sch_pie.c

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