[TCP]: Congestion control API update.
[deliverable/linux.git] / net / ipv4 / tcp_illinois.c
CommitLineData
c462238d
SH
1/*
2 * TCP Illinois congestion control.
3 * Home page:
4 * http://www.ews.uiuc.edu/~shaoliu/tcpillinois/index.html
5 *
6 * The algorithm is described in:
7 * "TCP-Illinois: A Loss and Delay-Based Congestion Control Algorithm
8 * for High-Speed Networks"
9 * http://www.ews.uiuc.edu/~shaoliu/papersandslides/liubassri06perf.pdf
10 *
11 * Implemented from description in paper and ns-2 simulation.
12 * Copyright (C) 2007 Stephen Hemminger <shemminger@linux-foundation.org>
13 */
14
15#include <linux/module.h>
16#include <linux/skbuff.h>
17#include <linux/inet_diag.h>
18#include <asm/div64.h>
19#include <net/tcp.h>
20
21#define ALPHA_SHIFT 7
22#define ALPHA_SCALE (1u<<ALPHA_SHIFT)
23#define ALPHA_MIN ((3*ALPHA_SCALE)/10) /* ~0.3 */
24#define ALPHA_MAX (10*ALPHA_SCALE) /* 10.0 */
25#define ALPHA_BASE ALPHA_SCALE /* 1.0 */
65d1b4a7
SH
26#define U32_MAX ((u32)~0U)
27#define RTT_MAX (U32_MAX / ALPHA_MAX) /* 3.3 secs */
c462238d
SH
28
29#define BETA_SHIFT 6
30#define BETA_SCALE (1u<<BETA_SHIFT)
65d1b4a7
SH
31#define BETA_MIN (BETA_SCALE/8) /* 0.125 */
32#define BETA_MAX (BETA_SCALE/2) /* 0.5 */
33#define BETA_BASE BETA_MAX
c462238d
SH
34
35static int win_thresh __read_mostly = 15;
65d1b4a7 36module_param(win_thresh, int, 0);
c462238d
SH
37MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing");
38
65d1b4a7
SH
39static int theta __read_mostly = 5;
40module_param(theta, int, 0);
41MODULE_PARM_DESC(theta, "# of fast RTT's before full growth");
c462238d
SH
42
43/* TCP Illinois Parameters */
65d1b4a7
SH
44struct illinois {
45 u64 sum_rtt; /* sum of rtt's measured within last rtt */
46 u16 cnt_rtt; /* # of rtts measured within last rtt */
47 u32 base_rtt; /* min of all rtt in usec */
48 u32 max_rtt; /* max of all rtt in usec */
49 u32 end_seq; /* right edge of current RTT */
50 u32 alpha; /* Additive increase */
51 u32 beta; /* Muliplicative decrease */
52 u16 acked; /* # packets acked by current ACK */
53 u8 rtt_above; /* average rtt has gone above threshold */
54 u8 rtt_low; /* # of rtts measurements below threshold */
c462238d
SH
55};
56
65d1b4a7
SH
57static void rtt_reset(struct sock *sk)
58{
59 struct tcp_sock *tp = tcp_sk(sk);
60 struct illinois *ca = inet_csk_ca(sk);
61
62 ca->end_seq = tp->snd_nxt;
63 ca->cnt_rtt = 0;
64 ca->sum_rtt = 0;
65
66 /* TODO: age max_rtt? */
67}
68
c462238d
SH
69static void tcp_illinois_init(struct sock *sk)
70{
65d1b4a7
SH
71 struct illinois *ca = inet_csk_ca(sk);
72
73 ca->alpha = ALPHA_MAX;
74 ca->beta = BETA_BASE;
75 ca->base_rtt = 0x7fffffff;
76 ca->max_rtt = 0;
77
78 ca->acked = 0;
79 ca->rtt_low = 0;
80 ca->rtt_above = 0;
c462238d 81
65d1b4a7 82 rtt_reset(sk);
c462238d
SH
83}
84
65d1b4a7 85/* Measure RTT for each ack. */
164891aa 86static void tcp_illinois_acked(struct sock *sk, u32 pkts_acked, ktime_t last)
c462238d 87{
65d1b4a7 88 struct illinois *ca = inet_csk_ca(sk);
164891aa
SH
89 u32 rtt;
90
91 ca->acked = pkts_acked;
92
93 rtt = ktime_to_ns(net_timedelta(last)) / NSEC_PER_USEC;
65d1b4a7
SH
94
95 /* ignore bogus values, this prevents wraparound in alpha math */
96 if (rtt > RTT_MAX)
97 rtt = RTT_MAX;
c462238d 98
65d1b4a7
SH
99 /* keep track of minimum RTT seen so far */
100 if (ca->base_rtt > rtt)
101 ca->base_rtt = rtt;
102
103 /* and max */
104 if (ca->max_rtt < rtt)
c462238d
SH
105 ca->max_rtt = rtt;
106
65d1b4a7
SH
107 ++ca->cnt_rtt;
108 ca->sum_rtt += rtt;
c462238d
SH
109}
110
65d1b4a7
SH
111/* Maximum queuing delay */
112static inline u32 max_delay(const struct illinois *ca)
c462238d 113{
65d1b4a7
SH
114 return ca->max_rtt - ca->base_rtt;
115}
c462238d 116
65d1b4a7
SH
117/* Average queuing delay */
118static inline u32 avg_delay(const struct illinois *ca)
119{
120 u64 t = ca->sum_rtt;
c462238d 121
65d1b4a7
SH
122 do_div(t, ca->cnt_rtt);
123 return t - ca->base_rtt;
c462238d
SH
124}
125
126/*
127 * Compute value of alpha used for additive increase.
128 * If small window then use 1.0, equivalent to Reno.
129 *
130 * For larger windows, adjust based on average delay.
131 * A. If average delay is at minimum (we are uncongested),
132 * then use large alpha (10.0) to increase faster.
133 * B. If average delay is at maximum (getting congested)
65d1b4a7 134 * then use small alpha (0.3)
c462238d
SH
135 *
136 * The result is a convex window growth curve.
137 */
65d1b4a7 138static u32 alpha(struct illinois *ca, u32 da, u32 dm)
c462238d 139{
65d1b4a7 140 u32 d1 = dm / 100; /* Low threshold */
c462238d 141
c462238d 142 if (da <= d1) {
65d1b4a7
SH
143 /* If never got out of low delay zone, then use max */
144 if (!ca->rtt_above)
c462238d 145 return ALPHA_MAX;
65d1b4a7
SH
146
147 /* Wait for 5 good RTT's before allowing alpha to go alpha max.
148 * This prevents one good RTT from causing sudden window increase.
149 */
150 if (++ca->rtt_low < theta)
151 return ca->alpha;
152
153 ca->rtt_low = 0;
154 ca->rtt_above = 0;
155 return ALPHA_MAX;
c462238d
SH
156 }
157
65d1b4a7 158 ca->rtt_above = 1;
c462238d
SH
159
160 /*
161 * Based on:
162 *
163 * (dm - d1) amin amax
164 * k1 = -------------------
165 * amax - amin
166 *
167 * (dm - d1) amin
168 * k2 = ---------------- - d1
169 * amax - amin
170 *
171 * k1
172 * alpha = ----------
173 * k2 + da
174 */
175
176 dm -= d1;
177 da -= d1;
65d1b4a7
SH
178 return (dm * ALPHA_MAX) /
179 (dm + (da * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN);
c462238d
SH
180}
181
182/*
183 * Beta used for multiplicative decrease.
184 * For small window sizes returns same value as Reno (0.5)
185 *
186 * If delay is small (10% of max) then beta = 1/8
187 * If delay is up to 80% of max then beta = 1/2
188 * In between is a linear function
189 */
65d1b4a7 190static u32 beta(u32 da, u32 dm)
c462238d 191{
c462238d
SH
192 u32 d2, d3;
193
c462238d
SH
194 d2 = dm / 10;
195 if (da <= d2)
196 return BETA_MIN;
65d1b4a7 197
c462238d
SH
198 d3 = (8 * dm) / 10;
199 if (da >= d3 || d3 <= d2)
200 return BETA_MAX;
201
202 /*
203 * Based on:
204 *
205 * bmin d3 - bmax d2
206 * k3 = -------------------
207 * d3 - d2
208 *
209 * bmax - bmin
210 * k4 = -------------
211 * d3 - d2
212 *
213 * b = k3 + k4 da
214 */
215 return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da)
216 / (d3 - d2);
217}
218
65d1b4a7
SH
219/* Update alpha and beta values once per RTT */
220static void update_params(struct sock *sk)
221{
222 struct tcp_sock *tp = tcp_sk(sk);
223 struct illinois *ca = inet_csk_ca(sk);
224
225 if (tp->snd_cwnd < win_thresh) {
226 ca->alpha = ALPHA_BASE;
227 ca->beta = BETA_BASE;
228 } else if (ca->cnt_rtt > 0) {
229 u32 dm = max_delay(ca);
230 u32 da = avg_delay(ca);
231
232 ca->alpha = alpha(ca, da, dm);
233 ca->beta = beta(da, dm);
234 }
235
236 rtt_reset(sk);
237}
238
239/*
240 * In case of loss, reset to default values
241 */
242static void tcp_illinois_state(struct sock *sk, u8 new_state)
243{
244 struct illinois *ca = inet_csk_ca(sk);
245
246 if (new_state == TCP_CA_Loss) {
247 ca->alpha = ALPHA_BASE;
248 ca->beta = BETA_BASE;
249 ca->rtt_low = 0;
250 ca->rtt_above = 0;
251 rtt_reset(sk);
252 }
253}
254
255/*
256 * Increase window in response to successful acknowledgment.
257 */
258static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
259 u32 in_flight, int flag)
260{
261 struct tcp_sock *tp = tcp_sk(sk);
262 struct illinois *ca = inet_csk_ca(sk);
263
264 if (after(ack, ca->end_seq))
265 update_params(sk);
266
267 /* RFC2861 only increase cwnd if fully utilized */
268 if (!tcp_is_cwnd_limited(sk, in_flight))
269 return;
270
271 /* In slow start */
272 if (tp->snd_cwnd <= tp->snd_ssthresh)
273 tcp_slow_start(tp);
274
275 else {
276 u32 delta;
277
278 /* snd_cwnd_cnt is # of packets since last cwnd increment */
279 tp->snd_cwnd_cnt += ca->acked;
280 ca->acked = 1;
281
282 /* This is close approximation of:
283 * tp->snd_cwnd += alpha/tp->snd_cwnd
284 */
285 delta = (tp->snd_cwnd_cnt * ca->alpha) >> ALPHA_SHIFT;
286 if (delta >= tp->snd_cwnd) {
287 tp->snd_cwnd = min(tp->snd_cwnd + delta / tp->snd_cwnd,
288 (u32) tp->snd_cwnd_clamp);
289 tp->snd_cwnd_cnt = 0;
290 }
291 }
292}
293
c462238d
SH
294static u32 tcp_illinois_ssthresh(struct sock *sk)
295{
296 struct tcp_sock *tp = tcp_sk(sk);
65d1b4a7 297 struct illinois *ca = inet_csk_ca(sk);
c462238d
SH
298
299 /* Multiplicative decrease */
65d1b4a7 300 return max((tp->snd_cwnd * ca->beta) >> BETA_SHIFT, 2U);
c462238d
SH
301}
302
65d1b4a7
SH
303
304/* Extract info for Tcp socket info provided via netlink. */
305static void tcp_illinois_info(struct sock *sk, u32 ext,
306 struct sk_buff *skb)
c462238d 307{
65d1b4a7 308 const struct illinois *ca = inet_csk_ca(sk);
c462238d
SH
309
310 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
311 struct tcpvegas_info info = {
312 .tcpv_enabled = 1,
65d1b4a7
SH
313 .tcpv_rttcnt = ca->cnt_rtt,
314 .tcpv_minrtt = ca->base_rtt,
c462238d 315 };
65d1b4a7
SH
316 u64 t = ca->sum_rtt;
317
318 do_div(t, ca->cnt_rtt);
319 info.tcpv_rtt = t;
c462238d
SH
320
321 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
322 }
323}
324
325static struct tcp_congestion_ops tcp_illinois = {
164891aa 326 .flags = TCP_CONG_RTT_STAMP,
c462238d
SH
327 .init = tcp_illinois_init,
328 .ssthresh = tcp_illinois_ssthresh,
329 .min_cwnd = tcp_reno_min_cwnd,
330 .cong_avoid = tcp_illinois_cong_avoid,
65d1b4a7 331 .set_state = tcp_illinois_state,
65d1b4a7
SH
332 .get_info = tcp_illinois_info,
333 .pkts_acked = tcp_illinois_acked,
c462238d
SH
334
335 .owner = THIS_MODULE,
336 .name = "illinois",
337};
338
339static int __init tcp_illinois_register(void)
340{
65d1b4a7 341 BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE);
c462238d
SH
342 return tcp_register_congestion_control(&tcp_illinois);
343}
344
345static void __exit tcp_illinois_unregister(void)
346{
347 tcp_unregister_congestion_control(&tcp_illinois);
348}
349
350module_init(tcp_illinois_register);
351module_exit(tcp_illinois_unregister);
352
353MODULE_AUTHOR("Stephen Hemminger, Shao Liu");
354MODULE_LICENSE("GPL");
355MODULE_DESCRIPTION("TCP Illinois");
65d1b4a7 356MODULE_VERSION("1.0");
This page took 0.101502 seconds and 5 git commands to generate.