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	struct pie_sched_data *q = qdisc_priv(sch);
89	bool enqueue = false;
90
91	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
92		q->stats.overlimit++;
93		goto out;
94	}
95
96	if (!pie_drop_early(sch, &q->params, &q->vars, sch->qstats.backlog,
97			    skb->len)) {
98		enqueue = true;
99	} else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
100		   INET_ECN_set_ce(skb)) {
101		/* If packet is ecn capable, mark it if drop probability
102		 * is lower than 10%, else drop it.
103		 */
104		q->stats.ecn_mark++;
105		enqueue = true;
106	}
107
108	/* we can enqueue the packet */
109	if (enqueue) {
110		/* Set enqueue time only when dq_rate_estimator is disabled. */
111		if (!q->params.dq_rate_estimator)
112			pie_set_enqueue_time(skb);
113
114		q->stats.packets_in++;
115		if (qdisc_qlen(sch) > q->stats.maxq)
116			q->stats.maxq = qdisc_qlen(sch);
117
118		return qdisc_enqueue_tail(skb, sch);
119	}
120
121out:
122	q->stats.dropped++;
123	q->vars.accu_prob = 0;
124	return qdisc_drop(skb, sch, to_free);
125}
126
127static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
128	[TCA_PIE_TARGET]		= {.type = NLA_U32},
129	[TCA_PIE_LIMIT]			= {.type = NLA_U32},
130	[TCA_PIE_TUPDATE]		= {.type = NLA_U32},
131	[TCA_PIE_ALPHA]			= {.type = NLA_U32},
132	[TCA_PIE_BETA]			= {.type = NLA_U32},
133	[TCA_PIE_ECN]			= {.type = NLA_U32},
134	[TCA_PIE_BYTEMODE]		= {.type = NLA_U32},
135	[TCA_PIE_DQ_RATE_ESTIMATOR]	= {.type = NLA_U32},
136};
137
138static int pie_change(struct Qdisc *sch, struct nlattr *opt,
139		      struct netlink_ext_ack *extack)
140{
141	struct pie_sched_data *q = qdisc_priv(sch);
142	struct nlattr *tb[TCA_PIE_MAX + 1];
143	unsigned int qlen, dropped = 0;
144	int err;
145
146	err = nla_parse_nested_deprecated(tb, TCA_PIE_MAX, opt, pie_policy,
147					  NULL);
148	if (err < 0)
149		return err;
150
151	sch_tree_lock(sch);
152
153	/* convert from microseconds to pschedtime */
154	if (tb[TCA_PIE_TARGET]) {
155		/* target is in us */
156		u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
157
158		/* convert to pschedtime */
159		q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
160	}
161
162	/* tupdate is in jiffies */
163	if (tb[TCA_PIE_TUPDATE])
164		q->params.tupdate =
165			usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
166
167	if (tb[TCA_PIE_LIMIT]) {
168		u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
169
170		q->params.limit = limit;
171		sch->limit = limit;
172	}
173
174	if (tb[TCA_PIE_ALPHA])
175		q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
176
177	if (tb[TCA_PIE_BETA])
178		q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
179
180	if (tb[TCA_PIE_ECN])
181		q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
182
183	if (tb[TCA_PIE_BYTEMODE])
184		q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
185
186	if (tb[TCA_PIE_DQ_RATE_ESTIMATOR])
187		q->params.dq_rate_estimator =
188				nla_get_u32(tb[TCA_PIE_DQ_RATE_ESTIMATOR]);
189
190	/* Drop excess packets if new limit is lower */
191	qlen = sch->q.qlen;
192	while (sch->q.qlen > sch->limit) {
193		struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
194
195		dropped += qdisc_pkt_len(skb);
196		qdisc_qstats_backlog_dec(sch, skb);
197		rtnl_qdisc_drop(skb, sch);
198	}
199	qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
200
201	sch_tree_unlock(sch);
202	return 0;
203}
204
205void pie_process_dequeue(struct sk_buff *skb, struct pie_params *params,
206			 struct pie_vars *vars, u32 backlog)
207{
208	psched_time_t now = psched_get_time();
209	u32 dtime = 0;
210
211	/* If dq_rate_estimator is disabled, calculate qdelay using the
212	 * packet timestamp.
213	 */
214	if (!params->dq_rate_estimator) {
215		vars->qdelay = now - pie_get_enqueue_time(skb);
216
217		if (vars->dq_tstamp != DTIME_INVALID)
218			dtime = now - vars->dq_tstamp;
219
220		vars->dq_tstamp = now;
221
222		if (backlog == 0)
223			vars->qdelay = 0;
224
225		if (dtime == 0)
226			return;
227
228		goto burst_allowance_reduction;
229	}
230
231	/* If current queue is about 10 packets or more and dq_count is unset
232	 * we have enough packets to calculate the drain rate. Save
233	 * current time as dq_tstamp and start measurement cycle.
234	 */
235	if (backlog >= QUEUE_THRESHOLD && vars->dq_count == DQCOUNT_INVALID) {
236		vars->dq_tstamp = psched_get_time();
237		vars->dq_count = 0;
238	}
239
240	/* Calculate the average drain rate from this value. If queue length
241	 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes, reset
242	 * the dq_count to -1 as we don't have enough packets to calculate the
243	 * drain rate anymore. The following if block is entered only when we
244	 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
245	 * and we calculate the drain rate for the threshold here.  dq_count is
246	 * in bytes, time difference in psched_time, hence rate is in
247	 * bytes/psched_time.
248	 */
249	if (vars->dq_count != DQCOUNT_INVALID) {
250		vars->dq_count += skb->len;
251
252		if (vars->dq_count >= QUEUE_THRESHOLD) {
253			u32 count = vars->dq_count << PIE_SCALE;
254
255			dtime = now - vars->dq_tstamp;
256
257			if (dtime == 0)
258				return;
259
260			count = count / dtime;
261
262			if (vars->avg_dq_rate == 0)
263				vars->avg_dq_rate = count;
264			else
265				vars->avg_dq_rate =
266				    (vars->avg_dq_rate -
267				     (vars->avg_dq_rate >> 3)) + (count >> 3);
268
269			/* If the queue has receded below the threshold, we hold
270			 * on to the last drain rate calculated, else we reset
271			 * dq_count to 0 to re-enter the if block when the next
272			 * packet is dequeued
273			 */
274			if (backlog < QUEUE_THRESHOLD) {
275				vars->dq_count = DQCOUNT_INVALID;
276			} else {
277				vars->dq_count = 0;
278				vars->dq_tstamp = psched_get_time();
279			}
280
281			goto burst_allowance_reduction;
282		}
283	}
284
285	return;
286
287burst_allowance_reduction:
288	if (vars->burst_time > 0) {
289		if (vars->burst_time > dtime)
290			vars->burst_time -= dtime;
291		else
292			vars->burst_time = 0;
293	}
294}
295EXPORT_SYMBOL_GPL(pie_process_dequeue);
296
297void pie_calculate_probability(struct pie_params *params, struct pie_vars *vars,
298			       u32 backlog)
299{
300	psched_time_t qdelay = 0;	/* in pschedtime */
301	psched_time_t qdelay_old = 0;	/* in pschedtime */
302	s64 delta = 0;		/* determines the change in probability */
303	u64 oldprob;
304	u64 alpha, beta;
305	u32 power;
306	bool update_prob = true;
307
308	if (params->dq_rate_estimator) {
309		qdelay_old = vars->qdelay;
310		vars->qdelay_old = vars->qdelay;
311
312		if (vars->avg_dq_rate > 0)
313			qdelay = (backlog << PIE_SCALE) / vars->avg_dq_rate;
314		else
315			qdelay = 0;
316	} else {
317		qdelay = vars->qdelay;
318		qdelay_old = vars->qdelay_old;
319	}
320
321	/* If qdelay is zero and backlog is not, it means backlog is very small,
322	 * so we do not update probability in this round.
323	 */
324	if (qdelay == 0 && backlog != 0)
325		update_prob = false;
326
327	/* In the algorithm, alpha and beta are between 0 and 2 with typical
328	 * value for alpha as 0.125. In this implementation, we use values 0-32
329	 * passed from user space to represent this. Also, alpha and beta have
330	 * unit of HZ and need to be scaled before they can used to update
331	 * probability. alpha/beta are updated locally below by scaling down
332	 * by 16 to come to 0-2 range.
333	 */
334	alpha = ((u64)params->alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
335	beta = ((u64)params->beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
336
337	/* We scale alpha and beta differently depending on how heavy the
338	 * congestion is. Please see RFC 8033 for details.
339	 */
340	if (vars->prob < MAX_PROB / 10) {
341		alpha >>= 1;
342		beta >>= 1;
343
344		power = 100;
345		while (vars->prob < div_u64(MAX_PROB, power) &&
346		       power <= 1000000) {
347			alpha >>= 2;
348			beta >>= 2;
349			power *= 10;
350		}
351	}
352
353	/* alpha and beta should be between 0 and 32, in multiples of 1/16 */
354	delta += alpha * (qdelay - params->target);
355	delta += beta * (qdelay - qdelay_old);
356
357	oldprob = vars->prob;
358
359	/* to ensure we increase probability in steps of no more than 2% */
360	if (delta > (s64)(MAX_PROB / (100 / 2)) &&
361	    vars->prob >= MAX_PROB / 10)
362		delta = (MAX_PROB / 100) * 2;
363
364	/* Non-linear drop:
365	 * Tune drop probability to increase quickly for high delays(>= 250ms)
366	 * 250ms is derived through experiments and provides error protection
367	 */
368
369	if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
370		delta += MAX_PROB / (100 / 2);
371
372	vars->prob += delta;
373
374	if (delta > 0) {
375		/* prevent overflow */
376		if (vars->prob < oldprob) {
377			vars->prob = MAX_PROB;
378			/* Prevent normalization error. If probability is at
379			 * maximum value already, we normalize it here, and
380			 * skip the check to do a non-linear drop in the next
381			 * section.
382			 */
383			update_prob = false;
384		}
385	} else {
386		/* prevent underflow */
387		if (vars->prob > oldprob)
388			vars->prob = 0;
389	}
390
391	/* Non-linear drop in probability: Reduce drop probability quickly if
392	 * delay is 0 for 2 consecutive Tupdate periods.
393	 */
394
395	if (qdelay == 0 && qdelay_old == 0 && update_prob)
396		/* Reduce drop probability to 98.4% */
397		vars->prob -= vars->prob / 64;
398
399	vars->qdelay = qdelay;
400	vars->backlog_old = backlog;
401
402	/* We restart the measurement cycle if the following conditions are met
403	 * 1. If the delay has been low for 2 consecutive Tupdate periods
404	 * 2. Calculated drop probability is zero
405	 * 3. If average dq_rate_estimator is enabled, we have at least one
406	 *    estimate for the avg_dq_rate ie., is a non-zero value
407	 */
408	if ((vars->qdelay < params->target / 2) &&
409	    (vars->qdelay_old < params->target / 2) &&
410	    vars->prob == 0 &&
411	    (!params->dq_rate_estimator || vars->avg_dq_rate > 0)) {
412		pie_vars_init(vars);
413	}
414
415	if (!params->dq_rate_estimator)
416		vars->qdelay_old = qdelay;
417}
418EXPORT_SYMBOL_GPL(pie_calculate_probability);
419
420static void pie_timer(struct timer_list *t)
421{
422	struct pie_sched_data *q = from_timer(q, t, adapt_timer);
423	struct Qdisc *sch = q->sch;
424	spinlock_t *root_lock;
425
426	rcu_read_lock();
427	root_lock = qdisc_lock(qdisc_root_sleeping(sch));
428	spin_lock(root_lock);
429	pie_calculate_probability(&q->params, &q->vars, sch->qstats.backlog);
430
431	/* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
432	if (q->params.tupdate)
433		mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
434	spin_unlock(root_lock);
435	rcu_read_unlock();
436}
437
438static int pie_init(struct Qdisc *sch, struct nlattr *opt,
439		    struct netlink_ext_ack *extack)
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	q->sch = sch;
448	timer_setup(&q->adapt_timer, pie_timer, 0);
449
450	if (opt) {
451		int err = pie_change(sch, opt, extack);
452
453		if (err)
454			return err;
455	}
456
457	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
458	return 0;
459}
460
461static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
462{
463	struct pie_sched_data *q = qdisc_priv(sch);
464	struct nlattr *opts;
465
466	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
467	if (!opts)
468		goto nla_put_failure;
469
470	/* convert target from pschedtime to us */
471	if (nla_put_u32(skb, TCA_PIE_TARGET,
472			((u32)PSCHED_TICKS2NS(q->params.target)) /
473			NSEC_PER_USEC) ||
474	    nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
475	    nla_put_u32(skb, TCA_PIE_TUPDATE,
476			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	    nla_put_u32(skb, TCA_PIE_DQ_RATE_ESTIMATOR,
482			q->params.dq_rate_estimator))
483		goto nla_put_failure;
484
485	return nla_nest_end(skb, opts);
486
487nla_put_failure:
488	nla_nest_cancel(skb, opts);
489	return -1;
490}
491
492static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
493{
494	struct pie_sched_data *q = qdisc_priv(sch);
495	struct tc_pie_xstats st = {
496		.prob		= q->vars.prob << BITS_PER_BYTE,
497		.delay		= ((u32)PSCHED_TICKS2NS(q->vars.qdelay)) /
498				   NSEC_PER_USEC,
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	/* avg_dq_rate is only valid if dq_rate_estimator is enabled */
507	st.dq_rate_estimating = q->params.dq_rate_estimator;
508
509	/* unscale and return dq_rate in bytes per sec */
510	if (q->params.dq_rate_estimator)
511		st.avg_dq_rate = q->vars.avg_dq_rate *
512				 (PSCHED_TICKS_PER_SEC) >> PIE_SCALE;
513
514	return gnet_stats_copy_app(d, &st, sizeof(st));
515}
516
517static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
518{
519	struct pie_sched_data *q = qdisc_priv(sch);
520	struct sk_buff *skb = qdisc_dequeue_head(sch);
521
522	if (!skb)
523		return NULL;
524
525	pie_process_dequeue(skb, &q->params, &q->vars, sch->qstats.backlog);
526	return skb;
527}
528
529static void pie_reset(struct Qdisc *sch)
530{
531	struct pie_sched_data *q = qdisc_priv(sch);
532
533	qdisc_reset_queue(sch);
534	pie_vars_init(&q->vars);
535}
536
537static void pie_destroy(struct Qdisc *sch)
538{
539	struct pie_sched_data *q = qdisc_priv(sch);
540
541	q->params.tupdate = 0;
542	del_timer_sync(&q->adapt_timer);
543}
544
545static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
546	.id		= "pie",
547	.priv_size	= sizeof(struct pie_sched_data),
548	.enqueue	= pie_qdisc_enqueue,
549	.dequeue	= pie_qdisc_dequeue,
550	.peek		= qdisc_peek_dequeued,
551	.init		= pie_init,
552	.destroy	= pie_destroy,
553	.reset		= pie_reset,
554	.change		= pie_change,
555	.dump		= pie_dump,
556	.dump_stats	= pie_dump_stats,
557	.owner		= THIS_MODULE,
558};
559MODULE_ALIAS_NET_SCH("pie");
560
561static int __init pie_module_init(void)
562{
563	return register_qdisc(&pie_qdisc_ops);
564}
565
566static void __exit pie_module_exit(void)
567{
568	unregister_qdisc(&pie_qdisc_ops);
569}
570
571module_init(pie_module_init);
572module_exit(pie_module_exit);
573
574MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
575MODULE_AUTHOR("Vijay Subramanian");
576MODULE_AUTHOR("Mythili Prabhu");
577MODULE_LICENSE("GPL");
578