2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
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.
10 * http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
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.
17 * http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
19 * All testing results are available from:
20 * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
22 * Unless CUBIC is enabled and congestion window is large
23 * this behaves the same as the original Reno.
27 #include <linux/module.h>
28 #include <linux/math64.h>
31 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
32 * max_cwnd = snd_cwnd * beta
34 #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
36 /* Two methods of hybrid slow start */
37 #define HYSTART_ACK_TRAIN 0x1
38 #define HYSTART_DELAY 0x2
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)
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;
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;
57 static u32 cube_rtt_scale __read_mostly;
58 static u32 beta_scale __read_mostly;
59 static u64 cube_factor __read_mostly;
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)");
82 /* BIC TCP Parameters */
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 */
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 */
105 static inline void bictcp_reset(struct bictcp *ca)
108 ca->last_max_cwnd = 0;
111 ca->bic_origin_point = 0;
120 static inline u32 bictcp_clock(void)
123 return ktime_to_ms(ktime_get_real());
125 return jiffies_to_msecs(jiffies);
129 static inline void bictcp_hystart_reset(struct sock *sk)
131 struct tcp_sock *tp = tcp_sk(sk);
132 struct bictcp *ca = inet_csk_ca(sk);
134 ca->round_start = ca->last_ack = bictcp_clock();
135 ca->end_seq = tp->snd_nxt;
140 static void bictcp_init(struct sock *sk)
142 struct bictcp *ca = inet_csk_ca(sk);
148 bictcp_hystart_reset(sk);
150 if (!hystart && initial_ssthresh)
151 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
154 /* calculate the cubic root of x using a table lookup followed by one
155 * Newton-Raphson iteration.
158 static u32 cubic_root(u64 a)
162 * cbrt(x) MSB values for x MSB values in [0..63].
163 * Precomputed then refined by hand - Willy Tarreau
166 * v = cbrt(x << 18) - 1
167 * cbrt(x) = (v[x] + 10) >> 6
169 static const u8 v[] = {
170 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
171 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
172 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
173 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
174 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
175 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
176 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
177 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
183 return ((u32)v[(u32)a] + 35) >> 6;
186 b = ((b * 84) >> 8) - 1;
187 shift = (a >> (b * 3));
189 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
192 * Newton-Raphson iteration
194 * x = ( 2 * x + a / x ) / 3
197 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
198 x = ((x * 341) >> 10);
203 * Compute congestion window to use.
205 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
207 u32 delta, bic_target, max_cnt;
210 ca->ack_cnt += acked; /* count the number of ACKed packets */
212 if (ca->last_cwnd == cwnd &&
213 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
216 /* The CUBIC function can update ca->cnt at most once per jiffy.
217 * On all cwnd reduction events, ca->epoch_start is set to 0,
218 * which will force a recalculation of ca->cnt.
220 if (ca->epoch_start && tcp_time_stamp == ca->last_time)
221 goto tcp_friendliness;
223 ca->last_cwnd = cwnd;
224 ca->last_time = tcp_time_stamp;
226 if (ca->epoch_start == 0) {
227 ca->epoch_start = tcp_time_stamp; /* record beginning */
228 ca->ack_cnt = acked; /* start counting */
229 ca->tcp_cwnd = cwnd; /* syn with cubic */
231 if (ca->last_max_cwnd <= cwnd) {
233 ca->bic_origin_point = cwnd;
235 /* Compute new K based on
236 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
238 ca->bic_K = cubic_root(cube_factor
239 * (ca->last_max_cwnd - cwnd));
240 ca->bic_origin_point = ca->last_max_cwnd;
244 /* cubic function - calc*/
245 /* calculate c * time^3 / rtt,
246 * while considering overflow in calculation of time^3
247 * (so time^3 is done by using 64 bit)
248 * and without the support of division of 64bit numbers
249 * (so all divisions are done by using 32 bit)
250 * also NOTE the unit of those veriables
251 * time = (t - K) / 2^bictcp_HZ
252 * c = bic_scale >> 10
253 * rtt = (srtt >> 3) / HZ
254 * !!! The following code does not have overflow problems,
255 * if the cwnd < 1 million packets !!!
258 t = (s32)(tcp_time_stamp - ca->epoch_start);
259 t += msecs_to_jiffies(ca->delay_min >> 3);
260 /* change the unit from HZ to bictcp_HZ */
264 if (t < ca->bic_K) /* t - K */
265 offs = ca->bic_K - t;
267 offs = t - ca->bic_K;
269 /* c/rtt * (t-K)^3 */
270 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
271 if (t < ca->bic_K) /* below origin*/
272 bic_target = ca->bic_origin_point - delta;
273 else /* above origin*/
274 bic_target = ca->bic_origin_point + delta;
276 /* cubic function - calc bictcp_cnt*/
277 if (bic_target > cwnd) {
278 ca->cnt = cwnd / (bic_target - cwnd);
280 ca->cnt = 100 * cwnd; /* very small increment*/
284 * The initial growth of cubic function may be too conservative
285 * when the available bandwidth is still unknown.
287 if (ca->last_max_cwnd == 0 && ca->cnt > 20)
288 ca->cnt = 20; /* increase cwnd 5% per RTT */
292 if (tcp_friendliness) {
293 u32 scale = beta_scale;
295 delta = (cwnd * scale) >> 3;
296 while (ca->ack_cnt > delta) { /* update tcp cwnd */
297 ca->ack_cnt -= delta;
301 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */
302 delta = ca->tcp_cwnd - cwnd;
303 max_cnt = cwnd / delta;
304 if (ca->cnt > max_cnt)
309 if (ca->cnt == 0) /* cannot be zero */
313 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
315 struct tcp_sock *tp = tcp_sk(sk);
316 struct bictcp *ca = inet_csk_ca(sk);
318 if (!tcp_is_cwnd_limited(sk))
321 if (tp->snd_cwnd <= tp->snd_ssthresh) {
322 if (hystart && after(ack, ca->end_seq))
323 bictcp_hystart_reset(sk);
324 acked = tcp_slow_start(tp, acked);
328 bictcp_update(ca, tp->snd_cwnd, acked);
329 tcp_cong_avoid_ai(tp, ca->cnt, acked);
332 static u32 bictcp_recalc_ssthresh(struct sock *sk)
334 const struct tcp_sock *tp = tcp_sk(sk);
335 struct bictcp *ca = inet_csk_ca(sk);
337 ca->epoch_start = 0; /* end of epoch */
339 /* Wmax and fast convergence */
340 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
341 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
342 / (2 * BICTCP_BETA_SCALE);
344 ca->last_max_cwnd = tp->snd_cwnd;
346 ca->loss_cwnd = tp->snd_cwnd;
348 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
351 static u32 bictcp_undo_cwnd(struct sock *sk)
353 struct bictcp *ca = inet_csk_ca(sk);
355 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
358 static void bictcp_state(struct sock *sk, u8 new_state)
360 if (new_state == TCP_CA_Loss) {
361 bictcp_reset(inet_csk_ca(sk));
362 bictcp_hystart_reset(sk);
366 static void hystart_update(struct sock *sk, u32 delay)
368 struct tcp_sock *tp = tcp_sk(sk);
369 struct bictcp *ca = inet_csk_ca(sk);
371 if (ca->found & hystart_detect)
374 if (hystart_detect & HYSTART_ACK_TRAIN) {
375 u32 now = bictcp_clock();
377 /* first detection parameter - ack-train detection */
378 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
380 if ((s32)(now - ca->round_start) > ca->delay_min >> 4) {
381 ca->found |= HYSTART_ACK_TRAIN;
382 NET_INC_STATS_BH(sock_net(sk),
383 LINUX_MIB_TCPHYSTARTTRAINDETECT);
384 NET_ADD_STATS_BH(sock_net(sk),
385 LINUX_MIB_TCPHYSTARTTRAINCWND,
387 tp->snd_ssthresh = tp->snd_cwnd;
392 if (hystart_detect & HYSTART_DELAY) {
393 /* obtain the minimum delay of more than sampling packets */
394 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
395 if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
396 ca->curr_rtt = delay;
400 if (ca->curr_rtt > ca->delay_min +
401 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
402 ca->found |= HYSTART_DELAY;
403 NET_INC_STATS_BH(sock_net(sk),
404 LINUX_MIB_TCPHYSTARTDELAYDETECT);
405 NET_ADD_STATS_BH(sock_net(sk),
406 LINUX_MIB_TCPHYSTARTDELAYCWND,
408 tp->snd_ssthresh = tp->snd_cwnd;
414 /* Track delayed acknowledgment ratio using sliding window
415 * ratio = (15*ratio + sample) / 16
417 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
419 const struct tcp_sock *tp = tcp_sk(sk);
420 struct bictcp *ca = inet_csk_ca(sk);
423 /* Some calls are for duplicates without timetamps */
427 /* Discard delay samples right after fast recovery */
428 if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
431 delay = (rtt_us << 3) / USEC_PER_MSEC;
435 /* first time call or link delay decreases */
436 if (ca->delay_min == 0 || ca->delay_min > delay)
437 ca->delay_min = delay;
439 /* hystart triggers when cwnd is larger than some threshold */
440 if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
441 tp->snd_cwnd >= hystart_low_window)
442 hystart_update(sk, delay);
445 static struct tcp_congestion_ops cubictcp __read_mostly = {
447 .ssthresh = bictcp_recalc_ssthresh,
448 .cong_avoid = bictcp_cong_avoid,
449 .set_state = bictcp_state,
450 .undo_cwnd = bictcp_undo_cwnd,
451 .pkts_acked = bictcp_acked,
452 .owner = THIS_MODULE,
456 static int __init cubictcp_register(void)
458 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
460 /* Precompute a bunch of the scaling factors that are used per-packet
461 * based on SRTT of 100ms
464 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
465 / (BICTCP_BETA_SCALE - beta);
467 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
469 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
470 * so K = cubic_root( (wmax-cwnd)*rtt/c )
471 * the unit of K is bictcp_HZ=2^10, not HZ
473 * c = bic_scale >> 10
476 * the following code has been designed and tested for
477 * cwnd < 1 million packets
479 * HZ < 1,000,00 (corresponding to 10 nano-second)
482 /* 1/c * 2^2*bictcp_HZ * srtt */
483 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
485 /* divide by bic_scale and by constant Srtt (100ms) */
486 do_div(cube_factor, bic_scale * 10);
488 return tcp_register_congestion_control(&cubictcp);
491 static void __exit cubictcp_unregister(void)
493 tcp_unregister_congestion_control(&cubictcp);
496 module_init(cubictcp_register);
497 module_exit(cubictcp_unregister);
499 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
500 MODULE_LICENSE("GPL");
501 MODULE_DESCRIPTION("CUBIC TCP");
502 MODULE_VERSION("2.3");