bpf_cubic.c

  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, struct sock *sk, enum tcp_ca_event event)
189{
190	if (event == CA_EVENT_TX_START) {
191		struct bpf_bictcp *ca = inet_csk_ca(sk);
192		__u32 now = tcp_jiffies32;
193		__s32 delta;
194
195		delta = now - tcp_sk(sk)->lsndtime;
196
197		/* We were application limited (idle) for a while.
198		 * Shift epoch_start to keep cwnd growth to cubic curve.
199		 */
200		if (ca->epoch_start && delta > 0) {
201			ca->epoch_start += delta;
202			if (after(ca->epoch_start, now))
203				ca->epoch_start = now;
204		}
205		return;
206	}
207}
208
209/*
210 * cbrt(x) MSB values for x MSB values in [0..63].
211 * Precomputed then refined by hand - Willy Tarreau
212 *
213 * For x in [0..63],
214 *   v = cbrt(x << 18) - 1
215 *   cbrt(x) = (v[x] + 10) >> 6
216 */
217static const __u8 v[] = {
218	/* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
219	/* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
220	/* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
221	/* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
222	/* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
223	/* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
224	/* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
225	/* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
226};
227
228/* calculate the cubic root of x using a table lookup followed by one
229 * Newton-Raphson iteration.
230 * Avg err ~= 0.195%
231 */
232static __u32 cubic_root(__u64 a)
233{
234	__u32 x, b, shift;
235
236	if (a < 64) {
237		/* a in [0..63] */
238		return ((__u32)v[(__u32)a] + 35) >> 6;
239	}
240
241	b = fls64(a);
242	b = ((b * 84) >> 8) - 1;
243	shift = (a >> (b * 3));
244
245	/* it is needed for verifier's bound check on v */
246	if (shift >= 64)
247		return 0;
248
249	x = ((__u32)(((__u32)v[shift] + 10) << b)) >> 6;
250
251	/*
252	 * Newton-Raphson iteration
253	 *                         2
254	 * x    = ( 2 * x  +  a / x  ) / 3
255	 *  k+1          k         k
256	 */
257	x = (2 * x + (__u32)div64_u64(a, (__u64)x * (__u64)(x - 1)));
258	x = ((x * 341) >> 10);
259	return x;
260}
261
262/*
263 * Compute congestion window to use.
264 */
265static void bictcp_update(struct bpf_bictcp *ca, __u32 cwnd, __u32 acked)
266{
267	__u32 delta, bic_target, max_cnt;
268	__u64 offs, t;
269
270	ca->ack_cnt += acked;	/* count the number of ACKed packets */
271
272	if (ca->last_cwnd == cwnd &&
273	    (__s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
274		return;
275
276	/* The CUBIC function can update ca->cnt at most once per jiffy.
277	 * On all cwnd reduction events, ca->epoch_start is set to 0,
278	 * which will force a recalculation of ca->cnt.
279	 */
280	if (ca->epoch_start && tcp_jiffies32 == ca->last_time)
281		goto tcp_friendliness;
282
283	ca->last_cwnd = cwnd;
284	ca->last_time = tcp_jiffies32;
285
286	if (ca->epoch_start == 0) {
287		ca->epoch_start = tcp_jiffies32;	/* record beginning */
288		ca->ack_cnt = acked;			/* start counting */
289		ca->tcp_cwnd = cwnd;			/* syn with cubic */
290
291		if (ca->last_max_cwnd <= cwnd) {
292			ca->bic_K = 0;
293			ca->bic_origin_point = cwnd;
294		} else {
295			/* Compute new K based on
296			 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
297			 */
298			ca->bic_K = cubic_root(cube_factor
299					       * (ca->last_max_cwnd - cwnd));
300			ca->bic_origin_point = ca->last_max_cwnd;
301		}
302	}
303
304	/* cubic function - calc*/
305	/* calculate c * time^3 / rtt,
306	 *  while considering overflow in calculation of time^3
307	 * (so time^3 is done by using 64 bit)
308	 * and without the support of division of 64bit numbers
309	 * (so all divisions are done by using 32 bit)
310	 *  also NOTE the unit of those variables
311	 *	  time  = (t - K) / 2^bictcp_HZ
312	 *	  c = bic_scale >> 10
313	 * rtt  = (srtt >> 3) / HZ
314	 * !!! The following code does not have overflow problems,
315	 * if the cwnd < 1 million packets !!!
316	 */
317
318	t = (__s32)(tcp_jiffies32 - ca->epoch_start) * USEC_PER_JIFFY;
319	t += ca->delay_min;
320	/* change the unit from usec to bictcp_HZ */
321	t <<= BICTCP_HZ;
322	t /= USEC_PER_SEC;
323
324	if (t < ca->bic_K)		/* t - K */
325		offs = ca->bic_K - t;
326	else
327		offs = t - ca->bic_K;
328
329	/* c/rtt * (t-K)^3 */
330	delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
331	if (t < ca->bic_K)                            /* below origin*/
332		bic_target = ca->bic_origin_point - delta;
333	else                                          /* above origin*/
334		bic_target = ca->bic_origin_point + delta;
335
336	/* cubic function - calc bictcp_cnt*/
337	if (bic_target > cwnd) {
338		ca->cnt = cwnd / (bic_target - cwnd);
339	} else {
340		ca->cnt = 100 * cwnd;              /* very small increment*/
341	}
342
343	/*
344	 * The initial growth of cubic function may be too conservative
345	 * when the available bandwidth is still unknown.
346	 */
347	if (ca->last_max_cwnd == 0 && ca->cnt > 20)
348		ca->cnt = 20;	/* increase cwnd 5% per RTT */
349
350tcp_friendliness:
351	/* TCP Friendly */
352	if (tcp_friendliness) {
353		__u32 scale = beta_scale;
354		__u32 n;
355
356		/* update tcp cwnd */
357		delta = (cwnd * scale) >> 3;
358		if (ca->ack_cnt > delta && delta) {
359			n = ca->ack_cnt / delta;
360			ca->ack_cnt -= n * delta;
361			ca->tcp_cwnd += n;
362		}
363
364		if (ca->tcp_cwnd > cwnd) {	/* if bic is slower than tcp */
365			delta = ca->tcp_cwnd - cwnd;
366			max_cnt = cwnd / delta;
367			if (ca->cnt > max_cnt)
368				ca->cnt = max_cnt;
369		}
370	}
371
372	/* The maximum rate of cwnd increase CUBIC allows is 1 packet per
373	 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT.
374	 */
375	ca->cnt = max(ca->cnt, 2U);
376}
377
378SEC("struct_ops")
379void BPF_PROG(bpf_cubic_cong_avoid, struct sock *sk, __u32 ack, __u32 acked)
380{
381	struct tcp_sock *tp = tcp_sk(sk);
382	struct bpf_bictcp *ca = inet_csk_ca(sk);
383
384	if (!tcp_is_cwnd_limited(sk))
385		return;
386
387	if (tcp_in_slow_start(tp)) {
388		if (hystart && after(ack, ca->end_seq))
389			bictcp_hystart_reset(sk);
390		acked = tcp_slow_start(tp, acked);
391		if (!acked)
392			return;
393	}
394	bictcp_update(ca, tp->snd_cwnd, acked);
395	tcp_cong_avoid_ai(tp, ca->cnt, acked);
396}
397
398SEC("struct_ops")
399__u32 BPF_PROG(bpf_cubic_recalc_ssthresh, struct sock *sk)
400{
401	const struct tcp_sock *tp = tcp_sk(sk);
402	struct bpf_bictcp *ca = inet_csk_ca(sk);
403
404	ca->epoch_start = 0;	/* end of epoch */
405
406	/* Wmax and fast convergence */
407	if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
408		ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
409			/ (2 * BICTCP_BETA_SCALE);
410	else
411		ca->last_max_cwnd = tp->snd_cwnd;
412
413	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
414}
415
416SEC("struct_ops")
417void BPF_PROG(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 */
436static __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
447static void hystart_update(struct sock *sk, __u32 delay)
448{
449	struct tcp_sock *tp = tcp_sk(sk);
450	struct bpf_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
493int bpf_cubic_acked_called = 0;
494
495SEC("struct_ops")
496void BPF_PROG(bpf_cubic_acked, struct sock *sk, const struct ack_sample *sample)
497{
498	const struct tcp_sock *tp = tcp_sk(sk);
499	struct bpf_bictcp *ca = inet_csk_ca(sk);
500	__u32 delay;
501
502	bpf_cubic_acked_called = 1;
503	/* Some calls are for duplicates without timestamps */
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
525extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym;
526
527SEC("struct_ops")
528__u32 BPF_PROG(bpf_cubic_undo_cwnd, struct sock *sk)
529{
530	return tcp_reno_undo_cwnd(sk);
531}
532
533SEC(".struct_ops")
534struct tcp_congestion_ops cubic = {
535	.init		= (void *)bpf_cubic_init,
536	.ssthresh	= (void *)bpf_cubic_recalc_ssthresh,
537	.cong_avoid	= (void *)bpf_cubic_cong_avoid,
538	.set_state	= (void *)bpf_cubic_state,
539	.undo_cwnd	= (void *)bpf_cubic_undo_cwnd,
540	.cwnd_event	= (void *)bpf_cubic_cwnd_event,
541	.pkts_acked     = (void *)bpf_cubic_acked,
542	.name		= "bpf_cubic",
543};