xref: /linux/net/sched/sch_generic.c (revision 9ee0034b8f49aaaa7e7c2da8db1038915db99c19)
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 */
34 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35 EXPORT_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 
48 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
49 {
50 	q->gso_skb = skb;
51 	q->qstats.requeues++;
52 	qdisc_qstats_backlog_inc(q, skb);
53 	q->q.qlen++;	/* it's still part of the queue */
54 	__netif_schedule(q);
55 
56 	return 0;
57 }
58 
59 static void try_bulk_dequeue_skb(struct Qdisc *q,
60 				 struct sk_buff *skb,
61 				 const struct netdev_queue *txq,
62 				 int *packets)
63 {
64 	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
65 
66 	while (bytelimit > 0) {
67 		struct sk_buff *nskb = q->dequeue(q);
68 
69 		if (!nskb)
70 			break;
71 
72 		bytelimit -= nskb->len; /* covers GSO len */
73 		skb->next = nskb;
74 		skb = nskb;
75 		(*packets)++; /* GSO counts as one pkt */
76 	}
77 	skb->next = NULL;
78 }
79 
80 /* This variant of try_bulk_dequeue_skb() makes sure
81  * all skbs in the chain are for the same txq
82  */
83 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
84 				      struct sk_buff *skb,
85 				      int *packets)
86 {
87 	int mapping = skb_get_queue_mapping(skb);
88 	struct sk_buff *nskb;
89 	int cnt = 0;
90 
91 	do {
92 		nskb = q->dequeue(q);
93 		if (!nskb)
94 			break;
95 		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
96 			q->skb_bad_txq = nskb;
97 			qdisc_qstats_backlog_inc(q, nskb);
98 			q->q.qlen++;
99 			break;
100 		}
101 		skb->next = nskb;
102 		skb = nskb;
103 	} while (++cnt < 8);
104 	(*packets) += cnt;
105 	skb->next = NULL;
106 }
107 
108 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
109  * A requeued skb (via q->gso_skb) can also be a SKB list.
110  */
111 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
112 				   int *packets)
113 {
114 	struct sk_buff *skb = q->gso_skb;
115 	const struct netdev_queue *txq = q->dev_queue;
116 
117 	*packets = 1;
118 	if (unlikely(skb)) {
119 		/* skb in gso_skb were already validated */
120 		*validate = false;
121 		/* check the reason of requeuing without tx lock first */
122 		txq = skb_get_tx_queue(txq->dev, skb);
123 		if (!netif_xmit_frozen_or_stopped(txq)) {
124 			q->gso_skb = NULL;
125 			qdisc_qstats_backlog_dec(q, skb);
126 			q->q.qlen--;
127 		} else
128 			skb = NULL;
129 		return skb;
130 	}
131 	*validate = true;
132 	skb = q->skb_bad_txq;
133 	if (unlikely(skb)) {
134 		/* check the reason of requeuing without tx lock first */
135 		txq = skb_get_tx_queue(txq->dev, skb);
136 		if (!netif_xmit_frozen_or_stopped(txq)) {
137 			q->skb_bad_txq = NULL;
138 			qdisc_qstats_backlog_dec(q, skb);
139 			q->q.qlen--;
140 			goto bulk;
141 		}
142 		return NULL;
143 	}
144 	if (!(q->flags & TCQ_F_ONETXQUEUE) ||
145 	    !netif_xmit_frozen_or_stopped(txq))
146 		skb = q->dequeue(q);
147 	if (skb) {
148 bulk:
149 		if (qdisc_may_bulk(q))
150 			try_bulk_dequeue_skb(q, skb, txq, packets);
151 		else
152 			try_bulk_dequeue_skb_slow(q, skb, packets);
153 	}
154 	return skb;
155 }
156 
157 /*
158  * Transmit possibly several skbs, and handle the return status as
159  * required. Owning running seqcount bit guarantees that
160  * only one CPU can execute this function.
161  *
162  * Returns to the caller:
163  *				0  - queue is empty or throttled.
164  *				>0 - queue is not empty.
165  */
166 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
167 		    struct net_device *dev, struct netdev_queue *txq,
168 		    spinlock_t *root_lock, bool validate)
169 {
170 	int ret = NETDEV_TX_BUSY;
171 
172 	/* And release qdisc */
173 	spin_unlock(root_lock);
174 
175 	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
176 	if (validate)
177 		skb = validate_xmit_skb_list(skb, dev);
178 
179 	if (likely(skb)) {
180 		HARD_TX_LOCK(dev, txq, smp_processor_id());
181 		if (!netif_xmit_frozen_or_stopped(txq))
182 			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
183 
184 		HARD_TX_UNLOCK(dev, txq);
185 	} else {
186 		spin_lock(root_lock);
187 		return qdisc_qlen(q);
188 	}
189 	spin_lock(root_lock);
190 
191 	if (dev_xmit_complete(ret)) {
192 		/* Driver sent out skb successfully or skb was consumed */
193 		ret = qdisc_qlen(q);
194 	} else {
195 		/* Driver returned NETDEV_TX_BUSY - requeue skb */
196 		if (unlikely(ret != NETDEV_TX_BUSY))
197 			net_warn_ratelimited("BUG %s code %d qlen %d\n",
198 					     dev->name, ret, q->q.qlen);
199 
200 		ret = dev_requeue_skb(skb, q);
201 	}
202 
203 	if (ret && netif_xmit_frozen_or_stopped(txq))
204 		ret = 0;
205 
206 	return ret;
207 }
208 
209 /*
210  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
211  *
212  * running seqcount guarantees only one CPU can process
213  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
214  * this queue.
215  *
216  *  netif_tx_lock serializes accesses to device driver.
217  *
218  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
219  *  if one is grabbed, another must be free.
220  *
221  * Note, that this procedure can be called by a watchdog timer
222  *
223  * Returns to the caller:
224  *				0  - queue is empty or throttled.
225  *				>0 - queue is not empty.
226  *
227  */
228 static inline int qdisc_restart(struct Qdisc *q, int *packets)
229 {
230 	struct netdev_queue *txq;
231 	struct net_device *dev;
232 	spinlock_t *root_lock;
233 	struct sk_buff *skb;
234 	bool validate;
235 
236 	/* Dequeue packet */
237 	skb = dequeue_skb(q, &validate, packets);
238 	if (unlikely(!skb))
239 		return 0;
240 
241 	root_lock = qdisc_lock(q);
242 	dev = qdisc_dev(q);
243 	txq = skb_get_tx_queue(dev, skb);
244 
245 	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
246 }
247 
248 void __qdisc_run(struct Qdisc *q)
249 {
250 	int quota = weight_p;
251 	int packets;
252 
253 	while (qdisc_restart(q, &packets)) {
254 		/*
255 		 * Ordered by possible occurrence: Postpone processing if
256 		 * 1. we've exceeded packet quota
257 		 * 2. another process needs the CPU;
258 		 */
259 		quota -= packets;
260 		if (quota <= 0 || need_resched()) {
261 			__netif_schedule(q);
262 			break;
263 		}
264 	}
265 
266 	qdisc_run_end(q);
267 }
268 
269 unsigned long dev_trans_start(struct net_device *dev)
270 {
271 	unsigned long val, res;
272 	unsigned int i;
273 
274 	if (is_vlan_dev(dev))
275 		dev = vlan_dev_real_dev(dev);
276 	res = netdev_get_tx_queue(dev, 0)->trans_start;
277 	for (i = 1; i < dev->num_tx_queues; i++) {
278 		val = netdev_get_tx_queue(dev, i)->trans_start;
279 		if (val && time_after(val, res))
280 			res = val;
281 	}
282 
283 	return res;
284 }
285 EXPORT_SYMBOL(dev_trans_start);
286 
287 static void dev_watchdog(unsigned long arg)
288 {
289 	struct net_device *dev = (struct net_device *)arg;
290 
291 	netif_tx_lock(dev);
292 	if (!qdisc_tx_is_noop(dev)) {
293 		if (netif_device_present(dev) &&
294 		    netif_running(dev) &&
295 		    netif_carrier_ok(dev)) {
296 			int some_queue_timedout = 0;
297 			unsigned int i;
298 			unsigned long trans_start;
299 
300 			for (i = 0; i < dev->num_tx_queues; i++) {
301 				struct netdev_queue *txq;
302 
303 				txq = netdev_get_tx_queue(dev, i);
304 				trans_start = txq->trans_start;
305 				if (netif_xmit_stopped(txq) &&
306 				    time_after(jiffies, (trans_start +
307 							 dev->watchdog_timeo))) {
308 					some_queue_timedout = 1;
309 					txq->trans_timeout++;
310 					break;
311 				}
312 			}
313 
314 			if (some_queue_timedout) {
315 				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
316 				       dev->name, netdev_drivername(dev), i);
317 				dev->netdev_ops->ndo_tx_timeout(dev);
318 			}
319 			if (!mod_timer(&dev->watchdog_timer,
320 				       round_jiffies(jiffies +
321 						     dev->watchdog_timeo)))
322 				dev_hold(dev);
323 		}
324 	}
325 	netif_tx_unlock(dev);
326 
327 	dev_put(dev);
328 }
329 
330 void __netdev_watchdog_up(struct net_device *dev)
331 {
332 	if (dev->netdev_ops->ndo_tx_timeout) {
333 		if (dev->watchdog_timeo <= 0)
334 			dev->watchdog_timeo = 5*HZ;
335 		if (!mod_timer(&dev->watchdog_timer,
336 			       round_jiffies(jiffies + dev->watchdog_timeo)))
337 			dev_hold(dev);
338 	}
339 }
340 
341 static void dev_watchdog_up(struct net_device *dev)
342 {
343 	__netdev_watchdog_up(dev);
344 }
345 
346 static void dev_watchdog_down(struct net_device *dev)
347 {
348 	netif_tx_lock_bh(dev);
349 	if (del_timer(&dev->watchdog_timer))
350 		dev_put(dev);
351 	netif_tx_unlock_bh(dev);
352 }
353 
354 /**
355  *	netif_carrier_on - set carrier
356  *	@dev: network device
357  *
358  * Device has detected that carrier.
359  */
360 void netif_carrier_on(struct net_device *dev)
361 {
362 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
363 		if (dev->reg_state == NETREG_UNINITIALIZED)
364 			return;
365 		atomic_inc(&dev->carrier_changes);
366 		linkwatch_fire_event(dev);
367 		if (netif_running(dev))
368 			__netdev_watchdog_up(dev);
369 	}
370 }
371 EXPORT_SYMBOL(netif_carrier_on);
372 
373 /**
374  *	netif_carrier_off - clear carrier
375  *	@dev: network device
376  *
377  * Device has detected loss of carrier.
378  */
379 void netif_carrier_off(struct net_device *dev)
380 {
381 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
382 		if (dev->reg_state == NETREG_UNINITIALIZED)
383 			return;
384 		atomic_inc(&dev->carrier_changes);
385 		linkwatch_fire_event(dev);
386 	}
387 }
388 EXPORT_SYMBOL(netif_carrier_off);
389 
390 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
391    under all circumstances. It is difficult to invent anything faster or
392    cheaper.
393  */
394 
395 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
396 			struct sk_buff **to_free)
397 {
398 	__qdisc_drop(skb, to_free);
399 	return NET_XMIT_CN;
400 }
401 
402 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
403 {
404 	return NULL;
405 }
406 
407 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
408 	.id		=	"noop",
409 	.priv_size	=	0,
410 	.enqueue	=	noop_enqueue,
411 	.dequeue	=	noop_dequeue,
412 	.peek		=	noop_dequeue,
413 	.owner		=	THIS_MODULE,
414 };
415 
416 static struct netdev_queue noop_netdev_queue = {
417 	.qdisc		=	&noop_qdisc,
418 	.qdisc_sleeping	=	&noop_qdisc,
419 };
420 
421 struct Qdisc noop_qdisc = {
422 	.enqueue	=	noop_enqueue,
423 	.dequeue	=	noop_dequeue,
424 	.flags		=	TCQ_F_BUILTIN,
425 	.ops		=	&noop_qdisc_ops,
426 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
427 	.dev_queue	=	&noop_netdev_queue,
428 	.running	=	SEQCNT_ZERO(noop_qdisc.running),
429 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
430 };
431 EXPORT_SYMBOL(noop_qdisc);
432 
433 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
434 {
435 	/* register_qdisc() assigns a default of noop_enqueue if unset,
436 	 * but __dev_queue_xmit() treats noqueue only as such
437 	 * if this is NULL - so clear it here. */
438 	qdisc->enqueue = NULL;
439 	return 0;
440 }
441 
442 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
443 	.id		=	"noqueue",
444 	.priv_size	=	0,
445 	.init		=	noqueue_init,
446 	.enqueue	=	noop_enqueue,
447 	.dequeue	=	noop_dequeue,
448 	.peek		=	noop_dequeue,
449 	.owner		=	THIS_MODULE,
450 };
451 
452 static const u8 prio2band[TC_PRIO_MAX + 1] = {
453 	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
454 };
455 
456 /* 3-band FIFO queue: old style, but should be a bit faster than
457    generic prio+fifo combination.
458  */
459 
460 #define PFIFO_FAST_BANDS 3
461 
462 /*
463  * Private data for a pfifo_fast scheduler containing:
464  * 	- queues for the three band
465  * 	- bitmap indicating which of the bands contain skbs
466  */
467 struct pfifo_fast_priv {
468 	u32 bitmap;
469 	struct sk_buff_head q[PFIFO_FAST_BANDS];
470 };
471 
472 /*
473  * Convert a bitmap to the first band number where an skb is queued, where:
474  * 	bitmap=0 means there are no skbs on any band.
475  * 	bitmap=1 means there is an skb on band 0.
476  *	bitmap=7 means there are skbs on all 3 bands, etc.
477  */
478 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
479 
480 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
481 					     int band)
482 {
483 	return priv->q + band;
484 }
485 
486 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
487 			      struct sk_buff **to_free)
488 {
489 	if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
490 		int band = prio2band[skb->priority & TC_PRIO_MAX];
491 		struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
492 		struct sk_buff_head *list = band2list(priv, band);
493 
494 		priv->bitmap |= (1 << band);
495 		qdisc->q.qlen++;
496 		return __qdisc_enqueue_tail(skb, qdisc, list);
497 	}
498 
499 	return qdisc_drop(skb, qdisc, to_free);
500 }
501 
502 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
503 {
504 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
505 	int band = bitmap2band[priv->bitmap];
506 
507 	if (likely(band >= 0)) {
508 		struct sk_buff_head *list = band2list(priv, band);
509 		struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
510 
511 		qdisc->q.qlen--;
512 		if (skb_queue_empty(list))
513 			priv->bitmap &= ~(1 << band);
514 
515 		return skb;
516 	}
517 
518 	return NULL;
519 }
520 
521 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
522 {
523 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
524 	int band = bitmap2band[priv->bitmap];
525 
526 	if (band >= 0) {
527 		struct sk_buff_head *list = band2list(priv, band);
528 
529 		return skb_peek(list);
530 	}
531 
532 	return NULL;
533 }
534 
535 static void pfifo_fast_reset(struct Qdisc *qdisc)
536 {
537 	int prio;
538 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
539 
540 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
541 		__qdisc_reset_queue(band2list(priv, prio));
542 
543 	priv->bitmap = 0;
544 	qdisc->qstats.backlog = 0;
545 	qdisc->q.qlen = 0;
546 }
547 
548 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
549 {
550 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
551 
552 	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
553 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
554 		goto nla_put_failure;
555 	return skb->len;
556 
557 nla_put_failure:
558 	return -1;
559 }
560 
561 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
562 {
563 	int prio;
564 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
565 
566 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
567 		__skb_queue_head_init(band2list(priv, prio));
568 
569 	/* Can by-pass the queue discipline */
570 	qdisc->flags |= TCQ_F_CAN_BYPASS;
571 	return 0;
572 }
573 
574 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
575 	.id		=	"pfifo_fast",
576 	.priv_size	=	sizeof(struct pfifo_fast_priv),
577 	.enqueue	=	pfifo_fast_enqueue,
578 	.dequeue	=	pfifo_fast_dequeue,
579 	.peek		=	pfifo_fast_peek,
580 	.init		=	pfifo_fast_init,
581 	.reset		=	pfifo_fast_reset,
582 	.dump		=	pfifo_fast_dump,
583 	.owner		=	THIS_MODULE,
584 };
585 EXPORT_SYMBOL(pfifo_fast_ops);
586 
587 static struct lock_class_key qdisc_tx_busylock;
588 static struct lock_class_key qdisc_running_key;
589 
590 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
591 			  const struct Qdisc_ops *ops)
592 {
593 	void *p;
594 	struct Qdisc *sch;
595 	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
596 	int err = -ENOBUFS;
597 	struct net_device *dev = dev_queue->dev;
598 
599 	p = kzalloc_node(size, GFP_KERNEL,
600 			 netdev_queue_numa_node_read(dev_queue));
601 
602 	if (!p)
603 		goto errout;
604 	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
605 	/* if we got non aligned memory, ask more and do alignment ourself */
606 	if (sch != p) {
607 		kfree(p);
608 		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
609 				 netdev_queue_numa_node_read(dev_queue));
610 		if (!p)
611 			goto errout;
612 		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
613 		sch->padded = (char *) sch - (char *) p;
614 	}
615 	skb_queue_head_init(&sch->q);
616 
617 	spin_lock_init(&sch->busylock);
618 	lockdep_set_class(&sch->busylock,
619 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
620 
621 	seqcount_init(&sch->running);
622 	lockdep_set_class(&sch->running,
623 			  dev->qdisc_running_key ?: &qdisc_running_key);
624 
625 	sch->ops = ops;
626 	sch->enqueue = ops->enqueue;
627 	sch->dequeue = ops->dequeue;
628 	sch->dev_queue = dev_queue;
629 	dev_hold(dev);
630 	atomic_set(&sch->refcnt, 1);
631 
632 	return sch;
633 errout:
634 	return ERR_PTR(err);
635 }
636 
637 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
638 				const struct Qdisc_ops *ops,
639 				unsigned int parentid)
640 {
641 	struct Qdisc *sch;
642 
643 	if (!try_module_get(ops->owner))
644 		return NULL;
645 
646 	sch = qdisc_alloc(dev_queue, ops);
647 	if (IS_ERR(sch)) {
648 		module_put(ops->owner);
649 		return NULL;
650 	}
651 	sch->parent = parentid;
652 
653 	if (!ops->init || ops->init(sch, NULL) == 0)
654 		return sch;
655 
656 	qdisc_destroy(sch);
657 	return NULL;
658 }
659 EXPORT_SYMBOL(qdisc_create_dflt);
660 
661 /* Under qdisc_lock(qdisc) and BH! */
662 
663 void qdisc_reset(struct Qdisc *qdisc)
664 {
665 	const struct Qdisc_ops *ops = qdisc->ops;
666 
667 	if (ops->reset)
668 		ops->reset(qdisc);
669 
670 	kfree_skb(qdisc->skb_bad_txq);
671 	qdisc->skb_bad_txq = NULL;
672 
673 	if (qdisc->gso_skb) {
674 		kfree_skb_list(qdisc->gso_skb);
675 		qdisc->gso_skb = NULL;
676 	}
677 	qdisc->q.qlen = 0;
678 }
679 EXPORT_SYMBOL(qdisc_reset);
680 
681 static void qdisc_rcu_free(struct rcu_head *head)
682 {
683 	struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
684 
685 	if (qdisc_is_percpu_stats(qdisc)) {
686 		free_percpu(qdisc->cpu_bstats);
687 		free_percpu(qdisc->cpu_qstats);
688 	}
689 
690 	kfree((char *) qdisc - qdisc->padded);
691 }
692 
693 void qdisc_destroy(struct Qdisc *qdisc)
694 {
695 	const struct Qdisc_ops  *ops = qdisc->ops;
696 
697 	if (qdisc->flags & TCQ_F_BUILTIN ||
698 	    !atomic_dec_and_test(&qdisc->refcnt))
699 		return;
700 
701 #ifdef CONFIG_NET_SCHED
702 	qdisc_hash_del(qdisc);
703 
704 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
705 #endif
706 	gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
707 	if (ops->reset)
708 		ops->reset(qdisc);
709 	if (ops->destroy)
710 		ops->destroy(qdisc);
711 
712 	module_put(ops->owner);
713 	dev_put(qdisc_dev(qdisc));
714 
715 	kfree_skb_list(qdisc->gso_skb);
716 	kfree_skb(qdisc->skb_bad_txq);
717 	/*
718 	 * gen_estimator est_timer() might access qdisc->q.lock,
719 	 * wait a RCU grace period before freeing qdisc.
720 	 */
721 	call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
722 }
723 EXPORT_SYMBOL(qdisc_destroy);
724 
725 /* Attach toplevel qdisc to device queue. */
726 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
727 			      struct Qdisc *qdisc)
728 {
729 	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
730 	spinlock_t *root_lock;
731 
732 	root_lock = qdisc_lock(oqdisc);
733 	spin_lock_bh(root_lock);
734 
735 	/* Prune old scheduler */
736 	if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
737 		qdisc_reset(oqdisc);
738 
739 	/* ... and graft new one */
740 	if (qdisc == NULL)
741 		qdisc = &noop_qdisc;
742 	dev_queue->qdisc_sleeping = qdisc;
743 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
744 
745 	spin_unlock_bh(root_lock);
746 
747 	return oqdisc;
748 }
749 EXPORT_SYMBOL(dev_graft_qdisc);
750 
751 static void attach_one_default_qdisc(struct net_device *dev,
752 				     struct netdev_queue *dev_queue,
753 				     void *_unused)
754 {
755 	struct Qdisc *qdisc;
756 	const struct Qdisc_ops *ops = default_qdisc_ops;
757 
758 	if (dev->priv_flags & IFF_NO_QUEUE)
759 		ops = &noqueue_qdisc_ops;
760 
761 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
762 	if (!qdisc) {
763 		netdev_info(dev, "activation failed\n");
764 		return;
765 	}
766 	if (!netif_is_multiqueue(dev))
767 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
768 	dev_queue->qdisc_sleeping = qdisc;
769 }
770 
771 static void attach_default_qdiscs(struct net_device *dev)
772 {
773 	struct netdev_queue *txq;
774 	struct Qdisc *qdisc;
775 
776 	txq = netdev_get_tx_queue(dev, 0);
777 
778 	if (!netif_is_multiqueue(dev) ||
779 	    dev->priv_flags & IFF_NO_QUEUE) {
780 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
781 		dev->qdisc = txq->qdisc_sleeping;
782 		atomic_inc(&dev->qdisc->refcnt);
783 	} else {
784 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
785 		if (qdisc) {
786 			dev->qdisc = qdisc;
787 			qdisc->ops->attach(qdisc);
788 		}
789 	}
790 #ifdef CONFIG_NET_SCHED
791 	if (dev->qdisc)
792 		qdisc_hash_add(dev->qdisc);
793 #endif
794 }
795 
796 static void transition_one_qdisc(struct net_device *dev,
797 				 struct netdev_queue *dev_queue,
798 				 void *_need_watchdog)
799 {
800 	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
801 	int *need_watchdog_p = _need_watchdog;
802 
803 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
804 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
805 
806 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
807 	if (need_watchdog_p) {
808 		dev_queue->trans_start = 0;
809 		*need_watchdog_p = 1;
810 	}
811 }
812 
813 void dev_activate(struct net_device *dev)
814 {
815 	int need_watchdog;
816 
817 	/* No queueing discipline is attached to device;
818 	 * create default one for devices, which need queueing
819 	 * and noqueue_qdisc for virtual interfaces
820 	 */
821 
822 	if (dev->qdisc == &noop_qdisc)
823 		attach_default_qdiscs(dev);
824 
825 	if (!netif_carrier_ok(dev))
826 		/* Delay activation until next carrier-on event */
827 		return;
828 
829 	need_watchdog = 0;
830 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
831 	if (dev_ingress_queue(dev))
832 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
833 
834 	if (need_watchdog) {
835 		netif_trans_update(dev);
836 		dev_watchdog_up(dev);
837 	}
838 }
839 EXPORT_SYMBOL(dev_activate);
840 
841 static void dev_deactivate_queue(struct net_device *dev,
842 				 struct netdev_queue *dev_queue,
843 				 void *_qdisc_default)
844 {
845 	struct Qdisc *qdisc_default = _qdisc_default;
846 	struct Qdisc *qdisc;
847 
848 	qdisc = rtnl_dereference(dev_queue->qdisc);
849 	if (qdisc) {
850 		spin_lock_bh(qdisc_lock(qdisc));
851 
852 		if (!(qdisc->flags & TCQ_F_BUILTIN))
853 			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
854 
855 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
856 		qdisc_reset(qdisc);
857 
858 		spin_unlock_bh(qdisc_lock(qdisc));
859 	}
860 }
861 
862 static bool some_qdisc_is_busy(struct net_device *dev)
863 {
864 	unsigned int i;
865 
866 	for (i = 0; i < dev->num_tx_queues; i++) {
867 		struct netdev_queue *dev_queue;
868 		spinlock_t *root_lock;
869 		struct Qdisc *q;
870 		int val;
871 
872 		dev_queue = netdev_get_tx_queue(dev, i);
873 		q = dev_queue->qdisc_sleeping;
874 		root_lock = qdisc_lock(q);
875 
876 		spin_lock_bh(root_lock);
877 
878 		val = (qdisc_is_running(q) ||
879 		       test_bit(__QDISC_STATE_SCHED, &q->state));
880 
881 		spin_unlock_bh(root_lock);
882 
883 		if (val)
884 			return true;
885 	}
886 	return false;
887 }
888 
889 /**
890  * 	dev_deactivate_many - deactivate transmissions on several devices
891  * 	@head: list of devices to deactivate
892  *
893  *	This function returns only when all outstanding transmissions
894  *	have completed, unless all devices are in dismantle phase.
895  */
896 void dev_deactivate_many(struct list_head *head)
897 {
898 	struct net_device *dev;
899 	bool sync_needed = false;
900 
901 	list_for_each_entry(dev, head, close_list) {
902 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
903 					 &noop_qdisc);
904 		if (dev_ingress_queue(dev))
905 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
906 					     &noop_qdisc);
907 
908 		dev_watchdog_down(dev);
909 		sync_needed |= !dev->dismantle;
910 	}
911 
912 	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
913 	 * This is avoided if all devices are in dismantle phase :
914 	 * Caller will call synchronize_net() for us
915 	 */
916 	if (sync_needed)
917 		synchronize_net();
918 
919 	/* Wait for outstanding qdisc_run calls. */
920 	list_for_each_entry(dev, head, close_list)
921 		while (some_qdisc_is_busy(dev))
922 			yield();
923 }
924 
925 void dev_deactivate(struct net_device *dev)
926 {
927 	LIST_HEAD(single);
928 
929 	list_add(&dev->close_list, &single);
930 	dev_deactivate_many(&single);
931 	list_del(&single);
932 }
933 EXPORT_SYMBOL(dev_deactivate);
934 
935 static void dev_init_scheduler_queue(struct net_device *dev,
936 				     struct netdev_queue *dev_queue,
937 				     void *_qdisc)
938 {
939 	struct Qdisc *qdisc = _qdisc;
940 
941 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
942 	dev_queue->qdisc_sleeping = qdisc;
943 }
944 
945 void dev_init_scheduler(struct net_device *dev)
946 {
947 	dev->qdisc = &noop_qdisc;
948 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
949 	if (dev_ingress_queue(dev))
950 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
951 
952 	setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
953 }
954 
955 static void shutdown_scheduler_queue(struct net_device *dev,
956 				     struct netdev_queue *dev_queue,
957 				     void *_qdisc_default)
958 {
959 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
960 	struct Qdisc *qdisc_default = _qdisc_default;
961 
962 	if (qdisc) {
963 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
964 		dev_queue->qdisc_sleeping = qdisc_default;
965 
966 		qdisc_destroy(qdisc);
967 	}
968 }
969 
970 void dev_shutdown(struct net_device *dev)
971 {
972 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
973 	if (dev_ingress_queue(dev))
974 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
975 	qdisc_destroy(dev->qdisc);
976 	dev->qdisc = &noop_qdisc;
977 
978 	WARN_ON(timer_pending(&dev->watchdog_timer));
979 }
980 
981 void psched_ratecfg_precompute(struct psched_ratecfg *r,
982 			       const struct tc_ratespec *conf,
983 			       u64 rate64)
984 {
985 	memset(r, 0, sizeof(*r));
986 	r->overhead = conf->overhead;
987 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
988 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
989 	r->mult = 1;
990 	/*
991 	 * The deal here is to replace a divide by a reciprocal one
992 	 * in fast path (a reciprocal divide is a multiply and a shift)
993 	 *
994 	 * Normal formula would be :
995 	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
996 	 *
997 	 * We compute mult/shift to use instead :
998 	 *  time_in_ns = (len * mult) >> shift;
999 	 *
1000 	 * We try to get the highest possible mult value for accuracy,
1001 	 * but have to make sure no overflows will ever happen.
1002 	 */
1003 	if (r->rate_bytes_ps > 0) {
1004 		u64 factor = NSEC_PER_SEC;
1005 
1006 		for (;;) {
1007 			r->mult = div64_u64(factor, r->rate_bytes_ps);
1008 			if (r->mult & (1U << 31) || factor & (1ULL << 63))
1009 				break;
1010 			factor <<= 1;
1011 			r->shift++;
1012 		}
1013 	}
1014 }
1015 EXPORT_SYMBOL(psched_ratecfg_precompute);
1016