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; | |
59 | }; | |
60 | ||
61 | static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */ | |
62 | module_param(dctcp_shift_g, uint, 0644); | |
63 | MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha"); | |
64 | ||
65 | static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA; | |
66 | module_param(dctcp_alpha_on_init, uint, 0644); | |
67 | MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value"); | |
68 | ||
69 | static unsigned int dctcp_clamp_alpha_on_loss __read_mostly; | |
70 | module_param(dctcp_clamp_alpha_on_loss, uint, 0644); | |
71 | MODULE_PARM_DESC(dctcp_clamp_alpha_on_loss, | |
72 | "parameter for clamping alpha on loss"); | |
73 | ||
74 | static struct tcp_congestion_ops dctcp_reno; | |
75 | ||
76 | static 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 | ||
84 | static 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 | ||
112 | static 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 | ||
126 | static 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 | ||
156 | static 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 | ||
186 | static 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 | /* For avoiding denominator == 1. */ | |
208 | if (ca->acked_bytes_total == 0) | |
209 | ca->acked_bytes_total = 1; | |
210 | ||
211 | /* alpha = (1 - g) * alpha + g * F */ | |
212 | ca->dctcp_alpha = ca->dctcp_alpha - | |
213 | (ca->dctcp_alpha >> dctcp_shift_g) + | |
214 | (ca->acked_bytes_ecn << (10U - dctcp_shift_g)) / | |
215 | ca->acked_bytes_total; | |
216 | ||
217 | if (ca->dctcp_alpha > DCTCP_MAX_ALPHA) | |
218 | /* Clamp dctcp_alpha to max. */ | |
219 | ca->dctcp_alpha = DCTCP_MAX_ALPHA; | |
220 | ||
221 | dctcp_reset(tp, ca); | |
222 | } | |
223 | } | |
224 | ||
225 | static void dctcp_state(struct sock *sk, u8 new_state) | |
226 | { | |
227 | if (dctcp_clamp_alpha_on_loss && new_state == TCP_CA_Loss) { | |
228 | struct dctcp *ca = inet_csk_ca(sk); | |
229 | ||
230 | /* If this extension is enabled, we clamp dctcp_alpha to | |
231 | * max on packet loss; the motivation is that dctcp_alpha | |
232 | * is an indicator to the extend of congestion and packet | |
233 | * loss is an indicator of extreme congestion; setting | |
234 | * this in practice turned out to be beneficial, and | |
235 | * effectively assumes total congestion which reduces the | |
236 | * window by half. | |
237 | */ | |
238 | ca->dctcp_alpha = DCTCP_MAX_ALPHA; | |
239 | } | |
240 | } | |
241 | ||
242 | static void dctcp_update_ack_reserved(struct sock *sk, enum tcp_ca_event ev) | |
243 | { | |
244 | struct dctcp *ca = inet_csk_ca(sk); | |
245 | ||
246 | switch (ev) { | |
247 | case CA_EVENT_DELAYED_ACK: | |
248 | if (!ca->delayed_ack_reserved) | |
249 | ca->delayed_ack_reserved = 1; | |
250 | break; | |
251 | case CA_EVENT_NON_DELAYED_ACK: | |
252 | if (ca->delayed_ack_reserved) | |
253 | ca->delayed_ack_reserved = 0; | |
254 | break; | |
255 | default: | |
256 | /* Don't care for the rest. */ | |
257 | break; | |
258 | } | |
259 | } | |
260 | ||
261 | static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev) | |
262 | { | |
263 | switch (ev) { | |
264 | case CA_EVENT_ECN_IS_CE: | |
265 | dctcp_ce_state_0_to_1(sk); | |
266 | break; | |
267 | case CA_EVENT_ECN_NO_CE: | |
268 | dctcp_ce_state_1_to_0(sk); | |
269 | break; | |
270 | case CA_EVENT_DELAYED_ACK: | |
271 | case CA_EVENT_NON_DELAYED_ACK: | |
272 | dctcp_update_ack_reserved(sk, ev); | |
273 | break; | |
274 | default: | |
275 | /* Don't care for the rest. */ | |
276 | break; | |
277 | } | |
278 | } | |
279 | ||
280 | static void dctcp_get_info(struct sock *sk, u32 ext, struct sk_buff *skb) | |
281 | { | |
282 | const struct dctcp *ca = inet_csk_ca(sk); | |
283 | ||
284 | /* Fill it also in case of VEGASINFO due to req struct limits. | |
285 | * We can still correctly retrieve it later. | |
286 | */ | |
287 | if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) || | |
288 | ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | |
289 | struct tcp_dctcp_info info; | |
290 | ||
291 | memset(&info, 0, sizeof(info)); | |
292 | if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) { | |
293 | info.dctcp_enabled = 1; | |
294 | info.dctcp_ce_state = (u16) ca->ce_state; | |
295 | info.dctcp_alpha = ca->dctcp_alpha; | |
296 | info.dctcp_ab_ecn = ca->acked_bytes_ecn; | |
297 | info.dctcp_ab_tot = ca->acked_bytes_total; | |
298 | } | |
299 | ||
300 | nla_put(skb, INET_DIAG_DCTCPINFO, sizeof(info), &info); | |
301 | } | |
302 | } | |
303 | ||
304 | static struct tcp_congestion_ops dctcp __read_mostly = { | |
305 | .init = dctcp_init, | |
306 | .in_ack_event = dctcp_update_alpha, | |
307 | .cwnd_event = dctcp_cwnd_event, | |
308 | .ssthresh = dctcp_ssthresh, | |
309 | .cong_avoid = tcp_reno_cong_avoid, | |
310 | .set_state = dctcp_state, | |
311 | .get_info = dctcp_get_info, | |
312 | .flags = TCP_CONG_NEEDS_ECN, | |
313 | .owner = THIS_MODULE, | |
314 | .name = "dctcp", | |
315 | }; | |
316 | ||
317 | static struct tcp_congestion_ops dctcp_reno __read_mostly = { | |
318 | .ssthresh = tcp_reno_ssthresh, | |
319 | .cong_avoid = tcp_reno_cong_avoid, | |
320 | .get_info = dctcp_get_info, | |
321 | .owner = THIS_MODULE, | |
322 | .name = "dctcp-reno", | |
323 | }; | |
324 | ||
325 | static int __init dctcp_register(void) | |
326 | { | |
327 | BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE); | |
328 | return tcp_register_congestion_control(&dctcp); | |
329 | } | |
330 | ||
331 | static void __exit dctcp_unregister(void) | |
332 | { | |
333 | tcp_unregister_congestion_control(&dctcp); | |
334 | } | |
335 | ||
336 | module_init(dctcp_register); | |
337 | module_exit(dctcp_unregister); | |
338 | ||
339 | MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>"); | |
340 | MODULE_AUTHOR("Florian Westphal <fw@strlen.de>"); | |
341 | MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>"); | |
342 | ||
343 | MODULE_LICENSE("GPL v2"); | |
344 | MODULE_DESCRIPTION("DataCenter TCP (DCTCP)"); |