xref: /linux/net/sched/sch_pie.c (revision 4ba9920e5e9c0e16b5ed24292d45322907bb9035)
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 
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/errno.h>
25 #include <linux/skbuff.h>
26 #include <net/pkt_sched.h>
27 #include <net/inet_ecn.h>
28 
29 #define QUEUE_THRESHOLD 10000
30 #define DQCOUNT_INVALID -1
31 #define MAX_PROB  0xffffffff
32 #define PIE_SCALE 8
33 
34 /* parameters used */
35 struct pie_params {
36 	psched_time_t target;	/* user specified target delay in pschedtime */
37 	u32 tupdate;		/* timer frequency (in jiffies) */
38 	u32 limit;		/* number of packets that can be enqueued */
39 	u32 alpha;		/* alpha and beta are between -4 and 4 */
40 	u32 beta;		/* and are used for shift relative to 1 */
41 	bool ecn;		/* true if ecn is enabled */
42 	bool bytemode;		/* to scale drop early prob based on pkt size */
43 };
44 
45 /* variables used */
46 struct pie_vars {
47 	u32 prob;		/* probability but scaled by u32 limit. */
48 	psched_time_t burst_time;
49 	psched_time_t qdelay;
50 	psched_time_t qdelay_old;
51 	u64 dq_count;		/* measured in bytes */
52 	psched_time_t dq_tstamp;	/* drain rate */
53 	u32 avg_dq_rate;	/* bytes per pschedtime tick,scaled */
54 	u32 qlen_old;		/* in bytes */
55 };
56 
57 /* statistics gathering */
58 struct pie_stats {
59 	u32 packets_in;		/* total number of packets enqueued */
60 	u32 dropped;		/* packets dropped due to pie_action */
61 	u32 overlimit;		/* dropped due to lack of space in queue */
62 	u32 maxq;		/* maximum queue size */
63 	u32 ecn_mark;		/* packets marked with ECN */
64 };
65 
66 /* private data for the Qdisc */
67 struct pie_sched_data {
68 	struct pie_params params;
69 	struct pie_vars vars;
70 	struct pie_stats stats;
71 	struct timer_list adapt_timer;
72 };
73 
74 static void pie_params_init(struct pie_params *params)
75 {
76 	params->alpha = 2;
77 	params->beta = 20;
78 	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
79 	params->limit = 1000;	/* default of 1000 packets */
80 	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);	/* 20 ms */
81 	params->ecn = false;
82 	params->bytemode = false;
83 }
84 
85 static void pie_vars_init(struct pie_vars *vars)
86 {
87 	vars->dq_count = DQCOUNT_INVALID;
88 	vars->avg_dq_rate = 0;
89 	/* default of 100 ms in pschedtime */
90 	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
91 }
92 
93 static bool drop_early(struct Qdisc *sch, u32 packet_size)
94 {
95 	struct pie_sched_data *q = qdisc_priv(sch);
96 	u32 rnd;
97 	u32 local_prob = q->vars.prob;
98 	u32 mtu = psched_mtu(qdisc_dev(sch));
99 
100 	/* If there is still burst allowance left skip random early drop */
101 	if (q->vars.burst_time > 0)
102 		return false;
103 
104 	/* If current delay is less than half of target, and
105 	 * if drop prob is low already, disable early_drop
106 	 */
107 	if ((q->vars.qdelay < q->params.target / 2)
108 	    && (q->vars.prob < MAX_PROB / 5))
109 		return false;
110 
111 	/* If we have fewer than 2 mtu-sized packets, disable drop_early,
112 	 * similar to min_th in RED
113 	 */
114 	if (sch->qstats.backlog < 2 * mtu)
115 		return false;
116 
117 	/* If bytemode is turned on, use packet size to compute new
118 	 * probablity. Smaller packets will have lower drop prob in this case
119 	 */
120 	if (q->params.bytemode && packet_size <= mtu)
121 		local_prob = (local_prob / mtu) * packet_size;
122 	else
123 		local_prob = q->vars.prob;
124 
125 	rnd = prandom_u32();
126 	if (rnd < local_prob)
127 		return true;
128 
129 	return false;
130 }
131 
132 static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
133 {
134 	struct pie_sched_data *q = qdisc_priv(sch);
135 	bool enqueue = false;
136 
137 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
138 		q->stats.overlimit++;
139 		goto out;
140 	}
141 
142 	if (!drop_early(sch, skb->len)) {
143 		enqueue = true;
144 	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
145 		   INET_ECN_set_ce(skb)) {
146 		/* If packet is ecn capable, mark it if drop probability
147 		 * is lower than 10%, else drop it.
148 		 */
149 		q->stats.ecn_mark++;
150 		enqueue = true;
151 	}
152 
153 	/* we can enqueue the packet */
154 	if (enqueue) {
155 		q->stats.packets_in++;
156 		if (qdisc_qlen(sch) > q->stats.maxq)
157 			q->stats.maxq = qdisc_qlen(sch);
158 
159 		return qdisc_enqueue_tail(skb, sch);
160 	}
161 
162 out:
163 	q->stats.dropped++;
164 	return qdisc_drop(skb, sch);
165 }
166 
167 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
168 	[TCA_PIE_TARGET] = {.type = NLA_U32},
169 	[TCA_PIE_LIMIT] = {.type = NLA_U32},
170 	[TCA_PIE_TUPDATE] = {.type = NLA_U32},
171 	[TCA_PIE_ALPHA] = {.type = NLA_U32},
172 	[TCA_PIE_BETA] = {.type = NLA_U32},
173 	[TCA_PIE_ECN] = {.type = NLA_U32},
174 	[TCA_PIE_BYTEMODE] = {.type = NLA_U32},
175 };
176 
177 static int pie_change(struct Qdisc *sch, struct nlattr *opt)
178 {
179 	struct pie_sched_data *q = qdisc_priv(sch);
180 	struct nlattr *tb[TCA_PIE_MAX + 1];
181 	unsigned int qlen;
182 	int err;
183 
184 	if (!opt)
185 		return -EINVAL;
186 
187 	err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
188 	if (err < 0)
189 		return err;
190 
191 	sch_tree_lock(sch);
192 
193 	/* convert from microseconds to pschedtime */
194 	if (tb[TCA_PIE_TARGET]) {
195 		/* target is in us */
196 		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
197 
198 		/* convert to pschedtime */
199 		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
200 	}
201 
202 	/* tupdate is in jiffies */
203 	if (tb[TCA_PIE_TUPDATE])
204 		q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
205 
206 	if (tb[TCA_PIE_LIMIT]) {
207 		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
208 
209 		q->params.limit = limit;
210 		sch->limit = limit;
211 	}
212 
213 	if (tb[TCA_PIE_ALPHA])
214 		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
215 
216 	if (tb[TCA_PIE_BETA])
217 		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
218 
219 	if (tb[TCA_PIE_ECN])
220 		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
221 
222 	if (tb[TCA_PIE_BYTEMODE])
223 		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
224 
225 	/* Drop excess packets if new limit is lower */
226 	qlen = sch->q.qlen;
227 	while (sch->q.qlen > sch->limit) {
228 		struct sk_buff *skb = __skb_dequeue(&sch->q);
229 
230 		sch->qstats.backlog -= qdisc_pkt_len(skb);
231 		qdisc_drop(skb, sch);
232 	}
233 	qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
234 
235 	sch_tree_unlock(sch);
236 	return 0;
237 }
238 
239 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
240 {
241 
242 	struct pie_sched_data *q = qdisc_priv(sch);
243 	int qlen = sch->qstats.backlog;	/* current queue size in bytes */
244 
245 	/* If current queue is about 10 packets or more and dq_count is unset
246 	 * we have enough packets to calculate the drain rate. Save
247 	 * current time as dq_tstamp and start measurement cycle.
248 	 */
249 	if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
250 		q->vars.dq_tstamp = psched_get_time();
251 		q->vars.dq_count = 0;
252 	}
253 
254 	/* Calculate the average drain rate from this value.  If queue length
255 	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
256 	 * the dq_count to -1 as we don't have enough packets to calculate the
257 	 * drain rate anymore The following if block is entered only when we
258 	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
259 	 * and we calculate the drain rate for the threshold here.  dq_count is
260 	 * in bytes, time difference in psched_time, hence rate is in
261 	 * bytes/psched_time.
262 	 */
263 	if (q->vars.dq_count != DQCOUNT_INVALID) {
264 		q->vars.dq_count += skb->len;
265 
266 		if (q->vars.dq_count >= QUEUE_THRESHOLD) {
267 			psched_time_t now = psched_get_time();
268 			u32 dtime = now - q->vars.dq_tstamp;
269 			u32 count = q->vars.dq_count << PIE_SCALE;
270 
271 			if (dtime == 0)
272 				return;
273 
274 			count = count / dtime;
275 
276 			if (q->vars.avg_dq_rate == 0)
277 				q->vars.avg_dq_rate = count;
278 			else
279 				q->vars.avg_dq_rate =
280 				    (q->vars.avg_dq_rate -
281 				     (q->vars.avg_dq_rate >> 3)) + (count >> 3);
282 
283 			/* If the queue has receded below the threshold, we hold
284 			 * on to the last drain rate calculated, else we reset
285 			 * dq_count to 0 to re-enter the if block when the next
286 			 * packet is dequeued
287 			 */
288 			if (qlen < QUEUE_THRESHOLD)
289 				q->vars.dq_count = DQCOUNT_INVALID;
290 			else {
291 				q->vars.dq_count = 0;
292 				q->vars.dq_tstamp = psched_get_time();
293 			}
294 
295 			if (q->vars.burst_time > 0) {
296 				if (q->vars.burst_time > dtime)
297 					q->vars.burst_time -= dtime;
298 				else
299 					q->vars.burst_time = 0;
300 			}
301 		}
302 	}
303 }
304 
305 static void calculate_probability(struct Qdisc *sch)
306 {
307 	struct pie_sched_data *q = qdisc_priv(sch);
308 	u32 qlen = sch->qstats.backlog;	/* queue size in bytes */
309 	psched_time_t qdelay = 0;	/* in pschedtime */
310 	psched_time_t qdelay_old = q->vars.qdelay;	/* in pschedtime */
311 	s32 delta = 0;		/* determines the change in probability */
312 	u32 oldprob;
313 	u32 alpha, beta;
314 	bool update_prob = true;
315 
316 	q->vars.qdelay_old = q->vars.qdelay;
317 
318 	if (q->vars.avg_dq_rate > 0)
319 		qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
320 	else
321 		qdelay = 0;
322 
323 	/* If qdelay is zero and qlen is not, it means qlen is very small, less
324 	 * than dequeue_rate, so we do not update probabilty in this round
325 	 */
326 	if (qdelay == 0 && qlen != 0)
327 		update_prob = false;
328 
329 	/* Add ranges for alpha and beta, more aggressive for high dropping
330 	 * mode and gentle steps for light dropping mode
331 	 * In light dropping mode, take gentle steps; in medium dropping mode,
332 	 * take medium steps; in high dropping mode, take big steps.
333 	 */
334 	if (q->vars.prob < MAX_PROB / 100) {
335 		alpha =
336 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
337 		beta =
338 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
339 	} else if (q->vars.prob < MAX_PROB / 10) {
340 		alpha =
341 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
342 		beta =
343 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
344 	} else {
345 		alpha =
346 		    (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
347 		beta =
348 		    (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
349 	}
350 
351 	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
352 	delta += alpha * ((qdelay - q->params.target));
353 	delta += beta * ((qdelay - qdelay_old));
354 
355 	oldprob = q->vars.prob;
356 
357 	/* to ensure we increase probability in steps of no more than 2% */
358 	if (delta > (s32) (MAX_PROB / (100 / 2)) &&
359 	    q->vars.prob >= MAX_PROB / 10)
360 		delta = (MAX_PROB / 100) * 2;
361 
362 	/* Non-linear drop:
363 	 * Tune drop probability to increase quickly for high delays(>= 250ms)
364 	 * 250ms is derived through experiments and provides error protection
365 	 */
366 
367 	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
368 		delta += MAX_PROB / (100 / 2);
369 
370 	q->vars.prob += delta;
371 
372 	if (delta > 0) {
373 		/* prevent overflow */
374 		if (q->vars.prob < oldprob) {
375 			q->vars.prob = MAX_PROB;
376 			/* Prevent normalization error. If probability is at
377 			 * maximum value already, we normalize it here, and
378 			 * skip the check to do a non-linear drop in the next
379 			 * section.
380 			 */
381 			update_prob = false;
382 		}
383 	} else {
384 		/* prevent underflow */
385 		if (q->vars.prob > oldprob)
386 			q->vars.prob = 0;
387 	}
388 
389 	/* Non-linear drop in probability: Reduce drop probability quickly if
390 	 * delay is 0 for 2 consecutive Tupdate periods.
391 	 */
392 
393 	if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
394 		q->vars.prob = (q->vars.prob * 98) / 100;
395 
396 	q->vars.qdelay = qdelay;
397 	q->vars.qlen_old = qlen;
398 
399 	/* We restart the measurement cycle if the following conditions are met
400 	 * 1. If the delay has been low for 2 consecutive Tupdate periods
401 	 * 2. Calculated drop probability is zero
402 	 * 3. We have atleast one estimate for the avg_dq_rate ie.,
403 	 *    is a non-zero value
404 	 */
405 	if ((q->vars.qdelay < q->params.target / 2) &&
406 	    (q->vars.qdelay_old < q->params.target / 2) &&
407 	    (q->vars.prob == 0) &&
408 	    (q->vars.avg_dq_rate > 0))
409 		pie_vars_init(&q->vars);
410 }
411 
412 static void pie_timer(unsigned long arg)
413 {
414 	struct Qdisc *sch = (struct Qdisc *)arg;
415 	struct pie_sched_data *q = qdisc_priv(sch);
416 	spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
417 
418 	spin_lock(root_lock);
419 	calculate_probability(sch);
420 
421 	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
422 	if (q->params.tupdate)
423 		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
424 	spin_unlock(root_lock);
425 
426 }
427 
428 static int pie_init(struct Qdisc *sch, struct nlattr *opt)
429 {
430 	struct pie_sched_data *q = qdisc_priv(sch);
431 
432 	pie_params_init(&q->params);
433 	pie_vars_init(&q->vars);
434 	sch->limit = q->params.limit;
435 
436 	setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
437 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
438 
439 	if (opt) {
440 		int err = pie_change(sch, opt);
441 
442 		if (err)
443 			return err;
444 	}
445 
446 	return 0;
447 }
448 
449 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
450 {
451 	struct pie_sched_data *q = qdisc_priv(sch);
452 	struct nlattr *opts;
453 
454 	opts = nla_nest_start(skb, TCA_OPTIONS);
455 	if (opts == NULL)
456 		goto nla_put_failure;
457 
458 	/* convert target from pschedtime to us */
459 	if (nla_put_u32(skb, TCA_PIE_TARGET,
460 			((u32) PSCHED_TICKS2NS(q->params.target)) /
461 			NSEC_PER_USEC) ||
462 	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
463 	    nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
464 	    nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
465 	    nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
466 	    nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
467 	    nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
468 		goto nla_put_failure;
469 
470 	return nla_nest_end(skb, opts);
471 
472 nla_put_failure:
473 	nla_nest_cancel(skb, opts);
474 	return -1;
475 
476 }
477 
478 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
479 {
480 	struct pie_sched_data *q = qdisc_priv(sch);
481 	struct tc_pie_xstats st = {
482 		.prob		= q->vars.prob,
483 		.delay		= ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
484 				   NSEC_PER_USEC,
485 		/* unscale and return dq_rate in bytes per sec */
486 		.avg_dq_rate	= q->vars.avg_dq_rate *
487 				  (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
488 		.packets_in	= q->stats.packets_in,
489 		.overlimit	= q->stats.overlimit,
490 		.maxq		= q->stats.maxq,
491 		.dropped	= q->stats.dropped,
492 		.ecn_mark	= q->stats.ecn_mark,
493 	};
494 
495 	return gnet_stats_copy_app(d, &st, sizeof(st));
496 }
497 
498 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
499 {
500 	struct sk_buff *skb;
501 	skb = __qdisc_dequeue_head(sch, &sch->q);
502 
503 	if (!skb)
504 		return NULL;
505 
506 	pie_process_dequeue(sch, skb);
507 	return skb;
508 }
509 
510 static void pie_reset(struct Qdisc *sch)
511 {
512 	struct pie_sched_data *q = qdisc_priv(sch);
513 	qdisc_reset_queue(sch);
514 	pie_vars_init(&q->vars);
515 }
516 
517 static void pie_destroy(struct Qdisc *sch)
518 {
519 	struct pie_sched_data *q = qdisc_priv(sch);
520 	q->params.tupdate = 0;
521 	del_timer_sync(&q->adapt_timer);
522 }
523 
524 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
525 	.id = "pie",
526 	.priv_size	= sizeof(struct pie_sched_data),
527 	.enqueue	= pie_qdisc_enqueue,
528 	.dequeue	= pie_qdisc_dequeue,
529 	.peek		= qdisc_peek_dequeued,
530 	.init		= pie_init,
531 	.destroy	= pie_destroy,
532 	.reset		= pie_reset,
533 	.change		= pie_change,
534 	.dump		= pie_dump,
535 	.dump_stats	= pie_dump_stats,
536 	.owner		= THIS_MODULE,
537 };
538 
539 static int __init pie_module_init(void)
540 {
541 	return register_qdisc(&pie_qdisc_ops);
542 }
543 
544 static void __exit pie_module_exit(void)
545 {
546 	unregister_qdisc(&pie_qdisc_ops);
547 }
548 
549 module_init(pie_module_init);
550 module_exit(pie_module_exit);
551 
552 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
553 MODULE_AUTHOR("Vijay Subramanian");
554 MODULE_AUTHOR("Mythili Prabhu");
555 MODULE_LICENSE("GPL");
556