xref: /openbmc/linux/net/ipv4/tcp_cubic.c (revision 565d76cb)
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 	u16	delayed_ack;	/* estimate the ratio of Packets/ACKs << 4 */
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->loss_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 	bictcp_reset(inet_csk_ca(sk));
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 /* calculate the cubic root of x using a table lookup followed by one
154  * Newton-Raphson iteration.
155  * Avg err ~= 0.195%
156  */
157 static u32 cubic_root(u64 a)
158 {
159 	u32 x, b, shift;
160 	/*
161 	 * cbrt(x) MSB values for x MSB values in [0..63].
162 	 * Precomputed then refined by hand - Willy Tarreau
163 	 *
164 	 * For x in [0..63],
165 	 *   v = cbrt(x << 18) - 1
166 	 *   cbrt(x) = (v[x] + 10) >> 6
167 	 */
168 	static const u8 v[] = {
169 		/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
170 		/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
171 		/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
172 		/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
173 		/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
174 		/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
175 		/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
176 		/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
177 	};
178 
179 	b = fls64(a);
180 	if (b < 7) {
181 		/* a in [0..63] */
182 		return ((u32)v[(u32)a] + 35) >> 6;
183 	}
184 
185 	b = ((b * 84) >> 8) - 1;
186 	shift = (a >> (b * 3));
187 
188 	x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
189 
190 	/*
191 	 * Newton-Raphson iteration
192 	 *                         2
193 	 * x    = ( 2 * x  +  a / x  ) / 3
194 	 *  k+1          k         k
195 	 */
196 	x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
197 	x = ((x * 341) >> 10);
198 	return x;
199 }
200 
201 /*
202  * Compute congestion window to use.
203  */
204 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
205 {
206 	u64 offs;
207 	u32 delta, t, bic_target, max_cnt;
208 
209 	ca->ack_cnt++;	/* count the number of ACKs */
210 
211 	if (ca->last_cwnd == cwnd &&
212 	    (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
213 		return;
214 
215 	ca->last_cwnd = cwnd;
216 	ca->last_time = tcp_time_stamp;
217 
218 	if (ca->epoch_start == 0) {
219 		ca->epoch_start = tcp_time_stamp;	/* record the beginning of an epoch */
220 		ca->ack_cnt = 1;			/* start counting */
221 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
222 
223 		if (ca->last_max_cwnd <= cwnd) {
224 			ca->bic_K = 0;
225 			ca->bic_origin_point = cwnd;
226 		} else {
227 			/* Compute new K based on
228 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
229 			 */
230 			ca->bic_K = cubic_root(cube_factor
231 					       * (ca->last_max_cwnd - cwnd));
232 			ca->bic_origin_point = ca->last_max_cwnd;
233 		}
234 	}
235 
236 	/* cubic function - calc*/
237 	/* calculate c * time^3 / rtt,
238 	 *  while considering overflow in calculation of time^3
239 	 * (so time^3 is done by using 64 bit)
240 	 * and without the support of division of 64bit numbers
241 	 * (so all divisions are done by using 32 bit)
242 	 *  also NOTE the unit of those veriables
243 	 *	  time  = (t - K) / 2^bictcp_HZ
244 	 *	  c = bic_scale >> 10
245 	 * rtt  = (srtt >> 3) / HZ
246 	 * !!! The following code does not have overflow problems,
247 	 * if the cwnd < 1 million packets !!!
248 	 */
249 
250 	/* change the unit from HZ to bictcp_HZ */
251 	t = ((tcp_time_stamp + msecs_to_jiffies(ca->delay_min>>3)
252 	      - ca->epoch_start) << BICTCP_HZ) / HZ;
253 
254 	if (t < ca->bic_K)		/* t - K */
255 		offs = ca->bic_K - t;
256 	else
257 		offs = t - ca->bic_K;
258 
259 	/* c/rtt * (t-K)^3 */
260 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
261 	if (t < ca->bic_K)                                	/* below origin*/
262 		bic_target = ca->bic_origin_point - delta;
263 	else                                                	/* above origin*/
264 		bic_target = ca->bic_origin_point + delta;
265 
266 	/* cubic function - calc bictcp_cnt*/
267 	if (bic_target > cwnd) {
268 		ca->cnt = cwnd / (bic_target - cwnd);
269 	} else {
270 		ca->cnt = 100 * cwnd;              /* very small increment*/
271 	}
272 
273 	/*
274 	 * The initial growth of cubic function may be too conservative
275 	 * when the available bandwidth is still unknown.
276 	 */
277 	if (ca->loss_cwnd == 0 && ca->cnt > 20)
278 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
279 
280 	/* TCP Friendly */
281 	if (tcp_friendliness) {
282 		u32 scale = beta_scale;
283 		delta = (cwnd * scale) >> 3;
284 		while (ca->ack_cnt > delta) {		/* update tcp cwnd */
285 			ca->ack_cnt -= delta;
286 			ca->tcp_cwnd++;
287 		}
288 
289 		if (ca->tcp_cwnd > cwnd){	/* if bic is slower than tcp */
290 			delta = ca->tcp_cwnd - cwnd;
291 			max_cnt = cwnd / delta;
292 			if (ca->cnt > max_cnt)
293 				ca->cnt = max_cnt;
294 		}
295 	}
296 
297 	ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
298 	if (ca->cnt == 0)			/* cannot be zero */
299 		ca->cnt = 1;
300 }
301 
302 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
303 {
304 	struct tcp_sock *tp = tcp_sk(sk);
305 	struct bictcp *ca = inet_csk_ca(sk);
306 
307 	if (!tcp_is_cwnd_limited(sk, in_flight))
308 		return;
309 
310 	if (tp->snd_cwnd <= tp->snd_ssthresh) {
311 		if (hystart && after(ack, ca->end_seq))
312 			bictcp_hystart_reset(sk);
313 		tcp_slow_start(tp);
314 	} else {
315 		bictcp_update(ca, tp->snd_cwnd);
316 		tcp_cong_avoid_ai(tp, ca->cnt);
317 	}
318 
319 }
320 
321 static u32 bictcp_recalc_ssthresh(struct sock *sk)
322 {
323 	const struct tcp_sock *tp = tcp_sk(sk);
324 	struct bictcp *ca = inet_csk_ca(sk);
325 
326 	ca->epoch_start = 0;	/* end of epoch */
327 
328 	/* Wmax and fast convergence */
329 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
330 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
331 			/ (2 * BICTCP_BETA_SCALE);
332 	else
333 		ca->last_max_cwnd = tp->snd_cwnd;
334 
335 	ca->loss_cwnd = tp->snd_cwnd;
336 
337 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
338 }
339 
340 static u32 bictcp_undo_cwnd(struct sock *sk)
341 {
342 	struct bictcp *ca = inet_csk_ca(sk);
343 
344 	return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
345 }
346 
347 static void bictcp_state(struct sock *sk, u8 new_state)
348 {
349 	if (new_state == TCP_CA_Loss) {
350 		bictcp_reset(inet_csk_ca(sk));
351 		bictcp_hystart_reset(sk);
352 	}
353 }
354 
355 static void hystart_update(struct sock *sk, u32 delay)
356 {
357 	struct tcp_sock *tp = tcp_sk(sk);
358 	struct bictcp *ca = inet_csk_ca(sk);
359 
360 	if (!(ca->found & hystart_detect)) {
361 		u32 now = bictcp_clock();
362 
363 		/* first detection parameter - ack-train detection */
364 		if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
365 			ca->last_ack = now;
366 			if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
367 				ca->found |= HYSTART_ACK_TRAIN;
368 		}
369 
370 		/* obtain the minimum delay of more than sampling packets */
371 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
372 			if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
373 				ca->curr_rtt = delay;
374 
375 			ca->sample_cnt++;
376 		} else {
377 			if (ca->curr_rtt > ca->delay_min +
378 			    HYSTART_DELAY_THRESH(ca->delay_min>>4))
379 				ca->found |= HYSTART_DELAY;
380 		}
381 		/*
382 		 * Either one of two conditions are met,
383 		 * we exit from slow start immediately.
384 		 */
385 		if (ca->found & hystart_detect)
386 			tp->snd_ssthresh = tp->snd_cwnd;
387 	}
388 }
389 
390 /* Track delayed acknowledgment ratio using sliding window
391  * ratio = (15*ratio + sample) / 16
392  */
393 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
394 {
395 	const struct inet_connection_sock *icsk = inet_csk(sk);
396 	const struct tcp_sock *tp = tcp_sk(sk);
397 	struct bictcp *ca = inet_csk_ca(sk);
398 	u32 delay;
399 
400 	if (icsk->icsk_ca_state == TCP_CA_Open) {
401 		cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
402 		ca->delayed_ack += cnt;
403 	}
404 
405 	/* Some calls are for duplicates without timetamps */
406 	if (rtt_us < 0)
407 		return;
408 
409 	/* Discard delay samples right after fast recovery */
410 	if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
411 		return;
412 
413 	delay = (rtt_us << 3) / USEC_PER_MSEC;
414 	if (delay == 0)
415 		delay = 1;
416 
417 	/* first time call or link delay decreases */
418 	if (ca->delay_min == 0 || ca->delay_min > delay)
419 		ca->delay_min = delay;
420 
421 	/* hystart triggers when cwnd is larger than some threshold */
422 	if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
423 	    tp->snd_cwnd >= hystart_low_window)
424 		hystart_update(sk, delay);
425 }
426 
427 static struct tcp_congestion_ops cubictcp __read_mostly = {
428 	.init		= bictcp_init,
429 	.ssthresh	= bictcp_recalc_ssthresh,
430 	.cong_avoid	= bictcp_cong_avoid,
431 	.set_state	= bictcp_state,
432 	.undo_cwnd	= bictcp_undo_cwnd,
433 	.pkts_acked     = bictcp_acked,
434 	.owner		= THIS_MODULE,
435 	.name		= "cubic",
436 };
437 
438 static int __init cubictcp_register(void)
439 {
440 	BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
441 
442 	/* Precompute a bunch of the scaling factors that are used per-packet
443 	 * based on SRTT of 100ms
444 	 */
445 
446 	beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
447 
448 	cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
449 
450 	/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
451 	 *  so K = cubic_root( (wmax-cwnd)*rtt/c )
452 	 * the unit of K is bictcp_HZ=2^10, not HZ
453 	 *
454 	 *  c = bic_scale >> 10
455 	 *  rtt = 100ms
456 	 *
457 	 * the following code has been designed and tested for
458 	 * cwnd < 1 million packets
459 	 * RTT < 100 seconds
460 	 * HZ < 1,000,00  (corresponding to 10 nano-second)
461 	 */
462 
463 	/* 1/c * 2^2*bictcp_HZ * srtt */
464 	cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
465 
466 	/* divide by bic_scale and by constant Srtt (100ms) */
467 	do_div(cube_factor, bic_scale * 10);
468 
469 	/* hystart needs ms clock resolution */
470 	if (hystart && HZ < 1000)
471 		cubictcp.flags |= TCP_CONG_RTT_STAMP;
472 
473 	return tcp_register_congestion_control(&cubictcp);
474 }
475 
476 static void __exit cubictcp_unregister(void)
477 {
478 	tcp_unregister_congestion_control(&cubictcp);
479 }
480 
481 module_init(cubictcp_register);
482 module_exit(cubictcp_unregister);
483 
484 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
485 MODULE_LICENSE("GPL");
486 MODULE_DESCRIPTION("CUBIC TCP");
487 MODULE_VERSION("2.3");
488