xref: /openbmc/linux/net/ipv4/tcp_illinois.c (revision d0e22329)
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://tamerbasar.csl.illinois.edu/LiuBasarSrikantPerfEvalArtJun2008.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 */
26 #define RTT_MAX		(U32_MAX / ALPHA_MAX)	/* 3.3 secs */
27 
28 #define BETA_SHIFT	6
29 #define BETA_SCALE	(1u<<BETA_SHIFT)
30 #define BETA_MIN	(BETA_SCALE/8)		/* 0.125 */
31 #define BETA_MAX	(BETA_SCALE/2)		/* 0.5 */
32 #define BETA_BASE	BETA_MAX
33 
34 static int win_thresh __read_mostly = 15;
35 module_param(win_thresh, int, 0);
36 MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing");
37 
38 static int theta __read_mostly = 5;
39 module_param(theta, int, 0);
40 MODULE_PARM_DESC(theta, "# of fast RTT's before full growth");
41 
42 /* TCP Illinois Parameters */
43 struct illinois {
44 	u64	sum_rtt;	/* sum of rtt's measured within last rtt */
45 	u16	cnt_rtt;	/* # of rtts measured within last rtt */
46 	u32	base_rtt;	/* min of all rtt in usec */
47 	u32	max_rtt;	/* max of all rtt in usec */
48 	u32	end_seq;	/* right edge of current RTT */
49 	u32	alpha;		/* Additive increase */
50 	u32	beta;		/* Muliplicative decrease */
51 	u16	acked;		/* # packets acked by current ACK */
52 	u8	rtt_above;	/* average rtt has gone above threshold */
53 	u8	rtt_low;	/* # of rtts measurements below threshold */
54 };
55 
56 static void rtt_reset(struct sock *sk)
57 {
58 	struct tcp_sock *tp = tcp_sk(sk);
59 	struct illinois *ca = inet_csk_ca(sk);
60 
61 	ca->end_seq = tp->snd_nxt;
62 	ca->cnt_rtt = 0;
63 	ca->sum_rtt = 0;
64 
65 	/* TODO: age max_rtt? */
66 }
67 
68 static void tcp_illinois_init(struct sock *sk)
69 {
70 	struct illinois *ca = inet_csk_ca(sk);
71 
72 	ca->alpha = ALPHA_MAX;
73 	ca->beta = BETA_BASE;
74 	ca->base_rtt = 0x7fffffff;
75 	ca->max_rtt = 0;
76 
77 	ca->acked = 0;
78 	ca->rtt_low = 0;
79 	ca->rtt_above = 0;
80 
81 	rtt_reset(sk);
82 }
83 
84 /* Measure RTT for each ack. */
85 static void tcp_illinois_acked(struct sock *sk, const struct ack_sample *sample)
86 {
87 	struct illinois *ca = inet_csk_ca(sk);
88 	s32 rtt_us = sample->rtt_us;
89 
90 	ca->acked = sample->pkts_acked;
91 
92 	/* dup ack, no rtt sample */
93 	if (rtt_us < 0)
94 		return;
95 
96 	/* ignore bogus values, this prevents wraparound in alpha math */
97 	if (rtt_us > RTT_MAX)
98 		rtt_us = RTT_MAX;
99 
100 	/* keep track of minimum RTT seen so far */
101 	if (ca->base_rtt > rtt_us)
102 		ca->base_rtt = rtt_us;
103 
104 	/* and max */
105 	if (ca->max_rtt < rtt_us)
106 		ca->max_rtt = rtt_us;
107 
108 	++ca->cnt_rtt;
109 	ca->sum_rtt += rtt_us;
110 }
111 
112 /* Maximum queuing delay */
113 static inline u32 max_delay(const struct illinois *ca)
114 {
115 	return ca->max_rtt - ca->base_rtt;
116 }
117 
118 /* Average queuing delay */
119 static inline u32 avg_delay(const struct illinois *ca)
120 {
121 	u64 t = ca->sum_rtt;
122 
123 	do_div(t, ca->cnt_rtt);
124 	return t - ca->base_rtt;
125 }
126 
127 /*
128  * Compute value of alpha used for additive increase.
129  * If small window then use 1.0, equivalent to Reno.
130  *
131  * For larger windows, adjust based on average delay.
132  * A. If average delay is at minimum (we are uncongested),
133  *    then use large alpha (10.0) to increase faster.
134  * B. If average delay is at maximum (getting congested)
135  *    then use small alpha (0.3)
136  *
137  * The result is a convex window growth curve.
138  */
139 static u32 alpha(struct illinois *ca, u32 da, u32 dm)
140 {
141 	u32 d1 = dm / 100;	/* Low threshold */
142 
143 	if (da <= d1) {
144 		/* If never got out of low delay zone, then use max */
145 		if (!ca->rtt_above)
146 			return ALPHA_MAX;
147 
148 		/* Wait for 5 good RTT's before allowing alpha to go alpha max.
149 		 * This prevents one good RTT from causing sudden window increase.
150 		 */
151 		if (++ca->rtt_low < theta)
152 			return ca->alpha;
153 
154 		ca->rtt_low = 0;
155 		ca->rtt_above = 0;
156 		return ALPHA_MAX;
157 	}
158 
159 	ca->rtt_above = 1;
160 
161 	/*
162 	 * Based on:
163 	 *
164 	 *      (dm - d1) amin amax
165 	 * k1 = -------------------
166 	 *         amax - amin
167 	 *
168 	 *       (dm - d1) amin
169 	 * k2 = ----------------  - d1
170 	 *        amax - amin
171 	 *
172 	 *             k1
173 	 * alpha = ----------
174 	 *          k2 + da
175 	 */
176 
177 	dm -= d1;
178 	da -= d1;
179 	return (dm * ALPHA_MAX) /
180 		(dm + (da  * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN);
181 }
182 
183 /*
184  * Beta used for multiplicative decrease.
185  * For small window sizes returns same value as Reno (0.5)
186  *
187  * If delay is small (10% of max) then beta = 1/8
188  * If delay is up to 80% of max then beta = 1/2
189  * In between is a linear function
190  */
191 static u32 beta(u32 da, u32 dm)
192 {
193 	u32 d2, d3;
194 
195 	d2 = dm / 10;
196 	if (da <= d2)
197 		return BETA_MIN;
198 
199 	d3 = (8 * dm) / 10;
200 	if (da >= d3 || d3 <= d2)
201 		return BETA_MAX;
202 
203 	/*
204 	 * Based on:
205 	 *
206 	 *       bmin d3 - bmax d2
207 	 * k3 = -------------------
208 	 *         d3 - d2
209 	 *
210 	 *       bmax - bmin
211 	 * k4 = -------------
212 	 *         d3 - d2
213 	 *
214 	 * b = k3 + k4 da
215 	 */
216 	return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da)
217 		/ (d3 - d2);
218 }
219 
220 /* Update alpha and beta values once per RTT */
221 static void update_params(struct sock *sk)
222 {
223 	struct tcp_sock *tp = tcp_sk(sk);
224 	struct illinois *ca = inet_csk_ca(sk);
225 
226 	if (tp->snd_cwnd < win_thresh) {
227 		ca->alpha = ALPHA_BASE;
228 		ca->beta = BETA_BASE;
229 	} else if (ca->cnt_rtt > 0) {
230 		u32 dm = max_delay(ca);
231 		u32 da = avg_delay(ca);
232 
233 		ca->alpha = alpha(ca, da, dm);
234 		ca->beta = beta(da, dm);
235 	}
236 
237 	rtt_reset(sk);
238 }
239 
240 /*
241  * In case of loss, reset to default values
242  */
243 static void tcp_illinois_state(struct sock *sk, u8 new_state)
244 {
245 	struct illinois *ca = inet_csk_ca(sk);
246 
247 	if (new_state == TCP_CA_Loss) {
248 		ca->alpha = ALPHA_BASE;
249 		ca->beta = BETA_BASE;
250 		ca->rtt_low = 0;
251 		ca->rtt_above = 0;
252 		rtt_reset(sk);
253 	}
254 }
255 
256 /*
257  * Increase window in response to successful acknowledgment.
258  */
259 static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 acked)
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))
269 		return;
270 
271 	/* In slow start */
272 	if (tcp_in_slow_start(tp))
273 		tcp_slow_start(tp, acked);
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 
294 static u32 tcp_illinois_ssthresh(struct sock *sk)
295 {
296 	struct tcp_sock *tp = tcp_sk(sk);
297 	struct illinois *ca = inet_csk_ca(sk);
298 
299 	/* Multiplicative decrease */
300 	return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->beta) >> BETA_SHIFT), 2U);
301 }
302 
303 /* Extract info for Tcp socket info provided via netlink. */
304 static size_t tcp_illinois_info(struct sock *sk, u32 ext, int *attr,
305 				union tcp_cc_info *info)
306 {
307 	const struct illinois *ca = inet_csk_ca(sk);
308 
309 	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
310 		info->vegas.tcpv_enabled = 1;
311 		info->vegas.tcpv_rttcnt = ca->cnt_rtt;
312 		info->vegas.tcpv_minrtt = ca->base_rtt;
313 		info->vegas.tcpv_rtt = 0;
314 
315 		if (info->vegas.tcpv_rttcnt > 0) {
316 			u64 t = ca->sum_rtt;
317 
318 			do_div(t, info->vegas.tcpv_rttcnt);
319 			info->vegas.tcpv_rtt = t;
320 		}
321 		*attr = INET_DIAG_VEGASINFO;
322 		return sizeof(struct tcpvegas_info);
323 	}
324 	return 0;
325 }
326 
327 static struct tcp_congestion_ops tcp_illinois __read_mostly = {
328 	.init		= tcp_illinois_init,
329 	.ssthresh	= tcp_illinois_ssthresh,
330 	.undo_cwnd	= tcp_reno_undo_cwnd,
331 	.cong_avoid	= tcp_illinois_cong_avoid,
332 	.set_state	= tcp_illinois_state,
333 	.get_info	= tcp_illinois_info,
334 	.pkts_acked	= tcp_illinois_acked,
335 
336 	.owner		= THIS_MODULE,
337 	.name		= "illinois",
338 };
339 
340 static int __init tcp_illinois_register(void)
341 {
342 	BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE);
343 	return tcp_register_congestion_control(&tcp_illinois);
344 }
345 
346 static void __exit tcp_illinois_unregister(void)
347 {
348 	tcp_unregister_congestion_control(&tcp_illinois);
349 }
350 
351 module_init(tcp_illinois_register);
352 module_exit(tcp_illinois_unregister);
353 
354 MODULE_AUTHOR("Stephen Hemminger, Shao Liu");
355 MODULE_LICENSE("GPL");
356 MODULE_DESCRIPTION("TCP Illinois");
357 MODULE_VERSION("1.0");
358