Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1// SPDX-License-Identifier: GPL-2.0-only
2
3/* WARNING: This implementation is not necessarily the same
4 * as the tcp_cubic.c. The purpose is mainly for testing
5 * the kernel BPF logic.
6 *
7 * Highlights:
8 * 1. CONFIG_HZ .kconfig map is used.
9 * 2. In bictcp_update(), calculation is changed to use usec
10 * resolution (i.e. USEC_PER_JIFFY) instead of using jiffies.
11 * Thus, usecs_to_jiffies() is not used in the bpf_cubic.c.
12 * 3. In bitctcp_update() [under tcp_friendliness], the original
13 * "while (ca->ack_cnt > delta)" loop is changed to the equivalent
14 * "ca->ack_cnt / delta" operation.
15 */
16
17#include "bpf_tracing_net.h"
18#include <bpf/bpf_tracing.h>
19
20char _license[] SEC("license") = "GPL";
21
22#define clamp(val, lo, hi) min((typeof(val))max(val, lo), hi)
23
24extern __u32 tcp_slow_start(struct tcp_sock *tp, __u32 acked) __ksym;
25extern void tcp_cong_avoid_ai(struct tcp_sock *tp, __u32 w, __u32 acked) __ksym;
26
27#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
28 * max_cwnd = snd_cwnd * beta
29 */
30#define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
31
32/* Two methods of hybrid slow start */
33#define HYSTART_ACK_TRAIN 0x1
34#define HYSTART_DELAY 0x2
35
36/* Number of delay samples for detecting the increase of delay */
37#define HYSTART_MIN_SAMPLES 8
38#define HYSTART_DELAY_MIN (4000U) /* 4ms */
39#define HYSTART_DELAY_MAX (16000U) /* 16 ms */
40#define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
41
42static int fast_convergence = 1;
43static const int beta = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
44static int initial_ssthresh;
45static const int bic_scale = 41;
46static int tcp_friendliness = 1;
47
48static int hystart = 1;
49static int hystart_detect = HYSTART_ACK_TRAIN | HYSTART_DELAY;
50static int hystart_low_window = 16;
51static int hystart_ack_delta_us = 2000;
52
53static const __u32 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
54static const __u32 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
55 / (BICTCP_BETA_SCALE - beta);
56/* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
57 * so K = cubic_root( (wmax-cwnd)*rtt/c )
58 * the unit of K is bictcp_HZ=2^10, not HZ
59 *
60 * c = bic_scale >> 10
61 * rtt = 100ms
62 *
63 * the following code has been designed and tested for
64 * cwnd < 1 million packets
65 * RTT < 100 seconds
66 * HZ < 1,000,00 (corresponding to 10 nano-second)
67 */
68
69/* 1/c * 2^2*bictcp_HZ * srtt, 2^40 */
70static const __u64 cube_factor = (__u64)(1ull << (10+3*BICTCP_HZ))
71 / (bic_scale * 10);
72
73/* BIC TCP Parameters */
74struct bpf_bictcp {
75 __u32 cnt; /* increase cwnd by 1 after ACKs */
76 __u32 last_max_cwnd; /* last maximum snd_cwnd */
77 __u32 last_cwnd; /* the last snd_cwnd */
78 __u32 last_time; /* time when updated last_cwnd */
79 __u32 bic_origin_point;/* origin point of bic function */
80 __u32 bic_K; /* time to origin point
81 from the beginning of the current epoch */
82 __u32 delay_min; /* min delay (usec) */
83 __u32 epoch_start; /* beginning of an epoch */
84 __u32 ack_cnt; /* number of acks */
85 __u32 tcp_cwnd; /* estimated tcp cwnd */
86 __u16 unused;
87 __u8 sample_cnt; /* number of samples to decide curr_rtt */
88 __u8 found; /* the exit point is found? */
89 __u32 round_start; /* beginning of each round */
90 __u32 end_seq; /* end_seq of the round */
91 __u32 last_ack; /* last time when the ACK spacing is close */
92 __u32 curr_rtt; /* the minimum rtt of current round */
93};
94
95static void bictcp_reset(struct bpf_bictcp *ca)
96{
97 ca->cnt = 0;
98 ca->last_max_cwnd = 0;
99 ca->last_cwnd = 0;
100 ca->last_time = 0;
101 ca->bic_origin_point = 0;
102 ca->bic_K = 0;
103 ca->delay_min = 0;
104 ca->epoch_start = 0;
105 ca->ack_cnt = 0;
106 ca->tcp_cwnd = 0;
107 ca->found = 0;
108}
109
110extern unsigned long CONFIG_HZ __kconfig;
111#define HZ CONFIG_HZ
112#define USEC_PER_MSEC 1000UL
113#define USEC_PER_SEC 1000000UL
114#define USEC_PER_JIFFY (USEC_PER_SEC / HZ)
115
116static __u64 div64_u64(__u64 dividend, __u64 divisor)
117{
118 return dividend / divisor;
119}
120
121#define div64_ul div64_u64
122
123#define BITS_PER_U64 (sizeof(__u64) * 8)
124static int fls64(__u64 x)
125{
126 int num = BITS_PER_U64 - 1;
127
128 if (x == 0)
129 return 0;
130
131 if (!(x & (~0ull << (BITS_PER_U64-32)))) {
132 num -= 32;
133 x <<= 32;
134 }
135 if (!(x & (~0ull << (BITS_PER_U64-16)))) {
136 num -= 16;
137 x <<= 16;
138 }
139 if (!(x & (~0ull << (BITS_PER_U64-8)))) {
140 num -= 8;
141 x <<= 8;
142 }
143 if (!(x & (~0ull << (BITS_PER_U64-4)))) {
144 num -= 4;
145 x <<= 4;
146 }
147 if (!(x & (~0ull << (BITS_PER_U64-2)))) {
148 num -= 2;
149 x <<= 2;
150 }
151 if (!(x & (~0ull << (BITS_PER_U64-1))))
152 num -= 1;
153
154 return num + 1;
155}
156
157static __u32 bictcp_clock_us(const struct sock *sk)
158{
159 return tcp_sk(sk)->tcp_mstamp;
160}
161
162static void bictcp_hystart_reset(struct sock *sk)
163{
164 struct tcp_sock *tp = tcp_sk(sk);
165 struct bpf_bictcp *ca = inet_csk_ca(sk);
166
167 ca->round_start = ca->last_ack = bictcp_clock_us(sk);
168 ca->end_seq = tp->snd_nxt;
169 ca->curr_rtt = ~0U;
170 ca->sample_cnt = 0;
171}
172
173SEC("struct_ops")
174void BPF_PROG(bpf_cubic_init, struct sock *sk)
175{
176 struct bpf_bictcp *ca = inet_csk_ca(sk);
177
178 bictcp_reset(ca);
179
180 if (hystart)
181 bictcp_hystart_reset(sk);
182
183 if (!hystart && initial_ssthresh)
184 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
185}
186
187SEC("struct_ops")
188void BPF_PROG(bpf_cubic_cwnd_event_tx_start, struct sock *sk)
189{
190 struct bpf_bictcp *ca = inet_csk_ca(sk);
191 __u32 now = tcp_jiffies32;
192 __s32 delta;
193
194 delta = now - tcp_sk(sk)->lsndtime;
195
196 /* We were application limited (idle) for a while.
197 * Shift epoch_start to keep cwnd growth to cubic curve.
198 */
199 if (ca->epoch_start && delta > 0) {
200 ca->epoch_start += delta;
201 if (after(ca->epoch_start, now))
202 ca->epoch_start = now;
203 }
204}
205
206/*
207 * cbrt(x) MSB values for x MSB values in [0..63].
208 * Precomputed then refined by hand - Willy Tarreau
209 *
210 * For x in [0..63],
211 * v = cbrt(x << 18) - 1
212 * cbrt(x) = (v[x] + 10) >> 6
213 */
214static const __u8 v[] = {
215 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
216 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
217 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
218 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
219 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
220 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
221 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
222 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
223};
224
225/* calculate the cubic root of x using a table lookup followed by one
226 * Newton-Raphson iteration.
227 * Avg err ~= 0.195%
228 */
229static __u32 cubic_root(__u64 a)
230{
231 __u32 x, b, shift;
232
233 if (a < 64) {
234 /* a in [0..63] */
235 return ((__u32)v[(__u32)a] + 35) >> 6;
236 }
237
238 b = fls64(a);
239 b = ((b * 84) >> 8) - 1;
240 shift = (a >> (b * 3));
241
242 /* it is needed for verifier's bound check on v */
243 if (shift >= 64)
244 return 0;
245
246 x = ((__u32)(((__u32)v[shift] + 10) << b)) >> 6;
247
248 /*
249 * Newton-Raphson iteration
250 * 2
251 * x = ( 2 * x + a / x ) / 3
252 * k+1 k k
253 */
254 x = (2 * x + (__u32)div64_u64(a, (__u64)x * (__u64)(x - 1)));
255 x = ((x * 341) >> 10);
256 return x;
257}
258
259/*
260 * Compute congestion window to use.
261 */
262static void bictcp_update(struct bpf_bictcp *ca, __u32 cwnd, __u32 acked)
263{
264 __u32 delta, bic_target, max_cnt;
265 __u64 offs, t;
266
267 ca->ack_cnt += acked; /* count the number of ACKed packets */
268
269 if (ca->last_cwnd == cwnd &&
270 (__s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
271 return;
272
273 /* The CUBIC function can update ca->cnt at most once per jiffy.
274 * On all cwnd reduction events, ca->epoch_start is set to 0,
275 * which will force a recalculation of ca->cnt.
276 */
277 if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
278 goto tcp_friendliness;
279
280 ca->last_cwnd = cwnd;
281 ca->last_time = tcp_jiffies32;
282
283 if (ca->epoch_start == 0) {
284 ca->epoch_start = tcp_jiffies32; /* record beginning */
285 ca->ack_cnt = acked; /* start counting */
286 ca->tcp_cwnd = cwnd; /* syn with cubic */
287
288 if (ca->last_max_cwnd <= cwnd) {
289 ca->bic_K = 0;
290 ca->bic_origin_point = cwnd;
291 } else {
292 /* Compute new K based on
293 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
294 */
295 ca->bic_K = cubic_root(cube_factor
296 * (ca->last_max_cwnd - cwnd));
297 ca->bic_origin_point = ca->last_max_cwnd;
298 }
299 }
300
301 /* cubic function - calc*/
302 /* calculate c * time^3 / rtt,
303 * while considering overflow in calculation of time^3
304 * (so time^3 is done by using 64 bit)
305 * and without the support of division of 64bit numbers
306 * (so all divisions are done by using 32 bit)
307 * also NOTE the unit of those variables
308 * time = (t - K) / 2^bictcp_HZ
309 * c = bic_scale >> 10
310 * rtt = (srtt >> 3) / HZ
311 * !!! The following code does not have overflow problems,
312 * if the cwnd < 1 million packets !!!
313 */
314
315 t = (__s32)(tcp_jiffies32 - ca->epoch_start) * USEC_PER_JIFFY;
316 t += ca->delay_min;
317 /* change the unit from usec to bictcp_HZ */
318 t <<= BICTCP_HZ;
319 t /= USEC_PER_SEC;
320
321 if (t < ca->bic_K) /* t - K */
322 offs = ca->bic_K - t;
323 else
324 offs = t - ca->bic_K;
325
326 /* c/rtt * (t-K)^3 */
327 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
328 if (t < ca->bic_K) /* below origin*/
329 bic_target = ca->bic_origin_point - delta;
330 else /* above origin*/
331 bic_target = ca->bic_origin_point + delta;
332
333 /* cubic function - calc bictcp_cnt*/
334 if (bic_target > cwnd) {
335 ca->cnt = cwnd / (bic_target - cwnd);
336 } else {
337 ca->cnt = 100 * cwnd; /* very small increment*/
338 }
339
340 /*
341 * The initial growth of cubic function may be too conservative
342 * when the available bandwidth is still unknown.
343 */
344 if (ca->last_max_cwnd == 0 && ca->cnt > 20)
345 ca->cnt = 20; /* increase cwnd 5% per RTT */
346
347tcp_friendliness:
348 /* TCP Friendly */
349 if (tcp_friendliness) {
350 __u32 scale = beta_scale;
351 __u32 n;
352
353 /* update tcp cwnd */
354 delta = (cwnd * scale) >> 3;
355 if (ca->ack_cnt > delta && delta) {
356 n = ca->ack_cnt / delta;
357 ca->ack_cnt -= n * delta;
358 ca->tcp_cwnd += n;
359 }
360
361 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */
362 delta = ca->tcp_cwnd - cwnd;
363 max_cnt = cwnd / delta;
364 if (ca->cnt > max_cnt)
365 ca->cnt = max_cnt;
366 }
367 }
368
369 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per
370 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
371 */
372 ca->cnt = max(ca->cnt, 2U);
373}
374
375SEC("struct_ops")
376void BPF_PROG(bpf_cubic_cong_avoid, struct sock *sk, __u32 ack, __u32 acked)
377{
378 struct tcp_sock *tp = tcp_sk(sk);
379 struct bpf_bictcp *ca = inet_csk_ca(sk);
380
381 if (!tcp_is_cwnd_limited(sk))
382 return;
383
384 if (tcp_in_slow_start(tp)) {
385 if (hystart && after(ack, ca->end_seq))
386 bictcp_hystart_reset(sk);
387 acked = tcp_slow_start(tp, acked);
388 if (!acked)
389 return;
390 }
391 bictcp_update(ca, tp->snd_cwnd, acked);
392 tcp_cong_avoid_ai(tp, ca->cnt, acked);
393}
394
395SEC("struct_ops")
396__u32 BPF_PROG(bpf_cubic_recalc_ssthresh, struct sock *sk)
397{
398 const struct tcp_sock *tp = tcp_sk(sk);
399 struct bpf_bictcp *ca = inet_csk_ca(sk);
400
401 ca->epoch_start = 0; /* end of epoch */
402
403 /* Wmax and fast convergence */
404 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
405 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
406 / (2 * BICTCP_BETA_SCALE);
407 else
408 ca->last_max_cwnd = tp->snd_cwnd;
409
410 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
411}
412
413SEC("struct_ops")
414void BPF_PROG(bpf_cubic_state, struct sock *sk, __u8 new_state)
415{
416 if (new_state == TCP_CA_Loss) {
417 bictcp_reset(inet_csk_ca(sk));
418 bictcp_hystart_reset(sk);
419 }
420}
421
422#define GSO_MAX_SIZE 65536
423
424/* Account for TSO/GRO delays.
425 * Otherwise short RTT flows could get too small ssthresh, since during
426 * slow start we begin with small TSO packets and ca->delay_min would
427 * not account for long aggregation delay when TSO packets get bigger.
428 * Ideally even with a very small RTT we would like to have at least one
429 * TSO packet being sent and received by GRO, and another one in qdisc layer.
430 * We apply another 100% factor because @rate is doubled at this point.
431 * We cap the cushion to 1ms.
432 */
433static __u32 hystart_ack_delay(struct sock *sk)
434{
435 unsigned long rate;
436
437 rate = sk->sk_pacing_rate;
438 if (!rate)
439 return 0;
440 return min((__u64)USEC_PER_MSEC,
441 div64_ul((__u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate));
442}
443
444static void hystart_update(struct sock *sk, __u32 delay)
445{
446 struct tcp_sock *tp = tcp_sk(sk);
447 struct bpf_bictcp *ca = inet_csk_ca(sk);
448 __u32 threshold;
449
450 if (hystart_detect & HYSTART_ACK_TRAIN) {
451 __u32 now = bictcp_clock_us(sk);
452
453 /* first detection parameter - ack-train detection */
454 if ((__s32)(now - ca->last_ack) <= hystart_ack_delta_us) {
455 ca->last_ack = now;
456
457 threshold = ca->delay_min + hystart_ack_delay(sk);
458
459 /* Hystart ack train triggers if we get ack past
460 * ca->delay_min/2.
461 * Pacing might have delayed packets up to RTT/2
462 * during slow start.
463 */
464 if (sk->sk_pacing_status == SK_PACING_NONE)
465 threshold >>= 1;
466
467 if ((__s32)(now - ca->round_start) > threshold) {
468 ca->found = 1;
469 tp->snd_ssthresh = tp->snd_cwnd;
470 }
471 }
472 }
473
474 if (hystart_detect & HYSTART_DELAY) {
475 /* obtain the minimum delay of more than sampling packets */
476 if (ca->curr_rtt > delay)
477 ca->curr_rtt = delay;
478 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
479 ca->sample_cnt++;
480 } else {
481 if (ca->curr_rtt > ca->delay_min +
482 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
483 ca->found = 1;
484 tp->snd_ssthresh = tp->snd_cwnd;
485 }
486 }
487 }
488}
489
490int bpf_cubic_acked_called = 0;
491
492SEC("struct_ops")
493void BPF_PROG(bpf_cubic_acked, struct sock *sk, const struct ack_sample *sample)
494{
495 const struct tcp_sock *tp = tcp_sk(sk);
496 struct bpf_bictcp *ca = inet_csk_ca(sk);
497 __u32 delay;
498
499 bpf_cubic_acked_called = 1;
500 /* Some calls are for duplicates without timestamps */
501 if (sample->rtt_us < 0)
502 return;
503
504 /* Discard delay samples right after fast recovery */
505 if (ca->epoch_start && (__s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
506 return;
507
508 delay = sample->rtt_us;
509 if (delay == 0)
510 delay = 1;
511
512 /* first time call or link delay decreases */
513 if (ca->delay_min == 0 || ca->delay_min > delay)
514 ca->delay_min = delay;
515
516 /* hystart triggers when cwnd is larger than some threshold */
517 if (!ca->found && tcp_in_slow_start(tp) && hystart &&
518 tp->snd_cwnd >= hystart_low_window)
519 hystart_update(sk, delay);
520}
521
522extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym;
523
524SEC("struct_ops")
525__u32 BPF_PROG(bpf_cubic_undo_cwnd, struct sock *sk)
526{
527 return tcp_reno_undo_cwnd(sk);
528}
529
530SEC(".struct_ops")
531struct tcp_congestion_ops cubic = {
532 .init = (void *)bpf_cubic_init,
533 .ssthresh = (void *)bpf_cubic_recalc_ssthresh,
534 .cong_avoid = (void *)bpf_cubic_cong_avoid,
535 .set_state = (void *)bpf_cubic_state,
536 .undo_cwnd = (void *)bpf_cubic_undo_cwnd,
537 .cwnd_event_tx_start = (void *)bpf_cubic_cwnd_event_tx_start,
538 .pkts_acked = (void *)bpf_cubic_acked,
539 .name = "bpf_cubic",
540};