xref: /linux/net/ipv4/tcp_vegas.c (revision d67b569f5f620c0fb95d5212642746b7ba9d29e4)
1 /*
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 
34 #include <linux/config.h>
35 #include <linux/mm.h>
36 #include <linux/module.h>
37 #include <linux/skbuff.h>
38 #include <linux/tcp_diag.h>
39 
40 #include <net/tcp.h>
41 
42 /* Default values of the Vegas variables, in fixed-point representation
43  * with V_PARAM_SHIFT bits to the right of the binary point.
44  */
45 #define V_PARAM_SHIFT 1
46 static int alpha = 1<<V_PARAM_SHIFT;
47 static int beta  = 3<<V_PARAM_SHIFT;
48 static int gamma = 1<<V_PARAM_SHIFT;
49 
50 module_param(alpha, int, 0644);
51 MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)");
52 module_param(beta, int, 0644);
53 MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)");
54 module_param(gamma, int, 0644);
55 MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
56 
57 
58 /* Vegas variables */
59 struct vegas {
60 	u32	beg_snd_nxt;	/* right edge during last RTT */
61 	u32	beg_snd_una;	/* left edge  during last RTT */
62 	u32	beg_snd_cwnd;	/* saves the size of the cwnd */
63 	u8	doing_vegas_now;/* if true, do vegas for this RTT */
64 	u16	cntRTT;		/* # of RTTs measured within last RTT */
65 	u32	minRTT;		/* min of RTTs measured within last RTT (in usec) */
66 	u32	baseRTT;	/* the min of all Vegas RTT measurements seen (in usec) */
67 };
68 
69 /* There are several situations when we must "re-start" Vegas:
70  *
71  *  o when a connection is established
72  *  o after an RTO
73  *  o after fast recovery
74  *  o when we send a packet and there is no outstanding
75  *    unacknowledged data (restarting an idle connection)
76  *
77  * In these circumstances we cannot do a Vegas calculation at the
78  * end of the first RTT, because any calculation we do is using
79  * stale info -- both the saved cwnd and congestion feedback are
80  * stale.
81  *
82  * Instead we must wait until the completion of an RTT during
83  * which we actually receive ACKs.
84  */
85 static inline void vegas_enable(struct tcp_sock *tp)
86 {
87 	struct vegas *vegas = tcp_ca(tp);
88 
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. */
100 static inline void vegas_disable(struct tcp_sock *tp)
101 {
102 	struct vegas *vegas = tcp_ca(tp);
103 
104 	vegas->doing_vegas_now = 0;
105 }
106 
107 static void tcp_vegas_init(struct tcp_sock *tp)
108 {
109 	struct vegas *vegas = tcp_ca(tp);
110 
111 	vegas->baseRTT = 0x7fffffff;
112 	vegas_enable(tp);
113 }
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  */
123 static void tcp_vegas_rtt_calc(struct tcp_sock *tp, u32 usrtt)
124 {
125 	struct vegas *vegas = tcp_ca(tp);
126 	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 
139 static void tcp_vegas_state(struct tcp_sock *tp, u8 ca_state)
140 {
141 
142 	if (ca_state == TCP_CA_Open)
143 		vegas_enable(tp);
144 	else
145 		vegas_disable(tp);
146 }
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  */
157 static void tcp_vegas_cwnd_event(struct tcp_sock *tp, enum tcp_ca_event event)
158 {
159 	if (event == CA_EVENT_CWND_RESTART ||
160 	    event == CA_EVENT_TX_START)
161 		tcp_vegas_init(tp);
162 }
163 
164 static void tcp_vegas_cong_avoid(struct tcp_sock *tp, u32 ack,
165 				 u32 seq_rtt, u32 in_flight, int flag)
166 {
167 	struct vegas *vegas = tcp_ca(tp);
168 
169 	if (!vegas->doing_vegas_now)
170 		return tcp_reno_cong_avoid(tp, ack, seq_rtt, in_flight, flag);
171 
172 	/* The key players are v_beg_snd_una and v_beg_snd_nxt.
173 	 *
174 	 * These are so named because they represent the approximate values
175 	 * of snd_una and snd_nxt at the beginning of the current RTT. More
176 	 * precisely, they represent the amount of data sent during the RTT.
177 	 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
178 	 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
179 	 * bytes of data have been ACKed during the course of the RTT, giving
180 	 * an "actual" rate of:
181 	 *
182 	 *     (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
183 	 *
184 	 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
185 	 * because delayed ACKs can cover more than one segment, so they
186 	 * don't line up nicely with the boundaries of RTTs.
187 	 *
188 	 * Another unfortunate fact of life is that delayed ACKs delay the
189 	 * advance of the left edge of our send window, so that the number
190 	 * of bytes we send in an RTT is often less than our cwnd will allow.
191 	 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
192 	 */
193 
194 	if (after(ack, vegas->beg_snd_nxt)) {
195 		/* Do the Vegas once-per-RTT cwnd adjustment. */
196 		u32 old_wnd, old_snd_cwnd;
197 
198 
199 		/* Here old_wnd is essentially the window of data that was
200 		 * sent during the previous RTT, and has all
201 		 * been acknowledged in the course of the RTT that ended
202 		 * with the ACK we just received. Likewise, old_snd_cwnd
203 		 * is the cwnd during the previous RTT.
204 		 */
205 		old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
206 			tp->mss_cache;
207 		old_snd_cwnd = vegas->beg_snd_cwnd;
208 
209 		/* Save the extent of the current window so we can use this
210 		 * at the end of the next RTT.
211 		 */
212 		vegas->beg_snd_una  = vegas->beg_snd_nxt;
213 		vegas->beg_snd_nxt  = tp->snd_nxt;
214 		vegas->beg_snd_cwnd = tp->snd_cwnd;
215 
216 		/* Take into account the current RTT sample too, to
217 		 * decrease the impact of delayed acks. This double counts
218 		 * this sample since we count it for the next window as well,
219 		 * but that's not too awful, since we're taking the min,
220 		 * rather than averaging.
221 		 */
222 		tcp_vegas_rtt_calc(tp, seq_rtt*1000);
223 
224 		/* We do the Vegas calculations only if we got enough RTT
225 		 * samples that we can be reasonably sure that we got
226 		 * at least one RTT sample that wasn't from a delayed ACK.
227 		 * If we only had 2 samples total,
228 		 * then that means we're getting only 1 ACK per RTT, which
229 		 * means they're almost certainly delayed ACKs.
230 		 * If  we have 3 samples, we should be OK.
231 		 */
232 
233 		if (vegas->cntRTT <= 2) {
234 			/* We don't have enough RTT samples to do the Vegas
235 			 * calculation, so we'll behave like Reno.
236 			 */
237 			if (tp->snd_cwnd > tp->snd_ssthresh)
238 				tp->snd_cwnd++;
239 		} else {
240 			u32 rtt, target_cwnd, diff;
241 
242 			/* We have enough RTT samples, so, using the Vegas
243 			 * algorithm, we determine if we should increase or
244 			 * decrease cwnd, and by how much.
245 			 */
246 
247 			/* Pluck out the RTT we are using for the Vegas
248 			 * calculations. This is the min RTT seen during the
249 			 * last RTT. Taking the min filters out the effects
250 			 * of delayed ACKs, at the cost of noticing congestion
251 			 * a bit later.
252 			 */
253 			rtt = vegas->minRTT;
254 
255 			/* Calculate the cwnd we should have, if we weren't
256 			 * going too fast.
257 			 *
258 			 * This is:
259 			 *     (actual rate in segments) * baseRTT
260 			 * We keep it as a fixed point number with
261 			 * V_PARAM_SHIFT bits to the right of the binary point.
262 			 */
263 			target_cwnd = ((old_wnd * vegas->baseRTT)
264 				       << V_PARAM_SHIFT) / rtt;
265 
266 			/* Calculate the difference between the window we had,
267 			 * and the window we would like to have. This quantity
268 			 * is the "Diff" from the Arizona Vegas papers.
269 			 *
270 			 * Again, this is a fixed point number with
271 			 * V_PARAM_SHIFT bits to the right of the binary
272 			 * point.
273 			 */
274 			diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
275 
276 			if (tp->snd_cwnd < tp->snd_ssthresh) {
277 				/* Slow start.  */
278 				if (diff > gamma) {
279 					/* Going too fast. Time to slow down
280 					 * and switch to congestion avoidance.
281 					 */
282 					tp->snd_ssthresh = 2;
283 
284 					/* Set cwnd to match the actual rate
285 					 * exactly:
286 					 *   cwnd = (actual rate) * baseRTT
287 					 * Then we add 1 because the integer
288 					 * truncation robs us of full link
289 					 * utilization.
290 					 */
291 					tp->snd_cwnd = min(tp->snd_cwnd,
292 							   (target_cwnd >>
293 							    V_PARAM_SHIFT)+1);
294 
295 				}
296 			} else {
297 				/* Congestion avoidance. */
298 				u32 next_snd_cwnd;
299 
300 				/* Figure out where we would like cwnd
301 				 * to be.
302 				 */
303 				if (diff > beta) {
304 					/* The old window was too fast, so
305 					 * we slow down.
306 					 */
307 					next_snd_cwnd = old_snd_cwnd - 1;
308 				} else if (diff < alpha) {
309 					/* We don't have enough extra packets
310 					 * in the network, so speed up.
311 					 */
312 					next_snd_cwnd = old_snd_cwnd + 1;
313 				} else {
314 					/* Sending just as fast as we
315 					 * should be.
316 					 */
317 					next_snd_cwnd = old_snd_cwnd;
318 				}
319 
320 				/* Adjust cwnd upward or downward, toward the
321 				 * desired value.
322 				 */
323 				if (next_snd_cwnd > tp->snd_cwnd)
324 					tp->snd_cwnd++;
325 				else if (next_snd_cwnd < tp->snd_cwnd)
326 					tp->snd_cwnd--;
327 			}
328 		}
329 
330 		/* Wipe the slate clean for the next RTT. */
331 		vegas->cntRTT = 0;
332 		vegas->minRTT = 0x7fffffff;
333 	}
334 
335 	/* The following code is executed for every ack we receive,
336 	 * except for conditions checked in should_advance_cwnd()
337 	 * before the call to tcp_cong_avoid(). Mainly this means that
338 	 * we only execute this code if the ack actually acked some
339 	 * data.
340 	 */
341 
342 	/* If we are in slow start, increase our cwnd in response to this ACK.
343 	 * (If we are not in slow start then we are in congestion avoidance,
344 	 * and adjust our congestion window only once per RTT. See the code
345 	 * above.)
346 	 */
347 	if (tp->snd_cwnd <= tp->snd_ssthresh)
348 		tp->snd_cwnd++;
349 
350 	/* to keep cwnd from growing without bound */
351 	tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
352 
353 	/* Make sure that we are never so timid as to reduce our cwnd below
354 	 * 2 MSS.
355 	 *
356 	 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
357 	 */
358 	tp->snd_cwnd = max(tp->snd_cwnd, 2U);
359 }
360 
361 /* Extract info for Tcp socket info provided via netlink. */
362 static void tcp_vegas_get_info(struct tcp_sock *tp, u32 ext,
363 			       struct sk_buff *skb)
364 {
365 	const struct vegas *ca = tcp_ca(tp);
366 	if (ext & (1<<(TCPDIAG_VEGASINFO-1))) {
367 		struct tcpvegas_info *info;
368 
369 		info = RTA_DATA(__RTA_PUT(skb, TCPDIAG_VEGASINFO,
370 					  sizeof(*info)));
371 
372 		info->tcpv_enabled = ca->doing_vegas_now;
373 		info->tcpv_rttcnt = ca->cntRTT;
374 		info->tcpv_rtt = ca->baseRTT;
375 		info->tcpv_minrtt = ca->minRTT;
376 	rtattr_failure:	;
377 	}
378 }
379 
380 static struct tcp_congestion_ops tcp_vegas = {
381 	.init		= tcp_vegas_init,
382 	.ssthresh	= tcp_reno_ssthresh,
383 	.cong_avoid	= tcp_vegas_cong_avoid,
384 	.min_cwnd	= tcp_reno_min_cwnd,
385 	.rtt_sample	= tcp_vegas_rtt_calc,
386 	.set_state	= tcp_vegas_state,
387 	.cwnd_event	= tcp_vegas_cwnd_event,
388 	.get_info	= tcp_vegas_get_info,
389 
390 	.owner		= THIS_MODULE,
391 	.name		= "vegas",
392 };
393 
394 static int __init tcp_vegas_register(void)
395 {
396 	BUG_ON(sizeof(struct vegas) > TCP_CA_PRIV_SIZE);
397 	tcp_register_congestion_control(&tcp_vegas);
398 	return 0;
399 }
400 
401 static void __exit tcp_vegas_unregister(void)
402 {
403 	tcp_unregister_congestion_control(&tcp_vegas);
404 }
405 
406 module_init(tcp_vegas_register);
407 module_exit(tcp_vegas_unregister);
408 
409 MODULE_AUTHOR("Stephen Hemminger");
410 MODULE_LICENSE("GPL");
411 MODULE_DESCRIPTION("TCP Vegas");
412