Merge remote-tracking branch 'sysctl/master'
[linux-2.6/next.git] / net / ipv4 / tcp_cubic.c
blobf376b05cca818fd9496fe0cb8540a051e3e7e50a
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
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
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.
26 #include <linux/mm.h>
27 #include <linux/module.h>
28 #include <linux/math64.h>
29 #include <net/tcp.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 */
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 #define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT)
97 u16 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
98 u8 sample_cnt; /* number of samples to decide curr_rtt */
99 u8 found; /* the exit point is found? */
100 u32 round_start; /* beginning of each round */
101 u32 end_seq; /* end_seq of the round */
102 u32 last_ack; /* last time when the ACK spacing is close */
103 u32 curr_rtt; /* the minimum rtt of current round */
106 static inline void bictcp_reset(struct bictcp *ca)
108 ca->cnt = 0;
109 ca->last_max_cwnd = 0;
110 ca->loss_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;
123 static inline u32 bictcp_clock(void)
125 #if HZ < 1000
126 return ktime_to_ms(ktime_get_real());
127 #else
128 return jiffies_to_msecs(jiffies);
129 #endif
132 static inline void bictcp_hystart_reset(struct sock *sk)
134 struct tcp_sock *tp = tcp_sk(sk);
135 struct bictcp *ca = inet_csk_ca(sk);
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;
143 static void bictcp_init(struct sock *sk)
145 bictcp_reset(inet_csk_ca(sk));
147 if (hystart)
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.
156 * Avg err ~= 0.195%
158 static u32 cubic_root(u64 a)
160 u32 x, b, shift;
162 * cbrt(x) MSB values for x MSB values in [0..63].
163 * Precomputed then refined by hand - Willy Tarreau
165 * For x in [0..63],
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,
180 b = fls64(a);
181 if (b < 7) {
182 /* a in [0..63] */
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
195 * k+1 k k
197 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
198 x = ((x * 341) >> 10);
199 return x;
203 * Compute congestion window to use.
205 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
207 u64 offs;
208 u32 delta, t, bic_target, max_cnt;
210 ca->ack_cnt++; /* count the number of ACKs */
212 if (ca->last_cwnd == cwnd &&
213 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
214 return;
216 ca->last_cwnd = cwnd;
217 ca->last_time = tcp_time_stamp;
219 if (ca->epoch_start == 0) {
220 ca->epoch_start = tcp_time_stamp; /* record the beginning of an epoch */
221 ca->ack_cnt = 1; /* start counting */
222 ca->tcp_cwnd = cwnd; /* syn with cubic */
224 if (ca->last_max_cwnd <= cwnd) {
225 ca->bic_K = 0;
226 ca->bic_origin_point = cwnd;
227 } else {
228 /* Compute new K based on
229 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
231 ca->bic_K = cubic_root(cube_factor
232 * (ca->last_max_cwnd - cwnd));
233 ca->bic_origin_point = ca->last_max_cwnd;
237 /* cubic function - calc*/
238 /* calculate c * time^3 / rtt,
239 * while considering overflow in calculation of time^3
240 * (so time^3 is done by using 64 bit)
241 * and without the support of division of 64bit numbers
242 * (so all divisions are done by using 32 bit)
243 * also NOTE the unit of those veriables
244 * time = (t - K) / 2^bictcp_HZ
245 * c = bic_scale >> 10
246 * rtt = (srtt >> 3) / HZ
247 * !!! The following code does not have overflow problems,
248 * if the cwnd < 1 million packets !!!
251 /* change the unit from HZ to bictcp_HZ */
252 t = ((tcp_time_stamp + msecs_to_jiffies(ca->delay_min>>3)
253 - ca->epoch_start) << BICTCP_HZ) / HZ;
255 if (t < ca->bic_K) /* t - K */
256 offs = ca->bic_K - t;
257 else
258 offs = t - ca->bic_K;
260 /* c/rtt * (t-K)^3 */
261 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
262 if (t < ca->bic_K) /* below origin*/
263 bic_target = ca->bic_origin_point - delta;
264 else /* above origin*/
265 bic_target = ca->bic_origin_point + delta;
267 /* cubic function - calc bictcp_cnt*/
268 if (bic_target > cwnd) {
269 ca->cnt = cwnd / (bic_target - cwnd);
270 } else {
271 ca->cnt = 100 * cwnd; /* very small increment*/
275 * The initial growth of cubic function may be too conservative
276 * when the available bandwidth is still unknown.
278 if (ca->loss_cwnd == 0 && ca->cnt > 20)
279 ca->cnt = 20; /* increase cwnd 5% per RTT */
281 /* TCP Friendly */
282 if (tcp_friendliness) {
283 u32 scale = beta_scale;
284 delta = (cwnd * scale) >> 3;
285 while (ca->ack_cnt > delta) { /* update tcp cwnd */
286 ca->ack_cnt -= delta;
287 ca->tcp_cwnd++;
290 if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
291 delta = ca->tcp_cwnd - cwnd;
292 max_cnt = cwnd / delta;
293 if (ca->cnt > max_cnt)
294 ca->cnt = max_cnt;
298 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
299 if (ca->cnt == 0) /* cannot be zero */
300 ca->cnt = 1;
303 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
305 struct tcp_sock *tp = tcp_sk(sk);
306 struct bictcp *ca = inet_csk_ca(sk);
308 if (!tcp_is_cwnd_limited(sk, in_flight))
309 return;
311 if (tp->snd_cwnd <= tp->snd_ssthresh) {
312 if (hystart && after(ack, ca->end_seq))
313 bictcp_hystart_reset(sk);
314 tcp_slow_start(tp);
315 } else {
316 bictcp_update(ca, tp->snd_cwnd);
317 tcp_cong_avoid_ai(tp, ca->cnt);
322 static u32 bictcp_recalc_ssthresh(struct sock *sk)
324 const struct tcp_sock *tp = tcp_sk(sk);
325 struct bictcp *ca = inet_csk_ca(sk);
327 ca->epoch_start = 0; /* end of epoch */
329 /* Wmax and fast convergence */
330 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
331 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
332 / (2 * BICTCP_BETA_SCALE);
333 else
334 ca->last_max_cwnd = tp->snd_cwnd;
336 ca->loss_cwnd = tp->snd_cwnd;
338 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
341 static u32 bictcp_undo_cwnd(struct sock *sk)
343 struct bictcp *ca = inet_csk_ca(sk);
345 return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
348 static void bictcp_state(struct sock *sk, u8 new_state)
350 if (new_state == TCP_CA_Loss) {
351 bictcp_reset(inet_csk_ca(sk));
352 bictcp_hystart_reset(sk);
356 static void hystart_update(struct sock *sk, u32 delay)
358 struct tcp_sock *tp = tcp_sk(sk);
359 struct bictcp *ca = inet_csk_ca(sk);
361 if (!(ca->found & hystart_detect)) {
362 u32 now = bictcp_clock();
364 /* first detection parameter - ack-train detection */
365 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
366 ca->last_ack = now;
367 if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
368 ca->found |= HYSTART_ACK_TRAIN;
371 /* obtain the minimum delay of more than sampling packets */
372 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
373 if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
374 ca->curr_rtt = delay;
376 ca->sample_cnt++;
377 } else {
378 if (ca->curr_rtt > ca->delay_min +
379 HYSTART_DELAY_THRESH(ca->delay_min>>4))
380 ca->found |= HYSTART_DELAY;
383 * Either one of two conditions are met,
384 * we exit from slow start immediately.
386 if (ca->found & hystart_detect)
387 tp->snd_ssthresh = tp->snd_cwnd;
391 /* Track delayed acknowledgment ratio using sliding window
392 * ratio = (15*ratio + sample) / 16
394 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
396 const struct inet_connection_sock *icsk = inet_csk(sk);
397 const struct tcp_sock *tp = tcp_sk(sk);
398 struct bictcp *ca = inet_csk_ca(sk);
399 u32 delay;
401 if (icsk->icsk_ca_state == TCP_CA_Open) {
402 u32 ratio = ca->delayed_ack;
404 ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT;
405 ratio += cnt;
407 ca->delayed_ack = min(ratio, ACK_RATIO_LIMIT);
410 /* Some calls are for duplicates without timetamps */
411 if (rtt_us < 0)
412 return;
414 /* Discard delay samples right after fast recovery */
415 if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
416 return;
418 delay = (rtt_us << 3) / USEC_PER_MSEC;
419 if (delay == 0)
420 delay = 1;
422 /* first time call or link delay decreases */
423 if (ca->delay_min == 0 || ca->delay_min > delay)
424 ca->delay_min = delay;
426 /* hystart triggers when cwnd is larger than some threshold */
427 if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
428 tp->snd_cwnd >= hystart_low_window)
429 hystart_update(sk, delay);
432 static struct tcp_congestion_ops cubictcp __read_mostly = {
433 .init = bictcp_init,
434 .ssthresh = bictcp_recalc_ssthresh,
435 .cong_avoid = bictcp_cong_avoid,
436 .set_state = bictcp_state,
437 .undo_cwnd = bictcp_undo_cwnd,
438 .pkts_acked = bictcp_acked,
439 .owner = THIS_MODULE,
440 .name = "cubic",
443 static int __init cubictcp_register(void)
445 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
447 /* Precompute a bunch of the scaling factors that are used per-packet
448 * based on SRTT of 100ms
451 beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
453 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
455 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
456 * so K = cubic_root( (wmax-cwnd)*rtt/c )
457 * the unit of K is bictcp_HZ=2^10, not HZ
459 * c = bic_scale >> 10
460 * rtt = 100ms
462 * the following code has been designed and tested for
463 * cwnd < 1 million packets
464 * RTT < 100 seconds
465 * HZ < 1,000,00 (corresponding to 10 nano-second)
468 /* 1/c * 2^2*bictcp_HZ * srtt */
469 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
471 /* divide by bic_scale and by constant Srtt (100ms) */
472 do_div(cube_factor, bic_scale * 10);
474 /* hystart needs ms clock resolution */
475 if (hystart && HZ < 1000)
476 cubictcp.flags |= TCP_CONG_RTT_STAMP;
478 return tcp_register_congestion_control(&cubictcp);
481 static void __exit cubictcp_unregister(void)
483 tcp_unregister_congestion_control(&cubictcp);
486 module_init(cubictcp_register);
487 module_exit(cubictcp_unregister);
489 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
490 MODULE_LICENSE("GPL");
491 MODULE_DESCRIPTION("CUBIC TCP");
492 MODULE_VERSION("2.3");