1/*
2 * net/sched/sch_generic.c	Generic packet scheduler routines.
3 *
4 *		This program is free software; you can redistribute it and/or
5 *		modify it under the terms of the GNU General Public License
6 *		as published by the Free Software Foundation; either version
7 *		2 of the License, or (at your option) any later version.
8 *
9 * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 *              - Ingress support
12 */
13
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/sched.h>
19#include <linux/string.h>
20#include <linux/errno.h>
21#include <linux/netdevice.h>
22#include <linux/skbuff.h>
23#include <linux/rtnetlink.h>
24#include <linux/init.h>
25#include <linux/rcupdate.h>
26#include <linux/list.h>
27#include <linux/slab.h>
28#include <linux/if_vlan.h>
29#include <net/sch_generic.h>
30#include <net/pkt_sched.h>
31#include <net/dst.h>
32
33/* Qdisc to use by default */
34const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35EXPORT_SYMBOL(default_qdisc_ops);
36
37/* Main transmission queue. */
38
39/* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
41 *
42 * The idea is the following:
43 * - enqueue, dequeue are serialized via qdisc root lock
44 * - ingress filtering is also serialized via qdisc root lock
45 * - updates to tree and tree walking are only done under the rtnl mutex.
46 */
47
48static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
49{
50	q->gso_skb = skb;
51	q->qstats.requeues++;
52	q->q.qlen++;	/* it's still part of the queue */
53	__netif_schedule(q);
54
55	return 0;
56}
57
58static void try_bulk_dequeue_skb(struct Qdisc *q,
59				 struct sk_buff *skb,
60				 const struct netdev_queue *txq,
61				 int *packets)
62{
63	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
64
65	while (bytelimit > 0) {
66		struct sk_buff *nskb = q->dequeue(q);
67
68		if (!nskb)
69			break;
70
71		bytelimit -= nskb->len; /* covers GSO len */
72		skb->next = nskb;
73		skb = nskb;
74		(*packets)++; /* GSO counts as one pkt */
75	}
76	skb->next = NULL;
77}
78
79/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
80 * A requeued skb (via q->gso_skb) can also be a SKB list.
81 */
82static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
83				   int *packets)
84{
85	struct sk_buff *skb = q->gso_skb;
86	const struct netdev_queue *txq = q->dev_queue;
87
88	*packets = 1;
89	*validate = true;
90	if (unlikely(skb)) {
91		/* check the reason of requeuing without tx lock first */
92		txq = skb_get_tx_queue(txq->dev, skb);
93		if (!netif_xmit_frozen_or_stopped(txq)) {
94			q->gso_skb = NULL;
95			q->q.qlen--;
96		} else
97			skb = NULL;
98		/* skb in gso_skb were already validated */
99		*validate = false;
100	} else {
101		if (!(q->flags & TCQ_F_ONETXQUEUE) ||
102		    !netif_xmit_frozen_or_stopped(txq)) {
103			skb = q->dequeue(q);
104			if (skb && qdisc_may_bulk(q))
105				try_bulk_dequeue_skb(q, skb, txq, packets);
106		}
107	}
108	return skb;
109}
110
111static inline int handle_dev_cpu_collision(struct sk_buff *skb,
112					   struct netdev_queue *dev_queue,
113					   struct Qdisc *q)
114{
115	int ret;
116
117	if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
118		/*
119		 * Same CPU holding the lock. It may be a transient
120		 * configuration error, when hard_start_xmit() recurses. We
121		 * detect it by checking xmit owner and drop the packet when
122		 * deadloop is detected. Return OK to try the next skb.
123		 */
124		kfree_skb_list(skb);
125		net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
126				     dev_queue->dev->name);
127		ret = qdisc_qlen(q);
128	} else {
129		/*
130		 * Another cpu is holding lock, requeue & delay xmits for
131		 * some time.
132		 */
133		__this_cpu_inc(softnet_data.cpu_collision);
134		ret = dev_requeue_skb(skb, q);
135	}
136
137	return ret;
138}
139
140/*
141 * Transmit possibly several skbs, and handle the return status as
142 * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
143 * only one CPU can execute this function.
144 *
145 * Returns to the caller:
146 *				0  - queue is empty or throttled.
147 *				>0 - queue is not empty.
148 */
149int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
150		    struct net_device *dev, struct netdev_queue *txq,
151		    spinlock_t *root_lock, bool validate)
152{
153	int ret = NETDEV_TX_BUSY;
154
155	/* And release qdisc */
156	spin_unlock(root_lock);
157
158	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
159	if (validate)
160		skb = validate_xmit_skb_list(skb, dev);
161
162	if (skb) {
163		HARD_TX_LOCK(dev, txq, smp_processor_id());
164		if (!netif_xmit_frozen_or_stopped(txq))
165			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
166
167		HARD_TX_UNLOCK(dev, txq);
168	}
169	spin_lock(root_lock);
170
171	if (dev_xmit_complete(ret)) {
172		/* Driver sent out skb successfully or skb was consumed */
173		ret = qdisc_qlen(q);
174	} else if (ret == NETDEV_TX_LOCKED) {
175		/* Driver try lock failed */
176		ret = handle_dev_cpu_collision(skb, txq, q);
177	} else {
178		/* Driver returned NETDEV_TX_BUSY - requeue skb */
179		if (unlikely(ret != NETDEV_TX_BUSY))
180			net_warn_ratelimited("BUG %s code %d qlen %d\n",
181					     dev->name, ret, q->q.qlen);
182
183		ret = dev_requeue_skb(skb, q);
184	}
185
186	if (ret && netif_xmit_frozen_or_stopped(txq))
187		ret = 0;
188
189	return ret;
190}
191
192/*
193 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
194 *
195 * __QDISC___STATE_RUNNING guarantees only one CPU can process
196 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
197 * this queue.
198 *
199 *  netif_tx_lock serializes accesses to device driver.
200 *
201 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
202 *  if one is grabbed, another must be free.
203 *
204 * Note, that this procedure can be called by a watchdog timer
205 *
206 * Returns to the caller:
207 *				0  - queue is empty or throttled.
208 *				>0 - queue is not empty.
209 *
210 */
211static inline int qdisc_restart(struct Qdisc *q, int *packets)
212{
213	struct netdev_queue *txq;
214	struct net_device *dev;
215	spinlock_t *root_lock;
216	struct sk_buff *skb;
217	bool validate;
218
219	/* Dequeue packet */
220	skb = dequeue_skb(q, &validate, packets);
221	if (unlikely(!skb))
222		return 0;
223
224	root_lock = qdisc_lock(q);
225	dev = qdisc_dev(q);
226	txq = skb_get_tx_queue(dev, skb);
227
228	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
229}
230
231void __qdisc_run(struct Qdisc *q)
232{
233	int quota = weight_p;
234	int packets;
235
236	while (qdisc_restart(q, &packets)) {
237		/*
238		 * Ordered by possible occurrence: Postpone processing if
239		 * 1. we've exceeded packet quota
240		 * 2. another process needs the CPU;
241		 */
242		quota -= packets;
243		if (quota <= 0 || need_resched()) {
244			__netif_schedule(q);
245			break;
246		}
247	}
248
249	qdisc_run_end(q);
250}
251
252unsigned long dev_trans_start(struct net_device *dev)
253{
254	unsigned long val, res;
255	unsigned int i;
256
257	if (is_vlan_dev(dev))
258		dev = vlan_dev_real_dev(dev);
259	res = dev->trans_start;
260	for (i = 0; i < dev->num_tx_queues; i++) {
261		val = netdev_get_tx_queue(dev, i)->trans_start;
262		if (val && time_after(val, res))
263			res = val;
264	}
265	dev->trans_start = res;
266
267	return res;
268}
269EXPORT_SYMBOL(dev_trans_start);
270
271static void dev_watchdog(unsigned long arg)
272{
273	struct net_device *dev = (struct net_device *)arg;
274
275	netif_tx_lock(dev);
276	if (!qdisc_tx_is_noop(dev)) {
277		if (netif_device_present(dev) &&
278		    netif_running(dev) &&
279		    netif_carrier_ok(dev)) {
280			int some_queue_timedout = 0;
281			unsigned int i;
282			unsigned long trans_start;
283
284			for (i = 0; i < dev->num_tx_queues; i++) {
285				struct netdev_queue *txq;
286
287				txq = netdev_get_tx_queue(dev, i);
288				/*
289				 * old device drivers set dev->trans_start
290				 */
291				trans_start = txq->trans_start ? : dev->trans_start;
292				if (netif_xmit_stopped(txq) &&
293				    time_after(jiffies, (trans_start +
294							 dev->watchdog_timeo))) {
295					some_queue_timedout = 1;
296					txq->trans_timeout++;
297					break;
298				}
299			}
300
301			if (some_queue_timedout) {
302				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
303				       dev->name, netdev_drivername(dev), i);
304				dev->netdev_ops->ndo_tx_timeout(dev);
305			}
306			if (!mod_timer(&dev->watchdog_timer,
307				       round_jiffies(jiffies +
308						     dev->watchdog_timeo)))
309				dev_hold(dev);
310		}
311	}
312	netif_tx_unlock(dev);
313
314	dev_put(dev);
315}
316
317void __netdev_watchdog_up(struct net_device *dev)
318{
319	if (dev->netdev_ops->ndo_tx_timeout) {
320		if (dev->watchdog_timeo <= 0)
321			dev->watchdog_timeo = 5*HZ;
322		if (!mod_timer(&dev->watchdog_timer,
323			       round_jiffies(jiffies + dev->watchdog_timeo)))
324			dev_hold(dev);
325	}
326}
327
328static void dev_watchdog_up(struct net_device *dev)
329{
330	__netdev_watchdog_up(dev);
331}
332
333static void dev_watchdog_down(struct net_device *dev)
334{
335	netif_tx_lock_bh(dev);
336	if (del_timer(&dev->watchdog_timer))
337		dev_put(dev);
338	netif_tx_unlock_bh(dev);
339}
340
341/**
342 *	netif_carrier_on - set carrier
343 *	@dev: network device
344 *
345 * Device has detected that carrier.
346 */
347void netif_carrier_on(struct net_device *dev)
348{
349	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
350		if (dev->reg_state == NETREG_UNINITIALIZED)
351			return;
352		atomic_inc(&dev->carrier_changes);
353		linkwatch_fire_event(dev);
354		if (netif_running(dev))
355			__netdev_watchdog_up(dev);
356	}
357}
358EXPORT_SYMBOL(netif_carrier_on);
359
360/**
361 *	netif_carrier_off - clear carrier
362 *	@dev: network device
363 *
364 * Device has detected loss of carrier.
365 */
366void netif_carrier_off(struct net_device *dev)
367{
368	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
369		if (dev->reg_state == NETREG_UNINITIALIZED)
370			return;
371		atomic_inc(&dev->carrier_changes);
372		linkwatch_fire_event(dev);
373	}
374}
375EXPORT_SYMBOL(netif_carrier_off);
376
377/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
378   under all circumstances. It is difficult to invent anything faster or
379   cheaper.
380 */
381
382static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
383{
384	kfree_skb(skb);
385	return NET_XMIT_CN;
386}
387
388static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
389{
390	return NULL;
391}
392
393struct Qdisc_ops noop_qdisc_ops __read_mostly = {
394	.id		=	"noop",
395	.priv_size	=	0,
396	.enqueue	=	noop_enqueue,
397	.dequeue	=	noop_dequeue,
398	.peek		=	noop_dequeue,
399	.owner		=	THIS_MODULE,
400};
401
402static struct netdev_queue noop_netdev_queue = {
403	.qdisc		=	&noop_qdisc,
404	.qdisc_sleeping	=	&noop_qdisc,
405};
406
407struct Qdisc noop_qdisc = {
408	.enqueue	=	noop_enqueue,
409	.dequeue	=	noop_dequeue,
410	.flags		=	TCQ_F_BUILTIN,
411	.ops		=	&noop_qdisc_ops,
412	.list		=	LIST_HEAD_INIT(noop_qdisc.list),
413	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
414	.dev_queue	=	&noop_netdev_queue,
415	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
416};
417EXPORT_SYMBOL(noop_qdisc);
418
419static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
420	.id		=	"noqueue",
421	.priv_size	=	0,
422	.enqueue	=	noop_enqueue,
423	.dequeue	=	noop_dequeue,
424	.peek		=	noop_dequeue,
425	.owner		=	THIS_MODULE,
426};
427
428static struct Qdisc noqueue_qdisc;
429static struct netdev_queue noqueue_netdev_queue = {
430	.qdisc		=	&noqueue_qdisc,
431	.qdisc_sleeping	=	&noqueue_qdisc,
432};
433
434static struct Qdisc noqueue_qdisc = {
435	.enqueue	=	NULL,
436	.dequeue	=	noop_dequeue,
437	.flags		=	TCQ_F_BUILTIN,
438	.ops		=	&noqueue_qdisc_ops,
439	.list		=	LIST_HEAD_INIT(noqueue_qdisc.list),
440	.q.lock		=	__SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
441	.dev_queue	=	&noqueue_netdev_queue,
442	.busylock	=	__SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
443};
444
445
446static const u8 prio2band[TC_PRIO_MAX + 1] = {
447	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
448};
449
450/* 3-band FIFO queue: old style, but should be a bit faster than
451   generic prio+fifo combination.
452 */
453
454#define PFIFO_FAST_BANDS 3
455
456/*
457 * Private data for a pfifo_fast scheduler containing:
458 * 	- queues for the three band
459 * 	- bitmap indicating which of the bands contain skbs
460 */
461struct pfifo_fast_priv {
462	u32 bitmap;
463	struct sk_buff_head q[PFIFO_FAST_BANDS];
464};
465
466/*
467 * Convert a bitmap to the first band number where an skb is queued, where:
468 * 	bitmap=0 means there are no skbs on any band.
469 * 	bitmap=1 means there is an skb on band 0.
470 *	bitmap=7 means there are skbs on all 3 bands, etc.
471 */
472static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
473
474static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
475					     int band)
476{
477	return priv->q + band;
478}
479
480static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
481{
482	if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
483		int band = prio2band[skb->priority & TC_PRIO_MAX];
484		struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
485		struct sk_buff_head *list = band2list(priv, band);
486
487		priv->bitmap |= (1 << band);
488		qdisc->q.qlen++;
489		return __qdisc_enqueue_tail(skb, qdisc, list);
490	}
491
492	return qdisc_drop(skb, qdisc);
493}
494
495static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
496{
497	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
498	int band = bitmap2band[priv->bitmap];
499
500	if (likely(band >= 0)) {
501		struct sk_buff_head *list = band2list(priv, band);
502		struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
503
504		qdisc->q.qlen--;
505		if (skb_queue_empty(list))
506			priv->bitmap &= ~(1 << band);
507
508		return skb;
509	}
510
511	return NULL;
512}
513
514static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
515{
516	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
517	int band = bitmap2band[priv->bitmap];
518
519	if (band >= 0) {
520		struct sk_buff_head *list = band2list(priv, band);
521
522		return skb_peek(list);
523	}
524
525	return NULL;
526}
527
528static void pfifo_fast_reset(struct Qdisc *qdisc)
529{
530	int prio;
531	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
532
533	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
534		__qdisc_reset_queue(qdisc, band2list(priv, prio));
535
536	priv->bitmap = 0;
537	qdisc->qstats.backlog = 0;
538	qdisc->q.qlen = 0;
539}
540
541static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
542{
543	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
544
545	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
546	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
547		goto nla_put_failure;
548	return skb->len;
549
550nla_put_failure:
551	return -1;
552}
553
554static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
555{
556	int prio;
557	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
558
559	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
560		__skb_queue_head_init(band2list(priv, prio));
561
562	/* Can by-pass the queue discipline */
563	qdisc->flags |= TCQ_F_CAN_BYPASS;
564	return 0;
565}
566
567struct Qdisc_ops pfifo_fast_ops __read_mostly = {
568	.id		=	"pfifo_fast",
569	.priv_size	=	sizeof(struct pfifo_fast_priv),
570	.enqueue	=	pfifo_fast_enqueue,
571	.dequeue	=	pfifo_fast_dequeue,
572	.peek		=	pfifo_fast_peek,
573	.init		=	pfifo_fast_init,
574	.reset		=	pfifo_fast_reset,
575	.dump		=	pfifo_fast_dump,
576	.owner		=	THIS_MODULE,
577};
578
579static struct lock_class_key qdisc_tx_busylock;
580
581struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
582			  const struct Qdisc_ops *ops)
583{
584	void *p;
585	struct Qdisc *sch;
586	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
587	int err = -ENOBUFS;
588	struct net_device *dev = dev_queue->dev;
589
590	p = kzalloc_node(size, GFP_KERNEL,
591			 netdev_queue_numa_node_read(dev_queue));
592
593	if (!p)
594		goto errout;
595	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
596	/* if we got non aligned memory, ask more and do alignment ourself */
597	if (sch != p) {
598		kfree(p);
599		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
600				 netdev_queue_numa_node_read(dev_queue));
601		if (!p)
602			goto errout;
603		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
604		sch->padded = (char *) sch - (char *) p;
605	}
606	INIT_LIST_HEAD(&sch->list);
607	skb_queue_head_init(&sch->q);
608
609	spin_lock_init(&sch->busylock);
610	lockdep_set_class(&sch->busylock,
611			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
612
613	sch->ops = ops;
614	sch->enqueue = ops->enqueue;
615	sch->dequeue = ops->dequeue;
616	sch->dev_queue = dev_queue;
617	dev_hold(dev);
618	atomic_set(&sch->refcnt, 1);
619
620	return sch;
621errout:
622	return ERR_PTR(err);
623}
624
625struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
626				const struct Qdisc_ops *ops,
627				unsigned int parentid)
628{
629	struct Qdisc *sch;
630
631	if (!try_module_get(ops->owner))
632		goto errout;
633
634	sch = qdisc_alloc(dev_queue, ops);
635	if (IS_ERR(sch))
636		goto errout;
637	sch->parent = parentid;
638
639	if (!ops->init || ops->init(sch, NULL) == 0)
640		return sch;
641
642	qdisc_destroy(sch);
643errout:
644	return NULL;
645}
646EXPORT_SYMBOL(qdisc_create_dflt);
647
648/* Under qdisc_lock(qdisc) and BH! */
649
650void qdisc_reset(struct Qdisc *qdisc)
651{
652	const struct Qdisc_ops *ops = qdisc->ops;
653
654	if (ops->reset)
655		ops->reset(qdisc);
656
657	if (qdisc->gso_skb) {
658		kfree_skb_list(qdisc->gso_skb);
659		qdisc->gso_skb = NULL;
660		qdisc->q.qlen = 0;
661	}
662}
663EXPORT_SYMBOL(qdisc_reset);
664
665static void qdisc_rcu_free(struct rcu_head *head)
666{
667	struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
668
669	if (qdisc_is_percpu_stats(qdisc)) {
670		free_percpu(qdisc->cpu_bstats);
671		free_percpu(qdisc->cpu_qstats);
672	}
673
674	kfree((char *) qdisc - qdisc->padded);
675}
676
677void qdisc_destroy(struct Qdisc *qdisc)
678{
679	const struct Qdisc_ops  *ops = qdisc->ops;
680
681	if (qdisc->flags & TCQ_F_BUILTIN ||
682	    !atomic_dec_and_test(&qdisc->refcnt))
683		return;
684
685#ifdef CONFIG_NET_SCHED
686	qdisc_list_del(qdisc);
687
688	qdisc_put_stab(rtnl_dereference(qdisc->stab));
689#endif
690	gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
691	if (ops->reset)
692		ops->reset(qdisc);
693	if (ops->destroy)
694		ops->destroy(qdisc);
695
696	module_put(ops->owner);
697	dev_put(qdisc_dev(qdisc));
698
699	kfree_skb_list(qdisc->gso_skb);
700	/*
701	 * gen_estimator est_timer() might access qdisc->q.lock,
702	 * wait a RCU grace period before freeing qdisc.
703	 */
704	call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
705}
706EXPORT_SYMBOL(qdisc_destroy);
707
708/* Attach toplevel qdisc to device queue. */
709struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
710			      struct Qdisc *qdisc)
711{
712	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
713	spinlock_t *root_lock;
714
715	root_lock = qdisc_lock(oqdisc);
716	spin_lock_bh(root_lock);
717
718	/* Prune old scheduler */
719	if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
720		qdisc_reset(oqdisc);
721
722	/* ... and graft new one */
723	if (qdisc == NULL)
724		qdisc = &noop_qdisc;
725	dev_queue->qdisc_sleeping = qdisc;
726	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
727
728	spin_unlock_bh(root_lock);
729
730	return oqdisc;
731}
732EXPORT_SYMBOL(dev_graft_qdisc);
733
734static void attach_one_default_qdisc(struct net_device *dev,
735				     struct netdev_queue *dev_queue,
736				     void *_unused)
737{
738	struct Qdisc *qdisc = &noqueue_qdisc;
739
740	if (dev->tx_queue_len) {
741		qdisc = qdisc_create_dflt(dev_queue,
742					  default_qdisc_ops, TC_H_ROOT);
743		if (!qdisc) {
744			netdev_info(dev, "activation failed\n");
745			return;
746		}
747		if (!netif_is_multiqueue(dev))
748			qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
749	}
750	dev_queue->qdisc_sleeping = qdisc;
751}
752
753static void attach_default_qdiscs(struct net_device *dev)
754{
755	struct netdev_queue *txq;
756	struct Qdisc *qdisc;
757
758	txq = netdev_get_tx_queue(dev, 0);
759
760	if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
761		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
762		dev->qdisc = txq->qdisc_sleeping;
763		atomic_inc(&dev->qdisc->refcnt);
764	} else {
765		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
766		if (qdisc) {
767			dev->qdisc = qdisc;
768			qdisc->ops->attach(qdisc);
769		}
770	}
771}
772
773static void transition_one_qdisc(struct net_device *dev,
774				 struct netdev_queue *dev_queue,
775				 void *_need_watchdog)
776{
777	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
778	int *need_watchdog_p = _need_watchdog;
779
780	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
781		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
782
783	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
784	if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
785		dev_queue->trans_start = 0;
786		*need_watchdog_p = 1;
787	}
788}
789
790void dev_activate(struct net_device *dev)
791{
792	int need_watchdog;
793
794	/* No queueing discipline is attached to device;
795	 * create default one for devices, which need queueing
796	 * and noqueue_qdisc for virtual interfaces
797	 */
798
799	if (dev->qdisc == &noop_qdisc)
800		attach_default_qdiscs(dev);
801
802	if (!netif_carrier_ok(dev))
803		/* Delay activation until next carrier-on event */
804		return;
805
806	need_watchdog = 0;
807	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
808	if (dev_ingress_queue(dev))
809		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
810
811	if (need_watchdog) {
812		dev->trans_start = jiffies;
813		dev_watchdog_up(dev);
814	}
815}
816EXPORT_SYMBOL(dev_activate);
817
818static void dev_deactivate_queue(struct net_device *dev,
819				 struct netdev_queue *dev_queue,
820				 void *_qdisc_default)
821{
822	struct Qdisc *qdisc_default = _qdisc_default;
823	struct Qdisc *qdisc;
824
825	qdisc = rtnl_dereference(dev_queue->qdisc);
826	if (qdisc) {
827		spin_lock_bh(qdisc_lock(qdisc));
828
829		if (!(qdisc->flags & TCQ_F_BUILTIN))
830			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
831
832		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
833		qdisc_reset(qdisc);
834
835		spin_unlock_bh(qdisc_lock(qdisc));
836	}
837}
838
839static bool some_qdisc_is_busy(struct net_device *dev)
840{
841	unsigned int i;
842
843	for (i = 0; i < dev->num_tx_queues; i++) {
844		struct netdev_queue *dev_queue;
845		spinlock_t *root_lock;
846		struct Qdisc *q;
847		int val;
848
849		dev_queue = netdev_get_tx_queue(dev, i);
850		q = dev_queue->qdisc_sleeping;
851		root_lock = qdisc_lock(q);
852
853		spin_lock_bh(root_lock);
854
855		val = (qdisc_is_running(q) ||
856		       test_bit(__QDISC_STATE_SCHED, &q->state));
857
858		spin_unlock_bh(root_lock);
859
860		if (val)
861			return true;
862	}
863	return false;
864}
865
866/**
867 * 	dev_deactivate_many - deactivate transmissions on several devices
868 * 	@head: list of devices to deactivate
869 *
870 *	This function returns only when all outstanding transmissions
871 *	have completed, unless all devices are in dismantle phase.
872 */
873void dev_deactivate_many(struct list_head *head)
874{
875	struct net_device *dev;
876	bool sync_needed = false;
877
878	list_for_each_entry(dev, head, close_list) {
879		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
880					 &noop_qdisc);
881		if (dev_ingress_queue(dev))
882			dev_deactivate_queue(dev, dev_ingress_queue(dev),
883					     &noop_qdisc);
884
885		dev_watchdog_down(dev);
886		sync_needed |= !dev->dismantle;
887	}
888
889	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
890	 * This is avoided if all devices are in dismantle phase :
891	 * Caller will call synchronize_net() for us
892	 */
893	if (sync_needed)
894		synchronize_net();
895
896	/* Wait for outstanding qdisc_run calls. */
897	list_for_each_entry(dev, head, close_list)
898		while (some_qdisc_is_busy(dev))
899			yield();
900}
901
902void dev_deactivate(struct net_device *dev)
903{
904	LIST_HEAD(single);
905
906	list_add(&dev->close_list, &single);
907	dev_deactivate_many(&single);
908	list_del(&single);
909}
910EXPORT_SYMBOL(dev_deactivate);
911
912static void dev_init_scheduler_queue(struct net_device *dev,
913				     struct netdev_queue *dev_queue,
914				     void *_qdisc)
915{
916	struct Qdisc *qdisc = _qdisc;
917
918	rcu_assign_pointer(dev_queue->qdisc, qdisc);
919	dev_queue->qdisc_sleeping = qdisc;
920}
921
922void dev_init_scheduler(struct net_device *dev)
923{
924	dev->qdisc = &noop_qdisc;
925	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
926	if (dev_ingress_queue(dev))
927		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
928
929	setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
930}
931
932static void shutdown_scheduler_queue(struct net_device *dev,
933				     struct netdev_queue *dev_queue,
934				     void *_qdisc_default)
935{
936	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
937	struct Qdisc *qdisc_default = _qdisc_default;
938
939	if (qdisc) {
940		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
941		dev_queue->qdisc_sleeping = qdisc_default;
942
943		qdisc_destroy(qdisc);
944	}
945}
946
947void dev_shutdown(struct net_device *dev)
948{
949	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
950	if (dev_ingress_queue(dev))
951		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
952	qdisc_destroy(dev->qdisc);
953	dev->qdisc = &noop_qdisc;
954
955	WARN_ON(timer_pending(&dev->watchdog_timer));
956}
957
958void psched_ratecfg_precompute(struct psched_ratecfg *r,
959			       const struct tc_ratespec *conf,
960			       u64 rate64)
961{
962	memset(r, 0, sizeof(*r));
963	r->overhead = conf->overhead;
964	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
965	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
966	r->mult = 1;
967	/*
968	 * The deal here is to replace a divide by a reciprocal one
969	 * in fast path (a reciprocal divide is a multiply and a shift)
970	 *
971	 * Normal formula would be :
972	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
973	 *
974	 * We compute mult/shift to use instead :
975	 *  time_in_ns = (len * mult) >> shift;
976	 *
977	 * We try to get the highest possible mult value for accuracy,
978	 * but have to make sure no overflows will ever happen.
979	 */
980	if (r->rate_bytes_ps > 0) {
981		u64 factor = NSEC_PER_SEC;
982
983		for (;;) {
984			r->mult = div64_u64(factor, r->rate_bytes_ps);
985			if (r->mult & (1U << 31) || factor & (1ULL << 63))
986				break;
987			factor <<= 1;
988			r->shift++;
989		}
990	}
991}
992EXPORT_SYMBOL(psched_ratecfg_precompute);
993