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
46static int fast_convergence __read_mostly = 1;
47static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
48static int initial_ssthresh __read_mostly;
49static int bic_scale __read_mostly = 41;
50static int tcp_friendliness __read_mostly = 1;
51
52static int hystart __read_mostly = 1;
53static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
54static int hystart_low_window __read_mostly = 16;
55static int hystart_ack_delta __read_mostly = 2;
56
57static u32 cube_rtt_scale __read_mostly;
58static u32 beta_scale __read_mostly;
59static u64 cube_factor __read_mostly;
60
61/* Note parameters that are used for precomputing scale factors are read-only */
62module_param(fast_convergence, int, 0644);
63MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
64module_param(beta, int, 0644);
65MODULE_PARM_DESC(beta, "beta for multiplicative increase");
66module_param(initial_ssthresh, int, 0644);
67MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
68module_param(bic_scale, int, 0444);
69MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
70module_param(tcp_friendliness, int, 0644);
71MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
72module_param(hystart, int, 0644);
73MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
74module_param(hystart_detect, int, 0644);
75MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms"
76 " 1: packet-train 2: delay 3: both packet-train and delay");
77module_param(hystart_low_window, int, 0644);
78MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
79module_param(hystart_ack_delta, int, 0644);
80MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
81
82/* BIC TCP Parameters */
83struct 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
104static 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
119static 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
128static 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
139static 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
152static 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 */
177static 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 */
224static 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
309tcp_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
334static 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
353static 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
370static 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
378static 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 */
429static 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
457static 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
469static 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
504static void __exit cubictcp_unregister(void)
505{
506 tcp_unregister_congestion_control(&cubictcp);
507}
508
509module_init(cubictcp_register);
510module_exit(cubictcp_unregister);
511
512MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
513MODULE_LICENSE("GPL");
514MODULE_DESCRIPTION("CUBIC TCP");
515MODULE_VERSION("2.3");
516