Linux Audio

Check our new training course

Embedded Linux training

Mar 31-Apr 8, 2025
Register
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);
v5.9
   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
  38/* Main transmission queue. */
  39
  40/* Modifications to data participating in scheduling must be protected with
  41 * qdisc_lock(qdisc) spinlock.
  42 *
  43 * The idea is the following:
  44 * - enqueue, dequeue are serialized via qdisc root lock
  45 * - ingress filtering is also serialized via qdisc root lock
  46 * - updates to tree and tree walking are only done under the rtnl mutex.
  47 */
  48
  49#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  50
  51static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  52{
  53	const struct netdev_queue *txq = q->dev_queue;
  54	spinlock_t *lock = NULL;
  55	struct sk_buff *skb;
 
  56
  57	if (q->flags & TCQ_F_NOLOCK) {
  58		lock = qdisc_lock(q);
  59		spin_lock(lock);
  60	}
  61
  62	skb = skb_peek(&q->skb_bad_txq);
  63	if (skb) {
  64		/* check the reason of requeuing without tx lock first */
  65		txq = skb_get_tx_queue(txq->dev, skb);
  66		if (!netif_xmit_frozen_or_stopped(txq)) {
  67			skb = __skb_dequeue(&q->skb_bad_txq);
  68			if (qdisc_is_percpu_stats(q)) {
  69				qdisc_qstats_cpu_backlog_dec(q, skb);
  70				qdisc_qstats_cpu_qlen_dec(q);
  71			} else {
  72				qdisc_qstats_backlog_dec(q, skb);
  73				q->q.qlen--;
  74			}
  75		} else {
  76			skb = SKB_XOFF_MAGIC;
  77		}
  78	}
  79
  80	if (lock)
  81		spin_unlock(lock);
  82
  83	return skb;
  84}
  85
  86static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
  87{
  88	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
  89
  90	if (unlikely(skb))
  91		skb = __skb_dequeue_bad_txq(q);
  92
  93	return skb;
  94}
  95
  96static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
  97					     struct sk_buff *skb)
  98{
  99	spinlock_t *lock = NULL;
 100
 101	if (q->flags & TCQ_F_NOLOCK) {
 102		lock = qdisc_lock(q);
 103		spin_lock(lock);
 104	}
 105
 106	__skb_queue_tail(&q->skb_bad_txq, skb);
 107
 108	if (qdisc_is_percpu_stats(q)) {
 109		qdisc_qstats_cpu_backlog_inc(q, skb);
 110		qdisc_qstats_cpu_qlen_inc(q);
 111	} else {
 112		qdisc_qstats_backlog_inc(q, skb);
 113		q->q.qlen++;
 114	}
 115
 116	if (lock)
 117		spin_unlock(lock);
 118}
 119
 120static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 121{
 122	spinlock_t *lock = NULL;
 123
 124	if (q->flags & TCQ_F_NOLOCK) {
 125		lock = qdisc_lock(q);
 126		spin_lock(lock);
 127	}
 128
 129	while (skb) {
 130		struct sk_buff *next = skb->next;
 131
 132		__skb_queue_tail(&q->gso_skb, skb);
 133
 134		/* it's still part of the queue */
 135		if (qdisc_is_percpu_stats(q)) {
 136			qdisc_qstats_cpu_requeues_inc(q);
 137			qdisc_qstats_cpu_backlog_inc(q, skb);
 138			qdisc_qstats_cpu_qlen_inc(q);
 139		} else {
 140			q->qstats.requeues++;
 141			qdisc_qstats_backlog_inc(q, skb);
 142			q->q.qlen++;
 143		}
 144
 145		skb = next;
 146	}
 147	if (lock)
 148		spin_unlock(lock);
 149	__netif_schedule(q);
 150}
 151
 152static void try_bulk_dequeue_skb(struct Qdisc *q,
 153				 struct sk_buff *skb,
 154				 const struct netdev_queue *txq,
 155				 int *packets)
 156{
 157	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 158
 159	while (bytelimit > 0) {
 160		struct sk_buff *nskb = q->dequeue(q);
 161
 162		if (!nskb)
 163			break;
 164
 165		bytelimit -= nskb->len; /* covers GSO len */
 166		skb->next = nskb;
 167		skb = nskb;
 168		(*packets)++; /* GSO counts as one pkt */
 169	}
 170	skb_mark_not_on_list(skb);
 171}
 172
 173/* This variant of try_bulk_dequeue_skb() makes sure
 174 * all skbs in the chain are for the same txq
 175 */
 176static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 177				      struct sk_buff *skb,
 178				      int *packets)
 179{
 180	int mapping = skb_get_queue_mapping(skb);
 181	struct sk_buff *nskb;
 182	int cnt = 0;
 183
 184	do {
 185		nskb = q->dequeue(q);
 186		if (!nskb)
 187			break;
 188		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 189			qdisc_enqueue_skb_bad_txq(q, nskb);
 190			break;
 191		}
 192		skb->next = nskb;
 193		skb = nskb;
 194	} while (++cnt < 8);
 195	(*packets) += cnt;
 196	skb_mark_not_on_list(skb);
 197}
 198
 199/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 200 * A requeued skb (via q->gso_skb) can also be a SKB list.
 201 */
 202static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 203				   int *packets)
 204{
 
 205	const struct netdev_queue *txq = q->dev_queue;
 206	struct sk_buff *skb = NULL;
 207
 208	*packets = 1;
 209	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 210		spinlock_t *lock = NULL;
 211
 212		if (q->flags & TCQ_F_NOLOCK) {
 213			lock = qdisc_lock(q);
 214			spin_lock(lock);
 215		}
 216
 217		skb = skb_peek(&q->gso_skb);
 218
 219		/* skb may be null if another cpu pulls gso_skb off in between
 220		 * empty check and lock.
 221		 */
 222		if (!skb) {
 223			if (lock)
 224				spin_unlock(lock);
 225			goto validate;
 226		}
 227
 228		/* skb in gso_skb were already validated */
 229		*validate = false;
 230		if (xfrm_offload(skb))
 231			*validate = true;
 232		/* check the reason of requeuing without tx lock first */
 233		txq = skb_get_tx_queue(txq->dev, skb);
 234		if (!netif_xmit_frozen_or_stopped(txq)) {
 235			skb = __skb_dequeue(&q->gso_skb);
 236			if (qdisc_is_percpu_stats(q)) {
 237				qdisc_qstats_cpu_backlog_dec(q, skb);
 238				qdisc_qstats_cpu_qlen_dec(q);
 239			} else {
 240				qdisc_qstats_backlog_dec(q, skb);
 241				q->q.qlen--;
 242			}
 243		} else {
 244			skb = NULL;
 
 
 
 
 
 
 
 
 245		}
 246		if (lock)
 247			spin_unlock(lock);
 248		goto trace;
 249	}
 250validate:
 251	*validate = true;
 252
 253	if ((q->flags & TCQ_F_ONETXQUEUE) &&
 254	    netif_xmit_frozen_or_stopped(txq))
 255		return skb;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 256
 257	skb = qdisc_dequeue_skb_bad_txq(q);
 258	if (unlikely(skb)) {
 259		if (skb == SKB_XOFF_MAGIC)
 260			return NULL;
 261		goto bulk;
 262	}
 263	skb = q->dequeue(q);
 264	if (skb) {
 265bulk:
 266		if (qdisc_may_bulk(q))
 267			try_bulk_dequeue_skb(q, skb, txq, packets);
 268		else
 269			try_bulk_dequeue_skb_slow(q, skb, packets);
 270	}
 271trace:
 272	trace_qdisc_dequeue(q, txq, *packets, skb);
 273	return skb;
 274}
 275
 276/*
 277 * Transmit possibly several skbs, and handle the return status as
 278 * required. Owning running seqcount bit guarantees that
 279 * only one CPU can execute this function.
 280 *
 281 * Returns to the caller:
 282 *				false  - hardware queue frozen backoff
 283 *				true   - feel free to send more pkts
 284 */
 285bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 286		     struct net_device *dev, struct netdev_queue *txq,
 287		     spinlock_t *root_lock, bool validate)
 288{
 289	int ret = NETDEV_TX_BUSY;
 290	bool again = false;
 291
 292	/* And release qdisc */
 293	if (root_lock)
 294		spin_unlock(root_lock);
 295
 296	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
 297	if (validate)
 298		skb = validate_xmit_skb_list(skb, dev, &again);
 299
 300#ifdef CONFIG_XFRM_OFFLOAD
 301	if (unlikely(again)) {
 302		if (root_lock)
 303			spin_lock(root_lock);
 304
 305		dev_requeue_skb(skb, q);
 306		return false;
 307	}
 308#endif
 309
 310	if (likely(skb)) {
 311		HARD_TX_LOCK(dev, txq, smp_processor_id());
 312		if (!netif_xmit_frozen_or_stopped(txq))
 313			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 314
 315		HARD_TX_UNLOCK(dev, txq);
 316	} else {
 317		if (root_lock)
 318			spin_lock(root_lock);
 319		return true;
 320	}
 
 321
 322	if (root_lock)
 323		spin_lock(root_lock);
 324
 325	if (!dev_xmit_complete(ret)) {
 
 
 
 326		/* Driver returned NETDEV_TX_BUSY - requeue skb */
 327		if (unlikely(ret != NETDEV_TX_BUSY))
 328			net_warn_ratelimited("BUG %s code %d qlen %d\n",
 329					     dev->name, ret, q->q.qlen);
 330
 331		dev_requeue_skb(skb, q);
 332		return false;
 333	}
 334
 335	return true;
 
 
 
 336}
 337
 338/*
 339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 340 *
 341 * running seqcount guarantees only one CPU can process
 342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 343 * this queue.
 344 *
 345 *  netif_tx_lock serializes accesses to device driver.
 346 *
 347 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 348 *  if one is grabbed, another must be free.
 349 *
 350 * Note, that this procedure can be called by a watchdog timer
 351 *
 352 * Returns to the caller:
 353 *				0  - queue is empty or throttled.
 354 *				>0 - queue is not empty.
 355 *
 356 */
 357static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 358{
 359	spinlock_t *root_lock = NULL;
 360	struct netdev_queue *txq;
 361	struct net_device *dev;
 
 362	struct sk_buff *skb;
 363	bool validate;
 364
 365	/* Dequeue packet */
 366	skb = dequeue_skb(q, &validate, packets);
 367	if (unlikely(!skb))
 368		return false;
 369
 370	if (!(q->flags & TCQ_F_NOLOCK))
 371		root_lock = qdisc_lock(q);
 372
 
 373	dev = qdisc_dev(q);
 374	txq = skb_get_tx_queue(dev, skb);
 375
 376	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 377}
 378
 379void __qdisc_run(struct Qdisc *q)
 380{
 381	int quota = dev_tx_weight;
 382	int packets;
 383
 384	while (qdisc_restart(q, &packets)) {
 
 
 
 
 
 385		quota -= packets;
 386		if (quota <= 0) {
 387			__netif_schedule(q);
 388			break;
 389		}
 390	}
 
 
 391}
 392
 393unsigned long dev_trans_start(struct net_device *dev)
 394{
 395	unsigned long val, res;
 396	unsigned int i;
 397
 398	if (is_vlan_dev(dev))
 399		dev = vlan_dev_real_dev(dev);
 400	else if (netif_is_macvlan(dev))
 401		dev = macvlan_dev_real_dev(dev);
 402	res = netdev_get_tx_queue(dev, 0)->trans_start;
 403	for (i = 1; i < dev->num_tx_queues; i++) {
 404		val = netdev_get_tx_queue(dev, i)->trans_start;
 405		if (val && time_after(val, res))
 406			res = val;
 407	}
 
 408
 409	return res;
 410}
 411EXPORT_SYMBOL(dev_trans_start);
 412
 413static void dev_watchdog(struct timer_list *t)
 414{
 415	struct net_device *dev = from_timer(dev, t, watchdog_timer);
 416
 417	netif_tx_lock(dev);
 418	if (!qdisc_tx_is_noop(dev)) {
 419		if (netif_device_present(dev) &&
 420		    netif_running(dev) &&
 421		    netif_carrier_ok(dev)) {
 422			int some_queue_timedout = 0;
 423			unsigned int i;
 424			unsigned long trans_start;
 425
 426			for (i = 0; i < dev->num_tx_queues; i++) {
 427				struct netdev_queue *txq;
 428
 429				txq = netdev_get_tx_queue(dev, i);
 430				trans_start = txq->trans_start;
 
 
 
 431				if (netif_xmit_stopped(txq) &&
 432				    time_after(jiffies, (trans_start +
 433							 dev->watchdog_timeo))) {
 434					some_queue_timedout = 1;
 435					txq->trans_timeout++;
 436					break;
 437				}
 438			}
 439
 440			if (some_queue_timedout) {
 441				trace_net_dev_xmit_timeout(dev, i);
 442				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 443				       dev->name, netdev_drivername(dev), i);
 444				dev->netdev_ops->ndo_tx_timeout(dev, i);
 445			}
 446			if (!mod_timer(&dev->watchdog_timer,
 447				       round_jiffies(jiffies +
 448						     dev->watchdog_timeo)))
 449				dev_hold(dev);
 450		}
 451	}
 452	netif_tx_unlock(dev);
 453
 454	dev_put(dev);
 455}
 456
 457void __netdev_watchdog_up(struct net_device *dev)
 458{
 459	if (dev->netdev_ops->ndo_tx_timeout) {
 460		if (dev->watchdog_timeo <= 0)
 461			dev->watchdog_timeo = 5*HZ;
 462		if (!mod_timer(&dev->watchdog_timer,
 463			       round_jiffies(jiffies + dev->watchdog_timeo)))
 464			dev_hold(dev);
 465	}
 466}
 467EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 468
 469static void dev_watchdog_up(struct net_device *dev)
 470{
 471	__netdev_watchdog_up(dev);
 472}
 473
 474static void dev_watchdog_down(struct net_device *dev)
 475{
 476	netif_tx_lock_bh(dev);
 477	if (del_timer(&dev->watchdog_timer))
 478		dev_put(dev);
 479	netif_tx_unlock_bh(dev);
 480}
 481
 482/**
 483 *	netif_carrier_on - set carrier
 484 *	@dev: network device
 485 *
 486 * Device has detected acquisition of carrier.
 487 */
 488void netif_carrier_on(struct net_device *dev)
 489{
 490	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 491		if (dev->reg_state == NETREG_UNINITIALIZED)
 492			return;
 493		atomic_inc(&dev->carrier_up_count);
 494		linkwatch_fire_event(dev);
 495		if (netif_running(dev))
 496			__netdev_watchdog_up(dev);
 497	}
 498}
 499EXPORT_SYMBOL(netif_carrier_on);
 500
 501/**
 502 *	netif_carrier_off - clear carrier
 503 *	@dev: network device
 504 *
 505 * Device has detected loss of carrier.
 506 */
 507void netif_carrier_off(struct net_device *dev)
 508{
 509	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 510		if (dev->reg_state == NETREG_UNINITIALIZED)
 511			return;
 512		atomic_inc(&dev->carrier_down_count);
 513		linkwatch_fire_event(dev);
 514	}
 515}
 516EXPORT_SYMBOL(netif_carrier_off);
 517
 518/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 519   under all circumstances. It is difficult to invent anything faster or
 520   cheaper.
 521 */
 522
 523static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 524			struct sk_buff **to_free)
 525{
 526	__qdisc_drop(skb, to_free);
 527	return NET_XMIT_CN;
 528}
 529
 530static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 531{
 532	return NULL;
 533}
 534
 535struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 536	.id		=	"noop",
 537	.priv_size	=	0,
 538	.enqueue	=	noop_enqueue,
 539	.dequeue	=	noop_dequeue,
 540	.peek		=	noop_dequeue,
 541	.owner		=	THIS_MODULE,
 542};
 543
 544static struct netdev_queue noop_netdev_queue = {
 545	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 546	.qdisc_sleeping	=	&noop_qdisc,
 547};
 548
 549struct Qdisc noop_qdisc = {
 550	.enqueue	=	noop_enqueue,
 551	.dequeue	=	noop_dequeue,
 552	.flags		=	TCQ_F_BUILTIN,
 553	.ops		=	&noop_qdisc_ops,
 
 554	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 555	.dev_queue	=	&noop_netdev_queue,
 556	.running	=	SEQCNT_ZERO(noop_qdisc.running),
 557	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 558	.gso_skb = {
 559		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
 560		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 561		.qlen = 0,
 562		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 563	},
 564	.skb_bad_txq = {
 565		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 566		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 567		.qlen = 0,
 568		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 569	},
 570};
 571EXPORT_SYMBOL(noop_qdisc);
 572
 573static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 574			struct netlink_ext_ack *extack)
 575{
 576	/* register_qdisc() assigns a default of noop_enqueue if unset,
 577	 * but __dev_queue_xmit() treats noqueue only as such
 578	 * if this is NULL - so clear it here. */
 579	qdisc->enqueue = NULL;
 580	return 0;
 581}
 582
 583struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 584	.id		=	"noqueue",
 585	.priv_size	=	0,
 586	.init		=	noqueue_init,
 587	.enqueue	=	noop_enqueue,
 588	.dequeue	=	noop_dequeue,
 589	.peek		=	noop_dequeue,
 590	.owner		=	THIS_MODULE,
 591};
 592
 593static const u8 prio2band[TC_PRIO_MAX + 1] = {
 594	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 595};
 596
 597/* 3-band FIFO queue: old style, but should be a bit faster than
 598   generic prio+fifo combination.
 599 */
 600
 601#define PFIFO_FAST_BANDS 3
 602
 603/*
 604 * Private data for a pfifo_fast scheduler containing:
 605 *	- rings for priority bands
 
 606 */
 607struct pfifo_fast_priv {
 608	struct skb_array q[PFIFO_FAST_BANDS];
 
 609};
 610
 611static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 612					  int band)
 
 
 
 
 
 
 
 
 613{
 614	return &priv->q[band];
 615}
 616
 617static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 618			      struct sk_buff **to_free)
 619{
 620	int band = prio2band[skb->priority & TC_PRIO_MAX];
 621	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 622	struct skb_array *q = band2list(priv, band);
 623	unsigned int pkt_len = qdisc_pkt_len(skb);
 624	int err;
 625
 626	err = skb_array_produce(q, skb);
 627
 628	if (unlikely(err)) {
 629		if (qdisc_is_percpu_stats(qdisc))
 630			return qdisc_drop_cpu(skb, qdisc, to_free);
 631		else
 632			return qdisc_drop(skb, qdisc, to_free);
 633	}
 634
 635	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 636	return NET_XMIT_SUCCESS;
 637}
 638
 639static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 640{
 641	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 642	struct sk_buff *skb = NULL;
 643	int band;
 644
 645	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 646		struct skb_array *q = band2list(priv, band);
 
 
 
 
 
 647
 648		if (__skb_array_empty(q))
 649			continue;
 650
 651		skb = __skb_array_consume(q);
 652	}
 653	if (likely(skb)) {
 654		qdisc_update_stats_at_dequeue(qdisc, skb);
 655	} else {
 656		WRITE_ONCE(qdisc->empty, true);
 657	}
 658
 659	return skb;
 660}
 661
 662static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 663{
 664	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 665	struct sk_buff *skb = NULL;
 666	int band;
 667
 668	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 669		struct skb_array *q = band2list(priv, band);
 670
 671		skb = __skb_array_peek(q);
 672	}
 673
 674	return skb;
 675}
 676
 677static void pfifo_fast_reset(struct Qdisc *qdisc)
 678{
 679	int i, band;
 680	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 681
 682	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 683		struct skb_array *q = band2list(priv, band);
 684		struct sk_buff *skb;
 685
 686		/* NULL ring is possible if destroy path is due to a failed
 687		 * skb_array_init() in pfifo_fast_init() case.
 688		 */
 689		if (!q->ring.queue)
 690			continue;
 691
 692		while ((skb = __skb_array_consume(q)) != NULL)
 693			kfree_skb(skb);
 694	}
 695
 696	if (qdisc_is_percpu_stats(qdisc)) {
 697		for_each_possible_cpu(i) {
 698			struct gnet_stats_queue *q;
 699
 700			q = per_cpu_ptr(qdisc->cpu_qstats, i);
 701			q->backlog = 0;
 702			q->qlen = 0;
 703		}
 704	}
 705}
 706
 707static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 708{
 709	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 710
 711	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 712	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 713		goto nla_put_failure;
 714	return skb->len;
 715
 716nla_put_failure:
 717	return -1;
 718}
 719
 720static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 721			   struct netlink_ext_ack *extack)
 722{
 723	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 724	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 725	int prio;
 726
 727	/* guard against zero length rings */
 728	if (!qlen)
 729		return -EINVAL;
 730
 731	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 732		struct skb_array *q = band2list(priv, prio);
 733		int err;
 734
 735		err = skb_array_init(q, qlen, GFP_KERNEL);
 736		if (err)
 737			return -ENOMEM;
 738	}
 739
 740	/* Can by-pass the queue discipline */
 741	qdisc->flags |= TCQ_F_CAN_BYPASS;
 742	return 0;
 743}
 744
 745static void pfifo_fast_destroy(struct Qdisc *sch)
 746{
 747	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 748	int prio;
 749
 750	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 751		struct skb_array *q = band2list(priv, prio);
 752
 753		/* NULL ring is possible if destroy path is due to a failed
 754		 * skb_array_init() in pfifo_fast_init() case.
 755		 */
 756		if (!q->ring.queue)
 757			continue;
 758		/* Destroy ring but no need to kfree_skb because a call to
 759		 * pfifo_fast_reset() has already done that work.
 760		 */
 761		ptr_ring_cleanup(&q->ring, NULL);
 762	}
 763}
 764
 765static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 766					  unsigned int new_len)
 767{
 768	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 769	struct skb_array *bands[PFIFO_FAST_BANDS];
 770	int prio;
 771
 772	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 773		struct skb_array *q = band2list(priv, prio);
 774
 775		bands[prio] = q;
 776	}
 777
 778	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 779					 GFP_KERNEL);
 780}
 781
 782struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 783	.id		=	"pfifo_fast",
 784	.priv_size	=	sizeof(struct pfifo_fast_priv),
 785	.enqueue	=	pfifo_fast_enqueue,
 786	.dequeue	=	pfifo_fast_dequeue,
 787	.peek		=	pfifo_fast_peek,
 788	.init		=	pfifo_fast_init,
 789	.destroy	=	pfifo_fast_destroy,
 790	.reset		=	pfifo_fast_reset,
 791	.dump		=	pfifo_fast_dump,
 792	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 793	.owner		=	THIS_MODULE,
 794	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 795};
 796EXPORT_SYMBOL(pfifo_fast_ops);
 797
 798static struct lock_class_key qdisc_tx_busylock;
 799static struct lock_class_key qdisc_running_key;
 800
 801struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 802			  const struct Qdisc_ops *ops,
 803			  struct netlink_ext_ack *extack)
 804{
 805	void *p;
 806	struct Qdisc *sch;
 807	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 808	int err = -ENOBUFS;
 809	struct net_device *dev;
 810
 811	if (!dev_queue) {
 812		NL_SET_ERR_MSG(extack, "No device queue given");
 813		err = -EINVAL;
 814		goto errout;
 815	}
 816
 817	dev = dev_queue->dev;
 818	p = kzalloc_node(size, GFP_KERNEL,
 819			 netdev_queue_numa_node_read(dev_queue));
 820
 821	if (!p)
 822		goto errout;
 823	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 824	/* if we got non aligned memory, ask more and do alignment ourself */
 825	if (sch != p) {
 826		kfree(p);
 827		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 828				 netdev_queue_numa_node_read(dev_queue));
 829		if (!p)
 830			goto errout;
 831		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 832		sch->padded = (char *) sch - (char *) p;
 833	}
 834	__skb_queue_head_init(&sch->gso_skb);
 835	__skb_queue_head_init(&sch->skb_bad_txq);
 836	qdisc_skb_head_init(&sch->q);
 837	spin_lock_init(&sch->q.lock);
 838
 839	if (ops->static_flags & TCQ_F_CPUSTATS) {
 840		sch->cpu_bstats =
 841			netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 842		if (!sch->cpu_bstats)
 843			goto errout1;
 844
 845		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 846		if (!sch->cpu_qstats) {
 847			free_percpu(sch->cpu_bstats);
 848			goto errout1;
 849		}
 850	}
 851
 852	spin_lock_init(&sch->busylock);
 853	lockdep_set_class(&sch->busylock,
 854			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 855
 856	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
 857	spin_lock_init(&sch->seqlock);
 858	lockdep_set_class(&sch->busylock,
 859			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 860
 861	seqcount_init(&sch->running);
 862	lockdep_set_class(&sch->running,
 863			  dev->qdisc_running_key ?: &qdisc_running_key);
 864
 865	sch->ops = ops;
 866	sch->flags = ops->static_flags;
 867	sch->enqueue = ops->enqueue;
 868	sch->dequeue = ops->dequeue;
 869	sch->dev_queue = dev_queue;
 870	sch->empty = true;
 871	dev_hold(dev);
 872	refcount_set(&sch->refcnt, 1);
 873
 874	return sch;
 875errout1:
 876	kfree(p);
 877errout:
 878	return ERR_PTR(err);
 879}
 880
 881struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 882				const struct Qdisc_ops *ops,
 883				unsigned int parentid,
 884				struct netlink_ext_ack *extack)
 885{
 886	struct Qdisc *sch;
 887
 888	if (!try_module_get(ops->owner)) {
 889		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 890		return NULL;
 891	}
 892
 893	sch = qdisc_alloc(dev_queue, ops, extack);
 894	if (IS_ERR(sch)) {
 895		module_put(ops->owner);
 896		return NULL;
 897	}
 898	sch->parent = parentid;
 899
 900	if (!ops->init || ops->init(sch, NULL, extack) == 0) {
 901		trace_qdisc_create(ops, dev_queue->dev, parentid);
 902		return sch;
 903	}
 904
 905	qdisc_put(sch);
 
 906	return NULL;
 907}
 908EXPORT_SYMBOL(qdisc_create_dflt);
 909
 910/* Under qdisc_lock(qdisc) and BH! */
 911
 912void qdisc_reset(struct Qdisc *qdisc)
 913{
 914	const struct Qdisc_ops *ops = qdisc->ops;
 915	struct sk_buff *skb, *tmp;
 916
 917	trace_qdisc_reset(qdisc);
 918
 919	if (ops->reset)
 920		ops->reset(qdisc);
 921
 922	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 923		__skb_unlink(skb, &qdisc->gso_skb);
 924		kfree_skb_list(skb);
 
 925	}
 926
 927	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 928		__skb_unlink(skb, &qdisc->skb_bad_txq);
 929		kfree_skb_list(skb);
 930	}
 931
 932	qdisc->q.qlen = 0;
 933	qdisc->qstats.backlog = 0;
 934}
 935EXPORT_SYMBOL(qdisc_reset);
 936
 937void qdisc_free(struct Qdisc *qdisc)
 938{
 
 
 939	if (qdisc_is_percpu_stats(qdisc)) {
 940		free_percpu(qdisc->cpu_bstats);
 941		free_percpu(qdisc->cpu_qstats);
 942	}
 943
 944	kfree((char *) qdisc - qdisc->padded);
 945}
 946
 947static void qdisc_free_cb(struct rcu_head *head)
 948{
 949	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
 950
 951	qdisc_free(q);
 952}
 953
 954static void qdisc_destroy(struct Qdisc *qdisc)
 955{
 956	const struct Qdisc_ops  *ops = qdisc->ops;
 957
 958#ifdef CONFIG_NET_SCHED
 959	qdisc_hash_del(qdisc);
 960
 961	qdisc_put_stab(rtnl_dereference(qdisc->stab));
 962#endif
 963	gen_kill_estimator(&qdisc->rate_est);
 964
 965	qdisc_reset(qdisc);
 966
 967	if (ops->destroy)
 968		ops->destroy(qdisc);
 969
 970	module_put(ops->owner);
 971	dev_put(qdisc_dev(qdisc));
 972
 973	trace_qdisc_destroy(qdisc);
 974
 975	call_rcu(&qdisc->rcu, qdisc_free_cb);
 976}
 977
 978void qdisc_put(struct Qdisc *qdisc)
 979{
 980	if (!qdisc)
 981		return;
 982
 983	if (qdisc->flags & TCQ_F_BUILTIN ||
 984	    !refcount_dec_and_test(&qdisc->refcnt))
 985		return;
 986
 987	qdisc_destroy(qdisc);
 988}
 989EXPORT_SYMBOL(qdisc_put);
 990
 991/* Version of qdisc_put() that is called with rtnl mutex unlocked.
 992 * Intended to be used as optimization, this function only takes rtnl lock if
 993 * qdisc reference counter reached zero.
 994 */
 995
 996void qdisc_put_unlocked(struct Qdisc *qdisc)
 997{
 998	if (qdisc->flags & TCQ_F_BUILTIN ||
 999	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1000		return;
1001
1002	qdisc_destroy(qdisc);
1003	rtnl_unlock();
1004}
1005EXPORT_SYMBOL(qdisc_put_unlocked);
1006
1007/* Attach toplevel qdisc to device queue. */
1008struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1009			      struct Qdisc *qdisc)
1010{
1011	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1012	spinlock_t *root_lock;
1013
1014	root_lock = qdisc_lock(oqdisc);
1015	spin_lock_bh(root_lock);
1016
 
 
 
 
1017	/* ... and graft new one */
1018	if (qdisc == NULL)
1019		qdisc = &noop_qdisc;
1020	dev_queue->qdisc_sleeping = qdisc;
1021	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1022
1023	spin_unlock_bh(root_lock);
1024
1025	return oqdisc;
1026}
1027EXPORT_SYMBOL(dev_graft_qdisc);
1028
1029static void attach_one_default_qdisc(struct net_device *dev,
1030				     struct netdev_queue *dev_queue,
1031				     void *_unused)
1032{
1033	struct Qdisc *qdisc;
1034	const struct Qdisc_ops *ops = default_qdisc_ops;
1035
1036	if (dev->priv_flags & IFF_NO_QUEUE)
1037		ops = &noqueue_qdisc_ops;
1038	else if(dev->type == ARPHRD_CAN)
1039		ops = &pfifo_fast_ops;
1040
1041	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1042	if (!qdisc)
 
1043		return;
1044
1045	if (!netif_is_multiqueue(dev))
1046		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1047	dev_queue->qdisc_sleeping = qdisc;
1048}
1049
1050static void attach_default_qdiscs(struct net_device *dev)
1051{
1052	struct netdev_queue *txq;
1053	struct Qdisc *qdisc;
1054
1055	txq = netdev_get_tx_queue(dev, 0);
1056
1057	if (!netif_is_multiqueue(dev) ||
1058	    dev->priv_flags & IFF_NO_QUEUE) {
1059		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1060		dev->qdisc = txq->qdisc_sleeping;
1061		qdisc_refcount_inc(dev->qdisc);
1062	} else {
1063		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1064		if (qdisc) {
1065			dev->qdisc = qdisc;
1066			qdisc->ops->attach(qdisc);
1067		}
1068	}
1069
1070	/* Detect default qdisc setup/init failed and fallback to "noqueue" */
1071	if (dev->qdisc == &noop_qdisc) {
1072		netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1073			    default_qdisc_ops->id, noqueue_qdisc_ops.id);
1074		dev->priv_flags |= IFF_NO_QUEUE;
1075		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1076		dev->qdisc = txq->qdisc_sleeping;
1077		qdisc_refcount_inc(dev->qdisc);
1078		dev->priv_flags ^= IFF_NO_QUEUE;
1079	}
1080
1081#ifdef CONFIG_NET_SCHED
1082	if (dev->qdisc != &noop_qdisc)
1083		qdisc_hash_add(dev->qdisc, false);
1084#endif
1085}
1086
1087static void transition_one_qdisc(struct net_device *dev,
1088				 struct netdev_queue *dev_queue,
1089				 void *_need_watchdog)
1090{
1091	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1092	int *need_watchdog_p = _need_watchdog;
1093
1094	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1095		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1096
1097	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1098	if (need_watchdog_p) {
1099		dev_queue->trans_start = 0;
1100		*need_watchdog_p = 1;
1101	}
1102}
1103
1104void dev_activate(struct net_device *dev)
1105{
1106	int need_watchdog;
1107
1108	/* No queueing discipline is attached to device;
1109	 * create default one for devices, which need queueing
1110	 * and noqueue_qdisc for virtual interfaces
1111	 */
1112
1113	if (dev->qdisc == &noop_qdisc)
1114		attach_default_qdiscs(dev);
1115
1116	if (!netif_carrier_ok(dev))
1117		/* Delay activation until next carrier-on event */
1118		return;
1119
1120	need_watchdog = 0;
1121	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1122	if (dev_ingress_queue(dev))
1123		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1124
1125	if (need_watchdog) {
1126		netif_trans_update(dev);
1127		dev_watchdog_up(dev);
1128	}
1129}
1130EXPORT_SYMBOL(dev_activate);
1131
1132static void qdisc_deactivate(struct Qdisc *qdisc)
1133{
1134	if (qdisc->flags & TCQ_F_BUILTIN)
1135		return;
1136
1137	set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1138}
1139
1140static void dev_deactivate_queue(struct net_device *dev,
1141				 struct netdev_queue *dev_queue,
1142				 void *_qdisc_default)
1143{
1144	struct Qdisc *qdisc_default = _qdisc_default;
1145	struct Qdisc *qdisc;
1146
1147	qdisc = rtnl_dereference(dev_queue->qdisc);
1148	if (qdisc) {
1149		qdisc_deactivate(qdisc);
1150		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1151	}
1152}
1153
1154static void dev_reset_queue(struct net_device *dev,
1155			    struct netdev_queue *dev_queue,
1156			    void *_unused)
1157{
1158	struct Qdisc *qdisc;
1159	bool nolock;
1160
1161	qdisc = dev_queue->qdisc_sleeping;
1162	if (!qdisc)
1163		return;
1164
1165	nolock = qdisc->flags & TCQ_F_NOLOCK;
1166
1167	if (nolock)
1168		spin_lock_bh(&qdisc->seqlock);
1169	spin_lock_bh(qdisc_lock(qdisc));
1170
1171	qdisc_reset(qdisc);
1172
1173	spin_unlock_bh(qdisc_lock(qdisc));
1174	if (nolock)
1175		spin_unlock_bh(&qdisc->seqlock);
1176}
1177
1178static bool some_qdisc_is_busy(struct net_device *dev)
1179{
1180	unsigned int i;
1181
1182	for (i = 0; i < dev->num_tx_queues; i++) {
1183		struct netdev_queue *dev_queue;
1184		spinlock_t *root_lock;
1185		struct Qdisc *q;
1186		int val;
1187
1188		dev_queue = netdev_get_tx_queue(dev, i);
1189		q = dev_queue->qdisc_sleeping;
 
1190
1191		root_lock = qdisc_lock(q);
1192		spin_lock_bh(root_lock);
1193
1194		val = (qdisc_is_running(q) ||
1195		       test_bit(__QDISC_STATE_SCHED, &q->state));
1196
1197		spin_unlock_bh(root_lock);
1198
1199		if (val)
1200			return true;
1201	}
1202	return false;
1203}
1204
1205/**
1206 * 	dev_deactivate_many - deactivate transmissions on several devices
1207 * 	@head: list of devices to deactivate
1208 *
1209 *	This function returns only when all outstanding transmissions
1210 *	have completed, unless all devices are in dismantle phase.
1211 */
1212void dev_deactivate_many(struct list_head *head)
1213{
1214	struct net_device *dev;
 
1215
1216	list_for_each_entry(dev, head, close_list) {
1217		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1218					 &noop_qdisc);
1219		if (dev_ingress_queue(dev))
1220			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1221					     &noop_qdisc);
1222
1223		dev_watchdog_down(dev);
 
1224	}
1225
1226	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
1227	 * outstanding qdisc enqueuing calls.
1228	 * This is avoided if all devices are in dismantle phase :
1229	 * Caller will call synchronize_net() for us
1230	 */
1231	synchronize_net();
1232
1233	list_for_each_entry(dev, head, close_list) {
1234		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1235
1236		if (dev_ingress_queue(dev))
1237			dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1238	}
1239
1240	/* Wait for outstanding qdisc_run calls. */
1241	list_for_each_entry(dev, head, close_list) {
1242		while (some_qdisc_is_busy(dev)) {
1243			/* wait_event() would avoid this sleep-loop but would
1244			 * require expensive checks in the fast paths of packet
1245			 * processing which isn't worth it.
1246			 */
1247			schedule_timeout_uninterruptible(1);
1248		}
1249	}
1250}
1251
1252void dev_deactivate(struct net_device *dev)
1253{
1254	LIST_HEAD(single);
1255
1256	list_add(&dev->close_list, &single);
1257	dev_deactivate_many(&single);
1258	list_del(&single);
1259}
1260EXPORT_SYMBOL(dev_deactivate);
1261
1262static int qdisc_change_tx_queue_len(struct net_device *dev,
1263				     struct netdev_queue *dev_queue)
1264{
1265	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1266	const struct Qdisc_ops *ops = qdisc->ops;
1267
1268	if (ops->change_tx_queue_len)
1269		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1270	return 0;
1271}
1272
1273int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1274{
1275	bool up = dev->flags & IFF_UP;
1276	unsigned int i;
1277	int ret = 0;
1278
1279	if (up)
1280		dev_deactivate(dev);
1281
1282	for (i = 0; i < dev->num_tx_queues; i++) {
1283		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1284
1285		/* TODO: revert changes on a partial failure */
1286		if (ret)
1287			break;
1288	}
1289
1290	if (up)
1291		dev_activate(dev);
1292	return ret;
1293}
1294
1295static void dev_init_scheduler_queue(struct net_device *dev,
1296				     struct netdev_queue *dev_queue,
1297				     void *_qdisc)
1298{
1299	struct Qdisc *qdisc = _qdisc;
1300
1301	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1302	dev_queue->qdisc_sleeping = qdisc;
1303}
1304
1305void dev_init_scheduler(struct net_device *dev)
1306{
1307	dev->qdisc = &noop_qdisc;
1308	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1309	if (dev_ingress_queue(dev))
1310		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311
1312	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1313}
1314
1315static void shutdown_scheduler_queue(struct net_device *dev,
1316				     struct netdev_queue *dev_queue,
1317				     void *_qdisc_default)
1318{
1319	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1320	struct Qdisc *qdisc_default = _qdisc_default;
1321
1322	if (qdisc) {
1323		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1324		dev_queue->qdisc_sleeping = qdisc_default;
1325
1326		qdisc_put(qdisc);
1327	}
1328}
1329
1330void dev_shutdown(struct net_device *dev)
1331{
1332	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1333	if (dev_ingress_queue(dev))
1334		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1335	qdisc_put(dev->qdisc);
1336	dev->qdisc = &noop_qdisc;
1337
1338	WARN_ON(timer_pending(&dev->watchdog_timer));
1339}
1340
1341void psched_ratecfg_precompute(struct psched_ratecfg *r,
1342			       const struct tc_ratespec *conf,
1343			       u64 rate64)
1344{
1345	memset(r, 0, sizeof(*r));
1346	r->overhead = conf->overhead;
1347	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1348	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1349	r->mult = 1;
1350	/*
1351	 * The deal here is to replace a divide by a reciprocal one
1352	 * in fast path (a reciprocal divide is a multiply and a shift)
1353	 *
1354	 * Normal formula would be :
1355	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1356	 *
1357	 * We compute mult/shift to use instead :
1358	 *  time_in_ns = (len * mult) >> shift;
1359	 *
1360	 * We try to get the highest possible mult value for accuracy,
1361	 * but have to make sure no overflows will ever happen.
1362	 */
1363	if (r->rate_bytes_ps > 0) {
1364		u64 factor = NSEC_PER_SEC;
1365
1366		for (;;) {
1367			r->mult = div64_u64(factor, r->rate_bytes_ps);
1368			if (r->mult & (1U << 31) || factor & (1ULL << 63))
1369				break;
1370			factor <<= 1;
1371			r->shift++;
1372		}
1373	}
1374}
1375EXPORT_SYMBOL(psched_ratecfg_precompute);
1376
1377static void mini_qdisc_rcu_func(struct rcu_head *head)
1378{
1379}
1380
1381void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1382			  struct tcf_proto *tp_head)
1383{
1384	/* Protected with chain0->filter_chain_lock.
1385	 * Can't access chain directly because tp_head can be NULL.
1386	 */
1387	struct mini_Qdisc *miniq_old =
1388		rcu_dereference_protected(*miniqp->p_miniq, 1);
1389	struct mini_Qdisc *miniq;
1390
1391	if (!tp_head) {
1392		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1393		/* Wait for flying RCU callback before it is freed. */
1394		rcu_barrier();
1395		return;
1396	}
1397
1398	miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1399		&miniqp->miniq1 : &miniqp->miniq2;
1400
1401	/* We need to make sure that readers won't see the miniq
1402	 * we are about to modify. So wait until previous call_rcu callback
1403	 * is done.
1404	 */
1405	rcu_barrier();
1406	miniq->filter_list = tp_head;
1407	rcu_assign_pointer(*miniqp->p_miniq, miniq);
1408
1409	if (miniq_old)
1410		/* This is counterpart of the rcu barriers above. We need to
1411		 * block potential new user of miniq_old until all readers
1412		 * are not seeing it.
1413		 */
1414		call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1415}
1416EXPORT_SYMBOL(mini_qdisc_pair_swap);
1417
1418void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1419				struct tcf_block *block)
1420{
1421	miniqp->miniq1.block = block;
1422	miniqp->miniq2.block = block;
1423}
1424EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1425
1426void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1427			  struct mini_Qdisc __rcu **p_miniq)
1428{
1429	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1430	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1431	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1432	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1433	miniqp->p_miniq = p_miniq;
1434}
1435EXPORT_SYMBOL(mini_qdisc_pair_init);