Linux Audio

Check our new training course

Loading...
v4.6
 
  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 (likely(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	} else {
169		spin_lock(root_lock);
170		return qdisc_qlen(q);
 
171	}
172	spin_lock(root_lock);
173
174	if (dev_xmit_complete(ret)) {
175		/* Driver sent out skb successfully or skb was consumed */
176		ret = qdisc_qlen(q);
177	} else if (ret == NETDEV_TX_LOCKED) {
178		/* Driver try lock failed */
179		ret = handle_dev_cpu_collision(skb, txq, q);
180	} else {
181		/* Driver returned NETDEV_TX_BUSY - requeue skb */
182		if (unlikely(ret != NETDEV_TX_BUSY))
183			net_warn_ratelimited("BUG %s code %d qlen %d\n",
184					     dev->name, ret, q->q.qlen);
185
186		ret = dev_requeue_skb(skb, q);
 
187	}
188
189	if (ret && netif_xmit_frozen_or_stopped(txq))
190		ret = 0;
191
192	return ret;
193}
194
195/*
196 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
197 *
198 * __QDISC___STATE_RUNNING guarantees only one CPU can process
199 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
200 * this queue.
201 *
202 *  netif_tx_lock serializes accesses to device driver.
203 *
204 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
205 *  if one is grabbed, another must be free.
206 *
207 * Note, that this procedure can be called by a watchdog timer
208 *
209 * Returns to the caller:
210 *				0  - queue is empty or throttled.
211 *				>0 - queue is not empty.
212 *
213 */
214static inline int qdisc_restart(struct Qdisc *q, int *packets)
215{
 
216	struct netdev_queue *txq;
217	struct net_device *dev;
218	spinlock_t *root_lock;
219	struct sk_buff *skb;
220	bool validate;
221
222	/* Dequeue packet */
223	skb = dequeue_skb(q, &validate, packets);
224	if (unlikely(!skb))
225		return 0;
 
 
 
226
227	root_lock = qdisc_lock(q);
228	dev = qdisc_dev(q);
229	txq = skb_get_tx_queue(dev, skb);
230
231	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
232}
233
234void __qdisc_run(struct Qdisc *q)
235{
236	int quota = weight_p;
237	int packets;
238
239	while (qdisc_restart(q, &packets)) {
240		/*
241		 * Ordered by possible occurrence: Postpone processing if
242		 * 1. we've exceeded packet quota
243		 * 2. another process needs the CPU;
244		 */
245		quota -= packets;
246		if (quota <= 0 || need_resched()) {
247			__netif_schedule(q);
 
 
 
 
248			break;
249		}
250	}
251
252	qdisc_run_end(q);
253}
254
255unsigned long dev_trans_start(struct net_device *dev)
256{
257	unsigned long val, res;
258	unsigned int i;
259
260	if (is_vlan_dev(dev))
261		dev = vlan_dev_real_dev(dev);
262	res = dev->trans_start;
263	for (i = 0; i < dev->num_tx_queues; i++) {
 
 
264		val = netdev_get_tx_queue(dev, i)->trans_start;
265		if (val && time_after(val, res))
266			res = val;
267	}
268	dev->trans_start = res;
269
270	return res;
271}
272EXPORT_SYMBOL(dev_trans_start);
273
274static void dev_watchdog(unsigned long arg)
275{
276	struct net_device *dev = (struct net_device *)arg;
277
278	netif_tx_lock(dev);
279	if (!qdisc_tx_is_noop(dev)) {
280		if (netif_device_present(dev) &&
281		    netif_running(dev) &&
282		    netif_carrier_ok(dev)) {
283			int some_queue_timedout = 0;
284			unsigned int i;
285			unsigned long trans_start;
286
287			for (i = 0; i < dev->num_tx_queues; i++) {
288				struct netdev_queue *txq;
289
290				txq = netdev_get_tx_queue(dev, i);
291				/*
292				 * old device drivers set dev->trans_start
293				 */
294				trans_start = txq->trans_start ? : dev->trans_start;
295				if (netif_xmit_stopped(txq) &&
296				    time_after(jiffies, (trans_start +
297							 dev->watchdog_timeo))) {
298					some_queue_timedout = 1;
299					txq->trans_timeout++;
300					break;
301				}
302			}
303
304			if (some_queue_timedout) {
 
305				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
306				       dev->name, netdev_drivername(dev), i);
307				dev->netdev_ops->ndo_tx_timeout(dev);
308			}
309			if (!mod_timer(&dev->watchdog_timer,
310				       round_jiffies(jiffies +
311						     dev->watchdog_timeo)))
312				dev_hold(dev);
313		}
314	}
315	netif_tx_unlock(dev);
316
317	dev_put(dev);
318}
319
320void __netdev_watchdog_up(struct net_device *dev)
321{
322	if (dev->netdev_ops->ndo_tx_timeout) {
323		if (dev->watchdog_timeo <= 0)
324			dev->watchdog_timeo = 5*HZ;
325		if (!mod_timer(&dev->watchdog_timer,
326			       round_jiffies(jiffies + dev->watchdog_timeo)))
327			dev_hold(dev);
328	}
329}
 
330
331static void dev_watchdog_up(struct net_device *dev)
332{
333	__netdev_watchdog_up(dev);
334}
335
336static void dev_watchdog_down(struct net_device *dev)
337{
338	netif_tx_lock_bh(dev);
339	if (del_timer(&dev->watchdog_timer))
340		dev_put(dev);
341	netif_tx_unlock_bh(dev);
342}
343
344/**
345 *	netif_carrier_on - set carrier
346 *	@dev: network device
347 *
348 * Device has detected that carrier.
349 */
350void netif_carrier_on(struct net_device *dev)
351{
352	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
353		if (dev->reg_state == NETREG_UNINITIALIZED)
354			return;
355		atomic_inc(&dev->carrier_changes);
356		linkwatch_fire_event(dev);
357		if (netif_running(dev))
358			__netdev_watchdog_up(dev);
359	}
360}
361EXPORT_SYMBOL(netif_carrier_on);
362
363/**
364 *	netif_carrier_off - clear carrier
365 *	@dev: network device
366 *
367 * Device has detected loss of carrier.
368 */
369void netif_carrier_off(struct net_device *dev)
370{
371	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
372		if (dev->reg_state == NETREG_UNINITIALIZED)
373			return;
374		atomic_inc(&dev->carrier_changes);
375		linkwatch_fire_event(dev);
376	}
377}
378EXPORT_SYMBOL(netif_carrier_off);
379
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
380/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
381   under all circumstances. It is difficult to invent anything faster or
382   cheaper.
383 */
384
385static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 
386{
387	kfree_skb(skb);
388	return NET_XMIT_CN;
389}
390
391static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
392{
393	return NULL;
394}
395
396struct Qdisc_ops noop_qdisc_ops __read_mostly = {
397	.id		=	"noop",
398	.priv_size	=	0,
399	.enqueue	=	noop_enqueue,
400	.dequeue	=	noop_dequeue,
401	.peek		=	noop_dequeue,
402	.owner		=	THIS_MODULE,
403};
404
405static struct netdev_queue noop_netdev_queue = {
406	.qdisc		=	&noop_qdisc,
407	.qdisc_sleeping	=	&noop_qdisc,
408};
409
410struct Qdisc noop_qdisc = {
411	.enqueue	=	noop_enqueue,
412	.dequeue	=	noop_dequeue,
413	.flags		=	TCQ_F_BUILTIN,
414	.ops		=	&noop_qdisc_ops,
415	.list		=	LIST_HEAD_INIT(noop_qdisc.list),
416	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
417	.dev_queue	=	&noop_netdev_queue,
 
418	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 
 
 
 
 
 
 
 
 
 
 
 
419};
420EXPORT_SYMBOL(noop_qdisc);
421
422static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 
423{
424	/* register_qdisc() assigns a default of noop_enqueue if unset,
425	 * but __dev_queue_xmit() treats noqueue only as such
426	 * if this is NULL - so clear it here. */
427	qdisc->enqueue = NULL;
428	return 0;
429}
430
431struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
432	.id		=	"noqueue",
433	.priv_size	=	0,
434	.init		=	noqueue_init,
435	.enqueue	=	noop_enqueue,
436	.dequeue	=	noop_dequeue,
437	.peek		=	noop_dequeue,
438	.owner		=	THIS_MODULE,
439};
440
441static const u8 prio2band[TC_PRIO_MAX + 1] = {
442	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
443};
444
445/* 3-band FIFO queue: old style, but should be a bit faster than
446   generic prio+fifo combination.
447 */
448
449#define PFIFO_FAST_BANDS 3
450
451/*
452 * Private data for a pfifo_fast scheduler containing:
453 * 	- queues for the three band
454 * 	- bitmap indicating which of the bands contain skbs
455 */
456struct pfifo_fast_priv {
457	u32 bitmap;
458	struct sk_buff_head q[PFIFO_FAST_BANDS];
459};
460
461/*
462 * Convert a bitmap to the first band number where an skb is queued, where:
463 * 	bitmap=0 means there are no skbs on any band.
464 * 	bitmap=1 means there is an skb on band 0.
465 *	bitmap=7 means there are skbs on all 3 bands, etc.
466 */
467static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
468
469static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
470					     int band)
471{
472	return priv->q + band;
473}
474
475static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 
476{
477	if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
478		int band = prio2band[skb->priority & TC_PRIO_MAX];
479		struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
480		struct sk_buff_head *list = band2list(priv, band);
 
 
 
481
482		priv->bitmap |= (1 << band);
483		qdisc->q.qlen++;
484		return __qdisc_enqueue_tail(skb, qdisc, list);
 
 
485	}
486
487	return qdisc_drop(skb, qdisc);
 
488}
489
490static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
491{
492	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493	int band = bitmap2band[priv->bitmap];
 
 
 
 
 
 
494
495	if (likely(band >= 0)) {
496		struct sk_buff_head *list = band2list(priv, band);
497		struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
498
499		qdisc->q.qlen--;
500		if (skb_queue_empty(list))
501			priv->bitmap &= ~(1 << band);
502
503		return skb;
504	}
 
 
 
 
 
 
 
 
 
 
 
505
506	return NULL;
 
 
 
 
 
 
 
 
 
 
507}
508
509static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
510{
511	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
512	int band = bitmap2band[priv->bitmap];
 
513
514	if (band >= 0) {
515		struct sk_buff_head *list = band2list(priv, band);
516
517		return skb_peek(list);
518	}
519
520	return NULL;
521}
522
523static void pfifo_fast_reset(struct Qdisc *qdisc)
524{
525	int prio;
526	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
527
528	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
529		__qdisc_reset_queue(qdisc, band2list(priv, prio));
 
530
531	priv->bitmap = 0;
532	qdisc->qstats.backlog = 0;
533	qdisc->q.qlen = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
534}
535
536static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
537{
538	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
539
540	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
541	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
542		goto nla_put_failure;
543	return skb->len;
544
545nla_put_failure:
546	return -1;
547}
548
549static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 
550{
551	int prio;
552	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 
553
554	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
555		__skb_queue_head_init(band2list(priv, prio));
 
 
 
 
 
 
 
 
 
 
556
557	/* Can by-pass the queue discipline */
558	qdisc->flags |= TCQ_F_CAN_BYPASS;
559	return 0;
560}
561
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
562struct Qdisc_ops pfifo_fast_ops __read_mostly = {
563	.id		=	"pfifo_fast",
564	.priv_size	=	sizeof(struct pfifo_fast_priv),
565	.enqueue	=	pfifo_fast_enqueue,
566	.dequeue	=	pfifo_fast_dequeue,
567	.peek		=	pfifo_fast_peek,
568	.init		=	pfifo_fast_init,
 
569	.reset		=	pfifo_fast_reset,
570	.dump		=	pfifo_fast_dump,
 
571	.owner		=	THIS_MODULE,
 
572};
573EXPORT_SYMBOL(pfifo_fast_ops);
574
575static struct lock_class_key qdisc_tx_busylock;
 
576
577struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
578			  const struct Qdisc_ops *ops)
 
579{
580	void *p;
581	struct Qdisc *sch;
582	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
583	int err = -ENOBUFS;
584	struct net_device *dev = dev_queue->dev;
585
586	p = kzalloc_node(size, GFP_KERNEL,
587			 netdev_queue_numa_node_read(dev_queue));
 
 
 
588
589	if (!p)
 
 
 
590		goto errout;
591	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
592	/* if we got non aligned memory, ask more and do alignment ourself */
593	if (sch != p) {
594		kfree(p);
595		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
596				 netdev_queue_numa_node_read(dev_queue));
597		if (!p)
598			goto errout;
599		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
600		sch->padded = (char *) sch - (char *) p;
 
 
 
 
 
 
601	}
602	INIT_LIST_HEAD(&sch->list);
603	skb_queue_head_init(&sch->q);
604
605	spin_lock_init(&sch->busylock);
606	lockdep_set_class(&sch->busylock,
607			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
608
 
 
 
 
 
 
 
 
 
609	sch->ops = ops;
 
610	sch->enqueue = ops->enqueue;
611	sch->dequeue = ops->dequeue;
612	sch->dev_queue = dev_queue;
613	dev_hold(dev);
614	atomic_set(&sch->refcnt, 1);
615
616	return sch;
 
 
617errout:
618	return ERR_PTR(err);
619}
620
621struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
622				const struct Qdisc_ops *ops,
623				unsigned int parentid)
 
624{
625	struct Qdisc *sch;
626
627	if (!try_module_get(ops->owner))
628		goto errout;
 
 
629
630	sch = qdisc_alloc(dev_queue, ops);
631	if (IS_ERR(sch))
632		goto errout;
 
 
633	sch->parent = parentid;
634
635	if (!ops->init || ops->init(sch, NULL) == 0)
 
636		return sch;
 
637
638	qdisc_destroy(sch);
639errout:
640	return NULL;
641}
642EXPORT_SYMBOL(qdisc_create_dflt);
643
644/* Under qdisc_lock(qdisc) and BH! */
645
646void qdisc_reset(struct Qdisc *qdisc)
647{
648	const struct Qdisc_ops *ops = qdisc->ops;
 
 
 
649
650	if (ops->reset)
651		ops->reset(qdisc);
652
653	if (qdisc->gso_skb) {
654		kfree_skb_list(qdisc->gso_skb);
655		qdisc->gso_skb = NULL;
656		qdisc->q.qlen = 0;
657	}
 
 
 
 
 
 
 
 
658}
659EXPORT_SYMBOL(qdisc_reset);
660
661static void qdisc_rcu_free(struct rcu_head *head)
662{
663	struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
664
665	if (qdisc_is_percpu_stats(qdisc)) {
666		free_percpu(qdisc->cpu_bstats);
667		free_percpu(qdisc->cpu_qstats);
668	}
669
670	kfree((char *) qdisc - qdisc->padded);
671}
672
673void qdisc_destroy(struct Qdisc *qdisc)
674{
675	const struct Qdisc_ops  *ops = qdisc->ops;
676
677	if (qdisc->flags & TCQ_F_BUILTIN ||
678	    !atomic_dec_and_test(&qdisc->refcnt))
679		return;
 
 
 
680
681#ifdef CONFIG_NET_SCHED
682	qdisc_list_del(qdisc);
683
684	qdisc_put_stab(rtnl_dereference(qdisc->stab));
685#endif
686	gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
687	if (ops->reset)
688		ops->reset(qdisc);
 
689	if (ops->destroy)
690		ops->destroy(qdisc);
691
692	module_put(ops->owner);
693	dev_put(qdisc_dev(qdisc));
694
695	kfree_skb_list(qdisc->gso_skb);
696	/*
697	 * gen_estimator est_timer() might access qdisc->q.lock,
698	 * wait a RCU grace period before freeing qdisc.
699	 */
700	call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 
 
 
 
 
 
 
 
 
701}
702EXPORT_SYMBOL(qdisc_destroy);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
703
704/* Attach toplevel qdisc to device queue. */
705struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
706			      struct Qdisc *qdisc)
707{
708	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
709	spinlock_t *root_lock;
710
711	root_lock = qdisc_lock(oqdisc);
712	spin_lock_bh(root_lock);
713
714	/* Prune old scheduler */
715	if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
716		qdisc_reset(oqdisc);
717
718	/* ... and graft new one */
719	if (qdisc == NULL)
720		qdisc = &noop_qdisc;
721	dev_queue->qdisc_sleeping = qdisc;
722	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
723
724	spin_unlock_bh(root_lock);
725
726	return oqdisc;
727}
728EXPORT_SYMBOL(dev_graft_qdisc);
729
730static void attach_one_default_qdisc(struct net_device *dev,
731				     struct netdev_queue *dev_queue,
732				     void *_unused)
733{
734	struct Qdisc *qdisc;
735	const struct Qdisc_ops *ops = default_qdisc_ops;
736
737	if (dev->priv_flags & IFF_NO_QUEUE)
738		ops = &noqueue_qdisc_ops;
 
 
739
740	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
741	if (!qdisc) {
742		netdev_info(dev, "activation failed\n");
743		return;
744	}
745	if (!netif_is_multiqueue(dev))
746		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
747	dev_queue->qdisc_sleeping = qdisc;
748}
749
750static void attach_default_qdiscs(struct net_device *dev)
751{
752	struct netdev_queue *txq;
753	struct Qdisc *qdisc;
754
755	txq = netdev_get_tx_queue(dev, 0);
756
757	if (!netif_is_multiqueue(dev) ||
758	    dev->priv_flags & IFF_NO_QUEUE) {
759		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
760		dev->qdisc = txq->qdisc_sleeping;
761		atomic_inc(&dev->qdisc->refcnt);
762	} else {
763		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
764		if (qdisc) {
765			dev->qdisc = qdisc;
766			qdisc->ops->attach(qdisc);
767		}
768	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
769}
770
771static void transition_one_qdisc(struct net_device *dev,
772				 struct netdev_queue *dev_queue,
773				 void *_need_watchdog)
774{
775	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
776	int *need_watchdog_p = _need_watchdog;
777
778	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
779		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
780
781	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
782	if (need_watchdog_p) {
783		dev_queue->trans_start = 0;
784		*need_watchdog_p = 1;
785	}
786}
787
788void dev_activate(struct net_device *dev)
789{
790	int need_watchdog;
791
792	/* No queueing discipline is attached to device;
793	 * create default one for devices, which need queueing
794	 * and noqueue_qdisc for virtual interfaces
795	 */
796
797	if (dev->qdisc == &noop_qdisc)
798		attach_default_qdiscs(dev);
799
800	if (!netif_carrier_ok(dev))
801		/* Delay activation until next carrier-on event */
802		return;
803
804	need_watchdog = 0;
805	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
806	if (dev_ingress_queue(dev))
807		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
808
809	if (need_watchdog) {
810		dev->trans_start = jiffies;
811		dev_watchdog_up(dev);
812	}
813}
814EXPORT_SYMBOL(dev_activate);
815
 
 
 
 
 
 
 
 
816static void dev_deactivate_queue(struct net_device *dev,
817				 struct netdev_queue *dev_queue,
818				 void *_qdisc_default)
819{
820	struct Qdisc *qdisc_default = _qdisc_default;
821	struct Qdisc *qdisc;
822
823	qdisc = rtnl_dereference(dev_queue->qdisc);
824	if (qdisc) {
825		spin_lock_bh(qdisc_lock(qdisc));
 
 
 
826
827		if (!(qdisc->flags & TCQ_F_BUILTIN))
828			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 
 
 
 
829
830		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
831		qdisc_reset(qdisc);
 
 
 
832
833		spin_unlock_bh(qdisc_lock(qdisc));
 
 
 
 
 
 
 
 
 
 
834	}
835}
836
837static bool some_qdisc_is_busy(struct net_device *dev)
838{
839	unsigned int i;
840
841	for (i = 0; i < dev->num_tx_queues; i++) {
842		struct netdev_queue *dev_queue;
843		spinlock_t *root_lock;
844		struct Qdisc *q;
845		int val;
846
847		dev_queue = netdev_get_tx_queue(dev, i);
848		q = dev_queue->qdisc_sleeping;
849		root_lock = qdisc_lock(q);
850
 
851		spin_lock_bh(root_lock);
852
853		val = (qdisc_is_running(q) ||
854		       test_bit(__QDISC_STATE_SCHED, &q->state));
855
856		spin_unlock_bh(root_lock);
857
858		if (val)
859			return true;
860	}
861	return false;
862}
863
864/**
865 * 	dev_deactivate_many - deactivate transmissions on several devices
866 * 	@head: list of devices to deactivate
867 *
868 *	This function returns only when all outstanding transmissions
869 *	have completed, unless all devices are in dismantle phase.
870 */
871void dev_deactivate_many(struct list_head *head)
872{
873	struct net_device *dev;
874	bool sync_needed = false;
875
876	list_for_each_entry(dev, head, close_list) {
877		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
878					 &noop_qdisc);
879		if (dev_ingress_queue(dev))
880			dev_deactivate_queue(dev, dev_ingress_queue(dev),
881					     &noop_qdisc);
882
883		dev_watchdog_down(dev);
884		sync_needed |= !dev->dismantle;
885	}
886
887	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
 
888	 * This is avoided if all devices are in dismantle phase :
889	 * Caller will call synchronize_net() for us
890	 */
891	if (sync_needed)
892		synchronize_net();
 
 
 
 
 
 
893
894	/* Wait for outstanding qdisc_run calls. */
895	list_for_each_entry(dev, head, close_list)
896		while (some_qdisc_is_busy(dev))
897			yield();
 
 
 
 
 
 
898}
899
900void dev_deactivate(struct net_device *dev)
901{
902	LIST_HEAD(single);
903
904	list_add(&dev->close_list, &single);
905	dev_deactivate_many(&single);
906	list_del(&single);
907}
908EXPORT_SYMBOL(dev_deactivate);
909
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
910static void dev_init_scheduler_queue(struct net_device *dev,
911				     struct netdev_queue *dev_queue,
912				     void *_qdisc)
913{
914	struct Qdisc *qdisc = _qdisc;
915
916	rcu_assign_pointer(dev_queue->qdisc, qdisc);
917	dev_queue->qdisc_sleeping = qdisc;
918}
919
920void dev_init_scheduler(struct net_device *dev)
921{
922	dev->qdisc = &noop_qdisc;
923	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
924	if (dev_ingress_queue(dev))
925		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
926
927	setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
928}
929
930static void shutdown_scheduler_queue(struct net_device *dev,
931				     struct netdev_queue *dev_queue,
932				     void *_qdisc_default)
933{
934	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
935	struct Qdisc *qdisc_default = _qdisc_default;
936
937	if (qdisc) {
938		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
939		dev_queue->qdisc_sleeping = qdisc_default;
940
941		qdisc_destroy(qdisc);
942	}
943}
944
945void dev_shutdown(struct net_device *dev)
946{
947	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
948	if (dev_ingress_queue(dev))
949		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
950	qdisc_destroy(dev->qdisc);
951	dev->qdisc = &noop_qdisc;
952
953	WARN_ON(timer_pending(&dev->watchdog_timer));
954}
955
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
956void psched_ratecfg_precompute(struct psched_ratecfg *r,
957			       const struct tc_ratespec *conf,
958			       u64 rate64)
959{
960	memset(r, 0, sizeof(*r));
961	r->overhead = conf->overhead;
962	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
963	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
964	r->mult = 1;
965	/*
966	 * The deal here is to replace a divide by a reciprocal one
967	 * in fast path (a reciprocal divide is a multiply and a shift)
968	 *
969	 * Normal formula would be :
970	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
971	 *
972	 * We compute mult/shift to use instead :
973	 *  time_in_ns = (len * mult) >> shift;
974	 *
975	 * We try to get the highest possible mult value for accuracy,
976	 * but have to make sure no overflows will ever happen.
977	 */
978	if (r->rate_bytes_ps > 0) {
979		u64 factor = NSEC_PER_SEC;
980
981		for (;;) {
982			r->mult = div64_u64(factor, r->rate_bytes_ps);
983			if (r->mult & (1U << 31) || factor & (1ULL << 63))
984				break;
985			factor <<= 1;
986			r->shift++;
987		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
988	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
989}
990EXPORT_SYMBOL(psched_ratecfg_precompute);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * net/sched/sch_generic.c	Generic packet scheduler routines.
   4 *
 
 
 
 
 
   5 * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   6 *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
   7 *              - Ingress support
   8 */
   9
  10#include <linux/bitops.h>
  11#include <linux/module.h>
  12#include <linux/types.h>
  13#include <linux/kernel.h>
  14#include <linux/sched.h>
  15#include <linux/string.h>
  16#include <linux/errno.h>
  17#include <linux/netdevice.h>
  18#include <linux/skbuff.h>
  19#include <linux/rtnetlink.h>
  20#include <linux/init.h>
  21#include <linux/rcupdate.h>
  22#include <linux/list.h>
  23#include <linux/slab.h>
  24#include <linux/if_vlan.h>
  25#include <linux/skb_array.h>
  26#include <linux/if_macvlan.h>
  27#include <net/sch_generic.h>
  28#include <net/pkt_sched.h>
  29#include <net/dst.h>
  30#include <trace/events/qdisc.h>
  31#include <trace/events/net.h>
  32#include <net/xfrm.h>
  33
  34/* Qdisc to use by default */
  35const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  36EXPORT_SYMBOL(default_qdisc_ops);
  37
  38static void qdisc_maybe_clear_missed(struct Qdisc *q,
  39				     const struct netdev_queue *txq)
  40{
  41	clear_bit(__QDISC_STATE_MISSED, &q->state);
  42
  43	/* Make sure the below netif_xmit_frozen_or_stopped()
  44	 * checking happens after clearing STATE_MISSED.
  45	 */
  46	smp_mb__after_atomic();
  47
  48	/* Checking netif_xmit_frozen_or_stopped() again to
  49	 * make sure STATE_MISSED is set if the STATE_MISSED
  50	 * set by netif_tx_wake_queue()'s rescheduling of
  51	 * net_tx_action() is cleared by the above clear_bit().
  52	 */
  53	if (!netif_xmit_frozen_or_stopped(txq))
  54		set_bit(__QDISC_STATE_MISSED, &q->state);
  55	else
  56		set_bit(__QDISC_STATE_DRAINING, &q->state);
  57}
  58
  59/* Main transmission queue. */
  60
  61/* Modifications to data participating in scheduling must be protected with
  62 * qdisc_lock(qdisc) spinlock.
  63 *
  64 * The idea is the following:
  65 * - enqueue, dequeue are serialized via qdisc root lock
  66 * - ingress filtering is also serialized via qdisc root lock
  67 * - updates to tree and tree walking are only done under the rtnl mutex.
  68 */
  69
  70#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  71
  72static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  73{
  74	const struct netdev_queue *txq = q->dev_queue;
  75	spinlock_t *lock = NULL;
  76	struct sk_buff *skb;
 
  77
  78	if (q->flags & TCQ_F_NOLOCK) {
  79		lock = qdisc_lock(q);
  80		spin_lock(lock);
  81	}
  82
  83	skb = skb_peek(&q->skb_bad_txq);
  84	if (skb) {
  85		/* check the reason of requeuing without tx lock first */
  86		txq = skb_get_tx_queue(txq->dev, skb);
  87		if (!netif_xmit_frozen_or_stopped(txq)) {
  88			skb = __skb_dequeue(&q->skb_bad_txq);
  89			if (qdisc_is_percpu_stats(q)) {
  90				qdisc_qstats_cpu_backlog_dec(q, skb);
  91				qdisc_qstats_cpu_qlen_dec(q);
  92			} else {
  93				qdisc_qstats_backlog_dec(q, skb);
  94				q->q.qlen--;
  95			}
  96		} else {
  97			skb = SKB_XOFF_MAGIC;
  98			qdisc_maybe_clear_missed(q, txq);
  99		}
 100	}
 101
 102	if (lock)
 103		spin_unlock(lock);
 104
 105	return skb;
 106}
 107
 108static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
 109{
 110	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
 111
 112	if (unlikely(skb))
 113		skb = __skb_dequeue_bad_txq(q);
 114
 115	return skb;
 116}
 117
 118static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
 119					     struct sk_buff *skb)
 120{
 121	spinlock_t *lock = NULL;
 122
 123	if (q->flags & TCQ_F_NOLOCK) {
 124		lock = qdisc_lock(q);
 125		spin_lock(lock);
 126	}
 127
 128	__skb_queue_tail(&q->skb_bad_txq, skb);
 129
 130	if (qdisc_is_percpu_stats(q)) {
 131		qdisc_qstats_cpu_backlog_inc(q, skb);
 132		qdisc_qstats_cpu_qlen_inc(q);
 133	} else {
 134		qdisc_qstats_backlog_inc(q, skb);
 135		q->q.qlen++;
 136	}
 137
 138	if (lock)
 139		spin_unlock(lock);
 140}
 141
 142static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 143{
 144	spinlock_t *lock = NULL;
 145
 146	if (q->flags & TCQ_F_NOLOCK) {
 147		lock = qdisc_lock(q);
 148		spin_lock(lock);
 149	}
 150
 151	while (skb) {
 152		struct sk_buff *next = skb->next;
 153
 154		__skb_queue_tail(&q->gso_skb, skb);
 155
 156		/* it's still part of the queue */
 157		if (qdisc_is_percpu_stats(q)) {
 158			qdisc_qstats_cpu_requeues_inc(q);
 159			qdisc_qstats_cpu_backlog_inc(q, skb);
 160			qdisc_qstats_cpu_qlen_inc(q);
 161		} else {
 162			q->qstats.requeues++;
 163			qdisc_qstats_backlog_inc(q, skb);
 164			q->q.qlen++;
 165		}
 166
 167		skb = next;
 168	}
 169
 170	if (lock) {
 171		spin_unlock(lock);
 172		set_bit(__QDISC_STATE_MISSED, &q->state);
 173	} else {
 174		__netif_schedule(q);
 175	}
 176}
 177
 178static void try_bulk_dequeue_skb(struct Qdisc *q,
 179				 struct sk_buff *skb,
 180				 const struct netdev_queue *txq,
 181				 int *packets)
 182{
 183	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 184
 185	while (bytelimit > 0) {
 186		struct sk_buff *nskb = q->dequeue(q);
 187
 188		if (!nskb)
 189			break;
 190
 191		bytelimit -= nskb->len; /* covers GSO len */
 192		skb->next = nskb;
 193		skb = nskb;
 194		(*packets)++; /* GSO counts as one pkt */
 195	}
 196	skb_mark_not_on_list(skb);
 197}
 198
 199/* This variant of try_bulk_dequeue_skb() makes sure
 200 * all skbs in the chain are for the same txq
 201 */
 202static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 203				      struct sk_buff *skb,
 204				      int *packets)
 205{
 206	int mapping = skb_get_queue_mapping(skb);
 207	struct sk_buff *nskb;
 208	int cnt = 0;
 209
 210	do {
 211		nskb = q->dequeue(q);
 212		if (!nskb)
 213			break;
 214		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 215			qdisc_enqueue_skb_bad_txq(q, nskb);
 216			break;
 217		}
 218		skb->next = nskb;
 219		skb = nskb;
 220	} while (++cnt < 8);
 221	(*packets) += cnt;
 222	skb_mark_not_on_list(skb);
 223}
 224
 225/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 226 * A requeued skb (via q->gso_skb) can also be a SKB list.
 227 */
 228static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 229				   int *packets)
 230{
 
 231	const struct netdev_queue *txq = q->dev_queue;
 232	struct sk_buff *skb = NULL;
 233
 234	*packets = 1;
 235	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 236		spinlock_t *lock = NULL;
 237
 238		if (q->flags & TCQ_F_NOLOCK) {
 239			lock = qdisc_lock(q);
 240			spin_lock(lock);
 241		}
 242
 243		skb = skb_peek(&q->gso_skb);
 244
 245		/* skb may be null if another cpu pulls gso_skb off in between
 246		 * empty check and lock.
 247		 */
 248		if (!skb) {
 249			if (lock)
 250				spin_unlock(lock);
 251			goto validate;
 252		}
 253
 254		/* skb in gso_skb were already validated */
 255		*validate = false;
 256		if (xfrm_offload(skb))
 257			*validate = true;
 258		/* check the reason of requeuing without tx lock first */
 259		txq = skb_get_tx_queue(txq->dev, skb);
 260		if (!netif_xmit_frozen_or_stopped(txq)) {
 261			skb = __skb_dequeue(&q->gso_skb);
 262			if (qdisc_is_percpu_stats(q)) {
 263				qdisc_qstats_cpu_backlog_dec(q, skb);
 264				qdisc_qstats_cpu_qlen_dec(q);
 265			} else {
 266				qdisc_qstats_backlog_dec(q, skb);
 267				q->q.qlen--;
 268			}
 269		} else {
 270			skb = NULL;
 271			qdisc_maybe_clear_missed(q, txq);
 
 
 
 
 
 
 
 272		}
 273		if (lock)
 274			spin_unlock(lock);
 275		goto trace;
 276	}
 277validate:
 278	*validate = true;
 279
 280	if ((q->flags & TCQ_F_ONETXQUEUE) &&
 281	    netif_xmit_frozen_or_stopped(txq)) {
 282		qdisc_maybe_clear_missed(q, txq);
 283		return skb;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 284	}
 285
 286	skb = qdisc_dequeue_skb_bad_txq(q);
 287	if (unlikely(skb)) {
 288		if (skb == SKB_XOFF_MAGIC)
 289			return NULL;
 290		goto bulk;
 291	}
 292	skb = q->dequeue(q);
 293	if (skb) {
 294bulk:
 295		if (qdisc_may_bulk(q))
 296			try_bulk_dequeue_skb(q, skb, txq, packets);
 297		else
 298			try_bulk_dequeue_skb_slow(q, skb, packets);
 299	}
 300trace:
 301	trace_qdisc_dequeue(q, txq, *packets, skb);
 302	return skb;
 303}
 304
 305/*
 306 * Transmit possibly several skbs, and handle the return status as
 307 * required. Owning running seqcount bit guarantees that
 308 * only one CPU can execute this function.
 309 *
 310 * Returns to the caller:
 311 *				false  - hardware queue frozen backoff
 312 *				true   - feel free to send more pkts
 313 */
 314bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 315		     struct net_device *dev, struct netdev_queue *txq,
 316		     spinlock_t *root_lock, bool validate)
 317{
 318	int ret = NETDEV_TX_BUSY;
 319	bool again = false;
 320
 321	/* And release qdisc */
 322	if (root_lock)
 323		spin_unlock(root_lock);
 324
 325	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
 326	if (validate)
 327		skb = validate_xmit_skb_list(skb, dev, &again);
 328
 329#ifdef CONFIG_XFRM_OFFLOAD
 330	if (unlikely(again)) {
 331		if (root_lock)
 332			spin_lock(root_lock);
 333
 334		dev_requeue_skb(skb, q);
 335		return false;
 336	}
 337#endif
 338
 339	if (likely(skb)) {
 340		HARD_TX_LOCK(dev, txq, smp_processor_id());
 341		if (!netif_xmit_frozen_or_stopped(txq))
 342			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 343		else
 344			qdisc_maybe_clear_missed(q, txq);
 345
 346		HARD_TX_UNLOCK(dev, txq);
 347	} else {
 348		if (root_lock)
 349			spin_lock(root_lock);
 350		return true;
 351	}
 
 352
 353	if (root_lock)
 354		spin_lock(root_lock);
 355
 356	if (!dev_xmit_complete(ret)) {
 
 
 
 357		/* Driver returned NETDEV_TX_BUSY - requeue skb */
 358		if (unlikely(ret != NETDEV_TX_BUSY))
 359			net_warn_ratelimited("BUG %s code %d qlen %d\n",
 360					     dev->name, ret, q->q.qlen);
 361
 362		dev_requeue_skb(skb, q);
 363		return false;
 364	}
 365
 366	return true;
 
 
 
 367}
 368
 369/*
 370 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 371 *
 372 * running seqcount guarantees only one CPU can process
 373 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 374 * this queue.
 375 *
 376 *  netif_tx_lock serializes accesses to device driver.
 377 *
 378 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 379 *  if one is grabbed, another must be free.
 380 *
 381 * Note, that this procedure can be called by a watchdog timer
 382 *
 383 * Returns to the caller:
 384 *				0  - queue is empty or throttled.
 385 *				>0 - queue is not empty.
 386 *
 387 */
 388static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 389{
 390	spinlock_t *root_lock = NULL;
 391	struct netdev_queue *txq;
 392	struct net_device *dev;
 
 393	struct sk_buff *skb;
 394	bool validate;
 395
 396	/* Dequeue packet */
 397	skb = dequeue_skb(q, &validate, packets);
 398	if (unlikely(!skb))
 399		return false;
 400
 401	if (!(q->flags & TCQ_F_NOLOCK))
 402		root_lock = qdisc_lock(q);
 403
 
 404	dev = qdisc_dev(q);
 405	txq = skb_get_tx_queue(dev, skb);
 406
 407	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 408}
 409
 410void __qdisc_run(struct Qdisc *q)
 411{
 412	int quota = dev_tx_weight;
 413	int packets;
 414
 415	while (qdisc_restart(q, &packets)) {
 
 
 
 
 
 416		quota -= packets;
 417		if (quota <= 0) {
 418			if (q->flags & TCQ_F_NOLOCK)
 419				set_bit(__QDISC_STATE_MISSED, &q->state);
 420			else
 421				__netif_schedule(q);
 422
 423			break;
 424		}
 425	}
 
 
 426}
 427
 428unsigned long dev_trans_start(struct net_device *dev)
 429{
 430	unsigned long val, res;
 431	unsigned int i;
 432
 433	if (is_vlan_dev(dev))
 434		dev = vlan_dev_real_dev(dev);
 435	else if (netif_is_macvlan(dev))
 436		dev = macvlan_dev_real_dev(dev);
 437	res = netdev_get_tx_queue(dev, 0)->trans_start;
 438	for (i = 1; i < dev->num_tx_queues; i++) {
 439		val = netdev_get_tx_queue(dev, i)->trans_start;
 440		if (val && time_after(val, res))
 441			res = val;
 442	}
 
 443
 444	return res;
 445}
 446EXPORT_SYMBOL(dev_trans_start);
 447
 448static void dev_watchdog(struct timer_list *t)
 449{
 450	struct net_device *dev = from_timer(dev, t, watchdog_timer);
 451
 452	netif_tx_lock(dev);
 453	if (!qdisc_tx_is_noop(dev)) {
 454		if (netif_device_present(dev) &&
 455		    netif_running(dev) &&
 456		    netif_carrier_ok(dev)) {
 457			int some_queue_timedout = 0;
 458			unsigned int i;
 459			unsigned long trans_start;
 460
 461			for (i = 0; i < dev->num_tx_queues; i++) {
 462				struct netdev_queue *txq;
 463
 464				txq = netdev_get_tx_queue(dev, i);
 465				trans_start = txq->trans_start;
 
 
 
 466				if (netif_xmit_stopped(txq) &&
 467				    time_after(jiffies, (trans_start +
 468							 dev->watchdog_timeo))) {
 469					some_queue_timedout = 1;
 470					txq->trans_timeout++;
 471					break;
 472				}
 473			}
 474
 475			if (some_queue_timedout) {
 476				trace_net_dev_xmit_timeout(dev, i);
 477				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 478				       dev->name, netdev_drivername(dev), i);
 479				dev->netdev_ops->ndo_tx_timeout(dev, i);
 480			}
 481			if (!mod_timer(&dev->watchdog_timer,
 482				       round_jiffies(jiffies +
 483						     dev->watchdog_timeo)))
 484				dev_hold(dev);
 485		}
 486	}
 487	netif_tx_unlock(dev);
 488
 489	dev_put(dev);
 490}
 491
 492void __netdev_watchdog_up(struct net_device *dev)
 493{
 494	if (dev->netdev_ops->ndo_tx_timeout) {
 495		if (dev->watchdog_timeo <= 0)
 496			dev->watchdog_timeo = 5*HZ;
 497		if (!mod_timer(&dev->watchdog_timer,
 498			       round_jiffies(jiffies + dev->watchdog_timeo)))
 499			dev_hold(dev);
 500	}
 501}
 502EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 503
 504static void dev_watchdog_up(struct net_device *dev)
 505{
 506	__netdev_watchdog_up(dev);
 507}
 508
 509static void dev_watchdog_down(struct net_device *dev)
 510{
 511	netif_tx_lock_bh(dev);
 512	if (del_timer(&dev->watchdog_timer))
 513		dev_put(dev);
 514	netif_tx_unlock_bh(dev);
 515}
 516
 517/**
 518 *	netif_carrier_on - set carrier
 519 *	@dev: network device
 520 *
 521 * Device has detected acquisition of carrier.
 522 */
 523void netif_carrier_on(struct net_device *dev)
 524{
 525	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 526		if (dev->reg_state == NETREG_UNINITIALIZED)
 527			return;
 528		atomic_inc(&dev->carrier_up_count);
 529		linkwatch_fire_event(dev);
 530		if (netif_running(dev))
 531			__netdev_watchdog_up(dev);
 532	}
 533}
 534EXPORT_SYMBOL(netif_carrier_on);
 535
 536/**
 537 *	netif_carrier_off - clear carrier
 538 *	@dev: network device
 539 *
 540 * Device has detected loss of carrier.
 541 */
 542void netif_carrier_off(struct net_device *dev)
 543{
 544	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 545		if (dev->reg_state == NETREG_UNINITIALIZED)
 546			return;
 547		atomic_inc(&dev->carrier_down_count);
 548		linkwatch_fire_event(dev);
 549	}
 550}
 551EXPORT_SYMBOL(netif_carrier_off);
 552
 553/**
 554 *	netif_carrier_event - report carrier state event
 555 *	@dev: network device
 556 *
 557 * Device has detected a carrier event but the carrier state wasn't changed.
 558 * Use in drivers when querying carrier state asynchronously, to avoid missing
 559 * events (link flaps) if link recovers before it's queried.
 560 */
 561void netif_carrier_event(struct net_device *dev)
 562{
 563	if (dev->reg_state == NETREG_UNINITIALIZED)
 564		return;
 565	atomic_inc(&dev->carrier_up_count);
 566	atomic_inc(&dev->carrier_down_count);
 567	linkwatch_fire_event(dev);
 568}
 569EXPORT_SYMBOL_GPL(netif_carrier_event);
 570
 571/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 572   under all circumstances. It is difficult to invent anything faster or
 573   cheaper.
 574 */
 575
 576static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 577			struct sk_buff **to_free)
 578{
 579	__qdisc_drop(skb, to_free);
 580	return NET_XMIT_CN;
 581}
 582
 583static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 584{
 585	return NULL;
 586}
 587
 588struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 589	.id		=	"noop",
 590	.priv_size	=	0,
 591	.enqueue	=	noop_enqueue,
 592	.dequeue	=	noop_dequeue,
 593	.peek		=	noop_dequeue,
 594	.owner		=	THIS_MODULE,
 595};
 596
 597static struct netdev_queue noop_netdev_queue = {
 598	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 599	.qdisc_sleeping	=	&noop_qdisc,
 600};
 601
 602struct Qdisc noop_qdisc = {
 603	.enqueue	=	noop_enqueue,
 604	.dequeue	=	noop_dequeue,
 605	.flags		=	TCQ_F_BUILTIN,
 606	.ops		=	&noop_qdisc_ops,
 
 607	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 608	.dev_queue	=	&noop_netdev_queue,
 609	.running	=	SEQCNT_ZERO(noop_qdisc.running),
 610	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 611	.gso_skb = {
 612		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
 613		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 614		.qlen = 0,
 615		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 616	},
 617	.skb_bad_txq = {
 618		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 619		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 620		.qlen = 0,
 621		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 622	},
 623};
 624EXPORT_SYMBOL(noop_qdisc);
 625
 626static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 627			struct netlink_ext_ack *extack)
 628{
 629	/* register_qdisc() assigns a default of noop_enqueue if unset,
 630	 * but __dev_queue_xmit() treats noqueue only as such
 631	 * if this is NULL - so clear it here. */
 632	qdisc->enqueue = NULL;
 633	return 0;
 634}
 635
 636struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 637	.id		=	"noqueue",
 638	.priv_size	=	0,
 639	.init		=	noqueue_init,
 640	.enqueue	=	noop_enqueue,
 641	.dequeue	=	noop_dequeue,
 642	.peek		=	noop_dequeue,
 643	.owner		=	THIS_MODULE,
 644};
 645
 646static const u8 prio2band[TC_PRIO_MAX + 1] = {
 647	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 648};
 649
 650/* 3-band FIFO queue: old style, but should be a bit faster than
 651   generic prio+fifo combination.
 652 */
 653
 654#define PFIFO_FAST_BANDS 3
 655
 656/*
 657 * Private data for a pfifo_fast scheduler containing:
 658 *	- rings for priority bands
 
 659 */
 660struct pfifo_fast_priv {
 661	struct skb_array q[PFIFO_FAST_BANDS];
 
 662};
 663
 664static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 665					  int band)
 
 
 
 
 
 
 
 
 666{
 667	return &priv->q[band];
 668}
 669
 670static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 671			      struct sk_buff **to_free)
 672{
 673	int band = prio2band[skb->priority & TC_PRIO_MAX];
 674	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 675	struct skb_array *q = band2list(priv, band);
 676	unsigned int pkt_len = qdisc_pkt_len(skb);
 677	int err;
 678
 679	err = skb_array_produce(q, skb);
 680
 681	if (unlikely(err)) {
 682		if (qdisc_is_percpu_stats(qdisc))
 683			return qdisc_drop_cpu(skb, qdisc, to_free);
 684		else
 685			return qdisc_drop(skb, qdisc, to_free);
 686	}
 687
 688	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 689	return NET_XMIT_SUCCESS;
 690}
 691
 692static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 693{
 694	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 695	struct sk_buff *skb = NULL;
 696	bool need_retry = true;
 697	int band;
 698
 699retry:
 700	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 701		struct skb_array *q = band2list(priv, band);
 702
 703		if (__skb_array_empty(q))
 704			continue;
 
 
 
 
 
 705
 706		skb = __skb_array_consume(q);
 707	}
 708	if (likely(skb)) {
 709		qdisc_update_stats_at_dequeue(qdisc, skb);
 710	} else if (need_retry &&
 711		   READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
 712		/* Delay clearing the STATE_MISSED here to reduce
 713		 * the overhead of the second spin_trylock() in
 714		 * qdisc_run_begin() and __netif_schedule() calling
 715		 * in qdisc_run_end().
 716		 */
 717		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
 718		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
 719
 720		/* Make sure dequeuing happens after clearing
 721		 * STATE_MISSED.
 722		 */
 723		smp_mb__after_atomic();
 724
 725		need_retry = false;
 726
 727		goto retry;
 728	}
 729
 730	return skb;
 731}
 732
 733static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 734{
 735	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 736	struct sk_buff *skb = NULL;
 737	int band;
 738
 739	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 740		struct skb_array *q = band2list(priv, band);
 741
 742		skb = __skb_array_peek(q);
 743	}
 744
 745	return skb;
 746}
 747
 748static void pfifo_fast_reset(struct Qdisc *qdisc)
 749{
 750	int i, band;
 751	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 752
 753	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 754		struct skb_array *q = band2list(priv, band);
 755		struct sk_buff *skb;
 756
 757		/* NULL ring is possible if destroy path is due to a failed
 758		 * skb_array_init() in pfifo_fast_init() case.
 759		 */
 760		if (!q->ring.queue)
 761			continue;
 762
 763		while ((skb = __skb_array_consume(q)) != NULL)
 764			kfree_skb(skb);
 765	}
 766
 767	if (qdisc_is_percpu_stats(qdisc)) {
 768		for_each_possible_cpu(i) {
 769			struct gnet_stats_queue *q;
 770
 771			q = per_cpu_ptr(qdisc->cpu_qstats, i);
 772			q->backlog = 0;
 773			q->qlen = 0;
 774		}
 775	}
 776}
 777
 778static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 779{
 780	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 781
 782	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 783	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 784		goto nla_put_failure;
 785	return skb->len;
 786
 787nla_put_failure:
 788	return -1;
 789}
 790
 791static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 792			   struct netlink_ext_ack *extack)
 793{
 794	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 795	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 796	int prio;
 797
 798	/* guard against zero length rings */
 799	if (!qlen)
 800		return -EINVAL;
 801
 802	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 803		struct skb_array *q = band2list(priv, prio);
 804		int err;
 805
 806		err = skb_array_init(q, qlen, GFP_KERNEL);
 807		if (err)
 808			return -ENOMEM;
 809	}
 810
 811	/* Can by-pass the queue discipline */
 812	qdisc->flags |= TCQ_F_CAN_BYPASS;
 813	return 0;
 814}
 815
 816static void pfifo_fast_destroy(struct Qdisc *sch)
 817{
 818	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 819	int prio;
 820
 821	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 822		struct skb_array *q = band2list(priv, prio);
 823
 824		/* NULL ring is possible if destroy path is due to a failed
 825		 * skb_array_init() in pfifo_fast_init() case.
 826		 */
 827		if (!q->ring.queue)
 828			continue;
 829		/* Destroy ring but no need to kfree_skb because a call to
 830		 * pfifo_fast_reset() has already done that work.
 831		 */
 832		ptr_ring_cleanup(&q->ring, NULL);
 833	}
 834}
 835
 836static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 837					  unsigned int new_len)
 838{
 839	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 840	struct skb_array *bands[PFIFO_FAST_BANDS];
 841	int prio;
 842
 843	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 844		struct skb_array *q = band2list(priv, prio);
 845
 846		bands[prio] = q;
 847	}
 848
 849	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 850					 GFP_KERNEL);
 851}
 852
 853struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 854	.id		=	"pfifo_fast",
 855	.priv_size	=	sizeof(struct pfifo_fast_priv),
 856	.enqueue	=	pfifo_fast_enqueue,
 857	.dequeue	=	pfifo_fast_dequeue,
 858	.peek		=	pfifo_fast_peek,
 859	.init		=	pfifo_fast_init,
 860	.destroy	=	pfifo_fast_destroy,
 861	.reset		=	pfifo_fast_reset,
 862	.dump		=	pfifo_fast_dump,
 863	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 864	.owner		=	THIS_MODULE,
 865	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 866};
 867EXPORT_SYMBOL(pfifo_fast_ops);
 868
 869static struct lock_class_key qdisc_tx_busylock;
 870static struct lock_class_key qdisc_running_key;
 871
 872struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 873			  const struct Qdisc_ops *ops,
 874			  struct netlink_ext_ack *extack)
 875{
 
 876	struct Qdisc *sch;
 877	unsigned int size = sizeof(*sch) + ops->priv_size;
 878	int err = -ENOBUFS;
 879	struct net_device *dev;
 880
 881	if (!dev_queue) {
 882		NL_SET_ERR_MSG(extack, "No device queue given");
 883		err = -EINVAL;
 884		goto errout;
 885	}
 886
 887	dev = dev_queue->dev;
 888	sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
 889
 890	if (!sch)
 891		goto errout;
 892	__skb_queue_head_init(&sch->gso_skb);
 893	__skb_queue_head_init(&sch->skb_bad_txq);
 894	qdisc_skb_head_init(&sch->q);
 895	spin_lock_init(&sch->q.lock);
 896
 897	if (ops->static_flags & TCQ_F_CPUSTATS) {
 898		sch->cpu_bstats =
 899			netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 900		if (!sch->cpu_bstats)
 901			goto errout1;
 902
 903		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 904		if (!sch->cpu_qstats) {
 905			free_percpu(sch->cpu_bstats);
 906			goto errout1;
 907		}
 908	}
 
 
 909
 910	spin_lock_init(&sch->busylock);
 911	lockdep_set_class(&sch->busylock,
 912			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 913
 914	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
 915	spin_lock_init(&sch->seqlock);
 916	lockdep_set_class(&sch->seqlock,
 917			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 918
 919	seqcount_init(&sch->running);
 920	lockdep_set_class(&sch->running,
 921			  dev->qdisc_running_key ?: &qdisc_running_key);
 922
 923	sch->ops = ops;
 924	sch->flags = ops->static_flags;
 925	sch->enqueue = ops->enqueue;
 926	sch->dequeue = ops->dequeue;
 927	sch->dev_queue = dev_queue;
 928	dev_hold(dev);
 929	refcount_set(&sch->refcnt, 1);
 930
 931	return sch;
 932errout1:
 933	kfree(sch);
 934errout:
 935	return ERR_PTR(err);
 936}
 937
 938struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 939				const struct Qdisc_ops *ops,
 940				unsigned int parentid,
 941				struct netlink_ext_ack *extack)
 942{
 943	struct Qdisc *sch;
 944
 945	if (!try_module_get(ops->owner)) {
 946		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 947		return NULL;
 948	}
 949
 950	sch = qdisc_alloc(dev_queue, ops, extack);
 951	if (IS_ERR(sch)) {
 952		module_put(ops->owner);
 953		return NULL;
 954	}
 955	sch->parent = parentid;
 956
 957	if (!ops->init || ops->init(sch, NULL, extack) == 0) {
 958		trace_qdisc_create(ops, dev_queue->dev, parentid);
 959		return sch;
 960	}
 961
 962	qdisc_put(sch);
 
 963	return NULL;
 964}
 965EXPORT_SYMBOL(qdisc_create_dflt);
 966
 967/* Under qdisc_lock(qdisc) and BH! */
 968
 969void qdisc_reset(struct Qdisc *qdisc)
 970{
 971	const struct Qdisc_ops *ops = qdisc->ops;
 972	struct sk_buff *skb, *tmp;
 973
 974	trace_qdisc_reset(qdisc);
 975
 976	if (ops->reset)
 977		ops->reset(qdisc);
 978
 979	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 980		__skb_unlink(skb, &qdisc->gso_skb);
 981		kfree_skb_list(skb);
 
 982	}
 983
 984	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 985		__skb_unlink(skb, &qdisc->skb_bad_txq);
 986		kfree_skb_list(skb);
 987	}
 988
 989	qdisc->q.qlen = 0;
 990	qdisc->qstats.backlog = 0;
 991}
 992EXPORT_SYMBOL(qdisc_reset);
 993
 994void qdisc_free(struct Qdisc *qdisc)
 995{
 
 
 996	if (qdisc_is_percpu_stats(qdisc)) {
 997		free_percpu(qdisc->cpu_bstats);
 998		free_percpu(qdisc->cpu_qstats);
 999	}
1000
1001	kfree(qdisc);
1002}
1003
1004static void qdisc_free_cb(struct rcu_head *head)
1005{
1006	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1007
1008	qdisc_free(q);
1009}
1010
1011static void qdisc_destroy(struct Qdisc *qdisc)
1012{
1013	const struct Qdisc_ops  *ops = qdisc->ops;
1014
1015#ifdef CONFIG_NET_SCHED
1016	qdisc_hash_del(qdisc);
1017
1018	qdisc_put_stab(rtnl_dereference(qdisc->stab));
1019#endif
1020	gen_kill_estimator(&qdisc->rate_est);
1021
1022	qdisc_reset(qdisc);
1023
1024	if (ops->destroy)
1025		ops->destroy(qdisc);
1026
1027	module_put(ops->owner);
1028	dev_put(qdisc_dev(qdisc));
1029
1030	trace_qdisc_destroy(qdisc);
1031
1032	call_rcu(&qdisc->rcu, qdisc_free_cb);
1033}
1034
1035void qdisc_put(struct Qdisc *qdisc)
1036{
1037	if (!qdisc)
1038		return;
1039
1040	if (qdisc->flags & TCQ_F_BUILTIN ||
1041	    !refcount_dec_and_test(&qdisc->refcnt))
1042		return;
1043
1044	qdisc_destroy(qdisc);
1045}
1046EXPORT_SYMBOL(qdisc_put);
1047
1048/* Version of qdisc_put() that is called with rtnl mutex unlocked.
1049 * Intended to be used as optimization, this function only takes rtnl lock if
1050 * qdisc reference counter reached zero.
1051 */
1052
1053void qdisc_put_unlocked(struct Qdisc *qdisc)
1054{
1055	if (qdisc->flags & TCQ_F_BUILTIN ||
1056	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1057		return;
1058
1059	qdisc_destroy(qdisc);
1060	rtnl_unlock();
1061}
1062EXPORT_SYMBOL(qdisc_put_unlocked);
1063
1064/* Attach toplevel qdisc to device queue. */
1065struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1066			      struct Qdisc *qdisc)
1067{
1068	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1069	spinlock_t *root_lock;
1070
1071	root_lock = qdisc_lock(oqdisc);
1072	spin_lock_bh(root_lock);
1073
 
 
 
 
1074	/* ... and graft new one */
1075	if (qdisc == NULL)
1076		qdisc = &noop_qdisc;
1077	dev_queue->qdisc_sleeping = qdisc;
1078	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1079
1080	spin_unlock_bh(root_lock);
1081
1082	return oqdisc;
1083}
1084EXPORT_SYMBOL(dev_graft_qdisc);
1085
1086static void attach_one_default_qdisc(struct net_device *dev,
1087				     struct netdev_queue *dev_queue,
1088				     void *_unused)
1089{
1090	struct Qdisc *qdisc;
1091	const struct Qdisc_ops *ops = default_qdisc_ops;
1092
1093	if (dev->priv_flags & IFF_NO_QUEUE)
1094		ops = &noqueue_qdisc_ops;
1095	else if(dev->type == ARPHRD_CAN)
1096		ops = &pfifo_fast_ops;
1097
1098	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1099	if (!qdisc)
 
1100		return;
1101
1102	if (!netif_is_multiqueue(dev))
1103		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1104	dev_queue->qdisc_sleeping = qdisc;
1105}
1106
1107static void attach_default_qdiscs(struct net_device *dev)
1108{
1109	struct netdev_queue *txq;
1110	struct Qdisc *qdisc;
1111
1112	txq = netdev_get_tx_queue(dev, 0);
1113
1114	if (!netif_is_multiqueue(dev) ||
1115	    dev->priv_flags & IFF_NO_QUEUE) {
1116		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1117		dev->qdisc = txq->qdisc_sleeping;
1118		qdisc_refcount_inc(dev->qdisc);
1119	} else {
1120		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1121		if (qdisc) {
1122			dev->qdisc = qdisc;
1123			qdisc->ops->attach(qdisc);
1124		}
1125	}
1126
1127	/* Detect default qdisc setup/init failed and fallback to "noqueue" */
1128	if (dev->qdisc == &noop_qdisc) {
1129		netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1130			    default_qdisc_ops->id, noqueue_qdisc_ops.id);
1131		dev->priv_flags |= IFF_NO_QUEUE;
1132		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1133		dev->qdisc = txq->qdisc_sleeping;
1134		qdisc_refcount_inc(dev->qdisc);
1135		dev->priv_flags ^= IFF_NO_QUEUE;
1136	}
1137
1138#ifdef CONFIG_NET_SCHED
1139	if (dev->qdisc != &noop_qdisc)
1140		qdisc_hash_add(dev->qdisc, false);
1141#endif
1142}
1143
1144static void transition_one_qdisc(struct net_device *dev,
1145				 struct netdev_queue *dev_queue,
1146				 void *_need_watchdog)
1147{
1148	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1149	int *need_watchdog_p = _need_watchdog;
1150
1151	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1152		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1153
1154	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1155	if (need_watchdog_p) {
1156		dev_queue->trans_start = 0;
1157		*need_watchdog_p = 1;
1158	}
1159}
1160
1161void dev_activate(struct net_device *dev)
1162{
1163	int need_watchdog;
1164
1165	/* No queueing discipline is attached to device;
1166	 * create default one for devices, which need queueing
1167	 * and noqueue_qdisc for virtual interfaces
1168	 */
1169
1170	if (dev->qdisc == &noop_qdisc)
1171		attach_default_qdiscs(dev);
1172
1173	if (!netif_carrier_ok(dev))
1174		/* Delay activation until next carrier-on event */
1175		return;
1176
1177	need_watchdog = 0;
1178	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1179	if (dev_ingress_queue(dev))
1180		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1181
1182	if (need_watchdog) {
1183		netif_trans_update(dev);
1184		dev_watchdog_up(dev);
1185	}
1186}
1187EXPORT_SYMBOL(dev_activate);
1188
1189static void qdisc_deactivate(struct Qdisc *qdisc)
1190{
1191	if (qdisc->flags & TCQ_F_BUILTIN)
1192		return;
1193
1194	set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1195}
1196
1197static void dev_deactivate_queue(struct net_device *dev,
1198				 struct netdev_queue *dev_queue,
1199				 void *_qdisc_default)
1200{
1201	struct Qdisc *qdisc_default = _qdisc_default;
1202	struct Qdisc *qdisc;
1203
1204	qdisc = rtnl_dereference(dev_queue->qdisc);
1205	if (qdisc) {
1206		qdisc_deactivate(qdisc);
1207		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1208	}
1209}
1210
1211static void dev_reset_queue(struct net_device *dev,
1212			    struct netdev_queue *dev_queue,
1213			    void *_unused)
1214{
1215	struct Qdisc *qdisc;
1216	bool nolock;
1217
1218	qdisc = dev_queue->qdisc_sleeping;
1219	if (!qdisc)
1220		return;
1221
1222	nolock = qdisc->flags & TCQ_F_NOLOCK;
1223
1224	if (nolock)
1225		spin_lock_bh(&qdisc->seqlock);
1226	spin_lock_bh(qdisc_lock(qdisc));
1227
1228	qdisc_reset(qdisc);
1229
1230	spin_unlock_bh(qdisc_lock(qdisc));
1231	if (nolock) {
1232		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1233		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1234		spin_unlock_bh(&qdisc->seqlock);
1235	}
1236}
1237
1238static bool some_qdisc_is_busy(struct net_device *dev)
1239{
1240	unsigned int i;
1241
1242	for (i = 0; i < dev->num_tx_queues; i++) {
1243		struct netdev_queue *dev_queue;
1244		spinlock_t *root_lock;
1245		struct Qdisc *q;
1246		int val;
1247
1248		dev_queue = netdev_get_tx_queue(dev, i);
1249		q = dev_queue->qdisc_sleeping;
 
1250
1251		root_lock = qdisc_lock(q);
1252		spin_lock_bh(root_lock);
1253
1254		val = (qdisc_is_running(q) ||
1255		       test_bit(__QDISC_STATE_SCHED, &q->state));
1256
1257		spin_unlock_bh(root_lock);
1258
1259		if (val)
1260			return true;
1261	}
1262	return false;
1263}
1264
1265/**
1266 * 	dev_deactivate_many - deactivate transmissions on several devices
1267 * 	@head: list of devices to deactivate
1268 *
1269 *	This function returns only when all outstanding transmissions
1270 *	have completed, unless all devices are in dismantle phase.
1271 */
1272void dev_deactivate_many(struct list_head *head)
1273{
1274	struct net_device *dev;
 
1275
1276	list_for_each_entry(dev, head, close_list) {
1277		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1278					 &noop_qdisc);
1279		if (dev_ingress_queue(dev))
1280			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1281					     &noop_qdisc);
1282
1283		dev_watchdog_down(dev);
 
1284	}
1285
1286	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
1287	 * outstanding qdisc enqueuing calls.
1288	 * This is avoided if all devices are in dismantle phase :
1289	 * Caller will call synchronize_net() for us
1290	 */
1291	synchronize_net();
1292
1293	list_for_each_entry(dev, head, close_list) {
1294		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1295
1296		if (dev_ingress_queue(dev))
1297			dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1298	}
1299
1300	/* Wait for outstanding qdisc_run calls. */
1301	list_for_each_entry(dev, head, close_list) {
1302		while (some_qdisc_is_busy(dev)) {
1303			/* wait_event() would avoid this sleep-loop but would
1304			 * require expensive checks in the fast paths of packet
1305			 * processing which isn't worth it.
1306			 */
1307			schedule_timeout_uninterruptible(1);
1308		}
1309	}
1310}
1311
1312void dev_deactivate(struct net_device *dev)
1313{
1314	LIST_HEAD(single);
1315
1316	list_add(&dev->close_list, &single);
1317	dev_deactivate_many(&single);
1318	list_del(&single);
1319}
1320EXPORT_SYMBOL(dev_deactivate);
1321
1322static int qdisc_change_tx_queue_len(struct net_device *dev,
1323				     struct netdev_queue *dev_queue)
1324{
1325	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1326	const struct Qdisc_ops *ops = qdisc->ops;
1327
1328	if (ops->change_tx_queue_len)
1329		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1330	return 0;
1331}
1332
1333int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1334{
1335	bool up = dev->flags & IFF_UP;
1336	unsigned int i;
1337	int ret = 0;
1338
1339	if (up)
1340		dev_deactivate(dev);
1341
1342	for (i = 0; i < dev->num_tx_queues; i++) {
1343		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1344
1345		/* TODO: revert changes on a partial failure */
1346		if (ret)
1347			break;
1348	}
1349
1350	if (up)
1351		dev_activate(dev);
1352	return ret;
1353}
1354
1355static void dev_init_scheduler_queue(struct net_device *dev,
1356				     struct netdev_queue *dev_queue,
1357				     void *_qdisc)
1358{
1359	struct Qdisc *qdisc = _qdisc;
1360
1361	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1362	dev_queue->qdisc_sleeping = qdisc;
1363}
1364
1365void dev_init_scheduler(struct net_device *dev)
1366{
1367	dev->qdisc = &noop_qdisc;
1368	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1369	if (dev_ingress_queue(dev))
1370		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1371
1372	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1373}
1374
1375static void shutdown_scheduler_queue(struct net_device *dev,
1376				     struct netdev_queue *dev_queue,
1377				     void *_qdisc_default)
1378{
1379	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1380	struct Qdisc *qdisc_default = _qdisc_default;
1381
1382	if (qdisc) {
1383		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1384		dev_queue->qdisc_sleeping = qdisc_default;
1385
1386		qdisc_put(qdisc);
1387	}
1388}
1389
1390void dev_shutdown(struct net_device *dev)
1391{
1392	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1393	if (dev_ingress_queue(dev))
1394		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1395	qdisc_put(dev->qdisc);
1396	dev->qdisc = &noop_qdisc;
1397
1398	WARN_ON(timer_pending(&dev->watchdog_timer));
1399}
1400
1401/**
1402 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1403 * @rate:   Rate to compute reciprocal division values of
1404 * @mult:   Multiplier for reciprocal division
1405 * @shift:  Shift for reciprocal division
1406 *
1407 * The multiplier and shift for reciprocal division by rate are stored
1408 * in mult and shift.
1409 *
1410 * The deal here is to replace a divide by a reciprocal one
1411 * in fast path (a reciprocal divide is a multiply and a shift)
1412 *
1413 * Normal formula would be :
1414 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1415 *
1416 * We compute mult/shift to use instead :
1417 *  time_in_ns = (len * mult) >> shift;
1418 *
1419 * We try to get the highest possible mult value for accuracy,
1420 * but have to make sure no overflows will ever happen.
1421 *
1422 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1423 */
1424static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1425{
1426	u64 factor = NSEC_PER_SEC;
1427
1428	*mult = 1;
1429	*shift = 0;
1430
1431	if (rate <= 0)
1432		return;
1433
1434	for (;;) {
1435		*mult = div64_u64(factor, rate);
1436		if (*mult & (1U << 31) || factor & (1ULL << 63))
1437			break;
1438		factor <<= 1;
1439		(*shift)++;
1440	}
1441}
1442
1443void psched_ratecfg_precompute(struct psched_ratecfg *r,
1444			       const struct tc_ratespec *conf,
1445			       u64 rate64)
1446{
1447	memset(r, 0, sizeof(*r));
1448	r->overhead = conf->overhead;
1449	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1450	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1451	psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1452}
1453EXPORT_SYMBOL(psched_ratecfg_precompute);
 
 
 
 
 
 
 
 
 
 
 
 
 
1454
1455void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1456{
1457	r->rate_pkts_ps = pktrate64;
1458	psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1459}
1460EXPORT_SYMBOL(psched_ppscfg_precompute);
1461
1462static void mini_qdisc_rcu_func(struct rcu_head *head)
1463{
1464}
1465
1466void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1467			  struct tcf_proto *tp_head)
1468{
1469	/* Protected with chain0->filter_chain_lock.
1470	 * Can't access chain directly because tp_head can be NULL.
1471	 */
1472	struct mini_Qdisc *miniq_old =
1473		rcu_dereference_protected(*miniqp->p_miniq, 1);
1474	struct mini_Qdisc *miniq;
1475
1476	if (!tp_head) {
1477		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1478		/* Wait for flying RCU callback before it is freed. */
1479		rcu_barrier();
1480		return;
1481	}
1482
1483	miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1484		&miniqp->miniq1 : &miniqp->miniq2;
1485
1486	/* We need to make sure that readers won't see the miniq
1487	 * we are about to modify. So wait until previous call_rcu callback
1488	 * is done.
1489	 */
1490	rcu_barrier();
1491	miniq->filter_list = tp_head;
1492	rcu_assign_pointer(*miniqp->p_miniq, miniq);
1493
1494	if (miniq_old)
1495		/* This is counterpart of the rcu barriers above. We need to
1496		 * block potential new user of miniq_old until all readers
1497		 * are not seeing it.
1498		 */
1499		call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1500}
1501EXPORT_SYMBOL(mini_qdisc_pair_swap);
1502
1503void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1504				struct tcf_block *block)
1505{
1506	miniqp->miniq1.block = block;
1507	miniqp->miniq2.block = block;
1508}
1509EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1510
1511void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1512			  struct mini_Qdisc __rcu **p_miniq)
1513{
1514	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1515	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1516	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1517	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1518	miniqp->p_miniq = p_miniq;
1519}
1520EXPORT_SYMBOL(mini_qdisc_pair_init);