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
bictcp_reset(struct bictcp * ca)105 static inline void bictcp_reset(struct bictcp *ca)
106 {
107 memset(ca, 0, offsetof(struct bictcp, unused));
108 ca->found = 0;
109 }
110
bictcp_clock_us(const struct sock * sk)111 static inline u32 bictcp_clock_us(const struct sock *sk)
112 {
113 return tcp_sk(sk)->tcp_mstamp;
114 }
115
bictcp_hystart_reset(struct sock * sk)116 static inline void bictcp_hystart_reset(struct sock *sk)
117 {
118 struct tcp_sock *tp = tcp_sk(sk);
119 struct bictcp *ca = inet_csk_ca(sk);
120
121 ca->round_start = ca->last_ack = bictcp_clock_us(sk);
122 ca->end_seq = tp->snd_nxt;
123 ca->curr_rtt = ~0U;
124 ca->sample_cnt = 0;
125 }
126
cubictcp_init(struct sock * sk)127 static void cubictcp_init(struct sock *sk)
128 {
129 struct bictcp *ca = inet_csk_ca(sk);
130
131 bictcp_reset(ca);
132
133 if (hystart)
134 bictcp_hystart_reset(sk);
135
136 if (!hystart && initial_ssthresh)
137 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
138 }
139
cubictcp_cwnd_event(struct sock * sk,enum tcp_ca_event event)140 static void cubictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
141 {
142 if (event == CA_EVENT_TX_START) {
143 struct bictcp *ca = inet_csk_ca(sk);
144 u32 now = tcp_jiffies32;
145 s32 delta;
146
147 delta = now - tcp_sk(sk)->lsndtime;
148
149 /* We were application limited (idle) for a while.
150 * Shift epoch_start to keep cwnd growth to cubic curve.
151 */
152 if (ca->epoch_start && delta > 0) {
153 ca->epoch_start += delta;
154 if (after(ca->epoch_start, now))
155 ca->epoch_start = now;
156 }
157 return;
158 }
159 }
160
161 /* calculate the cubic root of x using a table lookup followed by one
162 * Newton-Raphson iteration.
163 * Avg err ~= 0.195%
164 */
cubic_root(u64 a)165 static u32 cubic_root(u64 a)
166 {
167 u32 x, b, shift;
168 /*
169 * cbrt(x) MSB values for x MSB values in [0..63].
170 * Precomputed then refined by hand - Willy Tarreau
171 *
172 * For x in [0..63],
173 * v = cbrt(x << 18) - 1
174 * cbrt(x) = (v[x] + 10) >> 6
175 */
176 static const u8 v[] = {
177 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
178 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
179 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
180 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
181 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
182 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
183 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
184 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
185 };
186
187 b = fls64(a);
188 if (b < 7) {
189 /* a in [0..63] */
190 return ((u32)v[(u32)a] + 35) >> 6;
191 }
192
193 b = ((b * 84) >> 8) - 1;
194 shift = (a >> (b * 3));
195
196 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
197
198 /*
199 * Newton-Raphson iteration
200 * 2
201 * x = ( 2 * x + a / x ) / 3
202 * k+1 k k
203 */
204 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
205 x = ((x * 341) >> 10);
206 return x;
207 }
208
209 /*
210 * Compute congestion window to use.
211 */
bictcp_update(struct bictcp * ca,u32 cwnd,u32 acked)212 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
213 {
214 u32 delta, bic_target, max_cnt;
215 u64 offs, t;
216
217 ca->ack_cnt += acked; /* count the number of ACKed packets */
218
219 if (ca->last_cwnd == cwnd &&
220 (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
221 return;
222
223 /* The CUBIC function can update ca->cnt at most once per jiffy.
224 * On all cwnd reduction events, ca->epoch_start is set to 0,
225 * which will force a recalculation of ca->cnt.
226 */
227 if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
228 goto tcp_friendliness;
229
230 ca->last_cwnd = cwnd;
231 ca->last_time = tcp_jiffies32;
232
233 if (ca->epoch_start == 0) {
234 ca->epoch_start = tcp_jiffies32; /* record beginning */
235 ca->ack_cnt = acked; /* start counting */
236 ca->tcp_cwnd = cwnd; /* syn with cubic */
237
238 if (ca->last_max_cwnd <= cwnd) {
239 ca->bic_K = 0;
240 ca->bic_origin_point = cwnd;
241 } else {
242 /* Compute new K based on
243 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
244 */
245 ca->bic_K = cubic_root(cube_factor
246 * (ca->last_max_cwnd - cwnd));
247 ca->bic_origin_point = ca->last_max_cwnd;
248 }
249 }
250
251 /* cubic function - calc*/
252 /* calculate c * time^3 / rtt,
253 * while considering overflow in calculation of time^3
254 * (so time^3 is done by using 64 bit)
255 * and without the support of division of 64bit numbers
256 * (so all divisions are done by using 32 bit)
257 * also NOTE the unit of those veriables
258 * time = (t - K) / 2^bictcp_HZ
259 * c = bic_scale >> 10
260 * rtt = (srtt >> 3) / HZ
261 * !!! The following code does not have overflow problems,
262 * if the cwnd < 1 million packets !!!
263 */
264
265 t = (s32)(tcp_jiffies32 - ca->epoch_start);
266 t += usecs_to_jiffies(ca->delay_min);
267 /* change the unit from HZ to bictcp_HZ */
268 t <<= BICTCP_HZ;
269 do_div(t, HZ);
270
271 if (t < ca->bic_K) /* t - K */
272 offs = ca->bic_K - t;
273 else
274 offs = t - ca->bic_K;
275
276 /* c/rtt * (t-K)^3 */
277 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
278 if (t < ca->bic_K) /* below origin*/
279 bic_target = ca->bic_origin_point - delta;
280 else /* above origin*/
281 bic_target = ca->bic_origin_point + delta;
282
283 /* cubic function - calc bictcp_cnt*/
284 if (bic_target > cwnd) {
285 ca->cnt = cwnd / (bic_target - cwnd);
286 } else {
287 ca->cnt = 100 * cwnd; /* very small increment*/
288 }
289
290 /*
291 * The initial growth of cubic function may be too conservative
292 * when the available bandwidth is still unknown.
293 */
294 if (ca->last_max_cwnd == 0 && ca->cnt > 20)
295 ca->cnt = 20; /* increase cwnd 5% per RTT */
296
297 tcp_friendliness:
298 /* TCP Friendly */
299 if (tcp_friendliness) {
300 u32 scale = beta_scale;
301
302 delta = (cwnd * scale) >> 3;
303 while (ca->ack_cnt > delta) { /* update tcp cwnd */
304 ca->ack_cnt -= delta;
305 ca->tcp_cwnd++;
306 }
307
308 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */
309 delta = ca->tcp_cwnd - cwnd;
310 max_cnt = cwnd / delta;
311 if (ca->cnt > max_cnt)
312 ca->cnt = max_cnt;
313 }
314 }
315
316 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per
317 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
318 */
319 ca->cnt = max(ca->cnt, 2U);
320 }
321
cubictcp_cong_avoid(struct sock * sk,u32 ack,u32 acked)322 static void cubictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
323 {
324 struct tcp_sock *tp = tcp_sk(sk);
325 struct bictcp *ca = inet_csk_ca(sk);
326
327 if (!tcp_is_cwnd_limited(sk))
328 return;
329
330 if (tcp_in_slow_start(tp)) {
331 if (hystart && after(ack, ca->end_seq))
332 bictcp_hystart_reset(sk);
333 acked = tcp_slow_start(tp, acked);
334 if (!acked)
335 return;
336 }
337 bictcp_update(ca, tp->snd_cwnd, acked);
338 tcp_cong_avoid_ai(tp, ca->cnt, acked);
339 }
340
cubictcp_recalc_ssthresh(struct sock * sk)341 static u32 cubictcp_recalc_ssthresh(struct sock *sk)
342 {
343 const struct tcp_sock *tp = tcp_sk(sk);
344 struct bictcp *ca = inet_csk_ca(sk);
345
346 ca->epoch_start = 0; /* end of epoch */
347
348 /* Wmax and fast convergence */
349 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
350 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
351 / (2 * BICTCP_BETA_SCALE);
352 else
353 ca->last_max_cwnd = tp->snd_cwnd;
354
355 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
356 }
357
cubictcp_state(struct sock * sk,u8 new_state)358 static void cubictcp_state(struct sock *sk, u8 new_state)
359 {
360 if (new_state == TCP_CA_Loss) {
361 bictcp_reset(inet_csk_ca(sk));
362 bictcp_hystart_reset(sk);
363 }
364 }
365
366 /* Account for TSO/GRO delays.
367 * Otherwise short RTT flows could get too small ssthresh, since during
368 * slow start we begin with small TSO packets and ca->delay_min would
369 * not account for long aggregation delay when TSO packets get bigger.
370 * Ideally even with a very small RTT we would like to have at least one
371 * TSO packet being sent and received by GRO, and another one in qdisc layer.
372 * We apply another 100% factor because @rate is doubled at this point.
373 * We cap the cushion to 1ms.
374 */
hystart_ack_delay(struct sock * sk)375 static u32 hystart_ack_delay(struct sock *sk)
376 {
377 unsigned long rate;
378
379 rate = READ_ONCE(sk->sk_pacing_rate);
380 if (!rate)
381 return 0;
382 return min_t(u64, USEC_PER_MSEC,
383 div64_ul((u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate));
384 }
385
hystart_update(struct sock * sk,u32 delay)386 static void hystart_update(struct sock *sk, u32 delay)
387 {
388 struct tcp_sock *tp = tcp_sk(sk);
389 struct bictcp *ca = inet_csk_ca(sk);
390 u32 threshold;
391
392 if (hystart_detect & HYSTART_ACK_TRAIN) {
393 u32 now = bictcp_clock_us(sk);
394
395 /* first detection parameter - ack-train detection */
396 if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) {
397 ca->last_ack = now;
398
399 threshold = ca->delay_min + hystart_ack_delay(sk);
400
401 /* Hystart ack train triggers if we get ack past
402 * ca->delay_min/2.
403 * Pacing might have delayed packets up to RTT/2
404 * during slow start.
405 */
406 if (sk->sk_pacing_status == SK_PACING_NONE)
407 threshold >>= 1;
408
409 if ((s32)(now - ca->round_start) > threshold) {
410 ca->found = 1;
411 pr_debug("hystart_ack_train (%u > %u) delay_min %u (+ ack_delay %u) cwnd %u\n",
412 now - ca->round_start, threshold,
413 ca->delay_min, hystart_ack_delay(sk), tp->snd_cwnd);
414 NET_INC_STATS(sock_net(sk),
415 LINUX_MIB_TCPHYSTARTTRAINDETECT);
416 NET_ADD_STATS(sock_net(sk),
417 LINUX_MIB_TCPHYSTARTTRAINCWND,
418 tp->snd_cwnd);
419 tp->snd_ssthresh = tp->snd_cwnd;
420 }
421 }
422 }
423
424 if (hystart_detect & HYSTART_DELAY) {
425 /* obtain the minimum delay of more than sampling packets */
426 if (ca->curr_rtt > delay)
427 ca->curr_rtt = delay;
428 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
429 ca->sample_cnt++;
430 } else {
431 if (ca->curr_rtt > ca->delay_min +
432 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
433 ca->found = 1;
434 NET_INC_STATS(sock_net(sk),
435 LINUX_MIB_TCPHYSTARTDELAYDETECT);
436 NET_ADD_STATS(sock_net(sk),
437 LINUX_MIB_TCPHYSTARTDELAYCWND,
438 tp->snd_cwnd);
439 tp->snd_ssthresh = tp->snd_cwnd;
440 }
441 }
442 }
443 }
444
cubictcp_acked(struct sock * sk,const struct ack_sample * sample)445 static void cubictcp_acked(struct sock *sk, const struct ack_sample *sample)
446 {
447 const struct tcp_sock *tp = tcp_sk(sk);
448 struct bictcp *ca = inet_csk_ca(sk);
449 u32 delay;
450
451 /* Some calls are for duplicates without timetamps */
452 if (sample->rtt_us < 0)
453 return;
454
455 /* Discard delay samples right after fast recovery */
456 if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
457 return;
458
459 delay = sample->rtt_us;
460 if (delay == 0)
461 delay = 1;
462
463 /* first time call or link delay decreases */
464 if (ca->delay_min == 0 || ca->delay_min > delay)
465 ca->delay_min = delay;
466
467 /* hystart triggers when cwnd is larger than some threshold */
468 if (!ca->found && tcp_in_slow_start(tp) && hystart &&
469 tp->snd_cwnd >= hystart_low_window)
470 hystart_update(sk, delay);
471 }
472
473 static struct tcp_congestion_ops cubictcp __read_mostly = {
474 .init = cubictcp_init,
475 .ssthresh = cubictcp_recalc_ssthresh,
476 .cong_avoid = cubictcp_cong_avoid,
477 .set_state = cubictcp_state,
478 .undo_cwnd = tcp_reno_undo_cwnd,
479 .cwnd_event = cubictcp_cwnd_event,
480 .pkts_acked = cubictcp_acked,
481 .owner = THIS_MODULE,
482 .name = "cubic",
483 };
484
cubictcp_register(void)485 static int __init cubictcp_register(void)
486 {
487 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
488
489 /* Precompute a bunch of the scaling factors that are used per-packet
490 * based on SRTT of 100ms
491 */
492
493 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
494 / (BICTCP_BETA_SCALE - beta);
495
496 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
497
498 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
499 * so K = cubic_root( (wmax-cwnd)*rtt/c )
500 * the unit of K is bictcp_HZ=2^10, not HZ
501 *
502 * c = bic_scale >> 10
503 * rtt = 100ms
504 *
505 * the following code has been designed and tested for
506 * cwnd < 1 million packets
507 * RTT < 100 seconds
508 * HZ < 1,000,00 (corresponding to 10 nano-second)
509 */
510
511 /* 1/c * 2^2*bictcp_HZ * srtt */
512 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
513
514 /* divide by bic_scale and by constant Srtt (100ms) */
515 do_div(cube_factor, bic_scale * 10);
516
517 return tcp_register_congestion_control(&cubictcp);
518 }
519
cubictcp_unregister(void)520 static void __exit cubictcp_unregister(void)
521 {
522 tcp_unregister_congestion_control(&cubictcp);
523 }
524
525 module_init(cubictcp_register);
526 module_exit(cubictcp_unregister);
527
528 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
529 MODULE_LICENSE("GPL");
530 MODULE_DESCRIPTION("CUBIC TCP");
531 MODULE_VERSION("2.3");
532