xref: /linux/net/sched/sch_generic.c (revision 8795a739e5c72abeec51caf36b6df2b37e5720c5)
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 */
35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36 EXPORT_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 
51 static 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 
86 static 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 
96 static 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 
120 static 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 
152 static 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  */
176 static 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  */
202 static 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 	}
250 validate:
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) {
265 bulk:
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 	}
271 trace:
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  */
285 bool 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  */
357 static 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 
379 void __qdisc_run(struct Qdisc *q)
380 {
381 	int quota = dev_tx_weight;
382 	int packets;
383 
384 	while (qdisc_restart(q, &packets)) {
385 		/*
386 		 * Ordered by possible occurrence: Postpone processing if
387 		 * 1. we've exceeded packet quota
388 		 * 2. another process needs the CPU;
389 		 */
390 		quota -= packets;
391 		if (quota <= 0 || need_resched()) {
392 			__netif_schedule(q);
393 			break;
394 		}
395 	}
396 }
397 
398 unsigned long dev_trans_start(struct net_device *dev)
399 {
400 	unsigned long val, res;
401 	unsigned int i;
402 
403 	if (is_vlan_dev(dev))
404 		dev = vlan_dev_real_dev(dev);
405 	else if (netif_is_macvlan(dev))
406 		dev = macvlan_dev_real_dev(dev);
407 	res = netdev_get_tx_queue(dev, 0)->trans_start;
408 	for (i = 1; i < dev->num_tx_queues; i++) {
409 		val = netdev_get_tx_queue(dev, i)->trans_start;
410 		if (val && time_after(val, res))
411 			res = val;
412 	}
413 
414 	return res;
415 }
416 EXPORT_SYMBOL(dev_trans_start);
417 
418 static void dev_watchdog(struct timer_list *t)
419 {
420 	struct net_device *dev = from_timer(dev, t, watchdog_timer);
421 
422 	netif_tx_lock(dev);
423 	if (!qdisc_tx_is_noop(dev)) {
424 		if (netif_device_present(dev) &&
425 		    netif_running(dev) &&
426 		    netif_carrier_ok(dev)) {
427 			int some_queue_timedout = 0;
428 			unsigned int i;
429 			unsigned long trans_start;
430 
431 			for (i = 0; i < dev->num_tx_queues; i++) {
432 				struct netdev_queue *txq;
433 
434 				txq = netdev_get_tx_queue(dev, i);
435 				trans_start = txq->trans_start;
436 				if (netif_xmit_stopped(txq) &&
437 				    time_after(jiffies, (trans_start +
438 							 dev->watchdog_timeo))) {
439 					some_queue_timedout = 1;
440 					txq->trans_timeout++;
441 					break;
442 				}
443 			}
444 
445 			if (some_queue_timedout) {
446 				trace_net_dev_xmit_timeout(dev, i);
447 				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
448 				       dev->name, netdev_drivername(dev), i);
449 				dev->netdev_ops->ndo_tx_timeout(dev);
450 			}
451 			if (!mod_timer(&dev->watchdog_timer,
452 				       round_jiffies(jiffies +
453 						     dev->watchdog_timeo)))
454 				dev_hold(dev);
455 		}
456 	}
457 	netif_tx_unlock(dev);
458 
459 	dev_put(dev);
460 }
461 
462 void __netdev_watchdog_up(struct net_device *dev)
463 {
464 	if (dev->netdev_ops->ndo_tx_timeout) {
465 		if (dev->watchdog_timeo <= 0)
466 			dev->watchdog_timeo = 5*HZ;
467 		if (!mod_timer(&dev->watchdog_timer,
468 			       round_jiffies(jiffies + dev->watchdog_timeo)))
469 			dev_hold(dev);
470 	}
471 }
472 
473 static void dev_watchdog_up(struct net_device *dev)
474 {
475 	__netdev_watchdog_up(dev);
476 }
477 
478 static void dev_watchdog_down(struct net_device *dev)
479 {
480 	netif_tx_lock_bh(dev);
481 	if (del_timer(&dev->watchdog_timer))
482 		dev_put(dev);
483 	netif_tx_unlock_bh(dev);
484 }
485 
486 /**
487  *	netif_carrier_on - set carrier
488  *	@dev: network device
489  *
490  * Device has detected acquisition of carrier.
491  */
492 void netif_carrier_on(struct net_device *dev)
493 {
494 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
495 		if (dev->reg_state == NETREG_UNINITIALIZED)
496 			return;
497 		atomic_inc(&dev->carrier_up_count);
498 		linkwatch_fire_event(dev);
499 		if (netif_running(dev))
500 			__netdev_watchdog_up(dev);
501 	}
502 }
503 EXPORT_SYMBOL(netif_carrier_on);
504 
505 /**
506  *	netif_carrier_off - clear carrier
507  *	@dev: network device
508  *
509  * Device has detected loss of carrier.
510  */
511 void netif_carrier_off(struct net_device *dev)
512 {
513 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
514 		if (dev->reg_state == NETREG_UNINITIALIZED)
515 			return;
516 		atomic_inc(&dev->carrier_down_count);
517 		linkwatch_fire_event(dev);
518 	}
519 }
520 EXPORT_SYMBOL(netif_carrier_off);
521 
522 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
523    under all circumstances. It is difficult to invent anything faster or
524    cheaper.
525  */
526 
527 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
528 			struct sk_buff **to_free)
529 {
530 	__qdisc_drop(skb, to_free);
531 	return NET_XMIT_CN;
532 }
533 
534 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
535 {
536 	return NULL;
537 }
538 
539 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
540 	.id		=	"noop",
541 	.priv_size	=	0,
542 	.enqueue	=	noop_enqueue,
543 	.dequeue	=	noop_dequeue,
544 	.peek		=	noop_dequeue,
545 	.owner		=	THIS_MODULE,
546 };
547 
548 static struct netdev_queue noop_netdev_queue = {
549 	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
550 	.qdisc_sleeping	=	&noop_qdisc,
551 };
552 
553 struct Qdisc noop_qdisc = {
554 	.enqueue	=	noop_enqueue,
555 	.dequeue	=	noop_dequeue,
556 	.flags		=	TCQ_F_BUILTIN,
557 	.ops		=	&noop_qdisc_ops,
558 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
559 	.dev_queue	=	&noop_netdev_queue,
560 	.running	=	SEQCNT_ZERO(noop_qdisc.running),
561 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
562 	.gso_skb = {
563 		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
564 		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
565 		.qlen = 0,
566 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
567 	},
568 	.skb_bad_txq = {
569 		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
570 		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
571 		.qlen = 0,
572 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
573 	},
574 };
575 EXPORT_SYMBOL(noop_qdisc);
576 
577 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
578 			struct netlink_ext_ack *extack)
579 {
580 	/* register_qdisc() assigns a default of noop_enqueue if unset,
581 	 * but __dev_queue_xmit() treats noqueue only as such
582 	 * if this is NULL - so clear it here. */
583 	qdisc->enqueue = NULL;
584 	return 0;
585 }
586 
587 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
588 	.id		=	"noqueue",
589 	.priv_size	=	0,
590 	.init		=	noqueue_init,
591 	.enqueue	=	noop_enqueue,
592 	.dequeue	=	noop_dequeue,
593 	.peek		=	noop_dequeue,
594 	.owner		=	THIS_MODULE,
595 };
596 
597 static const u8 prio2band[TC_PRIO_MAX + 1] = {
598 	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
599 };
600 
601 /* 3-band FIFO queue: old style, but should be a bit faster than
602    generic prio+fifo combination.
603  */
604 
605 #define PFIFO_FAST_BANDS 3
606 
607 /*
608  * Private data for a pfifo_fast scheduler containing:
609  *	- rings for priority bands
610  */
611 struct pfifo_fast_priv {
612 	struct skb_array q[PFIFO_FAST_BANDS];
613 };
614 
615 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
616 					  int band)
617 {
618 	return &priv->q[band];
619 }
620 
621 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
622 			      struct sk_buff **to_free)
623 {
624 	int band = prio2band[skb->priority & TC_PRIO_MAX];
625 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
626 	struct skb_array *q = band2list(priv, band);
627 	unsigned int pkt_len = qdisc_pkt_len(skb);
628 	int err;
629 
630 	err = skb_array_produce(q, skb);
631 
632 	if (unlikely(err)) {
633 		if (qdisc_is_percpu_stats(qdisc))
634 			return qdisc_drop_cpu(skb, qdisc, to_free);
635 		else
636 			return qdisc_drop(skb, qdisc, to_free);
637 	}
638 
639 	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
640 	return NET_XMIT_SUCCESS;
641 }
642 
643 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
644 {
645 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
646 	struct sk_buff *skb = NULL;
647 	int band;
648 
649 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
650 		struct skb_array *q = band2list(priv, band);
651 
652 		if (__skb_array_empty(q))
653 			continue;
654 
655 		skb = __skb_array_consume(q);
656 	}
657 	if (likely(skb)) {
658 		qdisc_update_stats_at_dequeue(qdisc, skb);
659 	} else {
660 		qdisc->empty = true;
661 	}
662 
663 	return skb;
664 }
665 
666 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
667 {
668 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
669 	struct sk_buff *skb = NULL;
670 	int band;
671 
672 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
673 		struct skb_array *q = band2list(priv, band);
674 
675 		skb = __skb_array_peek(q);
676 	}
677 
678 	return skb;
679 }
680 
681 static void pfifo_fast_reset(struct Qdisc *qdisc)
682 {
683 	int i, band;
684 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
685 
686 	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
687 		struct skb_array *q = band2list(priv, band);
688 		struct sk_buff *skb;
689 
690 		/* NULL ring is possible if destroy path is due to a failed
691 		 * skb_array_init() in pfifo_fast_init() case.
692 		 */
693 		if (!q->ring.queue)
694 			continue;
695 
696 		while ((skb = __skb_array_consume(q)) != NULL)
697 			kfree_skb(skb);
698 	}
699 
700 	if (qdisc_is_percpu_stats(qdisc)) {
701 		for_each_possible_cpu(i) {
702 			struct gnet_stats_queue *q;
703 
704 			q = per_cpu_ptr(qdisc->cpu_qstats, i);
705 			q->backlog = 0;
706 			q->qlen = 0;
707 		}
708 	}
709 }
710 
711 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
712 {
713 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
714 
715 	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
716 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
717 		goto nla_put_failure;
718 	return skb->len;
719 
720 nla_put_failure:
721 	return -1;
722 }
723 
724 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
725 			   struct netlink_ext_ack *extack)
726 {
727 	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
728 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
729 	int prio;
730 
731 	/* guard against zero length rings */
732 	if (!qlen)
733 		return -EINVAL;
734 
735 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
736 		struct skb_array *q = band2list(priv, prio);
737 		int err;
738 
739 		err = skb_array_init(q, qlen, GFP_KERNEL);
740 		if (err)
741 			return -ENOMEM;
742 	}
743 
744 	/* Can by-pass the queue discipline */
745 	qdisc->flags |= TCQ_F_CAN_BYPASS;
746 	return 0;
747 }
748 
749 static void pfifo_fast_destroy(struct Qdisc *sch)
750 {
751 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
752 	int prio;
753 
754 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
755 		struct skb_array *q = band2list(priv, prio);
756 
757 		/* NULL ring is possible if destroy path is due to a failed
758 		 * skb_array_init() in pfifo_fast_init() case.
759 		 */
760 		if (!q->ring.queue)
761 			continue;
762 		/* Destroy ring but no need to kfree_skb because a call to
763 		 * pfifo_fast_reset() has already done that work.
764 		 */
765 		ptr_ring_cleanup(&q->ring, NULL);
766 	}
767 }
768 
769 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
770 					  unsigned int new_len)
771 {
772 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
773 	struct skb_array *bands[PFIFO_FAST_BANDS];
774 	int prio;
775 
776 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
777 		struct skb_array *q = band2list(priv, prio);
778 
779 		bands[prio] = q;
780 	}
781 
782 	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
783 					 GFP_KERNEL);
784 }
785 
786 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
787 	.id		=	"pfifo_fast",
788 	.priv_size	=	sizeof(struct pfifo_fast_priv),
789 	.enqueue	=	pfifo_fast_enqueue,
790 	.dequeue	=	pfifo_fast_dequeue,
791 	.peek		=	pfifo_fast_peek,
792 	.init		=	pfifo_fast_init,
793 	.destroy	=	pfifo_fast_destroy,
794 	.reset		=	pfifo_fast_reset,
795 	.dump		=	pfifo_fast_dump,
796 	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
797 	.owner		=	THIS_MODULE,
798 	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
799 };
800 EXPORT_SYMBOL(pfifo_fast_ops);
801 
802 static struct lock_class_key qdisc_tx_busylock;
803 static struct lock_class_key qdisc_running_key;
804 
805 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
806 			  const struct Qdisc_ops *ops,
807 			  struct netlink_ext_ack *extack)
808 {
809 	void *p;
810 	struct Qdisc *sch;
811 	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
812 	int err = -ENOBUFS;
813 	struct net_device *dev;
814 
815 	if (!dev_queue) {
816 		NL_SET_ERR_MSG(extack, "No device queue given");
817 		err = -EINVAL;
818 		goto errout;
819 	}
820 
821 	dev = dev_queue->dev;
822 	p = kzalloc_node(size, GFP_KERNEL,
823 			 netdev_queue_numa_node_read(dev_queue));
824 
825 	if (!p)
826 		goto errout;
827 	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
828 	/* if we got non aligned memory, ask more and do alignment ourself */
829 	if (sch != p) {
830 		kfree(p);
831 		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
832 				 netdev_queue_numa_node_read(dev_queue));
833 		if (!p)
834 			goto errout;
835 		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
836 		sch->padded = (char *) sch - (char *) p;
837 	}
838 	__skb_queue_head_init(&sch->gso_skb);
839 	__skb_queue_head_init(&sch->skb_bad_txq);
840 	qdisc_skb_head_init(&sch->q);
841 	spin_lock_init(&sch->q.lock);
842 
843 	if (ops->static_flags & TCQ_F_CPUSTATS) {
844 		sch->cpu_bstats =
845 			netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
846 		if (!sch->cpu_bstats)
847 			goto errout1;
848 
849 		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
850 		if (!sch->cpu_qstats) {
851 			free_percpu(sch->cpu_bstats);
852 			goto errout1;
853 		}
854 	}
855 
856 	spin_lock_init(&sch->busylock);
857 	lockdep_set_class(&sch->busylock,
858 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
859 
860 	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
861 	spin_lock_init(&sch->seqlock);
862 	lockdep_set_class(&sch->busylock,
863 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
864 
865 	seqcount_init(&sch->running);
866 	lockdep_set_class(&sch->running,
867 			  dev->qdisc_running_key ?: &qdisc_running_key);
868 
869 	sch->ops = ops;
870 	sch->flags = ops->static_flags;
871 	sch->enqueue = ops->enqueue;
872 	sch->dequeue = ops->dequeue;
873 	sch->dev_queue = dev_queue;
874 	sch->empty = true;
875 	dev_hold(dev);
876 	refcount_set(&sch->refcnt, 1);
877 
878 	return sch;
879 errout1:
880 	kfree(p);
881 errout:
882 	return ERR_PTR(err);
883 }
884 
885 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
886 				const struct Qdisc_ops *ops,
887 				unsigned int parentid,
888 				struct netlink_ext_ack *extack)
889 {
890 	struct Qdisc *sch;
891 
892 	if (!try_module_get(ops->owner)) {
893 		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
894 		return NULL;
895 	}
896 
897 	sch = qdisc_alloc(dev_queue, ops, extack);
898 	if (IS_ERR(sch)) {
899 		module_put(ops->owner);
900 		return NULL;
901 	}
902 	sch->parent = parentid;
903 
904 	if (!ops->init || ops->init(sch, NULL, extack) == 0)
905 		return sch;
906 
907 	qdisc_put(sch);
908 	return NULL;
909 }
910 EXPORT_SYMBOL(qdisc_create_dflt);
911 
912 /* Under qdisc_lock(qdisc) and BH! */
913 
914 void qdisc_reset(struct Qdisc *qdisc)
915 {
916 	const struct Qdisc_ops *ops = qdisc->ops;
917 	struct sk_buff *skb, *tmp;
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 }
935 EXPORT_SYMBOL(qdisc_reset);
936 
937 void 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 
947 static 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 
954 static void qdisc_destroy(struct Qdisc *qdisc)
955 {
956 	const struct Qdisc_ops  *ops = qdisc->ops;
957 	struct sk_buff *skb, *tmp;
958 
959 #ifdef CONFIG_NET_SCHED
960 	qdisc_hash_del(qdisc);
961 
962 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
963 #endif
964 	gen_kill_estimator(&qdisc->rate_est);
965 	if (ops->reset)
966 		ops->reset(qdisc);
967 	if (ops->destroy)
968 		ops->destroy(qdisc);
969 
970 	module_put(ops->owner);
971 	dev_put(qdisc_dev(qdisc));
972 
973 	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
974 		__skb_unlink(skb, &qdisc->gso_skb);
975 		kfree_skb_list(skb);
976 	}
977 
978 	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
979 		__skb_unlink(skb, &qdisc->skb_bad_txq);
980 		kfree_skb_list(skb);
981 	}
982 
983 	call_rcu(&qdisc->rcu, qdisc_free_cb);
984 }
985 
986 void qdisc_put(struct Qdisc *qdisc)
987 {
988 	if (!qdisc)
989 		return;
990 
991 	if (qdisc->flags & TCQ_F_BUILTIN ||
992 	    !refcount_dec_and_test(&qdisc->refcnt))
993 		return;
994 
995 	qdisc_destroy(qdisc);
996 }
997 EXPORT_SYMBOL(qdisc_put);
998 
999 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1000  * Intended to be used as optimization, this function only takes rtnl lock if
1001  * qdisc reference counter reached zero.
1002  */
1003 
1004 void qdisc_put_unlocked(struct Qdisc *qdisc)
1005 {
1006 	if (qdisc->flags & TCQ_F_BUILTIN ||
1007 	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1008 		return;
1009 
1010 	qdisc_destroy(qdisc);
1011 	rtnl_unlock();
1012 }
1013 EXPORT_SYMBOL(qdisc_put_unlocked);
1014 
1015 /* Attach toplevel qdisc to device queue. */
1016 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1017 			      struct Qdisc *qdisc)
1018 {
1019 	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1020 	spinlock_t *root_lock;
1021 
1022 	root_lock = qdisc_lock(oqdisc);
1023 	spin_lock_bh(root_lock);
1024 
1025 	/* ... and graft new one */
1026 	if (qdisc == NULL)
1027 		qdisc = &noop_qdisc;
1028 	dev_queue->qdisc_sleeping = qdisc;
1029 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1030 
1031 	spin_unlock_bh(root_lock);
1032 
1033 	return oqdisc;
1034 }
1035 EXPORT_SYMBOL(dev_graft_qdisc);
1036 
1037 static void attach_one_default_qdisc(struct net_device *dev,
1038 				     struct netdev_queue *dev_queue,
1039 				     void *_unused)
1040 {
1041 	struct Qdisc *qdisc;
1042 	const struct Qdisc_ops *ops = default_qdisc_ops;
1043 
1044 	if (dev->priv_flags & IFF_NO_QUEUE)
1045 		ops = &noqueue_qdisc_ops;
1046 
1047 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1048 	if (!qdisc) {
1049 		netdev_info(dev, "activation failed\n");
1050 		return;
1051 	}
1052 	if (!netif_is_multiqueue(dev))
1053 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1054 	dev_queue->qdisc_sleeping = qdisc;
1055 }
1056 
1057 static void attach_default_qdiscs(struct net_device *dev)
1058 {
1059 	struct netdev_queue *txq;
1060 	struct Qdisc *qdisc;
1061 
1062 	txq = netdev_get_tx_queue(dev, 0);
1063 
1064 	if (!netif_is_multiqueue(dev) ||
1065 	    dev->priv_flags & IFF_NO_QUEUE) {
1066 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1067 		dev->qdisc = txq->qdisc_sleeping;
1068 		qdisc_refcount_inc(dev->qdisc);
1069 	} else {
1070 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1071 		if (qdisc) {
1072 			dev->qdisc = qdisc;
1073 			qdisc->ops->attach(qdisc);
1074 		}
1075 	}
1076 #ifdef CONFIG_NET_SCHED
1077 	if (dev->qdisc != &noop_qdisc)
1078 		qdisc_hash_add(dev->qdisc, false);
1079 #endif
1080 }
1081 
1082 static void transition_one_qdisc(struct net_device *dev,
1083 				 struct netdev_queue *dev_queue,
1084 				 void *_need_watchdog)
1085 {
1086 	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1087 	int *need_watchdog_p = _need_watchdog;
1088 
1089 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1090 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1091 
1092 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1093 	if (need_watchdog_p) {
1094 		dev_queue->trans_start = 0;
1095 		*need_watchdog_p = 1;
1096 	}
1097 }
1098 
1099 void dev_activate(struct net_device *dev)
1100 {
1101 	int need_watchdog;
1102 
1103 	/* No queueing discipline is attached to device;
1104 	 * create default one for devices, which need queueing
1105 	 * and noqueue_qdisc for virtual interfaces
1106 	 */
1107 
1108 	if (dev->qdisc == &noop_qdisc)
1109 		attach_default_qdiscs(dev);
1110 
1111 	if (!netif_carrier_ok(dev))
1112 		/* Delay activation until next carrier-on event */
1113 		return;
1114 
1115 	need_watchdog = 0;
1116 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1117 	if (dev_ingress_queue(dev))
1118 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1119 
1120 	if (need_watchdog) {
1121 		netif_trans_update(dev);
1122 		dev_watchdog_up(dev);
1123 	}
1124 }
1125 EXPORT_SYMBOL(dev_activate);
1126 
1127 static void dev_deactivate_queue(struct net_device *dev,
1128 				 struct netdev_queue *dev_queue,
1129 				 void *_qdisc_default)
1130 {
1131 	struct Qdisc *qdisc_default = _qdisc_default;
1132 	struct Qdisc *qdisc;
1133 
1134 	qdisc = rtnl_dereference(dev_queue->qdisc);
1135 	if (qdisc) {
1136 		bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1137 
1138 		if (nolock)
1139 			spin_lock_bh(&qdisc->seqlock);
1140 		spin_lock_bh(qdisc_lock(qdisc));
1141 
1142 		if (!(qdisc->flags & TCQ_F_BUILTIN))
1143 			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1144 
1145 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1146 		qdisc_reset(qdisc);
1147 
1148 		spin_unlock_bh(qdisc_lock(qdisc));
1149 		if (nolock)
1150 			spin_unlock_bh(&qdisc->seqlock);
1151 	}
1152 }
1153 
1154 static bool some_qdisc_is_busy(struct net_device *dev)
1155 {
1156 	unsigned int i;
1157 
1158 	for (i = 0; i < dev->num_tx_queues; i++) {
1159 		struct netdev_queue *dev_queue;
1160 		spinlock_t *root_lock;
1161 		struct Qdisc *q;
1162 		int val;
1163 
1164 		dev_queue = netdev_get_tx_queue(dev, i);
1165 		q = dev_queue->qdisc_sleeping;
1166 
1167 		root_lock = qdisc_lock(q);
1168 		spin_lock_bh(root_lock);
1169 
1170 		val = (qdisc_is_running(q) ||
1171 		       test_bit(__QDISC_STATE_SCHED, &q->state));
1172 
1173 		spin_unlock_bh(root_lock);
1174 
1175 		if (val)
1176 			return true;
1177 	}
1178 	return false;
1179 }
1180 
1181 static void dev_qdisc_reset(struct net_device *dev,
1182 			    struct netdev_queue *dev_queue,
1183 			    void *none)
1184 {
1185 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1186 
1187 	if (qdisc)
1188 		qdisc_reset(qdisc);
1189 }
1190 
1191 /**
1192  * 	dev_deactivate_many - deactivate transmissions on several devices
1193  * 	@head: list of devices to deactivate
1194  *
1195  *	This function returns only when all outstanding transmissions
1196  *	have completed, unless all devices are in dismantle phase.
1197  */
1198 void dev_deactivate_many(struct list_head *head)
1199 {
1200 	struct net_device *dev;
1201 
1202 	list_for_each_entry(dev, head, close_list) {
1203 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1204 					 &noop_qdisc);
1205 		if (dev_ingress_queue(dev))
1206 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1207 					     &noop_qdisc);
1208 
1209 		dev_watchdog_down(dev);
1210 	}
1211 
1212 	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
1213 	 * This is avoided if all devices are in dismantle phase :
1214 	 * Caller will call synchronize_net() for us
1215 	 */
1216 	synchronize_net();
1217 
1218 	/* Wait for outstanding qdisc_run calls. */
1219 	list_for_each_entry(dev, head, close_list) {
1220 		while (some_qdisc_is_busy(dev))
1221 			yield();
1222 		/* The new qdisc is assigned at this point so we can safely
1223 		 * unwind stale skb lists and qdisc statistics
1224 		 */
1225 		netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1226 		if (dev_ingress_queue(dev))
1227 			dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1228 	}
1229 }
1230 
1231 void dev_deactivate(struct net_device *dev)
1232 {
1233 	LIST_HEAD(single);
1234 
1235 	list_add(&dev->close_list, &single);
1236 	dev_deactivate_many(&single);
1237 	list_del(&single);
1238 }
1239 EXPORT_SYMBOL(dev_deactivate);
1240 
1241 static int qdisc_change_tx_queue_len(struct net_device *dev,
1242 				     struct netdev_queue *dev_queue)
1243 {
1244 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1245 	const struct Qdisc_ops *ops = qdisc->ops;
1246 
1247 	if (ops->change_tx_queue_len)
1248 		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1249 	return 0;
1250 }
1251 
1252 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1253 {
1254 	bool up = dev->flags & IFF_UP;
1255 	unsigned int i;
1256 	int ret = 0;
1257 
1258 	if (up)
1259 		dev_deactivate(dev);
1260 
1261 	for (i = 0; i < dev->num_tx_queues; i++) {
1262 		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1263 
1264 		/* TODO: revert changes on a partial failure */
1265 		if (ret)
1266 			break;
1267 	}
1268 
1269 	if (up)
1270 		dev_activate(dev);
1271 	return ret;
1272 }
1273 
1274 static void dev_init_scheduler_queue(struct net_device *dev,
1275 				     struct netdev_queue *dev_queue,
1276 				     void *_qdisc)
1277 {
1278 	struct Qdisc *qdisc = _qdisc;
1279 
1280 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1281 	dev_queue->qdisc_sleeping = qdisc;
1282 }
1283 
1284 void dev_init_scheduler(struct net_device *dev)
1285 {
1286 	dev->qdisc = &noop_qdisc;
1287 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1288 	if (dev_ingress_queue(dev))
1289 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1290 
1291 	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1292 }
1293 
1294 static void shutdown_scheduler_queue(struct net_device *dev,
1295 				     struct netdev_queue *dev_queue,
1296 				     void *_qdisc_default)
1297 {
1298 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1299 	struct Qdisc *qdisc_default = _qdisc_default;
1300 
1301 	if (qdisc) {
1302 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1303 		dev_queue->qdisc_sleeping = qdisc_default;
1304 
1305 		qdisc_put(qdisc);
1306 	}
1307 }
1308 
1309 void dev_shutdown(struct net_device *dev)
1310 {
1311 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1312 	if (dev_ingress_queue(dev))
1313 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1314 	qdisc_put(dev->qdisc);
1315 	dev->qdisc = &noop_qdisc;
1316 
1317 	WARN_ON(timer_pending(&dev->watchdog_timer));
1318 }
1319 
1320 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1321 			       const struct tc_ratespec *conf,
1322 			       u64 rate64)
1323 {
1324 	memset(r, 0, sizeof(*r));
1325 	r->overhead = conf->overhead;
1326 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1327 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1328 	r->mult = 1;
1329 	/*
1330 	 * The deal here is to replace a divide by a reciprocal one
1331 	 * in fast path (a reciprocal divide is a multiply and a shift)
1332 	 *
1333 	 * Normal formula would be :
1334 	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1335 	 *
1336 	 * We compute mult/shift to use instead :
1337 	 *  time_in_ns = (len * mult) >> shift;
1338 	 *
1339 	 * We try to get the highest possible mult value for accuracy,
1340 	 * but have to make sure no overflows will ever happen.
1341 	 */
1342 	if (r->rate_bytes_ps > 0) {
1343 		u64 factor = NSEC_PER_SEC;
1344 
1345 		for (;;) {
1346 			r->mult = div64_u64(factor, r->rate_bytes_ps);
1347 			if (r->mult & (1U << 31) || factor & (1ULL << 63))
1348 				break;
1349 			factor <<= 1;
1350 			r->shift++;
1351 		}
1352 	}
1353 }
1354 EXPORT_SYMBOL(psched_ratecfg_precompute);
1355 
1356 static void mini_qdisc_rcu_func(struct rcu_head *head)
1357 {
1358 }
1359 
1360 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1361 			  struct tcf_proto *tp_head)
1362 {
1363 	/* Protected with chain0->filter_chain_lock.
1364 	 * Can't access chain directly because tp_head can be NULL.
1365 	 */
1366 	struct mini_Qdisc *miniq_old =
1367 		rcu_dereference_protected(*miniqp->p_miniq, 1);
1368 	struct mini_Qdisc *miniq;
1369 
1370 	if (!tp_head) {
1371 		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1372 		/* Wait for flying RCU callback before it is freed. */
1373 		rcu_barrier();
1374 		return;
1375 	}
1376 
1377 	miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1378 		&miniqp->miniq1 : &miniqp->miniq2;
1379 
1380 	/* We need to make sure that readers won't see the miniq
1381 	 * we are about to modify. So wait until previous call_rcu callback
1382 	 * is done.
1383 	 */
1384 	rcu_barrier();
1385 	miniq->filter_list = tp_head;
1386 	rcu_assign_pointer(*miniqp->p_miniq, miniq);
1387 
1388 	if (miniq_old)
1389 		/* This is counterpart of the rcu barriers above. We need to
1390 		 * block potential new user of miniq_old until all readers
1391 		 * are not seeing it.
1392 		 */
1393 		call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1394 }
1395 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1396 
1397 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1398 			  struct mini_Qdisc __rcu **p_miniq)
1399 {
1400 	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1401 	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1402 	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1403 	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1404 	miniqp->p_miniq = p_miniq;
1405 }
1406 EXPORT_SYMBOL(mini_qdisc_pair_init);
1407