xref: /linux/net/sched/sch_generic.c (revision a1ff5a7d78a036d6c2178ee5acd6ba4946243800)
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 
qdisc_maybe_clear_missed(struct Qdisc * q,const struct netdev_queue * txq)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 
__skb_dequeue_bad_txq(struct Qdisc * q)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 
qdisc_dequeue_skb_bad_txq(struct Qdisc * q)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 
qdisc_enqueue_skb_bad_txq(struct Qdisc * q,struct sk_buff * skb)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 
dev_requeue_skb(struct sk_buff * skb,struct Qdisc * q)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 
try_bulk_dequeue_skb(struct Qdisc * q,struct sk_buff * skb,const struct netdev_queue * txq,int * packets)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  */
try_bulk_dequeue_skb_slow(struct Qdisc * q,struct sk_buff * skb,int * packets)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  */
dequeue_skb(struct Qdisc * q,bool * validate,int * packets)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  */
sch_direct_xmit(struct sk_buff * skb,struct Qdisc * q,struct net_device * dev,struct netdev_queue * txq,spinlock_t * root_lock,bool validate)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  */
qdisc_restart(struct Qdisc * q,int * packets)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 
__qdisc_run(struct Qdisc * q)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 
dev_trans_start(struct net_device * dev)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 
netif_freeze_queues(struct net_device * dev)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 
netif_tx_lock(struct net_device * dev)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 
netif_unfreeze_queues(struct net_device * dev)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 
netif_tx_unlock(struct net_device * dev)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 
dev_watchdog(struct timer_list * t)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 
__netdev_watchdog_up(struct net_device * dev)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 
dev_watchdog_up(struct net_device * dev)561 static void dev_watchdog_up(struct net_device *dev)
562 {
563 	__netdev_watchdog_up(dev);
564 }
565 
dev_watchdog_down(struct net_device * dev)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  */
netif_carrier_on(struct net_device * dev)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  */
netif_carrier_off(struct net_device * dev)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  */
netif_carrier_event(struct net_device * dev)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 
noop_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)633 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
634 			struct sk_buff **to_free)
635 {
636 	dev_core_stats_tx_dropped_inc(skb->dev);
637 	__qdisc_drop(skb, to_free);
638 	return NET_XMIT_CN;
639 }
640 
noop_dequeue(struct Qdisc * qdisc)641 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
642 {
643 	return NULL;
644 }
645 
646 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
647 	.id		=	"noop",
648 	.priv_size	=	0,
649 	.enqueue	=	noop_enqueue,
650 	.dequeue	=	noop_dequeue,
651 	.peek		=	noop_dequeue,
652 	.owner		=	THIS_MODULE,
653 };
654 
655 static struct netdev_queue noop_netdev_queue = {
656 	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
657 	RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
658 };
659 
660 struct Qdisc noop_qdisc = {
661 	.enqueue	=	noop_enqueue,
662 	.dequeue	=	noop_dequeue,
663 	.flags		=	TCQ_F_BUILTIN,
664 	.ops		=	&noop_qdisc_ops,
665 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
666 	.dev_queue	=	&noop_netdev_queue,
667 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
668 	.gso_skb = {
669 		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
670 		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
671 		.qlen = 0,
672 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
673 	},
674 	.skb_bad_txq = {
675 		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
676 		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
677 		.qlen = 0,
678 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
679 	},
680 	.owner = -1,
681 };
682 EXPORT_SYMBOL(noop_qdisc);
683 
noqueue_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)684 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
685 			struct netlink_ext_ack *extack)
686 {
687 	/* register_qdisc() assigns a default of noop_enqueue if unset,
688 	 * but __dev_queue_xmit() treats noqueue only as such
689 	 * if this is NULL - so clear it here. */
690 	qdisc->enqueue = NULL;
691 	return 0;
692 }
693 
694 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
695 	.id		=	"noqueue",
696 	.priv_size	=	0,
697 	.init		=	noqueue_init,
698 	.enqueue	=	noop_enqueue,
699 	.dequeue	=	noop_dequeue,
700 	.peek		=	noop_dequeue,
701 	.owner		=	THIS_MODULE,
702 };
703 
704 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
705 	1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
706 };
707 EXPORT_SYMBOL(sch_default_prio2band);
708 
709 /* 3-band FIFO queue: old style, but should be a bit faster than
710    generic prio+fifo combination.
711  */
712 
713 #define PFIFO_FAST_BANDS 3
714 
715 /*
716  * Private data for a pfifo_fast scheduler containing:
717  *	- rings for priority bands
718  */
719 struct pfifo_fast_priv {
720 	struct skb_array q[PFIFO_FAST_BANDS];
721 };
722 
band2list(struct pfifo_fast_priv * priv,int band)723 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
724 					  int band)
725 {
726 	return &priv->q[band];
727 }
728 
pfifo_fast_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)729 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
730 			      struct sk_buff **to_free)
731 {
732 	int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
733 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
734 	struct skb_array *q = band2list(priv, band);
735 	unsigned int pkt_len = qdisc_pkt_len(skb);
736 	int err;
737 
738 	err = skb_array_produce(q, skb);
739 
740 	if (unlikely(err)) {
741 		if (qdisc_is_percpu_stats(qdisc))
742 			return qdisc_drop_cpu(skb, qdisc, to_free);
743 		else
744 			return qdisc_drop(skb, qdisc, to_free);
745 	}
746 
747 	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
748 	return NET_XMIT_SUCCESS;
749 }
750 
pfifo_fast_dequeue(struct Qdisc * qdisc)751 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
752 {
753 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
754 	struct sk_buff *skb = NULL;
755 	bool need_retry = true;
756 	int band;
757 
758 retry:
759 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
760 		struct skb_array *q = band2list(priv, band);
761 
762 		if (__skb_array_empty(q))
763 			continue;
764 
765 		skb = __skb_array_consume(q);
766 	}
767 	if (likely(skb)) {
768 		qdisc_update_stats_at_dequeue(qdisc, skb);
769 	} else if (need_retry &&
770 		   READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
771 		/* Delay clearing the STATE_MISSED here to reduce
772 		 * the overhead of the second spin_trylock() in
773 		 * qdisc_run_begin() and __netif_schedule() calling
774 		 * in qdisc_run_end().
775 		 */
776 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
777 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
778 
779 		/* Make sure dequeuing happens after clearing
780 		 * STATE_MISSED.
781 		 */
782 		smp_mb__after_atomic();
783 
784 		need_retry = false;
785 
786 		goto retry;
787 	}
788 
789 	return skb;
790 }
791 
pfifo_fast_peek(struct Qdisc * qdisc)792 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
793 {
794 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
795 	struct sk_buff *skb = NULL;
796 	int band;
797 
798 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
799 		struct skb_array *q = band2list(priv, band);
800 
801 		skb = __skb_array_peek(q);
802 	}
803 
804 	return skb;
805 }
806 
pfifo_fast_reset(struct Qdisc * qdisc)807 static void pfifo_fast_reset(struct Qdisc *qdisc)
808 {
809 	int i, band;
810 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
811 
812 	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
813 		struct skb_array *q = band2list(priv, band);
814 		struct sk_buff *skb;
815 
816 		/* NULL ring is possible if destroy path is due to a failed
817 		 * skb_array_init() in pfifo_fast_init() case.
818 		 */
819 		if (!q->ring.queue)
820 			continue;
821 
822 		while ((skb = __skb_array_consume(q)) != NULL)
823 			kfree_skb(skb);
824 	}
825 
826 	if (qdisc_is_percpu_stats(qdisc)) {
827 		for_each_possible_cpu(i) {
828 			struct gnet_stats_queue *q;
829 
830 			q = per_cpu_ptr(qdisc->cpu_qstats, i);
831 			q->backlog = 0;
832 			q->qlen = 0;
833 		}
834 	}
835 }
836 
pfifo_fast_dump(struct Qdisc * qdisc,struct sk_buff * skb)837 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
838 {
839 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
840 
841 	memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
842 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
843 		goto nla_put_failure;
844 	return skb->len;
845 
846 nla_put_failure:
847 	return -1;
848 }
849 
pfifo_fast_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)850 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
851 			   struct netlink_ext_ack *extack)
852 {
853 	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
854 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
855 	int prio;
856 
857 	/* guard against zero length rings */
858 	if (!qlen)
859 		return -EINVAL;
860 
861 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
862 		struct skb_array *q = band2list(priv, prio);
863 		int err;
864 
865 		err = skb_array_init(q, qlen, GFP_KERNEL);
866 		if (err)
867 			return -ENOMEM;
868 	}
869 
870 	/* Can by-pass the queue discipline */
871 	qdisc->flags |= TCQ_F_CAN_BYPASS;
872 	return 0;
873 }
874 
pfifo_fast_destroy(struct Qdisc * sch)875 static void pfifo_fast_destroy(struct Qdisc *sch)
876 {
877 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
878 	int prio;
879 
880 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
881 		struct skb_array *q = band2list(priv, prio);
882 
883 		/* NULL ring is possible if destroy path is due to a failed
884 		 * skb_array_init() in pfifo_fast_init() case.
885 		 */
886 		if (!q->ring.queue)
887 			continue;
888 		/* Destroy ring but no need to kfree_skb because a call to
889 		 * pfifo_fast_reset() has already done that work.
890 		 */
891 		ptr_ring_cleanup(&q->ring, NULL);
892 	}
893 }
894 
pfifo_fast_change_tx_queue_len(struct Qdisc * sch,unsigned int new_len)895 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
896 					  unsigned int new_len)
897 {
898 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
899 	struct skb_array *bands[PFIFO_FAST_BANDS];
900 	int prio;
901 
902 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
903 		struct skb_array *q = band2list(priv, prio);
904 
905 		bands[prio] = q;
906 	}
907 
908 	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
909 					 GFP_KERNEL);
910 }
911 
912 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
913 	.id		=	"pfifo_fast",
914 	.priv_size	=	sizeof(struct pfifo_fast_priv),
915 	.enqueue	=	pfifo_fast_enqueue,
916 	.dequeue	=	pfifo_fast_dequeue,
917 	.peek		=	pfifo_fast_peek,
918 	.init		=	pfifo_fast_init,
919 	.destroy	=	pfifo_fast_destroy,
920 	.reset		=	pfifo_fast_reset,
921 	.dump		=	pfifo_fast_dump,
922 	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
923 	.owner		=	THIS_MODULE,
924 	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
925 };
926 EXPORT_SYMBOL(pfifo_fast_ops);
927 
928 static struct lock_class_key qdisc_tx_busylock;
929 
qdisc_alloc(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,struct netlink_ext_ack * extack)930 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
931 			  const struct Qdisc_ops *ops,
932 			  struct netlink_ext_ack *extack)
933 {
934 	struct Qdisc *sch;
935 	unsigned int size = sizeof(*sch) + ops->priv_size;
936 	int err = -ENOBUFS;
937 	struct net_device *dev;
938 
939 	if (!dev_queue) {
940 		NL_SET_ERR_MSG(extack, "No device queue given");
941 		err = -EINVAL;
942 		goto errout;
943 	}
944 
945 	dev = dev_queue->dev;
946 	sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
947 
948 	if (!sch)
949 		goto errout;
950 	__skb_queue_head_init(&sch->gso_skb);
951 	__skb_queue_head_init(&sch->skb_bad_txq);
952 	gnet_stats_basic_sync_init(&sch->bstats);
953 	lockdep_register_key(&sch->root_lock_key);
954 	spin_lock_init(&sch->q.lock);
955 	lockdep_set_class(&sch->q.lock, &sch->root_lock_key);
956 
957 	if (ops->static_flags & TCQ_F_CPUSTATS) {
958 		sch->cpu_bstats =
959 			netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
960 		if (!sch->cpu_bstats)
961 			goto errout1;
962 
963 		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
964 		if (!sch->cpu_qstats) {
965 			free_percpu(sch->cpu_bstats);
966 			goto errout1;
967 		}
968 	}
969 
970 	spin_lock_init(&sch->busylock);
971 	lockdep_set_class(&sch->busylock,
972 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
973 
974 	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
975 	spin_lock_init(&sch->seqlock);
976 	lockdep_set_class(&sch->seqlock,
977 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
978 
979 	sch->ops = ops;
980 	sch->flags = ops->static_flags;
981 	sch->enqueue = ops->enqueue;
982 	sch->dequeue = ops->dequeue;
983 	sch->dev_queue = dev_queue;
984 	sch->owner = -1;
985 	netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
986 	refcount_set(&sch->refcnt, 1);
987 
988 	return sch;
989 errout1:
990 	lockdep_unregister_key(&sch->root_lock_key);
991 	kfree(sch);
992 errout:
993 	return ERR_PTR(err);
994 }
995 
qdisc_create_dflt(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,unsigned int parentid,struct netlink_ext_ack * extack)996 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
997 				const struct Qdisc_ops *ops,
998 				unsigned int parentid,
999 				struct netlink_ext_ack *extack)
1000 {
1001 	struct Qdisc *sch;
1002 
1003 	if (!try_module_get(ops->owner)) {
1004 		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1005 		return NULL;
1006 	}
1007 
1008 	sch = qdisc_alloc(dev_queue, ops, extack);
1009 	if (IS_ERR(sch)) {
1010 		module_put(ops->owner);
1011 		return NULL;
1012 	}
1013 	sch->parent = parentid;
1014 
1015 	if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1016 		trace_qdisc_create(ops, dev_queue->dev, parentid);
1017 		return sch;
1018 	}
1019 
1020 	qdisc_put(sch);
1021 	return NULL;
1022 }
1023 EXPORT_SYMBOL(qdisc_create_dflt);
1024 
1025 /* Under qdisc_lock(qdisc) and BH! */
1026 
qdisc_reset(struct Qdisc * qdisc)1027 void qdisc_reset(struct Qdisc *qdisc)
1028 {
1029 	const struct Qdisc_ops *ops = qdisc->ops;
1030 
1031 	trace_qdisc_reset(qdisc);
1032 
1033 	if (ops->reset)
1034 		ops->reset(qdisc);
1035 
1036 	__skb_queue_purge(&qdisc->gso_skb);
1037 	__skb_queue_purge(&qdisc->skb_bad_txq);
1038 
1039 	qdisc->q.qlen = 0;
1040 	qdisc->qstats.backlog = 0;
1041 }
1042 EXPORT_SYMBOL(qdisc_reset);
1043 
qdisc_free(struct Qdisc * qdisc)1044 void qdisc_free(struct Qdisc *qdisc)
1045 {
1046 	if (qdisc_is_percpu_stats(qdisc)) {
1047 		free_percpu(qdisc->cpu_bstats);
1048 		free_percpu(qdisc->cpu_qstats);
1049 	}
1050 
1051 	kfree(qdisc);
1052 }
1053 
qdisc_free_cb(struct rcu_head * head)1054 static void qdisc_free_cb(struct rcu_head *head)
1055 {
1056 	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1057 
1058 	qdisc_free(q);
1059 }
1060 
__qdisc_destroy(struct Qdisc * qdisc)1061 static void __qdisc_destroy(struct Qdisc *qdisc)
1062 {
1063 	const struct Qdisc_ops  *ops = qdisc->ops;
1064 	struct net_device *dev = qdisc_dev(qdisc);
1065 
1066 #ifdef CONFIG_NET_SCHED
1067 	qdisc_hash_del(qdisc);
1068 
1069 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
1070 #endif
1071 	gen_kill_estimator(&qdisc->rate_est);
1072 
1073 	qdisc_reset(qdisc);
1074 
1075 
1076 	if (ops->destroy)
1077 		ops->destroy(qdisc);
1078 
1079 	lockdep_unregister_key(&qdisc->root_lock_key);
1080 	module_put(ops->owner);
1081 	netdev_put(dev, &qdisc->dev_tracker);
1082 
1083 	trace_qdisc_destroy(qdisc);
1084 
1085 	call_rcu(&qdisc->rcu, qdisc_free_cb);
1086 }
1087 
qdisc_destroy(struct Qdisc * qdisc)1088 void qdisc_destroy(struct Qdisc *qdisc)
1089 {
1090 	if (qdisc->flags & TCQ_F_BUILTIN)
1091 		return;
1092 
1093 	__qdisc_destroy(qdisc);
1094 }
1095 
qdisc_put(struct Qdisc * qdisc)1096 void qdisc_put(struct Qdisc *qdisc)
1097 {
1098 	if (!qdisc)
1099 		return;
1100 
1101 	if (qdisc->flags & TCQ_F_BUILTIN ||
1102 	    !refcount_dec_and_test(&qdisc->refcnt))
1103 		return;
1104 
1105 	__qdisc_destroy(qdisc);
1106 }
1107 EXPORT_SYMBOL(qdisc_put);
1108 
1109 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1110  * Intended to be used as optimization, this function only takes rtnl lock if
1111  * qdisc reference counter reached zero.
1112  */
1113 
qdisc_put_unlocked(struct Qdisc * qdisc)1114 void qdisc_put_unlocked(struct Qdisc *qdisc)
1115 {
1116 	if (qdisc->flags & TCQ_F_BUILTIN ||
1117 	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1118 		return;
1119 
1120 	__qdisc_destroy(qdisc);
1121 	rtnl_unlock();
1122 }
1123 EXPORT_SYMBOL(qdisc_put_unlocked);
1124 
1125 /* Attach toplevel qdisc to device queue. */
dev_graft_qdisc(struct netdev_queue * dev_queue,struct Qdisc * qdisc)1126 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1127 			      struct Qdisc *qdisc)
1128 {
1129 	struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1130 	spinlock_t *root_lock;
1131 
1132 	root_lock = qdisc_lock(oqdisc);
1133 	spin_lock_bh(root_lock);
1134 
1135 	/* ... and graft new one */
1136 	if (qdisc == NULL)
1137 		qdisc = &noop_qdisc;
1138 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1139 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1140 
1141 	spin_unlock_bh(root_lock);
1142 
1143 	return oqdisc;
1144 }
1145 EXPORT_SYMBOL(dev_graft_qdisc);
1146 
shutdown_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1147 static void shutdown_scheduler_queue(struct net_device *dev,
1148 				     struct netdev_queue *dev_queue,
1149 				     void *_qdisc_default)
1150 {
1151 	struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1152 	struct Qdisc *qdisc_default = _qdisc_default;
1153 
1154 	if (qdisc) {
1155 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1156 		rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1157 
1158 		qdisc_put(qdisc);
1159 	}
1160 }
1161 
attach_one_default_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1162 static void attach_one_default_qdisc(struct net_device *dev,
1163 				     struct netdev_queue *dev_queue,
1164 				     void *_unused)
1165 {
1166 	struct Qdisc *qdisc;
1167 	const struct Qdisc_ops *ops = default_qdisc_ops;
1168 
1169 	if (dev->priv_flags & IFF_NO_QUEUE)
1170 		ops = &noqueue_qdisc_ops;
1171 	else if(dev->type == ARPHRD_CAN)
1172 		ops = &pfifo_fast_ops;
1173 
1174 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1175 	if (!qdisc)
1176 		return;
1177 
1178 	if (!netif_is_multiqueue(dev))
1179 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1180 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1181 }
1182 
attach_default_qdiscs(struct net_device * dev)1183 static void attach_default_qdiscs(struct net_device *dev)
1184 {
1185 	struct netdev_queue *txq;
1186 	struct Qdisc *qdisc;
1187 
1188 	txq = netdev_get_tx_queue(dev, 0);
1189 
1190 	if (!netif_is_multiqueue(dev) ||
1191 	    dev->priv_flags & IFF_NO_QUEUE) {
1192 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1193 		qdisc = rtnl_dereference(txq->qdisc_sleeping);
1194 		rcu_assign_pointer(dev->qdisc, qdisc);
1195 		qdisc_refcount_inc(qdisc);
1196 	} else {
1197 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1198 		if (qdisc) {
1199 			rcu_assign_pointer(dev->qdisc, qdisc);
1200 			qdisc->ops->attach(qdisc);
1201 		}
1202 	}
1203 	qdisc = rtnl_dereference(dev->qdisc);
1204 
1205 	/* Detect default qdisc setup/init failed and fallback to "noqueue" */
1206 	if (qdisc == &noop_qdisc) {
1207 		netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1208 			    default_qdisc_ops->id, noqueue_qdisc_ops.id);
1209 		netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1210 		dev->priv_flags |= IFF_NO_QUEUE;
1211 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1212 		qdisc = rtnl_dereference(txq->qdisc_sleeping);
1213 		rcu_assign_pointer(dev->qdisc, qdisc);
1214 		qdisc_refcount_inc(qdisc);
1215 		dev->priv_flags ^= IFF_NO_QUEUE;
1216 	}
1217 
1218 #ifdef CONFIG_NET_SCHED
1219 	if (qdisc != &noop_qdisc)
1220 		qdisc_hash_add(qdisc, false);
1221 #endif
1222 }
1223 
transition_one_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _need_watchdog)1224 static void transition_one_qdisc(struct net_device *dev,
1225 				 struct netdev_queue *dev_queue,
1226 				 void *_need_watchdog)
1227 {
1228 	struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1229 	int *need_watchdog_p = _need_watchdog;
1230 
1231 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1232 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1233 
1234 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1235 	if (need_watchdog_p) {
1236 		WRITE_ONCE(dev_queue->trans_start, 0);
1237 		*need_watchdog_p = 1;
1238 	}
1239 }
1240 
dev_activate(struct net_device * dev)1241 void dev_activate(struct net_device *dev)
1242 {
1243 	int need_watchdog;
1244 
1245 	/* No queueing discipline is attached to device;
1246 	 * create default one for devices, which need queueing
1247 	 * and noqueue_qdisc for virtual interfaces
1248 	 */
1249 
1250 	if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1251 		attach_default_qdiscs(dev);
1252 
1253 	if (!netif_carrier_ok(dev))
1254 		/* Delay activation until next carrier-on event */
1255 		return;
1256 
1257 	need_watchdog = 0;
1258 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1259 	if (dev_ingress_queue(dev))
1260 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1261 
1262 	if (need_watchdog) {
1263 		netif_trans_update(dev);
1264 		dev_watchdog_up(dev);
1265 	}
1266 }
1267 EXPORT_SYMBOL(dev_activate);
1268 
qdisc_deactivate(struct Qdisc * qdisc)1269 static void qdisc_deactivate(struct Qdisc *qdisc)
1270 {
1271 	if (qdisc->flags & TCQ_F_BUILTIN)
1272 		return;
1273 
1274 	set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1275 }
1276 
dev_deactivate_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1277 static void dev_deactivate_queue(struct net_device *dev,
1278 				 struct netdev_queue *dev_queue,
1279 				 void *_qdisc_default)
1280 {
1281 	struct Qdisc *qdisc_default = _qdisc_default;
1282 	struct Qdisc *qdisc;
1283 
1284 	qdisc = rtnl_dereference(dev_queue->qdisc);
1285 	if (qdisc) {
1286 		qdisc_deactivate(qdisc);
1287 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1288 	}
1289 }
1290 
dev_reset_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1291 static void dev_reset_queue(struct net_device *dev,
1292 			    struct netdev_queue *dev_queue,
1293 			    void *_unused)
1294 {
1295 	struct Qdisc *qdisc;
1296 	bool nolock;
1297 
1298 	qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1299 	if (!qdisc)
1300 		return;
1301 
1302 	nolock = qdisc->flags & TCQ_F_NOLOCK;
1303 
1304 	if (nolock)
1305 		spin_lock_bh(&qdisc->seqlock);
1306 	spin_lock_bh(qdisc_lock(qdisc));
1307 
1308 	qdisc_reset(qdisc);
1309 
1310 	spin_unlock_bh(qdisc_lock(qdisc));
1311 	if (nolock) {
1312 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1313 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1314 		spin_unlock_bh(&qdisc->seqlock);
1315 	}
1316 }
1317 
some_qdisc_is_busy(struct net_device * dev)1318 static bool some_qdisc_is_busy(struct net_device *dev)
1319 {
1320 	unsigned int i;
1321 
1322 	for (i = 0; i < dev->num_tx_queues; i++) {
1323 		struct netdev_queue *dev_queue;
1324 		spinlock_t *root_lock;
1325 		struct Qdisc *q;
1326 		int val;
1327 
1328 		dev_queue = netdev_get_tx_queue(dev, i);
1329 		q = rtnl_dereference(dev_queue->qdisc_sleeping);
1330 
1331 		root_lock = qdisc_lock(q);
1332 		spin_lock_bh(root_lock);
1333 
1334 		val = (qdisc_is_running(q) ||
1335 		       test_bit(__QDISC_STATE_SCHED, &q->state));
1336 
1337 		spin_unlock_bh(root_lock);
1338 
1339 		if (val)
1340 			return true;
1341 	}
1342 	return false;
1343 }
1344 
1345 /**
1346  * 	dev_deactivate_many - deactivate transmissions on several devices
1347  * 	@head: list of devices to deactivate
1348  *
1349  *	This function returns only when all outstanding transmissions
1350  *	have completed, unless all devices are in dismantle phase.
1351  */
dev_deactivate_many(struct list_head * head)1352 void dev_deactivate_many(struct list_head *head)
1353 {
1354 	struct net_device *dev;
1355 
1356 	list_for_each_entry(dev, head, close_list) {
1357 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1358 					 &noop_qdisc);
1359 		if (dev_ingress_queue(dev))
1360 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1361 					     &noop_qdisc);
1362 
1363 		dev_watchdog_down(dev);
1364 	}
1365 
1366 	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
1367 	 * outstanding qdisc enqueuing calls.
1368 	 * This is avoided if all devices are in dismantle phase :
1369 	 * Caller will call synchronize_net() for us
1370 	 */
1371 	synchronize_net();
1372 
1373 	list_for_each_entry(dev, head, close_list) {
1374 		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1375 
1376 		if (dev_ingress_queue(dev))
1377 			dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1378 	}
1379 
1380 	/* Wait for outstanding qdisc_run calls. */
1381 	list_for_each_entry(dev, head, close_list) {
1382 		while (some_qdisc_is_busy(dev)) {
1383 			/* wait_event() would avoid this sleep-loop but would
1384 			 * require expensive checks in the fast paths of packet
1385 			 * processing which isn't worth it.
1386 			 */
1387 			schedule_timeout_uninterruptible(1);
1388 		}
1389 	}
1390 }
1391 
dev_deactivate(struct net_device * dev)1392 void dev_deactivate(struct net_device *dev)
1393 {
1394 	LIST_HEAD(single);
1395 
1396 	list_add(&dev->close_list, &single);
1397 	dev_deactivate_many(&single);
1398 	list_del(&single);
1399 }
1400 EXPORT_SYMBOL(dev_deactivate);
1401 
qdisc_change_tx_queue_len(struct net_device * dev,struct netdev_queue * dev_queue)1402 static int qdisc_change_tx_queue_len(struct net_device *dev,
1403 				     struct netdev_queue *dev_queue)
1404 {
1405 	struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1406 	const struct Qdisc_ops *ops = qdisc->ops;
1407 
1408 	if (ops->change_tx_queue_len)
1409 		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1410 	return 0;
1411 }
1412 
dev_qdisc_change_real_num_tx(struct net_device * dev,unsigned int new_real_tx)1413 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1414 				  unsigned int new_real_tx)
1415 {
1416 	struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1417 
1418 	if (qdisc->ops->change_real_num_tx)
1419 		qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1420 }
1421 
mq_change_real_num_tx(struct Qdisc * sch,unsigned int new_real_tx)1422 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1423 {
1424 #ifdef CONFIG_NET_SCHED
1425 	struct net_device *dev = qdisc_dev(sch);
1426 	struct Qdisc *qdisc;
1427 	unsigned int i;
1428 
1429 	for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1430 		qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1431 		/* Only update the default qdiscs we created,
1432 		 * qdiscs with handles are always hashed.
1433 		 */
1434 		if (qdisc != &noop_qdisc && !qdisc->handle)
1435 			qdisc_hash_del(qdisc);
1436 	}
1437 	for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1438 		qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1439 		if (qdisc != &noop_qdisc && !qdisc->handle)
1440 			qdisc_hash_add(qdisc, false);
1441 	}
1442 #endif
1443 }
1444 EXPORT_SYMBOL(mq_change_real_num_tx);
1445 
dev_qdisc_change_tx_queue_len(struct net_device * dev)1446 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1447 {
1448 	bool up = dev->flags & IFF_UP;
1449 	unsigned int i;
1450 	int ret = 0;
1451 
1452 	if (up)
1453 		dev_deactivate(dev);
1454 
1455 	for (i = 0; i < dev->num_tx_queues; i++) {
1456 		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1457 
1458 		/* TODO: revert changes on a partial failure */
1459 		if (ret)
1460 			break;
1461 	}
1462 
1463 	if (up)
1464 		dev_activate(dev);
1465 	return ret;
1466 }
1467 
dev_init_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc)1468 static void dev_init_scheduler_queue(struct net_device *dev,
1469 				     struct netdev_queue *dev_queue,
1470 				     void *_qdisc)
1471 {
1472 	struct Qdisc *qdisc = _qdisc;
1473 
1474 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1475 	rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1476 }
1477 
dev_init_scheduler(struct net_device * dev)1478 void dev_init_scheduler(struct net_device *dev)
1479 {
1480 	rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1481 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1482 	if (dev_ingress_queue(dev))
1483 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1484 
1485 	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1486 }
1487 
dev_shutdown(struct net_device * dev)1488 void dev_shutdown(struct net_device *dev)
1489 {
1490 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1491 	if (dev_ingress_queue(dev))
1492 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1493 	qdisc_put(rtnl_dereference(dev->qdisc));
1494 	rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1495 
1496 	WARN_ON(timer_pending(&dev->watchdog_timer));
1497 }
1498 
1499 /**
1500  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1501  * @rate:   Rate to compute reciprocal division values of
1502  * @mult:   Multiplier for reciprocal division
1503  * @shift:  Shift for reciprocal division
1504  *
1505  * The multiplier and shift for reciprocal division by rate are stored
1506  * in mult and shift.
1507  *
1508  * The deal here is to replace a divide by a reciprocal one
1509  * in fast path (a reciprocal divide is a multiply and a shift)
1510  *
1511  * Normal formula would be :
1512  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1513  *
1514  * We compute mult/shift to use instead :
1515  *  time_in_ns = (len * mult) >> shift;
1516  *
1517  * We try to get the highest possible mult value for accuracy,
1518  * but have to make sure no overflows will ever happen.
1519  *
1520  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1521  */
psched_ratecfg_precompute__(u64 rate,u32 * mult,u8 * shift)1522 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1523 {
1524 	u64 factor = NSEC_PER_SEC;
1525 
1526 	*mult = 1;
1527 	*shift = 0;
1528 
1529 	if (rate <= 0)
1530 		return;
1531 
1532 	for (;;) {
1533 		*mult = div64_u64(factor, rate);
1534 		if (*mult & (1U << 31) || factor & (1ULL << 63))
1535 			break;
1536 		factor <<= 1;
1537 		(*shift)++;
1538 	}
1539 }
1540 
psched_ratecfg_precompute(struct psched_ratecfg * r,const struct tc_ratespec * conf,u64 rate64)1541 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1542 			       const struct tc_ratespec *conf,
1543 			       u64 rate64)
1544 {
1545 	memset(r, 0, sizeof(*r));
1546 	r->overhead = conf->overhead;
1547 	r->mpu = conf->mpu;
1548 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1549 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1550 	psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1551 }
1552 EXPORT_SYMBOL(psched_ratecfg_precompute);
1553 
psched_ppscfg_precompute(struct psched_pktrate * r,u64 pktrate64)1554 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1555 {
1556 	r->rate_pkts_ps = pktrate64;
1557 	psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1558 }
1559 EXPORT_SYMBOL(psched_ppscfg_precompute);
1560 
mini_qdisc_pair_swap(struct mini_Qdisc_pair * miniqp,struct tcf_proto * tp_head)1561 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1562 			  struct tcf_proto *tp_head)
1563 {
1564 	/* Protected with chain0->filter_chain_lock.
1565 	 * Can't access chain directly because tp_head can be NULL.
1566 	 */
1567 	struct mini_Qdisc *miniq_old =
1568 		rcu_dereference_protected(*miniqp->p_miniq, 1);
1569 	struct mini_Qdisc *miniq;
1570 
1571 	if (!tp_head) {
1572 		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1573 	} else {
1574 		miniq = miniq_old != &miniqp->miniq1 ?
1575 			&miniqp->miniq1 : &miniqp->miniq2;
1576 
1577 		/* We need to make sure that readers won't see the miniq
1578 		 * we are about to modify. So ensure that at least one RCU
1579 		 * grace period has elapsed since the miniq was made
1580 		 * inactive.
1581 		 */
1582 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
1583 			cond_synchronize_rcu(miniq->rcu_state);
1584 		else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1585 			synchronize_rcu_expedited();
1586 
1587 		miniq->filter_list = tp_head;
1588 		rcu_assign_pointer(*miniqp->p_miniq, miniq);
1589 	}
1590 
1591 	if (miniq_old)
1592 		/* This is counterpart of the rcu sync above. We need to
1593 		 * block potential new user of miniq_old until all readers
1594 		 * are not seeing it.
1595 		 */
1596 		miniq_old->rcu_state = start_poll_synchronize_rcu();
1597 }
1598 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1599 
mini_qdisc_pair_block_init(struct mini_Qdisc_pair * miniqp,struct tcf_block * block)1600 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1601 				struct tcf_block *block)
1602 {
1603 	miniqp->miniq1.block = block;
1604 	miniqp->miniq2.block = block;
1605 }
1606 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1607 
mini_qdisc_pair_init(struct mini_Qdisc_pair * miniqp,struct Qdisc * qdisc,struct mini_Qdisc __rcu ** p_miniq)1608 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1609 			  struct mini_Qdisc __rcu **p_miniq)
1610 {
1611 	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1612 	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1613 	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1614 	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1615 	miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1616 	miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1617 	miniqp->p_miniq = p_miniq;
1618 }
1619 EXPORT_SYMBOL(mini_qdisc_pair_init);
1620