xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision ae213c44)
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	(4U<<3)
44 #define HYSTART_DELAY_MAX	(16U<<3)
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 __read_mostly = 2;
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, int, 0644);
81 MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
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 (msec << 3) */
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(void)
121 {
122 #if HZ < 1000
123 	return ktime_to_ms(ktime_get_real());
124 #else
125 	return jiffies_to_msecs(jiffies);
126 #endif
127 }
128 
129 static inline void bictcp_hystart_reset(struct sock *sk)
130 {
131 	struct tcp_sock *tp = tcp_sk(sk);
132 	struct bictcp *ca = inet_csk_ca(sk);
133 
134 	ca->round_start = ca->last_ack = bictcp_clock();
135 	ca->end_seq = tp->snd_nxt;
136 	ca->curr_rtt = 0;
137 	ca->sample_cnt = 0;
138 }
139 
140 static void bictcp_init(struct sock *sk)
141 {
142 	struct bictcp *ca = inet_csk_ca(sk);
143 
144 	bictcp_reset(ca);
145 
146 	if (hystart)
147 		bictcp_hystart_reset(sk);
148 
149 	if (!hystart && initial_ssthresh)
150 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
151 }
152 
153 static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
154 {
155 	if (event == CA_EVENT_TX_START) {
156 		struct bictcp *ca = inet_csk_ca(sk);
157 		u32 now = tcp_jiffies32;
158 		s32 delta;
159 
160 		delta = now - tcp_sk(sk)->lsndtime;
161 
162 		/* We were application limited (idle) for a while.
163 		 * Shift epoch_start to keep cwnd growth to cubic curve.
164 		 */
165 		if (ca->epoch_start && delta > 0) {
166 			ca->epoch_start += delta;
167 			if (after(ca->epoch_start, now))
168 				ca->epoch_start = now;
169 		}
170 		return;
171 	}
172 }
173 
174 /* calculate the cubic root of x using a table lookup followed by one
175  * Newton-Raphson iteration.
176  * Avg err ~= 0.195%
177  */
178 static u32 cubic_root(u64 a)
179 {
180 	u32 x, b, shift;
181 	/*
182 	 * cbrt(x) MSB values for x MSB values in [0..63].
183 	 * Precomputed then refined by hand - Willy Tarreau
184 	 *
185 	 * For x in [0..63],
186 	 *   v = cbrt(x << 18) - 1
187 	 *   cbrt(x) = (v[x] + 10) >> 6
188 	 */
189 	static const u8 v[] = {
190 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
191 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
192 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
193 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
194 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
195 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
196 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
197 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
198 	};
199 
200 	b = fls64(a);
201 	if (b < 7) {
202 		/* a in [0..63] */
203 		return ((u32)v[(u32)a] + 35) >> 6;
204 	}
205 
206 	b = ((b * 84) >> 8) - 1;
207 	shift = (a >> (b * 3));
208 
209 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
210 
211 	/*
212 	 * Newton-Raphson iteration
213 	 *                         2
214 	 * x    = ( 2 * x  +  a / x  ) / 3
215 	 *  k+1          k         k
216 	 */
217 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
218 	x = ((x * 341) >> 10);
219 	return x;
220 }
221 
222 /*
223  * Compute congestion window to use.
224  */
225 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
226 {
227 	u32 delta, bic_target, max_cnt;
228 	u64 offs, t;
229 
230 	ca->ack_cnt += acked;	/* count the number of ACKed packets */
231 
232 	if (ca->last_cwnd == cwnd &&
233 	    (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
234 		return;
235 
236 	/* The CUBIC function can update ca->cnt at most once per jiffy.
237 	 * On all cwnd reduction events, ca->epoch_start is set to 0,
238 	 * which will force a recalculation of ca->cnt.
239 	 */
240 	if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
241 		goto tcp_friendliness;
242 
243 	ca->last_cwnd = cwnd;
244 	ca->last_time = tcp_jiffies32;
245 
246 	if (ca->epoch_start == 0) {
247 		ca->epoch_start = tcp_jiffies32;	/* record beginning */
248 		ca->ack_cnt = acked;			/* start counting */
249 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
250 
251 		if (ca->last_max_cwnd <= cwnd) {
252 			ca->bic_K = 0;
253 			ca->bic_origin_point = cwnd;
254 		} else {
255 			/* Compute new K based on
256 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
257 			 */
258 			ca->bic_K = cubic_root(cube_factor
259 					       * (ca->last_max_cwnd - cwnd));
260 			ca->bic_origin_point = ca->last_max_cwnd;
261 		}
262 	}
263 
264 	/* cubic function - calc*/
265 	/* calculate c * time^3 / rtt,
266 	 *  while considering overflow in calculation of time^3
267 	 * (so time^3 is done by using 64 bit)
268 	 * and without the support of division of 64bit numbers
269 	 * (so all divisions are done by using 32 bit)
270 	 *  also NOTE the unit of those veriables
271 	 *	  time  = (t - K) / 2^bictcp_HZ
272 	 *	  c = bic_scale >> 10
273 	 * rtt  = (srtt >> 3) / HZ
274 	 * !!! The following code does not have overflow problems,
275 	 * if the cwnd < 1 million packets !!!
276 	 */
277 
278 	t = (s32)(tcp_jiffies32 - ca->epoch_start);
279 	t += msecs_to_jiffies(ca->delay_min >> 3);
280 	/* change the unit from HZ to bictcp_HZ */
281 	t <<= BICTCP_HZ;
282 	do_div(t, HZ);
283 
284 	if (t < ca->bic_K)		/* t - K */
285 		offs = ca->bic_K - t;
286 	else
287 		offs = t - ca->bic_K;
288 
289 	/* c/rtt * (t-K)^3 */
290 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
291 	if (t < ca->bic_K)                            /* below origin*/
292 		bic_target = ca->bic_origin_point - delta;
293 	else                                          /* above origin*/
294 		bic_target = ca->bic_origin_point + delta;
295 
296 	/* cubic function - calc bictcp_cnt*/
297 	if (bic_target > cwnd) {
298 		ca->cnt = cwnd / (bic_target - cwnd);
299 	} else {
300 		ca->cnt = 100 * cwnd;              /* very small increment*/
301 	}
302 
303 	/*
304 	 * The initial growth of cubic function may be too conservative
305 	 * when the available bandwidth is still unknown.
306 	 */
307 	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
308 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
309 
310 tcp_friendliness:
311 	/* TCP Friendly */
312 	if (tcp_friendliness) {
313 		u32 scale = beta_scale;
314 
315 		delta = (cwnd * scale) >> 3;
316 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
317 			ca->ack_cnt -= delta;
318 			ca->tcp_cwnd++;
319 		}
320 
321 		if (ca->tcp_cwnd > cwnd) {	/* if bic is slower than tcp */
322 			delta = ca->tcp_cwnd - cwnd;
323 			max_cnt = cwnd / delta;
324 			if (ca->cnt > max_cnt)
325 				ca->cnt = max_cnt;
326 		}
327 	}
328 
329 	/* The maximum rate of cwnd increase CUBIC allows is 1 packet per
330 	 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
331 	 */
332 	ca->cnt = max(ca->cnt, 2U);
333 }
334 
335 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
336 {
337 	struct tcp_sock *tp = tcp_sk(sk);
338 	struct bictcp *ca = inet_csk_ca(sk);
339 
340 	if (!tcp_is_cwnd_limited(sk))
341 		return;
342 
343 	if (tcp_in_slow_start(tp)) {
344 		if (hystart && after(ack, ca->end_seq))
345 			bictcp_hystart_reset(sk);
346 		acked = tcp_slow_start(tp, acked);
347 		if (!acked)
348 			return;
349 	}
350 	bictcp_update(ca, tp->snd_cwnd, acked);
351 	tcp_cong_avoid_ai(tp, ca->cnt, acked);
352 }
353 
354 static u32 bictcp_recalc_ssthresh(struct sock *sk)
355 {
356 	const struct tcp_sock *tp = tcp_sk(sk);
357 	struct bictcp *ca = inet_csk_ca(sk);
358 
359 	ca->epoch_start = 0;	/* end of epoch */
360 
361 	/* Wmax and fast convergence */
362 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
363 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
364 			/ (2 * BICTCP_BETA_SCALE);
365 	else
366 		ca->last_max_cwnd = tp->snd_cwnd;
367 
368 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
369 }
370 
371 static void bictcp_state(struct sock *sk, u8 new_state)
372 {
373 	if (new_state == TCP_CA_Loss) {
374 		bictcp_reset(inet_csk_ca(sk));
375 		bictcp_hystart_reset(sk);
376 	}
377 }
378 
379 static void hystart_update(struct sock *sk, u32 delay)
380 {
381 	struct tcp_sock *tp = tcp_sk(sk);
382 	struct bictcp *ca = inet_csk_ca(sk);
383 
384 	if (ca->found & hystart_detect)
385 		return;
386 
387 	if (hystart_detect & HYSTART_ACK_TRAIN) {
388 		u32 now = bictcp_clock();
389 
390 		/* first detection parameter - ack-train detection */
391 		if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
392 			ca->last_ack = now;
393 			if ((s32)(now - ca->round_start) > ca->delay_min >> 4) {
394 				ca->found |= HYSTART_ACK_TRAIN;
395 				NET_INC_STATS(sock_net(sk),
396 					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
397 				NET_ADD_STATS(sock_net(sk),
398 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
399 					      tp->snd_cwnd);
400 				tp->snd_ssthresh = tp->snd_cwnd;
401 			}
402 		}
403 	}
404 
405 	if (hystart_detect & HYSTART_DELAY) {
406 		/* obtain the minimum delay of more than sampling packets */
407 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
408 			if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
409 				ca->curr_rtt = delay;
410 
411 			ca->sample_cnt++;
412 		} else {
413 			if (ca->curr_rtt > ca->delay_min +
414 			    HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
415 				ca->found |= HYSTART_DELAY;
416 				NET_INC_STATS(sock_net(sk),
417 					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
418 				NET_ADD_STATS(sock_net(sk),
419 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
420 					      tp->snd_cwnd);
421 				tp->snd_ssthresh = tp->snd_cwnd;
422 			}
423 		}
424 	}
425 }
426 
427 /* Track delayed acknowledgment ratio using sliding window
428  * ratio = (15*ratio + sample) / 16
429  */
430 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
431 {
432 	const struct tcp_sock *tp = tcp_sk(sk);
433 	struct bictcp *ca = inet_csk_ca(sk);
434 	u32 delay;
435 
436 	/* Some calls are for duplicates without timetamps */
437 	if (sample->rtt_us < 0)
438 		return;
439 
440 	/* Discard delay samples right after fast recovery */
441 	if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
442 		return;
443 
444 	delay = (sample->rtt_us << 3) / USEC_PER_MSEC;
445 	if (delay == 0)
446 		delay = 1;
447 
448 	/* first time call or link delay decreases */
449 	if (ca->delay_min == 0 || ca->delay_min > delay)
450 		ca->delay_min = delay;
451 
452 	/* hystart triggers when cwnd is larger than some threshold */
453 	if (hystart && tcp_in_slow_start(tp) &&
454 	    tp->snd_cwnd >= hystart_low_window)
455 		hystart_update(sk, delay);
456 }
457 
458 static struct tcp_congestion_ops cubictcp __read_mostly = {
459 	.init		= bictcp_init,
460 	.ssthresh	= bictcp_recalc_ssthresh,
461 	.cong_avoid	= bictcp_cong_avoid,
462 	.set_state	= bictcp_state,
463 	.undo_cwnd	= tcp_reno_undo_cwnd,
464 	.cwnd_event	= bictcp_cwnd_event,
465 	.pkts_acked     = bictcp_acked,
466 	.owner		= THIS_MODULE,
467 	.name		= "cubic",
468 };
469 
470 static int __init cubictcp_register(void)
471 {
472 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
473 
474 	/* Precompute a bunch of the scaling factors that are used per-packet
475 	 * based on SRTT of 100ms
476 	 */
477 
478 	beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
479 		/ (BICTCP_BETA_SCALE - beta);
480 
481 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
482 
483 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
484 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
485 	 * the unit of K is bictcp_HZ=2^10, not HZ
486 	 *
487 	 *  c = bic_scale >> 10
488 	 *  rtt = 100ms
489 	 *
490 	 * the following code has been designed and tested for
491 	 * cwnd < 1 million packets
492 	 * RTT < 100 seconds
493 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
494 	 */
495 
496 	/* 1/c * 2^2*bictcp_HZ * srtt */
497 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
498 
499 	/* divide by bic_scale and by constant Srtt (100ms) */
500 	do_div(cube_factor, bic_scale * 10);
501 
502 	return tcp_register_congestion_control(&cubictcp);
503 }
504 
505 static void __exit cubictcp_unregister(void)
506 {
507 	tcp_unregister_congestion_control(&cubictcp);
508 }
509 
510 module_init(cubictcp_register);
511 module_exit(cubictcp_unregister);
512 
513 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
514 MODULE_LICENSE("GPL");
515 MODULE_DESCRIPTION("CUBIC TCP");
516 MODULE_VERSION("2.3");
517