blob: b6ae92a51f58607f7c32be2d0762a49380205012 [file] [log] [blame]
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -08001/*
Sangtae Haae27e982008-10-29 00:07:18 -04002 * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
Sangtae Ha6b3d6262008-03-04 14:17:41 -08003 * Home page:
4 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -08005 * This is from the implementation of CUBIC TCP in
Sangtae Haae27e982008-10-29 00:07:18 -04006 * 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.
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -08009 * Available from:
Sangtae Haae27e982008-10-29 00:07:18 -040010 * 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
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080021 *
22 * Unless CUBIC is enabled and congestion window is large
23 * this behaves the same as the original Reno.
24 */
25
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080026#include <linux/mm.h>
27#include <linux/module.h>
Roman Zippel6f6d6a12008-05-01 04:34:28 -070028#include <linux/math64.h>
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080029#include <net/tcp.h>
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080030
31#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
32 * max_cwnd = snd_cwnd * beta
33 */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080034#define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
35
Sangtae Haae27e982008-10-29 00:07:18 -040036/* 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
Sangtae Ha2b4636a2011-03-14 07:52:17 +000042#define HYSTART_DELAY_MIN (4U<<3)
Sangtae Haae27e982008-10-29 00:07:18 -040043#define HYSTART_DELAY_MAX (16U<<3)
44#define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
45
Stephen Hemminger59758f42007-02-12 13:15:20 -080046static int fast_convergence __read_mostly = 1;
Sangtae Ha6b3d6262008-03-04 14:17:41 -080047static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
David S. Miller66e1e3b2007-06-13 01:03:53 -070048static int initial_ssthresh __read_mostly;
Stephen Hemminger59758f42007-02-12 13:15:20 -080049static int bic_scale __read_mostly = 41;
50static int tcp_friendliness __read_mostly = 1;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080051
Sangtae Haae27e982008-10-29 00:07:18 -040052static int hystart __read_mostly = 1;
53static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
54static int hystart_low_window __read_mostly = 16;
stephen hemmingeraac46322011-03-14 07:52:14 +000055static int hystart_ack_delta __read_mostly = 2;
Sangtae Haae27e982008-10-29 00:07:18 -040056
Stephen Hemminger59758f42007-02-12 13:15:20 -080057static u32 cube_rtt_scale __read_mostly;
58static u32 beta_scale __read_mostly;
59static u64 cube_factor __read_mostly;
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080060
61/* Note parameters that are used for precomputing scale factors are read-only */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080062module_param(fast_convergence, int, 0644);
63MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
Sangtae Ha6b3d6262008-03-04 14:17:41 -080064module_param(beta, int, 0644);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080065MODULE_PARM_DESC(beta, "beta for multiplicative increase");
66module_param(initial_ssthresh, int, 0644);
67MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080068module_param(bic_scale, int, 0444);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080069MODULE_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");
Sangtae Haae27e982008-10-29 00:07:18 -040072module_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, "hyrbrid 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");
stephen hemmingeraac46322011-03-14 07:52:14 +000079module_param(hystart_ack_delta, int, 0644);
80MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080081
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080082/* 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 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 */
stephen hemminger17a6e9f2011-03-14 07:52:15 +000091 u32 delay_min; /* min delay (msec << 3) */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080092 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
stephen hemmingerb9f47a3a2011-05-04 10:04:56 +000096#define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT)
Sangtae Haae27e982008-10-29 00:07:18 -040097 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 */
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000102 u32 last_ack; /* last time when the ACK spacing is close */
Sangtae Haae27e982008-10-29 00:07:18 -0400103 u32 curr_rtt; /* the minimum rtt of current round */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800104};
105
106static inline void bictcp_reset(struct bictcp *ca)
107{
108 ca->cnt = 0;
109 ca->last_max_cwnd = 0;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800110 ca->last_cwnd = 0;
111 ca->last_time = 0;
112 ca->bic_origin_point = 0;
113 ca->bic_K = 0;
114 ca->delay_min = 0;
115 ca->epoch_start = 0;
116 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
117 ca->ack_cnt = 0;
118 ca->tcp_cwnd = 0;
Sangtae Haae27e982008-10-29 00:07:18 -0400119 ca->found = 0;
120}
121
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000122static inline u32 bictcp_clock(void)
123{
124#if HZ < 1000
125 return ktime_to_ms(ktime_get_real());
126#else
127 return jiffies_to_msecs(jiffies);
128#endif
129}
130
Sangtae Haae27e982008-10-29 00:07:18 -0400131static inline void bictcp_hystart_reset(struct sock *sk)
132{
133 struct tcp_sock *tp = tcp_sk(sk);
134 struct bictcp *ca = inet_csk_ca(sk);
135
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000136 ca->round_start = ca->last_ack = bictcp_clock();
Sangtae Haae27e982008-10-29 00:07:18 -0400137 ca->end_seq = tp->snd_nxt;
138 ca->curr_rtt = 0;
139 ca->sample_cnt = 0;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800140}
141
142static void bictcp_init(struct sock *sk)
143{
Neal Cardwell5a45f002012-01-18 17:47:59 +0000144 struct bictcp *ca = inet_csk_ca(sk);
145
146 bictcp_reset(ca);
147 ca->loss_cwnd = 0;
Sangtae Haae27e982008-10-29 00:07:18 -0400148
149 if (hystart)
150 bictcp_hystart_reset(sk);
151
152 if (!hystart && initial_ssthresh)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800153 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
154}
155
Stephen Hemminger7e588862007-03-22 12:10:58 -0700156/* calculate the cubic root of x using a table lookup followed by one
157 * Newton-Raphson iteration.
158 * Avg err ~= 0.195%
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800159 */
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800160static u32 cubic_root(u64 a)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800161{
Stephen Hemminger7e588862007-03-22 12:10:58 -0700162 u32 x, b, shift;
163 /*
164 * cbrt(x) MSB values for x MSB values in [0..63].
165 * Precomputed then refined by hand - Willy Tarreau
166 *
167 * For x in [0..63],
168 * v = cbrt(x << 18) - 1
169 * cbrt(x) = (v[x] + 10) >> 6
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800170 */
Stephen Hemminger7e588862007-03-22 12:10:58 -0700171 static const u8 v[] = {
172 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
173 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
174 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
175 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
176 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
177 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
178 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
179 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
180 };
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800181
Stephen Hemminger7e588862007-03-22 12:10:58 -0700182 b = fls64(a);
183 if (b < 7) {
184 /* a in [0..63] */
185 return ((u32)v[(u32)a] + 35) >> 6;
186 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800187
Stephen Hemminger7e588862007-03-22 12:10:58 -0700188 b = ((b * 84) >> 8) - 1;
189 shift = (a >> (b * 3));
190
191 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
192
193 /*
194 * Newton-Raphson iteration
195 * 2
196 * x = ( 2 * x + a / x ) / 3
197 * k+1 k k
198 */
Roman Zippel6f6d6a12008-05-01 04:34:28 -0700199 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
Stephen Hemminger7e588862007-03-22 12:10:58 -0700200 x = ((x * 341) >> 10);
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800201 return x;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800202}
203
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800204/*
205 * Compute congestion window to use.
206 */
207static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
208{
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700209 u32 delta, bic_target, max_cnt;
210 u64 offs, t;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800211
212 ca->ack_cnt++; /* count the number of ACKs */
213
214 if (ca->last_cwnd == cwnd &&
215 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
216 return;
217
218 ca->last_cwnd = cwnd;
219 ca->last_time = tcp_time_stamp;
220
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800221 if (ca->epoch_start == 0) {
222 ca->epoch_start = tcp_time_stamp; /* record the beginning of an epoch */
223 ca->ack_cnt = 1; /* start counting */
224 ca->tcp_cwnd = cwnd; /* syn with cubic */
225
226 if (ca->last_max_cwnd <= cwnd) {
227 ca->bic_K = 0;
228 ca->bic_origin_point = cwnd;
229 } else {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800230 /* Compute new K based on
231 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
232 */
233 ca->bic_K = cubic_root(cube_factor
234 * (ca->last_max_cwnd - cwnd));
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800235 ca->bic_origin_point = ca->last_max_cwnd;
236 }
237 }
238
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900239 /* cubic function - calc*/
240 /* calculate c * time^3 / rtt,
241 * while considering overflow in calculation of time^3
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800242 * (so time^3 is done by using 64 bit)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800243 * and without the support of division of 64bit numbers
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800244 * (so all divisions are done by using 32 bit)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900245 * also NOTE the unit of those veriables
246 * time = (t - K) / 2^bictcp_HZ
247 * c = bic_scale >> 10
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800248 * rtt = (srtt >> 3) / HZ
249 * !!! The following code does not have overflow problems,
250 * if the cwnd < 1 million packets !!!
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900251 */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800252
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700253 t = (s32)(tcp_time_stamp - ca->epoch_start);
254 t += msecs_to_jiffies(ca->delay_min >> 3);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800255 /* change the unit from HZ to bictcp_HZ */
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700256 t <<= BICTCP_HZ;
257 do_div(t, HZ);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800258
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900259 if (t < ca->bic_K) /* t - K */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800260 offs = ca->bic_K - t;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900261 else
262 offs = t - ca->bic_K;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800263
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800264 /* c/rtt * (t-K)^3 */
265 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900266 if (t < ca->bic_K) /* below origin*/
267 bic_target = ca->bic_origin_point - delta;
268 else /* above origin*/
269 bic_target = ca->bic_origin_point + delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800270
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900271 /* cubic function - calc bictcp_cnt*/
272 if (bic_target > cwnd) {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800273 ca->cnt = cwnd / (bic_target - cwnd);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900274 } else {
275 ca->cnt = 100 * cwnd; /* very small increment*/
276 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800277
Sangtae Hab5ccd072011-03-14 07:52:18 +0000278 /*
279 * The initial growth of cubic function may be too conservative
280 * when the available bandwidth is still unknown.
281 */
Neal Cardwell5a45f002012-01-18 17:47:59 +0000282 if (ca->last_max_cwnd == 0 && ca->cnt > 20)
Sangtae Hab5ccd072011-03-14 07:52:18 +0000283 ca->cnt = 20; /* increase cwnd 5% per RTT */
284
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800285 /* TCP Friendly */
286 if (tcp_friendliness) {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800287 u32 scale = beta_scale;
288 delta = (cwnd * scale) >> 3;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900289 while (ca->ack_cnt > delta) { /* update tcp cwnd */
290 ca->ack_cnt -= delta;
291 ca->tcp_cwnd++;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800292 }
293
294 if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800295 delta = ca->tcp_cwnd - cwnd;
296 max_cnt = cwnd / delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800297 if (ca->cnt > max_cnt)
298 ca->cnt = max_cnt;
299 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900300 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800301
302 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
303 if (ca->cnt == 0) /* cannot be zero */
304 ca->cnt = 1;
305}
306
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +0200307static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800308{
309 struct tcp_sock *tp = tcp_sk(sk);
310 struct bictcp *ca = inet_csk_ca(sk);
311
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800312 if (!tcp_is_cwnd_limited(sk, in_flight))
313 return;
314
Sangtae Haae27e982008-10-29 00:07:18 -0400315 if (tp->snd_cwnd <= tp->snd_ssthresh) {
316 if (hystart && after(ack, ca->end_seq))
317 bictcp_hystart_reset(sk);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800318 tcp_slow_start(tp);
Sangtae Haae27e982008-10-29 00:07:18 -0400319 } else {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800320 bictcp_update(ca, tp->snd_cwnd);
Ilpo Järvinen758ce5c2009-02-28 04:44:37 +0000321 tcp_cong_avoid_ai(tp, ca->cnt);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800322 }
323
324}
325
326static u32 bictcp_recalc_ssthresh(struct sock *sk)
327{
328 const struct tcp_sock *tp = tcp_sk(sk);
329 struct bictcp *ca = inet_csk_ca(sk);
330
331 ca->epoch_start = 0; /* end of epoch */
332
333 /* Wmax and fast convergence */
334 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
335 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
336 / (2 * BICTCP_BETA_SCALE);
337 else
338 ca->last_max_cwnd = tp->snd_cwnd;
339
340 ca->loss_cwnd = tp->snd_cwnd;
341
342 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
343}
344
345static u32 bictcp_undo_cwnd(struct sock *sk)
346{
347 struct bictcp *ca = inet_csk_ca(sk);
348
Neal Cardwell5a45f002012-01-18 17:47:59 +0000349 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800350}
351
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800352static void bictcp_state(struct sock *sk, u8 new_state)
353{
Sangtae Haae27e982008-10-29 00:07:18 -0400354 if (new_state == TCP_CA_Loss) {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800355 bictcp_reset(inet_csk_ca(sk));
Sangtae Haae27e982008-10-29 00:07:18 -0400356 bictcp_hystart_reset(sk);
357 }
358}
359
360static void hystart_update(struct sock *sk, u32 delay)
361{
362 struct tcp_sock *tp = tcp_sk(sk);
363 struct bictcp *ca = inet_csk_ca(sk);
364
365 if (!(ca->found & hystart_detect)) {
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000366 u32 now = bictcp_clock();
Sangtae Haae27e982008-10-29 00:07:18 -0400367
368 /* first detection parameter - ack-train detection */
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000369 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
370 ca->last_ack = now;
371 if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
Sangtae Haae27e982008-10-29 00:07:18 -0400372 ca->found |= HYSTART_ACK_TRAIN;
373 }
374
375 /* obtain the minimum delay of more than sampling packets */
376 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
377 if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
378 ca->curr_rtt = delay;
379
380 ca->sample_cnt++;
381 } else {
382 if (ca->curr_rtt > ca->delay_min +
383 HYSTART_DELAY_THRESH(ca->delay_min>>4))
384 ca->found |= HYSTART_DELAY;
385 }
386 /*
387 * Either one of two conditions are met,
388 * we exit from slow start immediately.
389 */
390 if (ca->found & hystart_detect)
391 tp->snd_ssthresh = tp->snd_cwnd;
392 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800393}
394
395/* Track delayed acknowledgment ratio using sliding window
396 * ratio = (15*ratio + sample) / 16
397 */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700398static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800399{
400 const struct inet_connection_sock *icsk = inet_csk(sk);
Sangtae Haae27e982008-10-29 00:07:18 -0400401 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700402 struct bictcp *ca = inet_csk_ca(sk);
403 u32 delay;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800404
Ilpo Järvinen35e86942007-05-31 10:16:47 +0300405 if (icsk->icsk_ca_state == TCP_CA_Open) {
stephen hemmingerb9f47a3a2011-05-04 10:04:56 +0000406 u32 ratio = ca->delayed_ack;
407
408 ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT;
409 ratio += cnt;
410
411 ca->delayed_ack = min(ratio, ACK_RATIO_LIMIT);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800412 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800413
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700414 /* Some calls are for duplicates without timetamps */
415 if (rtt_us < 0)
416 return;
417
418 /* Discard delay samples right after fast recovery */
Eric Dumazetcd6b4232013-08-05 20:05:12 -0700419 if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700420 return;
421
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000422 delay = (rtt_us << 3) / USEC_PER_MSEC;
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700423 if (delay == 0)
424 delay = 1;
425
426 /* first time call or link delay decreases */
427 if (ca->delay_min == 0 || ca->delay_min > delay)
428 ca->delay_min = delay;
Sangtae Haae27e982008-10-29 00:07:18 -0400429
430 /* hystart triggers when cwnd is larger than some threshold */
431 if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
432 tp->snd_cwnd >= hystart_low_window)
433 hystart_update(sk, delay);
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700434}
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800435
Stephen Hemmingera252beb2011-03-10 00:40:17 -0800436static struct tcp_congestion_ops cubictcp __read_mostly = {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800437 .init = bictcp_init,
438 .ssthresh = bictcp_recalc_ssthresh,
439 .cong_avoid = bictcp_cong_avoid,
440 .set_state = bictcp_state,
441 .undo_cwnd = bictcp_undo_cwnd,
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800442 .pkts_acked = bictcp_acked,
443 .owner = THIS_MODULE,
444 .name = "cubic",
445};
446
447static int __init cubictcp_register(void)
448{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700449 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800450
451 /* Precompute a bunch of the scaling factors that are used per-packet
452 * based on SRTT of 100ms
453 */
454
455 beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
456
Stephen Hemminger22119242006-10-25 23:04:12 -0700457 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800458
459 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
460 * so K = cubic_root( (wmax-cwnd)*rtt/c )
461 * the unit of K is bictcp_HZ=2^10, not HZ
462 *
463 * c = bic_scale >> 10
464 * rtt = 100ms
465 *
466 * the following code has been designed and tested for
467 * cwnd < 1 million packets
468 * RTT < 100 seconds
469 * HZ < 1,000,00 (corresponding to 10 nano-second)
470 */
471
472 /* 1/c * 2^2*bictcp_HZ * srtt */
473 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
474
475 /* divide by bic_scale and by constant Srtt (100ms) */
476 do_div(cube_factor, bic_scale * 10);
477
stephen hemminger3b585b32011-03-14 07:52:16 +0000478 /* hystart needs ms clock resolution */
479 if (hystart && HZ < 1000)
480 cubictcp.flags |= TCP_CONG_RTT_STAMP;
481
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800482 return tcp_register_congestion_control(&cubictcp);
483}
484
485static void __exit cubictcp_unregister(void)
486{
487 tcp_unregister_congestion_control(&cubictcp);
488}
489
490module_init(cubictcp_register);
491module_exit(cubictcp_unregister);
492
493MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
494MODULE_LICENSE("GPL");
495MODULE_DESCRIPTION("CUBIC TCP");
Sangtae Haae27e982008-10-29 00:07:18 -0400496MODULE_VERSION("2.3");