blob: 7cd22262de3af3fa0f2917eefaea885c068855ac [file] [log] [blame]
Stephen Hemmingerb87d8562005-06-23 12:27:19 -07001/*
2 * TCP Vegas congestion control
3 *
4 * This is based on the congestion detection/avoidance scheme described in
5 * Lawrence S. Brakmo and Larry L. Peterson.
6 * "TCP Vegas: End to end congestion avoidance on a global internet."
7 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
8 * October 1995. Available from:
9 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
10 *
11 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
12 * The main aspects that distinguish this implementation from the
13 * Arizona Vegas implementation are:
14 * o We do not change the loss detection or recovery mechanisms of
15 * Linux in any way. Linux already recovers from losses quite well,
16 * using fine-grained timers, NewReno, and FACK.
17 * o To avoid the performance penalty imposed by increasing cwnd
18 * only every-other RTT during slow start, we increase during
19 * every RTT during slow start, just like Reno.
20 * o Largely to allow continuous cwnd growth during slow start,
21 * we use the rate at which ACKs come back as the "actual"
22 * rate, rather than the rate at which data is sent.
23 * o To speed convergence to the right rate, we set the cwnd
24 * to achieve the right ("actual") rate when we exit slow start.
25 * o To filter out the noise caused by delayed ACKs, we use the
26 * minimum RTT sample observed during the last RTT to calculate
27 * the actual rate.
28 * o When the sender re-starts from idle, it waits until it has
29 * received ACKs for an entire flight of new data before making
30 * a cwnd adjustment decision. The original Vegas implementation
31 * assumed senders never went idle.
32 */
33
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070034#include <linux/mm.h>
35#include <linux/module.h>
36#include <linux/skbuff.h>
Arnaldo Carvalho de Meloa8c21902005-08-12 12:56:38 -030037#include <linux/inet_diag.h>
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070038
39#include <net/tcp.h>
40
Stephen Hemminger77522372007-04-23 22:28:23 -070041#include "tcp_vegas.h"
42
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070043/* Default values of the Vegas variables, in fixed-point representation
44 * with V_PARAM_SHIFT bits to the right of the binary point.
45 */
46#define V_PARAM_SHIFT 1
David S. Millerd2e4bdc2006-11-28 14:37:38 -080047static int alpha = 2<<V_PARAM_SHIFT;
48static int beta = 4<<V_PARAM_SHIFT;
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070049static int gamma = 1<<V_PARAM_SHIFT;
50
51module_param(alpha, int, 0644);
52MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)");
53module_param(beta, int, 0644);
54MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)");
55module_param(gamma, int, 0644);
56MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
57
58
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070059/* There are several situations when we must "re-start" Vegas:
60 *
61 * o when a connection is established
62 * o after an RTO
63 * o after fast recovery
64 * o when we send a packet and there is no outstanding
65 * unacknowledged data (restarting an idle connection)
66 *
67 * In these circumstances we cannot do a Vegas calculation at the
68 * end of the first RTT, because any calculation we do is using
69 * stale info -- both the saved cwnd and congestion feedback are
70 * stale.
71 *
72 * Instead we must wait until the completion of an RTT during
73 * which we actually receive ACKs.
74 */
Stephen Hemminger77522372007-04-23 22:28:23 -070075static void vegas_enable(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070076{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030077 const struct tcp_sock *tp = tcp_sk(sk);
78 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070079
80 /* Begin taking Vegas samples next time we send something. */
81 vegas->doing_vegas_now = 1;
82
83 /* Set the beginning of the next send window. */
84 vegas->beg_snd_nxt = tp->snd_nxt;
85
86 vegas->cntRTT = 0;
87 vegas->minRTT = 0x7fffffff;
88}
89
90/* Stop taking Vegas samples for now. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030091static inline void vegas_disable(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070092{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030093 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070094
95 vegas->doing_vegas_now = 0;
96}
97
Stephen Hemminger77522372007-04-23 22:28:23 -070098void tcp_vegas_init(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070099{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300100 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700101
102 vegas->baseRTT = 0x7fffffff;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300103 vegas_enable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700104}
Stephen Hemminger77522372007-04-23 22:28:23 -0700105EXPORT_SYMBOL_GPL(tcp_vegas_init);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700106
107/* Do RTT sampling needed for Vegas.
108 * Basically we:
109 * o min-filter RTT samples from within an RTT to get the current
110 * propagation delay + queuing delay (we are min-filtering to try to
111 * avoid the effects of delayed ACKs)
112 * o min-filter RTT samples from a much longer window (forever for now)
113 * to find the propagation delay (baseRTT)
114 */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700115void tcp_vegas_pkts_acked(struct sock *sk, u32 cnt, s32 rtt_us)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700116{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300117 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemminger164891a2007-04-23 22:26:16 -0700118 u32 vrtt;
119
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700120 if (rtt_us < 0)
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -0700121 return;
122
Stephen Hemminger164891a2007-04-23 22:26:16 -0700123 /* Never allow zero rtt or baseRTT */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700124 vrtt = rtt_us + 1;
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700125
126 /* Filter to find propagation delay: */
127 if (vrtt < vegas->baseRTT)
128 vegas->baseRTT = vrtt;
129
130 /* Find the min RTT during the last RTT to find
131 * the current prop. delay + queuing delay:
132 */
133 vegas->minRTT = min(vegas->minRTT, vrtt);
134 vegas->cntRTT++;
135}
Stephen Hemminger77522372007-04-23 22:28:23 -0700136EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700137
Stephen Hemminger77522372007-04-23 22:28:23 -0700138void tcp_vegas_state(struct sock *sk, u8 ca_state)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700139{
140
141 if (ca_state == TCP_CA_Open)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300142 vegas_enable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700143 else
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300144 vegas_disable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700145}
Stephen Hemminger77522372007-04-23 22:28:23 -0700146EXPORT_SYMBOL_GPL(tcp_vegas_state);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700147
148/*
149 * If the connection is idle and we are restarting,
150 * then we don't want to do any Vegas calculations
151 * until we get fresh RTT samples. So when we
152 * restart, we reset our Vegas state to a clean
153 * slate. After we get acks for this flight of
154 * packets, _then_ we can make Vegas calculations
155 * again.
156 */
Stephen Hemminger77522372007-04-23 22:28:23 -0700157void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700158{
159 if (event == CA_EVENT_CWND_RESTART ||
160 event == CA_EVENT_TX_START)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300161 tcp_vegas_init(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700162}
Stephen Hemminger77522372007-04-23 22:28:23 -0700163EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700164
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200165static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700166{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300167 struct tcp_sock *tp = tcp_sk(sk);
168 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700169
Harvey Harrisonab598592008-05-01 02:47:38 -0700170 if (!vegas->doing_vegas_now) {
171 tcp_reno_cong_avoid(sk, ack, in_flight);
172 return;
173 }
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700174
175 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
176 *
177 * These are so named because they represent the approximate values
178 * of snd_una and snd_nxt at the beginning of the current RTT. More
179 * precisely, they represent the amount of data sent during the RTT.
180 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
181 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
182 * bytes of data have been ACKed during the course of the RTT, giving
183 * an "actual" rate of:
184 *
185 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
186 *
187 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
188 * because delayed ACKs can cover more than one segment, so they
189 * don't line up nicely with the boundaries of RTTs.
190 *
191 * Another unfortunate fact of life is that delayed ACKs delay the
192 * advance of the left edge of our send window, so that the number
193 * of bytes we send in an RTT is often less than our cwnd will allow.
194 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
195 */
196
197 if (after(ack, vegas->beg_snd_nxt)) {
198 /* Do the Vegas once-per-RTT cwnd adjustment. */
199 u32 old_wnd, old_snd_cwnd;
200
201
202 /* Here old_wnd is essentially the window of data that was
203 * sent during the previous RTT, and has all
204 * been acknowledged in the course of the RTT that ended
205 * with the ACK we just received. Likewise, old_snd_cwnd
206 * is the cwnd during the previous RTT.
207 */
208 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
209 tp->mss_cache;
210 old_snd_cwnd = vegas->beg_snd_cwnd;
211
212 /* Save the extent of the current window so we can use this
213 * at the end of the next RTT.
214 */
215 vegas->beg_snd_una = vegas->beg_snd_nxt;
216 vegas->beg_snd_nxt = tp->snd_nxt;
217 vegas->beg_snd_cwnd = tp->snd_cwnd;
218
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700219 /* We do the Vegas calculations only if we got enough RTT
220 * samples that we can be reasonably sure that we got
221 * at least one RTT sample that wasn't from a delayed ACK.
222 * If we only had 2 samples total,
223 * then that means we're getting only 1 ACK per RTT, which
224 * means they're almost certainly delayed ACKs.
225 * If we have 3 samples, we should be OK.
226 */
227
228 if (vegas->cntRTT <= 2) {
229 /* We don't have enough RTT samples to do the Vegas
230 * calculation, so we'll behave like Reno.
231 */
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200232 tcp_reno_cong_avoid(sk, ack, in_flight);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700233 } else {
Lachlan Andrew15913112008-04-30 01:04:03 -0700234 u32 rtt, diff;
235 u64 target_cwnd;
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700236
237 /* We have enough RTT samples, so, using the Vegas
238 * algorithm, we determine if we should increase or
239 * decrease cwnd, and by how much.
240 */
241
242 /* Pluck out the RTT we are using for the Vegas
243 * calculations. This is the min RTT seen during the
244 * last RTT. Taking the min filters out the effects
245 * of delayed ACKs, at the cost of noticing congestion
246 * a bit later.
247 */
248 rtt = vegas->minRTT;
249
250 /* Calculate the cwnd we should have, if we weren't
251 * going too fast.
252 *
253 * This is:
254 * (actual rate in segments) * baseRTT
255 * We keep it as a fixed point number with
256 * V_PARAM_SHIFT bits to the right of the binary point.
257 */
Lachlan Andrew15913112008-04-30 01:04:03 -0700258 target_cwnd = ((u64)old_wnd * vegas->baseRTT);
259 target_cwnd <<= V_PARAM_SHIFT;
260 do_div(target_cwnd, rtt);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700261
262 /* Calculate the difference between the window we had,
263 * and the window we would like to have. This quantity
264 * is the "Diff" from the Arizona Vegas papers.
265 *
266 * Again, this is a fixed point number with
267 * V_PARAM_SHIFT bits to the right of the binary
268 * point.
269 */
270 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
271
Xiaoliang (David) Weic9405872007-10-29 20:24:36 -0700272 if (diff > gamma && tp->snd_ssthresh > 2 ) {
273 /* Going too fast. Time to slow down
274 * and switch to congestion avoidance.
275 */
276 tp->snd_ssthresh = 2;
277
278 /* Set cwnd to match the actual rate
279 * exactly:
280 * cwnd = (actual rate) * baseRTT
281 * Then we add 1 because the integer
282 * truncation robs us of full link
283 * utilization.
284 */
285 tp->snd_cwnd = min(tp->snd_cwnd,
Lachlan Andrew15913112008-04-30 01:04:03 -0700286 ((u32)target_cwnd >>
Xiaoliang (David) Weic9405872007-10-29 20:24:36 -0700287 V_PARAM_SHIFT)+1);
288
289 } else if (tp->snd_cwnd <= tp->snd_ssthresh) {
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700290 /* Slow start. */
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800291 tcp_slow_start(tp);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700292 } else {
293 /* Congestion avoidance. */
294 u32 next_snd_cwnd;
295
296 /* Figure out where we would like cwnd
297 * to be.
298 */
299 if (diff > beta) {
300 /* The old window was too fast, so
301 * we slow down.
302 */
303 next_snd_cwnd = old_snd_cwnd - 1;
304 } else if (diff < alpha) {
305 /* We don't have enough extra packets
306 * in the network, so speed up.
307 */
308 next_snd_cwnd = old_snd_cwnd + 1;
309 } else {
310 /* Sending just as fast as we
311 * should be.
312 */
313 next_snd_cwnd = old_snd_cwnd;
314 }
315
316 /* Adjust cwnd upward or downward, toward the
317 * desired value.
318 */
319 if (next_snd_cwnd > tp->snd_cwnd)
320 tp->snd_cwnd++;
321 else if (next_snd_cwnd < tp->snd_cwnd)
322 tp->snd_cwnd--;
323 }
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700324
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800325 if (tp->snd_cwnd < 2)
326 tp->snd_cwnd = 2;
327 else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
328 tp->snd_cwnd = tp->snd_cwnd_clamp;
Doug Leitha6af2d62008-12-04 17:17:18 -0800329
330 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800331 }
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700332
Thomas Young5b495612005-12-06 16:16:34 -0800333 /* Wipe the slate clean for the next RTT. */
334 vegas->cntRTT = 0;
335 vegas->minRTT = 0x7fffffff;
336 }
Thomas Young74cb8792006-01-04 13:59:32 -0800337 /* Use normal slow start */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900338 else if (tp->snd_cwnd <= tp->snd_ssthresh)
Thomas Young74cb8792006-01-04 13:59:32 -0800339 tcp_slow_start(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900340
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700341}
342
343/* Extract info for Tcp socket info provided via netlink. */
Stephen Hemminger77522372007-04-23 22:28:23 -0700344void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700345{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300346 const struct vegas *ca = inet_csk_ca(sk);
Arnaldo Carvalho de Melo73c1f4a2005-08-12 12:51:49 -0300347 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
Thomas Grafe9195d62007-03-22 23:27:01 -0700348 struct tcpvegas_info info = {
349 .tcpv_enabled = ca->doing_vegas_now,
350 .tcpv_rttcnt = ca->cntRTT,
351 .tcpv_rtt = ca->baseRTT,
352 .tcpv_minrtt = ca->minRTT,
353 };
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700354
Thomas Grafe9195d62007-03-22 23:27:01 -0700355 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700356 }
357}
Stephen Hemminger77522372007-04-23 22:28:23 -0700358EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700359
360static struct tcp_congestion_ops tcp_vegas = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700361 .flags = TCP_CONG_RTT_STAMP,
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700362 .init = tcp_vegas_init,
363 .ssthresh = tcp_reno_ssthresh,
364 .cong_avoid = tcp_vegas_cong_avoid,
365 .min_cwnd = tcp_reno_min_cwnd,
Stephen Hemminger164891a2007-04-23 22:26:16 -0700366 .pkts_acked = tcp_vegas_pkts_acked,
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700367 .set_state = tcp_vegas_state,
368 .cwnd_event = tcp_vegas_cwnd_event,
369 .get_info = tcp_vegas_get_info,
370
371 .owner = THIS_MODULE,
372 .name = "vegas",
373};
374
375static int __init tcp_vegas_register(void)
376{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700377 BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700378 tcp_register_congestion_control(&tcp_vegas);
379 return 0;
380}
381
382static void __exit tcp_vegas_unregister(void)
383{
384 tcp_unregister_congestion_control(&tcp_vegas);
385}
386
387module_init(tcp_vegas_register);
388module_exit(tcp_vegas_unregister);
389
390MODULE_AUTHOR("Stephen Hemminger");
391MODULE_LICENSE("GPL");
392MODULE_DESCRIPTION("TCP Vegas");