 228 };
264 229
265 static struct tcp_congestion_ops dctcp_reno __ 230 static struct tcp_congestion_ops dctcp_reno __read_mostly = {
266 .ssthresh = tcp_reno_ssthresh, 231 .ssthresh = tcp_reno_ssthresh,
267 .cong_avoid = tcp_reno_cong_avoid, 232 .cong_avoid = tcp_reno_cong_avoid,
268 .undo_cwnd = tcp_reno_undo_cwnd, 233 .undo_cwnd = tcp_reno_undo_cwnd,
269 .get_info = dctcp_get_info, 234 .get_info = dctcp_get_info,
270 .owner = THIS_MODULE, 235 .owner = THIS_MODULE,
271 .name = "dctcp-reno", 236 .name = "dctcp-reno",
272 }; 237 };
273 238
274 BTF_KFUNCS_START(tcp_dctcp_check_kfunc_ids) <<
275 BTF_ID_FLAGS(func, dctcp_init) <<
276 BTF_ID_FLAGS(func, dctcp_update_alpha) <<
277 BTF_ID_FLAGS(func, dctcp_cwnd_event) <<
278 BTF_ID_FLAGS(func, dctcp_ssthresh) <<
279 BTF_ID_FLAGS(func, dctcp_cwnd_undo) <<
280 BTF_ID_FLAGS(func, dctcp_state) <<
281 BTF_KFUNCS_END(tcp_dctcp_check_kfunc_ids) <<
282 <<
283 static const struct btf_kfunc_id_set tcp_dctcp <<
284 .owner = THIS_MODULE, <<
285 .set = &tcp_dctcp_check_kfunc_ids, <<
286 }; <<
287 <<
288 static int __init dctcp_register(void) 239 static int __init dctcp_register(void)
289 { 240 {
290 int ret; <<
291 <<
292 BUILD_BUG_ON(sizeof(struct dctcp) > IC 241 BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
293 <<
294 ret = register_btf_kfunc_id_set(BPF_PR <<
295 if (ret < 0) <<
296 return ret; <<
297 return tcp_register_congestion_control 242 return tcp_register_congestion_control(&dctcp);
298 } 243 }
299 244
300 static void __exit dctcp_unregister(void) 245 static void __exit dctcp_unregister(void)
301 { 246 {
302 tcp_unregister_congestion_control(&dct 247 tcp_unregister_congestion_control(&dctcp);
303 } 248 }
304 249
305 module_init(dctcp_register); 250 module_init(dctcp_register);
306 module_exit(dctcp_unregister); 251 module_exit(dctcp_unregister);
307 252
308 MODULE_AUTHOR("Daniel Borkmann <dborkman@redha 253 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
309 MODULE_AUTHOR("Florian Westphal <fw@strlen.de> 254 MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
310 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganst 255 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
311 256
312 MODULE_LICENSE("GPL v2"); 257 MODULE_LICENSE("GPL v2");
313 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)"); 258 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");
314 259
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.