xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision e3d786a3)
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	(4U<<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 static int hystart_ack_delta __read_mostly = 2;
56 
57 static u32 cube_rtt_scale __read_mostly;
58 static u32 beta_scale __read_mostly;
59 static u64 cube_factor __read_mostly;
60 
61 /* Note parameters that are used for precomputing scale factors are read-only */
62 module_param(fast_convergence, int, 0644);
63 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
64 module_param(beta, int, 0644);
65 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
66 module_param(initial_ssthresh, int, 0644);
67 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
68 module_param(bic_scale, int, 0444);
69 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
70 module_param(tcp_friendliness, int, 0644);
71 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
72 module_param(hystart, int, 0644);
73 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
74 module_param(hystart_detect, int, 0644);
75 MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms"
76 		 " 1: packet-train 2: delay 3: both packet-train and delay");
77 module_param(hystart_low_window, int, 0644);
78 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
79 module_param(hystart_ack_delta, int, 0644);
80 MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
81 
82 /* BIC TCP Parameters */
83 struct bictcp {
84 	u32	cnt;		/* increase cwnd by 1 after ACKs */
85 	u32	last_max_cwnd;	/* last maximum snd_cwnd */
86 	u32	last_cwnd;	/* the last snd_cwnd */
87 	u32	last_time;	/* time when updated last_cwnd */
88 	u32	bic_origin_point;/* origin point of bic function */
89 	u32	bic_K;		/* time to origin point
90 				   from the beginning of the current epoch */
91 	u32	delay_min;	/* min delay (msec << 3) */
92 	u32	epoch_start;	/* beginning of an epoch */
93 	u32	ack_cnt;	/* number of acks */
94 	u32	tcp_cwnd;	/* estimated tcp cwnd */
95 	u16	unused;
96 	u8	sample_cnt;	/* number of samples to decide curr_rtt */
97 	u8	found;		/* the exit point is found? */
98 	u32	round_start;	/* beginning of each round */
99 	u32	end_seq;	/* end_seq of the round */
100 	u32	last_ack;	/* last time when the ACK spacing is close */
101 	u32	curr_rtt;	/* the minimum rtt of current round */
102 };
103 
104 static inline void bictcp_reset(struct bictcp *ca)
105 {
106 	ca->cnt = 0;
107 	ca->last_max_cwnd = 0;
108 	ca->last_cwnd = 0;
109 	ca->last_time = 0;
110 	ca->bic_origin_point = 0;
111 	ca->bic_K = 0;
112 	ca->delay_min = 0;
113 	ca->epoch_start = 0;
114 	ca->ack_cnt = 0;
115 	ca->tcp_cwnd = 0;
116 	ca->found = 0;
117 }
118 
119 static inline u32 bictcp_clock(void)
120 {
121 #if HZ < 1000
122 	return ktime_to_ms(ktime_get_real());
123 #else
124 	return jiffies_to_msecs(jiffies);
125 #endif
126 }
127 
128 static inline void bictcp_hystart_reset(struct sock *sk)
129 {
130 	struct tcp_sock *tp = tcp_sk(sk);
131 	struct bictcp *ca = inet_csk_ca(sk);
132 
133 	ca->round_start = ca->last_ack = bictcp_clock();
134 	ca->end_seq = tp->snd_nxt;
135 	ca->curr_rtt = 0;
136 	ca->sample_cnt = 0;
137 }
138 
139 static void bictcp_init(struct sock *sk)
140 {
141 	struct bictcp *ca = inet_csk_ca(sk);
142 
143 	bictcp_reset(ca);
144 
145 	if (hystart)
146 		bictcp_hystart_reset(sk);
147 
148 	if (!hystart && initial_ssthresh)
149 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
150 }
151 
152 static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
153 {
154 	if (event == CA_EVENT_TX_START) {
155 		struct bictcp *ca = inet_csk_ca(sk);
156 		u32 now = tcp_jiffies32;
157 		s32 delta;
158 
159 		delta = now - tcp_sk(sk)->lsndtime;
160 
161 		/* We were application limited (idle) for a while.
162 		 * Shift epoch_start to keep cwnd growth to cubic curve.
163 		 */
164 		if (ca->epoch_start && delta > 0) {
165 			ca->epoch_start += delta;
166 			if (after(ca->epoch_start, now))
167 				ca->epoch_start = now;
168 		}
169 		return;
170 	}
171 }
172 
173 /* calculate the cubic root of x using a table lookup followed by one
174  * Newton-Raphson iteration.
175  * Avg err ~= 0.195%
176  */
177 static u32 cubic_root(u64 a)
178 {
179 	u32 x, b, shift;
180 	/*
181 	 * cbrt(x) MSB values for x MSB values in [0..63].
182 	 * Precomputed then refined by hand - Willy Tarreau
183 	 *
184 	 * For x in [0..63],
185 	 *   v = cbrt(x << 18) - 1
186 	 *   cbrt(x) = (v[x] + 10) >> 6
187 	 */
188 	static const u8 v[] = {
189 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
190 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
191 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
192 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
193 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
194 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
195 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
196 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
197 	};
198 
199 	b = fls64(a);
200 	if (b < 7) {
201 		/* a in [0..63] */
202 		return ((u32)v[(u32)a] + 35) >> 6;
203 	}
204 
205 	b = ((b * 84) >> 8) - 1;
206 	shift = (a >> (b * 3));
207 
208 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
209 
210 	/*
211 	 * Newton-Raphson iteration
212 	 *                         2
213 	 * x    = ( 2 * x  +  a / x  ) / 3
214 	 *  k+1          k         k
215 	 */
216 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
217 	x = ((x * 341) >> 10);
218 	return x;
219 }
220 
221 /*
222  * Compute congestion window to use.
223  */
224 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
225 {
226 	u32 delta, bic_target, max_cnt;
227 	u64 offs, t;
228 
229 	ca->ack_cnt += acked;	/* count the number of ACKed packets */
230 
231 	if (ca->last_cwnd == cwnd &&
232 	    (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
233 		return;
234 
235 	/* The CUBIC function can update ca->cnt at most once per jiffy.
236 	 * On all cwnd reduction events, ca->epoch_start is set to 0,
237 	 * which will force a recalculation of ca->cnt.
238 	 */
239 	if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
240 		goto tcp_friendliness;
241 
242 	ca->last_cwnd = cwnd;
243 	ca->last_time = tcp_jiffies32;
244 
245 	if (ca->epoch_start == 0) {
246 		ca->epoch_start = tcp_jiffies32;	/* record beginning */
247 		ca->ack_cnt = acked;			/* start counting */
248 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
249 
250 		if (ca->last_max_cwnd <= cwnd) {
251 			ca->bic_K = 0;
252 			ca->bic_origin_point = cwnd;
253 		} else {
254 			/* Compute new K based on
255 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
256 			 */
257 			ca->bic_K = cubic_root(cube_factor
258 					       * (ca->last_max_cwnd - cwnd));
259 			ca->bic_origin_point = ca->last_max_cwnd;
260 		}
261 	}
262 
263 	/* cubic function - calc*/
264 	/* calculate c * time^3 / rtt,
265 	 *  while considering overflow in calculation of time^3
266 	 * (so time^3 is done by using 64 bit)
267 	 * and without the support of division of 64bit numbers
268 	 * (so all divisions are done by using 32 bit)
269 	 *  also NOTE the unit of those veriables
270 	 *	  time  = (t - K) / 2^bictcp_HZ
271 	 *	  c = bic_scale >> 10
272 	 * rtt  = (srtt >> 3) / HZ
273 	 * !!! The following code does not have overflow problems,
274 	 * if the cwnd < 1 million packets !!!
275 	 */
276 
277 	t = (s32)(tcp_jiffies32 - ca->epoch_start);
278 	t += msecs_to_jiffies(ca->delay_min >> 3);
279 	/* change the unit from HZ to bictcp_HZ */
280 	t <<= BICTCP_HZ;
281 	do_div(t, HZ);
282 
283 	if (t < ca->bic_K)		/* t - K */
284 		offs = ca->bic_K - t;
285 	else
286 		offs = t - ca->bic_K;
287 
288 	/* c/rtt * (t-K)^3 */
289 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
290 	if (t < ca->bic_K)                            /* below origin*/
291 		bic_target = ca->bic_origin_point - delta;
292 	else                                          /* above origin*/
293 		bic_target = ca->bic_origin_point + delta;
294 
295 	/* cubic function - calc bictcp_cnt*/
296 	if (bic_target > cwnd) {
297 		ca->cnt = cwnd / (bic_target - cwnd);
298 	} else {
299 		ca->cnt = 100 * cwnd;              /* very small increment*/
300 	}
301 
302 	/*
303 	 * The initial growth of cubic function may be too conservative
304 	 * when the available bandwidth is still unknown.
305 	 */
306 	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
307 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
308 
309 tcp_friendliness:
310 	/* TCP Friendly */
311 	if (tcp_friendliness) {
312 		u32 scale = beta_scale;
313 
314 		delta = (cwnd * scale) >> 3;
315 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
316 			ca->ack_cnt -= delta;
317 			ca->tcp_cwnd++;
318 		}
319 
320 		if (ca->tcp_cwnd > cwnd) {	/* if bic is slower than tcp */
321 			delta = ca->tcp_cwnd - cwnd;
322 			max_cnt = cwnd / delta;
323 			if (ca->cnt > max_cnt)
324 				ca->cnt = max_cnt;
325 		}
326 	}
327 
328 	/* The maximum rate of cwnd increase CUBIC allows is 1 packet per
329 	 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
330 	 */
331 	ca->cnt = max(ca->cnt, 2U);
332 }
333 
334 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
335 {
336 	struct tcp_sock *tp = tcp_sk(sk);
337 	struct bictcp *ca = inet_csk_ca(sk);
338 
339 	if (!tcp_is_cwnd_limited(sk))
340 		return;
341 
342 	if (tcp_in_slow_start(tp)) {
343 		if (hystart && after(ack, ca->end_seq))
344 			bictcp_hystart_reset(sk);
345 		acked = tcp_slow_start(tp, acked);
346 		if (!acked)
347 			return;
348 	}
349 	bictcp_update(ca, tp->snd_cwnd, acked);
350 	tcp_cong_avoid_ai(tp, ca->cnt, acked);
351 }
352 
353 static u32 bictcp_recalc_ssthresh(struct sock *sk)
354 {
355 	const struct tcp_sock *tp = tcp_sk(sk);
356 	struct bictcp *ca = inet_csk_ca(sk);
357 
358 	ca->epoch_start = 0;	/* end of epoch */
359 
360 	/* Wmax and fast convergence */
361 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
362 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
363 			/ (2 * BICTCP_BETA_SCALE);
364 	else
365 		ca->last_max_cwnd = tp->snd_cwnd;
366 
367 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
368 }
369 
370 static void bictcp_state(struct sock *sk, u8 new_state)
371 {
372 	if (new_state == TCP_CA_Loss) {
373 		bictcp_reset(inet_csk_ca(sk));
374 		bictcp_hystart_reset(sk);
375 	}
376 }
377 
378 static void hystart_update(struct sock *sk, u32 delay)
379 {
380 	struct tcp_sock *tp = tcp_sk(sk);
381 	struct bictcp *ca = inet_csk_ca(sk);
382 
383 	if (ca->found & hystart_detect)
384 		return;
385 
386 	if (hystart_detect & HYSTART_ACK_TRAIN) {
387 		u32 now = bictcp_clock();
388 
389 		/* first detection parameter - ack-train detection */
390 		if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
391 			ca->last_ack = now;
392 			if ((s32)(now - ca->round_start) > ca->delay_min >> 4) {
393 				ca->found |= HYSTART_ACK_TRAIN;
394 				NET_INC_STATS(sock_net(sk),
395 					      LINUX_MIB_TCPHYSTARTTRAINDETECT);
396 				NET_ADD_STATS(sock_net(sk),
397 					      LINUX_MIB_TCPHYSTARTTRAINCWND,
398 					      tp->snd_cwnd);
399 				tp->snd_ssthresh = tp->snd_cwnd;
400 			}
401 		}
402 	}
403 
404 	if (hystart_detect & HYSTART_DELAY) {
405 		/* obtain the minimum delay of more than sampling packets */
406 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
407 			if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
408 				ca->curr_rtt = delay;
409 
410 			ca->sample_cnt++;
411 		} else {
412 			if (ca->curr_rtt > ca->delay_min +
413 			    HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
414 				ca->found |= HYSTART_DELAY;
415 				NET_INC_STATS(sock_net(sk),
416 					      LINUX_MIB_TCPHYSTARTDELAYDETECT);
417 				NET_ADD_STATS(sock_net(sk),
418 					      LINUX_MIB_TCPHYSTARTDELAYCWND,
419 					      tp->snd_cwnd);
420 				tp->snd_ssthresh = tp->snd_cwnd;
421 			}
422 		}
423 	}
424 }
425 
426 /* Track delayed acknowledgment ratio using sliding window
427  * ratio = (15*ratio + sample) / 16
428  */
429 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
430 {
431 	const struct tcp_sock *tp = tcp_sk(sk);
432 	struct bictcp *ca = inet_csk_ca(sk);
433 	u32 delay;
434 
435 	/* Some calls are for duplicates without timetamps */
436 	if (sample->rtt_us < 0)
437 		return;
438 
439 	/* Discard delay samples right after fast recovery */
440 	if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
441 		return;
442 
443 	delay = (sample->rtt_us << 3) / USEC_PER_MSEC;
444 	if (delay == 0)
445 		delay = 1;
446 
447 	/* first time call or link delay decreases */
448 	if (ca->delay_min == 0 || ca->delay_min > delay)
449 		ca->delay_min = delay;
450 
451 	/* hystart triggers when cwnd is larger than some threshold */
452 	if (hystart && tcp_in_slow_start(tp) &&
453 	    tp->snd_cwnd >= hystart_low_window)
454 		hystart_update(sk, delay);
455 }
456 
457 static struct tcp_congestion_ops cubictcp __read_mostly = {
458 	.init		= bictcp_init,
459 	.ssthresh	= bictcp_recalc_ssthresh,
460 	.cong_avoid	= bictcp_cong_avoid,
461 	.set_state	= bictcp_state,
462 	.undo_cwnd	= tcp_reno_undo_cwnd,
463 	.cwnd_event	= bictcp_cwnd_event,
464 	.pkts_acked     = bictcp_acked,
465 	.owner		= THIS_MODULE,
466 	.name		= "cubic",
467 };
468 
469 static int __init cubictcp_register(void)
470 {
471 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
472 
473 	/* Precompute a bunch of the scaling factors that are used per-packet
474 	 * based on SRTT of 100ms
475 	 */
476 
477 	beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
478 		/ (BICTCP_BETA_SCALE - beta);
479 
480 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
481 
482 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
483 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
484 	 * the unit of K is bictcp_HZ=2^10, not HZ
485 	 *
486 	 *  c = bic_scale >> 10
487 	 *  rtt = 100ms
488 	 *
489 	 * the following code has been designed and tested for
490 	 * cwnd < 1 million packets
491 	 * RTT < 100 seconds
492 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
493 	 */
494 
495 	/* 1/c * 2^2*bictcp_HZ * srtt */
496 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
497 
498 	/* divide by bic_scale and by constant Srtt (100ms) */
499 	do_div(cube_factor, bic_scale * 10);
500 
501 	return tcp_register_congestion_control(&cubictcp);
502 }
503 
504 static void __exit cubictcp_unregister(void)
505 {
506 	tcp_unregister_congestion_control(&cubictcp);
507 }
508 
509 module_init(cubictcp_register);
510 module_exit(cubictcp_unregister);
511 
512 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
513 MODULE_LICENSE("GPL");
514 MODULE_DESCRIPTION("CUBIC TCP");
515 MODULE_VERSION("2.3");
516