blob: 44fd40891acd1fc346c3b584d46e28a09a43d3da [file] [log] [blame]
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -08001/*
2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.0
3 *
4 * This is from the implementation of CUBIC TCP in
5 * Injong Rhee, Lisong Xu.
6 * "CUBIC: A New TCP-Friendly High-Speed TCP Variant
7 * in PFLDnet 2005
8 * Available from:
9 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
10 *
11 * Unless CUBIC is enabled and congestion window is large
12 * this behaves the same as the original Reno.
13 */
14
15#include <linux/config.h>
16#include <linux/mm.h>
17#include <linux/module.h>
18#include <net/tcp.h>
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080019#include <asm/div64.h>
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080020
21#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
22 * max_cwnd = snd_cwnd * beta
23 */
24#define BICTCP_B 4 /*
25 * In binary search,
26 * go to point (max+min)/N
27 */
28#define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
29
30static int fast_convergence = 1;
31static int max_increment = 16;
32static int beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
33static int initial_ssthresh = 100;
34static int bic_scale = 41;
35static int tcp_friendliness = 1;
36
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080037static u32 cube_rtt_scale;
38static u32 beta_scale;
39static u64 cube_factor;
40
41/* Note parameters that are used for precomputing scale factors are read-only */
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080042module_param(fast_convergence, int, 0644);
43MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
44module_param(max_increment, int, 0644);
45MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080046module_param(beta, int, 0444);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080047MODULE_PARM_DESC(beta, "beta for multiplicative increase");
48module_param(initial_ssthresh, int, 0644);
49MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -080050module_param(bic_scale, int, 0444);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -080051MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
52module_param(tcp_friendliness, int, 0644);
53MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
54
55
56/* BIC TCP Parameters */
57struct bictcp {
58 u32 cnt; /* increase cwnd by 1 after ACKs */
59 u32 last_max_cwnd; /* last maximum snd_cwnd */
60 u32 loss_cwnd; /* congestion window at last loss */
61 u32 last_cwnd; /* the last snd_cwnd */
62 u32 last_time; /* time when updated last_cwnd */
63 u32 bic_origin_point;/* origin point of bic function */
64 u32 bic_K; /* time to origin point from the beginning of the current epoch */
65 u32 delay_min; /* min delay */
66 u32 epoch_start; /* beginning of an epoch */
67 u32 ack_cnt; /* number of acks */
68 u32 tcp_cwnd; /* estimated tcp cwnd */
69#define ACK_RATIO_SHIFT 4
70 u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
71};
72
73static inline void bictcp_reset(struct bictcp *ca)
74{
75 ca->cnt = 0;
76 ca->last_max_cwnd = 0;
77 ca->loss_cwnd = 0;
78 ca->last_cwnd = 0;
79 ca->last_time = 0;
80 ca->bic_origin_point = 0;
81 ca->bic_K = 0;
82 ca->delay_min = 0;
83 ca->epoch_start = 0;
84 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
85 ca->ack_cnt = 0;
86 ca->tcp_cwnd = 0;
87}
88
89static void bictcp_init(struct sock *sk)
90{
91 bictcp_reset(inet_csk_ca(sk));
92 if (initial_ssthresh)
93 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
94}
95
96/* 65536 times the cubic root */
97static const u64 cubic_table[8]
98 = {0, 65536, 82570, 94519, 104030, 112063, 119087, 125367};
99
100/*
101 * calculate the cubic root of x
102 * the basic idea is that x can be expressed as i*8^j
103 * so cubic_root(x) = cubic_root(i)*2^j
104 * in the following code, x is i, and y is 2^j
105 * because of integer calculation, there are errors in calculation
106 * so finally use binary search to find out the exact solution
107 */
108static u32 cubic_root(u64 x)
109{
110 u64 y, app, target, start, end, mid, start_diff, end_diff;
111
112 if (x == 0)
113 return 0;
114
115 target = x;
116
117 /* first estimate lower and upper bound */
118 y = 1;
119 while (x >= 8){
120 x = (x >> 3);
121 y = (y << 1);
122 }
123 start = (y*cubic_table[x])>>16;
124 if (x==7)
125 end = (y<<1);
126 else
127 end = (y*cubic_table[x+1]+65535)>>16;
128
129 /* binary search for more accurate one */
130 while (start < end-1) {
131 mid = (start+end) >> 1;
132 app = mid*mid*mid;
133 if (app < target)
134 start = mid;
135 else if (app > target)
136 end = mid;
137 else
138 return mid;
139 }
140
141 /* find the most accurate one from start and end */
142 app = start*start*start;
143 if (app < target)
144 start_diff = target - app;
145 else
146 start_diff = app - target;
147 app = end*end*end;
148 if (app < target)
149 end_diff = target - app;
150 else
151 end_diff = app - target;
152
153 if (start_diff < end_diff)
154 return (u32)start;
155 else
156 return (u32)end;
157}
158
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800159/*
160 * Compute congestion window to use.
161 */
162static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
163{
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800164 u64 offs;
165 u32 delta, t, bic_target, min_cnt, max_cnt;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800166
167 ca->ack_cnt++; /* count the number of ACKs */
168
169 if (ca->last_cwnd == cwnd &&
170 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
171 return;
172
173 ca->last_cwnd = cwnd;
174 ca->last_time = tcp_time_stamp;
175
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800176 if (ca->epoch_start == 0) {
177 ca->epoch_start = tcp_time_stamp; /* record the beginning of an epoch */
178 ca->ack_cnt = 1; /* start counting */
179 ca->tcp_cwnd = cwnd; /* syn with cubic */
180
181 if (ca->last_max_cwnd <= cwnd) {
182 ca->bic_K = 0;
183 ca->bic_origin_point = cwnd;
184 } else {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800185 /* Compute new K based on
186 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
187 */
188 ca->bic_K = cubic_root(cube_factor
189 * (ca->last_max_cwnd - cwnd));
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800190 ca->bic_origin_point = ca->last_max_cwnd;
191 }
192 }
193
194 /* cubic function - calc*/
195 /* calculate c * time^3 / rtt,
196 * while considering overflow in calculation of time^3
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800197 * (so time^3 is done by using 64 bit)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800198 * and without the support of division of 64bit numbers
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800199 * (so all divisions are done by using 32 bit)
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800200 * also NOTE the unit of those veriables
201 * time = (t - K) / 2^bictcp_HZ
202 * c = bic_scale >> 10
203 * rtt = (srtt >> 3) / HZ
204 * !!! The following code does not have overflow problems,
205 * if the cwnd < 1 million packets !!!
206 */
207
208 /* change the unit from HZ to bictcp_HZ */
209 t = ((tcp_time_stamp + ca->delay_min - ca->epoch_start)
210 << BICTCP_HZ) / HZ;
211
212 if (t < ca->bic_K) /* t - K */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800213 offs = ca->bic_K - t;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800214 else
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800215 offs = t - ca->bic_K;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800216
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800217 /* c/rtt * (t-K)^3 */
218 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800219 if (t < ca->bic_K) /* below origin*/
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800220 bic_target = ca->bic_origin_point - delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800221 else /* above origin*/
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800222 bic_target = ca->bic_origin_point + delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800223
224 /* cubic function - calc bictcp_cnt*/
225 if (bic_target > cwnd) {
226 ca->cnt = cwnd / (bic_target - cwnd);
227 } else {
228 ca->cnt = 100 * cwnd; /* very small increment*/
229 }
230
231 if (ca->delay_min > 0) {
232 /* max increment = Smax * rtt / 0.1 */
233 min_cnt = (cwnd * HZ * 8)/(10 * max_increment * ca->delay_min);
234 if (ca->cnt < min_cnt)
235 ca->cnt = min_cnt;
236 }
237
238 /* slow start and low utilization */
239 if (ca->loss_cwnd == 0) /* could be aggressive in slow start */
240 ca->cnt = 50;
241
242 /* TCP Friendly */
243 if (tcp_friendliness) {
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800244 u32 scale = beta_scale;
245 delta = (cwnd * scale) >> 3;
246 while (ca->ack_cnt > delta) { /* update tcp cwnd */
247 ca->ack_cnt -= delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800248 ca->tcp_cwnd++;
249 }
250
251 if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800252 delta = ca->tcp_cwnd - cwnd;
253 max_cnt = cwnd / delta;
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800254 if (ca->cnt > max_cnt)
255 ca->cnt = max_cnt;
256 }
257 }
258
259 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
260 if (ca->cnt == 0) /* cannot be zero */
261 ca->cnt = 1;
262}
263
264
265/* Keep track of minimum rtt */
266static inline void measure_delay(struct sock *sk)
267{
268 const struct tcp_sock *tp = tcp_sk(sk);
269 struct bictcp *ca = inet_csk_ca(sk);
270 u32 delay;
271
272 /* No time stamp */
273 if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
274 /* Discard delay samples right after fast recovery */
275 (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
276 return;
277
278 delay = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
279 if (delay == 0)
280 delay = 1;
281
282 /* first time call or link delay decreases */
283 if (ca->delay_min == 0 || ca->delay_min > delay)
284 ca->delay_min = delay;
285}
286
287static void bictcp_cong_avoid(struct sock *sk, u32 ack,
288 u32 seq_rtt, u32 in_flight, int data_acked)
289{
290 struct tcp_sock *tp = tcp_sk(sk);
291 struct bictcp *ca = inet_csk_ca(sk);
292
293 if (data_acked)
294 measure_delay(sk);
295
296 if (!tcp_is_cwnd_limited(sk, in_flight))
297 return;
298
299 if (tp->snd_cwnd <= tp->snd_ssthresh)
300 tcp_slow_start(tp);
301 else {
302 bictcp_update(ca, tp->snd_cwnd);
303
304 /* In dangerous area, increase slowly.
305 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
306 */
307 if (tp->snd_cwnd_cnt >= ca->cnt) {
308 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
309 tp->snd_cwnd++;
310 tp->snd_cwnd_cnt = 0;
311 } else
312 tp->snd_cwnd_cnt++;
313 }
314
315}
316
317static u32 bictcp_recalc_ssthresh(struct sock *sk)
318{
319 const struct tcp_sock *tp = tcp_sk(sk);
320 struct bictcp *ca = inet_csk_ca(sk);
321
322 ca->epoch_start = 0; /* end of epoch */
323
324 /* Wmax and fast convergence */
325 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
326 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
327 / (2 * BICTCP_BETA_SCALE);
328 else
329 ca->last_max_cwnd = tp->snd_cwnd;
330
331 ca->loss_cwnd = tp->snd_cwnd;
332
333 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
334}
335
336static u32 bictcp_undo_cwnd(struct sock *sk)
337{
338 struct bictcp *ca = inet_csk_ca(sk);
339
340 return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
341}
342
343static u32 bictcp_min_cwnd(struct sock *sk)
344{
345 return tcp_sk(sk)->snd_ssthresh;
346}
347
348static void bictcp_state(struct sock *sk, u8 new_state)
349{
350 if (new_state == TCP_CA_Loss)
351 bictcp_reset(inet_csk_ca(sk));
352}
353
354/* Track delayed acknowledgment ratio using sliding window
355 * ratio = (15*ratio + sample) / 16
356 */
357static void bictcp_acked(struct sock *sk, u32 cnt)
358{
359 const struct inet_connection_sock *icsk = inet_csk(sk);
360
361 if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
362 struct bictcp *ca = inet_csk_ca(sk);
363 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
364 ca->delayed_ack += cnt;
365 }
366}
367
368
369static struct tcp_congestion_ops cubictcp = {
370 .init = bictcp_init,
371 .ssthresh = bictcp_recalc_ssthresh,
372 .cong_avoid = bictcp_cong_avoid,
373 .set_state = bictcp_state,
374 .undo_cwnd = bictcp_undo_cwnd,
375 .min_cwnd = bictcp_min_cwnd,
376 .pkts_acked = bictcp_acked,
377 .owner = THIS_MODULE,
378 .name = "cubic",
379};
380
381static int __init cubictcp_register(void)
382{
383 BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
Stephen Hemminger89b3d9a2005-12-21 19:32:08 -0800384
385 /* Precompute a bunch of the scaling factors that are used per-packet
386 * based on SRTT of 100ms
387 */
388
389 beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
390
391 cube_rtt_scale = (bic_scale << 3) / 10; /* 1024*c/rtt */
392
393 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
394 * so K = cubic_root( (wmax-cwnd)*rtt/c )
395 * the unit of K is bictcp_HZ=2^10, not HZ
396 *
397 * c = bic_scale >> 10
398 * rtt = 100ms
399 *
400 * the following code has been designed and tested for
401 * cwnd < 1 million packets
402 * RTT < 100 seconds
403 * HZ < 1,000,00 (corresponding to 10 nano-second)
404 */
405
406 /* 1/c * 2^2*bictcp_HZ * srtt */
407 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
408
409 /* divide by bic_scale and by constant Srtt (100ms) */
410 do_div(cube_factor, bic_scale * 10);
411
Stephen Hemmingerdf3271f2005-12-13 23:13:28 -0800412 return tcp_register_congestion_control(&cubictcp);
413}
414
415static void __exit cubictcp_unregister(void)
416{
417 tcp_unregister_congestion_control(&cubictcp);
418}
419
420module_init(cubictcp_register);
421module_exit(cubictcp_unregister);
422
423MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
424MODULE_LICENSE("GPL");
425MODULE_DESCRIPTION("CUBIC TCP");
426MODULE_VERSION("2.0");