xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision 63dc02bd)
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, "hyrbrid 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	loss_cwnd;	/* congestion window at last loss */
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 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 #define ACK_RATIO_SHIFT	4
96 #define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT)
97 	u16	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
98 	u8	sample_cnt;	/* number of samples to decide curr_rtt */
99 	u8	found;		/* the exit point is found? */
100 	u32	round_start;	/* beginning of each round */
101 	u32	end_seq;	/* end_seq of the round */
102 	u32	last_ack;	/* last time when the ACK spacing is close */
103 	u32	curr_rtt;	/* the minimum rtt of current round */
104 };
105 
106 static inline void bictcp_reset(struct bictcp *ca)
107 {
108 	ca->cnt = 0;
109 	ca->last_max_cwnd = 0;
110 	ca->last_cwnd = 0;
111 	ca->last_time = 0;
112 	ca->bic_origin_point = 0;
113 	ca->bic_K = 0;
114 	ca->delay_min = 0;
115 	ca->epoch_start = 0;
116 	ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
117 	ca->ack_cnt = 0;
118 	ca->tcp_cwnd = 0;
119 	ca->found = 0;
120 }
121 
122 static inline u32 bictcp_clock(void)
123 {
124 #if HZ < 1000
125 	return ktime_to_ms(ktime_get_real());
126 #else
127 	return jiffies_to_msecs(jiffies);
128 #endif
129 }
130 
131 static inline void bictcp_hystart_reset(struct sock *sk)
132 {
133 	struct tcp_sock *tp = tcp_sk(sk);
134 	struct bictcp *ca = inet_csk_ca(sk);
135 
136 	ca->round_start = ca->last_ack = bictcp_clock();
137 	ca->end_seq = tp->snd_nxt;
138 	ca->curr_rtt = 0;
139 	ca->sample_cnt = 0;
140 }
141 
142 static void bictcp_init(struct sock *sk)
143 {
144 	struct bictcp *ca = inet_csk_ca(sk);
145 
146 	bictcp_reset(ca);
147 	ca->loss_cwnd = 0;
148 
149 	if (hystart)
150 		bictcp_hystart_reset(sk);
151 
152 	if (!hystart && initial_ssthresh)
153 		tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
154 }
155 
156 /* calculate the cubic root of x using a table lookup followed by one
157  * Newton-Raphson iteration.
158  * Avg err ~= 0.195%
159  */
160 static u32 cubic_root(u64 a)
161 {
162 	u32 x, b, shift;
163 	/*
164 	 * cbrt(x) MSB values for x MSB values in [0..63].
165 	 * Precomputed then refined by hand - Willy Tarreau
166 	 *
167 	 * For x in [0..63],
168 	 *   v = cbrt(x << 18) - 1
169 	 *   cbrt(x) = (v[x] + 10) >> 6
170 	 */
171 	static const u8 v[] = {
172 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
173 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
174 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
175 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
176 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
177 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
178 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
179 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
180 	};
181 
182 	b = fls64(a);
183 	if (b < 7) {
184 		/* a in [0..63] */
185 		return ((u32)v[(u32)a] + 35) >> 6;
186 	}
187 
188 	b = ((b * 84) >> 8) - 1;
189 	shift = (a >> (b * 3));
190 
191 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
192 
193 	/*
194 	 * Newton-Raphson iteration
195 	 *                         2
196 	 * x    = ( 2 * x  +  a / x  ) / 3
197 	 *  k+1          k         k
198 	 */
199 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
200 	x = ((x * 341) >> 10);
201 	return x;
202 }
203 
204 /*
205  * Compute congestion window to use.
206  */
207 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
208 {
209 	u64 offs;
210 	u32 delta, t, bic_target, max_cnt;
211 
212 	ca->ack_cnt++;	/* count the number of ACKs */
213 
214 	if (ca->last_cwnd == cwnd &&
215 	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
216 		return;
217 
218 	ca->last_cwnd = cwnd;
219 	ca->last_time = tcp_time_stamp;
220 
221 	if (ca->epoch_start == 0) {
222 		ca->epoch_start = tcp_time_stamp;	/* record the beginning of an epoch */
223 		ca->ack_cnt = 1;			/* start counting */
224 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
225 
226 		if (ca->last_max_cwnd <= cwnd) {
227 			ca->bic_K = 0;
228 			ca->bic_origin_point = cwnd;
229 		} else {
230 			/* Compute new K based on
231 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
232 			 */
233 			ca->bic_K = cubic_root(cube_factor
234 					       * (ca->last_max_cwnd - cwnd));
235 			ca->bic_origin_point = ca->last_max_cwnd;
236 		}
237 	}
238 
239 	/* cubic function - calc*/
240 	/* calculate c * time^3 / rtt,
241 	 *  while considering overflow in calculation of time^3
242 	 * (so time^3 is done by using 64 bit)
243 	 * and without the support of division of 64bit numbers
244 	 * (so all divisions are done by using 32 bit)
245 	 *  also NOTE the unit of those veriables
246 	 *	  time  = (t - K) / 2^bictcp_HZ
247 	 *	  c = bic_scale >> 10
248 	 * rtt  = (srtt >> 3) / HZ
249 	 * !!! The following code does not have overflow problems,
250 	 * if the cwnd < 1 million packets !!!
251 	 */
252 
253 	/* change the unit from HZ to bictcp_HZ */
254 	t = ((tcp_time_stamp + msecs_to_jiffies(ca->delay_min>>3)
255 	      - ca->epoch_start) << BICTCP_HZ) / HZ;
256 
257 	if (t < ca->bic_K)		/* t - K */
258 		offs = ca->bic_K - t;
259 	else
260 		offs = t - ca->bic_K;
261 
262 	/* c/rtt * (t-K)^3 */
263 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
264 	if (t < ca->bic_K)                                	/* below origin*/
265 		bic_target = ca->bic_origin_point - delta;
266 	else                                                	/* above origin*/
267 		bic_target = ca->bic_origin_point + delta;
268 
269 	/* cubic function - calc bictcp_cnt*/
270 	if (bic_target > cwnd) {
271 		ca->cnt = cwnd / (bic_target - cwnd);
272 	} else {
273 		ca->cnt = 100 * cwnd;              /* very small increment*/
274 	}
275 
276 	/*
277 	 * The initial growth of cubic function may be too conservative
278 	 * when the available bandwidth is still unknown.
279 	 */
280 	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
281 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
282 
283 	/* TCP Friendly */
284 	if (tcp_friendliness) {
285 		u32 scale = beta_scale;
286 		delta = (cwnd * scale) >> 3;
287 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
288 			ca->ack_cnt -= delta;
289 			ca->tcp_cwnd++;
290 		}
291 
292 		if (ca->tcp_cwnd > cwnd){	/* if bic is slower than tcp */
293 			delta = ca->tcp_cwnd - cwnd;
294 			max_cnt = cwnd / delta;
295 			if (ca->cnt > max_cnt)
296 				ca->cnt = max_cnt;
297 		}
298 	}
299 
300 	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
301 	if (ca->cnt == 0)			/* cannot be zero */
302 		ca->cnt = 1;
303 }
304 
305 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
306 {
307 	struct tcp_sock *tp = tcp_sk(sk);
308 	struct bictcp *ca = inet_csk_ca(sk);
309 
310 	if (!tcp_is_cwnd_limited(sk, in_flight))
311 		return;
312 
313 	if (tp->snd_cwnd <= tp->snd_ssthresh) {
314 		if (hystart && after(ack, ca->end_seq))
315 			bictcp_hystart_reset(sk);
316 		tcp_slow_start(tp);
317 	} else {
318 		bictcp_update(ca, tp->snd_cwnd);
319 		tcp_cong_avoid_ai(tp, ca->cnt);
320 	}
321 
322 }
323 
324 static u32 bictcp_recalc_ssthresh(struct sock *sk)
325 {
326 	const struct tcp_sock *tp = tcp_sk(sk);
327 	struct bictcp *ca = inet_csk_ca(sk);
328 
329 	ca->epoch_start = 0;	/* end of epoch */
330 
331 	/* Wmax and fast convergence */
332 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
333 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
334 			/ (2 * BICTCP_BETA_SCALE);
335 	else
336 		ca->last_max_cwnd = tp->snd_cwnd;
337 
338 	ca->loss_cwnd = tp->snd_cwnd;
339 
340 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
341 }
342 
343 static u32 bictcp_undo_cwnd(struct sock *sk)
344 {
345 	struct bictcp *ca = inet_csk_ca(sk);
346 
347 	return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
348 }
349 
350 static void bictcp_state(struct sock *sk, u8 new_state)
351 {
352 	if (new_state == TCP_CA_Loss) {
353 		bictcp_reset(inet_csk_ca(sk));
354 		bictcp_hystart_reset(sk);
355 	}
356 }
357 
358 static void hystart_update(struct sock *sk, u32 delay)
359 {
360 	struct tcp_sock *tp = tcp_sk(sk);
361 	struct bictcp *ca = inet_csk_ca(sk);
362 
363 	if (!(ca->found & hystart_detect)) {
364 		u32 now = bictcp_clock();
365 
366 		/* first detection parameter - ack-train detection */
367 		if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
368 			ca->last_ack = now;
369 			if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
370 				ca->found |= HYSTART_ACK_TRAIN;
371 		}
372 
373 		/* obtain the minimum delay of more than sampling packets */
374 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
375 			if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
376 				ca->curr_rtt = delay;
377 
378 			ca->sample_cnt++;
379 		} else {
380 			if (ca->curr_rtt > ca->delay_min +
381 			    HYSTART_DELAY_THRESH(ca->delay_min>>4))
382 				ca->found |= HYSTART_DELAY;
383 		}
384 		/*
385 		 * Either one of two conditions are met,
386 		 * we exit from slow start immediately.
387 		 */
388 		if (ca->found & hystart_detect)
389 			tp->snd_ssthresh = tp->snd_cwnd;
390 	}
391 }
392 
393 /* Track delayed acknowledgment ratio using sliding window
394  * ratio = (15*ratio + sample) / 16
395  */
396 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
397 {
398 	const struct inet_connection_sock *icsk = inet_csk(sk);
399 	const struct tcp_sock *tp = tcp_sk(sk);
400 	struct bictcp *ca = inet_csk_ca(sk);
401 	u32 delay;
402 
403 	if (icsk->icsk_ca_state == TCP_CA_Open) {
404 		u32 ratio = ca->delayed_ack;
405 
406 		ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT;
407 		ratio += cnt;
408 
409 		ca->delayed_ack = min(ratio, ACK_RATIO_LIMIT);
410 	}
411 
412 	/* Some calls are for duplicates without timetamps */
413 	if (rtt_us < 0)
414 		return;
415 
416 	/* Discard delay samples right after fast recovery */
417 	if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
418 		return;
419 
420 	delay = (rtt_us << 3) / USEC_PER_MSEC;
421 	if (delay == 0)
422 		delay = 1;
423 
424 	/* first time call or link delay decreases */
425 	if (ca->delay_min == 0 || ca->delay_min > delay)
426 		ca->delay_min = delay;
427 
428 	/* hystart triggers when cwnd is larger than some threshold */
429 	if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
430 	    tp->snd_cwnd >= hystart_low_window)
431 		hystart_update(sk, delay);
432 }
433 
434 static struct tcp_congestion_ops cubictcp __read_mostly = {
435 	.init		= bictcp_init,
436 	.ssthresh	= bictcp_recalc_ssthresh,
437 	.cong_avoid	= bictcp_cong_avoid,
438 	.set_state	= bictcp_state,
439 	.undo_cwnd	= bictcp_undo_cwnd,
440 	.pkts_acked     = bictcp_acked,
441 	.owner		= THIS_MODULE,
442 	.name		= "cubic",
443 };
444 
445 static int __init cubictcp_register(void)
446 {
447 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
448 
449 	/* Precompute a bunch of the scaling factors that are used per-packet
450 	 * based on SRTT of 100ms
451 	 */
452 
453 	beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
454 
455 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
456 
457 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
458 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
459 	 * the unit of K is bictcp_HZ=2^10, not HZ
460 	 *
461 	 *  c = bic_scale >> 10
462 	 *  rtt = 100ms
463 	 *
464 	 * the following code has been designed and tested for
465 	 * cwnd < 1 million packets
466 	 * RTT < 100 seconds
467 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
468 	 */
469 
470 	/* 1/c * 2^2*bictcp_HZ * srtt */
471 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
472 
473 	/* divide by bic_scale and by constant Srtt (100ms) */
474 	do_div(cube_factor, bic_scale * 10);
475 
476 	/* hystart needs ms clock resolution */
477 	if (hystart && HZ < 1000)
478 		cubictcp.flags |= TCP_CONG_RTT_STAMP;
479 
480 	return tcp_register_congestion_control(&cubictcp);
481 }
482 
483 static void __exit cubictcp_unregister(void)
484 {
485 	tcp_unregister_congestion_control(&cubictcp);
486 }
487 
488 module_init(cubictcp_register);
489 module_exit(cubictcp_unregister);
490 
491 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
492 MODULE_LICENSE("GPL");
493 MODULE_DESCRIPTION("CUBIC TCP");
494 MODULE_VERSION("2.3");
495