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/mm.h>
35#include <linux/module.h>
36#include <linux/skbuff.h>
37#include <linux/inet_diag.h>
38
39#include <net/tcp.h>
40
41#include "tcp_vegas.h"
42
43static int alpha = 2;
44static int beta  = 4;
45static int gamma = 1;
46
47module_param(alpha, int, 0644);
48MODULE_PARM_DESC(alpha, "lower bound of packets in network");
49module_param(beta, int, 0644);
50MODULE_PARM_DESC(beta, "upper bound of packets in network");
51module_param(gamma, int, 0644);
52MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
53
54/* There are several situations when we must "re-start" Vegas:
55 *
56 *  o when a connection is established
57 *  o after an RTO
58 *  o after fast recovery
59 *  o when we send a packet and there is no outstanding
60 *    unacknowledged data (restarting an idle connection)
61 *
62 * In these circumstances we cannot do a Vegas calculation at the
63 * end of the first RTT, because any calculation we do is using
64 * stale info -- both the saved cwnd and congestion feedback are
65 * stale.
66 *
67 * Instead we must wait until the completion of an RTT during
68 * which we actually receive ACKs.
69 */
70static void vegas_enable(struct sock *sk)
71{
72	const struct tcp_sock *tp = tcp_sk(sk);
73	struct vegas *vegas = inet_csk_ca(sk);
74
75	/* Begin taking Vegas samples next time we send something. */
76	vegas->doing_vegas_now = 1;
77
78	/* Set the beginning of the next send window. */
79	vegas->beg_snd_nxt = tp->snd_nxt;
80
81	vegas->cntRTT = 0;
82	vegas->minRTT = 0x7fffffff;
83}
84
85/* Stop taking Vegas samples for now. */
86static inline void vegas_disable(struct sock *sk)
87{
88	struct vegas *vegas = inet_csk_ca(sk);
89
90	vegas->doing_vegas_now = 0;
91}
92
93void tcp_vegas_init(struct sock *sk)
94{
95	struct vegas *vegas = inet_csk_ca(sk);
96
97	vegas->baseRTT = 0x7fffffff;
98	vegas_enable(sk);
99}
100EXPORT_SYMBOL_GPL(tcp_vegas_init);
101
102/* Do RTT sampling needed for Vegas.
103 * Basically we:
104 *   o min-filter RTT samples from within an RTT to get the current
105 *     propagation delay + queuing delay (we are min-filtering to try to
106 *     avoid the effects of delayed ACKs)
107 *   o min-filter RTT samples from a much longer window (forever for now)
108 *     to find the propagation delay (baseRTT)
109 */
110void tcp_vegas_pkts_acked(struct sock *sk, u32 cnt, s32 rtt_us)
111{
112	struct vegas *vegas = inet_csk_ca(sk);
113	u32 vrtt;
114
115	if (rtt_us < 0)
116		return;
117
118	/* Never allow zero rtt or baseRTT */
119	vrtt = rtt_us + 1;
120
121	/* Filter to find propagation delay: */
122	if (vrtt < vegas->baseRTT)
123		vegas->baseRTT = vrtt;
124
125	/* Find the min RTT during the last RTT to find
126	 * the current prop. delay + queuing delay:
127	 */
128	vegas->minRTT = min(vegas->minRTT, vrtt);
129	vegas->cntRTT++;
130}
131EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked);
132
133void tcp_vegas_state(struct sock *sk, u8 ca_state)
134{
135	if (ca_state == TCP_CA_Open)
136		vegas_enable(sk);
137	else
138		vegas_disable(sk);
139}
140EXPORT_SYMBOL_GPL(tcp_vegas_state);
141
142/*
143 * If the connection is idle and we are restarting,
144 * then we don't want to do any Vegas calculations
145 * until we get fresh RTT samples.  So when we
146 * restart, we reset our Vegas state to a clean
147 * slate. After we get acks for this flight of
148 * packets, _then_ we can make Vegas calculations
149 * again.
150 */
151void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
152{
153	if (event == CA_EVENT_CWND_RESTART ||
154	    event == CA_EVENT_TX_START)
155		tcp_vegas_init(sk);
156}
157EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
158
159static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp)
160{
161	return  min(tp->snd_ssthresh, tp->snd_cwnd-1);
162}
163
164static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked)
165{
166	struct tcp_sock *tp = tcp_sk(sk);
167	struct vegas *vegas = inet_csk_ca(sk);
168
169	if (!vegas->doing_vegas_now) {
170		tcp_reno_cong_avoid(sk, ack, acked);
171		return;
172	}
173
174	if (after(ack, vegas->beg_snd_nxt)) {
175		/* Do the Vegas once-per-RTT cwnd adjustment. */
176
177		/* Save the extent of the current window so we can use this
178		 * at the end of the next RTT.
179		 */
180		vegas->beg_snd_nxt  = tp->snd_nxt;
181
182		/* We do the Vegas calculations only if we got enough RTT
183		 * samples that we can be reasonably sure that we got
184		 * at least one RTT sample that wasn't from a delayed ACK.
185		 * If we only had 2 samples total,
186		 * then that means we're getting only 1 ACK per RTT, which
187		 * means they're almost certainly delayed ACKs.
188		 * If  we have 3 samples, we should be OK.
189		 */
190
191		if (vegas->cntRTT <= 2) {
192			/* We don't have enough RTT samples to do the Vegas
193			 * calculation, so we'll behave like Reno.
194			 */
195			tcp_reno_cong_avoid(sk, ack, acked);
196		} else {
197			u32 rtt, diff;
198			u64 target_cwnd;
199
200			/* We have enough RTT samples, so, using the Vegas
201			 * algorithm, we determine if we should increase or
202			 * decrease cwnd, and by how much.
203			 */
204
205			/* Pluck out the RTT we are using for the Vegas
206			 * calculations. This is the min RTT seen during the
207			 * last RTT. Taking the min filters out the effects
208			 * of delayed ACKs, at the cost of noticing congestion
209			 * a bit later.
210			 */
211			rtt = vegas->minRTT;
212
213			/* Calculate the cwnd we should have, if we weren't
214			 * going too fast.
215			 *
216			 * This is:
217			 *     (actual rate in segments) * baseRTT
218			 */
219			target_cwnd = (u64)tp->snd_cwnd * vegas->baseRTT;
220			do_div(target_cwnd, rtt);
221
222			/* Calculate the difference between the window we had,
223			 * and the window we would like to have. This quantity
224			 * is the "Diff" from the Arizona Vegas papers.
225			 */
226			diff = tp->snd_cwnd * (rtt-vegas->baseRTT) / vegas->baseRTT;
227
228			if (diff > gamma && tp->snd_cwnd <= tp->snd_ssthresh) {
229				/* Going too fast. Time to slow down
230				 * and switch to congestion avoidance.
231				 */
232
233				/* Set cwnd to match the actual rate
234				 * exactly:
235				 *   cwnd = (actual rate) * baseRTT
236				 * Then we add 1 because the integer
237				 * truncation robs us of full link
238				 * utilization.
239				 */
240				tp->snd_cwnd = min(tp->snd_cwnd, (u32)target_cwnd+1);
241				tp->snd_ssthresh = tcp_vegas_ssthresh(tp);
242
243			} else if (tp->snd_cwnd <= tp->snd_ssthresh) {
244				/* Slow start.  */
245				tcp_slow_start(tp, acked);
246			} else {
247				/* Congestion avoidance. */
248
249				/* Figure out where we would like cwnd
250				 * to be.
251				 */
252				if (diff > beta) {
253					/* The old window was too fast, so
254					 * we slow down.
255					 */
256					tp->snd_cwnd--;
257					tp->snd_ssthresh
258						= tcp_vegas_ssthresh(tp);
259				} else if (diff < alpha) {
260					/* We don't have enough extra packets
261					 * in the network, so speed up.
262					 */
263					tp->snd_cwnd++;
264				} else {
265					/* Sending just as fast as we
266					 * should be.
267					 */
268				}
269			}
270
271			if (tp->snd_cwnd < 2)
272				tp->snd_cwnd = 2;
273			else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
274				tp->snd_cwnd = tp->snd_cwnd_clamp;
275
276			tp->snd_ssthresh = tcp_current_ssthresh(sk);
277		}
278
279		/* Wipe the slate clean for the next RTT. */
280		vegas->cntRTT = 0;
281		vegas->minRTT = 0x7fffffff;
282	}
283	/* Use normal slow start */
284	else if (tp->snd_cwnd <= tp->snd_ssthresh)
285		tcp_slow_start(tp, acked);
286}
287
288/* Extract info for Tcp socket info provided via netlink. */
289void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
290{
291	const struct vegas *ca = inet_csk_ca(sk);
292	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
293		struct tcpvegas_info info = {
294			.tcpv_enabled = ca->doing_vegas_now,
295			.tcpv_rttcnt = ca->cntRTT,
296			.tcpv_rtt = ca->baseRTT,
297			.tcpv_minrtt = ca->minRTT,
298		};
299
300		nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
301	}
302}
303EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
304
305static struct tcp_congestion_ops tcp_vegas __read_mostly = {
306	.init		= tcp_vegas_init,
307	.ssthresh	= tcp_reno_ssthresh,
308	.cong_avoid	= tcp_vegas_cong_avoid,
309	.pkts_acked	= tcp_vegas_pkts_acked,
310	.set_state	= tcp_vegas_state,
311	.cwnd_event	= tcp_vegas_cwnd_event,
312	.get_info	= tcp_vegas_get_info,
313
314	.owner		= THIS_MODULE,
315	.name		= "vegas",
316};
317
318static int __init tcp_vegas_register(void)
319{
320	BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
321	tcp_register_congestion_control(&tcp_vegas);
322	return 0;
323}
324
325static void __exit tcp_vegas_unregister(void)
326{
327	tcp_unregister_congestion_control(&tcp_vegas);
328}
329
330module_init(tcp_vegas_register);
331module_exit(tcp_vegas_unregister);
332
333MODULE_AUTHOR("Stephen Hemminger");
334MODULE_LICENSE("GPL");
335MODULE_DESCRIPTION("TCP Vegas");
336