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 | 27 | #include <linux/mm.h> |
0e32dfc8 KKD |
28 | #include <linux/btf.h> |
29 | #include <linux/btf_ids.h> | |
df3271f3 | 30 | #include <linux/module.h> |
6f6d6a1a | 31 | #include <linux/math64.h> |
df3271f3 | 32 | #include <net/tcp.h> |
df3271f3 SH |
33 | |
34 | #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation | |
35 | * max_cwnd = snd_cwnd * beta | |
36 | */ | |
df3271f3 SH |
37 | #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ |
38 | ||
ae27e98a SH |
39 | /* Two methods of hybrid slow start */ |
40 | #define HYSTART_ACK_TRAIN 0x1 | |
41 | #define HYSTART_DELAY 0x2 | |
42 | ||
43 | /* Number of delay samples for detecting the increase of delay */ | |
44 | #define HYSTART_MIN_SAMPLES 8 | |
cff04e2d ED |
45 | #define HYSTART_DELAY_MIN (4000U) /* 4 ms */ |
46 | #define HYSTART_DELAY_MAX (16000U) /* 16 ms */ | |
ae27e98a SH |
47 | #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) |
48 | ||
59758f44 | 49 | static int fast_convergence __read_mostly = 1; |
6b3d6263 | 50 | static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ |
66e1e3b2 | 51 | static int initial_ssthresh __read_mostly; |
59758f44 SH |
52 | static int bic_scale __read_mostly = 41; |
53 | static int tcp_friendliness __read_mostly = 1; | |
df3271f3 | 54 | |
ae27e98a SH |
55 | static int hystart __read_mostly = 1; |
56 | static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY; | |
57 | static int hystart_low_window __read_mostly = 16; | |
cff04e2d | 58 | static int hystart_ack_delta_us __read_mostly = 2000; |
ae27e98a | 59 | |
59758f44 SH |
60 | static u32 cube_rtt_scale __read_mostly; |
61 | static u32 beta_scale __read_mostly; | |
62 | static u64 cube_factor __read_mostly; | |
89b3d9aa SH |
63 | |
64 | /* Note parameters that are used for precomputing scale factors are read-only */ | |
df3271f3 SH |
65 | module_param(fast_convergence, int, 0644); |
66 | MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence"); | |
6b3d6263 | 67 | module_param(beta, int, 0644); |
df3271f3 SH |
68 | MODULE_PARM_DESC(beta, "beta for multiplicative increase"); |
69 | module_param(initial_ssthresh, int, 0644); | |
70 | MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold"); | |
89b3d9aa | 71 | module_param(bic_scale, int, 0444); |
df3271f3 SH |
72 | MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)"); |
73 | module_param(tcp_friendliness, int, 0644); | |
74 | MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness"); | |
ae27e98a SH |
75 | module_param(hystart, int, 0644); |
76 | MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm"); | |
77 | module_param(hystart_detect, int, 0644); | |
d6ecf328 | 78 | MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms" |
ae27e98a SH |
79 | " 1: packet-train 2: delay 3: both packet-train and delay"); |
80 | module_param(hystart_low_window, int, 0644); | |
81 | MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start"); | |
cff04e2d ED |
82 | module_param(hystart_ack_delta_us, int, 0644); |
83 | MODULE_PARM_DESC(hystart_ack_delta_us, "spacing between ack's indicating train (usecs)"); | |
df3271f3 | 84 | |
df3271f3 SH |
85 | /* BIC TCP Parameters */ |
86 | struct bictcp { | |
87 | u32 cnt; /* increase cwnd by 1 after ACKs */ | |
688d1945 | 88 | u32 last_max_cwnd; /* last maximum snd_cwnd */ |
df3271f3 SH |
89 | u32 last_cwnd; /* the last snd_cwnd */ |
90 | u32 last_time; /* time when updated last_cwnd */ | |
91 | u32 bic_origin_point;/* origin point of bic function */ | |
688d1945 | 92 | u32 bic_K; /* time to origin point |
93 | from the beginning of the current epoch */ | |
cff04e2d | 94 | u32 delay_min; /* min delay (usec) */ |
df3271f3 SH |
95 | u32 epoch_start; /* beginning of an epoch */ |
96 | u32 ack_cnt; /* number of acks */ | |
97 | u32 tcp_cwnd; /* estimated tcp cwnd */ | |
9cd981dc | 98 | u16 unused; |
ae27e98a SH |
99 | u8 sample_cnt; /* number of samples to decide curr_rtt */ |
100 | u8 found; /* the exit point is found? */ | |
101 | u32 round_start; /* beginning of each round */ | |
102 | u32 end_seq; /* end_seq of the round */ | |
17a6e9f1 | 103 | u32 last_ack; /* last time when the ACK spacing is close */ |
ae27e98a | 104 | u32 curr_rtt; /* the minimum rtt of current round */ |
df3271f3 SH |
105 | }; |
106 | ||
107 | static inline void bictcp_reset(struct bictcp *ca) | |
108 | { | |
f4d133d8 | 109 | memset(ca, 0, offsetof(struct bictcp, unused)); |
ae27e98a SH |
110 | ca->found = 0; |
111 | } | |
112 | ||
cff04e2d | 113 | static inline u32 bictcp_clock_us(const struct sock *sk) |
17a6e9f1 | 114 | { |
cff04e2d | 115 | return tcp_sk(sk)->tcp_mstamp; |
17a6e9f1 | 116 | } |
117 | ||
ae27e98a SH |
118 | static inline void bictcp_hystart_reset(struct sock *sk) |
119 | { | |
120 | struct tcp_sock *tp = tcp_sk(sk); | |
121 | struct bictcp *ca = inet_csk_ca(sk); | |
122 | ||
cff04e2d | 123 | ca->round_start = ca->last_ack = bictcp_clock_us(sk); |
ae27e98a | 124 | ca->end_seq = tp->snd_nxt; |
35821fc2 | 125 | ca->curr_rtt = ~0U; |
ae27e98a | 126 | ca->sample_cnt = 0; |
df3271f3 SH |
127 | } |
128 | ||
d22f6ad1 | 129 | static void cubictcp_init(struct sock *sk) |
df3271f3 | 130 | { |
5a45f008 NC |
131 | struct bictcp *ca = inet_csk_ca(sk); |
132 | ||
133 | bictcp_reset(ca); | |
ae27e98a SH |
134 | |
135 | if (hystart) | |
136 | bictcp_hystart_reset(sk); | |
137 | ||
138 | if (!hystart && initial_ssthresh) | |
df3271f3 SH |
139 | tcp_sk(sk)->snd_ssthresh = initial_ssthresh; |
140 | } | |
141 | ||
d22f6ad1 | 142 | static void cubictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event) |
30927520 ED |
143 | { |
144 | if (event == CA_EVENT_TX_START) { | |
30927520 | 145 | struct bictcp *ca = inet_csk_ca(sk); |
d635fbe2 | 146 | u32 now = tcp_jiffies32; |
c2e7204d ED |
147 | s32 delta; |
148 | ||
149 | delta = now - tcp_sk(sk)->lsndtime; | |
30927520 ED |
150 | |
151 | /* We were application limited (idle) for a while. | |
152 | * Shift epoch_start to keep cwnd growth to cubic curve. | |
153 | */ | |
c2e7204d | 154 | if (ca->epoch_start && delta > 0) { |
30927520 | 155 | ca->epoch_start += delta; |
c2e7204d ED |
156 | if (after(ca->epoch_start, now)) |
157 | ca->epoch_start = now; | |
158 | } | |
30927520 ED |
159 | return; |
160 | } | |
161 | } | |
162 | ||
7e58886b SH |
163 | /* calculate the cubic root of x using a table lookup followed by one |
164 | * Newton-Raphson iteration. | |
165 | * Avg err ~= 0.195% | |
df3271f3 | 166 | */ |
9eb2d627 | 167 | static u32 cubic_root(u64 a) |
df3271f3 | 168 | { |
7e58886b SH |
169 | u32 x, b, shift; |
170 | /* | |
171 | * cbrt(x) MSB values for x MSB values in [0..63]. | |
172 | * Precomputed then refined by hand - Willy Tarreau | |
173 | * | |
174 | * For x in [0..63], | |
175 | * v = cbrt(x << 18) - 1 | |
176 | * cbrt(x) = (v[x] + 10) >> 6 | |
9eb2d627 | 177 | */ |
7e58886b SH |
178 | static const u8 v[] = { |
179 | /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, | |
180 | /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, | |
181 | /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, | |
182 | /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, | |
183 | /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, | |
184 | /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, | |
185 | /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, | |
186 | /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, | |
187 | }; | |
188 | ||
189 | b = fls64(a); | |
190 | if (b < 7) { | |
191 | /* a in [0..63] */ | |
192 | return ((u32)v[(u32)a] + 35) >> 6; | |
193 | } | |
194 | ||
195 | b = ((b * 84) >> 8) - 1; | |
196 | shift = (a >> (b * 3)); | |
197 | ||
198 | x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; | |
199 | ||
200 | /* | |
201 | * Newton-Raphson iteration | |
202 | * 2 | |
203 | * x = ( 2 * x + a / x ) / 3 | |
204 | * k+1 k k | |
205 | */ | |
6f6d6a1a | 206 | x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); |
7e58886b | 207 | x = ((x * 341) >> 10); |
9eb2d627 | 208 | return x; |
df3271f3 SH |
209 | } |
210 | ||
df3271f3 SH |
211 | /* |
212 | * Compute congestion window to use. | |
213 | */ | |
9cd981dc | 214 | static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked) |
df3271f3 | 215 | { |
2ed0edf9 ED |
216 | u32 delta, bic_target, max_cnt; |
217 | u64 offs, t; | |
df3271f3 | 218 | |
9cd981dc | 219 | ca->ack_cnt += acked; /* count the number of ACKed packets */ |
df3271f3 SH |
220 | |
221 | if (ca->last_cwnd == cwnd && | |
ac35f562 | 222 | (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) |
df3271f3 SH |
223 | return; |
224 | ||
d6b1a8a9 NC |
225 | /* The CUBIC function can update ca->cnt at most once per jiffy. |
226 | * On all cwnd reduction events, ca->epoch_start is set to 0, | |
227 | * which will force a recalculation of ca->cnt. | |
228 | */ | |
ac35f562 | 229 | if (ca->epoch_start && tcp_jiffies32 == ca->last_time) |
d6b1a8a9 NC |
230 | goto tcp_friendliness; |
231 | ||
df3271f3 | 232 | ca->last_cwnd = cwnd; |
ac35f562 | 233 | ca->last_time = tcp_jiffies32; |
df3271f3 | 234 | |
df3271f3 | 235 | if (ca->epoch_start == 0) { |
ac35f562 | 236 | ca->epoch_start = tcp_jiffies32; /* record beginning */ |
9cd981dc | 237 | ca->ack_cnt = acked; /* start counting */ |
df3271f3 SH |
238 | ca->tcp_cwnd = cwnd; /* syn with cubic */ |
239 | ||
240 | if (ca->last_max_cwnd <= cwnd) { | |
241 | ca->bic_K = 0; | |
242 | ca->bic_origin_point = cwnd; | |
243 | } else { | |
89b3d9aa SH |
244 | /* Compute new K based on |
245 | * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) | |
246 | */ | |
247 | ca->bic_K = cubic_root(cube_factor | |
248 | * (ca->last_max_cwnd - cwnd)); | |
df3271f3 SH |
249 | ca->bic_origin_point = ca->last_max_cwnd; |
250 | } | |
251 | } | |
252 | ||
e905a9ed YH |
253 | /* cubic function - calc*/ |
254 | /* calculate c * time^3 / rtt, | |
255 | * while considering overflow in calculation of time^3 | |
89b3d9aa | 256 | * (so time^3 is done by using 64 bit) |
df3271f3 | 257 | * and without the support of division of 64bit numbers |
89b3d9aa | 258 | * (so all divisions are done by using 32 bit) |
e905a9ed YH |
259 | * also NOTE the unit of those veriables |
260 | * time = (t - K) / 2^bictcp_HZ | |
261 | * c = bic_scale >> 10 | |
df3271f3 SH |
262 | * rtt = (srtt >> 3) / HZ |
263 | * !!! The following code does not have overflow problems, | |
264 | * if the cwnd < 1 million packets !!! | |
e905a9ed | 265 | */ |
df3271f3 | 266 | |
ac35f562 | 267 | t = (s32)(tcp_jiffies32 - ca->epoch_start); |
cff04e2d | 268 | t += usecs_to_jiffies(ca->delay_min); |
df3271f3 | 269 | /* change the unit from HZ to bictcp_HZ */ |
2ed0edf9 ED |
270 | t <<= BICTCP_HZ; |
271 | do_div(t, HZ); | |
df3271f3 | 272 | |
e905a9ed | 273 | if (t < ca->bic_K) /* t - K */ |
89b3d9aa | 274 | offs = ca->bic_K - t; |
e905a9ed YH |
275 | else |
276 | offs = t - ca->bic_K; | |
df3271f3 | 277 | |
89b3d9aa SH |
278 | /* c/rtt * (t-K)^3 */ |
279 | delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); | |
688d1945 | 280 | if (t < ca->bic_K) /* below origin*/ |
e905a9ed | 281 | bic_target = ca->bic_origin_point - delta; |
688d1945 | 282 | else /* above origin*/ |
e905a9ed | 283 | bic_target = ca->bic_origin_point + delta; |
df3271f3 | 284 | |
e905a9ed YH |
285 | /* cubic function - calc bictcp_cnt*/ |
286 | if (bic_target > cwnd) { | |
df3271f3 | 287 | ca->cnt = cwnd / (bic_target - cwnd); |
e905a9ed YH |
288 | } else { |
289 | ca->cnt = 100 * cwnd; /* very small increment*/ | |
290 | } | |
df3271f3 | 291 | |
b5ccd073 SH |
292 | /* |
293 | * The initial growth of cubic function may be too conservative | |
294 | * when the available bandwidth is still unknown. | |
295 | */ | |
5a45f008 | 296 | if (ca->last_max_cwnd == 0 && ca->cnt > 20) |
b5ccd073 SH |
297 | ca->cnt = 20; /* increase cwnd 5% per RTT */ |
298 | ||
d6b1a8a9 | 299 | tcp_friendliness: |
df3271f3 SH |
300 | /* TCP Friendly */ |
301 | if (tcp_friendliness) { | |
89b3d9aa | 302 | u32 scale = beta_scale; |
688d1945 | 303 | |
89b3d9aa | 304 | delta = (cwnd * scale) >> 3; |
e905a9ed YH |
305 | while (ca->ack_cnt > delta) { /* update tcp cwnd */ |
306 | ca->ack_cnt -= delta; | |
307 | ca->tcp_cwnd++; | |
df3271f3 SH |
308 | } |
309 | ||
688d1945 | 310 | if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ |
89b3d9aa SH |
311 | delta = ca->tcp_cwnd - cwnd; |
312 | max_cnt = cwnd / delta; | |
df3271f3 SH |
313 | if (ca->cnt > max_cnt) |
314 | ca->cnt = max_cnt; | |
315 | } | |
e905a9ed | 316 | } |
df3271f3 | 317 | |
d578e18c NC |
318 | /* The maximum rate of cwnd increase CUBIC allows is 1 packet per |
319 | * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. | |
320 | */ | |
321 | ca->cnt = max(ca->cnt, 2U); | |
df3271f3 SH |
322 | } |
323 | ||
d22f6ad1 | 324 | static void cubictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
df3271f3 SH |
325 | { |
326 | struct tcp_sock *tp = tcp_sk(sk); | |
327 | struct bictcp *ca = inet_csk_ca(sk); | |
328 | ||
24901551 | 329 | if (!tcp_is_cwnd_limited(sk)) |
df3271f3 SH |
330 | return; |
331 | ||
071d5080 | 332 | if (tcp_in_slow_start(tp)) { |
ae27e98a SH |
333 | if (hystart && after(ack, ca->end_seq)) |
334 | bictcp_hystart_reset(sk); | |
9cd981dc NC |
335 | acked = tcp_slow_start(tp, acked); |
336 | if (!acked) | |
337 | return; | |
df3271f3 | 338 | } |
9cd981dc NC |
339 | bictcp_update(ca, tp->snd_cwnd, acked); |
340 | tcp_cong_avoid_ai(tp, ca->cnt, acked); | |
df3271f3 SH |
341 | } |
342 | ||
d22f6ad1 | 343 | static u32 cubictcp_recalc_ssthresh(struct sock *sk) |
df3271f3 SH |
344 | { |
345 | const struct tcp_sock *tp = tcp_sk(sk); | |
346 | struct bictcp *ca = inet_csk_ca(sk); | |
347 | ||
348 | ca->epoch_start = 0; /* end of epoch */ | |
349 | ||
350 | /* Wmax and fast convergence */ | |
351 | if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) | |
352 | ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) | |
353 | / (2 * BICTCP_BETA_SCALE); | |
354 | else | |
355 | ca->last_max_cwnd = tp->snd_cwnd; | |
356 | ||
df3271f3 SH |
357 | return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); |
358 | } | |
359 | ||
d22f6ad1 | 360 | static void cubictcp_state(struct sock *sk, u8 new_state) |
df3271f3 | 361 | { |
ae27e98a | 362 | if (new_state == TCP_CA_Loss) { |
df3271f3 | 363 | bictcp_reset(inet_csk_ca(sk)); |
ae27e98a SH |
364 | bictcp_hystart_reset(sk); |
365 | } | |
366 | } | |
367 | ||
f278b99c ED |
368 | /* Account for TSO/GRO delays. |
369 | * Otherwise short RTT flows could get too small ssthresh, since during | |
370 | * slow start we begin with small TSO packets and ca->delay_min would | |
371 | * not account for long aggregation delay when TSO packets get bigger. | |
372 | * Ideally even with a very small RTT we would like to have at least one | |
373 | * TSO packet being sent and received by GRO, and another one in qdisc layer. | |
374 | * We apply another 100% factor because @rate is doubled at this point. | |
375 | * We cap the cushion to 1ms. | |
376 | */ | |
377 | static u32 hystart_ack_delay(struct sock *sk) | |
378 | { | |
379 | unsigned long rate; | |
380 | ||
381 | rate = READ_ONCE(sk->sk_pacing_rate); | |
382 | if (!rate) | |
383 | return 0; | |
384 | return min_t(u64, USEC_PER_MSEC, | |
385 | div64_ul((u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate)); | |
386 | } | |
387 | ||
ae27e98a SH |
388 | static void hystart_update(struct sock *sk, u32 delay) |
389 | { | |
390 | struct tcp_sock *tp = tcp_sk(sk); | |
391 | struct bictcp *ca = inet_csk_ca(sk); | |
ede656e8 | 392 | u32 threshold; |
ae27e98a | 393 | |
6e3a8a93 | 394 | if (hystart_detect & HYSTART_ACK_TRAIN) { |
cff04e2d | 395 | u32 now = bictcp_clock_us(sk); |
ae27e98a SH |
396 | |
397 | /* first detection parameter - ack-train detection */ | |
cff04e2d | 398 | if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) { |
17a6e9f1 | 399 | ca->last_ack = now; |
ede656e8 | 400 | |
f278b99c ED |
401 | threshold = ca->delay_min + hystart_ack_delay(sk); |
402 | ||
ede656e8 ED |
403 | /* Hystart ack train triggers if we get ack past |
404 | * ca->delay_min/2. | |
405 | * Pacing might have delayed packets up to RTT/2 | |
406 | * during slow start. | |
407 | */ | |
408 | if (sk->sk_pacing_status == SK_PACING_NONE) | |
409 | threshold >>= 1; | |
410 | ||
411 | if ((s32)(now - ca->round_start) > threshold) { | |
473900a5 | 412 | ca->found = 1; |
f278b99c ED |
413 | pr_debug("hystart_ack_train (%u > %u) delay_min %u (+ ack_delay %u) cwnd %u\n", |
414 | now - ca->round_start, threshold, | |
415 | ca->delay_min, hystart_ack_delay(sk), tp->snd_cwnd); | |
c10d9310 ED |
416 | NET_INC_STATS(sock_net(sk), |
417 | LINUX_MIB_TCPHYSTARTTRAINDETECT); | |
418 | NET_ADD_STATS(sock_net(sk), | |
419 | LINUX_MIB_TCPHYSTARTTRAINCWND, | |
420 | tp->snd_cwnd); | |
6e3a8a93 ED |
421 | tp->snd_ssthresh = tp->snd_cwnd; |
422 | } | |
ae27e98a | 423 | } |
6e3a8a93 | 424 | } |
ae27e98a | 425 | |
6e3a8a93 | 426 | if (hystart_detect & HYSTART_DELAY) { |
ae27e98a | 427 | /* obtain the minimum delay of more than sampling packets */ |
b344579c NC |
428 | if (ca->curr_rtt > delay) |
429 | ca->curr_rtt = delay; | |
ae27e98a | 430 | if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { |
ae27e98a SH |
431 | ca->sample_cnt++; |
432 | } else { | |
433 | if (ca->curr_rtt > ca->delay_min + | |
42eef7a0 | 434 | HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { |
473900a5 | 435 | ca->found = 1; |
c10d9310 ED |
436 | NET_INC_STATS(sock_net(sk), |
437 | LINUX_MIB_TCPHYSTARTDELAYDETECT); | |
438 | NET_ADD_STATS(sock_net(sk), | |
439 | LINUX_MIB_TCPHYSTARTDELAYCWND, | |
440 | tp->snd_cwnd); | |
6e3a8a93 ED |
441 | tp->snd_ssthresh = tp->snd_cwnd; |
442 | } | |
ae27e98a | 443 | } |
ae27e98a | 444 | } |
df3271f3 SH |
445 | } |
446 | ||
d22f6ad1 | 447 | static void cubictcp_acked(struct sock *sk, const struct ack_sample *sample) |
df3271f3 | 448 | { |
ae27e98a | 449 | const struct tcp_sock *tp = tcp_sk(sk); |
e7d0c885 SH |
450 | struct bictcp *ca = inet_csk_ca(sk); |
451 | u32 delay; | |
df3271f3 | 452 | |
e7d0c885 | 453 | /* Some calls are for duplicates without timetamps */ |
756ee172 | 454 | if (sample->rtt_us < 0) |
e7d0c885 SH |
455 | return; |
456 | ||
457 | /* Discard delay samples right after fast recovery */ | |
ac35f562 | 458 | if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ) |
e7d0c885 SH |
459 | return; |
460 | ||
cff04e2d | 461 | delay = sample->rtt_us; |
e7d0c885 SH |
462 | if (delay == 0) |
463 | delay = 1; | |
464 | ||
465 | /* first time call or link delay decreases */ | |
f278b99c ED |
466 | if (ca->delay_min == 0 || ca->delay_min > delay) |
467 | ca->delay_min = delay; | |
ae27e98a SH |
468 | |
469 | /* hystart triggers when cwnd is larger than some threshold */ | |
f278b99c | 470 | if (!ca->found && tcp_in_slow_start(tp) && hystart && |
ae27e98a SH |
471 | tp->snd_cwnd >= hystart_low_window) |
472 | hystart_update(sk, delay); | |
e7d0c885 | 473 | } |
df3271f3 | 474 | |
a252bebe | 475 | static struct tcp_congestion_ops cubictcp __read_mostly = { |
d22f6ad1 MKL |
476 | .init = cubictcp_init, |
477 | .ssthresh = cubictcp_recalc_ssthresh, | |
478 | .cong_avoid = cubictcp_cong_avoid, | |
479 | .set_state = cubictcp_state, | |
f1722a1b | 480 | .undo_cwnd = tcp_reno_undo_cwnd, |
d22f6ad1 MKL |
481 | .cwnd_event = cubictcp_cwnd_event, |
482 | .pkts_acked = cubictcp_acked, | |
df3271f3 SH |
483 | .owner = THIS_MODULE, |
484 | .name = "cubic", | |
485 | }; | |
486 | ||
0e32dfc8 KKD |
487 | BTF_SET_START(tcp_cubic_kfunc_ids) |
488 | #ifdef CONFIG_X86 | |
489 | #ifdef CONFIG_DYNAMIC_FTRACE | |
490 | BTF_ID(func, cubictcp_init) | |
491 | BTF_ID(func, cubictcp_recalc_ssthresh) | |
492 | BTF_ID(func, cubictcp_cong_avoid) | |
493 | BTF_ID(func, cubictcp_state) | |
494 | BTF_ID(func, cubictcp_cwnd_event) | |
495 | BTF_ID(func, cubictcp_acked) | |
496 | #endif | |
497 | #endif | |
498 | BTF_SET_END(tcp_cubic_kfunc_ids) | |
499 | ||
500 | static DEFINE_KFUNC_BTF_ID_SET(&tcp_cubic_kfunc_ids, tcp_cubic_kfunc_btf_set); | |
501 | ||
df3271f3 SH |
502 | static int __init cubictcp_register(void) |
503 | { | |
0e32dfc8 KKD |
504 | int ret; |
505 | ||
74975d40 | 506 | BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); |
89b3d9aa SH |
507 | |
508 | /* Precompute a bunch of the scaling factors that are used per-packet | |
509 | * based on SRTT of 100ms | |
510 | */ | |
511 | ||
688d1945 | 512 | beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 |
513 | / (BICTCP_BETA_SCALE - beta); | |
89b3d9aa | 514 | |
22119240 | 515 | cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ |
89b3d9aa SH |
516 | |
517 | /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 | |
518 | * so K = cubic_root( (wmax-cwnd)*rtt/c ) | |
519 | * the unit of K is bictcp_HZ=2^10, not HZ | |
520 | * | |
521 | * c = bic_scale >> 10 | |
522 | * rtt = 100ms | |
523 | * | |
524 | * the following code has been designed and tested for | |
525 | * cwnd < 1 million packets | |
526 | * RTT < 100 seconds | |
527 | * HZ < 1,000,00 (corresponding to 10 nano-second) | |
528 | */ | |
529 | ||
530 | /* 1/c * 2^2*bictcp_HZ * srtt */ | |
531 | cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ | |
532 | ||
533 | /* divide by bic_scale and by constant Srtt (100ms) */ | |
534 | do_div(cube_factor, bic_scale * 10); | |
535 | ||
0e32dfc8 KKD |
536 | ret = tcp_register_congestion_control(&cubictcp); |
537 | if (ret) | |
538 | return ret; | |
539 | register_kfunc_btf_id_set(&bpf_tcp_ca_kfunc_list, &tcp_cubic_kfunc_btf_set); | |
540 | return 0; | |
df3271f3 SH |
541 | } |
542 | ||
543 | static void __exit cubictcp_unregister(void) | |
544 | { | |
0e32dfc8 | 545 | unregister_kfunc_btf_id_set(&bpf_tcp_ca_kfunc_list, &tcp_cubic_kfunc_btf_set); |
df3271f3 SH |
546 | tcp_unregister_congestion_control(&cubictcp); |
547 | } | |
548 | ||
549 | module_init(cubictcp_register); | |
550 | module_exit(cubictcp_unregister); | |
551 | ||
552 | MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); | |
553 | MODULE_LICENSE("GPL"); | |
554 | MODULE_DESCRIPTION("CUBIC TCP"); | |
ae27e98a | 555 | MODULE_VERSION("2.3"); |