]>
Commit | Line | Data |
---|---|---|
09c434b8 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
df3271f3 | 2 | /* |
ae27e98a | 3 | * TCP CUBIC: Binary Increase Congestion control for TCP v2.3 |
6b3d6263 SH |
4 | * Home page: |
5 | * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC | |
df3271f3 | 6 | * This is from the implementation of CUBIC TCP in |
ae27e98a SH |
7 | * Sangtae Ha, Injong Rhee and Lisong Xu, |
8 | * "CUBIC: A New TCP-Friendly High-Speed TCP Variant" | |
9 | * in ACM SIGOPS Operating System Review, July 2008. | |
df3271f3 | 10 | * Available from: |
ae27e98a SH |
11 | * http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf |
12 | * | |
13 | * CUBIC integrates a new slow start algorithm, called HyStart. | |
14 | * The details of HyStart are presented in | |
15 | * Sangtae Ha and Injong Rhee, | |
16 | * "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008. | |
17 | * Available from: | |
18 | * http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf | |
19 | * | |
20 | * All testing results are available from: | |
21 | * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing | |
df3271f3 SH |
22 | * |
23 | * Unless CUBIC is enabled and congestion window is large | |
24 | * this behaves the same as the original Reno. | |
25 | */ | |
26 | ||
df3271f3 SH |
27 | #include <linux/mm.h> |
28 | #include <linux/module.h> | |
6f6d6a1a | 29 | #include <linux/math64.h> |
df3271f3 | 30 | #include <net/tcp.h> |
df3271f3 SH |
31 | |
32 | #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation | |
33 | * max_cwnd = snd_cwnd * beta | |
34 | */ | |
df3271f3 SH |
35 | #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ |
36 | ||
ae27e98a SH |
37 | /* Two methods of hybrid slow start */ |
38 | #define HYSTART_ACK_TRAIN 0x1 | |
39 | #define HYSTART_DELAY 0x2 | |
40 | ||
41 | /* Number of delay samples for detecting the increase of delay */ | |
42 | #define HYSTART_MIN_SAMPLES 8 | |
cff04e2d ED |
43 | #define HYSTART_DELAY_MIN (4000U) /* 4 ms */ |
44 | #define HYSTART_DELAY_MAX (16000U) /* 16 ms */ | |
ae27e98a SH |
45 | #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) |
46 | ||
59758f44 | 47 | static int fast_convergence __read_mostly = 1; |
6b3d6263 | 48 | static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ |
66e1e3b2 | 49 | static int initial_ssthresh __read_mostly; |
59758f44 SH |
50 | static int bic_scale __read_mostly = 41; |
51 | static int tcp_friendliness __read_mostly = 1; | |
df3271f3 | 52 | |
ae27e98a SH |
53 | static int hystart __read_mostly = 1; |
54 | static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY; | |
55 | static int hystart_low_window __read_mostly = 16; | |
cff04e2d | 56 | static int hystart_ack_delta_us __read_mostly = 2000; |
ae27e98a | 57 | |
59758f44 SH |
58 | static u32 cube_rtt_scale __read_mostly; |
59 | static u32 beta_scale __read_mostly; | |
60 | static u64 cube_factor __read_mostly; | |
89b3d9aa SH |
61 | |
62 | /* Note parameters that are used for precomputing scale factors are read-only */ | |
df3271f3 SH |
63 | module_param(fast_convergence, int, 0644); |
64 | MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence"); | |
6b3d6263 | 65 | module_param(beta, int, 0644); |
df3271f3 SH |
66 | MODULE_PARM_DESC(beta, "beta for multiplicative increase"); |
67 | module_param(initial_ssthresh, int, 0644); | |
68 | MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold"); | |
89b3d9aa | 69 | module_param(bic_scale, int, 0444); |
df3271f3 SH |
70 | MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)"); |
71 | module_param(tcp_friendliness, int, 0644); | |
72 | MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness"); | |
ae27e98a SH |
73 | module_param(hystart, int, 0644); |
74 | MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm"); | |
75 | module_param(hystart_detect, int, 0644); | |
d6ecf328 | 76 | MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms" |
ae27e98a SH |
77 | " 1: packet-train 2: delay 3: both packet-train and delay"); |
78 | module_param(hystart_low_window, int, 0644); | |
79 | MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start"); | |
cff04e2d ED |
80 | module_param(hystart_ack_delta_us, int, 0644); |
81 | MODULE_PARM_DESC(hystart_ack_delta_us, "spacing between ack's indicating train (usecs)"); | |
df3271f3 | 82 | |
df3271f3 SH |
83 | /* BIC TCP Parameters */ |
84 | struct bictcp { | |
85 | u32 cnt; /* increase cwnd by 1 after ACKs */ | |
688d1945 | 86 | u32 last_max_cwnd; /* last maximum snd_cwnd */ |
df3271f3 SH |
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 */ | |
688d1945 | 90 | u32 bic_K; /* time to origin point |
91 | from the beginning of the current epoch */ | |
cff04e2d | 92 | u32 delay_min; /* min delay (usec) */ |
df3271f3 SH |
93 | u32 epoch_start; /* beginning of an epoch */ |
94 | u32 ack_cnt; /* number of acks */ | |
95 | u32 tcp_cwnd; /* estimated tcp cwnd */ | |
9cd981dc | 96 | u16 unused; |
ae27e98a SH |
97 | 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 */ | |
17a6e9f1 | 101 | u32 last_ack; /* last time when the ACK spacing is close */ |
ae27e98a | 102 | u32 curr_rtt; /* the minimum rtt of current round */ |
df3271f3 SH |
103 | }; |
104 | ||
105 | static inline void bictcp_reset(struct bictcp *ca) | |
106 | { | |
107 | ca->cnt = 0; | |
108 | ca->last_max_cwnd = 0; | |
df3271f3 SH |
109 | 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; | |
df3271f3 SH |
115 | ca->ack_cnt = 0; |
116 | ca->tcp_cwnd = 0; | |
ae27e98a SH |
117 | ca->found = 0; |
118 | } | |
119 | ||
cff04e2d | 120 | static inline u32 bictcp_clock_us(const struct sock *sk) |
17a6e9f1 | 121 | { |
cff04e2d | 122 | return tcp_sk(sk)->tcp_mstamp; |
17a6e9f1 | 123 | } |
124 | ||
ae27e98a SH |
125 | static inline void bictcp_hystart_reset(struct sock *sk) |
126 | { | |
127 | struct tcp_sock *tp = tcp_sk(sk); | |
128 | struct bictcp *ca = inet_csk_ca(sk); | |
129 | ||
cff04e2d | 130 | ca->round_start = ca->last_ack = bictcp_clock_us(sk); |
ae27e98a | 131 | ca->end_seq = tp->snd_nxt; |
35821fc2 | 132 | ca->curr_rtt = ~0U; |
ae27e98a | 133 | ca->sample_cnt = 0; |
df3271f3 SH |
134 | } |
135 | ||
136 | static void bictcp_init(struct sock *sk) | |
137 | { | |
5a45f008 NC |
138 | struct bictcp *ca = inet_csk_ca(sk); |
139 | ||
140 | bictcp_reset(ca); | |
ae27e98a SH |
141 | |
142 | if (hystart) | |
143 | bictcp_hystart_reset(sk); | |
144 | ||
145 | if (!hystart && initial_ssthresh) | |
df3271f3 SH |
146 | tcp_sk(sk)->snd_ssthresh = initial_ssthresh; |
147 | } | |
148 | ||
30927520 ED |
149 | static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event) |
150 | { | |
151 | if (event == CA_EVENT_TX_START) { | |
30927520 | 152 | struct bictcp *ca = inet_csk_ca(sk); |
d635fbe2 | 153 | u32 now = tcp_jiffies32; |
c2e7204d ED |
154 | s32 delta; |
155 | ||
156 | delta = now - tcp_sk(sk)->lsndtime; | |
30927520 ED |
157 | |
158 | /* We were application limited (idle) for a while. | |
159 | * Shift epoch_start to keep cwnd growth to cubic curve. | |
160 | */ | |
c2e7204d | 161 | if (ca->epoch_start && delta > 0) { |
30927520 | 162 | ca->epoch_start += delta; |
c2e7204d ED |
163 | if (after(ca->epoch_start, now)) |
164 | ca->epoch_start = now; | |
165 | } | |
30927520 ED |
166 | return; |
167 | } | |
168 | } | |
169 | ||
7e58886b SH |
170 | /* calculate the cubic root of x using a table lookup followed by one |
171 | * Newton-Raphson iteration. | |
172 | * Avg err ~= 0.195% | |
df3271f3 | 173 | */ |
9eb2d627 | 174 | static u32 cubic_root(u64 a) |
df3271f3 | 175 | { |
7e58886b SH |
176 | u32 x, b, shift; |
177 | /* | |
178 | * cbrt(x) MSB values for x MSB values in [0..63]. | |
179 | * Precomputed then refined by hand - Willy Tarreau | |
180 | * | |
181 | * For x in [0..63], | |
182 | * v = cbrt(x << 18) - 1 | |
183 | * cbrt(x) = (v[x] + 10) >> 6 | |
9eb2d627 | 184 | */ |
7e58886b SH |
185 | static const u8 v[] = { |
186 | /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, | |
187 | /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, | |
188 | /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, | |
189 | /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, | |
190 | /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, | |
191 | /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, | |
192 | /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, | |
193 | /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, | |
194 | }; | |
195 | ||
196 | b = fls64(a); | |
197 | if (b < 7) { | |
198 | /* a in [0..63] */ | |
199 | return ((u32)v[(u32)a] + 35) >> 6; | |
200 | } | |
201 | ||
202 | b = ((b * 84) >> 8) - 1; | |
203 | shift = (a >> (b * 3)); | |
204 | ||
205 | x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; | |
206 | ||
207 | /* | |
208 | * Newton-Raphson iteration | |
209 | * 2 | |
210 | * x = ( 2 * x + a / x ) / 3 | |
211 | * k+1 k k | |
212 | */ | |
6f6d6a1a | 213 | x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); |
7e58886b | 214 | x = ((x * 341) >> 10); |
9eb2d627 | 215 | return x; |
df3271f3 SH |
216 | } |
217 | ||
df3271f3 SH |
218 | /* |
219 | * Compute congestion window to use. | |
220 | */ | |
9cd981dc | 221 | static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked) |
df3271f3 | 222 | { |
2ed0edf9 ED |
223 | u32 delta, bic_target, max_cnt; |
224 | u64 offs, t; | |
df3271f3 | 225 | |
9cd981dc | 226 | ca->ack_cnt += acked; /* count the number of ACKed packets */ |
df3271f3 SH |
227 | |
228 | if (ca->last_cwnd == cwnd && | |
ac35f562 | 229 | (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) |
df3271f3 SH |
230 | return; |
231 | ||
d6b1a8a9 NC |
232 | /* The CUBIC function can update ca->cnt at most once per jiffy. |
233 | * On all cwnd reduction events, ca->epoch_start is set to 0, | |
234 | * which will force a recalculation of ca->cnt. | |
235 | */ | |
ac35f562 | 236 | if (ca->epoch_start && tcp_jiffies32 == ca->last_time) |
d6b1a8a9 NC |
237 | goto tcp_friendliness; |
238 | ||
df3271f3 | 239 | ca->last_cwnd = cwnd; |
ac35f562 | 240 | ca->last_time = tcp_jiffies32; |
df3271f3 | 241 | |
df3271f3 | 242 | if (ca->epoch_start == 0) { |
ac35f562 | 243 | ca->epoch_start = tcp_jiffies32; /* record beginning */ |
9cd981dc | 244 | ca->ack_cnt = acked; /* start counting */ |
df3271f3 SH |
245 | ca->tcp_cwnd = cwnd; /* syn with cubic */ |
246 | ||
247 | if (ca->last_max_cwnd <= cwnd) { | |
248 | ca->bic_K = 0; | |
249 | ca->bic_origin_point = cwnd; | |
250 | } else { | |
89b3d9aa SH |
251 | /* Compute new K based on |
252 | * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) | |
253 | */ | |
254 | ca->bic_K = cubic_root(cube_factor | |
255 | * (ca->last_max_cwnd - cwnd)); | |
df3271f3 SH |
256 | ca->bic_origin_point = ca->last_max_cwnd; |
257 | } | |
258 | } | |
259 | ||
e905a9ed YH |
260 | /* cubic function - calc*/ |
261 | /* calculate c * time^3 / rtt, | |
262 | * while considering overflow in calculation of time^3 | |
89b3d9aa | 263 | * (so time^3 is done by using 64 bit) |
df3271f3 | 264 | * and without the support of division of 64bit numbers |
89b3d9aa | 265 | * (so all divisions are done by using 32 bit) |
e905a9ed YH |
266 | * also NOTE the unit of those veriables |
267 | * time = (t - K) / 2^bictcp_HZ | |
268 | * c = bic_scale >> 10 | |
df3271f3 SH |
269 | * rtt = (srtt >> 3) / HZ |
270 | * !!! The following code does not have overflow problems, | |
271 | * if the cwnd < 1 million packets !!! | |
e905a9ed | 272 | */ |
df3271f3 | 273 | |
ac35f562 | 274 | t = (s32)(tcp_jiffies32 - ca->epoch_start); |
cff04e2d | 275 | t += usecs_to_jiffies(ca->delay_min); |
df3271f3 | 276 | /* change the unit from HZ to bictcp_HZ */ |
2ed0edf9 ED |
277 | t <<= BICTCP_HZ; |
278 | do_div(t, HZ); | |
df3271f3 | 279 | |
e905a9ed | 280 | if (t < ca->bic_K) /* t - K */ |
89b3d9aa | 281 | offs = ca->bic_K - t; |
e905a9ed YH |
282 | else |
283 | offs = t - ca->bic_K; | |
df3271f3 | 284 | |
89b3d9aa SH |
285 | /* c/rtt * (t-K)^3 */ |
286 | delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); | |
688d1945 | 287 | if (t < ca->bic_K) /* below origin*/ |
e905a9ed | 288 | bic_target = ca->bic_origin_point - delta; |
688d1945 | 289 | else /* above origin*/ |
e905a9ed | 290 | bic_target = ca->bic_origin_point + delta; |
df3271f3 | 291 | |
e905a9ed YH |
292 | /* cubic function - calc bictcp_cnt*/ |
293 | if (bic_target > cwnd) { | |
df3271f3 | 294 | ca->cnt = cwnd / (bic_target - cwnd); |
e905a9ed YH |
295 | } else { |
296 | ca->cnt = 100 * cwnd; /* very small increment*/ | |
297 | } | |
df3271f3 | 298 | |
b5ccd073 SH |
299 | /* |
300 | * The initial growth of cubic function may be too conservative | |
301 | * when the available bandwidth is still unknown. | |
302 | */ | |
5a45f008 | 303 | if (ca->last_max_cwnd == 0 && ca->cnt > 20) |
b5ccd073 SH |
304 | ca->cnt = 20; /* increase cwnd 5% per RTT */ |
305 | ||
d6b1a8a9 | 306 | tcp_friendliness: |
df3271f3 SH |
307 | /* TCP Friendly */ |
308 | if (tcp_friendliness) { | |
89b3d9aa | 309 | u32 scale = beta_scale; |
688d1945 | 310 | |
89b3d9aa | 311 | delta = (cwnd * scale) >> 3; |
e905a9ed YH |
312 | while (ca->ack_cnt > delta) { /* update tcp cwnd */ |
313 | ca->ack_cnt -= delta; | |
314 | ca->tcp_cwnd++; | |
df3271f3 SH |
315 | } |
316 | ||
688d1945 | 317 | if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ |
89b3d9aa SH |
318 | delta = ca->tcp_cwnd - cwnd; |
319 | max_cnt = cwnd / delta; | |
df3271f3 SH |
320 | if (ca->cnt > max_cnt) |
321 | ca->cnt = max_cnt; | |
322 | } | |
e905a9ed | 323 | } |
df3271f3 | 324 | |
d578e18c NC |
325 | /* The maximum rate of cwnd increase CUBIC allows is 1 packet per |
326 | * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. | |
327 | */ | |
328 | ca->cnt = max(ca->cnt, 2U); | |
df3271f3 SH |
329 | } |
330 | ||
24901551 | 331 | static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
df3271f3 SH |
332 | { |
333 | struct tcp_sock *tp = tcp_sk(sk); | |
334 | struct bictcp *ca = inet_csk_ca(sk); | |
335 | ||
24901551 | 336 | if (!tcp_is_cwnd_limited(sk)) |
df3271f3 SH |
337 | return; |
338 | ||
071d5080 | 339 | if (tcp_in_slow_start(tp)) { |
ae27e98a SH |
340 | if (hystart && after(ack, ca->end_seq)) |
341 | bictcp_hystart_reset(sk); | |
9cd981dc NC |
342 | acked = tcp_slow_start(tp, acked); |
343 | if (!acked) | |
344 | return; | |
df3271f3 | 345 | } |
9cd981dc NC |
346 | bictcp_update(ca, tp->snd_cwnd, acked); |
347 | tcp_cong_avoid_ai(tp, ca->cnt, acked); | |
df3271f3 SH |
348 | } |
349 | ||
350 | static u32 bictcp_recalc_ssthresh(struct sock *sk) | |
351 | { | |
352 | const struct tcp_sock *tp = tcp_sk(sk); | |
353 | struct bictcp *ca = inet_csk_ca(sk); | |
354 | ||
355 | ca->epoch_start = 0; /* end of epoch */ | |
356 | ||
357 | /* Wmax and fast convergence */ | |
358 | if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) | |
359 | ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) | |
360 | / (2 * BICTCP_BETA_SCALE); | |
361 | else | |
362 | ca->last_max_cwnd = tp->snd_cwnd; | |
363 | ||
df3271f3 SH |
364 | return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); |
365 | } | |
366 | ||
df3271f3 SH |
367 | static void bictcp_state(struct sock *sk, u8 new_state) |
368 | { | |
ae27e98a | 369 | if (new_state == TCP_CA_Loss) { |
df3271f3 | 370 | bictcp_reset(inet_csk_ca(sk)); |
ae27e98a SH |
371 | bictcp_hystart_reset(sk); |
372 | } | |
373 | } | |
374 | ||
375 | static void hystart_update(struct sock *sk, u32 delay) | |
376 | { | |
377 | struct tcp_sock *tp = tcp_sk(sk); | |
378 | struct bictcp *ca = inet_csk_ca(sk); | |
379 | ||
6e3a8a93 | 380 | if (hystart_detect & HYSTART_ACK_TRAIN) { |
cff04e2d | 381 | u32 now = bictcp_clock_us(sk); |
ae27e98a SH |
382 | |
383 | /* first detection parameter - ack-train detection */ | |
cff04e2d | 384 | if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) { |
17a6e9f1 | 385 | ca->last_ack = now; |
cff04e2d | 386 | if ((s32)(now - ca->round_start) > ca->delay_min >> 1) { |
473900a5 | 387 | ca->found = 1; |
c10d9310 ED |
388 | NET_INC_STATS(sock_net(sk), |
389 | LINUX_MIB_TCPHYSTARTTRAINDETECT); | |
390 | NET_ADD_STATS(sock_net(sk), | |
391 | LINUX_MIB_TCPHYSTARTTRAINCWND, | |
392 | tp->snd_cwnd); | |
6e3a8a93 ED |
393 | tp->snd_ssthresh = tp->snd_cwnd; |
394 | } | |
ae27e98a | 395 | } |
6e3a8a93 | 396 | } |
ae27e98a | 397 | |
6e3a8a93 | 398 | if (hystart_detect & HYSTART_DELAY) { |
ae27e98a SH |
399 | /* obtain the minimum delay of more than sampling packets */ |
400 | if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { | |
35821fc2 | 401 | if (ca->curr_rtt > delay) |
ae27e98a SH |
402 | ca->curr_rtt = delay; |
403 | ||
404 | ca->sample_cnt++; | |
405 | } else { | |
406 | if (ca->curr_rtt > ca->delay_min + | |
42eef7a0 | 407 | HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { |
473900a5 | 408 | ca->found = 1; |
c10d9310 ED |
409 | NET_INC_STATS(sock_net(sk), |
410 | LINUX_MIB_TCPHYSTARTDELAYDETECT); | |
411 | NET_ADD_STATS(sock_net(sk), | |
412 | LINUX_MIB_TCPHYSTARTDELAYCWND, | |
413 | tp->snd_cwnd); | |
6e3a8a93 ED |
414 | tp->snd_ssthresh = tp->snd_cwnd; |
415 | } | |
ae27e98a | 416 | } |
ae27e98a | 417 | } |
df3271f3 SH |
418 | } |
419 | ||
756ee172 | 420 | static void bictcp_acked(struct sock *sk, const struct ack_sample *sample) |
df3271f3 | 421 | { |
ae27e98a | 422 | const struct tcp_sock *tp = tcp_sk(sk); |
e7d0c885 SH |
423 | struct bictcp *ca = inet_csk_ca(sk); |
424 | u32 delay; | |
df3271f3 | 425 | |
e7d0c885 | 426 | /* Some calls are for duplicates without timetamps */ |
756ee172 | 427 | if (sample->rtt_us < 0) |
e7d0c885 SH |
428 | return; |
429 | ||
430 | /* Discard delay samples right after fast recovery */ | |
ac35f562 | 431 | if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ) |
e7d0c885 SH |
432 | return; |
433 | ||
cff04e2d | 434 | delay = sample->rtt_us; |
e7d0c885 SH |
435 | if (delay == 0) |
436 | delay = 1; | |
437 | ||
438 | /* first time call or link delay decreases */ | |
42f3a8aa ED |
439 | if (ca->delay_min == 0 || ca->delay_min > delay) { |
440 | unsigned long rate = READ_ONCE(sk->sk_pacing_rate); | |
441 | ||
442 | /* Account for TSO/GRO delays. | |
443 | * Otherwise short RTT flows could get too small ssthresh, | |
444 | * since during slow start we begin with small TSO packets | |
445 | * and could lower ca->delay_min too much. | |
446 | * Ideally even with a very small RTT we would like to have | |
447 | * at least one TSO packet being sent and received by GRO, | |
448 | * and another one in qdisc layer. | |
449 | * We apply another 100% factor because @rate is doubled at | |
450 | * this point. | |
451 | * We cap the cushion to 1ms. | |
452 | */ | |
453 | if (rate) | |
454 | delay += min_t(u64, USEC_PER_MSEC, | |
455 | div64_ul((u64)GSO_MAX_SIZE * | |
456 | 4 * USEC_PER_SEC, rate)); | |
457 | if (ca->delay_min == 0 || ca->delay_min > delay) | |
458 | ca->delay_min = delay; | |
459 | } | |
ae27e98a SH |
460 | |
461 | /* hystart triggers when cwnd is larger than some threshold */ | |
473900a5 | 462 | if (!ca->found && hystart && tcp_in_slow_start(tp) && |
ae27e98a SH |
463 | tp->snd_cwnd >= hystart_low_window) |
464 | hystart_update(sk, delay); | |
e7d0c885 | 465 | } |
df3271f3 | 466 | |
a252bebe | 467 | static struct tcp_congestion_ops cubictcp __read_mostly = { |
df3271f3 SH |
468 | .init = bictcp_init, |
469 | .ssthresh = bictcp_recalc_ssthresh, | |
470 | .cong_avoid = bictcp_cong_avoid, | |
471 | .set_state = bictcp_state, | |
f1722a1b | 472 | .undo_cwnd = tcp_reno_undo_cwnd, |
30927520 | 473 | .cwnd_event = bictcp_cwnd_event, |
df3271f3 SH |
474 | .pkts_acked = bictcp_acked, |
475 | .owner = THIS_MODULE, | |
476 | .name = "cubic", | |
477 | }; | |
478 | ||
479 | static int __init cubictcp_register(void) | |
480 | { | |
74975d40 | 481 | BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); |
89b3d9aa SH |
482 | |
483 | /* Precompute a bunch of the scaling factors that are used per-packet | |
484 | * based on SRTT of 100ms | |
485 | */ | |
486 | ||
688d1945 | 487 | beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 |
488 | / (BICTCP_BETA_SCALE - beta); | |
89b3d9aa | 489 | |
22119240 | 490 | cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ |
89b3d9aa SH |
491 | |
492 | /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 | |
493 | * so K = cubic_root( (wmax-cwnd)*rtt/c ) | |
494 | * the unit of K is bictcp_HZ=2^10, not HZ | |
495 | * | |
496 | * c = bic_scale >> 10 | |
497 | * rtt = 100ms | |
498 | * | |
499 | * the following code has been designed and tested for | |
500 | * cwnd < 1 million packets | |
501 | * RTT < 100 seconds | |
502 | * HZ < 1,000,00 (corresponding to 10 nano-second) | |
503 | */ | |
504 | ||
505 | /* 1/c * 2^2*bictcp_HZ * srtt */ | |
506 | cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ | |
507 | ||
508 | /* divide by bic_scale and by constant Srtt (100ms) */ | |
509 | do_div(cube_factor, bic_scale * 10); | |
510 | ||
df3271f3 SH |
511 | return tcp_register_congestion_control(&cubictcp); |
512 | } | |
513 | ||
514 | static void __exit cubictcp_unregister(void) | |
515 | { | |
516 | tcp_unregister_congestion_control(&cubictcp); | |
517 | } | |
518 | ||
519 | module_init(cubictcp_register); | |
520 | module_exit(cubictcp_unregister); | |
521 | ||
522 | MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); | |
523 | MODULE_LICENSE("GPL"); | |
524 | MODULE_DESCRIPTION("CUBIC TCP"); | |
ae27e98a | 525 | MODULE_VERSION("2.3"); |