2 * TCP Vegas congestion control
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
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
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.
35 #include <linux/module.h>
36 #include <linux/skbuff.h>
37 #include <linux/inet_diag.h>
41 #include "tcp_vegas.h"
43 /* Default values of the Vegas variables, in fixed-point representation
44 * with V_PARAM_SHIFT bits to the right of the binary point.
46 #define V_PARAM_SHIFT 1
47 static int alpha
= 2<<V_PARAM_SHIFT
;
48 static int beta
= 4<<V_PARAM_SHIFT
;
49 static int gamma
= 1<<V_PARAM_SHIFT
;
51 module_param(alpha
, int, 0644);
52 MODULE_PARM_DESC(alpha
, "lower bound of packets in network (scale by 2)");
53 module_param(beta
, int, 0644);
54 MODULE_PARM_DESC(beta
, "upper bound of packets in network (scale by 2)");
55 module_param(gamma
, int, 0644);
56 MODULE_PARM_DESC(gamma
, "limit on increase (scale by 2)");
59 /* There are several situations when we must "re-start" Vegas:
61 * o when a connection is established
63 * o after fast recovery
64 * o when we send a packet and there is no outstanding
65 * unacknowledged data (restarting an idle connection)
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
72 * Instead we must wait until the completion of an RTT during
73 * which we actually receive ACKs.
75 static void vegas_enable(struct sock
*sk
)
77 const struct tcp_sock
*tp
= tcp_sk(sk
);
78 struct vegas
*vegas
= inet_csk_ca(sk
);
80 /* Begin taking Vegas samples next time we send something. */
81 vegas
->doing_vegas_now
= 1;
83 /* Set the beginning of the next send window. */
84 vegas
->beg_snd_nxt
= tp
->snd_nxt
;
87 vegas
->minRTT
= 0x7fffffff;
90 /* Stop taking Vegas samples for now. */
91 static inline void vegas_disable(struct sock
*sk
)
93 struct vegas
*vegas
= inet_csk_ca(sk
);
95 vegas
->doing_vegas_now
= 0;
98 void tcp_vegas_init(struct sock
*sk
)
100 struct vegas
*vegas
= inet_csk_ca(sk
);
102 vegas
->baseRTT
= 0x7fffffff;
105 EXPORT_SYMBOL_GPL(tcp_vegas_init
);
107 /* Do RTT sampling needed for Vegas.
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)
115 void tcp_vegas_pkts_acked(struct sock
*sk
, u32 cnt
, ktime_t last
)
117 struct vegas
*vegas
= inet_csk_ca(sk
);
120 /* Never allow zero rtt or baseRTT */
121 vrtt
= ktime_to_us(net_timedelta(last
)) + 1;
123 /* Filter to find propagation delay: */
124 if (vrtt
< vegas
->baseRTT
)
125 vegas
->baseRTT
= vrtt
;
127 /* Find the min RTT during the last RTT to find
128 * the current prop. delay + queuing delay:
130 vegas
->minRTT
= min(vegas
->minRTT
, vrtt
);
133 EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked
);
135 void tcp_vegas_state(struct sock
*sk
, u8 ca_state
)
138 if (ca_state
== TCP_CA_Open
)
143 EXPORT_SYMBOL_GPL(tcp_vegas_state
);
146 * If the connection is idle and we are restarting,
147 * then we don't want to do any Vegas calculations
148 * until we get fresh RTT samples. So when we
149 * restart, we reset our Vegas state to a clean
150 * slate. After we get acks for this flight of
151 * packets, _then_ we can make Vegas calculations
154 void tcp_vegas_cwnd_event(struct sock
*sk
, enum tcp_ca_event event
)
156 if (event
== CA_EVENT_CWND_RESTART
||
157 event
== CA_EVENT_TX_START
)
160 EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event
);
162 static void tcp_vegas_cong_avoid(struct sock
*sk
, u32 ack
,
163 u32 seq_rtt
, u32 in_flight
, int flag
)
165 struct tcp_sock
*tp
= tcp_sk(sk
);
166 struct vegas
*vegas
= inet_csk_ca(sk
);
168 if (!vegas
->doing_vegas_now
)
169 return tcp_reno_cong_avoid(sk
, ack
, seq_rtt
, in_flight
, flag
);
171 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
173 * These are so named because they represent the approximate values
174 * of snd_una and snd_nxt at the beginning of the current RTT. More
175 * precisely, they represent the amount of data sent during the RTT.
176 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
177 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
178 * bytes of data have been ACKed during the course of the RTT, giving
179 * an "actual" rate of:
181 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
183 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
184 * because delayed ACKs can cover more than one segment, so they
185 * don't line up nicely with the boundaries of RTTs.
187 * Another unfortunate fact of life is that delayed ACKs delay the
188 * advance of the left edge of our send window, so that the number
189 * of bytes we send in an RTT is often less than our cwnd will allow.
190 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
193 if (after(ack
, vegas
->beg_snd_nxt
)) {
194 /* Do the Vegas once-per-RTT cwnd adjustment. */
195 u32 old_wnd
, old_snd_cwnd
;
198 /* Here old_wnd is essentially the window of data that was
199 * sent during the previous RTT, and has all
200 * been acknowledged in the course of the RTT that ended
201 * with the ACK we just received. Likewise, old_snd_cwnd
202 * is the cwnd during the previous RTT.
204 old_wnd
= (vegas
->beg_snd_nxt
- vegas
->beg_snd_una
) /
206 old_snd_cwnd
= vegas
->beg_snd_cwnd
;
208 /* Save the extent of the current window so we can use this
209 * at the end of the next RTT.
211 vegas
->beg_snd_una
= vegas
->beg_snd_nxt
;
212 vegas
->beg_snd_nxt
= tp
->snd_nxt
;
213 vegas
->beg_snd_cwnd
= tp
->snd_cwnd
;
215 /* We do the Vegas calculations only if we got enough RTT
216 * samples that we can be reasonably sure that we got
217 * at least one RTT sample that wasn't from a delayed ACK.
218 * If we only had 2 samples total,
219 * then that means we're getting only 1 ACK per RTT, which
220 * means they're almost certainly delayed ACKs.
221 * If we have 3 samples, we should be OK.
224 if (vegas
->cntRTT
<= 2) {
225 /* We don't have enough RTT samples to do the Vegas
226 * calculation, so we'll behave like Reno.
228 tcp_reno_cong_avoid(sk
, ack
, seq_rtt
, in_flight
, flag
);
230 u32 rtt
, target_cwnd
, diff
;
232 /* We have enough RTT samples, so, using the Vegas
233 * algorithm, we determine if we should increase or
234 * decrease cwnd, and by how much.
237 /* Pluck out the RTT we are using for the Vegas
238 * calculations. This is the min RTT seen during the
239 * last RTT. Taking the min filters out the effects
240 * of delayed ACKs, at the cost of noticing congestion
245 /* Calculate the cwnd we should have, if we weren't
249 * (actual rate in segments) * baseRTT
250 * We keep it as a fixed point number with
251 * V_PARAM_SHIFT bits to the right of the binary point.
253 target_cwnd
= ((old_wnd
* vegas
->baseRTT
)
254 << V_PARAM_SHIFT
) / rtt
;
256 /* Calculate the difference between the window we had,
257 * and the window we would like to have. This quantity
258 * is the "Diff" from the Arizona Vegas papers.
260 * Again, this is a fixed point number with
261 * V_PARAM_SHIFT bits to the right of the binary
264 diff
= (old_wnd
<< V_PARAM_SHIFT
) - target_cwnd
;
266 if (tp
->snd_cwnd
<= tp
->snd_ssthresh
) {
269 /* Going too fast. Time to slow down
270 * and switch to congestion avoidance.
272 tp
->snd_ssthresh
= 2;
274 /* Set cwnd to match the actual rate
276 * cwnd = (actual rate) * baseRTT
277 * Then we add 1 because the integer
278 * truncation robs us of full link
281 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
288 /* Congestion avoidance. */
291 /* Figure out where we would like cwnd
295 /* The old window was too fast, so
298 next_snd_cwnd
= old_snd_cwnd
- 1;
299 } else if (diff
< alpha
) {
300 /* We don't have enough extra packets
301 * in the network, so speed up.
303 next_snd_cwnd
= old_snd_cwnd
+ 1;
305 /* Sending just as fast as we
308 next_snd_cwnd
= old_snd_cwnd
;
311 /* Adjust cwnd upward or downward, toward the
314 if (next_snd_cwnd
> tp
->snd_cwnd
)
316 else if (next_snd_cwnd
< tp
->snd_cwnd
)
320 if (tp
->snd_cwnd
< 2)
322 else if (tp
->snd_cwnd
> tp
->snd_cwnd_clamp
)
323 tp
->snd_cwnd
= tp
->snd_cwnd_clamp
;
326 /* Wipe the slate clean for the next RTT. */
328 vegas
->minRTT
= 0x7fffffff;
330 /* Use normal slow start */
331 else if (tp
->snd_cwnd
<= tp
->snd_ssthresh
)
336 /* Extract info for Tcp socket info provided via netlink. */
337 void tcp_vegas_get_info(struct sock
*sk
, u32 ext
, struct sk_buff
*skb
)
339 const struct vegas
*ca
= inet_csk_ca(sk
);
340 if (ext
& (1 << (INET_DIAG_VEGASINFO
- 1))) {
341 struct tcpvegas_info info
= {
342 .tcpv_enabled
= ca
->doing_vegas_now
,
343 .tcpv_rttcnt
= ca
->cntRTT
,
344 .tcpv_rtt
= ca
->baseRTT
,
345 .tcpv_minrtt
= ca
->minRTT
,
348 nla_put(skb
, INET_DIAG_VEGASINFO
, sizeof(info
), &info
);
351 EXPORT_SYMBOL_GPL(tcp_vegas_get_info
);
353 static struct tcp_congestion_ops tcp_vegas
= {
354 .flags
= TCP_CONG_RTT_STAMP
,
355 .init
= tcp_vegas_init
,
356 .ssthresh
= tcp_reno_ssthresh
,
357 .cong_avoid
= tcp_vegas_cong_avoid
,
358 .min_cwnd
= tcp_reno_min_cwnd
,
359 .pkts_acked
= tcp_vegas_pkts_acked
,
360 .set_state
= tcp_vegas_state
,
361 .cwnd_event
= tcp_vegas_cwnd_event
,
362 .get_info
= tcp_vegas_get_info
,
364 .owner
= THIS_MODULE
,
368 static int __init
tcp_vegas_register(void)
370 BUILD_BUG_ON(sizeof(struct vegas
) > ICSK_CA_PRIV_SIZE
);
371 tcp_register_congestion_control(&tcp_vegas
);
375 static void __exit
tcp_vegas_unregister(void)
377 tcp_unregister_congestion_control(&tcp_vegas
);
380 module_init(tcp_vegas_register
);
381 module_exit(tcp_vegas_unregister
);
383 MODULE_AUTHOR("Stephen Hemminger");
384 MODULE_LICENSE("GPL");
385 MODULE_DESCRIPTION("TCP Vegas");