]>
Commit | Line | Data |
---|---|---|
df3271f3 | 1 | /* |
ae27e98a | 2 | * TCP CUBIC: Binary Increase Congestion control for TCP v2.3 |
6b3d6263 SH |
3 | * Home page: |
4 | * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC | |
df3271f3 | 5 | * This is from the implementation of CUBIC TCP in |
ae27e98a SH |
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. | |
df3271f3 | 9 | * Available from: |
ae27e98a SH |
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 | |
df3271f3 SH |
21 | * |
22 | * Unless CUBIC is enabled and congestion window is large | |
23 | * this behaves the same as the original Reno. | |
24 | */ | |
25 | ||
df3271f3 SH |
26 | #include <linux/mm.h> |
27 | #include <linux/module.h> | |
6f6d6a1a | 28 | #include <linux/math64.h> |
df3271f3 | 29 | #include <net/tcp.h> |
df3271f3 SH |
30 | |
31 | #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation | |
32 | * max_cwnd = snd_cwnd * beta | |
33 | */ | |
df3271f3 SH |
34 | #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ |
35 | ||
ae27e98a SH |
36 | /* 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 | |
2b4636a5 | 42 | #define HYSTART_DELAY_MIN (4U<<3) |
ae27e98a SH |
43 | #define HYSTART_DELAY_MAX (16U<<3) |
44 | #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) | |
45 | ||
59758f44 | 46 | static int fast_convergence __read_mostly = 1; |
6b3d6263 | 47 | static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ |
66e1e3b2 | 48 | static int initial_ssthresh __read_mostly; |
59758f44 SH |
49 | static int bic_scale __read_mostly = 41; |
50 | static int tcp_friendliness __read_mostly = 1; | |
df3271f3 | 51 | |
ae27e98a SH |
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; | |
aac46324 | 55 | static int hystart_ack_delta __read_mostly = 2; |
ae27e98a | 56 | |
59758f44 SH |
57 | static u32 cube_rtt_scale __read_mostly; |
58 | static u32 beta_scale __read_mostly; | |
59 | static u64 cube_factor __read_mostly; | |
89b3d9aa SH |
60 | |
61 | /* Note parameters that are used for precomputing scale factors are read-only */ | |
df3271f3 SH |
62 | module_param(fast_convergence, int, 0644); |
63 | MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence"); | |
6b3d6263 | 64 | module_param(beta, int, 0644); |
df3271f3 SH |
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"); | |
89b3d9aa | 68 | module_param(bic_scale, int, 0444); |
df3271f3 SH |
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"); | |
ae27e98a SH |
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); | |
d6ecf328 | 75 | MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms" |
ae27e98a SH |
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"); | |
aac46324 | 79 | module_param(hystart_ack_delta, int, 0644); |
80 | MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)"); | |
df3271f3 | 81 | |
df3271f3 SH |
82 | /* BIC TCP Parameters */ |
83 | struct bictcp { | |
84 | u32 cnt; /* increase cwnd by 1 after ACKs */ | |
688d1945 | 85 | u32 last_max_cwnd; /* last maximum snd_cwnd */ |
df3271f3 SH |
86 | u32 last_cwnd; /* the last snd_cwnd */ |
87 | u32 last_time; /* time when updated last_cwnd */ | |
88 | u32 bic_origin_point;/* origin point of bic function */ | |
688d1945 | 89 | u32 bic_K; /* time to origin point |
90 | from the beginning of the current epoch */ | |
17a6e9f1 | 91 | u32 delay_min; /* min delay (msec << 3) */ |
df3271f3 SH |
92 | u32 epoch_start; /* beginning of an epoch */ |
93 | u32 ack_cnt; /* number of acks */ | |
94 | u32 tcp_cwnd; /* estimated tcp cwnd */ | |
9cd981dc | 95 | u16 unused; |
ae27e98a SH |
96 | u8 sample_cnt; /* number of samples to decide curr_rtt */ |
97 | u8 found; /* the exit point is found? */ | |
98 | u32 round_start; /* beginning of each round */ | |
99 | u32 end_seq; /* end_seq of the round */ | |
17a6e9f1 | 100 | u32 last_ack; /* last time when the ACK spacing is close */ |
ae27e98a | 101 | u32 curr_rtt; /* the minimum rtt of current round */ |
df3271f3 SH |
102 | }; |
103 | ||
104 | static inline void bictcp_reset(struct bictcp *ca) | |
105 | { | |
106 | ca->cnt = 0; | |
107 | ca->last_max_cwnd = 0; | |
df3271f3 SH |
108 | ca->last_cwnd = 0; |
109 | ca->last_time = 0; | |
110 | ca->bic_origin_point = 0; | |
111 | ca->bic_K = 0; | |
112 | ca->delay_min = 0; | |
113 | ca->epoch_start = 0; | |
df3271f3 SH |
114 | ca->ack_cnt = 0; |
115 | ca->tcp_cwnd = 0; | |
ae27e98a SH |
116 | ca->found = 0; |
117 | } | |
118 | ||
17a6e9f1 | 119 | static inline u32 bictcp_clock(void) |
120 | { | |
121 | #if HZ < 1000 | |
122 | return ktime_to_ms(ktime_get_real()); | |
123 | #else | |
124 | return jiffies_to_msecs(jiffies); | |
125 | #endif | |
126 | } | |
127 | ||
ae27e98a SH |
128 | static inline void bictcp_hystart_reset(struct sock *sk) |
129 | { | |
130 | struct tcp_sock *tp = tcp_sk(sk); | |
131 | struct bictcp *ca = inet_csk_ca(sk); | |
132 | ||
17a6e9f1 | 133 | ca->round_start = ca->last_ack = bictcp_clock(); |
ae27e98a SH |
134 | ca->end_seq = tp->snd_nxt; |
135 | ca->curr_rtt = 0; | |
136 | ca->sample_cnt = 0; | |
df3271f3 SH |
137 | } |
138 | ||
139 | static void bictcp_init(struct sock *sk) | |
140 | { | |
5a45f008 NC |
141 | struct bictcp *ca = inet_csk_ca(sk); |
142 | ||
143 | bictcp_reset(ca); | |
ae27e98a SH |
144 | |
145 | if (hystart) | |
146 | bictcp_hystart_reset(sk); | |
147 | ||
148 | if (!hystart && initial_ssthresh) | |
df3271f3 SH |
149 | tcp_sk(sk)->snd_ssthresh = initial_ssthresh; |
150 | } | |
151 | ||
30927520 ED |
152 | static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event) |
153 | { | |
154 | if (event == CA_EVENT_TX_START) { | |
30927520 | 155 | struct bictcp *ca = inet_csk_ca(sk); |
d635fbe2 | 156 | u32 now = tcp_jiffies32; |
c2e7204d ED |
157 | s32 delta; |
158 | ||
159 | delta = now - tcp_sk(sk)->lsndtime; | |
30927520 ED |
160 | |
161 | /* We were application limited (idle) for a while. | |
162 | * Shift epoch_start to keep cwnd growth to cubic curve. | |
163 | */ | |
c2e7204d | 164 | if (ca->epoch_start && delta > 0) { |
30927520 | 165 | ca->epoch_start += delta; |
c2e7204d ED |
166 | if (after(ca->epoch_start, now)) |
167 | ca->epoch_start = now; | |
168 | } | |
30927520 ED |
169 | return; |
170 | } | |
171 | } | |
172 | ||
7e58886b SH |
173 | /* calculate the cubic root of x using a table lookup followed by one |
174 | * Newton-Raphson iteration. | |
175 | * Avg err ~= 0.195% | |
df3271f3 | 176 | */ |
9eb2d627 | 177 | static u32 cubic_root(u64 a) |
df3271f3 | 178 | { |
7e58886b SH |
179 | u32 x, b, shift; |
180 | /* | |
181 | * cbrt(x) MSB values for x MSB values in [0..63]. | |
182 | * Precomputed then refined by hand - Willy Tarreau | |
183 | * | |
184 | * For x in [0..63], | |
185 | * v = cbrt(x << 18) - 1 | |
186 | * cbrt(x) = (v[x] + 10) >> 6 | |
9eb2d627 | 187 | */ |
7e58886b SH |
188 | static const u8 v[] = { |
189 | /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, | |
190 | /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, | |
191 | /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, | |
192 | /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, | |
193 | /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, | |
194 | /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, | |
195 | /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, | |
196 | /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, | |
197 | }; | |
198 | ||
199 | b = fls64(a); | |
200 | if (b < 7) { | |
201 | /* a in [0..63] */ | |
202 | return ((u32)v[(u32)a] + 35) >> 6; | |
203 | } | |
204 | ||
205 | b = ((b * 84) >> 8) - 1; | |
206 | shift = (a >> (b * 3)); | |
207 | ||
208 | x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; | |
209 | ||
210 | /* | |
211 | * Newton-Raphson iteration | |
212 | * 2 | |
213 | * x = ( 2 * x + a / x ) / 3 | |
214 | * k+1 k k | |
215 | */ | |
6f6d6a1a | 216 | x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); |
7e58886b | 217 | x = ((x * 341) >> 10); |
9eb2d627 | 218 | return x; |
df3271f3 SH |
219 | } |
220 | ||
df3271f3 SH |
221 | /* |
222 | * Compute congestion window to use. | |
223 | */ | |
9cd981dc | 224 | static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked) |
df3271f3 | 225 | { |
2ed0edf9 ED |
226 | u32 delta, bic_target, max_cnt; |
227 | u64 offs, t; | |
df3271f3 | 228 | |
9cd981dc | 229 | ca->ack_cnt += acked; /* count the number of ACKed packets */ |
df3271f3 SH |
230 | |
231 | if (ca->last_cwnd == cwnd && | |
ac35f562 | 232 | (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) |
df3271f3 SH |
233 | return; |
234 | ||
d6b1a8a9 NC |
235 | /* The CUBIC function can update ca->cnt at most once per jiffy. |
236 | * On all cwnd reduction events, ca->epoch_start is set to 0, | |
237 | * which will force a recalculation of ca->cnt. | |
238 | */ | |
ac35f562 | 239 | if (ca->epoch_start && tcp_jiffies32 == ca->last_time) |
d6b1a8a9 NC |
240 | goto tcp_friendliness; |
241 | ||
df3271f3 | 242 | ca->last_cwnd = cwnd; |
ac35f562 | 243 | ca->last_time = tcp_jiffies32; |
df3271f3 | 244 | |
df3271f3 | 245 | if (ca->epoch_start == 0) { |
ac35f562 | 246 | ca->epoch_start = tcp_jiffies32; /* record beginning */ |
9cd981dc | 247 | ca->ack_cnt = acked; /* start counting */ |
df3271f3 SH |
248 | ca->tcp_cwnd = cwnd; /* syn with cubic */ |
249 | ||
250 | if (ca->last_max_cwnd <= cwnd) { | |
251 | ca->bic_K = 0; | |
252 | ca->bic_origin_point = cwnd; | |
253 | } else { | |
89b3d9aa SH |
254 | /* Compute new K based on |
255 | * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) | |
256 | */ | |
257 | ca->bic_K = cubic_root(cube_factor | |
258 | * (ca->last_max_cwnd - cwnd)); | |
df3271f3 SH |
259 | ca->bic_origin_point = ca->last_max_cwnd; |
260 | } | |
261 | } | |
262 | ||
e905a9ed YH |
263 | /* cubic function - calc*/ |
264 | /* calculate c * time^3 / rtt, | |
265 | * while considering overflow in calculation of time^3 | |
89b3d9aa | 266 | * (so time^3 is done by using 64 bit) |
df3271f3 | 267 | * and without the support of division of 64bit numbers |
89b3d9aa | 268 | * (so all divisions are done by using 32 bit) |
e905a9ed YH |
269 | * also NOTE the unit of those veriables |
270 | * time = (t - K) / 2^bictcp_HZ | |
271 | * c = bic_scale >> 10 | |
df3271f3 SH |
272 | * rtt = (srtt >> 3) / HZ |
273 | * !!! The following code does not have overflow problems, | |
274 | * if the cwnd < 1 million packets !!! | |
e905a9ed | 275 | */ |
df3271f3 | 276 | |
ac35f562 | 277 | t = (s32)(tcp_jiffies32 - ca->epoch_start); |
2ed0edf9 | 278 | t += msecs_to_jiffies(ca->delay_min >> 3); |
df3271f3 | 279 | /* change the unit from HZ to bictcp_HZ */ |
2ed0edf9 ED |
280 | t <<= BICTCP_HZ; |
281 | do_div(t, HZ); | |
df3271f3 | 282 | |
e905a9ed | 283 | if (t < ca->bic_K) /* t - K */ |
89b3d9aa | 284 | offs = ca->bic_K - t; |
e905a9ed YH |
285 | else |
286 | offs = t - ca->bic_K; | |
df3271f3 | 287 | |
89b3d9aa SH |
288 | /* c/rtt * (t-K)^3 */ |
289 | delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); | |
688d1945 | 290 | if (t < ca->bic_K) /* below origin*/ |
e905a9ed | 291 | bic_target = ca->bic_origin_point - delta; |
688d1945 | 292 | else /* above origin*/ |
e905a9ed | 293 | bic_target = ca->bic_origin_point + delta; |
df3271f3 | 294 | |
e905a9ed YH |
295 | /* cubic function - calc bictcp_cnt*/ |
296 | if (bic_target > cwnd) { | |
df3271f3 | 297 | ca->cnt = cwnd / (bic_target - cwnd); |
e905a9ed YH |
298 | } else { |
299 | ca->cnt = 100 * cwnd; /* very small increment*/ | |
300 | } | |
df3271f3 | 301 | |
b5ccd073 SH |
302 | /* |
303 | * The initial growth of cubic function may be too conservative | |
304 | * when the available bandwidth is still unknown. | |
305 | */ | |
5a45f008 | 306 | if (ca->last_max_cwnd == 0 && ca->cnt > 20) |
b5ccd073 SH |
307 | ca->cnt = 20; /* increase cwnd 5% per RTT */ |
308 | ||
d6b1a8a9 | 309 | tcp_friendliness: |
df3271f3 SH |
310 | /* TCP Friendly */ |
311 | if (tcp_friendliness) { | |
89b3d9aa | 312 | u32 scale = beta_scale; |
688d1945 | 313 | |
89b3d9aa | 314 | delta = (cwnd * scale) >> 3; |
e905a9ed YH |
315 | while (ca->ack_cnt > delta) { /* update tcp cwnd */ |
316 | ca->ack_cnt -= delta; | |
317 | ca->tcp_cwnd++; | |
df3271f3 SH |
318 | } |
319 | ||
688d1945 | 320 | if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ |
89b3d9aa SH |
321 | delta = ca->tcp_cwnd - cwnd; |
322 | max_cnt = cwnd / delta; | |
df3271f3 SH |
323 | if (ca->cnt > max_cnt) |
324 | ca->cnt = max_cnt; | |
325 | } | |
e905a9ed | 326 | } |
df3271f3 | 327 | |
d578e18c NC |
328 | /* The maximum rate of cwnd increase CUBIC allows is 1 packet per |
329 | * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. | |
330 | */ | |
331 | ca->cnt = max(ca->cnt, 2U); | |
df3271f3 SH |
332 | } |
333 | ||
24901551 | 334 | static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
df3271f3 SH |
335 | { |
336 | struct tcp_sock *tp = tcp_sk(sk); | |
337 | struct bictcp *ca = inet_csk_ca(sk); | |
338 | ||
24901551 | 339 | if (!tcp_is_cwnd_limited(sk)) |
df3271f3 SH |
340 | return; |
341 | ||
071d5080 | 342 | if (tcp_in_slow_start(tp)) { |
ae27e98a SH |
343 | if (hystart && after(ack, ca->end_seq)) |
344 | bictcp_hystart_reset(sk); | |
9cd981dc NC |
345 | acked = tcp_slow_start(tp, acked); |
346 | if (!acked) | |
347 | return; | |
df3271f3 | 348 | } |
9cd981dc NC |
349 | bictcp_update(ca, tp->snd_cwnd, acked); |
350 | tcp_cong_avoid_ai(tp, ca->cnt, acked); | |
df3271f3 SH |
351 | } |
352 | ||
353 | static u32 bictcp_recalc_ssthresh(struct sock *sk) | |
354 | { | |
355 | const struct tcp_sock *tp = tcp_sk(sk); | |
356 | struct bictcp *ca = inet_csk_ca(sk); | |
357 | ||
358 | ca->epoch_start = 0; /* end of epoch */ | |
359 | ||
360 | /* Wmax and fast convergence */ | |
361 | if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) | |
362 | ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) | |
363 | / (2 * BICTCP_BETA_SCALE); | |
364 | else | |
365 | ca->last_max_cwnd = tp->snd_cwnd; | |
366 | ||
df3271f3 SH |
367 | return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); |
368 | } | |
369 | ||
df3271f3 SH |
370 | static void bictcp_state(struct sock *sk, u8 new_state) |
371 | { | |
ae27e98a | 372 | if (new_state == TCP_CA_Loss) { |
df3271f3 | 373 | bictcp_reset(inet_csk_ca(sk)); |
ae27e98a SH |
374 | bictcp_hystart_reset(sk); |
375 | } | |
376 | } | |
377 | ||
378 | static void hystart_update(struct sock *sk, u32 delay) | |
379 | { | |
380 | struct tcp_sock *tp = tcp_sk(sk); | |
381 | struct bictcp *ca = inet_csk_ca(sk); | |
382 | ||
6e3a8a93 ED |
383 | if (ca->found & hystart_detect) |
384 | return; | |
385 | ||
386 | if (hystart_detect & HYSTART_ACK_TRAIN) { | |
17a6e9f1 | 387 | u32 now = bictcp_clock(); |
ae27e98a SH |
388 | |
389 | /* first detection parameter - ack-train detection */ | |
17a6e9f1 | 390 | if ((s32)(now - ca->last_ack) <= hystart_ack_delta) { |
391 | ca->last_ack = now; | |
6e3a8a93 | 392 | if ((s32)(now - ca->round_start) > ca->delay_min >> 4) { |
ae27e98a | 393 | ca->found |= HYSTART_ACK_TRAIN; |
c10d9310 ED |
394 | NET_INC_STATS(sock_net(sk), |
395 | LINUX_MIB_TCPHYSTARTTRAINDETECT); | |
396 | NET_ADD_STATS(sock_net(sk), | |
397 | LINUX_MIB_TCPHYSTARTTRAINCWND, | |
398 | tp->snd_cwnd); | |
6e3a8a93 ED |
399 | tp->snd_ssthresh = tp->snd_cwnd; |
400 | } | |
ae27e98a | 401 | } |
6e3a8a93 | 402 | } |
ae27e98a | 403 | |
6e3a8a93 | 404 | if (hystart_detect & HYSTART_DELAY) { |
ae27e98a SH |
405 | /* obtain the minimum delay of more than sampling packets */ |
406 | if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { | |
407 | if (ca->curr_rtt == 0 || ca->curr_rtt > delay) | |
408 | ca->curr_rtt = delay; | |
409 | ||
410 | ca->sample_cnt++; | |
411 | } else { | |
412 | if (ca->curr_rtt > ca->delay_min + | |
42eef7a0 | 413 | HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { |
ae27e98a | 414 | ca->found |= HYSTART_DELAY; |
c10d9310 ED |
415 | NET_INC_STATS(sock_net(sk), |
416 | LINUX_MIB_TCPHYSTARTDELAYDETECT); | |
417 | NET_ADD_STATS(sock_net(sk), | |
418 | LINUX_MIB_TCPHYSTARTDELAYCWND, | |
419 | tp->snd_cwnd); | |
6e3a8a93 ED |
420 | tp->snd_ssthresh = tp->snd_cwnd; |
421 | } | |
ae27e98a | 422 | } |
ae27e98a | 423 | } |
df3271f3 SH |
424 | } |
425 | ||
426 | /* Track delayed acknowledgment ratio using sliding window | |
427 | * ratio = (15*ratio + sample) / 16 | |
428 | */ | |
756ee172 | 429 | static void bictcp_acked(struct sock *sk, const struct ack_sample *sample) |
df3271f3 | 430 | { |
ae27e98a | 431 | const struct tcp_sock *tp = tcp_sk(sk); |
e7d0c885 SH |
432 | struct bictcp *ca = inet_csk_ca(sk); |
433 | u32 delay; | |
df3271f3 | 434 | |
e7d0c885 | 435 | /* Some calls are for duplicates without timetamps */ |
756ee172 | 436 | if (sample->rtt_us < 0) |
e7d0c885 SH |
437 | return; |
438 | ||
439 | /* Discard delay samples right after fast recovery */ | |
ac35f562 | 440 | if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ) |
e7d0c885 SH |
441 | return; |
442 | ||
756ee172 | 443 | delay = (sample->rtt_us << 3) / USEC_PER_MSEC; |
e7d0c885 SH |
444 | if (delay == 0) |
445 | delay = 1; | |
446 | ||
447 | /* first time call or link delay decreases */ | |
448 | if (ca->delay_min == 0 || ca->delay_min > delay) | |
449 | ca->delay_min = delay; | |
ae27e98a SH |
450 | |
451 | /* hystart triggers when cwnd is larger than some threshold */ | |
071d5080 | 452 | if (hystart && tcp_in_slow_start(tp) && |
ae27e98a SH |
453 | tp->snd_cwnd >= hystart_low_window) |
454 | hystart_update(sk, delay); | |
e7d0c885 | 455 | } |
df3271f3 | 456 | |
a252bebe | 457 | static struct tcp_congestion_ops cubictcp __read_mostly = { |
df3271f3 SH |
458 | .init = bictcp_init, |
459 | .ssthresh = bictcp_recalc_ssthresh, | |
460 | .cong_avoid = bictcp_cong_avoid, | |
461 | .set_state = bictcp_state, | |
f1722a1b | 462 | .undo_cwnd = tcp_reno_undo_cwnd, |
30927520 | 463 | .cwnd_event = bictcp_cwnd_event, |
df3271f3 SH |
464 | .pkts_acked = bictcp_acked, |
465 | .owner = THIS_MODULE, | |
466 | .name = "cubic", | |
467 | }; | |
468 | ||
469 | static int __init cubictcp_register(void) | |
470 | { | |
74975d40 | 471 | BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); |
89b3d9aa SH |
472 | |
473 | /* Precompute a bunch of the scaling factors that are used per-packet | |
474 | * based on SRTT of 100ms | |
475 | */ | |
476 | ||
688d1945 | 477 | beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 |
478 | / (BICTCP_BETA_SCALE - beta); | |
89b3d9aa | 479 | |
22119240 | 480 | cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ |
89b3d9aa SH |
481 | |
482 | /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 | |
483 | * so K = cubic_root( (wmax-cwnd)*rtt/c ) | |
484 | * the unit of K is bictcp_HZ=2^10, not HZ | |
485 | * | |
486 | * c = bic_scale >> 10 | |
487 | * rtt = 100ms | |
488 | * | |
489 | * the following code has been designed and tested for | |
490 | * cwnd < 1 million packets | |
491 | * RTT < 100 seconds | |
492 | * HZ < 1,000,00 (corresponding to 10 nano-second) | |
493 | */ | |
494 | ||
495 | /* 1/c * 2^2*bictcp_HZ * srtt */ | |
496 | cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ | |
497 | ||
498 | /* divide by bic_scale and by constant Srtt (100ms) */ | |
499 | do_div(cube_factor, bic_scale * 10); | |
500 | ||
df3271f3 SH |
501 | return tcp_register_congestion_control(&cubictcp); |
502 | } | |
503 | ||
504 | static void __exit cubictcp_unregister(void) | |
505 | { | |
506 | tcp_unregister_congestion_control(&cubictcp); | |
507 | } | |
508 | ||
509 | module_init(cubictcp_register); | |
510 | module_exit(cubictcp_unregister); | |
511 | ||
512 | MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); | |
513 | MODULE_LICENSE("GPL"); | |
514 | MODULE_DESCRIPTION("CUBIC TCP"); | |
ae27e98a | 515 | MODULE_VERSION("2.3"); |