xref: /linux/net/sched/sch_generic.c (revision 90d32e92011eaae8e70a9169b4e7acf4ca8f9d3a)
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 <net/hotdata.h>
31 #include <trace/events/qdisc.h>
32 #include <trace/events/net.h>
33 #include <net/xfrm.h>
34 
35 /* Qdisc to use by default */
36 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
37 EXPORT_SYMBOL(default_qdisc_ops);
38 
39 static void qdisc_maybe_clear_missed(struct Qdisc *q,
40 				     const struct netdev_queue *txq)
41 {
42 	clear_bit(__QDISC_STATE_MISSED, &q->state);
43 
44 	/* Make sure the below netif_xmit_frozen_or_stopped()
45 	 * checking happens after clearing STATE_MISSED.
46 	 */
47 	smp_mb__after_atomic();
48 
49 	/* Checking netif_xmit_frozen_or_stopped() again to
50 	 * make sure STATE_MISSED is set if the STATE_MISSED
51 	 * set by netif_tx_wake_queue()'s rescheduling of
52 	 * net_tx_action() is cleared by the above clear_bit().
53 	 */
54 	if (!netif_xmit_frozen_or_stopped(txq))
55 		set_bit(__QDISC_STATE_MISSED, &q->state);
56 	else
57 		set_bit(__QDISC_STATE_DRAINING, &q->state);
58 }
59 
60 /* Main transmission queue. */
61 
62 /* Modifications to data participating in scheduling must be protected with
63  * qdisc_lock(qdisc) spinlock.
64  *
65  * The idea is the following:
66  * - enqueue, dequeue are serialized via qdisc root lock
67  * - ingress filtering is also serialized via qdisc root lock
68  * - updates to tree and tree walking are only done under the rtnl mutex.
69  */
70 
71 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
72 
73 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
74 {
75 	const struct netdev_queue *txq = q->dev_queue;
76 	spinlock_t *lock = NULL;
77 	struct sk_buff *skb;
78 
79 	if (q->flags & TCQ_F_NOLOCK) {
80 		lock = qdisc_lock(q);
81 		spin_lock(lock);
82 	}
83 
84 	skb = skb_peek(&q->skb_bad_txq);
85 	if (skb) {
86 		/* check the reason of requeuing without tx lock first */
87 		txq = skb_get_tx_queue(txq->dev, skb);
88 		if (!netif_xmit_frozen_or_stopped(txq)) {
89 			skb = __skb_dequeue(&q->skb_bad_txq);
90 			if (qdisc_is_percpu_stats(q)) {
91 				qdisc_qstats_cpu_backlog_dec(q, skb);
92 				qdisc_qstats_cpu_qlen_dec(q);
93 			} else {
94 				qdisc_qstats_backlog_dec(q, skb);
95 				q->q.qlen--;
96 			}
97 		} else {
98 			skb = SKB_XOFF_MAGIC;
99 			qdisc_maybe_clear_missed(q, txq);
100 		}
101 	}
102 
103 	if (lock)
104 		spin_unlock(lock);
105 
106 	return skb;
107 }
108 
109 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
110 {
111 	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
112 
113 	if (unlikely(skb))
114 		skb = __skb_dequeue_bad_txq(q);
115 
116 	return skb;
117 }
118 
119 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
120 					     struct sk_buff *skb)
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 	__skb_queue_tail(&q->skb_bad_txq, skb);
130 
131 	if (qdisc_is_percpu_stats(q)) {
132 		qdisc_qstats_cpu_backlog_inc(q, skb);
133 		qdisc_qstats_cpu_qlen_inc(q);
134 	} else {
135 		qdisc_qstats_backlog_inc(q, skb);
136 		q->q.qlen++;
137 	}
138 
139 	if (lock)
140 		spin_unlock(lock);
141 }
142 
143 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
144 {
145 	spinlock_t *lock = NULL;
146 
147 	if (q->flags & TCQ_F_NOLOCK) {
148 		lock = qdisc_lock(q);
149 		spin_lock(lock);
150 	}
151 
152 	while (skb) {
153 		struct sk_buff *next = skb->next;
154 
155 		__skb_queue_tail(&q->gso_skb, skb);
156 
157 		/* it's still part of the queue */
158 		if (qdisc_is_percpu_stats(q)) {
159 			qdisc_qstats_cpu_requeues_inc(q);
160 			qdisc_qstats_cpu_backlog_inc(q, skb);
161 			qdisc_qstats_cpu_qlen_inc(q);
162 		} else {
163 			q->qstats.requeues++;
164 			qdisc_qstats_backlog_inc(q, skb);
165 			q->q.qlen++;
166 		}
167 
168 		skb = next;
169 	}
170 
171 	if (lock) {
172 		spin_unlock(lock);
173 		set_bit(__QDISC_STATE_MISSED, &q->state);
174 	} else {
175 		__netif_schedule(q);
176 	}
177 }
178 
179 static void try_bulk_dequeue_skb(struct Qdisc *q,
180 				 struct sk_buff *skb,
181 				 const struct netdev_queue *txq,
182 				 int *packets)
183 {
184 	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
185 
186 	while (bytelimit > 0) {
187 		struct sk_buff *nskb = q->dequeue(q);
188 
189 		if (!nskb)
190 			break;
191 
192 		bytelimit -= nskb->len; /* covers GSO len */
193 		skb->next = nskb;
194 		skb = nskb;
195 		(*packets)++; /* GSO counts as one pkt */
196 	}
197 	skb_mark_not_on_list(skb);
198 }
199 
200 /* This variant of try_bulk_dequeue_skb() makes sure
201  * all skbs in the chain are for the same txq
202  */
203 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
204 				      struct sk_buff *skb,
205 				      int *packets)
206 {
207 	int mapping = skb_get_queue_mapping(skb);
208 	struct sk_buff *nskb;
209 	int cnt = 0;
210 
211 	do {
212 		nskb = q->dequeue(q);
213 		if (!nskb)
214 			break;
215 		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
216 			qdisc_enqueue_skb_bad_txq(q, nskb);
217 			break;
218 		}
219 		skb->next = nskb;
220 		skb = nskb;
221 	} while (++cnt < 8);
222 	(*packets) += cnt;
223 	skb_mark_not_on_list(skb);
224 }
225 
226 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
227  * A requeued skb (via q->gso_skb) can also be a SKB list.
228  */
229 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
230 				   int *packets)
231 {
232 	const struct netdev_queue *txq = q->dev_queue;
233 	struct sk_buff *skb = NULL;
234 
235 	*packets = 1;
236 	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
237 		spinlock_t *lock = NULL;
238 
239 		if (q->flags & TCQ_F_NOLOCK) {
240 			lock = qdisc_lock(q);
241 			spin_lock(lock);
242 		}
243 
244 		skb = skb_peek(&q->gso_skb);
245 
246 		/* skb may be null if another cpu pulls gso_skb off in between
247 		 * empty check and lock.
248 		 */
249 		if (!skb) {
250 			if (lock)
251 				spin_unlock(lock);
252 			goto validate;
253 		}
254 
255 		/* skb in gso_skb were already validated */
256 		*validate = false;
257 		if (xfrm_offload(skb))
258 			*validate = true;
259 		/* check the reason of requeuing without tx lock first */
260 		txq = skb_get_tx_queue(txq->dev, skb);
261 		if (!netif_xmit_frozen_or_stopped(txq)) {
262 			skb = __skb_dequeue(&q->gso_skb);
263 			if (qdisc_is_percpu_stats(q)) {
264 				qdisc_qstats_cpu_backlog_dec(q, skb);
265 				qdisc_qstats_cpu_qlen_dec(q);
266 			} else {
267 				qdisc_qstats_backlog_dec(q, skb);
268 				q->q.qlen--;
269 			}
270 		} else {
271 			skb = NULL;
272 			qdisc_maybe_clear_missed(q, txq);
273 		}
274 		if (lock)
275 			spin_unlock(lock);
276 		goto trace;
277 	}
278 validate:
279 	*validate = true;
280 
281 	if ((q->flags & TCQ_F_ONETXQUEUE) &&
282 	    netif_xmit_frozen_or_stopped(txq)) {
283 		qdisc_maybe_clear_missed(q, txq);
284 		return skb;
285 	}
286 
287 	skb = qdisc_dequeue_skb_bad_txq(q);
288 	if (unlikely(skb)) {
289 		if (skb == SKB_XOFF_MAGIC)
290 			return NULL;
291 		goto bulk;
292 	}
293 	skb = q->dequeue(q);
294 	if (skb) {
295 bulk:
296 		if (qdisc_may_bulk(q))
297 			try_bulk_dequeue_skb(q, skb, txq, packets);
298 		else
299 			try_bulk_dequeue_skb_slow(q, skb, packets);
300 	}
301 trace:
302 	trace_qdisc_dequeue(q, txq, *packets, skb);
303 	return skb;
304 }
305 
306 /*
307  * Transmit possibly several skbs, and handle the return status as
308  * required. Owning qdisc running bit guarantees that only one CPU
309  * can execute this function.
310  *
311  * Returns to the caller:
312  *				false  - hardware queue frozen backoff
313  *				true   - feel free to send more pkts
314  */
315 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
316 		     struct net_device *dev, struct netdev_queue *txq,
317 		     spinlock_t *root_lock, bool validate)
318 {
319 	int ret = NETDEV_TX_BUSY;
320 	bool again = false;
321 
322 	/* And release qdisc */
323 	if (root_lock)
324 		spin_unlock(root_lock);
325 
326 	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
327 	if (validate)
328 		skb = validate_xmit_skb_list(skb, dev, &again);
329 
330 #ifdef CONFIG_XFRM_OFFLOAD
331 	if (unlikely(again)) {
332 		if (root_lock)
333 			spin_lock(root_lock);
334 
335 		dev_requeue_skb(skb, q);
336 		return false;
337 	}
338 #endif
339 
340 	if (likely(skb)) {
341 		HARD_TX_LOCK(dev, txq, smp_processor_id());
342 		if (!netif_xmit_frozen_or_stopped(txq))
343 			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
344 		else
345 			qdisc_maybe_clear_missed(q, txq);
346 
347 		HARD_TX_UNLOCK(dev, txq);
348 	} else {
349 		if (root_lock)
350 			spin_lock(root_lock);
351 		return true;
352 	}
353 
354 	if (root_lock)
355 		spin_lock(root_lock);
356 
357 	if (!dev_xmit_complete(ret)) {
358 		/* Driver returned NETDEV_TX_BUSY - requeue skb */
359 		if (unlikely(ret != NETDEV_TX_BUSY))
360 			net_warn_ratelimited("BUG %s code %d qlen %d\n",
361 					     dev->name, ret, q->q.qlen);
362 
363 		dev_requeue_skb(skb, q);
364 		return false;
365 	}
366 
367 	return true;
368 }
369 
370 /*
371  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
372  *
373  * running seqcount guarantees only one CPU can process
374  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
375  * this queue.
376  *
377  *  netif_tx_lock serializes accesses to device driver.
378  *
379  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
380  *  if one is grabbed, another must be free.
381  *
382  * Note, that this procedure can be called by a watchdog timer
383  *
384  * Returns to the caller:
385  *				0  - queue is empty or throttled.
386  *				>0 - queue is not empty.
387  *
388  */
389 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
390 {
391 	spinlock_t *root_lock = NULL;
392 	struct netdev_queue *txq;
393 	struct net_device *dev;
394 	struct sk_buff *skb;
395 	bool validate;
396 
397 	/* Dequeue packet */
398 	skb = dequeue_skb(q, &validate, packets);
399 	if (unlikely(!skb))
400 		return false;
401 
402 	if (!(q->flags & TCQ_F_NOLOCK))
403 		root_lock = qdisc_lock(q);
404 
405 	dev = qdisc_dev(q);
406 	txq = skb_get_tx_queue(dev, skb);
407 
408 	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
409 }
410 
411 void __qdisc_run(struct Qdisc *q)
412 {
413 	int quota = READ_ONCE(net_hotdata.dev_tx_weight);
414 	int packets;
415 
416 	while (qdisc_restart(q, &packets)) {
417 		quota -= packets;
418 		if (quota <= 0) {
419 			if (q->flags & TCQ_F_NOLOCK)
420 				set_bit(__QDISC_STATE_MISSED, &q->state);
421 			else
422 				__netif_schedule(q);
423 
424 			break;
425 		}
426 	}
427 }
428 
429 unsigned long dev_trans_start(struct net_device *dev)
430 {
431 	unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
432 	unsigned long val;
433 	unsigned int i;
434 
435 	for (i = 1; i < dev->num_tx_queues; i++) {
436 		val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
437 		if (val && time_after(val, res))
438 			res = val;
439 	}
440 
441 	return res;
442 }
443 EXPORT_SYMBOL(dev_trans_start);
444 
445 static void netif_freeze_queues(struct net_device *dev)
446 {
447 	unsigned int i;
448 	int cpu;
449 
450 	cpu = smp_processor_id();
451 	for (i = 0; i < dev->num_tx_queues; i++) {
452 		struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
453 
454 		/* We are the only thread of execution doing a
455 		 * freeze, but we have to grab the _xmit_lock in
456 		 * order to synchronize with threads which are in
457 		 * the ->hard_start_xmit() handler and already
458 		 * checked the frozen bit.
459 		 */
460 		__netif_tx_lock(txq, cpu);
461 		set_bit(__QUEUE_STATE_FROZEN, &txq->state);
462 		__netif_tx_unlock(txq);
463 	}
464 }
465 
466 void netif_tx_lock(struct net_device *dev)
467 {
468 	spin_lock(&dev->tx_global_lock);
469 	netif_freeze_queues(dev);
470 }
471 EXPORT_SYMBOL(netif_tx_lock);
472 
473 static void netif_unfreeze_queues(struct net_device *dev)
474 {
475 	unsigned int i;
476 
477 	for (i = 0; i < dev->num_tx_queues; i++) {
478 		struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
479 
480 		/* No need to grab the _xmit_lock here.  If the
481 		 * queue is not stopped for another reason, we
482 		 * force a schedule.
483 		 */
484 		clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
485 		netif_schedule_queue(txq);
486 	}
487 }
488 
489 void netif_tx_unlock(struct net_device *dev)
490 {
491 	netif_unfreeze_queues(dev);
492 	spin_unlock(&dev->tx_global_lock);
493 }
494 EXPORT_SYMBOL(netif_tx_unlock);
495 
496 static void dev_watchdog(struct timer_list *t)
497 {
498 	struct net_device *dev = from_timer(dev, t, watchdog_timer);
499 	bool release = true;
500 
501 	spin_lock(&dev->tx_global_lock);
502 	if (!qdisc_tx_is_noop(dev)) {
503 		if (netif_device_present(dev) &&
504 		    netif_running(dev) &&
505 		    netif_carrier_ok(dev)) {
506 			unsigned int timedout_ms = 0;
507 			unsigned int i;
508 			unsigned long trans_start;
509 			unsigned long oldest_start = jiffies;
510 
511 			for (i = 0; i < dev->num_tx_queues; i++) {
512 				struct netdev_queue *txq;
513 
514 				txq = netdev_get_tx_queue(dev, i);
515 				trans_start = READ_ONCE(txq->trans_start);
516 				if (!netif_xmit_stopped(txq))
517 					continue;
518 				if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
519 					timedout_ms = jiffies_to_msecs(jiffies - trans_start);
520 					atomic_long_inc(&txq->trans_timeout);
521 					break;
522 				}
523 				if (time_after(oldest_start, trans_start))
524 					oldest_start = trans_start;
525 			}
526 
527 			if (unlikely(timedout_ms)) {
528 				trace_net_dev_xmit_timeout(dev, i);
529 				netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
530 					    raw_smp_processor_id(),
531 					    i, timedout_ms);
532 				netif_freeze_queues(dev);
533 				dev->netdev_ops->ndo_tx_timeout(dev, i);
534 				netif_unfreeze_queues(dev);
535 			}
536 			if (!mod_timer(&dev->watchdog_timer,
537 				       round_jiffies(oldest_start +
538 						     dev->watchdog_timeo)))
539 				release = false;
540 		}
541 	}
542 	spin_unlock(&dev->tx_global_lock);
543 
544 	if (release)
545 		netdev_put(dev, &dev->watchdog_dev_tracker);
546 }
547 
548 void __netdev_watchdog_up(struct net_device *dev)
549 {
550 	if (dev->netdev_ops->ndo_tx_timeout) {
551 		if (dev->watchdog_timeo <= 0)
552 			dev->watchdog_timeo = 5*HZ;
553 		if (!mod_timer(&dev->watchdog_timer,
554 			       round_jiffies(jiffies + dev->watchdog_timeo)))
555 			netdev_hold(dev, &dev->watchdog_dev_tracker,
556 				    GFP_ATOMIC);
557 	}
558 }
559 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
560 
561 static void dev_watchdog_up(struct net_device *dev)
562 {
563 	__netdev_watchdog_up(dev);
564 }
565 
566 static void dev_watchdog_down(struct net_device *dev)
567 {
568 	netif_tx_lock_bh(dev);
569 	if (del_timer(&dev->watchdog_timer))
570 		netdev_put(dev, &dev->watchdog_dev_tracker);
571 	netif_tx_unlock_bh(dev);
572 }
573 
574 /**
575  *	netif_carrier_on - set carrier
576  *	@dev: network device
577  *
578  * Device has detected acquisition of carrier.
579  */
580 void netif_carrier_on(struct net_device *dev)
581 {
582 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
583 		if (dev->reg_state == NETREG_UNINITIALIZED)
584 			return;
585 		atomic_inc(&dev->carrier_up_count);
586 		linkwatch_fire_event(dev);
587 		if (netif_running(dev))
588 			__netdev_watchdog_up(dev);
589 	}
590 }
591 EXPORT_SYMBOL(netif_carrier_on);
592 
593 /**
594  *	netif_carrier_off - clear carrier
595  *	@dev: network device
596  *
597  * Device has detected loss of carrier.
598  */
599 void netif_carrier_off(struct net_device *dev)
600 {
601 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
602 		if (dev->reg_state == NETREG_UNINITIALIZED)
603 			return;
604 		atomic_inc(&dev->carrier_down_count);
605 		linkwatch_fire_event(dev);
606 	}
607 }
608 EXPORT_SYMBOL(netif_carrier_off);
609 
610 /**
611  *	netif_carrier_event - report carrier state event
612  *	@dev: network device
613  *
614  * Device has detected a carrier event but the carrier state wasn't changed.
615  * Use in drivers when querying carrier state asynchronously, to avoid missing
616  * events (link flaps) if link recovers before it's queried.
617  */
618 void netif_carrier_event(struct net_device *dev)
619 {
620 	if (dev->reg_state == NETREG_UNINITIALIZED)
621 		return;
622 	atomic_inc(&dev->carrier_up_count);
623 	atomic_inc(&dev->carrier_down_count);
624 	linkwatch_fire_event(dev);
625 }
626 EXPORT_SYMBOL_GPL(netif_carrier_event);
627 
628 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
629    under all circumstances. It is difficult to invent anything faster or
630    cheaper.
631  */
632 
633 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
634 			struct sk_buff **to_free)
635 {
636 	__qdisc_drop(skb, to_free);
637 	return NET_XMIT_CN;
638 }
639 
640 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
641 {
642 	return NULL;
643 }
644 
645 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
646 	.id		=	"noop",
647 	.priv_size	=	0,
648 	.enqueue	=	noop_enqueue,
649 	.dequeue	=	noop_dequeue,
650 	.peek		=	noop_dequeue,
651 	.owner		=	THIS_MODULE,
652 };
653 
654 static struct netdev_queue noop_netdev_queue = {
655 	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
656 	RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
657 };
658 
659 struct Qdisc noop_qdisc = {
660 	.enqueue	=	noop_enqueue,
661 	.dequeue	=	noop_dequeue,
662 	.flags		=	TCQ_F_BUILTIN,
663 	.ops		=	&noop_qdisc_ops,
664 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
665 	.dev_queue	=	&noop_netdev_queue,
666 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
667 	.gso_skb = {
668 		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
669 		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
670 		.qlen = 0,
671 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
672 	},
673 	.skb_bad_txq = {
674 		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
675 		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
676 		.qlen = 0,
677 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
678 	},
679 };
680 EXPORT_SYMBOL(noop_qdisc);
681 
682 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
683 			struct netlink_ext_ack *extack)
684 {
685 	/* register_qdisc() assigns a default of noop_enqueue if unset,
686 	 * but __dev_queue_xmit() treats noqueue only as such
687 	 * if this is NULL - so clear it here. */
688 	qdisc->enqueue = NULL;
689 	return 0;
690 }
691 
692 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
693 	.id		=	"noqueue",
694 	.priv_size	=	0,
695 	.init		=	noqueue_init,
696 	.enqueue	=	noop_enqueue,
697 	.dequeue	=	noop_dequeue,
698 	.peek		=	noop_dequeue,
699 	.owner		=	THIS_MODULE,
700 };
701 
702 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
703 	1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
704 };
705 EXPORT_SYMBOL(sch_default_prio2band);
706 
707 /* 3-band FIFO queue: old style, but should be a bit faster than
708    generic prio+fifo combination.
709  */
710 
711 #define PFIFO_FAST_BANDS 3
712 
713 /*
714  * Private data for a pfifo_fast scheduler containing:
715  *	- rings for priority bands
716  */
717 struct pfifo_fast_priv {
718 	struct skb_array q[PFIFO_FAST_BANDS];
719 };
720 
721 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
722 					  int band)
723 {
724 	return &priv->q[band];
725 }
726 
727 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
728 			      struct sk_buff **to_free)
729 {
730 	int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
731 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
732 	struct skb_array *q = band2list(priv, band);
733 	unsigned int pkt_len = qdisc_pkt_len(skb);
734 	int err;
735 
736 	err = skb_array_produce(q, skb);
737 
738 	if (unlikely(err)) {
739 		if (qdisc_is_percpu_stats(qdisc))
740 			return qdisc_drop_cpu(skb, qdisc, to_free);
741 		else
742 			return qdisc_drop(skb, qdisc, to_free);
743 	}
744 
745 	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
746 	return NET_XMIT_SUCCESS;
747 }
748 
749 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
750 {
751 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
752 	struct sk_buff *skb = NULL;
753 	bool need_retry = true;
754 	int band;
755 
756 retry:
757 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
758 		struct skb_array *q = band2list(priv, band);
759 
760 		if (__skb_array_empty(q))
761 			continue;
762 
763 		skb = __skb_array_consume(q);
764 	}
765 	if (likely(skb)) {
766 		qdisc_update_stats_at_dequeue(qdisc, skb);
767 	} else if (need_retry &&
768 		   READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
769 		/* Delay clearing the STATE_MISSED here to reduce
770 		 * the overhead of the second spin_trylock() in
771 		 * qdisc_run_begin() and __netif_schedule() calling
772 		 * in qdisc_run_end().
773 		 */
774 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
775 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
776 
777 		/* Make sure dequeuing happens after clearing
778 		 * STATE_MISSED.
779 		 */
780 		smp_mb__after_atomic();
781 
782 		need_retry = false;
783 
784 		goto retry;
785 	}
786 
787 	return skb;
788 }
789 
790 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
791 {
792 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
793 	struct sk_buff *skb = NULL;
794 	int band;
795 
796 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
797 		struct skb_array *q = band2list(priv, band);
798 
799 		skb = __skb_array_peek(q);
800 	}
801 
802 	return skb;
803 }
804 
805 static void pfifo_fast_reset(struct Qdisc *qdisc)
806 {
807 	int i, band;
808 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
809 
810 	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
811 		struct skb_array *q = band2list(priv, band);
812 		struct sk_buff *skb;
813 
814 		/* NULL ring is possible if destroy path is due to a failed
815 		 * skb_array_init() in pfifo_fast_init() case.
816 		 */
817 		if (!q->ring.queue)
818 			continue;
819 
820 		while ((skb = __skb_array_consume(q)) != NULL)
821 			kfree_skb(skb);
822 	}
823 
824 	if (qdisc_is_percpu_stats(qdisc)) {
825 		for_each_possible_cpu(i) {
826 			struct gnet_stats_queue *q;
827 
828 			q = per_cpu_ptr(qdisc->cpu_qstats, i);
829 			q->backlog = 0;
830 			q->qlen = 0;
831 		}
832 	}
833 }
834 
835 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
836 {
837 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
838 
839 	memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
840 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
841 		goto nla_put_failure;
842 	return skb->len;
843 
844 nla_put_failure:
845 	return -1;
846 }
847 
848 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
849 			   struct netlink_ext_ack *extack)
850 {
851 	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
852 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
853 	int prio;
854 
855 	/* guard against zero length rings */
856 	if (!qlen)
857 		return -EINVAL;
858 
859 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
860 		struct skb_array *q = band2list(priv, prio);
861 		int err;
862 
863 		err = skb_array_init(q, qlen, GFP_KERNEL);
864 		if (err)
865 			return -ENOMEM;
866 	}
867 
868 	/* Can by-pass the queue discipline */
869 	qdisc->flags |= TCQ_F_CAN_BYPASS;
870 	return 0;
871 }
872 
873 static void pfifo_fast_destroy(struct Qdisc *sch)
874 {
875 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
876 	int prio;
877 
878 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
879 		struct skb_array *q = band2list(priv, prio);
880 
881 		/* NULL ring is possible if destroy path is due to a failed
882 		 * skb_array_init() in pfifo_fast_init() case.
883 		 */
884 		if (!q->ring.queue)
885 			continue;
886 		/* Destroy ring but no need to kfree_skb because a call to
887 		 * pfifo_fast_reset() has already done that work.
888 		 */
889 		ptr_ring_cleanup(&q->ring, NULL);
890 	}
891 }
892 
893 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
894 					  unsigned int new_len)
895 {
896 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
897 	struct skb_array *bands[PFIFO_FAST_BANDS];
898 	int prio;
899 
900 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
901 		struct skb_array *q = band2list(priv, prio);
902 
903 		bands[prio] = q;
904 	}
905 
906 	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
907 					 GFP_KERNEL);
908 }
909 
910 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
911 	.id		=	"pfifo_fast",
912 	.priv_size	=	sizeof(struct pfifo_fast_priv),
913 	.enqueue	=	pfifo_fast_enqueue,
914 	.dequeue	=	pfifo_fast_dequeue,
915 	.peek		=	pfifo_fast_peek,
916 	.init		=	pfifo_fast_init,
917 	.destroy	=	pfifo_fast_destroy,
918 	.reset		=	pfifo_fast_reset,
919 	.dump		=	pfifo_fast_dump,
920 	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
921 	.owner		=	THIS_MODULE,
922 	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
923 };
924 EXPORT_SYMBOL(pfifo_fast_ops);
925 
926 static struct lock_class_key qdisc_tx_busylock;
927 
928 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
929 			  const struct Qdisc_ops *ops,
930 			  struct netlink_ext_ack *extack)
931 {
932 	struct Qdisc *sch;
933 	unsigned int size = sizeof(*sch) + ops->priv_size;
934 	int err = -ENOBUFS;
935 	struct net_device *dev;
936 
937 	if (!dev_queue) {
938 		NL_SET_ERR_MSG(extack, "No device queue given");
939 		err = -EINVAL;
940 		goto errout;
941 	}
942 
943 	dev = dev_queue->dev;
944 	sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
945 
946 	if (!sch)
947 		goto errout;
948 	__skb_queue_head_init(&sch->gso_skb);
949 	__skb_queue_head_init(&sch->skb_bad_txq);
950 	gnet_stats_basic_sync_init(&sch->bstats);
951 	lockdep_register_key(&sch->root_lock_key);
952 	spin_lock_init(&sch->q.lock);
953 	lockdep_set_class(&sch->q.lock, &sch->root_lock_key);
954 
955 	if (ops->static_flags & TCQ_F_CPUSTATS) {
956 		sch->cpu_bstats =
957 			netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
958 		if (!sch->cpu_bstats)
959 			goto errout1;
960 
961 		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
962 		if (!sch->cpu_qstats) {
963 			free_percpu(sch->cpu_bstats);
964 			goto errout1;
965 		}
966 	}
967 
968 	spin_lock_init(&sch->busylock);
969 	lockdep_set_class(&sch->busylock,
970 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
971 
972 	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
973 	spin_lock_init(&sch->seqlock);
974 	lockdep_set_class(&sch->seqlock,
975 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
976 
977 	sch->ops = ops;
978 	sch->flags = ops->static_flags;
979 	sch->enqueue = ops->enqueue;
980 	sch->dequeue = ops->dequeue;
981 	sch->dev_queue = dev_queue;
982 	sch->owner = -1;
983 	netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
984 	refcount_set(&sch->refcnt, 1);
985 
986 	return sch;
987 errout1:
988 	lockdep_unregister_key(&sch->root_lock_key);
989 	kfree(sch);
990 errout:
991 	return ERR_PTR(err);
992 }
993 
994 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
995 				const struct Qdisc_ops *ops,
996 				unsigned int parentid,
997 				struct netlink_ext_ack *extack)
998 {
999 	struct Qdisc *sch;
1000 
1001 	if (!try_module_get(ops->owner)) {
1002 		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1003 		return NULL;
1004 	}
1005 
1006 	sch = qdisc_alloc(dev_queue, ops, extack);
1007 	if (IS_ERR(sch)) {
1008 		module_put(ops->owner);
1009 		return NULL;
1010 	}
1011 	sch->parent = parentid;
1012 
1013 	if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1014 		trace_qdisc_create(ops, dev_queue->dev, parentid);
1015 		return sch;
1016 	}
1017 
1018 	qdisc_put(sch);
1019 	return NULL;
1020 }
1021 EXPORT_SYMBOL(qdisc_create_dflt);
1022 
1023 /* Under qdisc_lock(qdisc) and BH! */
1024 
1025 void qdisc_reset(struct Qdisc *qdisc)
1026 {
1027 	const struct Qdisc_ops *ops = qdisc->ops;
1028 
1029 	trace_qdisc_reset(qdisc);
1030 
1031 	if (ops->reset)
1032 		ops->reset(qdisc);
1033 
1034 	__skb_queue_purge(&qdisc->gso_skb);
1035 	__skb_queue_purge(&qdisc->skb_bad_txq);
1036 
1037 	qdisc->q.qlen = 0;
1038 	qdisc->qstats.backlog = 0;
1039 }
1040 EXPORT_SYMBOL(qdisc_reset);
1041 
1042 void qdisc_free(struct Qdisc *qdisc)
1043 {
1044 	if (qdisc_is_percpu_stats(qdisc)) {
1045 		free_percpu(qdisc->cpu_bstats);
1046 		free_percpu(qdisc->cpu_qstats);
1047 	}
1048 
1049 	kfree(qdisc);
1050 }
1051 
1052 static void qdisc_free_cb(struct rcu_head *head)
1053 {
1054 	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1055 
1056 	qdisc_free(q);
1057 }
1058 
1059 static void __qdisc_destroy(struct Qdisc *qdisc)
1060 {
1061 	const struct Qdisc_ops  *ops = qdisc->ops;
1062 	struct net_device *dev = qdisc_dev(qdisc);
1063 
1064 #ifdef CONFIG_NET_SCHED
1065 	qdisc_hash_del(qdisc);
1066 
1067 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
1068 #endif
1069 	gen_kill_estimator(&qdisc->rate_est);
1070 
1071 	qdisc_reset(qdisc);
1072 
1073 
1074 	if (ops->destroy)
1075 		ops->destroy(qdisc);
1076 
1077 	lockdep_unregister_key(&qdisc->root_lock_key);
1078 	module_put(ops->owner);
1079 	netdev_put(dev, &qdisc->dev_tracker);
1080 
1081 	trace_qdisc_destroy(qdisc);
1082 
1083 	call_rcu(&qdisc->rcu, qdisc_free_cb);
1084 }
1085 
1086 void qdisc_destroy(struct Qdisc *qdisc)
1087 {
1088 	if (qdisc->flags & TCQ_F_BUILTIN)
1089 		return;
1090 
1091 	__qdisc_destroy(qdisc);
1092 }
1093 
1094 void qdisc_put(struct Qdisc *qdisc)
1095 {
1096 	if (!qdisc)
1097 		return;
1098 
1099 	if (qdisc->flags & TCQ_F_BUILTIN ||
1100 	    !refcount_dec_and_test(&qdisc->refcnt))
1101 		return;
1102 
1103 	__qdisc_destroy(qdisc);
1104 }
1105 EXPORT_SYMBOL(qdisc_put);
1106 
1107 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1108  * Intended to be used as optimization, this function only takes rtnl lock if
1109  * qdisc reference counter reached zero.
1110  */
1111 
1112 void qdisc_put_unlocked(struct Qdisc *qdisc)
1113 {
1114 	if (qdisc->flags & TCQ_F_BUILTIN ||
1115 	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1116 		return;
1117 
1118 	__qdisc_destroy(qdisc);
1119 	rtnl_unlock();
1120 }
1121 EXPORT_SYMBOL(qdisc_put_unlocked);
1122 
1123 /* Attach toplevel qdisc to device queue. */
1124 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1125 			      struct Qdisc *qdisc)
1126 {
1127 	struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1128 	spinlock_t *root_lock;
1129 
1130 	root_lock = qdisc_lock(oqdisc);
1131 	spin_lock_bh(root_lock);
1132 
1133 	/* ... and graft new one */
1134 	if (qdisc == NULL)
1135 		qdisc = &noop_qdisc;
1136 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1137 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1138 
1139 	spin_unlock_bh(root_lock);
1140 
1141 	return oqdisc;
1142 }
1143 EXPORT_SYMBOL(dev_graft_qdisc);
1144 
1145 static void shutdown_scheduler_queue(struct net_device *dev,
1146 				     struct netdev_queue *dev_queue,
1147 				     void *_qdisc_default)
1148 {
1149 	struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1150 	struct Qdisc *qdisc_default = _qdisc_default;
1151 
1152 	if (qdisc) {
1153 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1154 		rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1155 
1156 		qdisc_put(qdisc);
1157 	}
1158 }
1159 
1160 static void attach_one_default_qdisc(struct net_device *dev,
1161 				     struct netdev_queue *dev_queue,
1162 				     void *_unused)
1163 {
1164 	struct Qdisc *qdisc;
1165 	const struct Qdisc_ops *ops = default_qdisc_ops;
1166 
1167 	if (dev->priv_flags & IFF_NO_QUEUE)
1168 		ops = &noqueue_qdisc_ops;
1169 	else if(dev->type == ARPHRD_CAN)
1170 		ops = &pfifo_fast_ops;
1171 
1172 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1173 	if (!qdisc)
1174 		return;
1175 
1176 	if (!netif_is_multiqueue(dev))
1177 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1178 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1179 }
1180 
1181 static void attach_default_qdiscs(struct net_device *dev)
1182 {
1183 	struct netdev_queue *txq;
1184 	struct Qdisc *qdisc;
1185 
1186 	txq = netdev_get_tx_queue(dev, 0);
1187 
1188 	if (!netif_is_multiqueue(dev) ||
1189 	    dev->priv_flags & IFF_NO_QUEUE) {
1190 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1191 		qdisc = rtnl_dereference(txq->qdisc_sleeping);
1192 		rcu_assign_pointer(dev->qdisc, qdisc);
1193 		qdisc_refcount_inc(qdisc);
1194 	} else {
1195 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1196 		if (qdisc) {
1197 			rcu_assign_pointer(dev->qdisc, qdisc);
1198 			qdisc->ops->attach(qdisc);
1199 		}
1200 	}
1201 	qdisc = rtnl_dereference(dev->qdisc);
1202 
1203 	/* Detect default qdisc setup/init failed and fallback to "noqueue" */
1204 	if (qdisc == &noop_qdisc) {
1205 		netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1206 			    default_qdisc_ops->id, noqueue_qdisc_ops.id);
1207 		netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1208 		dev->priv_flags |= IFF_NO_QUEUE;
1209 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1210 		qdisc = rtnl_dereference(txq->qdisc_sleeping);
1211 		rcu_assign_pointer(dev->qdisc, qdisc);
1212 		qdisc_refcount_inc(qdisc);
1213 		dev->priv_flags ^= IFF_NO_QUEUE;
1214 	}
1215 
1216 #ifdef CONFIG_NET_SCHED
1217 	if (qdisc != &noop_qdisc)
1218 		qdisc_hash_add(qdisc, false);
1219 #endif
1220 }
1221 
1222 static void transition_one_qdisc(struct net_device *dev,
1223 				 struct netdev_queue *dev_queue,
1224 				 void *_need_watchdog)
1225 {
1226 	struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1227 	int *need_watchdog_p = _need_watchdog;
1228 
1229 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1230 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1231 
1232 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1233 	if (need_watchdog_p) {
1234 		WRITE_ONCE(dev_queue->trans_start, 0);
1235 		*need_watchdog_p = 1;
1236 	}
1237 }
1238 
1239 void dev_activate(struct net_device *dev)
1240 {
1241 	int need_watchdog;
1242 
1243 	/* No queueing discipline is attached to device;
1244 	 * create default one for devices, which need queueing
1245 	 * and noqueue_qdisc for virtual interfaces
1246 	 */
1247 
1248 	if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1249 		attach_default_qdiscs(dev);
1250 
1251 	if (!netif_carrier_ok(dev))
1252 		/* Delay activation until next carrier-on event */
1253 		return;
1254 
1255 	need_watchdog = 0;
1256 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1257 	if (dev_ingress_queue(dev))
1258 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1259 
1260 	if (need_watchdog) {
1261 		netif_trans_update(dev);
1262 		dev_watchdog_up(dev);
1263 	}
1264 }
1265 EXPORT_SYMBOL(dev_activate);
1266 
1267 static void qdisc_deactivate(struct Qdisc *qdisc)
1268 {
1269 	if (qdisc->flags & TCQ_F_BUILTIN)
1270 		return;
1271 
1272 	set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1273 }
1274 
1275 static void dev_deactivate_queue(struct net_device *dev,
1276 				 struct netdev_queue *dev_queue,
1277 				 void *_qdisc_default)
1278 {
1279 	struct Qdisc *qdisc_default = _qdisc_default;
1280 	struct Qdisc *qdisc;
1281 
1282 	qdisc = rtnl_dereference(dev_queue->qdisc);
1283 	if (qdisc) {
1284 		qdisc_deactivate(qdisc);
1285 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1286 	}
1287 }
1288 
1289 static void dev_reset_queue(struct net_device *dev,
1290 			    struct netdev_queue *dev_queue,
1291 			    void *_unused)
1292 {
1293 	struct Qdisc *qdisc;
1294 	bool nolock;
1295 
1296 	qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1297 	if (!qdisc)
1298 		return;
1299 
1300 	nolock = qdisc->flags & TCQ_F_NOLOCK;
1301 
1302 	if (nolock)
1303 		spin_lock_bh(&qdisc->seqlock);
1304 	spin_lock_bh(qdisc_lock(qdisc));
1305 
1306 	qdisc_reset(qdisc);
1307 
1308 	spin_unlock_bh(qdisc_lock(qdisc));
1309 	if (nolock) {
1310 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1311 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1312 		spin_unlock_bh(&qdisc->seqlock);
1313 	}
1314 }
1315 
1316 static bool some_qdisc_is_busy(struct net_device *dev)
1317 {
1318 	unsigned int i;
1319 
1320 	for (i = 0; i < dev->num_tx_queues; i++) {
1321 		struct netdev_queue *dev_queue;
1322 		spinlock_t *root_lock;
1323 		struct Qdisc *q;
1324 		int val;
1325 
1326 		dev_queue = netdev_get_tx_queue(dev, i);
1327 		q = rtnl_dereference(dev_queue->qdisc_sleeping);
1328 
1329 		root_lock = qdisc_lock(q);
1330 		spin_lock_bh(root_lock);
1331 
1332 		val = (qdisc_is_running(q) ||
1333 		       test_bit(__QDISC_STATE_SCHED, &q->state));
1334 
1335 		spin_unlock_bh(root_lock);
1336 
1337 		if (val)
1338 			return true;
1339 	}
1340 	return false;
1341 }
1342 
1343 /**
1344  * 	dev_deactivate_many - deactivate transmissions on several devices
1345  * 	@head: list of devices to deactivate
1346  *
1347  *	This function returns only when all outstanding transmissions
1348  *	have completed, unless all devices are in dismantle phase.
1349  */
1350 void dev_deactivate_many(struct list_head *head)
1351 {
1352 	struct net_device *dev;
1353 
1354 	list_for_each_entry(dev, head, close_list) {
1355 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1356 					 &noop_qdisc);
1357 		if (dev_ingress_queue(dev))
1358 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1359 					     &noop_qdisc);
1360 
1361 		dev_watchdog_down(dev);
1362 	}
1363 
1364 	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
1365 	 * outstanding qdisc enqueuing calls.
1366 	 * This is avoided if all devices are in dismantle phase :
1367 	 * Caller will call synchronize_net() for us
1368 	 */
1369 	synchronize_net();
1370 
1371 	list_for_each_entry(dev, head, close_list) {
1372 		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1373 
1374 		if (dev_ingress_queue(dev))
1375 			dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1376 	}
1377 
1378 	/* Wait for outstanding qdisc_run calls. */
1379 	list_for_each_entry(dev, head, close_list) {
1380 		while (some_qdisc_is_busy(dev)) {
1381 			/* wait_event() would avoid this sleep-loop but would
1382 			 * require expensive checks in the fast paths of packet
1383 			 * processing which isn't worth it.
1384 			 */
1385 			schedule_timeout_uninterruptible(1);
1386 		}
1387 	}
1388 }
1389 
1390 void dev_deactivate(struct net_device *dev)
1391 {
1392 	LIST_HEAD(single);
1393 
1394 	list_add(&dev->close_list, &single);
1395 	dev_deactivate_many(&single);
1396 	list_del(&single);
1397 }
1398 EXPORT_SYMBOL(dev_deactivate);
1399 
1400 static int qdisc_change_tx_queue_len(struct net_device *dev,
1401 				     struct netdev_queue *dev_queue)
1402 {
1403 	struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1404 	const struct Qdisc_ops *ops = qdisc->ops;
1405 
1406 	if (ops->change_tx_queue_len)
1407 		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1408 	return 0;
1409 }
1410 
1411 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1412 				  unsigned int new_real_tx)
1413 {
1414 	struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1415 
1416 	if (qdisc->ops->change_real_num_tx)
1417 		qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1418 }
1419 
1420 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1421 {
1422 #ifdef CONFIG_NET_SCHED
1423 	struct net_device *dev = qdisc_dev(sch);
1424 	struct Qdisc *qdisc;
1425 	unsigned int i;
1426 
1427 	for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1428 		qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1429 		/* Only update the default qdiscs we created,
1430 		 * qdiscs with handles are always hashed.
1431 		 */
1432 		if (qdisc != &noop_qdisc && !qdisc->handle)
1433 			qdisc_hash_del(qdisc);
1434 	}
1435 	for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1436 		qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1437 		if (qdisc != &noop_qdisc && !qdisc->handle)
1438 			qdisc_hash_add(qdisc, false);
1439 	}
1440 #endif
1441 }
1442 EXPORT_SYMBOL(mq_change_real_num_tx);
1443 
1444 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1445 {
1446 	bool up = dev->flags & IFF_UP;
1447 	unsigned int i;
1448 	int ret = 0;
1449 
1450 	if (up)
1451 		dev_deactivate(dev);
1452 
1453 	for (i = 0; i < dev->num_tx_queues; i++) {
1454 		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1455 
1456 		/* TODO: revert changes on a partial failure */
1457 		if (ret)
1458 			break;
1459 	}
1460 
1461 	if (up)
1462 		dev_activate(dev);
1463 	return ret;
1464 }
1465 
1466 static void dev_init_scheduler_queue(struct net_device *dev,
1467 				     struct netdev_queue *dev_queue,
1468 				     void *_qdisc)
1469 {
1470 	struct Qdisc *qdisc = _qdisc;
1471 
1472 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1473 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1474 }
1475 
1476 void dev_init_scheduler(struct net_device *dev)
1477 {
1478 	rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1479 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1480 	if (dev_ingress_queue(dev))
1481 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1482 
1483 	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1484 }
1485 
1486 void dev_shutdown(struct net_device *dev)
1487 {
1488 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1489 	if (dev_ingress_queue(dev))
1490 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1491 	qdisc_put(rtnl_dereference(dev->qdisc));
1492 	rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1493 
1494 	WARN_ON(timer_pending(&dev->watchdog_timer));
1495 }
1496 
1497 /**
1498  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1499  * @rate:   Rate to compute reciprocal division values of
1500  * @mult:   Multiplier for reciprocal division
1501  * @shift:  Shift for reciprocal division
1502  *
1503  * The multiplier and shift for reciprocal division by rate are stored
1504  * in mult and shift.
1505  *
1506  * The deal here is to replace a divide by a reciprocal one
1507  * in fast path (a reciprocal divide is a multiply and a shift)
1508  *
1509  * Normal formula would be :
1510  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1511  *
1512  * We compute mult/shift to use instead :
1513  *  time_in_ns = (len * mult) >> shift;
1514  *
1515  * We try to get the highest possible mult value for accuracy,
1516  * but have to make sure no overflows will ever happen.
1517  *
1518  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1519  */
1520 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1521 {
1522 	u64 factor = NSEC_PER_SEC;
1523 
1524 	*mult = 1;
1525 	*shift = 0;
1526 
1527 	if (rate <= 0)
1528 		return;
1529 
1530 	for (;;) {
1531 		*mult = div64_u64(factor, rate);
1532 		if (*mult & (1U << 31) || factor & (1ULL << 63))
1533 			break;
1534 		factor <<= 1;
1535 		(*shift)++;
1536 	}
1537 }
1538 
1539 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1540 			       const struct tc_ratespec *conf,
1541 			       u64 rate64)
1542 {
1543 	memset(r, 0, sizeof(*r));
1544 	r->overhead = conf->overhead;
1545 	r->mpu = conf->mpu;
1546 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1547 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1548 	psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1549 }
1550 EXPORT_SYMBOL(psched_ratecfg_precompute);
1551 
1552 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1553 {
1554 	r->rate_pkts_ps = pktrate64;
1555 	psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1556 }
1557 EXPORT_SYMBOL(psched_ppscfg_precompute);
1558 
1559 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1560 			  struct tcf_proto *tp_head)
1561 {
1562 	/* Protected with chain0->filter_chain_lock.
1563 	 * Can't access chain directly because tp_head can be NULL.
1564 	 */
1565 	struct mini_Qdisc *miniq_old =
1566 		rcu_dereference_protected(*miniqp->p_miniq, 1);
1567 	struct mini_Qdisc *miniq;
1568 
1569 	if (!tp_head) {
1570 		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1571 	} else {
1572 		miniq = miniq_old != &miniqp->miniq1 ?
1573 			&miniqp->miniq1 : &miniqp->miniq2;
1574 
1575 		/* We need to make sure that readers won't see the miniq
1576 		 * we are about to modify. So ensure that at least one RCU
1577 		 * grace period has elapsed since the miniq was made
1578 		 * inactive.
1579 		 */
1580 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
1581 			cond_synchronize_rcu(miniq->rcu_state);
1582 		else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1583 			synchronize_rcu_expedited();
1584 
1585 		miniq->filter_list = tp_head;
1586 		rcu_assign_pointer(*miniqp->p_miniq, miniq);
1587 	}
1588 
1589 	if (miniq_old)
1590 		/* This is counterpart of the rcu sync above. We need to
1591 		 * block potential new user of miniq_old until all readers
1592 		 * are not seeing it.
1593 		 */
1594 		miniq_old->rcu_state = start_poll_synchronize_rcu();
1595 }
1596 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1597 
1598 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1599 				struct tcf_block *block)
1600 {
1601 	miniqp->miniq1.block = block;
1602 	miniqp->miniq2.block = block;
1603 }
1604 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1605 
1606 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1607 			  struct mini_Qdisc __rcu **p_miniq)
1608 {
1609 	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1610 	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1611 	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1612 	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1613 	miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1614 	miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1615 	miniqp->p_miniq = p_miniq;
1616 }
1617 EXPORT_SYMBOL(mini_qdisc_pair_init);
1618