xref: /openbmc/linux/net/sched/sch_pie.c (revision e3b9f1e8)
1 /* Copyright (C) 2013 Cisco Systems, Inc, 2013.
2  *
3  * This program is free software; you can redistribute it and/or
4  * modify it under the terms of the GNU General Public License
5  * as published by the Free Software Foundation; either version 2
6  * of the License.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
11  * GNU General Public License for more details.
12  *
13  * Author: Vijay Subramanian <vijaynsu@cisco.com>
14  * Author: Mythili Prabhu <mysuryan@cisco.com>
15  *
16  * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
17  * University of Oslo, Norway.
18  *
19  * References:
20  * IETF draft submission: http://tools.ietf.org/html/draft-pan-aqm-pie-00
21  * IEEE  Conference on High Performance Switching and Routing 2013 :
22  * "PIE: A * Lightweight Control Scheme to Address the Bufferbloat Problem"
23  */
24 
25 #include <linux/module.h>
26 #include <linux/slab.h>
27 #include <linux/types.h>
28 #include <linux/kernel.h>
29 #include <linux/errno.h>
30 #include <linux/skbuff.h>
31 #include <net/pkt_sched.h>
32 #include <net/inet_ecn.h>
33 
34 #define QUEUE_THRESHOLD 10000
35 #define DQCOUNT_INVALID -1
36 #define MAX_PROB  0xffffffff
37 #define PIE_SCALE 8
38 
39 /* parameters used */
40 struct pie_params {
41 	psched_time_t target;	/* user specified target delay in pschedtime */
42 	u32 tupdate;		/* timer frequency (in jiffies) */
43 	u32 limit;		/* number of packets that can be enqueued */
44 	u32 alpha;		/* alpha and beta are between 0 and 32 */
45 	u32 beta;		/* and are used for shift relative to 1 */
46 	bool ecn;		/* true if ecn is enabled */
47 	bool bytemode;		/* to scale drop early prob based on pkt size */
48 };
49 
50 /* variables used */
51 struct pie_vars {
52 	u32 prob;		/* probability but scaled by u32 limit. */
53 	psched_time_t burst_time;
54 	psched_time_t qdelay;
55 	psched_time_t qdelay_old;
56 	u64 dq_count;		/* measured in bytes */
57 	psched_time_t dq_tstamp;	/* drain rate */
58 	u32 avg_dq_rate;	/* bytes per pschedtime tick,scaled */
59 	u32 qlen_old;		/* in bytes */
60 };
61 
62 /* statistics gathering */
63 struct pie_stats {
64 	u32 packets_in;		/* total number of packets enqueued */
65 	u32 dropped;		/* packets dropped due to pie_action */
66 	u32 overlimit;		/* dropped due to lack of space in queue */
67 	u32 maxq;		/* maximum queue size */
68 	u32 ecn_mark;		/* packets marked with ECN */
69 };
70 
71 /* private data for the Qdisc */
72 struct pie_sched_data {
73 	struct pie_params params;
74 	struct pie_vars vars;
75 	struct pie_stats stats;
76 	struct timer_list adapt_timer;
77 	struct Qdisc *sch;
78 };
79 
80 static void pie_params_init(struct pie_params *params)
81 {
82 	params->alpha = 2;
83 	params->beta = 20;
84 	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
85 	params->limit = 1000;	/* default of 1000 packets */
86 	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);	/* 20 ms */
87 	params->ecn = false;
88 	params->bytemode = false;
89 }
90 
91 static void pie_vars_init(struct pie_vars *vars)
92 {
93 	vars->dq_count = DQCOUNT_INVALID;
94 	vars->avg_dq_rate = 0;
95 	/* default of 100 ms in pschedtime */
96 	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
97 }
98 
99 static bool drop_early(struct Qdisc *sch, u32 packet_size)
100 {
101 	struct pie_sched_data *q = qdisc_priv(sch);
102 	u32 rnd;
103 	u32 local_prob = q->vars.prob;
104 	u32 mtu = psched_mtu(qdisc_dev(sch));
105 
106 	/* If there is still burst allowance left skip random early drop */
107 	if (q->vars.burst_time > 0)
108 		return false;
109 
110 	/* If current delay is less than half of target, and
111 	 * if drop prob is low already, disable early_drop
112 	 */
113 	if ((q->vars.qdelay < q->params.target / 2)
114 	    && (q->vars.prob < MAX_PROB / 5))
115 		return false;
116 
117 	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
118 	 * similar to min_th in RED
119 	 */
120 	if (sch->qstats.backlog < 2 * mtu)
121 		return false;
122 
123 	/* If bytemode is turned on, use packet size to compute new
124 	 * probablity. Smaller packets will have lower drop prob in this case
125 	 */
126 	if (q->params.bytemode && packet_size <= mtu)
127 		local_prob = (local_prob / mtu) * packet_size;
128 	else
129 		local_prob = q->vars.prob;
130 
131 	rnd = prandom_u32();
132 	if (rnd < local_prob)
133 		return true;
134 
135 	return false;
136 }
137 
138 static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
139 			     struct sk_buff **to_free)
140 {
141 	struct pie_sched_data *q = qdisc_priv(sch);
142 	bool enqueue = false;
143 
144 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
145 		q->stats.overlimit++;
146 		goto out;
147 	}
148 
149 	if (!drop_early(sch, skb->len)) {
150 		enqueue = true;
151 	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
152 		   INET_ECN_set_ce(skb)) {
153 		/* If packet is ecn capable, mark it if drop probability
154 		 * is lower than 10%, else drop it.
155 		 */
156 		q->stats.ecn_mark++;
157 		enqueue = true;
158 	}
159 
160 	/* we can enqueue the packet */
161 	if (enqueue) {
162 		q->stats.packets_in++;
163 		if (qdisc_qlen(sch) > q->stats.maxq)
164 			q->stats.maxq = qdisc_qlen(sch);
165 
166 		return qdisc_enqueue_tail(skb, sch);
167 	}
168 
169 out:
170 	q->stats.dropped++;
171 	return qdisc_drop(skb, sch, to_free);
172 }
173 
174 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
175 	[TCA_PIE_TARGET] = {.type = NLA_U32},
176 	[TCA_PIE_LIMIT] = {.type = NLA_U32},
177 	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
178 	[TCA_PIE_ALPHA] = {.type = NLA_U32},
179 	[TCA_PIE_BETA] = {.type = NLA_U32},
180 	[TCA_PIE_ECN] = {.type = NLA_U32},
181 	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
182 };
183 
184 static int pie_change(struct Qdisc *sch, struct nlattr *opt,
185 		      struct netlink_ext_ack *extack)
186 {
187 	struct pie_sched_data *q = qdisc_priv(sch);
188 	struct nlattr *tb[TCA_PIE_MAX + 1];
189 	unsigned int qlen, dropped = 0;
190 	int err;
191 
192 	if (!opt)
193 		return -EINVAL;
194 
195 	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy, NULL);
196 	if (err < 0)
197 		return err;
198 
199 	sch_tree_lock(sch);
200 
201 	/* convert from microseconds to pschedtime */
202 	if (tb[TCA_PIE_TARGET]) {
203 		/* target is in us */
204 		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
205 
206 		/* convert to pschedtime */
207 		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
208 	}
209 
210 	/* tupdate is in jiffies */
211 	if (tb[TCA_PIE_TUPDATE])
212 		q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
213 
214 	if (tb[TCA_PIE_LIMIT]) {
215 		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
216 
217 		q->params.limit = limit;
218 		sch->limit = limit;
219 	}
220 
221 	if (tb[TCA_PIE_ALPHA])
222 		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
223 
224 	if (tb[TCA_PIE_BETA])
225 		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
226 
227 	if (tb[TCA_PIE_ECN])
228 		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
229 
230 	if (tb[TCA_PIE_BYTEMODE])
231 		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
232 
233 	/* Drop excess packets if new limit is lower */
234 	qlen = sch->q.qlen;
235 	while (sch->q.qlen > sch->limit) {
236 		struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
237 
238 		dropped += qdisc_pkt_len(skb);
239 		qdisc_qstats_backlog_dec(sch, skb);
240 		rtnl_qdisc_drop(skb, sch);
241 	}
242 	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
243 
244 	sch_tree_unlock(sch);
245 	return 0;
246 }
247 
248 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
249 {
250 
251 	struct pie_sched_data *q = qdisc_priv(sch);
252 	int qlen = sch->qstats.backlog;	/* current queue size in bytes */
253 
254 	/* If current queue is about 10 packets or more and dq_count is unset
255 	 * we have enough packets to calculate the drain rate. Save
256 	 * current time as dq_tstamp and start measurement cycle.
257 	 */
258 	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
259 		q->vars.dq_tstamp = psched_get_time();
260 		q->vars.dq_count = 0;
261 	}
262 
263 	/* Calculate the average drain rate from this value.  If queue length
264 	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
265 	 * the dq_count to -1 as we don't have enough packets to calculate the
266 	 * drain rate anymore The following if block is entered only when we
267 	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
268 	 * and we calculate the drain rate for the threshold here.  dq_count is
269 	 * in bytes, time difference in psched_time, hence rate is in
270 	 * bytes/psched_time.
271 	 */
272 	if (q->vars.dq_count != DQCOUNT_INVALID) {
273 		q->vars.dq_count += skb->len;
274 
275 		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
276 			psched_time_t now = psched_get_time();
277 			u32 dtime = now - q->vars.dq_tstamp;
278 			u32 count = q->vars.dq_count << PIE_SCALE;
279 
280 			if (dtime == 0)
281 				return;
282 
283 			count = count / dtime;
284 
285 			if (q->vars.avg_dq_rate == 0)
286 				q->vars.avg_dq_rate = count;
287 			else
288 				q->vars.avg_dq_rate =
289 				    (q->vars.avg_dq_rate -
290 				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);
291 
292 			/* If the queue has receded below the threshold, we hold
293 			 * on to the last drain rate calculated, else we reset
294 			 * dq_count to 0 to re-enter the if block when the next
295 			 * packet is dequeued
296 			 */
297 			if (qlen < QUEUE_THRESHOLD)
298 				q->vars.dq_count = DQCOUNT_INVALID;
299 			else {
300 				q->vars.dq_count = 0;
301 				q->vars.dq_tstamp = psched_get_time();
302 			}
303 
304 			if (q->vars.burst_time > 0) {
305 				if (q->vars.burst_time > dtime)
306 					q->vars.burst_time -= dtime;
307 				else
308 					q->vars.burst_time = 0;
309 			}
310 		}
311 	}
312 }
313 
314 static void calculate_probability(struct Qdisc *sch)
315 {
316 	struct pie_sched_data *q = qdisc_priv(sch);
317 	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
318 	psched_time_t qdelay = 0;	/* in pschedtime */
319 	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
320 	s32 delta = 0;		/* determines the change in probability */
321 	u32 oldprob;
322 	u32 alpha, beta;
323 	bool update_prob = true;
324 
325 	q->vars.qdelay_old = q->vars.qdelay;
326 
327 	if (q->vars.avg_dq_rate > 0)
328 		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
329 	else
330 		qdelay = 0;
331 
332 	/* If qdelay is zero and qlen is not, it means qlen is very small, less
333 	 * than dequeue_rate, so we do not update probabilty in this round
334 	 */
335 	if (qdelay == 0 && qlen != 0)
336 		update_prob = false;
337 
338 	/* In the algorithm, alpha and beta are between 0 and 2 with typical
339 	 * value for alpha as 0.125. In this implementation, we use values 0-32
340 	 * passed from user space to represent this. Also, alpha and beta have
341 	 * unit of HZ and need to be scaled before they can used to update
342 	 * probability. alpha/beta are updated locally below by 1) scaling them
343 	 * appropriately 2) scaling down by 16 to come to 0-2 range.
344 	 * Please see paper for details.
345 	 *
346 	 * We scale alpha and beta differently depending on whether we are in
347 	 * light, medium or high dropping mode.
348 	 */
349 	if (q->vars.prob < MAX_PROB / 100) {
350 		alpha =
351 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
352 		beta =
353 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
354 	} else if (q->vars.prob < MAX_PROB / 10) {
355 		alpha =
356 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
357 		beta =
358 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
359 	} else {
360 		alpha =
361 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
362 		beta =
363 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
364 	}
365 
366 	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
367 	delta += alpha * ((qdelay - q->params.target));
368 	delta += beta * ((qdelay - qdelay_old));
369 
370 	oldprob = q->vars.prob;
371 
372 	/* to ensure we increase probability in steps of no more than 2% */
373 	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
374 	    q->vars.prob >= MAX_PROB / 10)
375 		delta = (MAX_PROB / 100) * 2;
376 
377 	/* Non-linear drop:
378 	 * Tune drop probability to increase quickly for high delays(>= 250ms)
379 	 * 250ms is derived through experiments and provides error protection
380 	 */
381 
382 	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
383 		delta += MAX_PROB / (100 / 2);
384 
385 	q->vars.prob += delta;
386 
387 	if (delta > 0) {
388 		/* prevent overflow */
389 		if (q->vars.prob < oldprob) {
390 			q->vars.prob = MAX_PROB;
391 			/* Prevent normalization error. If probability is at
392 			 * maximum value already, we normalize it here, and
393 			 * skip the check to do a non-linear drop in the next
394 			 * section.
395 			 */
396 			update_prob = false;
397 		}
398 	} else {
399 		/* prevent underflow */
400 		if (q->vars.prob > oldprob)
401 			q->vars.prob = 0;
402 	}
403 
404 	/* Non-linear drop in probability: Reduce drop probability quickly if
405 	 * delay is 0 for 2 consecutive Tupdate periods.
406 	 */
407 
408 	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
409 		q->vars.prob = (q->vars.prob * 98) / 100;
410 
411 	q->vars.qdelay = qdelay;
412 	q->vars.qlen_old = qlen;
413 
414 	/* We restart the measurement cycle if the following conditions are met
415 	 * 1. If the delay has been low for 2 consecutive Tupdate periods
416 	 * 2. Calculated drop probability is zero
417 	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
418 	 *    is a non-zero value
419 	 */
420 	if ((q->vars.qdelay < q->params.target / 2) &&
421 	    (q->vars.qdelay_old < q->params.target / 2) &&
422 	    (q->vars.prob == 0) &&
423 	    (q->vars.avg_dq_rate > 0))
424 		pie_vars_init(&q->vars);
425 }
426 
427 static void pie_timer(struct timer_list *t)
428 {
429 	struct pie_sched_data *q = from_timer(q, t, adapt_timer);
430 	struct Qdisc *sch = q->sch;
431 	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
432 
433 	spin_lock(root_lock);
434 	calculate_probability(sch);
435 
436 	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
437 	if (q->params.tupdate)
438 		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
439 	spin_unlock(root_lock);
440 
441 }
442 
443 static int pie_init(struct Qdisc *sch, struct nlattr *opt,
444 		    struct netlink_ext_ack *extack)
445 {
446 	struct pie_sched_data *q = qdisc_priv(sch);
447 
448 	pie_params_init(&q->params);
449 	pie_vars_init(&q->vars);
450 	sch->limit = q->params.limit;
451 
452 	q->sch = sch;
453 	timer_setup(&q->adapt_timer, pie_timer, 0);
454 
455 	if (opt) {
456 		int err = pie_change(sch, opt, extack);
457 
458 		if (err)
459 			return err;
460 	}
461 
462 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
463 	return 0;
464 }
465 
466 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
467 {
468 	struct pie_sched_data *q = qdisc_priv(sch);
469 	struct nlattr *opts;
470 
471 	opts = nla_nest_start(skb, TCA_OPTIONS);
472 	if (opts == NULL)
473 		goto nla_put_failure;
474 
475 	/* convert target from pschedtime to us */
476 	if (nla_put_u32(skb, TCA_PIE_TARGET,
477 			((u32) PSCHED_TICKS2NS(q->params.target)) /
478 			NSEC_PER_USEC) ||
479 	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
480 	    nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
481 	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
482 	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
483 	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
484 	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
485 		goto nla_put_failure;
486 
487 	return nla_nest_end(skb, opts);
488 
489 nla_put_failure:
490 	nla_nest_cancel(skb, opts);
491 	return -1;
492 
493 }
494 
495 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
496 {
497 	struct pie_sched_data *q = qdisc_priv(sch);
498 	struct tc_pie_xstats st = {
499 		.prob		= q->vars.prob,
500 		.delay		= ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
501 				   NSEC_PER_USEC,
502 		/* unscale and return dq_rate in bytes per sec */
503 		.avg_dq_rate	= q->vars.avg_dq_rate *
504 				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
505 		.packets_in	= q->stats.packets_in,
506 		.overlimit	= q->stats.overlimit,
507 		.maxq		= q->stats.maxq,
508 		.dropped	= q->stats.dropped,
509 		.ecn_mark	= q->stats.ecn_mark,
510 	};
511 
512 	return gnet_stats_copy_app(d, &st, sizeof(st));
513 }
514 
515 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
516 {
517 	struct sk_buff *skb;
518 	skb = qdisc_dequeue_head(sch);
519 
520 	if (!skb)
521 		return NULL;
522 
523 	pie_process_dequeue(sch, skb);
524 	return skb;
525 }
526 
527 static void pie_reset(struct Qdisc *sch)
528 {
529 	struct pie_sched_data *q = qdisc_priv(sch);
530 	qdisc_reset_queue(sch);
531 	pie_vars_init(&q->vars);
532 }
533 
534 static void pie_destroy(struct Qdisc *sch)
535 {
536 	struct pie_sched_data *q = qdisc_priv(sch);
537 	q->params.tupdate = 0;
538 	del_timer_sync(&q->adapt_timer);
539 }
540 
541 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
542 	.id = "pie",
543 	.priv_size	= sizeof(struct pie_sched_data),
544 	.enqueue	= pie_qdisc_enqueue,
545 	.dequeue	= pie_qdisc_dequeue,
546 	.peek		= qdisc_peek_dequeued,
547 	.init		= pie_init,
548 	.destroy	= pie_destroy,
549 	.reset		= pie_reset,
550 	.change		= pie_change,
551 	.dump		= pie_dump,
552 	.dump_stats	= pie_dump_stats,
553 	.owner		= THIS_MODULE,
554 };
555 
556 static int __init pie_module_init(void)
557 {
558 	return register_qdisc(&pie_qdisc_ops);
559 }
560 
561 static void __exit pie_module_exit(void)
562 {
563 	unregister_qdisc(&pie_qdisc_ops);
564 }
565 
566 module_init(pie_module_init);
567 module_exit(pie_module_exit);
568 
569 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
570 MODULE_AUTHOR("Vijay Subramanian");
571 MODULE_AUTHOR("Mythili Prabhu");
572 MODULE_LICENSE("GPL");
573