1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 /* WARNING: This implemenation 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 <linux/bpf.h>
18 #include <linux/stddef.h>
19 #include <linux/tcp.h>
20 #include "bpf_tcp_helpers.h"
21 
22 char _license[] SEC("license") = "GPL";
23 
24 #define clamp(val, lo, hi) min((typeof(val))max(val, lo), hi)
25 
26 #define BICTCP_BETA_SCALE    1024	/* Scale factor beta calculation
27 					 * max_cwnd = snd_cwnd * beta
28 					 */
29 #define	BICTCP_HZ		10	/* BIC HZ 2^10 = 1024 */
30 
31 /* Two methods of hybrid slow start */
32 #define HYSTART_ACK_TRAIN	0x1
33 #define HYSTART_DELAY		0x2
34 
35 /* Number of delay samples for detecting the increase of delay */
36 #define HYSTART_MIN_SAMPLES	8
37 #define HYSTART_DELAY_MIN	(4000U)	/* 4ms */
38 #define HYSTART_DELAY_MAX	(16000U)	/* 16 ms */
39 #define HYSTART_DELAY_THRESH(x)	clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
40 
41 static int fast_convergence = 1;
42 static const int beta = 717;	/* = 717/1024 (BICTCP_BETA_SCALE) */
43 static int initial_ssthresh;
44 static const int bic_scale = 41;
45 static int tcp_friendliness = 1;
46 
47 static int hystart = 1;
48 static int hystart_detect = HYSTART_ACK_TRAIN | HYSTART_DELAY;
49 static int hystart_low_window = 16;
50 static int hystart_ack_delta_us = 2000;
51 
52 static const __u32 cube_rtt_scale = (bic_scale * 10);	/* 1024*c/rtt */
53 static const __u32 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3
54 				/ (BICTCP_BETA_SCALE - beta);
55 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
56  *  so K = cubic_root( (wmax-cwnd)*rtt/c )
57  * the unit of K is bictcp_HZ=2^10, not HZ
58  *
59  *  c = bic_scale >> 10
60  *  rtt = 100ms
61  *
62  * the following code has been designed and tested for
63  * cwnd < 1 million packets
64  * RTT < 100 seconds
65  * HZ < 1,000,00  (corresponding to 10 nano-second)
66  */
67 
68 /* 1/c * 2^2*bictcp_HZ * srtt, 2^40 */
69 static const __u64 cube_factor = (__u64)(1ull << (10+3*BICTCP_HZ))
70 				/ (bic_scale * 10);
71 
72 /* BIC TCP Parameters */
73 struct bictcp {
74 	__u32	cnt;		/* increase cwnd by 1 after ACKs */
75 	__u32	last_max_cwnd;	/* last maximum snd_cwnd */
76 	__u32	last_cwnd;	/* the last snd_cwnd */
77 	__u32	last_time;	/* time when updated last_cwnd */
78 	__u32	bic_origin_point;/* origin point of bic function */
79 	__u32	bic_K;		/* time to origin point
80 				   from the beginning of the current epoch */
81 	__u32	delay_min;	/* min delay (usec) */
82 	__u32	epoch_start;	/* beginning of an epoch */
83 	__u32	ack_cnt;	/* number of acks */
84 	__u32	tcp_cwnd;	/* estimated tcp cwnd */
85 	__u16	unused;
86 	__u8	sample_cnt;	/* number of samples to decide curr_rtt */
87 	__u8	found;		/* the exit point is found? */
88 	__u32	round_start;	/* beginning of each round */
89 	__u32	end_seq;	/* end_seq of the round */
90 	__u32	last_ack;	/* last time when the ACK spacing is close */
91 	__u32	curr_rtt;	/* the minimum rtt of current round */
92 };
93 
94 static inline void bictcp_reset(struct bictcp *ca)
95 {
96 	ca->cnt = 0;
97 	ca->last_max_cwnd = 0;
98 	ca->last_cwnd = 0;
99 	ca->last_time = 0;
100 	ca->bic_origin_point = 0;
101 	ca->bic_K = 0;
102 	ca->delay_min = 0;
103 	ca->epoch_start = 0;
104 	ca->ack_cnt = 0;
105 	ca->tcp_cwnd = 0;
106 	ca->found = 0;
107 }
108 
109 extern unsigned long CONFIG_HZ __kconfig;
110 #define HZ CONFIG_HZ
111 #define USEC_PER_MSEC	1000UL
112 #define USEC_PER_SEC	1000000UL
113 #define USEC_PER_JIFFY	(USEC_PER_SEC / HZ)
114 
115 static __always_inline __u64 div64_u64(__u64 dividend, __u64 divisor)
116 {
117 	return dividend / divisor;
118 }
119 
120 #define div64_ul div64_u64
121 
122 #define BITS_PER_U64 (sizeof(__u64) * 8)
123 static __always_inline int fls64(__u64 x)
124 {
125 	int num = BITS_PER_U64 - 1;
126 
127 	if (x == 0)
128 		return 0;
129 
130 	if (!(x & (~0ull << (BITS_PER_U64-32)))) {
131 		num -= 32;
132 		x <<= 32;
133 	}
134 	if (!(x & (~0ull << (BITS_PER_U64-16)))) {
135 		num -= 16;
136 		x <<= 16;
137 	}
138 	if (!(x & (~0ull << (BITS_PER_U64-8)))) {
139 		num -= 8;
140 		x <<= 8;
141 	}
142 	if (!(x & (~0ull << (BITS_PER_U64-4)))) {
143 		num -= 4;
144 		x <<= 4;
145 	}
146 	if (!(x & (~0ull << (BITS_PER_U64-2)))) {
147 		num -= 2;
148 		x <<= 2;
149 	}
150 	if (!(x & (~0ull << (BITS_PER_U64-1))))
151 		num -= 1;
152 
153 	return num + 1;
154 }
155 
156 static __always_inline __u32 bictcp_clock_us(const struct sock *sk)
157 {
158 	return tcp_sk(sk)->tcp_mstamp;
159 }
160 
161 static __always_inline void bictcp_hystart_reset(struct sock *sk)
162 {
163 	struct tcp_sock *tp = tcp_sk(sk);
164 	struct bictcp *ca = inet_csk_ca(sk);
165 
166 	ca->round_start = ca->last_ack = bictcp_clock_us(sk);
167 	ca->end_seq = tp->snd_nxt;
168 	ca->curr_rtt = ~0U;
169 	ca->sample_cnt = 0;
170 }
171 
172 /* "struct_ops/" prefix is a requirement */
173 SEC("struct_ops/bpf_cubic_init")
174 void BPF_PROG(bpf_cubic_init, struct sock *sk)
175 {
176 	struct 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 
187 /* "struct_ops" prefix is a requirement */
188 SEC("struct_ops/bpf_cubic_cwnd_event")
189 void BPF_PROG(bpf_cubic_cwnd_event, struct sock *sk, enum tcp_ca_event event)
190 {
191 	if (event == CA_EVENT_TX_START) {
192 		struct bictcp *ca = inet_csk_ca(sk);
193 		__u32 now = tcp_jiffies32;
194 		__s32 delta;
195 
196 		delta = now - tcp_sk(sk)->lsndtime;
197 
198 		/* We were application limited (idle) for a while.
199 		 * Shift epoch_start to keep cwnd growth to cubic curve.
200 		 */
201 		if (ca->epoch_start && delta > 0) {
202 			ca->epoch_start += delta;
203 			if (after(ca->epoch_start, now))
204 				ca->epoch_start = now;
205 		}
206 		return;
207 	}
208 }
209 
210 /*
211  * cbrt(x) MSB values for x MSB values in [0..63].
212  * Precomputed then refined by hand - Willy Tarreau
213  *
214  * For x in [0..63],
215  *   v = cbrt(x << 18) - 1
216  *   cbrt(x) = (v[x] + 10) >> 6
217  */
218 static const __u8 v[] = {
219 	/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
220 	/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
221 	/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
222 	/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
223 	/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
224 	/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
225 	/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
226 	/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
227 };
228 
229 /* calculate the cubic root of x using a table lookup followed by one
230  * Newton-Raphson iteration.
231  * Avg err ~= 0.195%
232  */
233 static __always_inline __u32 cubic_root(__u64 a)
234 {
235 	__u32 x, b, shift;
236 
237 	if (a < 64) {
238 		/* a in [0..63] */
239 		return ((__u32)v[(__u32)a] + 35) >> 6;
240 	}
241 
242 	b = fls64(a);
243 	b = ((b * 84) >> 8) - 1;
244 	shift = (a >> (b * 3));
245 
246 	/* it is needed for verifier's bound check on v */
247 	if (shift >= 64)
248 		return 0;
249 
250 	x = ((__u32)(((__u32)v[shift] + 10) << b)) >> 6;
251 
252 	/*
253 	 * Newton-Raphson iteration
254 	 *                         2
255 	 * x    = ( 2 * x  +  a / x  ) / 3
256 	 *  k+1          k         k
257 	 */
258 	x = (2 * x + (__u32)div64_u64(a, (__u64)x * (__u64)(x - 1)));
259 	x = ((x * 341) >> 10);
260 	return x;
261 }
262 
263 /*
264  * Compute congestion window to use.
265  */
266 static __always_inline void bictcp_update(struct bictcp *ca, __u32 cwnd,
267 					  __u32 acked)
268 {
269 	__u32 delta, bic_target, max_cnt;
270 	__u64 offs, t;
271 
272 	ca->ack_cnt += acked;	/* count the number of ACKed packets */
273 
274 	if (ca->last_cwnd == cwnd &&
275 	    (__s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
276 		return;
277 
278 	/* The CUBIC function can update ca->cnt at most once per jiffy.
279 	 * On all cwnd reduction events, ca->epoch_start is set to 0,
280 	 * which will force a recalculation of ca->cnt.
281 	 */
282 	if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
283 		goto tcp_friendliness;
284 
285 	ca->last_cwnd = cwnd;
286 	ca->last_time = tcp_jiffies32;
287 
288 	if (ca->epoch_start == 0) {
289 		ca->epoch_start = tcp_jiffies32;	/* record beginning */
290 		ca->ack_cnt = acked;			/* start counting */
291 		ca->tcp_cwnd = cwnd;			/* syn with cubic */
292 
293 		if (ca->last_max_cwnd <= cwnd) {
294 			ca->bic_K = 0;
295 			ca->bic_origin_point = cwnd;
296 		} else {
297 			/* Compute new K based on
298 			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
299 			 */
300 			ca->bic_K = cubic_root(cube_factor
301 					       * (ca->last_max_cwnd - cwnd));
302 			ca->bic_origin_point = ca->last_max_cwnd;
303 		}
304 	}
305 
306 	/* cubic function - calc*/
307 	/* calculate c * time^3 / rtt,
308 	 *  while considering overflow in calculation of time^3
309 	 * (so time^3 is done by using 64 bit)
310 	 * and without the support of division of 64bit numbers
311 	 * (so all divisions are done by using 32 bit)
312 	 *  also NOTE the unit of those veriables
313 	 *	  time  = (t - K) / 2^bictcp_HZ
314 	 *	  c = bic_scale >> 10
315 	 * rtt  = (srtt >> 3) / HZ
316 	 * !!! The following code does not have overflow problems,
317 	 * if the cwnd < 1 million packets !!!
318 	 */
319 
320 	t = (__s32)(tcp_jiffies32 - ca->epoch_start) * USEC_PER_JIFFY;
321 	t += ca->delay_min;
322 	/* change the unit from usec to bictcp_HZ */
323 	t <<= BICTCP_HZ;
324 	t /= USEC_PER_SEC;
325 
326 	if (t < ca->bic_K)		/* t - K */
327 		offs = ca->bic_K - t;
328 	else
329 		offs = t - ca->bic_K;
330 
331 	/* c/rtt * (t-K)^3 */
332 	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
333 	if (t < ca->bic_K)                            /* below origin*/
334 		bic_target = ca->bic_origin_point - delta;
335 	else                                          /* above origin*/
336 		bic_target = ca->bic_origin_point + delta;
337 
338 	/* cubic function - calc bictcp_cnt*/
339 	if (bic_target > cwnd) {
340 		ca->cnt = cwnd / (bic_target - cwnd);
341 	} else {
342 		ca->cnt = 100 * cwnd;              /* very small increment*/
343 	}
344 
345 	/*
346 	 * The initial growth of cubic function may be too conservative
347 	 * when the available bandwidth is still unknown.
348 	 */
349 	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
350 		ca->cnt = 20;	/* increase cwnd 5% per RTT */
351 
352 tcp_friendliness:
353 	/* TCP Friendly */
354 	if (tcp_friendliness) {
355 		__u32 scale = beta_scale;
356 		__u32 n;
357 
358 		/* update tcp cwnd */
359 		delta = (cwnd * scale) >> 3;
360 		if (ca->ack_cnt > delta && delta) {
361 			n = ca->ack_cnt / delta;
362 			ca->ack_cnt -= n * delta;
363 			ca->tcp_cwnd += n;
364 		}
365 
366 		if (ca->tcp_cwnd > cwnd) {	/* if bic is slower than tcp */
367 			delta = ca->tcp_cwnd - cwnd;
368 			max_cnt = cwnd / delta;
369 			if (ca->cnt > max_cnt)
370 				ca->cnt = max_cnt;
371 		}
372 	}
373 
374 	/* The maximum rate of cwnd increase CUBIC allows is 1 packet per
375 	 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
376 	 */
377 	ca->cnt = max(ca->cnt, 2U);
378 }
379 
380 /* Or simply use the BPF_STRUCT_OPS to avoid the SEC boiler plate. */
381 void BPF_STRUCT_OPS(bpf_cubic_cong_avoid, struct sock *sk, __u32 ack, __u32 acked)
382 {
383 	struct tcp_sock *tp = tcp_sk(sk);
384 	struct bictcp *ca = inet_csk_ca(sk);
385 
386 	if (!tcp_is_cwnd_limited(sk))
387 		return;
388 
389 	if (tcp_in_slow_start(tp)) {
390 		if (hystart && after(ack, ca->end_seq))
391 			bictcp_hystart_reset(sk);
392 		acked = tcp_slow_start(tp, acked);
393 		if (!acked)
394 			return;
395 	}
396 	bictcp_update(ca, tp->snd_cwnd, acked);
397 	tcp_cong_avoid_ai(tp, ca->cnt, acked);
398 }
399 
400 __u32 BPF_STRUCT_OPS(bpf_cubic_recalc_ssthresh, struct sock *sk)
401 {
402 	const struct tcp_sock *tp = tcp_sk(sk);
403 	struct bictcp *ca = inet_csk_ca(sk);
404 
405 	ca->epoch_start = 0;	/* end of epoch */
406 
407 	/* Wmax and fast convergence */
408 	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
409 		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
410 			/ (2 * BICTCP_BETA_SCALE);
411 	else
412 		ca->last_max_cwnd = tp->snd_cwnd;
413 
414 	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
415 }
416 
417 void BPF_STRUCT_OPS(bpf_cubic_state, struct sock *sk, __u8 new_state)
418 {
419 	if (new_state == TCP_CA_Loss) {
420 		bictcp_reset(inet_csk_ca(sk));
421 		bictcp_hystart_reset(sk);
422 	}
423 }
424 
425 #define GSO_MAX_SIZE		65536
426 
427 /* Account for TSO/GRO delays.
428  * Otherwise short RTT flows could get too small ssthresh, since during
429  * slow start we begin with small TSO packets and ca->delay_min would
430  * not account for long aggregation delay when TSO packets get bigger.
431  * Ideally even with a very small RTT we would like to have at least one
432  * TSO packet being sent and received by GRO, and another one in qdisc layer.
433  * We apply another 100% factor because @rate is doubled at this point.
434  * We cap the cushion to 1ms.
435  */
436 static __always_inline __u32 hystart_ack_delay(struct sock *sk)
437 {
438 	unsigned long rate;
439 
440 	rate = sk->sk_pacing_rate;
441 	if (!rate)
442 		return 0;
443 	return min((__u64)USEC_PER_MSEC,
444 		   div64_ul((__u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate));
445 }
446 
447 static __always_inline void hystart_update(struct sock *sk, __u32 delay)
448 {
449 	struct tcp_sock *tp = tcp_sk(sk);
450 	struct bictcp *ca = inet_csk_ca(sk);
451 	__u32 threshold;
452 
453 	if (hystart_detect & HYSTART_ACK_TRAIN) {
454 		__u32 now = bictcp_clock_us(sk);
455 
456 		/* first detection parameter - ack-train detection */
457 		if ((__s32)(now - ca->last_ack) <= hystart_ack_delta_us) {
458 			ca->last_ack = now;
459 
460 			threshold = ca->delay_min + hystart_ack_delay(sk);
461 
462 			/* Hystart ack train triggers if we get ack past
463 			 * ca->delay_min/2.
464 			 * Pacing might have delayed packets up to RTT/2
465 			 * during slow start.
466 			 */
467 			if (sk->sk_pacing_status == SK_PACING_NONE)
468 				threshold >>= 1;
469 
470 			if ((__s32)(now - ca->round_start) > threshold) {
471 				ca->found = 1;
472 				tp->snd_ssthresh = tp->snd_cwnd;
473 			}
474 		}
475 	}
476 
477 	if (hystart_detect & HYSTART_DELAY) {
478 		/* obtain the minimum delay of more than sampling packets */
479 		if (ca->curr_rtt > delay)
480 			ca->curr_rtt = delay;
481 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
482 			ca->sample_cnt++;
483 		} else {
484 			if (ca->curr_rtt > ca->delay_min +
485 			    HYSTART_DELAY_THRESH(ca->delay_min >> 3)) {
486 				ca->found = 1;
487 				tp->snd_ssthresh = tp->snd_cwnd;
488 			}
489 		}
490 	}
491 }
492 
493 int bpf_cubic_acked_called = 0;
494 
495 void BPF_STRUCT_OPS(bpf_cubic_acked, struct sock *sk,
496 		    const struct ack_sample *sample)
497 {
498 	const struct tcp_sock *tp = tcp_sk(sk);
499 	struct bictcp *ca = inet_csk_ca(sk);
500 	__u32 delay;
501 
502 	bpf_cubic_acked_called = 1;
503 	/* Some calls are for duplicates without timetamps */
504 	if (sample->rtt_us < 0)
505 		return;
506 
507 	/* Discard delay samples right after fast recovery */
508 	if (ca->epoch_start && (__s32)(tcp_jiffies32 - ca->epoch_start) < HZ)
509 		return;
510 
511 	delay = sample->rtt_us;
512 	if (delay == 0)
513 		delay = 1;
514 
515 	/* first time call or link delay decreases */
516 	if (ca->delay_min == 0 || ca->delay_min > delay)
517 		ca->delay_min = delay;
518 
519 	/* hystart triggers when cwnd is larger than some threshold */
520 	if (!ca->found && tcp_in_slow_start(tp) && hystart &&
521 	    tp->snd_cwnd >= hystart_low_window)
522 		hystart_update(sk, delay);
523 }
524 
525 extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym;
526 
527 __u32 BPF_STRUCT_OPS(bpf_cubic_undo_cwnd, struct sock *sk)
528 {
529 	return tcp_reno_undo_cwnd(sk);
530 }
531 
532 SEC(".struct_ops")
533 struct tcp_congestion_ops cubic = {
534 	.init		= (void *)bpf_cubic_init,
535 	.ssthresh	= (void *)bpf_cubic_recalc_ssthresh,
536 	.cong_avoid	= (void *)bpf_cubic_cong_avoid,
537 	.set_state	= (void *)bpf_cubic_state,
538 	.undo_cwnd	= (void *)bpf_cubic_undo_cwnd,
539 	.cwnd_event	= (void *)bpf_cubic_cwnd_event,
540 	.pkts_acked     = (void *)bpf_cubic_acked,
541 	.name		= "bpf_cubic",
542 };
543