xref: /openbmc/linux/net/sched/sch_pie.c (revision 5d0e4d78)
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 };
78 
79 static void pie_params_init(struct pie_params *params)
80 {
81 	params->alpha = 2;
82 	params->beta = 20;
83 	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
84 	params->limit = 1000;	/* default of 1000 packets */
85 	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);	/* 20 ms */
86 	params->ecn = false;
87 	params->bytemode = false;
88 }
89 
90 static void pie_vars_init(struct pie_vars *vars)
91 {
92 	vars->dq_count = DQCOUNT_INVALID;
93 	vars->avg_dq_rate = 0;
94 	/* default of 100 ms in pschedtime */
95 	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
96 }
97 
98 static bool drop_early(struct Qdisc *sch, u32 packet_size)
99 {
100 	struct pie_sched_data *q = qdisc_priv(sch);
101 	u32 rnd;
102 	u32 local_prob = q->vars.prob;
103 	u32 mtu = psched_mtu(qdisc_dev(sch));
104 
105 	/* If there is still burst allowance left skip random early drop */
106 	if (q->vars.burst_time > 0)
107 		return false;
108 
109 	/* If current delay is less than half of target, and
110 	 * if drop prob is low already, disable early_drop
111 	 */
112 	if ((q->vars.qdelay < q->params.target / 2)
113 	    && (q->vars.prob < MAX_PROB / 5))
114 		return false;
115 
116 	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
117 	 * similar to min_th in RED
118 	 */
119 	if (sch->qstats.backlog < 2 * mtu)
120 		return false;
121 
122 	/* If bytemode is turned on, use packet size to compute new
123 	 * probablity. Smaller packets will have lower drop prob in this case
124 	 */
125 	if (q->params.bytemode && packet_size <= mtu)
126 		local_prob = (local_prob / mtu) * packet_size;
127 	else
128 		local_prob = q->vars.prob;
129 
130 	rnd = prandom_u32();
131 	if (rnd < local_prob)
132 		return true;
133 
134 	return false;
135 }
136 
137 static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
138 			     struct sk_buff **to_free)
139 {
140 	struct pie_sched_data *q = qdisc_priv(sch);
141 	bool enqueue = false;
142 
143 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
144 		q->stats.overlimit++;
145 		goto out;
146 	}
147 
148 	if (!drop_early(sch, skb->len)) {
149 		enqueue = true;
150 	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
151 		   INET_ECN_set_ce(skb)) {
152 		/* If packet is ecn capable, mark it if drop probability
153 		 * is lower than 10%, else drop it.
154 		 */
155 		q->stats.ecn_mark++;
156 		enqueue = true;
157 	}
158 
159 	/* we can enqueue the packet */
160 	if (enqueue) {
161 		q->stats.packets_in++;
162 		if (qdisc_qlen(sch) > q->stats.maxq)
163 			q->stats.maxq = qdisc_qlen(sch);
164 
165 		return qdisc_enqueue_tail(skb, sch);
166 	}
167 
168 out:
169 	q->stats.dropped++;
170 	return qdisc_drop(skb, sch, to_free);
171 }
172 
173 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
174 	[TCA_PIE_TARGET] = {.type = NLA_U32},
175 	[TCA_PIE_LIMIT] = {.type = NLA_U32},
176 	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
177 	[TCA_PIE_ALPHA] = {.type = NLA_U32},
178 	[TCA_PIE_BETA] = {.type = NLA_U32},
179 	[TCA_PIE_ECN] = {.type = NLA_U32},
180 	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
181 };
182 
183 static int pie_change(struct Qdisc *sch, struct nlattr *opt)
184 {
185 	struct pie_sched_data *q = qdisc_priv(sch);
186 	struct nlattr *tb[TCA_PIE_MAX + 1];
187 	unsigned int qlen, dropped = 0;
188 	int err;
189 
190 	if (!opt)
191 		return -EINVAL;
192 
193 	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy, NULL);
194 	if (err < 0)
195 		return err;
196 
197 	sch_tree_lock(sch);
198 
199 	/* convert from microseconds to pschedtime */
200 	if (tb[TCA_PIE_TARGET]) {
201 		/* target is in us */
202 		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
203 
204 		/* convert to pschedtime */
205 		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
206 	}
207 
208 	/* tupdate is in jiffies */
209 	if (tb[TCA_PIE_TUPDATE])
210 		q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
211 
212 	if (tb[TCA_PIE_LIMIT]) {
213 		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
214 
215 		q->params.limit = limit;
216 		sch->limit = limit;
217 	}
218 
219 	if (tb[TCA_PIE_ALPHA])
220 		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
221 
222 	if (tb[TCA_PIE_BETA])
223 		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
224 
225 	if (tb[TCA_PIE_ECN])
226 		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
227 
228 	if (tb[TCA_PIE_BYTEMODE])
229 		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
230 
231 	/* Drop excess packets if new limit is lower */
232 	qlen = sch->q.qlen;
233 	while (sch->q.qlen > sch->limit) {
234 		struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
235 
236 		dropped += qdisc_pkt_len(skb);
237 		qdisc_qstats_backlog_dec(sch, skb);
238 		rtnl_qdisc_drop(skb, sch);
239 	}
240 	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
241 
242 	sch_tree_unlock(sch);
243 	return 0;
244 }
245 
246 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
247 {
248 
249 	struct pie_sched_data *q = qdisc_priv(sch);
250 	int qlen = sch->qstats.backlog;	/* current queue size in bytes */
251 
252 	/* If current queue is about 10 packets or more and dq_count is unset
253 	 * we have enough packets to calculate the drain rate. Save
254 	 * current time as dq_tstamp and start measurement cycle.
255 	 */
256 	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
257 		q->vars.dq_tstamp = psched_get_time();
258 		q->vars.dq_count = 0;
259 	}
260 
261 	/* Calculate the average drain rate from this value.  If queue length
262 	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
263 	 * the dq_count to -1 as we don't have enough packets to calculate the
264 	 * drain rate anymore The following if block is entered only when we
265 	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
266 	 * and we calculate the drain rate for the threshold here.  dq_count is
267 	 * in bytes, time difference in psched_time, hence rate is in
268 	 * bytes/psched_time.
269 	 */
270 	if (q->vars.dq_count != DQCOUNT_INVALID) {
271 		q->vars.dq_count += skb->len;
272 
273 		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
274 			psched_time_t now = psched_get_time();
275 			u32 dtime = now - q->vars.dq_tstamp;
276 			u32 count = q->vars.dq_count << PIE_SCALE;
277 
278 			if (dtime == 0)
279 				return;
280 
281 			count = count / dtime;
282 
283 			if (q->vars.avg_dq_rate == 0)
284 				q->vars.avg_dq_rate = count;
285 			else
286 				q->vars.avg_dq_rate =
287 				    (q->vars.avg_dq_rate -
288 				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);
289 
290 			/* If the queue has receded below the threshold, we hold
291 			 * on to the last drain rate calculated, else we reset
292 			 * dq_count to 0 to re-enter the if block when the next
293 			 * packet is dequeued
294 			 */
295 			if (qlen < QUEUE_THRESHOLD)
296 				q->vars.dq_count = DQCOUNT_INVALID;
297 			else {
298 				q->vars.dq_count = 0;
299 				q->vars.dq_tstamp = psched_get_time();
300 			}
301 
302 			if (q->vars.burst_time > 0) {
303 				if (q->vars.burst_time > dtime)
304 					q->vars.burst_time -= dtime;
305 				else
306 					q->vars.burst_time = 0;
307 			}
308 		}
309 	}
310 }
311 
312 static void calculate_probability(struct Qdisc *sch)
313 {
314 	struct pie_sched_data *q = qdisc_priv(sch);
315 	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
316 	psched_time_t qdelay = 0;	/* in pschedtime */
317 	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
318 	s32 delta = 0;		/* determines the change in probability */
319 	u32 oldprob;
320 	u32 alpha, beta;
321 	bool update_prob = true;
322 
323 	q->vars.qdelay_old = q->vars.qdelay;
324 
325 	if (q->vars.avg_dq_rate > 0)
326 		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
327 	else
328 		qdelay = 0;
329 
330 	/* If qdelay is zero and qlen is not, it means qlen is very small, less
331 	 * than dequeue_rate, so we do not update probabilty in this round
332 	 */
333 	if (qdelay == 0 && qlen != 0)
334 		update_prob = false;
335 
336 	/* In the algorithm, alpha and beta are between 0 and 2 with typical
337 	 * value for alpha as 0.125. In this implementation, we use values 0-32
338 	 * passed from user space to represent this. Also, alpha and beta have
339 	 * unit of HZ and need to be scaled before they can used to update
340 	 * probability. alpha/beta are updated locally below by 1) scaling them
341 	 * appropriately 2) scaling down by 16 to come to 0-2 range.
342 	 * Please see paper for details.
343 	 *
344 	 * We scale alpha and beta differently depending on whether we are in
345 	 * light, medium or high dropping mode.
346 	 */
347 	if (q->vars.prob < MAX_PROB / 100) {
348 		alpha =
349 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
350 		beta =
351 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
352 	} else if (q->vars.prob < MAX_PROB / 10) {
353 		alpha =
354 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
355 		beta =
356 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
357 	} else {
358 		alpha =
359 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
360 		beta =
361 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
362 	}
363 
364 	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
365 	delta += alpha * ((qdelay - q->params.target));
366 	delta += beta * ((qdelay - qdelay_old));
367 
368 	oldprob = q->vars.prob;
369 
370 	/* to ensure we increase probability in steps of no more than 2% */
371 	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
372 	    q->vars.prob >= MAX_PROB / 10)
373 		delta = (MAX_PROB / 100) * 2;
374 
375 	/* Non-linear drop:
376 	 * Tune drop probability to increase quickly for high delays(>= 250ms)
377 	 * 250ms is derived through experiments and provides error protection
378 	 */
379 
380 	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
381 		delta += MAX_PROB / (100 / 2);
382 
383 	q->vars.prob += delta;
384 
385 	if (delta > 0) {
386 		/* prevent overflow */
387 		if (q->vars.prob < oldprob) {
388 			q->vars.prob = MAX_PROB;
389 			/* Prevent normalization error. If probability is at
390 			 * maximum value already, we normalize it here, and
391 			 * skip the check to do a non-linear drop in the next
392 			 * section.
393 			 */
394 			update_prob = false;
395 		}
396 	} else {
397 		/* prevent underflow */
398 		if (q->vars.prob > oldprob)
399 			q->vars.prob = 0;
400 	}
401 
402 	/* Non-linear drop in probability: Reduce drop probability quickly if
403 	 * delay is 0 for 2 consecutive Tupdate periods.
404 	 */
405 
406 	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
407 		q->vars.prob = (q->vars.prob * 98) / 100;
408 
409 	q->vars.qdelay = qdelay;
410 	q->vars.qlen_old = qlen;
411 
412 	/* We restart the measurement cycle if the following conditions are met
413 	 * 1. If the delay has been low for 2 consecutive Tupdate periods
414 	 * 2. Calculated drop probability is zero
415 	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
416 	 *    is a non-zero value
417 	 */
418 	if ((q->vars.qdelay < q->params.target / 2) &&
419 	    (q->vars.qdelay_old < q->params.target / 2) &&
420 	    (q->vars.prob == 0) &&
421 	    (q->vars.avg_dq_rate > 0))
422 		pie_vars_init(&q->vars);
423 }
424 
425 static void pie_timer(unsigned long arg)
426 {
427 	struct Qdisc *sch = (struct Qdisc *)arg;
428 	struct pie_sched_data *q = qdisc_priv(sch);
429 	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
430 
431 	spin_lock(root_lock);
432 	calculate_probability(sch);
433 
434 	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
435 	if (q->params.tupdate)
436 		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
437 	spin_unlock(root_lock);
438 
439 }
440 
441 static int pie_init(struct Qdisc *sch, struct nlattr *opt)
442 {
443 	struct pie_sched_data *q = qdisc_priv(sch);
444 
445 	pie_params_init(&q->params);
446 	pie_vars_init(&q->vars);
447 	sch->limit = q->params.limit;
448 
449 	setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
450 
451 	if (opt) {
452 		int err = pie_change(sch, opt);
453 
454 		if (err)
455 			return err;
456 	}
457 
458 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
459 	return 0;
460 }
461 
462 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
463 {
464 	struct pie_sched_data *q = qdisc_priv(sch);
465 	struct nlattr *opts;
466 
467 	opts = nla_nest_start(skb, TCA_OPTIONS);
468 	if (opts == NULL)
469 		goto nla_put_failure;
470 
471 	/* convert target from pschedtime to us */
472 	if (nla_put_u32(skb, TCA_PIE_TARGET,
473 			((u32) PSCHED_TICKS2NS(q->params.target)) /
474 			NSEC_PER_USEC) ||
475 	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
476 	    nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
477 	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
478 	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
479 	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
480 	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
481 		goto nla_put_failure;
482 
483 	return nla_nest_end(skb, opts);
484 
485 nla_put_failure:
486 	nla_nest_cancel(skb, opts);
487 	return -1;
488 
489 }
490 
491 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
492 {
493 	struct pie_sched_data *q = qdisc_priv(sch);
494 	struct tc_pie_xstats st = {
495 		.prob		= q->vars.prob,
496 		.delay		= ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
497 				   NSEC_PER_USEC,
498 		/* unscale and return dq_rate in bytes per sec */
499 		.avg_dq_rate	= q->vars.avg_dq_rate *
500 				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
501 		.packets_in	= q->stats.packets_in,
502 		.overlimit	= q->stats.overlimit,
503 		.maxq		= q->stats.maxq,
504 		.dropped	= q->stats.dropped,
505 		.ecn_mark	= q->stats.ecn_mark,
506 	};
507 
508 	return gnet_stats_copy_app(d, &st, sizeof(st));
509 }
510 
511 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
512 {
513 	struct sk_buff *skb;
514 	skb = qdisc_dequeue_head(sch);
515 
516 	if (!skb)
517 		return NULL;
518 
519 	pie_process_dequeue(sch, skb);
520 	return skb;
521 }
522 
523 static void pie_reset(struct Qdisc *sch)
524 {
525 	struct pie_sched_data *q = qdisc_priv(sch);
526 	qdisc_reset_queue(sch);
527 	pie_vars_init(&q->vars);
528 }
529 
530 static void pie_destroy(struct Qdisc *sch)
531 {
532 	struct pie_sched_data *q = qdisc_priv(sch);
533 	q->params.tupdate = 0;
534 	del_timer_sync(&q->adapt_timer);
535 }
536 
537 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
538 	.id = "pie",
539 	.priv_size	= sizeof(struct pie_sched_data),
540 	.enqueue	= pie_qdisc_enqueue,
541 	.dequeue	= pie_qdisc_dequeue,
542 	.peek		= qdisc_peek_dequeued,
543 	.init		= pie_init,
544 	.destroy	= pie_destroy,
545 	.reset		= pie_reset,
546 	.change		= pie_change,
547 	.dump		= pie_dump,
548 	.dump_stats	= pie_dump_stats,
549 	.owner		= THIS_MODULE,
550 };
551 
552 static int __init pie_module_init(void)
553 {
554 	return register_qdisc(&pie_qdisc_ops);
555 }
556 
557 static void __exit pie_module_exit(void)
558 {
559 	unregister_qdisc(&pie_qdisc_ops);
560 }
561 
562 module_init(pie_module_init);
563 module_exit(pie_module_exit);
564 
565 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
566 MODULE_AUTHOR("Vijay Subramanian");
567 MODULE_AUTHOR("Mythili Prabhu");
568 MODULE_LICENSE("GPL");
569