xref: /openbmc/linux/net/sched/sch_pie.c (revision d0b5e15f)
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 {
139 	struct pie_sched_data *q = qdisc_priv(sch);
140 	bool enqueue = false;
141 
142 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
143 		q->stats.overlimit++;
144 		goto out;
145 	}
146 
147 	if (!drop_early(sch, skb->len)) {
148 		enqueue = true;
149 	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
150 		   INET_ECN_set_ce(skb)) {
151 		/* If packet is ecn capable, mark it if drop probability
152 		 * is lower than 10%, else drop it.
153 		 */
154 		q->stats.ecn_mark++;
155 		enqueue = true;
156 	}
157 
158 	/* we can enqueue the packet */
159 	if (enqueue) {
160 		q->stats.packets_in++;
161 		if (qdisc_qlen(sch) > q->stats.maxq)
162 			q->stats.maxq = qdisc_qlen(sch);
163 
164 		return qdisc_enqueue_tail(skb, sch);
165 	}
166 
167 out:
168 	q->stats.dropped++;
169 	return qdisc_drop(skb, sch);
170 }
171 
172 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
173 	[TCA_PIE_TARGET] = {.type = NLA_U32},
174 	[TCA_PIE_LIMIT] = {.type = NLA_U32},
175 	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
176 	[TCA_PIE_ALPHA] = {.type = NLA_U32},
177 	[TCA_PIE_BETA] = {.type = NLA_U32},
178 	[TCA_PIE_ECN] = {.type = NLA_U32},
179 	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
180 };
181 
182 static int pie_change(struct Qdisc *sch, struct nlattr *opt)
183 {
184 	struct pie_sched_data *q = qdisc_priv(sch);
185 	struct nlattr *tb[TCA_PIE_MAX + 1];
186 	unsigned int qlen;
187 	int err;
188 
189 	if (!opt)
190 		return -EINVAL;
191 
192 	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
193 	if (err < 0)
194 		return err;
195 
196 	sch_tree_lock(sch);
197 
198 	/* convert from microseconds to pschedtime */
199 	if (tb[TCA_PIE_TARGET]) {
200 		/* target is in us */
201 		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
202 
203 		/* convert to pschedtime */
204 		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
205 	}
206 
207 	/* tupdate is in jiffies */
208 	if (tb[TCA_PIE_TUPDATE])
209 		q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
210 
211 	if (tb[TCA_PIE_LIMIT]) {
212 		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
213 
214 		q->params.limit = limit;
215 		sch->limit = limit;
216 	}
217 
218 	if (tb[TCA_PIE_ALPHA])
219 		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
220 
221 	if (tb[TCA_PIE_BETA])
222 		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
223 
224 	if (tb[TCA_PIE_ECN])
225 		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
226 
227 	if (tb[TCA_PIE_BYTEMODE])
228 		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
229 
230 	/* Drop excess packets if new limit is lower */
231 	qlen = sch->q.qlen;
232 	while (sch->q.qlen > sch->limit) {
233 		struct sk_buff *skb = __skb_dequeue(&sch->q);
234 
235 		qdisc_qstats_backlog_dec(sch, skb);
236 		qdisc_drop(skb, sch);
237 	}
238 	qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
239 
240 	sch_tree_unlock(sch);
241 	return 0;
242 }
243 
244 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
245 {
246 
247 	struct pie_sched_data *q = qdisc_priv(sch);
248 	int qlen = sch->qstats.backlog;	/* current queue size in bytes */
249 
250 	/* If current queue is about 10 packets or more and dq_count is unset
251 	 * we have enough packets to calculate the drain rate. Save
252 	 * current time as dq_tstamp and start measurement cycle.
253 	 */
254 	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
255 		q->vars.dq_tstamp = psched_get_time();
256 		q->vars.dq_count = 0;
257 	}
258 
259 	/* Calculate the average drain rate from this value.  If queue length
260 	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
261 	 * the dq_count to -1 as we don't have enough packets to calculate the
262 	 * drain rate anymore The following if block is entered only when we
263 	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
264 	 * and we calculate the drain rate for the threshold here.  dq_count is
265 	 * in bytes, time difference in psched_time, hence rate is in
266 	 * bytes/psched_time.
267 	 */
268 	if (q->vars.dq_count != DQCOUNT_INVALID) {
269 		q->vars.dq_count += skb->len;
270 
271 		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
272 			psched_time_t now = psched_get_time();
273 			u32 dtime = now - q->vars.dq_tstamp;
274 			u32 count = q->vars.dq_count << PIE_SCALE;
275 
276 			if (dtime == 0)
277 				return;
278 
279 			count = count / dtime;
280 
281 			if (q->vars.avg_dq_rate == 0)
282 				q->vars.avg_dq_rate = count;
283 			else
284 				q->vars.avg_dq_rate =
285 				    (q->vars.avg_dq_rate -
286 				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);
287 
288 			/* If the queue has receded below the threshold, we hold
289 			 * on to the last drain rate calculated, else we reset
290 			 * dq_count to 0 to re-enter the if block when the next
291 			 * packet is dequeued
292 			 */
293 			if (qlen < QUEUE_THRESHOLD)
294 				q->vars.dq_count = DQCOUNT_INVALID;
295 			else {
296 				q->vars.dq_count = 0;
297 				q->vars.dq_tstamp = psched_get_time();
298 			}
299 
300 			if (q->vars.burst_time > 0) {
301 				if (q->vars.burst_time > dtime)
302 					q->vars.burst_time -= dtime;
303 				else
304 					q->vars.burst_time = 0;
305 			}
306 		}
307 	}
308 }
309 
310 static void calculate_probability(struct Qdisc *sch)
311 {
312 	struct pie_sched_data *q = qdisc_priv(sch);
313 	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
314 	psched_time_t qdelay = 0;	/* in pschedtime */
315 	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
316 	s32 delta = 0;		/* determines the change in probability */
317 	u32 oldprob;
318 	u32 alpha, beta;
319 	bool update_prob = true;
320 
321 	q->vars.qdelay_old = q->vars.qdelay;
322 
323 	if (q->vars.avg_dq_rate > 0)
324 		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
325 	else
326 		qdelay = 0;
327 
328 	/* If qdelay is zero and qlen is not, it means qlen is very small, less
329 	 * than dequeue_rate, so we do not update probabilty in this round
330 	 */
331 	if (qdelay == 0 && qlen != 0)
332 		update_prob = false;
333 
334 	/* In the algorithm, alpha and beta are between 0 and 2 with typical
335 	 * value for alpha as 0.125. In this implementation, we use values 0-32
336 	 * passed from user space to represent this. Also, alpha and beta have
337 	 * unit of HZ and need to be scaled before they can used to update
338 	 * probability. alpha/beta are updated locally below by 1) scaling them
339 	 * appropriately 2) scaling down by 16 to come to 0-2 range.
340 	 * Please see paper for details.
341 	 *
342 	 * We scale alpha and beta differently depending on whether we are in
343 	 * light, medium or high dropping mode.
344 	 */
345 	if (q->vars.prob < MAX_PROB / 100) {
346 		alpha =
347 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
348 		beta =
349 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
350 	} else if (q->vars.prob < MAX_PROB / 10) {
351 		alpha =
352 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
353 		beta =
354 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
355 	} else {
356 		alpha =
357 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
358 		beta =
359 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
360 	}
361 
362 	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
363 	delta += alpha * ((qdelay - q->params.target));
364 	delta += beta * ((qdelay - qdelay_old));
365 
366 	oldprob = q->vars.prob;
367 
368 	/* to ensure we increase probability in steps of no more than 2% */
369 	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
370 	    q->vars.prob >= MAX_PROB / 10)
371 		delta = (MAX_PROB / 100) * 2;
372 
373 	/* Non-linear drop:
374 	 * Tune drop probability to increase quickly for high delays(>= 250ms)
375 	 * 250ms is derived through experiments and provides error protection
376 	 */
377 
378 	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
379 		delta += MAX_PROB / (100 / 2);
380 
381 	q->vars.prob += delta;
382 
383 	if (delta > 0) {
384 		/* prevent overflow */
385 		if (q->vars.prob < oldprob) {
386 			q->vars.prob = MAX_PROB;
387 			/* Prevent normalization error. If probability is at
388 			 * maximum value already, we normalize it here, and
389 			 * skip the check to do a non-linear drop in the next
390 			 * section.
391 			 */
392 			update_prob = false;
393 		}
394 	} else {
395 		/* prevent underflow */
396 		if (q->vars.prob > oldprob)
397 			q->vars.prob = 0;
398 	}
399 
400 	/* Non-linear drop in probability: Reduce drop probability quickly if
401 	 * delay is 0 for 2 consecutive Tupdate periods.
402 	 */
403 
404 	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
405 		q->vars.prob = (q->vars.prob * 98) / 100;
406 
407 	q->vars.qdelay = qdelay;
408 	q->vars.qlen_old = qlen;
409 
410 	/* We restart the measurement cycle if the following conditions are met
411 	 * 1. If the delay has been low for 2 consecutive Tupdate periods
412 	 * 2. Calculated drop probability is zero
413 	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
414 	 *    is a non-zero value
415 	 */
416 	if ((q->vars.qdelay < q->params.target / 2) &&
417 	    (q->vars.qdelay_old < q->params.target / 2) &&
418 	    (q->vars.prob == 0) &&
419 	    (q->vars.avg_dq_rate > 0))
420 		pie_vars_init(&q->vars);
421 }
422 
423 static void pie_timer(unsigned long arg)
424 {
425 	struct Qdisc *sch = (struct Qdisc *)arg;
426 	struct pie_sched_data *q = qdisc_priv(sch);
427 	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
428 
429 	spin_lock(root_lock);
430 	calculate_probability(sch);
431 
432 	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
433 	if (q->params.tupdate)
434 		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
435 	spin_unlock(root_lock);
436 
437 }
438 
439 static int pie_init(struct Qdisc *sch, struct nlattr *opt)
440 {
441 	struct pie_sched_data *q = qdisc_priv(sch);
442 
443 	pie_params_init(&q->params);
444 	pie_vars_init(&q->vars);
445 	sch->limit = q->params.limit;
446 
447 	setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
448 
449 	if (opt) {
450 		int err = pie_change(sch, opt);
451 
452 		if (err)
453 			return err;
454 	}
455 
456 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
457 	return 0;
458 }
459 
460 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
461 {
462 	struct pie_sched_data *q = qdisc_priv(sch);
463 	struct nlattr *opts;
464 
465 	opts = nla_nest_start(skb, TCA_OPTIONS);
466 	if (opts == NULL)
467 		goto nla_put_failure;
468 
469 	/* convert target from pschedtime to us */
470 	if (nla_put_u32(skb, TCA_PIE_TARGET,
471 			((u32) PSCHED_TICKS2NS(q->params.target)) /
472 			NSEC_PER_USEC) ||
473 	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
474 	    nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
475 	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
476 	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
477 	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
478 	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
479 		goto nla_put_failure;
480 
481 	return nla_nest_end(skb, opts);
482 
483 nla_put_failure:
484 	nla_nest_cancel(skb, opts);
485 	return -1;
486 
487 }
488 
489 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
490 {
491 	struct pie_sched_data *q = qdisc_priv(sch);
492 	struct tc_pie_xstats st = {
493 		.prob		= q->vars.prob,
494 		.delay		= ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
495 				   NSEC_PER_USEC,
496 		/* unscale and return dq_rate in bytes per sec */
497 		.avg_dq_rate	= q->vars.avg_dq_rate *
498 				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
499 		.packets_in	= q->stats.packets_in,
500 		.overlimit	= q->stats.overlimit,
501 		.maxq		= q->stats.maxq,
502 		.dropped	= q->stats.dropped,
503 		.ecn_mark	= q->stats.ecn_mark,
504 	};
505 
506 	return gnet_stats_copy_app(d, &st, sizeof(st));
507 }
508 
509 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
510 {
511 	struct sk_buff *skb;
512 	skb = __qdisc_dequeue_head(sch, &sch->q);
513 
514 	if (!skb)
515 		return NULL;
516 
517 	pie_process_dequeue(sch, skb);
518 	return skb;
519 }
520 
521 static void pie_reset(struct Qdisc *sch)
522 {
523 	struct pie_sched_data *q = qdisc_priv(sch);
524 	qdisc_reset_queue(sch);
525 	pie_vars_init(&q->vars);
526 }
527 
528 static void pie_destroy(struct Qdisc *sch)
529 {
530 	struct pie_sched_data *q = qdisc_priv(sch);
531 	q->params.tupdate = 0;
532 	del_timer_sync(&q->adapt_timer);
533 }
534 
535 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
536 	.id = "pie",
537 	.priv_size	= sizeof(struct pie_sched_data),
538 	.enqueue	= pie_qdisc_enqueue,
539 	.dequeue	= pie_qdisc_dequeue,
540 	.peek		= qdisc_peek_dequeued,
541 	.init		= pie_init,
542 	.destroy	= pie_destroy,
543 	.reset		= pie_reset,
544 	.change		= pie_change,
545 	.dump		= pie_dump,
546 	.dump_stats	= pie_dump_stats,
547 	.owner		= THIS_MODULE,
548 };
549 
550 static int __init pie_module_init(void)
551 {
552 	return register_qdisc(&pie_qdisc_ops);
553 }
554 
555 static void __exit pie_module_exit(void)
556 {
557 	unregister_qdisc(&pie_qdisc_ops);
558 }
559 
560 module_init(pie_module_init);
561 module_exit(pie_module_exit);
562 
563 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
564 MODULE_AUTHOR("Vijay Subramanian");
565 MODULE_AUTHOR("Mythili Prabhu");
566 MODULE_LICENSE("GPL");
567