]>
Commit | Line | Data |
---|---|---|
e3118e83 DB |
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 | ||
50 | struct 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; | |
ce6dd233 | 59 | u32 loss_cwnd; |
e3118e83 DB |
60 | }; |
61 | ||
62 | static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */ | |
63 | module_param(dctcp_shift_g, uint, 0644); | |
64 | MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha"); | |
65 | ||
66 | static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA; | |
67 | module_param(dctcp_alpha_on_init, uint, 0644); | |
68 | MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value"); | |
69 | ||
70 | static unsigned int dctcp_clamp_alpha_on_loss __read_mostly; | |
71 | module_param(dctcp_clamp_alpha_on_loss, uint, 0644); | |
72 | MODULE_PARM_DESC(dctcp_clamp_alpha_on_loss, | |
73 | "parameter for clamping alpha on loss"); | |
74 | ||
75 | static struct tcp_congestion_ops dctcp_reno; | |
76 | ||
77 | static 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 | ||
85 | static 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->delayed_ack_reserved = 0; | |
ce6dd233 | 100 | ca->loss_cwnd = 0; |
e3118e83 DB |
101 | ca->ce_state = 0; |
102 | ||
103 | dctcp_reset(tp, ca); | |
104 | return; | |
105 | } | |
106 | ||
107 | /* No ECN support? Fall back to Reno. Also need to clear | |
108 | * ECT from sk since it is set during 3WHS for DCTCP. | |
109 | */ | |
110 | inet_csk(sk)->icsk_ca_ops = &dctcp_reno; | |
111 | INET_ECN_dontxmit(sk); | |
112 | } | |
113 | ||
114 | static u32 dctcp_ssthresh(struct sock *sk) | |
115 | { | |
ce6dd233 | 116 | struct dctcp *ca = inet_csk_ca(sk); |
e3118e83 DB |
117 | struct tcp_sock *tp = tcp_sk(sk); |
118 | ||
ce6dd233 | 119 | ca->loss_cwnd = tp->snd_cwnd; |
e3118e83 DB |
120 | return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->dctcp_alpha) >> 11U), 2U); |
121 | } | |
122 | ||
123 | /* Minimal DCTP CE state machine: | |
124 | * | |
125 | * S: 0 <- last pkt was non-CE | |
126 | * 1 <- last pkt was CE | |
127 | */ | |
128 | ||
129 | static void dctcp_ce_state_0_to_1(struct sock *sk) | |
130 | { | |
131 | struct dctcp *ca = inet_csk_ca(sk); | |
132 | struct tcp_sock *tp = tcp_sk(sk); | |
133 | ||
134 | /* State has changed from CE=0 to CE=1 and delayed | |
135 | * ACK has not sent yet. | |
136 | */ | |
137 | if (!ca->ce_state && ca->delayed_ack_reserved) { | |
138 | u32 tmp_rcv_nxt; | |
139 | ||
140 | /* Save current rcv_nxt. */ | |
141 | tmp_rcv_nxt = tp->rcv_nxt; | |
142 | ||
143 | /* Generate previous ack with CE=0. */ | |
144 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
145 | tp->rcv_nxt = ca->prior_rcv_nxt; | |
146 | ||
147 | tcp_send_ack(sk); | |
148 | ||
149 | /* Recover current rcv_nxt. */ | |
150 | tp->rcv_nxt = tmp_rcv_nxt; | |
151 | } | |
152 | ||
153 | ca->prior_rcv_nxt = tp->rcv_nxt; | |
154 | ca->ce_state = 1; | |
155 | ||
156 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
157 | } | |
158 | ||
159 | static void dctcp_ce_state_1_to_0(struct sock *sk) | |
160 | { | |
161 | struct dctcp *ca = inet_csk_ca(sk); | |
162 | struct tcp_sock *tp = tcp_sk(sk); | |
163 | ||
164 | /* State has changed from CE=1 to CE=0 and delayed | |
165 | * ACK has not sent yet. | |
166 | */ | |
167 | if (ca->ce_state && ca->delayed_ack_reserved) { | |
168 | u32 tmp_rcv_nxt; | |
169 | ||
170 | /* Save current rcv_nxt. */ | |
171 | tmp_rcv_nxt = tp->rcv_nxt; | |
172 | ||
173 | /* Generate previous ack with CE=1. */ | |
174 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
175 | tp->rcv_nxt = ca->prior_rcv_nxt; | |
176 | ||
177 | tcp_send_ack(sk); | |
178 | ||
179 | /* Recover current rcv_nxt. */ | |
180 | tp->rcv_nxt = tmp_rcv_nxt; | |
181 | } | |
182 | ||
183 | ca->prior_rcv_nxt = tp->rcv_nxt; | |
184 | ca->ce_state = 0; | |
185 | ||
186 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
187 | } | |
188 | ||
189 | static void dctcp_update_alpha(struct sock *sk, u32 flags) | |
190 | { | |
343dfaa1 | 191 | const struct tcp_sock *tp = tcp_sk(sk); |
e3118e83 DB |
192 | struct dctcp *ca = inet_csk_ca(sk); |
193 | u32 acked_bytes = tp->snd_una - ca->prior_snd_una; | |
194 | ||
195 | /* If ack did not advance snd_una, count dupack as MSS size. | |
196 | * If ack did update window, do not count it at all. | |
197 | */ | |
198 | if (acked_bytes == 0 && !(flags & CA_ACK_WIN_UPDATE)) | |
199 | acked_bytes = inet_csk(sk)->icsk_ack.rcv_mss; | |
200 | if (acked_bytes) { | |
201 | ca->acked_bytes_total += acked_bytes; | |
202 | ca->prior_snd_una = tp->snd_una; | |
203 | ||
204 | if (flags & CA_ACK_ECE) | |
205 | ca->acked_bytes_ecn += acked_bytes; | |
206 | } | |
207 | ||
208 | /* Expired RTT */ | |
209 | if (!before(tp->snd_una, ca->next_seq)) { | |
f9c2ff22 ED |
210 | u64 bytes_ecn = ca->acked_bytes_ecn; |
211 | u32 alpha = ca->dctcp_alpha; | |
e3118e83 DB |
212 | |
213 | /* alpha = (1 - g) * alpha + g * F */ | |
e3118e83 | 214 | |
c80dbe04 | 215 | alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g); |
f9c2ff22 ED |
216 | if (bytes_ecn) { |
217 | /* If dctcp_shift_g == 1, a 32bit value would overflow | |
218 | * after 8 Mbytes. | |
219 | */ | |
220 | bytes_ecn <<= (10 - dctcp_shift_g); | |
221 | do_div(bytes_ecn, max(1U, ca->acked_bytes_total)); | |
222 | ||
223 | alpha = min(alpha + (u32)bytes_ecn, DCTCP_MAX_ALPHA); | |
224 | } | |
225 | /* dctcp_alpha can be read from dctcp_get_info() without | |
226 | * synchro, so we ask compiler to not use dctcp_alpha | |
227 | * as a temporary variable in prior operations. | |
228 | */ | |
229 | WRITE_ONCE(ca->dctcp_alpha, alpha); | |
e3118e83 DB |
230 | dctcp_reset(tp, ca); |
231 | } | |
232 | } | |
233 | ||
234 | static void dctcp_state(struct sock *sk, u8 new_state) | |
235 | { | |
236 | if (dctcp_clamp_alpha_on_loss && new_state == TCP_CA_Loss) { | |
237 | struct dctcp *ca = inet_csk_ca(sk); | |
238 | ||
239 | /* If this extension is enabled, we clamp dctcp_alpha to | |
240 | * max on packet loss; the motivation is that dctcp_alpha | |
241 | * is an indicator to the extend of congestion and packet | |
242 | * loss is an indicator of extreme congestion; setting | |
243 | * this in practice turned out to be beneficial, and | |
244 | * effectively assumes total congestion which reduces the | |
245 | * window by half. | |
246 | */ | |
247 | ca->dctcp_alpha = DCTCP_MAX_ALPHA; | |
248 | } | |
249 | } | |
250 | ||
251 | static void dctcp_update_ack_reserved(struct sock *sk, enum tcp_ca_event ev) | |
252 | { | |
253 | struct dctcp *ca = inet_csk_ca(sk); | |
254 | ||
255 | switch (ev) { | |
256 | case CA_EVENT_DELAYED_ACK: | |
257 | if (!ca->delayed_ack_reserved) | |
258 | ca->delayed_ack_reserved = 1; | |
259 | break; | |
260 | case CA_EVENT_NON_DELAYED_ACK: | |
261 | if (ca->delayed_ack_reserved) | |
262 | ca->delayed_ack_reserved = 0; | |
263 | break; | |
264 | default: | |
265 | /* Don't care for the rest. */ | |
266 | break; | |
267 | } | |
268 | } | |
269 | ||
270 | static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev) | |
271 | { | |
272 | switch (ev) { | |
273 | case CA_EVENT_ECN_IS_CE: | |
274 | dctcp_ce_state_0_to_1(sk); | |
275 | break; | |
276 | case CA_EVENT_ECN_NO_CE: | |
277 | dctcp_ce_state_1_to_0(sk); | |
278 | break; | |
279 | case CA_EVENT_DELAYED_ACK: | |
280 | case CA_EVENT_NON_DELAYED_ACK: | |
281 | dctcp_update_ack_reserved(sk, ev); | |
282 | break; | |
283 | default: | |
284 | /* Don't care for the rest. */ | |
285 | break; | |
286 | } | |
287 | } | |
288 | ||
64f40ff5 ED |
289 | static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr, |
290 | union tcp_cc_info *info) | |
e3118e83 DB |
291 | { |
292 | const struct dctcp *ca = inet_csk_ca(sk); | |
293 | ||
294 | /* Fill it also in case of VEGASINFO due to req struct limits. | |
295 | * We can still correctly retrieve it later. | |
296 | */ | |
297 | if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) || | |
298 | ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | |
dcf1158b | 299 | memset(&info->dctcp, 0, sizeof(info->dctcp)); |
e3118e83 | 300 | if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) { |
64f40ff5 ED |
301 | info->dctcp.dctcp_enabled = 1; |
302 | info->dctcp.dctcp_ce_state = (u16) ca->ce_state; | |
303 | info->dctcp.dctcp_alpha = ca->dctcp_alpha; | |
304 | info->dctcp.dctcp_ab_ecn = ca->acked_bytes_ecn; | |
305 | info->dctcp.dctcp_ab_tot = ca->acked_bytes_total; | |
e3118e83 DB |
306 | } |
307 | ||
64f40ff5 | 308 | *attr = INET_DIAG_DCTCPINFO; |
dcf1158b | 309 | return sizeof(info->dctcp); |
e3118e83 | 310 | } |
521f1cf1 | 311 | return 0; |
e3118e83 DB |
312 | } |
313 | ||
ce6dd233 FW |
314 | static u32 dctcp_cwnd_undo(struct sock *sk) |
315 | { | |
316 | const struct dctcp *ca = inet_csk_ca(sk); | |
317 | ||
318 | return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd); | |
319 | } | |
320 | ||
e3118e83 DB |
321 | static struct tcp_congestion_ops dctcp __read_mostly = { |
322 | .init = dctcp_init, | |
323 | .in_ack_event = dctcp_update_alpha, | |
324 | .cwnd_event = dctcp_cwnd_event, | |
325 | .ssthresh = dctcp_ssthresh, | |
326 | .cong_avoid = tcp_reno_cong_avoid, | |
ce6dd233 | 327 | .undo_cwnd = dctcp_cwnd_undo, |
e3118e83 DB |
328 | .set_state = dctcp_state, |
329 | .get_info = dctcp_get_info, | |
330 | .flags = TCP_CONG_NEEDS_ECN, | |
331 | .owner = THIS_MODULE, | |
332 | .name = "dctcp", | |
333 | }; | |
334 | ||
335 | static struct tcp_congestion_ops dctcp_reno __read_mostly = { | |
336 | .ssthresh = tcp_reno_ssthresh, | |
337 | .cong_avoid = tcp_reno_cong_avoid, | |
e9799183 | 338 | .undo_cwnd = tcp_reno_undo_cwnd, |
e3118e83 DB |
339 | .get_info = dctcp_get_info, |
340 | .owner = THIS_MODULE, | |
341 | .name = "dctcp-reno", | |
342 | }; | |
343 | ||
344 | static int __init dctcp_register(void) | |
345 | { | |
346 | BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE); | |
347 | return tcp_register_congestion_control(&dctcp); | |
348 | } | |
349 | ||
350 | static void __exit dctcp_unregister(void) | |
351 | { | |
352 | tcp_unregister_congestion_control(&dctcp); | |
353 | } | |
354 | ||
355 | module_init(dctcp_register); | |
356 | module_exit(dctcp_unregister); | |
357 | ||
358 | MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>"); | |
359 | MODULE_AUTHOR("Florian Westphal <fw@strlen.de>"); | |
360 | MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>"); | |
361 | ||
362 | MODULE_LICENSE("GPL v2"); | |
363 | MODULE_DESCRIPTION("DataCenter TCP (DCTCP)"); |