xref: /openbmc/linux/net/sched/sch_pie.c (revision 5e012745)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2013 Cisco Systems, Inc, 2013.
3  *
4  * Author: Vijay Subramanian <vijaynsu@cisco.com>
5  * Author: Mythili Prabhu <mysuryan@cisco.com>
6  *
7  * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
8  * University of Oslo, Norway.
9  *
10  * References:
11  * RFC 8033: https://tools.ietf.org/html/rfc8033
12  */
13 
14 #include <linux/module.h>
15 #include <linux/slab.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/errno.h>
19 #include <linux/skbuff.h>
20 #include <net/pkt_sched.h>
21 #include <net/inet_ecn.h>
22 
23 #define QUEUE_THRESHOLD 16384
24 #define DQCOUNT_INVALID -1
25 #define MAX_PROB 0xffffffffffffffff
26 #define PIE_SCALE 8
27 
28 /* parameters used */
29 struct pie_params {
30 	psched_time_t target;	/* user specified target delay in pschedtime */
31 	u32 tupdate;		/* timer frequency (in jiffies) */
32 	u32 limit;		/* number of packets that can be enqueued */
33 	u32 alpha;		/* alpha and beta are between 0 and 32 */
34 	u32 beta;		/* and are used for shift relative to 1 */
35 	bool ecn;		/* true if ecn is enabled */
36 	bool bytemode;		/* to scale drop early prob based on pkt size */
37 };
38 
39 /* variables used */
40 struct pie_vars {
41 	u64 prob;		/* probability but scaled by u64 limit. */
42 	psched_time_t burst_time;
43 	psched_time_t qdelay;
44 	psched_time_t qdelay_old;
45 	u64 dq_count;		/* measured in bytes */
46 	psched_time_t dq_tstamp;	/* drain rate */
47 	u64 accu_prob;		/* accumulated drop probability */
48 	u32 avg_dq_rate;	/* bytes per pschedtime tick,scaled */
49 	u32 qlen_old;		/* in bytes */
50 	u8 accu_prob_overflows;	/* overflows of accu_prob */
51 };
52 
53 /* statistics gathering */
54 struct pie_stats {
55 	u32 packets_in;		/* total number of packets enqueued */
56 	u32 dropped;		/* packets dropped due to pie_action */
57 	u32 overlimit;		/* dropped due to lack of space in queue */
58 	u32 maxq;		/* maximum queue size */
59 	u32 ecn_mark;		/* packets marked with ECN */
60 };
61 
62 /* private data for the Qdisc */
63 struct pie_sched_data {
64 	struct pie_params params;
65 	struct pie_vars vars;
66 	struct pie_stats stats;
67 	struct timer_list adapt_timer;
68 	struct Qdisc *sch;
69 };
70 
71 static void pie_params_init(struct pie_params *params)
72 {
73 	params->alpha = 2;
74 	params->beta = 20;
75 	params->tupdate = usecs_to_jiffies(15 * USEC_PER_MSEC);	/* 15 ms */
76 	params->limit = 1000;	/* default of 1000 packets */
77 	params->target = PSCHED_NS2TICKS(15 * NSEC_PER_MSEC);	/* 15 ms */
78 	params->ecn = false;
79 	params->bytemode = false;
80 }
81 
82 static void pie_vars_init(struct pie_vars *vars)
83 {
84 	vars->dq_count = DQCOUNT_INVALID;
85 	vars->accu_prob = 0;
86 	vars->avg_dq_rate = 0;
87 	/* default of 150 ms in pschedtime */
88 	vars->burst_time = PSCHED_NS2TICKS(150 * NSEC_PER_MSEC);
89 	vars->accu_prob_overflows = 0;
90 }
91 
92 static bool drop_early(struct Qdisc *sch, u32 packet_size)
93 {
94 	struct pie_sched_data *q = qdisc_priv(sch);
95 	u64 rnd;
96 	u64 local_prob = q->vars.prob;
97 	u32 mtu = psched_mtu(qdisc_dev(sch));
98 
99 	/* If there is still burst allowance left skip random early drop */
100 	if (q->vars.burst_time > 0)
101 		return false;
102 
103 	/* If current delay is less than half of target, and
104 	 * if drop prob is low already, disable early_drop
105 	 */
106 	if ((q->vars.qdelay < q->params.target / 2) &&
107 	    (q->vars.prob < MAX_PROB / 5))
108 		return false;
109 
110 	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
111 	 * similar to min_th in RED
112 	 */
113 	if (sch->qstats.backlog < 2 * mtu)
114 		return false;
115 
116 	/* If bytemode is turned on, use packet size to compute new
117 	 * probablity. Smaller packets will have lower drop prob in this case
118 	 */
119 	if (q->params.bytemode && packet_size <= mtu)
120 		local_prob = (u64)packet_size * div_u64(local_prob, mtu);
121 	else
122 		local_prob = q->vars.prob;
123 
124 	if (local_prob == 0) {
125 		q->vars.accu_prob = 0;
126 		q->vars.accu_prob_overflows = 0;
127 	}
128 
129 	if (local_prob > MAX_PROB - q->vars.accu_prob)
130 		q->vars.accu_prob_overflows++;
131 
132 	q->vars.accu_prob += local_prob;
133 
134 	if (q->vars.accu_prob_overflows == 0 &&
135 	    q->vars.accu_prob < (MAX_PROB / 100) * 85)
136 		return false;
137 	if (q->vars.accu_prob_overflows == 8 &&
138 	    q->vars.accu_prob >= MAX_PROB / 2)
139 		return true;
140 
141 	prandom_bytes(&rnd, 8);
142 	if (rnd < local_prob) {
143 		q->vars.accu_prob = 0;
144 		q->vars.accu_prob_overflows = 0;
145 		return true;
146 	}
147 
148 	return false;
149 }
150 
151 static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
152 			     struct sk_buff **to_free)
153 {
154 	struct pie_sched_data *q = qdisc_priv(sch);
155 	bool enqueue = false;
156 
157 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
158 		q->stats.overlimit++;
159 		goto out;
160 	}
161 
162 	if (!drop_early(sch, skb->len)) {
163 		enqueue = true;
164 	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
165 		   INET_ECN_set_ce(skb)) {
166 		/* If packet is ecn capable, mark it if drop probability
167 		 * is lower than 10%, else drop it.
168 		 */
169 		q->stats.ecn_mark++;
170 		enqueue = true;
171 	}
172 
173 	/* we can enqueue the packet */
174 	if (enqueue) {
175 		q->stats.packets_in++;
176 		if (qdisc_qlen(sch) > q->stats.maxq)
177 			q->stats.maxq = qdisc_qlen(sch);
178 
179 		return qdisc_enqueue_tail(skb, sch);
180 	}
181 
182 out:
183 	q->stats.dropped++;
184 	q->vars.accu_prob = 0;
185 	q->vars.accu_prob_overflows = 0;
186 	return qdisc_drop(skb, sch, to_free);
187 }
188 
189 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
190 	[TCA_PIE_TARGET] = {.type = NLA_U32},
191 	[TCA_PIE_LIMIT] = {.type = NLA_U32},
192 	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
193 	[TCA_PIE_ALPHA] = {.type = NLA_U32},
194 	[TCA_PIE_BETA] = {.type = NLA_U32},
195 	[TCA_PIE_ECN] = {.type = NLA_U32},
196 	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
197 };
198 
199 static int pie_change(struct Qdisc *sch, struct nlattr *opt,
200 		      struct netlink_ext_ack *extack)
201 {
202 	struct pie_sched_data *q = qdisc_priv(sch);
203 	struct nlattr *tb[TCA_PIE_MAX + 1];
204 	unsigned int qlen, dropped = 0;
205 	int err;
206 
207 	if (!opt)
208 		return -EINVAL;
209 
210 	err = nla_parse_nested_deprecated(tb, TCA_PIE_MAX, opt, pie_policy,
211 					  NULL);
212 	if (err < 0)
213 		return err;
214 
215 	sch_tree_lock(sch);
216 
217 	/* convert from microseconds to pschedtime */
218 	if (tb[TCA_PIE_TARGET]) {
219 		/* target is in us */
220 		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
221 
222 		/* convert to pschedtime */
223 		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
224 	}
225 
226 	/* tupdate is in jiffies */
227 	if (tb[TCA_PIE_TUPDATE])
228 		q->params.tupdate =
229 			usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
230 
231 	if (tb[TCA_PIE_LIMIT]) {
232 		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
233 
234 		q->params.limit = limit;
235 		sch->limit = limit;
236 	}
237 
238 	if (tb[TCA_PIE_ALPHA])
239 		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
240 
241 	if (tb[TCA_PIE_BETA])
242 		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
243 
244 	if (tb[TCA_PIE_ECN])
245 		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
246 
247 	if (tb[TCA_PIE_BYTEMODE])
248 		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
249 
250 	/* Drop excess packets if new limit is lower */
251 	qlen = sch->q.qlen;
252 	while (sch->q.qlen > sch->limit) {
253 		struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
254 
255 		dropped += qdisc_pkt_len(skb);
256 		qdisc_qstats_backlog_dec(sch, skb);
257 		rtnl_qdisc_drop(skb, sch);
258 	}
259 	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
260 
261 	sch_tree_unlock(sch);
262 	return 0;
263 }
264 
265 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
266 {
267 	struct pie_sched_data *q = qdisc_priv(sch);
268 	int qlen = sch->qstats.backlog;	/* current queue size in bytes */
269 
270 	/* If current queue is about 10 packets or more and dq_count is unset
271 	 * we have enough packets to calculate the drain rate. Save
272 	 * current time as dq_tstamp and start measurement cycle.
273 	 */
274 	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
275 		q->vars.dq_tstamp = psched_get_time();
276 		q->vars.dq_count = 0;
277 	}
278 
279 	/* Calculate the average drain rate from this value.  If queue length
280 	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
281 	 * the dq_count to -1 as we don't have enough packets to calculate the
282 	 * drain rate anymore The following if block is entered only when we
283 	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
284 	 * and we calculate the drain rate for the threshold here.  dq_count is
285 	 * in bytes, time difference in psched_time, hence rate is in
286 	 * bytes/psched_time.
287 	 */
288 	if (q->vars.dq_count != DQCOUNT_INVALID) {
289 		q->vars.dq_count += skb->len;
290 
291 		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
292 			psched_time_t now = psched_get_time();
293 			u32 dtime = now - q->vars.dq_tstamp;
294 			u32 count = q->vars.dq_count << PIE_SCALE;
295 
296 			if (dtime == 0)
297 				return;
298 
299 			count = count / dtime;
300 
301 			if (q->vars.avg_dq_rate == 0)
302 				q->vars.avg_dq_rate = count;
303 			else
304 				q->vars.avg_dq_rate =
305 				    (q->vars.avg_dq_rate -
306 				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);
307 
308 			/* If the queue has receded below the threshold, we hold
309 			 * on to the last drain rate calculated, else we reset
310 			 * dq_count to 0 to re-enter the if block when the next
311 			 * packet is dequeued
312 			 */
313 			if (qlen < QUEUE_THRESHOLD) {
314 				q->vars.dq_count = DQCOUNT_INVALID;
315 			} else {
316 				q->vars.dq_count = 0;
317 				q->vars.dq_tstamp = psched_get_time();
318 			}
319 
320 			if (q->vars.burst_time > 0) {
321 				if (q->vars.burst_time > dtime)
322 					q->vars.burst_time -= dtime;
323 				else
324 					q->vars.burst_time = 0;
325 			}
326 		}
327 	}
328 }
329 
330 static void calculate_probability(struct Qdisc *sch)
331 {
332 	struct pie_sched_data *q = qdisc_priv(sch);
333 	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
334 	psched_time_t qdelay = 0;	/* in pschedtime */
335 	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
336 	s64 delta = 0;		/* determines the change in probability */
337 	u64 oldprob;
338 	u64 alpha, beta;
339 	u32 power;
340 	bool update_prob = true;
341 
342 	q->vars.qdelay_old = q->vars.qdelay;
343 
344 	if (q->vars.avg_dq_rate > 0)
345 		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
346 	else
347 		qdelay = 0;
348 
349 	/* If qdelay is zero and qlen is not, it means qlen is very small, less
350 	 * than dequeue_rate, so we do not update probabilty in this round
351 	 */
352 	if (qdelay == 0 && qlen != 0)
353 		update_prob = false;
354 
355 	/* In the algorithm, alpha and beta are between 0 and 2 with typical
356 	 * value for alpha as 0.125. In this implementation, we use values 0-32
357 	 * passed from user space to represent this. Also, alpha and beta have
358 	 * unit of HZ and need to be scaled before they can used to update
359 	 * probability. alpha/beta are updated locally below by scaling down
360 	 * by 16 to come to 0-2 range.
361 	 */
362 	alpha = ((u64)q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
363 	beta = ((u64)q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
364 
365 	/* We scale alpha and beta differently depending on how heavy the
366 	 * congestion is. Please see RFC 8033 for details.
367 	 */
368 	if (q->vars.prob < MAX_PROB / 10) {
369 		alpha >>= 1;
370 		beta >>= 1;
371 
372 		power = 100;
373 		while (q->vars.prob < div_u64(MAX_PROB, power) &&
374 		       power <= 1000000) {
375 			alpha >>= 2;
376 			beta >>= 2;
377 			power *= 10;
378 		}
379 	}
380 
381 	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
382 	delta += alpha * (u64)(qdelay - q->params.target);
383 	delta += beta * (u64)(qdelay - qdelay_old);
384 
385 	oldprob = q->vars.prob;
386 
387 	/* to ensure we increase probability in steps of no more than 2% */
388 	if (delta > (s64)(MAX_PROB / (100 / 2)) &&
389 	    q->vars.prob >= MAX_PROB / 10)
390 		delta = (MAX_PROB / 100) * 2;
391 
392 	/* Non-linear drop:
393 	 * Tune drop probability to increase quickly for high delays(>= 250ms)
394 	 * 250ms is derived through experiments and provides error protection
395 	 */
396 
397 	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
398 		delta += MAX_PROB / (100 / 2);
399 
400 	q->vars.prob += delta;
401 
402 	if (delta > 0) {
403 		/* prevent overflow */
404 		if (q->vars.prob < oldprob) {
405 			q->vars.prob = MAX_PROB;
406 			/* Prevent normalization error. If probability is at
407 			 * maximum value already, we normalize it here, and
408 			 * skip the check to do a non-linear drop in the next
409 			 * section.
410 			 */
411 			update_prob = false;
412 		}
413 	} else {
414 		/* prevent underflow */
415 		if (q->vars.prob > oldprob)
416 			q->vars.prob = 0;
417 	}
418 
419 	/* Non-linear drop in probability: Reduce drop probability quickly if
420 	 * delay is 0 for 2 consecutive Tupdate periods.
421 	 */
422 
423 	if (qdelay == 0 && qdelay_old == 0 && update_prob)
424 		/* Reduce drop probability to 98.4% */
425 		q->vars.prob -= q->vars.prob / 64u;
426 
427 	q->vars.qdelay = qdelay;
428 	q->vars.qlen_old = qlen;
429 
430 	/* We restart the measurement cycle if the following conditions are met
431 	 * 1. If the delay has been low for 2 consecutive Tupdate periods
432 	 * 2. Calculated drop probability is zero
433 	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
434 	 *    is a non-zero value
435 	 */
436 	if ((q->vars.qdelay < q->params.target / 2) &&
437 	    (q->vars.qdelay_old < q->params.target / 2) &&
438 	    q->vars.prob == 0 &&
439 	    q->vars.avg_dq_rate > 0)
440 		pie_vars_init(&q->vars);
441 }
442 
443 static void pie_timer(struct timer_list *t)
444 {
445 	struct pie_sched_data *q = from_timer(q, t, adapt_timer);
446 	struct Qdisc *sch = q->sch;
447 	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
448 
449 	spin_lock(root_lock);
450 	calculate_probability(sch);
451 
452 	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
453 	if (q->params.tupdate)
454 		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
455 	spin_unlock(root_lock);
456 }
457 
458 static int pie_init(struct Qdisc *sch, struct nlattr *opt,
459 		    struct netlink_ext_ack *extack)
460 {
461 	struct pie_sched_data *q = qdisc_priv(sch);
462 
463 	pie_params_init(&q->params);
464 	pie_vars_init(&q->vars);
465 	sch->limit = q->params.limit;
466 
467 	q->sch = sch;
468 	timer_setup(&q->adapt_timer, pie_timer, 0);
469 
470 	if (opt) {
471 		int err = pie_change(sch, opt, extack);
472 
473 		if (err)
474 			return err;
475 	}
476 
477 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
478 	return 0;
479 }
480 
481 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
482 {
483 	struct pie_sched_data *q = qdisc_priv(sch);
484 	struct nlattr *opts;
485 
486 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
487 	if (!opts)
488 		goto nla_put_failure;
489 
490 	/* convert target from pschedtime to us */
491 	if (nla_put_u32(skb, TCA_PIE_TARGET,
492 			((u32)PSCHED_TICKS2NS(q->params.target)) /
493 			NSEC_PER_USEC) ||
494 	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
495 	    nla_put_u32(skb, TCA_PIE_TUPDATE,
496 			jiffies_to_usecs(q->params.tupdate)) ||
497 	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
498 	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
499 	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
500 	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
501 		goto nla_put_failure;
502 
503 	return nla_nest_end(skb, opts);
504 
505 nla_put_failure:
506 	nla_nest_cancel(skb, opts);
507 	return -1;
508 }
509 
510 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
511 {
512 	struct pie_sched_data *q = qdisc_priv(sch);
513 	struct tc_pie_xstats st = {
514 		.prob		= q->vars.prob,
515 		.delay		= ((u32)PSCHED_TICKS2NS(q->vars.qdelay)) /
516 				   NSEC_PER_USEC,
517 		/* unscale and return dq_rate in bytes per sec */
518 		.avg_dq_rate	= q->vars.avg_dq_rate *
519 				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
520 		.packets_in	= q->stats.packets_in,
521 		.overlimit	= q->stats.overlimit,
522 		.maxq		= q->stats.maxq,
523 		.dropped	= q->stats.dropped,
524 		.ecn_mark	= q->stats.ecn_mark,
525 	};
526 
527 	return gnet_stats_copy_app(d, &st, sizeof(st));
528 }
529 
530 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
531 {
532 	struct sk_buff *skb = qdisc_dequeue_head(sch);
533 
534 	if (!skb)
535 		return NULL;
536 
537 	pie_process_dequeue(sch, skb);
538 	return skb;
539 }
540 
541 static void pie_reset(struct Qdisc *sch)
542 {
543 	struct pie_sched_data *q = qdisc_priv(sch);
544 
545 	qdisc_reset_queue(sch);
546 	pie_vars_init(&q->vars);
547 }
548 
549 static void pie_destroy(struct Qdisc *sch)
550 {
551 	struct pie_sched_data *q = qdisc_priv(sch);
552 
553 	q->params.tupdate = 0;
554 	del_timer_sync(&q->adapt_timer);
555 }
556 
557 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
558 	.id = "pie",
559 	.priv_size	= sizeof(struct pie_sched_data),
560 	.enqueue	= pie_qdisc_enqueue,
561 	.dequeue	= pie_qdisc_dequeue,
562 	.peek		= qdisc_peek_dequeued,
563 	.init		= pie_init,
564 	.destroy	= pie_destroy,
565 	.reset		= pie_reset,
566 	.change		= pie_change,
567 	.dump		= pie_dump,
568 	.dump_stats	= pie_dump_stats,
569 	.owner		= THIS_MODULE,
570 };
571 
572 static int __init pie_module_init(void)
573 {
574 	return register_qdisc(&pie_qdisc_ops);
575 }
576 
577 static void __exit pie_module_exit(void)
578 {
579 	unregister_qdisc(&pie_qdisc_ops);
580 }
581 
582 module_init(pie_module_init);
583 module_exit(pie_module_exit);
584 
585 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
586 MODULE_AUTHOR("Vijay Subramanian");
587 MODULE_AUTHOR("Mythili Prabhu");
588 MODULE_LICENSE("GPL");
589