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.
34 * TCP Compound based on TCP Vegas
36 * further details can be found here:
37 * ftp://ftp.research.microsoft.com/pub/tr/TR-2005-86.pdf
40 #include <linux/config.h>
42 #include <linux/module.h>
43 #include <linux/skbuff.h>
44 #include <linux/inet_diag.h>
48 /* Default values of the Vegas variables, in fixed-point representation
49 * with V_PARAM_SHIFT bits to the right of the binary point.
51 #define V_PARAM_SHIFT 1
53 #define TCP_COMPOUND_ALPHA 3U
54 #define TCP_COMPOUND_BETA 1U
55 #define TCP_COMPOUND_KAPPA_POW 3
56 #define TCP_COMPOUND_KAPPA_NSQRT 2
57 #define TCP_COMPOUND_GAMMA 30
58 #define TCP_COMPOUND_ZETA 1
60 /* TCP compound variables */
62 u32 beg_snd_nxt; /* right edge during last RTT */
63 u32 beg_snd_una; /* left edge during last RTT */
64 u32 beg_snd_cwnd; /* saves the size of the cwnd */
65 u8 doing_vegas_now; /* if true, do vegas for this RTT */
66 u16 cntRTT; /* # of RTTs measured within last RTT */
67 u32 minRTT; /* min of RTTs measured within last RTT (in usec) */
68 u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */
74 /* There are several situations when we must "re-start" Vegas:
76 * o when a connection is established
78 * o after fast recovery
79 * o when we send a packet and there is no outstanding
80 * unacknowledged data (restarting an idle connection)
82 * In these circumstances we cannot do a Vegas calculation at the
83 * end of the first RTT, because any calculation we do is using
84 * stale info -- both the saved cwnd and congestion feedback are
87 * Instead we must wait until the completion of an RTT during
88 * which we actually receive ACKs.
90 static inline void vegas_enable(struct sock *sk)
92 const struct tcp_sock *tp = tcp_sk(sk);
93 struct compound *vegas = inet_csk_ca(sk);
95 /* Begin taking Vegas samples next time we send something. */
96 vegas->doing_vegas_now = 1;
98 /* Set the beginning of the next send window. */
99 vegas->beg_snd_nxt = tp->snd_nxt;
102 vegas->minRTT = 0x7fffffff;
105 /* Stop taking Vegas samples for now. */
106 static inline void vegas_disable(struct sock *sk)
108 struct compound *vegas = inet_csk_ca(sk);
110 vegas->doing_vegas_now = 0;
113 static void tcp_compound_init(struct sock *sk)
115 struct compound *vegas = inet_csk_ca(sk);
116 const struct tcp_sock *tp = tcp_sk(sk);
118 vegas->baseRTT = 0x7fffffff;
122 vegas->cwnd = tp->snd_cwnd;
125 /* Do RTT sampling needed for Vegas.
127 * o min-filter RTT samples from within an RTT to get the current
128 * propagation delay + queuing delay (we are min-filtering to try to
129 * avoid the effects of delayed ACKs)
130 * o min-filter RTT samples from a much longer window (forever for now)
131 * to find the propagation delay (baseRTT)
133 static void tcp_compound_rtt_calc(struct sock *sk, u32 usrtt)
135 struct compound *vegas = inet_csk_ca(sk);
136 u32 vrtt = usrtt + 1; /* Never allow zero rtt or baseRTT */
138 /* Filter to find propagation delay: */
139 if (vrtt < vegas->baseRTT)
140 vegas->baseRTT = vrtt;
142 /* Find the min RTT during the last RTT to find
143 * the current prop. delay + queuing delay:
146 vegas->minRTT = min(vegas->minRTT, vrtt);
150 static void tcp_compound_state(struct sock *sk, u8 ca_state)
153 if (ca_state == TCP_CA_Open)
160 * If the connection is idle and we are restarting,
161 * then we don't want to do any Vegas calculations
162 * until we get fresh RTT samples. So when we
163 * restart, we reset our Vegas state to a clean
164 * slate. After we get acks for this flight of
165 * packets, _then_ we can make Vegas calculations
168 static void tcp_compound_cwnd_event(struct sock *sk, enum tcp_ca_event event)
170 if (event == CA_EVENT_CWND_RESTART || event == CA_EVENT_TX_START)
171 tcp_compound_init(sk);
174 static void tcp_compound_cong_avoid(struct sock *sk, u32 ack,
175 u32 seq_rtt, u32 in_flight, int flag)
177 struct tcp_sock *tp = tcp_sk(sk);
178 struct compound *vegas = inet_csk_ca(sk);
181 if (vegas->cwnd + vegas->dwnd > tp->snd_cwnd) {
182 if (vegas->cwnd > tp->snd_cwnd || vegas->dwnd > tp->snd_cwnd) {
183 vegas->cwnd = tp->snd_cwnd;
186 vegas->cwnd = tp->snd_cwnd - vegas->dwnd;
190 if (!tcp_is_cwnd_limited(sk, in_flight))
193 if (vegas->cwnd <= tp->snd_ssthresh)
195 else if (tp->snd_cwnd_cnt < tp->snd_cwnd)
198 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
200 tp->snd_cwnd_cnt = 0;
203 if (inc && tp->snd_cwnd < tp->snd_cwnd_clamp)
206 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
208 * These are so named because they represent the approximate values
209 * of snd_una and snd_nxt at the beginning of the current RTT. More
210 * precisely, they represent the amount of data sent during the RTT.
211 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
212 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
213 * bytes of data have been ACKed during the course of the RTT, giving
214 * an "actual" rate of:
216 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
218 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
219 * because delayed ACKs can cover more than one segment, so they
220 * don't line up nicely with the boundaries of RTTs.
222 * Another unfortunate fact of life is that delayed ACKs delay the
223 * advance of the left edge of our send window, so that the number
224 * of bytes we send in an RTT is often less than our cwnd will allow.
225 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
228 if (after(ack, vegas->beg_snd_nxt)) {
229 /* Do the Vegas once-per-RTT cwnd adjustment. */
230 u32 old_wnd, old_snd_cwnd;
232 /* Here old_wnd is essentially the window of data that was
233 * sent during the previous RTT, and has all
234 * been acknowledged in the course of the RTT that ended
235 * with the ACK we just received. Likewise, old_snd_cwnd
236 * is the cwnd during the previous RTT.
241 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
243 old_snd_cwnd = vegas->beg_snd_cwnd;
245 /* Save the extent of the current window so we can use this
246 * at the end of the next RTT.
248 vegas->beg_snd_una = vegas->beg_snd_nxt;
249 vegas->beg_snd_nxt = tp->snd_nxt;
250 vegas->beg_snd_cwnd = tp->snd_cwnd;
252 /* We do the Vegas calculations only if we got enough RTT
253 * samples that we can be reasonably sure that we got
254 * at least one RTT sample that wasn't from a delayed ACK.
255 * If we only had 2 samples total,
256 * then that means we're getting only 1 ACK per RTT, which
257 * means they're almost certainly delayed ACKs.
258 * If we have 3 samples, we should be OK.
261 if (vegas->cntRTT > 2) {
262 u32 rtt, target_cwnd, diff;
265 /* We have enough RTT samples, so, using the Vegas
266 * algorithm, we determine if we should increase or
267 * decrease cwnd, and by how much.
270 /* Pluck out the RTT we are using for the Vegas
271 * calculations. This is the min RTT seen during the
272 * last RTT. Taking the min filters out the effects
273 * of delayed ACKs, at the cost of noticing congestion
278 /* Calculate the cwnd we should have, if we weren't
282 * (actual rate in segments) * baseRTT
283 * We keep it as a fixed point number with
284 * V_PARAM_SHIFT bits to the right of the binary point.
289 brtt = vegas->baseRTT;
290 target_cwnd = ((old_wnd * brtt)
291 << V_PARAM_SHIFT) / rtt;
293 /* Calculate the difference between the window we had,
294 * and the window we would like to have. This quantity
295 * is the "Diff" from the Arizona Vegas papers.
297 * Again, this is a fixed point number with
298 * V_PARAM_SHIFT bits to the right of the binary
302 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
306 if (diff < (TCP_COMPOUND_GAMMA << V_PARAM_SHIFT)) {
312 for (i = 0; i < TCP_COMPOUND_KAPPA_POW; i++)
315 for (i = 0; i < TCP_COMPOUND_KAPPA_NSQRT; i++) {
317 for (j = 0; j < 200; j++) {
318 x2 = (x + v / x) / 2;
328 x = (u32) v >> TCP_COMPOUND_ALPHA;
337 } else if ((dwnd << V_PARAM_SHIFT) <
338 (diff * TCP_COMPOUND_BETA))
342 ((dwnd << V_PARAM_SHIFT) -
344 TCP_COMPOUND_BETA)) >> V_PARAM_SHIFT;
350 /* Wipe the slate clean for the next RTT. */
352 vegas->minRTT = 0x7fffffff;
355 tp->snd_cwnd = vegas->cwnd + vegas->dwnd;
358 /* Extract info for Tcp socket info provided via netlink. */
359 static void tcp_compound_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
361 const struct compound *ca = inet_csk_ca(sk);
362 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
363 struct tcpvegas_info *info;
365 info = RTA_DATA(__RTA_PUT(skb, INET_DIAG_VEGASINFO,
368 info->tcpv_enabled = ca->doing_vegas_now;
369 info->tcpv_rttcnt = ca->cntRTT;
370 info->tcpv_rtt = ca->baseRTT;
371 info->tcpv_minrtt = ca->minRTT;
376 static struct tcp_congestion_ops tcp_compound = {
377 .init = tcp_compound_init,
378 .ssthresh = tcp_reno_ssthresh,
379 .cong_avoid = tcp_compound_cong_avoid,
380 .min_cwnd = tcp_reno_min_cwnd,
381 .rtt_sample = tcp_compound_rtt_calc,
382 .set_state = tcp_compound_state,
383 .cwnd_event = tcp_compound_cwnd_event,
384 .get_info = tcp_compound_get_info,
386 .owner = THIS_MODULE,
390 static int __init tcp_compound_register(void)
392 BUG_ON(sizeof(struct compound) > ICSK_CA_PRIV_SIZE);
393 tcp_register_congestion_control(&tcp_compound);
397 static void __exit tcp_compound_unregister(void)
399 tcp_unregister_congestion_control(&tcp_compound);
402 module_init(tcp_compound_register);
403 module_exit(tcp_compound_unregister);
405 MODULE_AUTHOR("Angelo P. Castellani, Stephen Hemminger");
406 MODULE_LICENSE("GPL");
407 MODULE_DESCRIPTION("TCP Compound");