xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision fd589a8f)
1 /*
2  * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
3  * Home page:
4  *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
5  * This is from the implementation of CUBIC TCP in
6  * Sangtae Ha, Injong Rhee and Lisong Xu,
7  *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant"
8  *  in ACM SIGOPS Operating System Review, July 2008.
9  * Available from:
10  *  http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
11  *
12  * CUBIC integrates a new slow start algorithm, called HyStart.
13  * The details of HyStart are presented in
14  *  Sangtae Ha and Injong Rhee,
15  *  "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008.
16  * Available from:
17  *  http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
18  *
19  * All testing results are available from:
20  * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
21  *
22  * Unless CUBIC is enabled and congestion window is large
23  * this behaves the same as the original Reno.
24  */
25 
26 #include <linux/mm.h>
27 #include <linux/module.h>
28 #include <linux/math64.h>
29 #include <net/tcp.h>
30 
31 #define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
32 					 * max_cwnd = snd_cwnd * beta
33 					 */
34 #define	BICTCP_HZ		10	/* BIC HZ 2^10 = 1024 */
35 
36 /* Two methods of hybrid slow start */
37 #define HYSTART_ACK_TRAIN	0x1
38 #define HYSTART_DELAY		0x2
39 
40 /* Number of delay samples for detecting the increase of delay */
41 #define HYSTART_MIN_SAMPLES	8
42 #define HYSTART_DELAY_MIN	(2U<<3)
43 #define HYSTART_DELAY_MAX	(16U<<3)
44 #define HYSTART_DELAY_THRESH(x)	clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
45 
46 static int fast_convergence __read_mostly = 1;
47 static int beta __read_mostly = 717;	/* = 717/1024 (BICTCP_BETA_SCALE) */
48 static int initial_ssthresh __read_mostly;
49 static int bic_scale __read_mostly = 41;
50 static int tcp_friendliness __read_mostly = 1;
51 
52 static int hystart __read_mostly = 1;
53 static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
54 static int hystart_low_window __read_mostly = 16;
55 
56 static u32 cube_rtt_scale __read_mostly;
57 static u32 beta_scale __read_mostly;
58 static u64 cube_factor __read_mostly;
59 
60 /* Note parameters that are used for precomputing scale factors are read-only */
61 module_param(fast_convergence, int, 0644);
62 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
63 module_param(beta, int, 0644);
64 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
65 module_param(initial_ssthresh, int, 0644);
66 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
67 module_param(bic_scale, int, 0444);
68 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
69 module_param(tcp_friendliness, int, 0644);
70 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
71 module_param(hystart, int, 0644);
72 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
73 module_param(hystart_detect, int, 0644);
74 MODULE_PARM_DESC(hystart_detect, "hyrbrid slow start detection mechanisms"
75 		 " 1: packet-train 2: delay 3: both packet-train and delay");
76 module_param(hystart_low_window, int, 0644);
77 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
78 
79 /* BIC TCP Parameters */
80 struct bictcp {
81 	u32	cnt;		/* increase cwnd by 1 after ACKs */
82 	u32 	last_max_cwnd;	/* last maximum snd_cwnd */
83 	u32	loss_cwnd;	/* congestion window at last loss */
84 	u32	last_cwnd;	/* the last snd_cwnd */
85 	u32	last_time;	/* time when updated last_cwnd */
86 	u32	bic_origin_point;/* origin point of bic function */
87 	u32	bic_K;		/* time to origin point from the beginning of the current epoch */
88 	u32	delay_min;	/* min delay */
89 	u32	epoch_start;	/* beginning of an epoch */
90 	u32	ack_cnt;	/* number of acks */
91 	u32	tcp_cwnd;	/* estimated tcp cwnd */
92 #define ACK_RATIO_SHIFT	4
93 	u16	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
94 	u8	sample_cnt;	/* number of samples to decide curr_rtt */
95 	u8	found;		/* the exit point is found? */
96 	u32	round_start;	/* beginning of each round */
97 	u32	end_seq;	/* end_seq of the round */
98 	u32	last_jiffies;	/* last time when the ACK spacing is close */
99 	u32	curr_rtt;	/* the minimum rtt of current round */
100 };
101 
102 static inline void bictcp_reset(struct bictcp *ca)
103 {
104 	ca->cnt = 0;
105 	ca->last_max_cwnd = 0;
106 	ca->loss_cwnd = 0;
107 	ca->last_cwnd = 0;
108 	ca->last_time = 0;
109 	ca->bic_origin_point = 0;
110 	ca->bic_K = 0;
111 	ca->delay_min = 0;
112 	ca->epoch_start = 0;
113 	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
114 	ca->ack_cnt = 0;
115 	ca->tcp_cwnd = 0;
116 	ca->found = 0;
117 }
118 
119 static inline void bictcp_hystart_reset(struct sock *sk)
120 {
121 	struct tcp_sock *tp = tcp_sk(sk);
122 	struct bictcp *ca = inet_csk_ca(sk);
123 
124 	ca->round_start = ca->last_jiffies = jiffies;
125 	ca->end_seq = tp->snd_nxt;
126 	ca->curr_rtt = 0;
127 	ca->sample_cnt = 0;
128 }
129 
130 static void bictcp_init(struct sock *sk)
131 {
132 	bictcp_reset(inet_csk_ca(sk));
133 
134 	if (hystart)
135 		bictcp_hystart_reset(sk);
136 
137 	if (!hystart && initial_ssthresh)
138 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
139 }
140 
141 /* calculate the cubic root of x using a table lookup followed by one
142  * Newton-Raphson iteration.
143  * Avg err ~= 0.195%
144  */
145 static u32 cubic_root(u64 a)
146 {
147 	u32 x, b, shift;
148 	/*
149 	 * cbrt(x) MSB values for x MSB values in [0..63].
150 	 * Precomputed then refined by hand - Willy Tarreau
151 	 *
152 	 * For x in [0..63],
153 	 *   v = cbrt(x << 18) - 1
154 	 *   cbrt(x) = (v[x] + 10) >> 6
155 	 */
156 	static const u8 v[] = {
157 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
158 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
159 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
160 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
161 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
162 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
163 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
164 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
165 	};
166 
167 	b = fls64(a);
168 	if (b < 7) {
169 		/* a in [0..63] */
170 		return ((u32)v[(u32)a] + 35) >> 6;
171 	}
172 
173 	b = ((b * 84) >> 8) - 1;
174 	shift = (a >> (b * 3));
175 
176 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
177 
178 	/*
179 	 * Newton-Raphson iteration
180 	 *                         2
181 	 * x    = ( 2 * x  +  a / x  ) / 3
182 	 *  k+1          k         k
183 	 */
184 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
185 	x = ((x * 341) >> 10);
186 	return x;
187 }
188 
189 /*
190  * Compute congestion window to use.
191  */
192 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
193 {
194 	u64 offs;
195 	u32 delta, t, bic_target, max_cnt;
196 
197 	ca->ack_cnt++;	/* count the number of ACKs */
198 
199 	if (ca->last_cwnd == cwnd &&
200 	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
201 		return;
202 
203 	ca->last_cwnd = cwnd;
204 	ca->last_time = tcp_time_stamp;
205 
206 	if (ca->epoch_start == 0) {
207 		ca->epoch_start = tcp_time_stamp;	/* record the beginning of an epoch */
208 		ca->ack_cnt = 1;			/* start counting */
209 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
210 
211 		if (ca->last_max_cwnd <= cwnd) {
212 			ca->bic_K = 0;
213 			ca->bic_origin_point = cwnd;
214 		} else {
215 			/* Compute new K based on
216 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
217 			 */
218 			ca->bic_K = cubic_root(cube_factor
219 					       * (ca->last_max_cwnd - cwnd));
220 			ca->bic_origin_point = ca->last_max_cwnd;
221 		}
222 	}
223 
224 	/* cubic function - calc*/
225 	/* calculate c * time^3 / rtt,
226 	 *  while considering overflow in calculation of time^3
227 	 * (so time^3 is done by using 64 bit)
228 	 * and without the support of division of 64bit numbers
229 	 * (so all divisions are done by using 32 bit)
230 	 *  also NOTE the unit of those veriables
231 	 *	  time  = (t - K) / 2^bictcp_HZ
232 	 *	  c = bic_scale >> 10
233 	 * rtt  = (srtt >> 3) / HZ
234 	 * !!! The following code does not have overflow problems,
235 	 * if the cwnd < 1 million packets !!!
236 	 */
237 
238 	/* change the unit from HZ to bictcp_HZ */
239 	t = ((tcp_time_stamp + (ca->delay_min>>3) - ca->epoch_start)
240 	     << BICTCP_HZ) / HZ;
241 
242 	if (t < ca->bic_K)		/* t - K */
243 		offs = ca->bic_K - t;
244 	else
245 		offs = t - ca->bic_K;
246 
247 	/* c/rtt * (t-K)^3 */
248 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
249 	if (t < ca->bic_K)                                	/* below origin*/
250 		bic_target = ca->bic_origin_point - delta;
251 	else                                                	/* above origin*/
252 		bic_target = ca->bic_origin_point + delta;
253 
254 	/* cubic function - calc bictcp_cnt*/
255 	if (bic_target > cwnd) {
256 		ca->cnt = cwnd / (bic_target - cwnd);
257 	} else {
258 		ca->cnt = 100 * cwnd;              /* very small increment*/
259 	}
260 
261 	/* TCP Friendly */
262 	if (tcp_friendliness) {
263 		u32 scale = beta_scale;
264 		delta = (cwnd * scale) >> 3;
265 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
266 			ca->ack_cnt -= delta;
267 			ca->tcp_cwnd++;
268 		}
269 
270 		if (ca->tcp_cwnd > cwnd){	/* if bic is slower than tcp */
271 			delta = ca->tcp_cwnd - cwnd;
272 			max_cnt = cwnd / delta;
273 			if (ca->cnt > max_cnt)
274 				ca->cnt = max_cnt;
275 		}
276 	}
277 
278 	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
279 	if (ca->cnt == 0)			/* cannot be zero */
280 		ca->cnt = 1;
281 }
282 
283 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
284 {
285 	struct tcp_sock *tp = tcp_sk(sk);
286 	struct bictcp *ca = inet_csk_ca(sk);
287 
288 	if (!tcp_is_cwnd_limited(sk, in_flight))
289 		return;
290 
291 	if (tp->snd_cwnd <= tp->snd_ssthresh) {
292 		if (hystart && after(ack, ca->end_seq))
293 			bictcp_hystart_reset(sk);
294 		tcp_slow_start(tp);
295 	} else {
296 		bictcp_update(ca, tp->snd_cwnd);
297 		tcp_cong_avoid_ai(tp, ca->cnt);
298 	}
299 
300 }
301 
302 static u32 bictcp_recalc_ssthresh(struct sock *sk)
303 {
304 	const struct tcp_sock *tp = tcp_sk(sk);
305 	struct bictcp *ca = inet_csk_ca(sk);
306 
307 	ca->epoch_start = 0;	/* end of epoch */
308 
309 	/* Wmax and fast convergence */
310 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
311 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
312 			/ (2 * BICTCP_BETA_SCALE);
313 	else
314 		ca->last_max_cwnd = tp->snd_cwnd;
315 
316 	ca->loss_cwnd = tp->snd_cwnd;
317 
318 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
319 }
320 
321 static u32 bictcp_undo_cwnd(struct sock *sk)
322 {
323 	struct bictcp *ca = inet_csk_ca(sk);
324 
325 	return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
326 }
327 
328 static void bictcp_state(struct sock *sk, u8 new_state)
329 {
330 	if (new_state == TCP_CA_Loss) {
331 		bictcp_reset(inet_csk_ca(sk));
332 		bictcp_hystart_reset(sk);
333 	}
334 }
335 
336 static void hystart_update(struct sock *sk, u32 delay)
337 {
338 	struct tcp_sock *tp = tcp_sk(sk);
339 	struct bictcp *ca = inet_csk_ca(sk);
340 
341 	if (!(ca->found & hystart_detect)) {
342 		u32 curr_jiffies = jiffies;
343 
344 		/* first detection parameter - ack-train detection */
345 		if (curr_jiffies - ca->last_jiffies <= msecs_to_jiffies(2)) {
346 			ca->last_jiffies = curr_jiffies;
347 			if (curr_jiffies - ca->round_start >= ca->delay_min>>4)
348 				ca->found |= HYSTART_ACK_TRAIN;
349 		}
350 
351 		/* obtain the minimum delay of more than sampling packets */
352 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
353 			if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
354 				ca->curr_rtt = delay;
355 
356 			ca->sample_cnt++;
357 		} else {
358 			if (ca->curr_rtt > ca->delay_min +
359 			    HYSTART_DELAY_THRESH(ca->delay_min>>4))
360 				ca->found |= HYSTART_DELAY;
361 		}
362 		/*
363 		 * Either one of two conditions are met,
364 		 * we exit from slow start immediately.
365 		 */
366 		if (ca->found & hystart_detect)
367 			tp->snd_ssthresh = tp->snd_cwnd;
368 	}
369 }
370 
371 /* Track delayed acknowledgment ratio using sliding window
372  * ratio = (15*ratio + sample) / 16
373  */
374 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
375 {
376 	const struct inet_connection_sock *icsk = inet_csk(sk);
377 	const struct tcp_sock *tp = tcp_sk(sk);
378 	struct bictcp *ca = inet_csk_ca(sk);
379 	u32 delay;
380 
381 	if (icsk->icsk_ca_state == TCP_CA_Open) {
382 		cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
383 		ca->delayed_ack += cnt;
384 	}
385 
386 	/* Some calls are for duplicates without timetamps */
387 	if (rtt_us < 0)
388 		return;
389 
390 	/* Discard delay samples right after fast recovery */
391 	if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
392 		return;
393 
394 	delay = usecs_to_jiffies(rtt_us) << 3;
395 	if (delay == 0)
396 		delay = 1;
397 
398 	/* first time call or link delay decreases */
399 	if (ca->delay_min == 0 || ca->delay_min > delay)
400 		ca->delay_min = delay;
401 
402 	/* hystart triggers when cwnd is larger than some threshold */
403 	if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
404 	    tp->snd_cwnd >= hystart_low_window)
405 		hystart_update(sk, delay);
406 }
407 
408 static struct tcp_congestion_ops cubictcp = {
409 	.init		= bictcp_init,
410 	.ssthresh	= bictcp_recalc_ssthresh,
411 	.cong_avoid	= bictcp_cong_avoid,
412 	.set_state	= bictcp_state,
413 	.undo_cwnd	= bictcp_undo_cwnd,
414 	.pkts_acked     = bictcp_acked,
415 	.owner		= THIS_MODULE,
416 	.name		= "cubic",
417 };
418 
419 static int __init cubictcp_register(void)
420 {
421 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
422 
423 	/* Precompute a bunch of the scaling factors that are used per-packet
424 	 * based on SRTT of 100ms
425 	 */
426 
427 	beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
428 
429 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
430 
431 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
432 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
433 	 * the unit of K is bictcp_HZ=2^10, not HZ
434 	 *
435 	 *  c = bic_scale >> 10
436 	 *  rtt = 100ms
437 	 *
438 	 * the following code has been designed and tested for
439 	 * cwnd < 1 million packets
440 	 * RTT < 100 seconds
441 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
442 	 */
443 
444 	/* 1/c * 2^2*bictcp_HZ * srtt */
445 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
446 
447 	/* divide by bic_scale and by constant Srtt (100ms) */
448 	do_div(cube_factor, bic_scale * 10);
449 
450 	return tcp_register_congestion_control(&cubictcp);
451 }
452 
453 static void __exit cubictcp_unregister(void)
454 {
455 	tcp_unregister_congestion_control(&cubictcp);
456 }
457 
458 module_init(cubictcp_register);
459 module_exit(cubictcp_unregister);
460 
461 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
462 MODULE_LICENSE("GPL");
463 MODULE_DESCRIPTION("CUBIC TCP");
464 MODULE_VERSION("2.3");
465