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