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
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
21 *
22 * Unless CUBIC is enabled and congestion window is large
23 * this behaves the same as the original Reno.
24 */
25
26 #include <linux/mm.h>
27 #include <linux/module.h>
28 #include <linux/math64.h>
29 #include <net/tcp.h>
30
31 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
32 * max_cwnd = snd_cwnd * beta
33 */
34 #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
35
36 /* Two methods of hybrid slow start */
37 //Both run independently at the same time and slow start exits when any of them detects an exit point.
38 //1. ACK train length
39 //2. Delay increase
40
41 #define HYSTART_ACK_TRAIN 0x1
42 #define HYSTART_DELAY 0x2
43 /* 注意：这里的delay_min没有放大8倍！
44 * 此宏用来计算Delay increase threshold
45 * delay_min <= 32ms，则threshold = 2ms
46 * 32ms < delay_min < 256ms，则threshold = delay_min / 16 ms
47 * delay_min >= 256ms，则threshold = 16ms
48 */
49 /* Number of delay samples for detecting the increase of delay */
50 #define HYSTART_MIN_SAMPLES 8
51 #define HYSTART_DELAY_MIN (2U<<3)
52 #define HYSTART_DELAY_MAX (16U<<3)
53 #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
54
55 static int fast_convergence __read_mostly = 1;
56 static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
57 //beta在BIC中为819，而CUBIC中为717，
58 //会导致在bictcp_recalc_ssthresh中，并且启用了fast convergence，
59 //cubic: last_max_cwnd = 0.85*snd_cwnd ，而慢启动阈值=0.7*snd_cwnd 。
60 //bic: last_max_cwnd = 0.95*snd_cwnd ，而慢启动阈值=0.8*snd_cwnd 。
61 //这样会导致更早的到达平衡值，对snd_cwnd有很大的影响。
62
63
64
65 static int initial_ssthresh __read_mostly;
66 static int bic_scale __read_mostly = 41;
67 static int tcp_friendliness __read_mostly = 1;
68
69
70
71 //hybrid slow start的开关
72 static int hystart __read_mostly = 1;
73 //HyStart状态描述
74 //1：packet-train 2: delay 3:both packet-train and delay
75 //默认2种方法都使用，故设为3
76 static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
77 //设置snd_ssthresh的最小拥塞窗口值,除非cwnd超过了这个值，才能使用HyStart
78 static int hystart_low_window __read_mostly = 16;
79
80 static u32 cube_rtt_scale __read_mostly;
81 static u32 beta_scale __read_mostly;
82 static u64 cube_factor __read_mostly;
83
84 /* Note parameters that are used for precomputing scale factors are read-only */
85 module_param(fast_convergence, int, 0644);
86 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
87 module_param(beta, int, 0644);
88 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
89 module_param(initial_ssthresh, int, 0644);
90 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
91 module_param(bic_scale, int, 0444);
92 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
93 module_param(tcp_friendliness, int, 0644);
94 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
95 module_param(hystart, int, 0644);
96 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
97 module_param(hystart_detect, int, 0644);
98 MODULE_PARM_DESC(hystart_detect, "hyrbrid slow start detection mechanisms"
99 " 1: packet-train 2: delay 3: both packet-train and delay");
100 module_param(hystart_low_window, int, 0644);
101 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
102
103 /* BIC TCP Parameters */
104 struct bictcp {
105 u32 cnt; /*用来控制snd_cwnd的增长 increase cwnd by 1 after ACKs */
106 //两个重要的count值:
107 //第一个是tcp_sock->snd_cwnd_cnt，表示在当前的拥塞窗口中已经
108 //发送(经过对方ack包确认)的数据段的个数，
109 //而第二个是bictcp->cnt，它是cubic拥塞算法的核心，
110 //主要用来控制在拥塞避免状态的时候，什么时候才能增大拥塞窗口，
111 //具体实现是通过比较cnt和snd_cwnd_cnt，来决定是否增大拥塞窗口，
112
113 u32 last_max_cwnd; /*上一次的最大拥塞窗口值 last maximum snd_cwnd */
114 u32 loss_cwnd; /* 拥塞状态切换时的拥塞窗口值congestion window at last loss */
115 u32 last_cwnd; /* 上一次的拥塞窗口值 the last snd_cwnd */
116 u32 last_time; /* time when updated last_cwnd */
117 u32 bic_origin_point;/*即新的Wmax饱和点，取Wlast_max_cwnd和snd_cwnd较大者 origin point of bic function */
118 u32 bic_K; /*即新Wmax所对应的时间点t，W(bic_K) = Wmax time to origin point from the beginning of the current epoch */
119 u32 delay_min; /*应该是最小RTT min delay */
120 u32 epoch_start; /*拥塞状态切换开始的时刻 beginning of an epoch */
121 u32 ack_cnt; /*在一个epoch中的ack包的数量 number of acks */
122 u32 tcp_cwnd; /*按照Reno算法计算得的cwnd estimated tcp cwnd */
123 #define ACK_RATIO_SHIFT 4
124 u16 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
125 u8 sample_cnt; /*第几个sample number of samples to decide curr_rtt */
126 u8 found; /* the exit point is found? */
127 u32 round_start; /*针对每个RTT beginning of each round */
128 u32 end_seq; /*用来标识每个RTT end_seq of the round */
129 u32 last_jiffies; /*超过2ms则不认为是连续的 last time when the ACK spacing is close */
130 u32 curr_rtt; /*由sampe中最小的决定 the minimum rtt of current round */
131 };
132
133 static inline void bictcp_reset(struct bictcp *ca)
134 {//论文说Time out时调用
135 ca->cnt = 0;
136 ca->last_max_cwnd = 0;
137 ca->loss_cwnd = 0;
138 ca->last_cwnd = 0;
139 ca->last_time = 0;
140 ca->bic_origin_point = 0;
141 ca->bic_K = 0;
142 ca->delay_min = 0;
143 ca->epoch_start = 0;
144 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
145 ca->ack_cnt = 0;
146 ca->tcp_cwnd = 0;
147 ca->found = 0;
148 }
149
150 static inline void bictcp_hystart_reset(struct sock *sk)
151 {
152 struct tcp_sock *tp = tcp_sk(sk);
153 struct bictcp *ca = inet_csk_ca(sk);
154
155 ca->round_start = ca->last_jiffies = jiffies;//记录时间戳
156 ca->end_seq = tp->snd_nxt;//记录待发送的下一个序列号
157 ca->curr_rtt = 0;
158 ca->sample_cnt = 0;
159
160 //bictcp_hystart_reset中并没有对ca->found置0。
161 //也就是说，只有在初始化时、LOSS状态时、开启hystart的慢启动时。
162 //HyStart才会派上用场，其它时间并不使用.
163 }
164
165 static void bictcp_init(struct sock *sk)
166 {
167 bictcp_reset(inet_csk_ca(sk));
168
169 if (hystart)//如果指定hystart
170 bictcp_hystart_reset(sk);
171
172 if (!hystart && initial_ssthresh)
173 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
174 }
175
176 /* calculate the cubic root of x using a table lookup followed by one
177 * Newton-Raphson iteration.
178 * Avg err ~= 0.195%
179 */
180 static u32 cubic_root(u64 a) //用来计算立方根
181 {
182 u32 x, b, shift;
183 /*
184 * cbrt(x) MSB values for x MSB values in [0..63].
185 * Precomputed then refined by hand - Willy Tarreau
186 *
187 * For x in [0..63],
188 * v = cbrt(x << 18) - 1
189 * cbrt(x) = (v[x] + 10) >> 6
190 */
191 static const u8 v[] = {
192 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
193 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
194 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
195 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
196 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
197 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
198 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
199 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
200 };
201
202 b = fls64(a);
203 if (b < 7) {
204 /* a in [0..63] */
205 return ((u32)v[(u32)a] + 35) >> 6;
206 }
207
208 b = ((b * 84) >> 8) - 1;
209 shift = (a >> (b * 3));
210
211 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
212
213 /*
214 * Newton-Raphson iteration
215 * 2
216 * x = ( 2 * x + a / x ) / 3
217 * k+1 k k
218 */
219 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
220 x = ((x * 341) >> 10);
221 return x;
222 }
223
224 /*
225 * Compute congestion window to use.
226 */ //从快速恢复退出并进入拥塞避免状态之后，更新cnt
227 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
228 {
229 u64 offs;//时间差|t - K|
230 //delta是cwnd差，bic_target是预测值，t为预测时间
231 u32 delta, t, bic_target, max_cnt;
232
233 ca->ack_cnt++; /*ack包计数器加1 count the number of ACKs */
234
235 if (ca->last_cwnd == cwnd && //当前窗口与历史窗口相同
236 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)//时间差小于1000/32ms
237 return; //直接结束
238
239 ca->last_cwnd = cwnd;//记录进入拥塞避免时的窗口值
240 ca->last_time = tcp_time_stamp;//记录进入拥塞避免时的时刻
241
242 if (ca->epoch_start == 0) {//丢包后，开启一个新的时段
243 ca->epoch_start = tcp_time_stamp; /*新时段的开始 record the beginning of an epoch */
244 ca->ack_cnt = 1; /*ack包计数器初始化 start counting */
245 ca->tcp_cwnd = cwnd; /*同步更新 syn with cubic */
246
247 //取max(last_max_cwnd , cwnd)作为当前Wmax饱和点
248 if (ca->last_max_cwnd <= cwnd) {
249 ca->bic_K = 0;
250 ca->bic_origin_point = cwnd;
251 } else {
252 /* Compute new K based on
253 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
254 */
255 ca->bic_K = cubic_root(cube_factor
256 * (ca->last_max_cwnd - cwnd));
257 ca->bic_origin_point = ca->last_max_cwnd;
258 }
259 }
260
261 /* cubic function - calc*/
262 /* calculate c * time^3 / rtt,
263 * while considering overflow in calculation of time^3
264 * (so time^3 is done by using 64 bit)
265 * and without the support of division of 64bit numbers
266 * (so all divisions are done by using 32 bit)
267 * also NOTE the unit of those veriables
268 * time = (t - K) / 2^bictcp_HZ
269 * c = bic_scale >> 10 == 0.04
270 * rtt = (srtt >> 3) / HZ
271 * !!! The following code does not have overflow problems,
272 * if the cwnd < 1 million packets !!!
273 */
274
275 /* change the unit from HZ to bictcp_HZ */
276 t = ((tcp_time_stamp + (ca->delay_min>>3) - ca->epoch_start)
277 << BICTCP_HZ) / HZ;
278
279 //求| t - bic_K |
280 if (t < ca->bic_K) // 还未达到Wmax
281 offs = ca->bic_K - t;
282 else
283 offs = t - ca->bic_K;//已经超过Wmax
284
285 /* c/rtt * (t-K)^3 */ //计算立方，delta =| W(t) - W(bic_K) |
286 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
287
288
289
290 //t为预测时间，bic_K为新Wmax所对应的时间，
291 //bic_target为cwnd预测值，bic_origin_point为当前Wmax饱和点
292 if (t < ca->bic_K) /* below origin*/
293 bic_target = ca->bic_origin_point - delta;
294 else /* above origin*/
295 bic_target = ca->bic_origin_point + delta;
296
297 /* cubic function - calc bictcp_cnt*/
298 if (bic_target > cwnd) {// 相差越多，增长越快，这就是函数形状由来
299 ca->cnt = cwnd / (bic_target - cwnd);//
300 } else {//目前cwnd已经超出预期了，应该降速
301 ca->cnt = 100 * cwnd; /* very small increment*/
302 }
303
304
305
306 /* TCP Friendly —如果bic比RENO慢，则提升cwnd增长速度，即减小cnt
307 * 以上次丢包以后的时间t算起，每次RTT增长 3B / ( 2 - B)，那么可以得到
308 * 采用RENO算法的cwnd。
309 * cwnd (RENO) = cwnd + 3B / (2 - B) * ack_cnt / cwnd
310 * B为乘性减少因子，在此算法中为0.3
311 */
312 if (tcp_friendliness) {
313 u32 scale = beta_scale;
314 delta = (cwnd * scale) >> 3; //delta代表多少ACK可使tcp_cwnd++
315 while (ca->ack_cnt > delta) { /* update tcp cwnd */
316 ca->ack_cnt -= delta;
317 ca->tcp_cwnd++;
318 }
319
320 if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
321 delta = ca->tcp_cwnd - cwnd;
322 max_cnt = cwnd / delta;
323 if (ca->cnt > max_cnt)
324 ca->cnt = max_cnt;
325 }
326 }
327
328 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
329 if (ca->cnt == 0) /* cannot be zero */
330 ca->cnt = 1; //此时代表cwnd远小于bic_target，增长速度最大
331 }
332
333 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
334 {
335 struct tcp_sock *tp = tcp_sk(sk);
336 struct bictcp *ca = inet_csk_ca(sk);
337
338 //判断发送拥塞窗口是否到达限制，如果到达限制则直接返回。
339 if (!tcp_is_cwnd_limited(sk, in_flight))
340 return;
341
342 if (tp->snd_cwnd <= tp->snd_ssthresh) {
343 //当snd_cwnd<=ssthresh的时候，进入慢启动状态
344 if (hystart && after(ack, ca->end_seq))//是否需要reset对应的bictcp的值
345 bictcp_hystart_reset(sk);
346 tcp_slow_start(tp);//进入slow start状态
347 } else {
348 //当snd_cwnd>ssthresh的时候，进入拥塞避免状态
349 bictcp_update(ca, tp->snd_cwnd);//首先会更新bictcp->cnt
350 tcp_cong_avoid_ai(tp, ca->cnt);//然后进入拥塞避免，更新tcp_sock->snd_cwnd_cnt
351 }
352
353 }
354
355
356 //每次发生拥塞状态切换时，就会重新计算慢启动阈值
357 //做了两件事：重赋值last_max_cwnd、返回新的慢启动阈值
358 static u32 bictcp_recalc_ssthresh(struct sock *sk)
359 {//论文说这个函数在Packet loss时调用
360 const struct tcp_sock *tp = tcp_sk(sk);
361 struct bictcp *ca = inet_csk_ca(sk);
362
363 ca->epoch_start = 0; /* 发生拥塞状态切换，标志一个epoch结束 end of epoch */
364
365 /* Wmax and fast convergence */
366 //当一个新的TCP流加入到网络，
367 //网络中已有TCP流需要放弃自己带宽，
368 //给新的TCP流提供一定的上升空间。
369 //为提高已有TCP流所释放的带宽而引入快速收敛机制。
370 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
371 //snd_cwnd<last_max_cwnd
372 //表示已有TCP流所经历的饱和点因为可用带宽改变而正在降低。
373 //然后，通过进一步降低Wmax让已有流释放更多带宽。
374 //这种行为有效地延长已有流增大其窗口的时间，
375 //因为降低后的Wmax强制已有流更早进入平稳状态。
376 //这允许新流有更多的时间来赶上其窗口尺寸。
377 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
378 / (2 * BICTCP_BETA_SCALE); //last_max_cwnd = 0.9 * snd_cwnd
379 else
380 ca->last_max_cwnd = tp->snd_cwnd;
381
382 ca->loss_cwnd = tp->snd_cwnd;
383
384 //修改snd_ssthresh，即max(0.7*snd_cwnd，2)
385 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
386
387 }
388
389 static u32 bictcp_undo_cwnd(struct sock *sk)
390 {
391 struct bictcp *ca = inet_csk_ca(sk);
392
393 return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
394 }
395
396 static void bictcp_state(struct sock *sk, u8 new_state)
397 {
398 if (new_state == TCP_CA_Loss) {//如果处于LOSS状态，丢包处理
399 bictcp_reset(inet_csk_ca(sk));
400 bictcp_hystart_reset(sk);
401 }
402 }
403
404 static void hystart_update(struct sock *sk, u32 delay)
405 {//会修改snd_ssthresh
406 struct tcp_sock *tp = tcp_sk(sk);
407 struct bictcp *ca = inet_csk_ca(sk);
408
409 if (!(ca->found & hystart_detect)) {
410 u32 curr_jiffies = jiffies;
411
412 /* first detection parameter - ack-train detection */
413 if (curr_jiffies - ca->last_jiffies <= msecs_to_jiffies(2)) {
414 ca->last_jiffies = curr_jiffies;
415 if (curr_jiffies - ca->round_start >= ca->delay_min>>4)
416 ca->found |= HYSTART_ACK_TRAIN;
417 }
418
419 /* obtain the minimum delay of more than sampling packets */
420 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
421 if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
422 ca->curr_rtt = delay;
423
424 ca->sample_cnt++;
425 } else {
426 if (ca->curr_rtt > ca->delay_min +
427 HYSTART_DELAY_THRESH(ca->delay_min>>4))
428 ca->found |= HYSTART_DELAY;
429 }
430 /*
431 * Either one of two conditions are met,
432 * we exit from slow start immediately.
433 */
434 if (ca->found & hystart_detect)//found是一个是否退出slow start的标记
435 tp->snd_ssthresh = tp->snd_cwnd;//修改snd_ssthresh
436 }
437 }
438
439 /* Track delayed acknowledgment ratio using sliding window
440 * ratio = (15*ratio + sample) / 16
441 */ //基本每次收到ack都会调用这个函数，更新snd_ssthresh和delayed_ack
442 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
443 {//论文说这个函数在On each ACK时调用
444 const struct inet_connection_sock *icsk = inet_csk(sk);
445 const struct tcp_sock *tp = tcp_sk(sk);
446 struct bictcp *ca = inet_csk_ca(sk);
447 u32 delay;
448
449 if (icsk->icsk_ca_state == TCP_CA_Open) {
450 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
451 ca->delayed_ack += cnt;
452 }
453
454 /* Some calls are for duplicates without timetamps */
455 if (rtt_us < 0)
456 return;
457
458 /* Discard delay samples right after fast recovery */
459 if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
460 return;
461
462 delay = usecs_to_jiffies(rtt_us) << 3;
463 if (delay == 0)
464 delay = 1;
465
466 /* first time call or link delay decreases */
467 if (ca->delay_min == 0 || ca->delay_min > delay)
468 ca->delay_min = delay;
469
470 /* hystart triggers when cwnd is larger than some threshold */
471 //tp->snd_ssthresh初始值是一个很大的值0x7fffffff
472
473 //当拥塞窗口增大到16的时候，
474 //调用hystart_update来修改更新snd_ssthresh
475 //hystart_update主要用于是否退出slow start
476 if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
477 tp->snd_cwnd >= hystart_low_window)
478 hystart_update(sk, delay);
479 }
480
481 static struct tcp_congestion_ops cubictcp = {
482
483 .init = bictcp_init,
484
485
486 //调用ssthresh函数的地方有：tcp_fastretrans_alert(), tcp_enter_cwr(),tcp_enter_frto(), tcp_enter_loss()
487 //看起来每次发生拥塞状态切换的时候，都会调整ssthresh。
488 //修改snd_ssthresh值的地方有bictcp_init,hystart_update以及上面列出的调用ssthresh函数处。
489 .ssthresh = bictcp_recalc_ssthresh,
490
491 //发送方发出一个data包之后，接收方回复一个ack包，发送方收到这个ack包之后，
492 //调用tcp_ack()->tcp_cong_avoid()->bictcp_cong_avoid()来更改拥塞窗口snd_cwnd大小.
493 .cong_avoid = bictcp_cong_avoid,
494
495 .set_state = bictcp_state,
496
497 //调用undo_cwnd函数的地方有:tcp_undo_cwr()用来撤销之前误判导致的"缩小拥塞窗口"
498 .undo_cwnd = bictcp_undo_cwnd,
499
500 //调用ptts_acked函数的路径为：tcp_ack() -->tcp_clean_rtx_queue()
501 .pkts_acked = bictcp_acked,
502
503 .owner = THIS_MODULE,
504 .name = "cubic",
505 };
506
507 static int __init cubictcp_register(void)
508 {
509 //bictcp参数的个数不能过多
510 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
511
512 /* Precompute a bunch of the scaling factors that are used per-packet
513 * based on SRTT of 100ms
514 */
515 //beta_scale == 8*(1024 + 717) / 3 / (1024 -717 )，大约为15
516 beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
517
518 //cube_rtt_scale == 41*10 = 410
519 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
520
521 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
522 * so K = cubic_root( (wmax-cwnd)*rtt/c )
523 * the unit of K is bictcp_HZ=2^10, not HZ
524 *
525 * c = bic_scale >> 10
526 * rtt = 100ms
527 *
528 * the following code has been designed and tested for
529 * cwnd < 1 million packets
530 * RTT < 100 seconds
531 * HZ < 1,000,00 (corresponding to 10 nano-second)
532 */
533
534 /* 1/c * 2^2*bictcp_HZ * srtt */
535 cube_factor = 1ull << (10+3*BICTCP_HZ); /* cube_factor == 2^40 */
536
537 /* divide by bic_scale and by constant Srtt (100ms) */
538 do_div(cube_factor, bic_scale * 10);//cube_factor == 2^40 / 410
539
540 return tcp_register_congestion_control(&cubictcp);
541 }
542
543 static void __exit cubictcp_unregister(void)
544 {
545 tcp_unregister_congestion_control(&cubictcp);
546 }
547
548 module_init(cubictcp_register);
549 module_exit(cubictcp_unregister);
550
551 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
552 MODULE_LICENSE("GPL");
553 MODULE_DESCRIPTION("CUBIC TCP");
554 MODULE_VERSION("2.3");
浙公网安备 33010602011771号