blob: ddc4bcc5785eed2737c2ddca93422026f86c3e0f [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
41/* Default values of the Vegas variables, in fixed-point representation
42 * with V_PARAM_SHIFT bits to the right of the binary point.
43 */
44#define V_PARAM_SHIFT 1
David S. Millerd2e4bdc2006-11-28 14:37:38 -080045static int alpha = 2<<V_PARAM_SHIFT;
46static int beta = 4<<V_PARAM_SHIFT;
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070047static int gamma = 1<<V_PARAM_SHIFT;
48
49module_param(alpha, int, 0644);
50MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)");
51module_param(beta, int, 0644);
52MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)");
53module_param(gamma, int, 0644);
54MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
55
56
57/* Vegas variables */
58struct vegas {
59 u32 beg_snd_nxt; /* right edge during last RTT */
60 u32 beg_snd_una; /* left edge during last RTT */
61 u32 beg_snd_cwnd; /* saves the size of the cwnd */
62 u8 doing_vegas_now;/* if true, do vegas for this RTT */
63 u16 cntRTT; /* # of RTTs measured within last RTT */
64 u32 minRTT; /* min of RTTs measured within last RTT (in usec) */
65 u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */
66};
67
68/* There are several situations when we must "re-start" Vegas:
69 *
70 * o when a connection is established
71 * o after an RTO
72 * o after fast recovery
73 * o when we send a packet and there is no outstanding
74 * unacknowledged data (restarting an idle connection)
75 *
76 * In these circumstances we cannot do a Vegas calculation at the
77 * end of the first RTT, because any calculation we do is using
78 * stale info -- both the saved cwnd and congestion feedback are
79 * stale.
80 *
81 * Instead we must wait until the completion of an RTT during
82 * which we actually receive ACKs.
83 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030084static inline void vegas_enable(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070085{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -030086 const struct tcp_sock *tp = tcp_sk(sk);
87 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -070088
89 /* Begin taking Vegas samples next time we send something. */
90 vegas->doing_vegas_now = 1;
91
92 /* Set the beginning of the next send window. */
93 vegas->beg_snd_nxt = tp->snd_nxt;
94
95 vegas->cntRTT = 0;
96 vegas->minRTT = 0x7fffffff;
97}
98
99/* Stop taking Vegas samples for now. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300100static inline void vegas_disable(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700101{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300102 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700103
104 vegas->doing_vegas_now = 0;
105}
106
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300107static void tcp_vegas_init(struct sock *sk)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700108{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300109 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700110
111 vegas->baseRTT = 0x7fffffff;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300112 vegas_enable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700113}
114
115/* Do RTT sampling needed for Vegas.
116 * Basically we:
117 * o min-filter RTT samples from within an RTT to get the current
118 * propagation delay + queuing delay (we are min-filtering to try to
119 * avoid the effects of delayed ACKs)
120 * o min-filter RTT samples from a much longer window (forever for now)
121 * to find the propagation delay (baseRTT)
122 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300123static void tcp_vegas_rtt_calc(struct sock *sk, u32 usrtt)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700124{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300125 struct vegas *vegas = inet_csk_ca(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700126 u32 vrtt = usrtt + 1; /* Never allow zero rtt or baseRTT */
127
128 /* Filter to find propagation delay: */
129 if (vrtt < vegas->baseRTT)
130 vegas->baseRTT = vrtt;
131
132 /* Find the min RTT during the last RTT to find
133 * the current prop. delay + queuing delay:
134 */
135 vegas->minRTT = min(vegas->minRTT, vrtt);
136 vegas->cntRTT++;
137}
138
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300139static void tcp_vegas_state(struct sock *sk, u8 ca_state)
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700140{
141
142 if (ca_state == TCP_CA_Open)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300143 vegas_enable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700144 else
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300145 vegas_disable(sk);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700146}
147
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 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300157static void 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}
163
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300164static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack,
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700165 u32 seq_rtt, u32 in_flight, int flag)
166{
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
170 if (!vegas->doing_vegas_now)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300171 return tcp_reno_cong_avoid(sk, ack, seq_rtt, in_flight, flag);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700172
173 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
174 *
175 * These are so named because they represent the approximate values
176 * of snd_una and snd_nxt at the beginning of the current RTT. More
177 * precisely, they represent the amount of data sent during the RTT.
178 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
179 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
180 * bytes of data have been ACKed during the course of the RTT, giving
181 * an "actual" rate of:
182 *
183 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
184 *
185 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
186 * because delayed ACKs can cover more than one segment, so they
187 * don't line up nicely with the boundaries of RTTs.
188 *
189 * Another unfortunate fact of life is that delayed ACKs delay the
190 * advance of the left edge of our send window, so that the number
191 * of bytes we send in an RTT is often less than our cwnd will allow.
192 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
193 */
194
195 if (after(ack, vegas->beg_snd_nxt)) {
196 /* Do the Vegas once-per-RTT cwnd adjustment. */
197 u32 old_wnd, old_snd_cwnd;
198
199
200 /* Here old_wnd is essentially the window of data that was
201 * sent during the previous RTT, and has all
202 * been acknowledged in the course of the RTT that ended
203 * with the ACK we just received. Likewise, old_snd_cwnd
204 * is the cwnd during the previous RTT.
205 */
206 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
207 tp->mss_cache;
208 old_snd_cwnd = vegas->beg_snd_cwnd;
209
210 /* Save the extent of the current window so we can use this
211 * at the end of the next RTT.
212 */
213 vegas->beg_snd_una = vegas->beg_snd_nxt;
214 vegas->beg_snd_nxt = tp->snd_nxt;
215 vegas->beg_snd_cwnd = tp->snd_cwnd;
216
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700217 /* We do the Vegas calculations only if we got enough RTT
218 * samples that we can be reasonably sure that we got
219 * at least one RTT sample that wasn't from a delayed ACK.
220 * If we only had 2 samples total,
221 * then that means we're getting only 1 ACK per RTT, which
222 * means they're almost certainly delayed ACKs.
223 * If we have 3 samples, we should be OK.
224 */
225
226 if (vegas->cntRTT <= 2) {
227 /* We don't have enough RTT samples to do the Vegas
228 * calculation, so we'll behave like Reno.
229 */
Jeff Garzikc0509702005-11-11 04:43:47 -0500230 tcp_reno_cong_avoid(sk, ack, seq_rtt, in_flight, flag);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700231 } else {
232 u32 rtt, target_cwnd, diff;
233
234 /* We have enough RTT samples, so, using the Vegas
235 * algorithm, we determine if we should increase or
236 * decrease cwnd, and by how much.
237 */
238
239 /* Pluck out the RTT we are using for the Vegas
240 * calculations. This is the min RTT seen during the
241 * last RTT. Taking the min filters out the effects
242 * of delayed ACKs, at the cost of noticing congestion
243 * a bit later.
244 */
245 rtt = vegas->minRTT;
246
247 /* Calculate the cwnd we should have, if we weren't
248 * going too fast.
249 *
250 * This is:
251 * (actual rate in segments) * baseRTT
252 * We keep it as a fixed point number with
253 * V_PARAM_SHIFT bits to the right of the binary point.
254 */
255 target_cwnd = ((old_wnd * vegas->baseRTT)
256 << V_PARAM_SHIFT) / rtt;
257
258 /* Calculate the difference between the window we had,
259 * and the window we would like to have. This quantity
260 * is the "Diff" from the Arizona Vegas papers.
261 *
262 * Again, this is a fixed point number with
263 * V_PARAM_SHIFT bits to the right of the binary
264 * point.
265 */
266 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
267
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800268 if (tp->snd_cwnd <= tp->snd_ssthresh) {
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700269 /* Slow start. */
270 if (diff > gamma) {
271 /* Going too fast. Time to slow down
272 * and switch to congestion avoidance.
273 */
274 tp->snd_ssthresh = 2;
275
276 /* Set cwnd to match the actual rate
277 * exactly:
278 * cwnd = (actual rate) * baseRTT
279 * Then we add 1 because the integer
280 * truncation robs us of full link
281 * utilization.
282 */
283 tp->snd_cwnd = min(tp->snd_cwnd,
284 (target_cwnd >>
285 V_PARAM_SHIFT)+1);
286
287 }
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800288 tcp_slow_start(tp);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700289 } else {
290 /* Congestion avoidance. */
291 u32 next_snd_cwnd;
292
293 /* Figure out where we would like cwnd
294 * to be.
295 */
296 if (diff > beta) {
297 /* The old window was too fast, so
298 * we slow down.
299 */
300 next_snd_cwnd = old_snd_cwnd - 1;
301 } else if (diff < alpha) {
302 /* We don't have enough extra packets
303 * in the network, so speed up.
304 */
305 next_snd_cwnd = old_snd_cwnd + 1;
306 } else {
307 /* Sending just as fast as we
308 * should be.
309 */
310 next_snd_cwnd = old_snd_cwnd;
311 }
312
313 /* Adjust cwnd upward or downward, toward the
314 * desired value.
315 */
316 if (next_snd_cwnd > tp->snd_cwnd)
317 tp->snd_cwnd++;
318 else if (next_snd_cwnd < tp->snd_cwnd)
319 tp->snd_cwnd--;
320 }
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700321
Stephen Hemminger7faffa12005-11-10 17:07:24 -0800322 if (tp->snd_cwnd < 2)
323 tp->snd_cwnd = 2;
324 else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
325 tp->snd_cwnd = tp->snd_cwnd_clamp;
326 }
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700327
Thomas Young5b495612005-12-06 16:16:34 -0800328 /* Wipe the slate clean for the next RTT. */
329 vegas->cntRTT = 0;
330 vegas->minRTT = 0x7fffffff;
331 }
Thomas Young74cb8792006-01-04 13:59:32 -0800332 /* Use normal slow start */
333 else if (tp->snd_cwnd <= tp->snd_ssthresh)
334 tcp_slow_start(tp);
335
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700336}
337
338/* Extract info for Tcp socket info provided via netlink. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300339static void tcp_vegas_get_info(struct sock *sk, u32 ext,
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700340 struct sk_buff *skb)
341{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300342 const struct vegas *ca = inet_csk_ca(sk);
Arnaldo Carvalho de Melo73c1f4a2005-08-12 12:51:49 -0300343 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700344 struct tcpvegas_info *info;
345
Arnaldo Carvalho de Melo73c1f4a2005-08-12 12:51:49 -0300346 info = RTA_DATA(__RTA_PUT(skb, INET_DIAG_VEGASINFO,
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700347 sizeof(*info)));
348
349 info->tcpv_enabled = ca->doing_vegas_now;
350 info->tcpv_rttcnt = ca->cntRTT;
351 info->tcpv_rtt = ca->baseRTT;
352 info->tcpv_minrtt = ca->minRTT;
353 rtattr_failure: ;
354 }
355}
356
357static struct tcp_congestion_ops tcp_vegas = {
358 .init = tcp_vegas_init,
359 .ssthresh = tcp_reno_ssthresh,
360 .cong_avoid = tcp_vegas_cong_avoid,
361 .min_cwnd = tcp_reno_min_cwnd,
362 .rtt_sample = tcp_vegas_rtt_calc,
363 .set_state = tcp_vegas_state,
364 .cwnd_event = tcp_vegas_cwnd_event,
365 .get_info = tcp_vegas_get_info,
366
367 .owner = THIS_MODULE,
368 .name = "vegas",
369};
370
371static int __init tcp_vegas_register(void)
372{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700373 BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
Stephen Hemmingerb87d8562005-06-23 12:27:19 -0700374 tcp_register_congestion_control(&tcp_vegas);
375 return 0;
376}
377
378static void __exit tcp_vegas_unregister(void)
379{
380 tcp_unregister_congestion_control(&tcp_vegas);
381}
382
383module_init(tcp_vegas_register);
384module_exit(tcp_vegas_unregister);
385
386MODULE_AUTHOR("Stephen Hemminger");
387MODULE_LICENSE("GPL");
388MODULE_DESCRIPTION("TCP Vegas");