xref: /openbmc/linux/net/ipv4/tcp_illinois.c (revision 3a83e4e6)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * TCP Illinois congestion control.
4  * Home page:
5  *	http://www.ews.uiuc.edu/~shaoliu/tcpillinois/index.html
6  *
7  * The algorithm is described in:
8  * "TCP-Illinois: A Loss and Delay-Based Congestion Control Algorithm
9  *  for High-Speed Networks"
10  * http://tamerbasar.csl.illinois.edu/LiuBasarSrikantPerfEvalArtJun2008.pdf
11  *
12  * Implemented from description in paper and ns-2 simulation.
13  * Copyright (C) 2007 Stephen Hemminger <shemminger@linux-foundation.org>
14  */
15 
16 #include <linux/module.h>
17 #include <linux/skbuff.h>
18 #include <linux/inet_diag.h>
19 #include <asm/div64.h>
20 #include <net/tcp.h>
21 
22 #define ALPHA_SHIFT	7
23 #define ALPHA_SCALE	(1u<<ALPHA_SHIFT)
24 #define ALPHA_MIN	((3*ALPHA_SCALE)/10)	/* ~0.3 */
25 #define ALPHA_MAX	(10*ALPHA_SCALE)	/* 10.0 */
26 #define ALPHA_BASE	ALPHA_SCALE		/* 1.0 */
27 #define RTT_MAX		(U32_MAX / ALPHA_MAX)	/* 3.3 secs */
28 
29 #define BETA_SHIFT	6
30 #define BETA_SCALE	(1u<<BETA_SHIFT)
31 #define BETA_MIN	(BETA_SCALE/8)		/* 0.125 */
32 #define BETA_MAX	(BETA_SCALE/2)		/* 0.5 */
33 #define BETA_BASE	BETA_MAX
34 
35 static int win_thresh __read_mostly = 15;
36 module_param(win_thresh, int, 0);
37 MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing");
38 
39 static int theta __read_mostly = 5;
40 module_param(theta, int, 0);
41 MODULE_PARM_DESC(theta, "# of fast RTT's before full growth");
42 
43 /* TCP Illinois Parameters */
44 struct 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 */
55 };
56 
57 static 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 
69 static void tcp_illinois_init(struct sock *sk)
70 {
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;
81 
82 	rtt_reset(sk);
83 }
84 
85 /* Measure RTT for each ack. */
86 static void tcp_illinois_acked(struct sock *sk, const struct ack_sample *sample)
87 {
88 	struct illinois *ca = inet_csk_ca(sk);
89 	s32 rtt_us = sample->rtt_us;
90 
91 	ca->acked = sample->pkts_acked;
92 
93 	/* dup ack, no rtt sample */
94 	if (rtt_us < 0)
95 		return;
96 
97 	/* ignore bogus values, this prevents wraparound in alpha math */
98 	if (rtt_us > RTT_MAX)
99 		rtt_us = RTT_MAX;
100 
101 	/* keep track of minimum RTT seen so far */
102 	if (ca->base_rtt > rtt_us)
103 		ca->base_rtt = rtt_us;
104 
105 	/* and max */
106 	if (ca->max_rtt < rtt_us)
107 		ca->max_rtt = rtt_us;
108 
109 	++ca->cnt_rtt;
110 	ca->sum_rtt += rtt_us;
111 }
112 
113 /* Maximum queuing delay */
114 static inline u32 max_delay(const struct illinois *ca)
115 {
116 	return ca->max_rtt - ca->base_rtt;
117 }
118 
119 /* Average queuing delay */
120 static inline u32 avg_delay(const struct illinois *ca)
121 {
122 	u64 t = ca->sum_rtt;
123 
124 	do_div(t, ca->cnt_rtt);
125 	return t - ca->base_rtt;
126 }
127 
128 /*
129  * Compute value of alpha used for additive increase.
130  * If small window then use 1.0, equivalent to Reno.
131  *
132  * For larger windows, adjust based on average delay.
133  * A. If average delay is at minimum (we are uncongested),
134  *    then use large alpha (10.0) to increase faster.
135  * B. If average delay is at maximum (getting congested)
136  *    then use small alpha (0.3)
137  *
138  * The result is a convex window growth curve.
139  */
140 static u32 alpha(struct illinois *ca, u32 da, u32 dm)
141 {
142 	u32 d1 = dm / 100;	/* Low threshold */
143 
144 	if (da <= d1) {
145 		/* If never got out of low delay zone, then use max */
146 		if (!ca->rtt_above)
147 			return ALPHA_MAX;
148 
149 		/* Wait for 5 good RTT's before allowing alpha to go alpha max.
150 		 * This prevents one good RTT from causing sudden window increase.
151 		 */
152 		if (++ca->rtt_low < theta)
153 			return ca->alpha;
154 
155 		ca->rtt_low = 0;
156 		ca->rtt_above = 0;
157 		return ALPHA_MAX;
158 	}
159 
160 	ca->rtt_above = 1;
161 
162 	/*
163 	 * Based on:
164 	 *
165 	 *      (dm - d1) amin amax
166 	 * k1 = -------------------
167 	 *         amax - amin
168 	 *
169 	 *       (dm - d1) amin
170 	 * k2 = ----------------  - d1
171 	 *        amax - amin
172 	 *
173 	 *             k1
174 	 * alpha = ----------
175 	 *          k2 + da
176 	 */
177 
178 	dm -= d1;
179 	da -= d1;
180 	return (dm * ALPHA_MAX) /
181 		(dm + (da  * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN);
182 }
183 
184 /*
185  * Beta used for multiplicative decrease.
186  * For small window sizes returns same value as Reno (0.5)
187  *
188  * If delay is small (10% of max) then beta = 1/8
189  * If delay is up to 80% of max then beta = 1/2
190  * In between is a linear function
191  */
192 static u32 beta(u32 da, u32 dm)
193 {
194 	u32 d2, d3;
195 
196 	d2 = dm / 10;
197 	if (da <= d2)
198 		return BETA_MIN;
199 
200 	d3 = (8 * dm) / 10;
201 	if (da >= d3 || d3 <= d2)
202 		return BETA_MAX;
203 
204 	/*
205 	 * Based on:
206 	 *
207 	 *       bmin d3 - bmax d2
208 	 * k3 = -------------------
209 	 *         d3 - d2
210 	 *
211 	 *       bmax - bmin
212 	 * k4 = -------------
213 	 *         d3 - d2
214 	 *
215 	 * b = k3 + k4 da
216 	 */
217 	return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da)
218 		/ (d3 - d2);
219 }
220 
221 /* Update alpha and beta values once per RTT */
222 static void update_params(struct sock *sk)
223 {
224 	struct tcp_sock *tp = tcp_sk(sk);
225 	struct illinois *ca = inet_csk_ca(sk);
226 
227 	if (tp->snd_cwnd < win_thresh) {
228 		ca->alpha = ALPHA_BASE;
229 		ca->beta = BETA_BASE;
230 	} else if (ca->cnt_rtt > 0) {
231 		u32 dm = max_delay(ca);
232 		u32 da = avg_delay(ca);
233 
234 		ca->alpha = alpha(ca, da, dm);
235 		ca->beta = beta(da, dm);
236 	}
237 
238 	rtt_reset(sk);
239 }
240 
241 /*
242  * In case of loss, reset to default values
243  */
244 static void tcp_illinois_state(struct sock *sk, u8 new_state)
245 {
246 	struct illinois *ca = inet_csk_ca(sk);
247 
248 	if (new_state == TCP_CA_Loss) {
249 		ca->alpha = ALPHA_BASE;
250 		ca->beta = BETA_BASE;
251 		ca->rtt_low = 0;
252 		ca->rtt_above = 0;
253 		rtt_reset(sk);
254 	}
255 }
256 
257 /*
258  * Increase window in response to successful acknowledgment.
259  */
260 static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 acked)
261 {
262 	struct tcp_sock *tp = tcp_sk(sk);
263 	struct illinois *ca = inet_csk_ca(sk);
264 
265 	if (after(ack, ca->end_seq))
266 		update_params(sk);
267 
268 	/* RFC2861 only increase cwnd if fully utilized */
269 	if (!tcp_is_cwnd_limited(sk))
270 		return;
271 
272 	/* In slow start */
273 	if (tcp_in_slow_start(tp))
274 		tcp_slow_start(tp, acked);
275 
276 	else {
277 		u32 delta;
278 
279 		/* snd_cwnd_cnt is # of packets since last cwnd increment */
280 		tp->snd_cwnd_cnt += ca->acked;
281 		ca->acked = 1;
282 
283 		/* This is close approximation of:
284 		 * tp->snd_cwnd += alpha/tp->snd_cwnd
285 		*/
286 		delta = (tp->snd_cwnd_cnt * ca->alpha) >> ALPHA_SHIFT;
287 		if (delta >= tp->snd_cwnd) {
288 			tp->snd_cwnd = min(tp->snd_cwnd + delta / tp->snd_cwnd,
289 					   (u32)tp->snd_cwnd_clamp);
290 			tp->snd_cwnd_cnt = 0;
291 		}
292 	}
293 }
294 
295 static u32 tcp_illinois_ssthresh(struct sock *sk)
296 {
297 	struct tcp_sock *tp = tcp_sk(sk);
298 	struct illinois *ca = inet_csk_ca(sk);
299 
300 	/* Multiplicative decrease */
301 	return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->beta) >> BETA_SHIFT), 2U);
302 }
303 
304 /* Extract info for Tcp socket info provided via netlink. */
305 static size_t tcp_illinois_info(struct sock *sk, u32 ext, int *attr,
306 				union tcp_cc_info *info)
307 {
308 	const struct illinois *ca = inet_csk_ca(sk);
309 
310 	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
311 		info->vegas.tcpv_enabled = 1;
312 		info->vegas.tcpv_rttcnt = ca->cnt_rtt;
313 		info->vegas.tcpv_minrtt = ca->base_rtt;
314 		info->vegas.tcpv_rtt = 0;
315 
316 		if (info->vegas.tcpv_rttcnt > 0) {
317 			u64 t = ca->sum_rtt;
318 
319 			do_div(t, info->vegas.tcpv_rttcnt);
320 			info->vegas.tcpv_rtt = t;
321 		}
322 		*attr = INET_DIAG_VEGASINFO;
323 		return sizeof(struct tcpvegas_info);
324 	}
325 	return 0;
326 }
327 
328 static struct tcp_congestion_ops tcp_illinois __read_mostly = {
329 	.init		= tcp_illinois_init,
330 	.ssthresh	= tcp_illinois_ssthresh,
331 	.undo_cwnd	= tcp_reno_undo_cwnd,
332 	.cong_avoid	= tcp_illinois_cong_avoid,
333 	.set_state	= tcp_illinois_state,
334 	.get_info	= tcp_illinois_info,
335 	.pkts_acked	= tcp_illinois_acked,
336 
337 	.owner		= THIS_MODULE,
338 	.name		= "illinois",
339 };
340 
341 static int __init tcp_illinois_register(void)
342 {
343 	BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE);
344 	return tcp_register_congestion_control(&tcp_illinois);
345 }
346 
347 static void __exit tcp_illinois_unregister(void)
348 {
349 	tcp_unregister_congestion_control(&tcp_illinois);
350 }
351 
352 module_init(tcp_illinois_register);
353 module_exit(tcp_illinois_unregister);
354 
355 MODULE_AUTHOR("Stephen Hemminger, Shao Liu");
356 MODULE_LICENSE("GPL");
357 MODULE_DESCRIPTION("TCP Illinois");
358 MODULE_VERSION("1.0");
359