xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision 842ed298)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
4  * Home page:
5  *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
6  * This is from the implementation of CUBIC TCP in
7  * Sangtae Ha, Injong Rhee and Lisong Xu,
8  *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant"
9  *  in ACM SIGOPS Operating System Review, July 2008.
10  * Available from:
11  *  http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
12  *
13  * CUBIC integrates a new slow start algorithm, called HyStart.
14  * The details of HyStart are presented in
15  *  Sangtae Ha and Injong Rhee,
16  *  "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008.
17  * Available from:
18  *  http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
19  *
20  * All testing results are available from:
21  * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
22  *
23  * Unless CUBIC is enabled and congestion window is large
24  * this behaves the same as the original Reno.
25  */
26 
27 #include <linux/mm.h>
28 #include <linux/module.h>
29 #include <linux/math64.h>
30 #include <net/tcp.h>
31 
32 #define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
33 					 * max_cwnd = snd_cwnd * beta
34 					 */
35 #define	BICTCP_HZ		10	/* BIC HZ 2^10 = 1024 */
36 
37 /* Two methods of hybrid slow start */
38 #define HYSTART_ACK_TRAIN	0x1
39 #define HYSTART_DELAY		0x2
40 
41 /* Number of delay samples for detecting the increase of delay */
42 #define HYSTART_MIN_SAMPLES	8
43 #define HYSTART_DELAY_MIN	(4000U)	/* 4 ms */
44 #define HYSTART_DELAY_MAX	(16000U)	/* 16 ms */
45 #define HYSTART_DELAY_THRESH(x)	clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
46 
47 static int fast_convergence __read_mostly = 1;
48 static int beta __read_mostly = 717;	/* = 717/1024 (BICTCP_BETA_SCALE) */
49 static int initial_ssthresh __read_mostly;
50 static int bic_scale __read_mostly = 41;
51 static int tcp_friendliness __read_mostly = 1;
52 
53 static int hystart __read_mostly = 1;
54 static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
55 static int hystart_low_window __read_mostly = 16;
56 static int hystart_ack_delta_us __read_mostly = 2000;
57 
58 static u32 cube_rtt_scale __read_mostly;
59 static u32 beta_scale __read_mostly;
60 static u64 cube_factor __read_mostly;
61 
62 /* Note parameters that are used for precomputing scale factors are read-only */
63 module_param(fast_convergence, int, 0644);
64 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
65 module_param(beta, int, 0644);
66 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
67 module_param(initial_ssthresh, int, 0644);
68 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
69 module_param(bic_scale, int, 0444);
70 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
71 module_param(tcp_friendliness, int, 0644);
72 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
73 module_param(hystart, int, 0644);
74 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
75 module_param(hystart_detect, int, 0644);
76 MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms"
77 		 " 1: packet-train 2: delay 3: both packet-train and delay");
78 module_param(hystart_low_window, int, 0644);
79 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
80 module_param(hystart_ack_delta_us, int, 0644);
81 MODULE_PARM_DESC(hystart_ack_delta_us, "spacing between ack's indicating train (usecs)");
82 
83 /* BIC TCP Parameters */
84 struct bictcp {
85 	u32	cnt;		/* increase cwnd by 1 after ACKs */
86 	u32	last_max_cwnd;	/* last maximum snd_cwnd */
87 	u32	last_cwnd;	/* the last snd_cwnd */
88 	u32	last_time;	/* time when updated last_cwnd */
89 	u32	bic_origin_point;/* origin point of bic function */
90 	u32	bic_K;		/* time to origin point
91 				   from the beginning of the current epoch */
92 	u32	delay_min;	/* min delay (usec) */
93 	u32	epoch_start;	/* beginning of an epoch */
94 	u32	ack_cnt;	/* number of acks */
95 	u32	tcp_cwnd;	/* estimated tcp cwnd */
96 	u16	unused;
97 	u8	sample_cnt;	/* number of samples to decide curr_rtt */
98 	u8	found;		/* the exit point is found? */
99 	u32	round_start;	/* beginning of each round */
100 	u32	end_seq;	/* end_seq of the round */
101 	u32	last_ack;	/* last time when the ACK spacing is close */
102 	u32	curr_rtt;	/* the minimum rtt of current round */
103 };
104 
105 static inline void bictcp_reset(struct bictcp *ca)
106 {
107 	ca->cnt = 0;
108 	ca->last_max_cwnd = 0;
109 	ca->last_cwnd = 0;
110 	ca->last_time = 0;
111 	ca->bic_origin_point = 0;
112 	ca->bic_K = 0;
113 	ca->delay_min = 0;
114 	ca->epoch_start = 0;
115 	ca->ack_cnt = 0;
116 	ca->tcp_cwnd = 0;
117 	ca->found = 0;
118 }
119 
120 static inline u32 bictcp_clock_us(const struct sock *sk)
121 {
122 	return tcp_sk(sk)->tcp_mstamp;
123 }
124 
125 static inline void bictcp_hystart_reset(struct sock *sk)
126 {
127 	struct tcp_sock *tp = tcp_sk(sk);
128 	struct bictcp *ca = inet_csk_ca(sk);
129 
130 	ca->round_start = ca->last_ack = bictcp_clock_us(sk);
131 	ca->end_seq = tp->snd_nxt;
132 	ca->curr_rtt = ~0U;
133 	ca->sample_cnt = 0;
134 }
135 
136 static void bictcp_init(struct sock *sk)
137 {
138 	struct bictcp *ca = inet_csk_ca(sk);
139 
140 	bictcp_reset(ca);
141 
142 	if (hystart)
143 		bictcp_hystart_reset(sk);
144 
145 	if (!hystart && initial_ssthresh)
146 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
147 }
148 
149 static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
150 {
151 	if (event == CA_EVENT_TX_START) {
152 		struct bictcp *ca = inet_csk_ca(sk);
153 		u32 now = tcp_jiffies32;
154 		s32 delta;
155 
156 		delta = now - tcp_sk(sk)->lsndtime;
157 
158 		/* We were application limited (idle) for a while.
159 		 * Shift epoch_start to keep cwnd growth to cubic curve.
160 		 */
161 		if (ca->epoch_start && delta > 0) {
162 			ca->epoch_start += delta;
163 			if (after(ca->epoch_start, now))
164 				ca->epoch_start = now;
165 		}
166 		return;
167 	}
168 }
169 
170 /* calculate the cubic root of x using a table lookup followed by one
171  * Newton-Raphson iteration.
172  * Avg err ~= 0.195%
173  */
174 static u32 cubic_root(u64 a)
175 {
176 	u32 x, b, shift;
177 	/*
178 	 * cbrt(x) MSB values for x MSB values in [0..63].
179 	 * Precomputed then refined by hand - Willy Tarreau
180 	 *
181 	 * For x in [0..63],
182 	 *   v = cbrt(x << 18) - 1
183 	 *   cbrt(x) = (v[x] + 10) >> 6
184 	 */
185 	static const u8 v[] = {
186 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
187 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
188 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
189 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
190 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
191 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
192 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
193 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
194 	};
195 
196 	b = fls64(a);
197 	if (b < 7) {
198 		/* a in [0..63] */
199 		return ((u32)v[(u32)a] + 35) >> 6;
200 	}
201 
202 	b = ((b * 84) >> 8) - 1;
203 	shift = (a >> (b * 3));
204 
205 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
206 
207 	/*
208 	 * Newton-Raphson iteration
209 	 *                         2
210 	 * x    = ( 2 * x  +  a / x  ) / 3
211 	 *  k+1          k         k
212 	 */
213 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
214 	x = ((x * 341) >> 10);
215 	return x;
216 }
217 
218 /*
219  * Compute congestion window to use.
220  */
221 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
222 {
223 	u32 delta, bic_target, max_cnt;
224 	u64 offs, t;
225 
226 	ca->ack_cnt += acked;	/* count the number of ACKed packets */
227 
228 	if (ca->last_cwnd == cwnd &&
229 	    (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
230 		return;
231 
232 	/* The CUBIC function can update ca->cnt at most once per jiffy.
233 	 * On all cwnd reduction events, ca->epoch_start is set to 0,
234 	 * which will force a recalculation of ca->cnt.
235 	 */
236 	if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
237 		goto tcp_friendliness;
238 
239 	ca->last_cwnd = cwnd;
240 	ca->last_time = tcp_jiffies32;
241 
242 	if (ca->epoch_start == 0) {
243 		ca->epoch_start = tcp_jiffies32;	/* record beginning */
244 		ca->ack_cnt = acked;			/* start counting */
245 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
246 
247 		if (ca->last_max_cwnd <= cwnd) {
248 			ca->bic_K = 0;
249 			ca->bic_origin_point = cwnd;
250 		} else {
251 			/* Compute new K based on
252 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
253 			 */
254 			ca->bic_K = cubic_root(cube_factor
255 					       * (ca->last_max_cwnd - cwnd));
256 			ca->bic_origin_point = ca->last_max_cwnd;
257 		}
258 	}
259 
260 	/* cubic function - calc*/
261 	/* calculate c * time^3 / rtt,
262 	 *  while considering overflow in calculation of time^3
263 	 * (so time^3 is done by using 64 bit)
264 	 * and without the support of division of 64bit numbers
265 	 * (so all divisions are done by using 32 bit)
266 	 *  also NOTE the unit of those veriables
267 	 *	  time  = (t - K) / 2^bictcp_HZ
268 	 *	  c = bic_scale >> 10
269 	 * rtt  = (srtt >> 3) / HZ
270 	 * !!! The following code does not have overflow problems,
271 	 * if the cwnd < 1 million packets !!!
272 	 */
273 
274 	t = (s32)(tcp_jiffies32 - ca->epoch_start);
275 	t += usecs_to_jiffies(ca->delay_min);
276 	/* change the unit from HZ to bictcp_HZ */
277 	t <<= BICTCP_HZ;
278 	do_div(t, HZ);
279 
280 	if (t < ca->bic_K)		/* t - K */
281 		offs = ca->bic_K - t;
282 	else
283 		offs = t - ca->bic_K;
284 
285 	/* c/rtt * (t-K)^3 */
286 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
287 	if (t < ca->bic_K)                            /* below origin*/
288 		bic_target = ca->bic_origin_point - delta;
289 	else                                          /* above origin*/
290 		bic_target = ca->bic_origin_point + delta;
291 
292 	/* cubic function - calc bictcp_cnt*/
293 	if (bic_target > cwnd) {
294 		ca->cnt = cwnd / (bic_target - cwnd);
295 	} else {
296 		ca->cnt = 100 * cwnd;              /* very small increment*/
297 	}
298 
299 	/*
300 	 * The initial growth of cubic function may be too conservative
301 	 * when the available bandwidth is still unknown.
302 	 */
303 	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
304 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
305 
306 tcp_friendliness:
307 	/* TCP Friendly */
308 	if (tcp_friendliness) {
309 		u32 scale = beta_scale;
310 
311 		delta = (cwnd * scale) >> 3;
312 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
313 			ca->ack_cnt -= delta;
314 			ca->tcp_cwnd++;
315 		}
316 
317 		if (ca->tcp_cwnd > cwnd) {	/* if bic is slower than tcp */
318 			delta = ca->tcp_cwnd - cwnd;
319 			max_cnt = cwnd / delta;
320 			if (ca->cnt > max_cnt)
321 				ca->cnt = max_cnt;
322 		}
323 	}
324 
325 	/* The maximum rate of cwnd increase CUBIC allows is 1 packet per
326 	 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
327 	 */
328 	ca->cnt = max(ca->cnt, 2U);
329 }
330 
331 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
332 {
333 	struct tcp_sock *tp = tcp_sk(sk);
334 	struct bictcp *ca = inet_csk_ca(sk);
335 
336 	if (!tcp_is_cwnd_limited(sk))
337 		return;
338 
339 	if (tcp_in_slow_start(tp)) {
340 		if (hystart && after(ack, ca->end_seq))
341 			bictcp_hystart_reset(sk);
342 		acked = tcp_slow_start(tp, acked);
343 		if (!acked)
344 			return;
345 	}
346 	bictcp_update(ca, tp->snd_cwnd, acked);
347 	tcp_cong_avoid_ai(tp, ca->cnt, acked);
348 }
349 
350 static u32 bictcp_recalc_ssthresh(struct sock *sk)
351 {
352 	const struct tcp_sock *tp = tcp_sk(sk);
353 	struct bictcp *ca = inet_csk_ca(sk);
354 
355 	ca->epoch_start = 0;	/* end of epoch */
356 
357 	/* Wmax and fast convergence */
358 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
359 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
360 			/ (2 * BICTCP_BETA_SCALE);
361 	else
362 		ca->last_max_cwnd = tp->snd_cwnd;
363 
364 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
365 }
366 
367 static void bictcp_state(struct sock *sk, u8 new_state)
368 {
369 	if (new_state == TCP_CA_Loss) {
370 		bictcp_reset(inet_csk_ca(sk));
371 		bictcp_hystart_reset(sk);
372 	}
373 }
374 
375 /* Account for TSO/GRO delays.
376  * Otherwise short RTT flows could get too small ssthresh, since during
377  * slow start we begin with small TSO packets and ca->delay_min would
378  * not account for long aggregation delay when TSO packets get bigger.
379  * Ideally even with a very small RTT we would like to have at least one
380  * TSO packet being sent and received by GRO, and another one in qdisc layer.
381  * We apply another 100% factor because @rate is doubled at this point.
382  * We cap the cushion to 1ms.
383  */
384 static u32 hystart_ack_delay(struct sock *sk)
385 {
386 	unsigned long rate;
387 
388 	rate = READ_ONCE(sk->sk_pacing_rate);
389 	if (!rate)
390 		return 0;
391 	return min_t(u64, USEC_PER_MSEC,
392 		     div64_ul((u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate));
393 }
394 
395 static void hystart_update(struct sock *sk, u32 delay)
396 {
397 	struct tcp_sock *tp = tcp_sk(sk);
398 	struct bictcp *ca = inet_csk_ca(sk);
399 	u32 threshold;
400 
401 	if (hystart_detect & HYSTART_ACK_TRAIN) {
402 		u32 now = bictcp_clock_us(sk);
403 
404 		/* first detection parameter - ack-train detection */
405 		if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) {
406 			ca->last_ack = now;
407 
408 			threshold = ca->delay_min + hystart_ack_delay(sk);
409 
410 			/* Hystart ack train triggers if we get ack past
411 			 * ca->delay_min/2.
412 			 * Pacing might have delayed packets up to RTT/2
413 			 * during slow start.
414 			 */
415 			if (sk->sk_pacing_status == SK_PACING_NONE)
416 				threshold >>= 1;
417 
418 			if ((s32)(now - ca->round_start) > threshold) {
419 				ca->found = 1;
420 				pr_debug("hystart_ack_train (%u > %u) delay_min %u (+ ack_delay %u) cwnd %u\n",
421 					 now - ca->round_start, threshold,
422 					 ca->delay_min, hystart_ack_delay(sk), tp->snd_cwnd);
423 				NET_INC_STATS(sock_net(sk),
424 					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
425 				NET_ADD_STATS(sock_net(sk),
426 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
427 					      tp->snd_cwnd);
428 				tp->snd_ssthresh = tp->snd_cwnd;
429 			}
430 		}
431 	}
432 
433 	if (hystart_detect & HYSTART_DELAY) {
434 		/* obtain the minimum delay of more than sampling packets */
435 		if (ca->curr_rtt > delay)
436 			ca->curr_rtt = delay;
437 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
438 			ca->sample_cnt++;
439 		} else {
440 			if (ca->curr_rtt > ca->delay_min +
441 			    HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
442 				ca->found = 1;
443 				NET_INC_STATS(sock_net(sk),
444 					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
445 				NET_ADD_STATS(sock_net(sk),
446 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
447 					      tp->snd_cwnd);
448 				tp->snd_ssthresh = tp->snd_cwnd;
449 			}
450 		}
451 	}
452 }
453 
454 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
455 {
456 	const struct tcp_sock *tp = tcp_sk(sk);
457 	struct bictcp *ca = inet_csk_ca(sk);
458 	u32 delay;
459 
460 	/* Some calls are for duplicates without timetamps */
461 	if (sample->rtt_us < 0)
462 		return;
463 
464 	/* Discard delay samples right after fast recovery */
465 	if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
466 		return;
467 
468 	delay = sample->rtt_us;
469 	if (delay == 0)
470 		delay = 1;
471 
472 	/* first time call or link delay decreases */
473 	if (ca->delay_min == 0 || ca->delay_min > delay)
474 		ca->delay_min = delay;
475 
476 	/* hystart triggers when cwnd is larger than some threshold */
477 	if (!ca->found && tcp_in_slow_start(tp) && hystart &&
478 	    tp->snd_cwnd >= hystart_low_window)
479 		hystart_update(sk, delay);
480 }
481 
482 static struct tcp_congestion_ops cubictcp __read_mostly = {
483 	.init		= bictcp_init,
484 	.ssthresh	= bictcp_recalc_ssthresh,
485 	.cong_avoid	= bictcp_cong_avoid,
486 	.set_state	= bictcp_state,
487 	.undo_cwnd	= tcp_reno_undo_cwnd,
488 	.cwnd_event	= bictcp_cwnd_event,
489 	.pkts_acked     = bictcp_acked,
490 	.owner		= THIS_MODULE,
491 	.name		= "cubic",
492 };
493 
494 static int __init cubictcp_register(void)
495 {
496 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
497 
498 	/* Precompute a bunch of the scaling factors that are used per-packet
499 	 * based on SRTT of 100ms
500 	 */
501 
502 	beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
503 		/ (BICTCP_BETA_SCALE - beta);
504 
505 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
506 
507 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
508 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
509 	 * the unit of K is bictcp_HZ=2^10, not HZ
510 	 *
511 	 *  c = bic_scale >> 10
512 	 *  rtt = 100ms
513 	 *
514 	 * the following code has been designed and tested for
515 	 * cwnd < 1 million packets
516 	 * RTT < 100 seconds
517 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
518 	 */
519 
520 	/* 1/c * 2^2*bictcp_HZ * srtt */
521 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
522 
523 	/* divide by bic_scale and by constant Srtt (100ms) */
524 	do_div(cube_factor, bic_scale * 10);
525 
526 	return tcp_register_congestion_control(&cubictcp);
527 }
528 
529 static void __exit cubictcp_unregister(void)
530 {
531 	tcp_unregister_congestion_control(&cubictcp);
532 }
533 
534 module_init(cubictcp_register);
535 module_exit(cubictcp_unregister);
536 
537 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
538 MODULE_LICENSE("GPL");
539 MODULE_DESCRIPTION("CUBIC TCP");
540 MODULE_VERSION("2.3");
541