blob: 448c2615fece946d7b5cc3136941efff92a2edd9 [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 */
stephen hemminger688d1942014-08-29 23:32:05 -070085 u32 last_max_cwnd; /* last maximum snd_cwnd */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080086 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 */
stephen hemminger688d1942014-08-29 23:32:05 -070090 u32 bic_K; /* time to origin point
91 from the beginning of the current epoch */
stephen hemminger17a6e9f2011-03-14 07:52:15 +000092 u32 delay_min; /* min delay (msec << 3) */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080093 u32 epoch_start; /* beginning of an epoch */
94 u32 ack_cnt; /* number of acks */
95 u32 tcp_cwnd; /* estimated tcp cwnd */
Neal Cardwell9cd981d2015-01-28 20:01:38 -050096 u16 unused;
Sangtae Haae27e982008-10-29 00:07:18 -040097 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 */
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000101 u32 last_ack; /* last time when the ACK spacing is close */
Sangtae Haae27e982008-10-29 00:07:18 -0400102 u32 curr_rtt; /* the minimum rtt of current round */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800103};
104
105static inline void bictcp_reset(struct bictcp *ca)
106{
107 ca->cnt = 0;
108 ca->last_max_cwnd = 0;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800109 ca->last_cwnd = 0;
110 ca->last_time = 0;
111 ca->bic_origin_point = 0;
112 ca->bic_K = 0;
113 ca->delay_min = 0;
114 ca->epoch_start = 0;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800115 ca->ack_cnt = 0;
116 ca->tcp_cwnd = 0;
Sangtae Haae27e982008-10-29 00:07:18 -0400117 ca->found = 0;
118}
119
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000120static inline u32 bictcp_clock(void)
121{
122#if HZ < 1000
123 return ktime_to_ms(ktime_get_real());
124#else
125 return jiffies_to_msecs(jiffies);
126#endif
127}
128
Sangtae Haae27e982008-10-29 00:07:18 -0400129static inline void bictcp_hystart_reset(struct sock *sk)
130{
131 struct tcp_sock *tp = tcp_sk(sk);
132 struct bictcp *ca = inet_csk_ca(sk);
133
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000134 ca->round_start = ca->last_ack = bictcp_clock();
Sangtae Haae27e982008-10-29 00:07:18 -0400135 ca->end_seq = tp->snd_nxt;
136 ca->curr_rtt = 0;
137 ca->sample_cnt = 0;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800138}
139
140static void bictcp_init(struct sock *sk)
141{
Neal Cardwell5a45f002012-01-18 17:47:59 +0000142 struct bictcp *ca = inet_csk_ca(sk);
143
144 bictcp_reset(ca);
145 ca->loss_cwnd = 0;
Sangtae Haae27e982008-10-29 00:07:18 -0400146
147 if (hystart)
148 bictcp_hystart_reset(sk);
149
150 if (!hystart && initial_ssthresh)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800151 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
152}
153
Eric Dumazet30927522015-09-09 21:55:07 -0700154static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
155{
156 if (event == CA_EVENT_TX_START) {
Eric Dumazet30927522015-09-09 21:55:07 -0700157 struct bictcp *ca = inet_csk_ca(sk);
Eric Dumazetc2e72042015-09-17 08:38:00 -0700158 u32 now = tcp_time_stamp;
159 s32 delta;
160
161 delta = now - tcp_sk(sk)->lsndtime;
Eric Dumazet30927522015-09-09 21:55:07 -0700162
163 /* We were application limited (idle) for a while.
164 * Shift epoch_start to keep cwnd growth to cubic curve.
165 */
Eric Dumazetc2e72042015-09-17 08:38:00 -0700166 if (ca->epoch_start && delta > 0) {
Eric Dumazet30927522015-09-09 21:55:07 -0700167 ca->epoch_start += delta;
Eric Dumazetc2e72042015-09-17 08:38:00 -0700168 if (after(ca->epoch_start, now))
169 ca->epoch_start = now;
170 }
Eric Dumazet30927522015-09-09 21:55:07 -0700171 return;
172 }
173}
174
Stephen Hemminger7e588862007-03-22 12:10:58 -0700175/* calculate the cubic root of x using a table lookup followed by one
176 * Newton-Raphson iteration.
177 * Avg err ~= 0.195%
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800178 */
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800179static u32 cubic_root(u64 a)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800180{
Stephen Hemminger7e588862007-03-22 12:10:58 -0700181 u32 x, b, shift;
182 /*
183 * cbrt(x) MSB values for x MSB values in [0..63].
184 * Precomputed then refined by hand - Willy Tarreau
185 *
186 * For x in [0..63],
187 * v = cbrt(x << 18) - 1
188 * cbrt(x) = (v[x] + 10) >> 6
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800189 */
Stephen Hemminger7e588862007-03-22 12:10:58 -0700190 static const u8 v[] = {
191 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
192 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
193 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
194 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
195 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
196 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
197 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
198 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
199 };
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800200
Stephen Hemminger7e588862007-03-22 12:10:58 -0700201 b = fls64(a);
202 if (b < 7) {
203 /* a in [0..63] */
204 return ((u32)v[(u32)a] + 35) >> 6;
205 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800206
Stephen Hemminger7e588862007-03-22 12:10:58 -0700207 b = ((b * 84) >> 8) - 1;
208 shift = (a >> (b * 3));
209
210 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
211
212 /*
213 * Newton-Raphson iteration
214 * 2
215 * x = ( 2 * x + a / x ) / 3
216 * k+1 k k
217 */
Roman Zippel6f6d6a12008-05-01 04:34:28 -0700218 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
Stephen Hemminger7e588862007-03-22 12:10:58 -0700219 x = ((x * 341) >> 10);
Stephen Hemminger9eb2d622005-12-21 19:32:36 -0800220 return x;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800221}
222
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800223/*
224 * Compute congestion window to use.
225 */
Neal Cardwell9cd981d2015-01-28 20:01:38 -0500226static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800227{
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700228 u32 delta, bic_target, max_cnt;
229 u64 offs, t;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800230
Neal Cardwell9cd981d2015-01-28 20:01:38 -0500231 ca->ack_cnt += acked; /* count the number of ACKed packets */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800232
233 if (ca->last_cwnd == cwnd &&
234 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
235 return;
236
Neal Cardwelld6b1a8a2015-01-28 20:01:39 -0500237 /* The CUBIC function can update ca->cnt at most once per jiffy.
238 * On all cwnd reduction events, ca->epoch_start is set to 0,
239 * which will force a recalculation of ca->cnt.
240 */
241 if (ca->epoch_start && tcp_time_stamp == ca->last_time)
242 goto tcp_friendliness;
243
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800244 ca->last_cwnd = cwnd;
245 ca->last_time = tcp_time_stamp;
246
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800247 if (ca->epoch_start == 0) {
stephen hemminger688d1942014-08-29 23:32:05 -0700248 ca->epoch_start = tcp_time_stamp; /* record beginning */
Neal Cardwell9cd981d2015-01-28 20:01:38 -0500249 ca->ack_cnt = acked; /* start counting */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800250 ca->tcp_cwnd = cwnd; /* syn with cubic */
251
252 if (ca->last_max_cwnd <= cwnd) {
253 ca->bic_K = 0;
254 ca->bic_origin_point = cwnd;
255 } else {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800256 /* Compute new K based on
257 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
258 */
259 ca->bic_K = cubic_root(cube_factor
260 * (ca->last_max_cwnd - cwnd));
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800261 ca->bic_origin_point = ca->last_max_cwnd;
262 }
263 }
264
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900265 /* cubic function - calc*/
266 /* calculate c * time^3 / rtt,
267 * while considering overflow in calculation of time^3
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800268 * (so time^3 is done by using 64 bit)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800269 * and without the support of division of 64bit numbers
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800270 * (so all divisions are done by using 32 bit)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900271 * also NOTE the unit of those veriables
272 * time = (t - K) / 2^bictcp_HZ
273 * c = bic_scale >> 10
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800274 * rtt = (srtt >> 3) / HZ
275 * !!! The following code does not have overflow problems,
276 * if the cwnd < 1 million packets !!!
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900277 */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800278
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700279 t = (s32)(tcp_time_stamp - ca->epoch_start);
280 t += msecs_to_jiffies(ca->delay_min >> 3);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800281 /* change the unit from HZ to bictcp_HZ */
Eric Dumazet2ed0edf2013-08-05 17:10:15 -0700282 t <<= BICTCP_HZ;
283 do_div(t, HZ);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800284
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900285 if (t < ca->bic_K) /* t - K */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800286 offs = ca->bic_K - t;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900287 else
288 offs = t - ca->bic_K;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800289
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800290 /* c/rtt * (t-K)^3 */
291 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
stephen hemminger688d1942014-08-29 23:32:05 -0700292 if (t < ca->bic_K) /* below origin*/
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900293 bic_target = ca->bic_origin_point - delta;
stephen hemminger688d1942014-08-29 23:32:05 -0700294 else /* above origin*/
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900295 bic_target = ca->bic_origin_point + delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800296
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900297 /* cubic function - calc bictcp_cnt*/
298 if (bic_target > cwnd) {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800299 ca->cnt = cwnd / (bic_target - cwnd);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900300 } else {
301 ca->cnt = 100 * cwnd; /* very small increment*/
302 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800303
Sangtae Hab5ccd072011-03-14 07:52:18 +0000304 /*
305 * The initial growth of cubic function may be too conservative
306 * when the available bandwidth is still unknown.
307 */
Neal Cardwell5a45f002012-01-18 17:47:59 +0000308 if (ca->last_max_cwnd == 0 && ca->cnt > 20)
Sangtae Hab5ccd072011-03-14 07:52:18 +0000309 ca->cnt = 20; /* increase cwnd 5% per RTT */
310
Neal Cardwelld6b1a8a2015-01-28 20:01:39 -0500311tcp_friendliness:
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800312 /* TCP Friendly */
313 if (tcp_friendliness) {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800314 u32 scale = beta_scale;
stephen hemminger688d1942014-08-29 23:32:05 -0700315
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800316 delta = (cwnd * scale) >> 3;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900317 while (ca->ack_cnt > delta) { /* update tcp cwnd */
318 ca->ack_cnt -= delta;
319 ca->tcp_cwnd++;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800320 }
321
stephen hemminger688d1942014-08-29 23:32:05 -0700322 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800323 delta = ca->tcp_cwnd - cwnd;
324 max_cnt = cwnd / delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800325 if (ca->cnt > max_cnt)
326 ca->cnt = max_cnt;
327 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900328 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800329
Neal Cardwelld578e182015-03-10 17:17:04 -0400330 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per
331 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
332 */
333 ca->cnt = max(ca->cnt, 2U);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800334}
335
Eric Dumazet24901552014-05-02 21:18:05 -0700336static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800337{
338 struct tcp_sock *tp = tcp_sk(sk);
339 struct bictcp *ca = inet_csk_ca(sk);
340
Eric Dumazet24901552014-05-02 21:18:05 -0700341 if (!tcp_is_cwnd_limited(sk))
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800342 return;
343
Yuchung Cheng071d5082015-07-09 13:16:29 -0700344 if (tcp_in_slow_start(tp)) {
Sangtae Haae27e982008-10-29 00:07:18 -0400345 if (hystart && after(ack, ca->end_seq))
346 bictcp_hystart_reset(sk);
Neal Cardwell9cd981d2015-01-28 20:01:38 -0500347 acked = tcp_slow_start(tp, acked);
348 if (!acked)
349 return;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800350 }
Neal Cardwell9cd981d2015-01-28 20:01:38 -0500351 bictcp_update(ca, tp->snd_cwnd, acked);
352 tcp_cong_avoid_ai(tp, ca->cnt, acked);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800353}
354
355static u32 bictcp_recalc_ssthresh(struct sock *sk)
356{
357 const struct tcp_sock *tp = tcp_sk(sk);
358 struct bictcp *ca = inet_csk_ca(sk);
359
360 ca->epoch_start = 0; /* end of epoch */
361
362 /* Wmax and fast convergence */
363 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
364 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
365 / (2 * BICTCP_BETA_SCALE);
366 else
367 ca->last_max_cwnd = tp->snd_cwnd;
368
369 ca->loss_cwnd = tp->snd_cwnd;
370
371 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
372}
373
374static u32 bictcp_undo_cwnd(struct sock *sk)
375{
376 struct bictcp *ca = inet_csk_ca(sk);
377
Neal Cardwell5a45f002012-01-18 17:47:59 +0000378 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800379}
380
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800381static void bictcp_state(struct sock *sk, u8 new_state)
382{
Sangtae Haae27e982008-10-29 00:07:18 -0400383 if (new_state == TCP_CA_Loss) {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800384 bictcp_reset(inet_csk_ca(sk));
Sangtae Haae27e982008-10-29 00:07:18 -0400385 bictcp_hystart_reset(sk);
386 }
387}
388
389static void hystart_update(struct sock *sk, u32 delay)
390{
391 struct tcp_sock *tp = tcp_sk(sk);
392 struct bictcp *ca = inet_csk_ca(sk);
393
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800394 if (ca->found & hystart_detect)
395 return;
396
397 if (hystart_detect & HYSTART_ACK_TRAIN) {
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000398 u32 now = bictcp_clock();
Sangtae Haae27e982008-10-29 00:07:18 -0400399
400 /* first detection parameter - ack-train detection */
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000401 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
402 ca->last_ack = now;
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800403 if ((s32)(now - ca->round_start) > ca->delay_min >> 4) {
Sangtae Haae27e982008-10-29 00:07:18 -0400404 ca->found |= HYSTART_ACK_TRAIN;
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800405 NET_INC_STATS_BH(sock_net(sk),
406 LINUX_MIB_TCPHYSTARTTRAINDETECT);
407 NET_ADD_STATS_BH(sock_net(sk),
408 LINUX_MIB_TCPHYSTARTTRAINCWND,
409 tp->snd_cwnd);
410 tp->snd_ssthresh = tp->snd_cwnd;
411 }
Sangtae Haae27e982008-10-29 00:07:18 -0400412 }
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800413 }
Sangtae Haae27e982008-10-29 00:07:18 -0400414
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800415 if (hystart_detect & HYSTART_DELAY) {
Sangtae Haae27e982008-10-29 00:07:18 -0400416 /* obtain the minimum delay of more than sampling packets */
417 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
418 if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
419 ca->curr_rtt = delay;
420
421 ca->sample_cnt++;
422 } else {
423 if (ca->curr_rtt > ca->delay_min +
Eric Dumazet42eef7a2014-12-04 16:13:49 -0800424 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
Sangtae Haae27e982008-10-29 00:07:18 -0400425 ca->found |= HYSTART_DELAY;
Eric Dumazet6e3a8a92014-12-04 16:13:23 -0800426 NET_INC_STATS_BH(sock_net(sk),
427 LINUX_MIB_TCPHYSTARTDELAYDETECT);
428 NET_ADD_STATS_BH(sock_net(sk),
429 LINUX_MIB_TCPHYSTARTDELAYCWND,
430 tp->snd_cwnd);
431 tp->snd_ssthresh = tp->snd_cwnd;
432 }
Sangtae Haae27e982008-10-29 00:07:18 -0400433 }
Sangtae Haae27e982008-10-29 00:07:18 -0400434 }
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800435}
436
437/* Track delayed acknowledgment ratio using sliding window
438 * ratio = (15*ratio + sample) / 16
439 */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700440static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800441{
Sangtae Haae27e982008-10-29 00:07:18 -0400442 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700443 struct bictcp *ca = inet_csk_ca(sk);
444 u32 delay;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800445
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700446 /* Some calls are for duplicates without timetamps */
447 if (rtt_us < 0)
448 return;
449
450 /* Discard delay samples right after fast recovery */
Eric Dumazetcd6b4232013-08-05 20:05:12 -0700451 if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700452 return;
453
stephen hemminger17a6e9f2011-03-14 07:52:15 +0000454 delay = (rtt_us << 3) / USEC_PER_MSEC;
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700455 if (delay == 0)
456 delay = 1;
457
458 /* first time call or link delay decreases */
459 if (ca->delay_min == 0 || ca->delay_min > delay)
460 ca->delay_min = delay;
Sangtae Haae27e982008-10-29 00:07:18 -0400461
462 /* hystart triggers when cwnd is larger than some threshold */
Yuchung Cheng071d5082015-07-09 13:16:29 -0700463 if (hystart && tcp_in_slow_start(tp) &&
Sangtae Haae27e982008-10-29 00:07:18 -0400464 tp->snd_cwnd >= hystart_low_window)
465 hystart_update(sk, delay);
Stephen Hemmingere7d0c8852007-07-25 23:50:06 -0700466}
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800467
Stephen Hemmingera252beb2011-03-10 00:40:17 -0800468static struct tcp_congestion_ops cubictcp __read_mostly = {
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800469 .init = bictcp_init,
470 .ssthresh = bictcp_recalc_ssthresh,
471 .cong_avoid = bictcp_cong_avoid,
472 .set_state = bictcp_state,
473 .undo_cwnd = bictcp_undo_cwnd,
Eric Dumazet30927522015-09-09 21:55:07 -0700474 .cwnd_event = bictcp_cwnd_event,
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800475 .pkts_acked = bictcp_acked,
476 .owner = THIS_MODULE,
477 .name = "cubic",
478};
479
480static int __init cubictcp_register(void)
481{
Alexey Dobriyan74975d42006-08-25 17:10:33 -0700482 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800483
484 /* Precompute a bunch of the scaling factors that are used per-packet
485 * based on SRTT of 100ms
486 */
487
stephen hemminger688d1942014-08-29 23:32:05 -0700488 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
489 / (BICTCP_BETA_SCALE - beta);
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800490
Stephen Hemminger22119242006-10-25 23:04:12 -0700491 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800492
493 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
494 * so K = cubic_root( (wmax-cwnd)*rtt/c )
495 * the unit of K is bictcp_HZ=2^10, not HZ
496 *
497 * c = bic_scale >> 10
498 * rtt = 100ms
499 *
500 * the following code has been designed and tested for
501 * cwnd < 1 million packets
502 * RTT < 100 seconds
503 * HZ < 1,000,00 (corresponding to 10 nano-second)
504 */
505
506 /* 1/c * 2^2*bictcp_HZ * srtt */
507 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
508
509 /* divide by bic_scale and by constant Srtt (100ms) */
510 do_div(cube_factor, bic_scale * 10);
511
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800512 return tcp_register_congestion_control(&cubictcp);
513}
514
515static void __exit cubictcp_unregister(void)
516{
517 tcp_unregister_congestion_control(&cubictcp);
518}
519
520module_init(cubictcp_register);
521module_exit(cubictcp_unregister);
522
523MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
524MODULE_LICENSE("GPL");
525MODULE_DESCRIPTION("CUBIC TCP");
Sangtae Haae27e982008-10-29 00:07:18 -0400526MODULE_VERSION("2.3");