1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * net/sched/sch_generic.c	Generic packet scheduler routines.
4  *
5  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
7  *              - Ingress support
8  */
9 
10 #include <linux/bitops.h>
11 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/netdevice.h>
18 #include <linux/skbuff.h>
19 #include <linux/rtnetlink.h>
20 #include <linux/init.h>
21 #include <linux/rcupdate.h>
22 #include <linux/list.h>
23 #include <linux/slab.h>
24 #include <linux/if_vlan.h>
25 #include <linux/skb_array.h>
26 #include <linux/if_macvlan.h>
27 #include <net/sch_generic.h>
28 #include <net/pkt_sched.h>
29 #include <net/dst.h>
30 #include <trace/events/qdisc.h>
31 #include <trace/events/net.h>
32 #include <net/xfrm.h>
33 
34 /* Qdisc to use by default */
35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36 EXPORT_SYMBOL(default_qdisc_ops);
37 
qdisc_maybe_clear_missed(struct Qdisc * q,const struct netdev_queue * txq)38 static void qdisc_maybe_clear_missed(struct Qdisc *q,
39 				     const struct netdev_queue *txq)
40 {
41 	clear_bit(__QDISC_STATE_MISSED, &q->state);
42 
43 	/* Make sure the below netif_xmit_frozen_or_stopped()
44 	 * checking happens after clearing STATE_MISSED.
45 	 */
46 	smp_mb__after_atomic();
47 
48 	/* Checking netif_xmit_frozen_or_stopped() again to
49 	 * make sure STATE_MISSED is set if the STATE_MISSED
50 	 * set by netif_tx_wake_queue()'s rescheduling of
51 	 * net_tx_action() is cleared by the above clear_bit().
52 	 */
53 	if (!netif_xmit_frozen_or_stopped(txq))
54 		set_bit(__QDISC_STATE_MISSED, &q->state);
55 	else
56 		set_bit(__QDISC_STATE_DRAINING, &q->state);
57 }
58 
59 /* Main transmission queue. */
60 
61 /* Modifications to data participating in scheduling must be protected with
62  * qdisc_lock(qdisc) spinlock.
63  *
64  * The idea is the following:
65  * - enqueue, dequeue are serialized via qdisc root lock
66  * - ingress filtering is also serialized via qdisc root lock
67  * - updates to tree and tree walking are only done under the rtnl mutex.
68  */
69 
70 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
71 
__skb_dequeue_bad_txq(struct Qdisc * q)72 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
73 {
74 	const struct netdev_queue *txq = q->dev_queue;
75 	spinlock_t *lock = NULL;
76 	struct sk_buff *skb;
77 
78 	if (q->flags & TCQ_F_NOLOCK) {
79 		lock = qdisc_lock(q);
80 		spin_lock(lock);
81 	}
82 
83 	skb = skb_peek(&q->skb_bad_txq);
84 	if (skb) {
85 		/* check the reason of requeuing without tx lock first */
86 		txq = skb_get_tx_queue(txq->dev, skb);
87 		if (!netif_xmit_frozen_or_stopped(txq)) {
88 			skb = __skb_dequeue(&q->skb_bad_txq);
89 			if (qdisc_is_percpu_stats(q)) {
90 				qdisc_qstats_cpu_backlog_dec(q, skb);
91 				qdisc_qstats_cpu_qlen_dec(q);
92 			} else {
93 				qdisc_qstats_backlog_dec(q, skb);
94 				q->q.qlen--;
95 			}
96 		} else {
97 			skb = SKB_XOFF_MAGIC;
98 			qdisc_maybe_clear_missed(q, txq);
99 		}
100 	}
101 
102 	if (lock)
103 		spin_unlock(lock);
104 
105 	return skb;
106 }
107 
qdisc_dequeue_skb_bad_txq(struct Qdisc * q)108 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
109 {
110 	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
111 
112 	if (unlikely(skb))
113 		skb = __skb_dequeue_bad_txq(q);
114 
115 	return skb;
116 }
117 
qdisc_enqueue_skb_bad_txq(struct Qdisc * q,struct sk_buff * skb)118 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
119 					     struct sk_buff *skb)
120 {
121 	spinlock_t *lock = NULL;
122 
123 	if (q->flags & TCQ_F_NOLOCK) {
124 		lock = qdisc_lock(q);
125 		spin_lock(lock);
126 	}
127 
128 	__skb_queue_tail(&q->skb_bad_txq, skb);
129 
130 	if (qdisc_is_percpu_stats(q)) {
131 		qdisc_qstats_cpu_backlog_inc(q, skb);
132 		qdisc_qstats_cpu_qlen_inc(q);
133 	} else {
134 		qdisc_qstats_backlog_inc(q, skb);
135 		q->q.qlen++;
136 	}
137 
138 	if (lock)
139 		spin_unlock(lock);
140 }
141 
dev_requeue_skb(struct sk_buff * skb,struct Qdisc * q)142 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
143 {
144 	spinlock_t *lock = NULL;
145 
146 	if (q->flags & TCQ_F_NOLOCK) {
147 		lock = qdisc_lock(q);
148 		spin_lock(lock);
149 	}
150 
151 	while (skb) {
152 		struct sk_buff *next = skb->next;
153 
154 		__skb_queue_tail(&q->gso_skb, skb);
155 
156 		/* it's still part of the queue */
157 		if (qdisc_is_percpu_stats(q)) {
158 			qdisc_qstats_cpu_requeues_inc(q);
159 			qdisc_qstats_cpu_backlog_inc(q, skb);
160 			qdisc_qstats_cpu_qlen_inc(q);
161 		} else {
162 			q->qstats.requeues++;
163 			qdisc_qstats_backlog_inc(q, skb);
164 			q->q.qlen++;
165 		}
166 
167 		skb = next;
168 	}
169 
170 	if (lock) {
171 		spin_unlock(lock);
172 		set_bit(__QDISC_STATE_MISSED, &q->state);
173 	} else {
174 		__netif_schedule(q);
175 	}
176 }
177 
try_bulk_dequeue_skb(struct Qdisc * q,struct sk_buff * skb,const struct netdev_queue * txq,int * packets)178 static void try_bulk_dequeue_skb(struct Qdisc *q,
179 				 struct sk_buff *skb,
180 				 const struct netdev_queue *txq,
181 				 int *packets)
182 {
183 	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
184 
185 	while (bytelimit > 0) {
186 		struct sk_buff *nskb = q->dequeue(q);
187 
188 		if (!nskb)
189 			break;
190 
191 		bytelimit -= nskb->len; /* covers GSO len */
192 		skb->next = nskb;
193 		skb = nskb;
194 		(*packets)++; /* GSO counts as one pkt */
195 	}
196 	skb_mark_not_on_list(skb);
197 }
198 
199 /* This variant of try_bulk_dequeue_skb() makes sure
200  * all skbs in the chain are for the same txq
201  */
try_bulk_dequeue_skb_slow(struct Qdisc * q,struct sk_buff * skb,int * packets)202 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
203 				      struct sk_buff *skb,
204 				      int *packets)
205 {
206 	int mapping = skb_get_queue_mapping(skb);
207 	struct sk_buff *nskb;
208 	int cnt = 0;
209 
210 	do {
211 		nskb = q->dequeue(q);
212 		if (!nskb)
213 			break;
214 		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
215 			qdisc_enqueue_skb_bad_txq(q, nskb);
216 			break;
217 		}
218 		skb->next = nskb;
219 		skb = nskb;
220 	} while (++cnt < 8);
221 	(*packets) += cnt;
222 	skb_mark_not_on_list(skb);
223 }
224 
225 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
226  * A requeued skb (via q->gso_skb) can also be a SKB list.
227  */
dequeue_skb(struct Qdisc * q,bool * validate,int * packets)228 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
229 				   int *packets)
230 {
231 	const struct netdev_queue *txq = q->dev_queue;
232 	struct sk_buff *skb = NULL;
233 
234 	*packets = 1;
235 	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
236 		spinlock_t *lock = NULL;
237 
238 		if (q->flags & TCQ_F_NOLOCK) {
239 			lock = qdisc_lock(q);
240 			spin_lock(lock);
241 		}
242 
243 		skb = skb_peek(&q->gso_skb);
244 
245 		/* skb may be null if another cpu pulls gso_skb off in between
246 		 * empty check and lock.
247 		 */
248 		if (!skb) {
249 			if (lock)
250 				spin_unlock(lock);
251 			goto validate;
252 		}
253 
254 		/* skb in gso_skb were already validated */
255 		*validate = false;
256 		if (xfrm_offload(skb))
257 			*validate = true;
258 		/* check the reason of requeuing without tx lock first */
259 		txq = skb_get_tx_queue(txq->dev, skb);
260 		if (!netif_xmit_frozen_or_stopped(txq)) {
261 			skb = __skb_dequeue(&q->gso_skb);
262 			if (qdisc_is_percpu_stats(q)) {
263 				qdisc_qstats_cpu_backlog_dec(q, skb);
264 				qdisc_qstats_cpu_qlen_dec(q);
265 			} else {
266 				qdisc_qstats_backlog_dec(q, skb);
267 				q->q.qlen--;
268 			}
269 		} else {
270 			skb = NULL;
271 			qdisc_maybe_clear_missed(q, txq);
272 		}
273 		if (lock)
274 			spin_unlock(lock);
275 		goto trace;
276 	}
277 validate:
278 	*validate = true;
279 
280 	if ((q->flags & TCQ_F_ONETXQUEUE) &&
281 	    netif_xmit_frozen_or_stopped(txq)) {
282 		qdisc_maybe_clear_missed(q, txq);
283 		return skb;
284 	}
285 
286 	skb = qdisc_dequeue_skb_bad_txq(q);
287 	if (unlikely(skb)) {
288 		if (skb == SKB_XOFF_MAGIC)
289 			return NULL;
290 		goto bulk;
291 	}
292 	skb = q->dequeue(q);
293 	if (skb) {
294 bulk:
295 		if (qdisc_may_bulk(q))
296 			try_bulk_dequeue_skb(q, skb, txq, packets);
297 		else
298 			try_bulk_dequeue_skb_slow(q, skb, packets);
299 	}
300 trace:
301 	trace_qdisc_dequeue(q, txq, *packets, skb);
302 	return skb;
303 }
304 
305 /*
306  * Transmit possibly several skbs, and handle the return status as
307  * required. Owning qdisc running bit guarantees that only one CPU
308  * can execute this function.
309  *
310  * Returns to the caller:
311  *				false  - hardware queue frozen backoff
312  *				true   - feel free to send more pkts
313  */
sch_direct_xmit(struct sk_buff * skb,struct Qdisc * q,struct net_device * dev,struct netdev_queue * txq,spinlock_t * root_lock,bool validate)314 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
315 		     struct net_device *dev, struct netdev_queue *txq,
316 		     spinlock_t *root_lock, bool validate)
317 {
318 	int ret = NETDEV_TX_BUSY;
319 	bool again = false;
320 
321 	/* And release qdisc */
322 	if (root_lock)
323 		spin_unlock(root_lock);
324 
325 	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
326 	if (validate)
327 		skb = validate_xmit_skb_list(skb, dev, &again);
328 
329 #ifdef CONFIG_XFRM_OFFLOAD
330 	if (unlikely(again)) {
331 		if (root_lock)
332 			spin_lock(root_lock);
333 
334 		dev_requeue_skb(skb, q);
335 		return false;
336 	}
337 #endif
338 
339 	if (likely(skb)) {
340 		HARD_TX_LOCK(dev, txq, smp_processor_id());
341 		if (!netif_xmit_frozen_or_stopped(txq))
342 			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
343 		else
344 			qdisc_maybe_clear_missed(q, txq);
345 
346 		HARD_TX_UNLOCK(dev, txq);
347 	} else {
348 		if (root_lock)
349 			spin_lock(root_lock);
350 		return true;
351 	}
352 
353 	if (root_lock)
354 		spin_lock(root_lock);
355 
356 	if (!dev_xmit_complete(ret)) {
357 		/* Driver returned NETDEV_TX_BUSY - requeue skb */
358 		if (unlikely(ret != NETDEV_TX_BUSY))
359 			net_warn_ratelimited("BUG %s code %d qlen %d\n",
360 					     dev->name, ret, q->q.qlen);
361 
362 		dev_requeue_skb(skb, q);
363 		return false;
364 	}
365 
366 	return true;
367 }
368 
369 /*
370  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
371  *
372  * running seqcount guarantees only one CPU can process
373  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
374  * this queue.
375  *
376  *  netif_tx_lock serializes accesses to device driver.
377  *
378  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
379  *  if one is grabbed, another must be free.
380  *
381  * Note, that this procedure can be called by a watchdog timer
382  *
383  * Returns to the caller:
384  *				0  - queue is empty or throttled.
385  *				>0 - queue is not empty.
386  *
387  */
qdisc_restart(struct Qdisc * q,int * packets)388 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
389 {
390 	spinlock_t *root_lock = NULL;
391 	struct netdev_queue *txq;
392 	struct net_device *dev;
393 	struct sk_buff *skb;
394 	bool validate;
395 
396 	/* Dequeue packet */
397 	skb = dequeue_skb(q, &validate, packets);
398 	if (unlikely(!skb))
399 		return false;
400 
401 	if (!(q->flags & TCQ_F_NOLOCK))
402 		root_lock = qdisc_lock(q);
403 
404 	dev = qdisc_dev(q);
405 	txq = skb_get_tx_queue(dev, skb);
406 
407 	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
408 }
409 
__qdisc_run(struct Qdisc * q)410 void __qdisc_run(struct Qdisc *q)
411 {
412 	int quota = dev_tx_weight;
413 	int packets;
414 
415 	while (qdisc_restart(q, &packets)) {
416 		quota -= packets;
417 		if (quota <= 0) {
418 			if (q->flags & TCQ_F_NOLOCK)
419 				set_bit(__QDISC_STATE_MISSED, &q->state);
420 			else
421 				__netif_schedule(q);
422 
423 			break;
424 		}
425 	}
426 }
427 
dev_trans_start(struct net_device * dev)428 unsigned long dev_trans_start(struct net_device *dev)
429 {
430 	unsigned long val, res;
431 	unsigned int i;
432 
433 	if (is_vlan_dev(dev))
434 		dev = vlan_dev_real_dev(dev);
435 	else if (netif_is_macvlan(dev))
436 		dev = macvlan_dev_real_dev(dev);
437 	res = netdev_get_tx_queue(dev, 0)->trans_start;
438 	for (i = 1; i < dev->num_tx_queues; i++) {
439 		val = netdev_get_tx_queue(dev, i)->trans_start;
440 		if (val && time_after(val, res))
441 			res = val;
442 	}
443 
444 	return res;
445 }
446 EXPORT_SYMBOL(dev_trans_start);
447 
dev_watchdog(struct timer_list * t)448 static void dev_watchdog(struct timer_list *t)
449 {
450 	struct net_device *dev = from_timer(dev, t, watchdog_timer);
451 
452 	netif_tx_lock(dev);
453 	if (!qdisc_tx_is_noop(dev)) {
454 		if (netif_device_present(dev) &&
455 		    netif_running(dev) &&
456 		    netif_carrier_ok(dev)) {
457 			int some_queue_timedout = 0;
458 			unsigned int i;
459 			unsigned long trans_start;
460 
461 			for (i = 0; i < dev->num_tx_queues; i++) {
462 				struct netdev_queue *txq;
463 
464 				txq = netdev_get_tx_queue(dev, i);
465 				trans_start = txq->trans_start;
466 				if (netif_xmit_stopped(txq) &&
467 				    time_after(jiffies, (trans_start +
468 							 dev->watchdog_timeo))) {
469 					some_queue_timedout = 1;
470 					txq->trans_timeout++;
471 					break;
472 				}
473 			}
474 
475 			if (some_queue_timedout) {
476 				trace_net_dev_xmit_timeout(dev, i);
477 				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
478 				       dev->name, netdev_drivername(dev), i);
479 				dev->netdev_ops->ndo_tx_timeout(dev, i);
480 			}
481 			if (!mod_timer(&dev->watchdog_timer,
482 				       round_jiffies(jiffies +
483 						     dev->watchdog_timeo)))
484 				dev_hold(dev);
485 		}
486 	}
487 	netif_tx_unlock(dev);
488 
489 	dev_put(dev);
490 }
491 
__netdev_watchdog_up(struct net_device * dev)492 void __netdev_watchdog_up(struct net_device *dev)
493 {
494 	if (dev->netdev_ops->ndo_tx_timeout) {
495 		if (dev->watchdog_timeo <= 0)
496 			dev->watchdog_timeo = 5*HZ;
497 		if (!mod_timer(&dev->watchdog_timer,
498 			       round_jiffies(jiffies + dev->watchdog_timeo)))
499 			dev_hold(dev);
500 	}
501 }
502 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
503 
dev_watchdog_up(struct net_device * dev)504 static void dev_watchdog_up(struct net_device *dev)
505 {
506 	__netdev_watchdog_up(dev);
507 }
508 
dev_watchdog_down(struct net_device * dev)509 static void dev_watchdog_down(struct net_device *dev)
510 {
511 	netif_tx_lock_bh(dev);
512 	if (del_timer(&dev->watchdog_timer))
513 		dev_put(dev);
514 	netif_tx_unlock_bh(dev);
515 }
516 
517 /**
518  *	netif_carrier_on - set carrier
519  *	@dev: network device
520  *
521  * Device has detected acquisition of carrier.
522  */
netif_carrier_on(struct net_device * dev)523 void netif_carrier_on(struct net_device *dev)
524 {
525 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
526 		if (dev->reg_state == NETREG_UNINITIALIZED)
527 			return;
528 		atomic_inc(&dev->carrier_up_count);
529 		linkwatch_fire_event(dev);
530 		if (netif_running(dev))
531 			__netdev_watchdog_up(dev);
532 	}
533 }
534 EXPORT_SYMBOL(netif_carrier_on);
535 
536 /**
537  *	netif_carrier_off - clear carrier
538  *	@dev: network device
539  *
540  * Device has detected loss of carrier.
541  */
netif_carrier_off(struct net_device * dev)542 void netif_carrier_off(struct net_device *dev)
543 {
544 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
545 		if (dev->reg_state == NETREG_UNINITIALIZED)
546 			return;
547 		atomic_inc(&dev->carrier_down_count);
548 		linkwatch_fire_event(dev);
549 	}
550 }
551 EXPORT_SYMBOL(netif_carrier_off);
552 
553 /**
554  *	netif_carrier_event - report carrier state event
555  *	@dev: network device
556  *
557  * Device has detected a carrier event but the carrier state wasn't changed.
558  * Use in drivers when querying carrier state asynchronously, to avoid missing
559  * events (link flaps) if link recovers before it's queried.
560  */
netif_carrier_event(struct net_device * dev)561 void netif_carrier_event(struct net_device *dev)
562 {
563 	if (dev->reg_state == NETREG_UNINITIALIZED)
564 		return;
565 	atomic_inc(&dev->carrier_up_count);
566 	atomic_inc(&dev->carrier_down_count);
567 	linkwatch_fire_event(dev);
568 }
569 EXPORT_SYMBOL_GPL(netif_carrier_event);
570 
571 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
572    under all circumstances. It is difficult to invent anything faster or
573    cheaper.
574  */
575 
noop_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)576 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
577 			struct sk_buff **to_free)
578 {
579 	__qdisc_drop(skb, to_free);
580 	return NET_XMIT_CN;
581 }
582 
noop_dequeue(struct Qdisc * qdisc)583 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
584 {
585 	return NULL;
586 }
587 
588 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
589 	.id		=	"noop",
590 	.priv_size	=	0,
591 	.enqueue	=	noop_enqueue,
592 	.dequeue	=	noop_dequeue,
593 	.peek		=	noop_dequeue,
594 	.owner		=	THIS_MODULE,
595 };
596 
597 static struct netdev_queue noop_netdev_queue = {
598 	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
599 	.qdisc_sleeping	=	&noop_qdisc,
600 };
601 
602 struct Qdisc noop_qdisc = {
603 	.enqueue	=	noop_enqueue,
604 	.dequeue	=	noop_dequeue,
605 	.flags		=	TCQ_F_BUILTIN,
606 	.ops		=	&noop_qdisc_ops,
607 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
608 	.dev_queue	=	&noop_netdev_queue,
609 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
610 	.gso_skb = {
611 		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
612 		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
613 		.qlen = 0,
614 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
615 	},
616 	.skb_bad_txq = {
617 		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
618 		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
619 		.qlen = 0,
620 		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
621 	},
622 };
623 EXPORT_SYMBOL(noop_qdisc);
624 
noqueue_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)625 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
626 			struct netlink_ext_ack *extack)
627 {
628 	/* register_qdisc() assigns a default of noop_enqueue if unset,
629 	 * but __dev_queue_xmit() treats noqueue only as such
630 	 * if this is NULL - so clear it here. */
631 	qdisc->enqueue = NULL;
632 	return 0;
633 }
634 
635 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
636 	.id		=	"noqueue",
637 	.priv_size	=	0,
638 	.init		=	noqueue_init,
639 	.enqueue	=	noop_enqueue,
640 	.dequeue	=	noop_dequeue,
641 	.peek		=	noop_dequeue,
642 	.owner		=	THIS_MODULE,
643 };
644 
645 static const u8 prio2band[TC_PRIO_MAX + 1] = {
646 	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
647 };
648 
649 /* 3-band FIFO queue: old style, but should be a bit faster than
650    generic prio+fifo combination.
651  */
652 
653 #define PFIFO_FAST_BANDS 3
654 
655 /*
656  * Private data for a pfifo_fast scheduler containing:
657  *	- rings for priority bands
658  */
659 struct pfifo_fast_priv {
660 	struct skb_array q[PFIFO_FAST_BANDS];
661 };
662 
band2list(struct pfifo_fast_priv * priv,int band)663 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
664 					  int band)
665 {
666 	return &priv->q[band];
667 }
668 
pfifo_fast_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)669 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
670 			      struct sk_buff **to_free)
671 {
672 	int band = prio2band[skb->priority & TC_PRIO_MAX];
673 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
674 	struct skb_array *q = band2list(priv, band);
675 	unsigned int pkt_len = qdisc_pkt_len(skb);
676 	int err;
677 
678 	err = skb_array_produce(q, skb);
679 
680 	if (unlikely(err)) {
681 		if (qdisc_is_percpu_stats(qdisc))
682 			return qdisc_drop_cpu(skb, qdisc, to_free);
683 		else
684 			return qdisc_drop(skb, qdisc, to_free);
685 	}
686 
687 	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
688 	return NET_XMIT_SUCCESS;
689 }
690 
pfifo_fast_dequeue(struct Qdisc * qdisc)691 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
692 {
693 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
694 	struct sk_buff *skb = NULL;
695 	bool need_retry = true;
696 	int band;
697 
698 retry:
699 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
700 		struct skb_array *q = band2list(priv, band);
701 
702 		if (__skb_array_empty(q))
703 			continue;
704 
705 		skb = __skb_array_consume(q);
706 	}
707 	if (likely(skb)) {
708 		qdisc_update_stats_at_dequeue(qdisc, skb);
709 	} else if (need_retry &&
710 		   READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
711 		/* Delay clearing the STATE_MISSED here to reduce
712 		 * the overhead of the second spin_trylock() in
713 		 * qdisc_run_begin() and __netif_schedule() calling
714 		 * in qdisc_run_end().
715 		 */
716 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
717 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
718 
719 		/* Make sure dequeuing happens after clearing
720 		 * STATE_MISSED.
721 		 */
722 		smp_mb__after_atomic();
723 
724 		need_retry = false;
725 
726 		goto retry;
727 	}
728 
729 	return skb;
730 }
731 
pfifo_fast_peek(struct Qdisc * qdisc)732 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
733 {
734 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
735 	struct sk_buff *skb = NULL;
736 	int band;
737 
738 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
739 		struct skb_array *q = band2list(priv, band);
740 
741 		skb = __skb_array_peek(q);
742 	}
743 
744 	return skb;
745 }
746 
pfifo_fast_reset(struct Qdisc * qdisc)747 static void pfifo_fast_reset(struct Qdisc *qdisc)
748 {
749 	int i, band;
750 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
751 
752 	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
753 		struct skb_array *q = band2list(priv, band);
754 		struct sk_buff *skb;
755 
756 		/* NULL ring is possible if destroy path is due to a failed
757 		 * skb_array_init() in pfifo_fast_init() case.
758 		 */
759 		if (!q->ring.queue)
760 			continue;
761 
762 		while ((skb = __skb_array_consume(q)) != NULL)
763 			kfree_skb(skb);
764 	}
765 
766 	if (qdisc_is_percpu_stats(qdisc)) {
767 		for_each_possible_cpu(i) {
768 			struct gnet_stats_queue *q;
769 
770 			q = per_cpu_ptr(qdisc->cpu_qstats, i);
771 			q->backlog = 0;
772 			q->qlen = 0;
773 		}
774 	}
775 }
776 
pfifo_fast_dump(struct Qdisc * qdisc,struct sk_buff * skb)777 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
778 {
779 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
780 
781 	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
782 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
783 		goto nla_put_failure;
784 	return skb->len;
785 
786 nla_put_failure:
787 	return -1;
788 }
789 
pfifo_fast_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)790 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
791 			   struct netlink_ext_ack *extack)
792 {
793 	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
794 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
795 	int prio;
796 
797 	/* guard against zero length rings */
798 	if (!qlen)
799 		return -EINVAL;
800 
801 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
802 		struct skb_array *q = band2list(priv, prio);
803 		int err;
804 
805 		err = skb_array_init(q, qlen, GFP_KERNEL);
806 		if (err)
807 			return -ENOMEM;
808 	}
809 
810 	/* Can by-pass the queue discipline */
811 	qdisc->flags |= TCQ_F_CAN_BYPASS;
812 	return 0;
813 }
814 
pfifo_fast_destroy(struct Qdisc * sch)815 static void pfifo_fast_destroy(struct Qdisc *sch)
816 {
817 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
818 	int prio;
819 
820 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
821 		struct skb_array *q = band2list(priv, prio);
822 
823 		/* NULL ring is possible if destroy path is due to a failed
824 		 * skb_array_init() in pfifo_fast_init() case.
825 		 */
826 		if (!q->ring.queue)
827 			continue;
828 		/* Destroy ring but no need to kfree_skb because a call to
829 		 * pfifo_fast_reset() has already done that work.
830 		 */
831 		ptr_ring_cleanup(&q->ring, NULL);
832 	}
833 }
834 
pfifo_fast_change_tx_queue_len(struct Qdisc * sch,unsigned int new_len)835 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
836 					  unsigned int new_len)
837 {
838 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
839 	struct skb_array *bands[PFIFO_FAST_BANDS];
840 	int prio;
841 
842 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
843 		struct skb_array *q = band2list(priv, prio);
844 
845 		bands[prio] = q;
846 	}
847 
848 	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
849 					 GFP_KERNEL);
850 }
851 
852 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
853 	.id		=	"pfifo_fast",
854 	.priv_size	=	sizeof(struct pfifo_fast_priv),
855 	.enqueue	=	pfifo_fast_enqueue,
856 	.dequeue	=	pfifo_fast_dequeue,
857 	.peek		=	pfifo_fast_peek,
858 	.init		=	pfifo_fast_init,
859 	.destroy	=	pfifo_fast_destroy,
860 	.reset		=	pfifo_fast_reset,
861 	.dump		=	pfifo_fast_dump,
862 	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
863 	.owner		=	THIS_MODULE,
864 	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
865 };
866 EXPORT_SYMBOL(pfifo_fast_ops);
867 
868 static struct lock_class_key qdisc_tx_busylock;
869 
qdisc_alloc(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,struct netlink_ext_ack * extack)870 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
871 			  const struct Qdisc_ops *ops,
872 			  struct netlink_ext_ack *extack)
873 {
874 	struct Qdisc *sch;
875 	unsigned int size = sizeof(*sch) + ops->priv_size;
876 	int err = -ENOBUFS;
877 	struct net_device *dev;
878 
879 	if (!dev_queue) {
880 		NL_SET_ERR_MSG(extack, "No device queue given");
881 		err = -EINVAL;
882 		goto errout;
883 	}
884 
885 	dev = dev_queue->dev;
886 	sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
887 
888 	if (!sch)
889 		goto errout;
890 	__skb_queue_head_init(&sch->gso_skb);
891 	__skb_queue_head_init(&sch->skb_bad_txq);
892 	qdisc_skb_head_init(&sch->q);
893 	gnet_stats_basic_sync_init(&sch->bstats);
894 	spin_lock_init(&sch->q.lock);
895 
896 	if (ops->static_flags & TCQ_F_CPUSTATS) {
897 		sch->cpu_bstats =
898 			netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
899 		if (!sch->cpu_bstats)
900 			goto errout1;
901 
902 		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
903 		if (!sch->cpu_qstats) {
904 			free_percpu(sch->cpu_bstats);
905 			goto errout1;
906 		}
907 	}
908 
909 	spin_lock_init(&sch->busylock);
910 	lockdep_set_class(&sch->busylock,
911 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
912 
913 	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
914 	spin_lock_init(&sch->seqlock);
915 	lockdep_set_class(&sch->seqlock,
916 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
917 
918 	sch->ops = ops;
919 	sch->flags = ops->static_flags;
920 	sch->enqueue = ops->enqueue;
921 	sch->dequeue = ops->dequeue;
922 	sch->dev_queue = dev_queue;
923 	dev_hold(dev);
924 	refcount_set(&sch->refcnt, 1);
925 
926 	return sch;
927 errout1:
928 	kfree(sch);
929 errout:
930 	return ERR_PTR(err);
931 }
932 
qdisc_create_dflt(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,unsigned int parentid,struct netlink_ext_ack * extack)933 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
934 				const struct Qdisc_ops *ops,
935 				unsigned int parentid,
936 				struct netlink_ext_ack *extack)
937 {
938 	struct Qdisc *sch;
939 
940 	if (!try_module_get(ops->owner)) {
941 		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
942 		return NULL;
943 	}
944 
945 	sch = qdisc_alloc(dev_queue, ops, extack);
946 	if (IS_ERR(sch)) {
947 		module_put(ops->owner);
948 		return NULL;
949 	}
950 	sch->parent = parentid;
951 
952 	if (!ops->init || ops->init(sch, NULL, extack) == 0) {
953 		trace_qdisc_create(ops, dev_queue->dev, parentid);
954 		return sch;
955 	}
956 
957 	qdisc_put(sch);
958 	return NULL;
959 }
960 EXPORT_SYMBOL(qdisc_create_dflt);
961 
962 /* Under qdisc_lock(qdisc) and BH! */
963 
qdisc_reset(struct Qdisc * qdisc)964 void qdisc_reset(struct Qdisc *qdisc)
965 {
966 	const struct Qdisc_ops *ops = qdisc->ops;
967 	struct sk_buff *skb, *tmp;
968 
969 	trace_qdisc_reset(qdisc);
970 
971 	if (ops->reset)
972 		ops->reset(qdisc);
973 
974 	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
975 		__skb_unlink(skb, &qdisc->gso_skb);
976 		kfree_skb_list(skb);
977 	}
978 
979 	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
980 		__skb_unlink(skb, &qdisc->skb_bad_txq);
981 		kfree_skb_list(skb);
982 	}
983 
984 	qdisc->q.qlen = 0;
985 	qdisc->qstats.backlog = 0;
986 }
987 EXPORT_SYMBOL(qdisc_reset);
988 
qdisc_free(struct Qdisc * qdisc)989 void qdisc_free(struct Qdisc *qdisc)
990 {
991 	if (qdisc_is_percpu_stats(qdisc)) {
992 		free_percpu(qdisc->cpu_bstats);
993 		free_percpu(qdisc->cpu_qstats);
994 	}
995 
996 	kfree(qdisc);
997 }
998 
qdisc_free_cb(struct rcu_head * head)999 static void qdisc_free_cb(struct rcu_head *head)
1000 {
1001 	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1002 
1003 	qdisc_free(q);
1004 }
1005 
qdisc_destroy(struct Qdisc * qdisc)1006 static void qdisc_destroy(struct Qdisc *qdisc)
1007 {
1008 	const struct Qdisc_ops  *ops = qdisc->ops;
1009 
1010 #ifdef CONFIG_NET_SCHED
1011 	qdisc_hash_del(qdisc);
1012 
1013 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
1014 #endif
1015 	gen_kill_estimator(&qdisc->rate_est);
1016 
1017 	qdisc_reset(qdisc);
1018 
1019 	if (ops->destroy)
1020 		ops->destroy(qdisc);
1021 
1022 	module_put(ops->owner);
1023 	dev_put(qdisc_dev(qdisc));
1024 
1025 	trace_qdisc_destroy(qdisc);
1026 
1027 	call_rcu(&qdisc->rcu, qdisc_free_cb);
1028 }
1029 
qdisc_put(struct Qdisc * qdisc)1030 void qdisc_put(struct Qdisc *qdisc)
1031 {
1032 	if (!qdisc)
1033 		return;
1034 
1035 	if (qdisc->flags & TCQ_F_BUILTIN ||
1036 	    !refcount_dec_and_test(&qdisc->refcnt))
1037 		return;
1038 
1039 	qdisc_destroy(qdisc);
1040 }
1041 EXPORT_SYMBOL(qdisc_put);
1042 
1043 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1044  * Intended to be used as optimization, this function only takes rtnl lock if
1045  * qdisc reference counter reached zero.
1046  */
1047 
qdisc_put_unlocked(struct Qdisc * qdisc)1048 void qdisc_put_unlocked(struct Qdisc *qdisc)
1049 {
1050 	if (qdisc->flags & TCQ_F_BUILTIN ||
1051 	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1052 		return;
1053 
1054 	qdisc_destroy(qdisc);
1055 	rtnl_unlock();
1056 }
1057 EXPORT_SYMBOL(qdisc_put_unlocked);
1058 
1059 /* Attach toplevel qdisc to device queue. */
dev_graft_qdisc(struct netdev_queue * dev_queue,struct Qdisc * qdisc)1060 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1061 			      struct Qdisc *qdisc)
1062 {
1063 	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1064 	spinlock_t *root_lock;
1065 
1066 	root_lock = qdisc_lock(oqdisc);
1067 	spin_lock_bh(root_lock);
1068 
1069 	/* ... and graft new one */
1070 	if (qdisc == NULL)
1071 		qdisc = &noop_qdisc;
1072 	dev_queue->qdisc_sleeping = qdisc;
1073 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1074 
1075 	spin_unlock_bh(root_lock);
1076 
1077 	return oqdisc;
1078 }
1079 EXPORT_SYMBOL(dev_graft_qdisc);
1080 
attach_one_default_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1081 static void attach_one_default_qdisc(struct net_device *dev,
1082 				     struct netdev_queue *dev_queue,
1083 				     void *_unused)
1084 {
1085 	struct Qdisc *qdisc;
1086 	const struct Qdisc_ops *ops = default_qdisc_ops;
1087 
1088 	if (dev->priv_flags & IFF_NO_QUEUE)
1089 		ops = &noqueue_qdisc_ops;
1090 	else if(dev->type == ARPHRD_CAN)
1091 		ops = &pfifo_fast_ops;
1092 
1093 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1094 	if (!qdisc)
1095 		return;
1096 
1097 	if (!netif_is_multiqueue(dev))
1098 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1099 	dev_queue->qdisc_sleeping = qdisc;
1100 }
1101 
attach_default_qdiscs(struct net_device * dev)1102 static void attach_default_qdiscs(struct net_device *dev)
1103 {
1104 	struct netdev_queue *txq;
1105 	struct Qdisc *qdisc;
1106 
1107 	txq = netdev_get_tx_queue(dev, 0);
1108 
1109 	if (!netif_is_multiqueue(dev) ||
1110 	    dev->priv_flags & IFF_NO_QUEUE) {
1111 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1112 		dev->qdisc = txq->qdisc_sleeping;
1113 		qdisc_refcount_inc(dev->qdisc);
1114 	} else {
1115 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1116 		if (qdisc) {
1117 			dev->qdisc = qdisc;
1118 			qdisc->ops->attach(qdisc);
1119 		}
1120 	}
1121 
1122 	/* Detect default qdisc setup/init failed and fallback to "noqueue" */
1123 	if (dev->qdisc == &noop_qdisc) {
1124 		netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1125 			    default_qdisc_ops->id, noqueue_qdisc_ops.id);
1126 		dev->priv_flags |= IFF_NO_QUEUE;
1127 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1128 		dev->qdisc = txq->qdisc_sleeping;
1129 		qdisc_refcount_inc(dev->qdisc);
1130 		dev->priv_flags ^= IFF_NO_QUEUE;
1131 	}
1132 
1133 #ifdef CONFIG_NET_SCHED
1134 	if (dev->qdisc != &noop_qdisc)
1135 		qdisc_hash_add(dev->qdisc, false);
1136 #endif
1137 }
1138 
transition_one_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _need_watchdog)1139 static void transition_one_qdisc(struct net_device *dev,
1140 				 struct netdev_queue *dev_queue,
1141 				 void *_need_watchdog)
1142 {
1143 	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1144 	int *need_watchdog_p = _need_watchdog;
1145 
1146 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1147 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1148 
1149 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1150 	if (need_watchdog_p) {
1151 		dev_queue->trans_start = 0;
1152 		*need_watchdog_p = 1;
1153 	}
1154 }
1155 
dev_activate(struct net_device * dev)1156 void dev_activate(struct net_device *dev)
1157 {
1158 	int need_watchdog;
1159 
1160 	/* No queueing discipline is attached to device;
1161 	 * create default one for devices, which need queueing
1162 	 * and noqueue_qdisc for virtual interfaces
1163 	 */
1164 
1165 	if (dev->qdisc == &noop_qdisc)
1166 		attach_default_qdiscs(dev);
1167 
1168 	if (!netif_carrier_ok(dev))
1169 		/* Delay activation until next carrier-on event */
1170 		return;
1171 
1172 	need_watchdog = 0;
1173 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1174 	if (dev_ingress_queue(dev))
1175 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1176 
1177 	if (need_watchdog) {
1178 		netif_trans_update(dev);
1179 		dev_watchdog_up(dev);
1180 	}
1181 }
1182 EXPORT_SYMBOL(dev_activate);
1183 
qdisc_deactivate(struct Qdisc * qdisc)1184 static void qdisc_deactivate(struct Qdisc *qdisc)
1185 {
1186 	if (qdisc->flags & TCQ_F_BUILTIN)
1187 		return;
1188 
1189 	set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1190 }
1191 
dev_deactivate_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1192 static void dev_deactivate_queue(struct net_device *dev,
1193 				 struct netdev_queue *dev_queue,
1194 				 void *_qdisc_default)
1195 {
1196 	struct Qdisc *qdisc_default = _qdisc_default;
1197 	struct Qdisc *qdisc;
1198 
1199 	qdisc = rtnl_dereference(dev_queue->qdisc);
1200 	if (qdisc) {
1201 		qdisc_deactivate(qdisc);
1202 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1203 	}
1204 }
1205 
dev_reset_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1206 static void dev_reset_queue(struct net_device *dev,
1207 			    struct netdev_queue *dev_queue,
1208 			    void *_unused)
1209 {
1210 	struct Qdisc *qdisc;
1211 	bool nolock;
1212 
1213 	qdisc = dev_queue->qdisc_sleeping;
1214 	if (!qdisc)
1215 		return;
1216 
1217 	nolock = qdisc->flags & TCQ_F_NOLOCK;
1218 
1219 	if (nolock)
1220 		spin_lock_bh(&qdisc->seqlock);
1221 	spin_lock_bh(qdisc_lock(qdisc));
1222 
1223 	qdisc_reset(qdisc);
1224 
1225 	spin_unlock_bh(qdisc_lock(qdisc));
1226 	if (nolock) {
1227 		clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1228 		clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1229 		spin_unlock_bh(&qdisc->seqlock);
1230 	}
1231 }
1232 
some_qdisc_is_busy(struct net_device * dev)1233 static bool some_qdisc_is_busy(struct net_device *dev)
1234 {
1235 	unsigned int i;
1236 
1237 	for (i = 0; i < dev->num_tx_queues; i++) {
1238 		struct netdev_queue *dev_queue;
1239 		spinlock_t *root_lock;
1240 		struct Qdisc *q;
1241 		int val;
1242 
1243 		dev_queue = netdev_get_tx_queue(dev, i);
1244 		q = dev_queue->qdisc_sleeping;
1245 
1246 		root_lock = qdisc_lock(q);
1247 		spin_lock_bh(root_lock);
1248 
1249 		val = (qdisc_is_running(q) ||
1250 		       test_bit(__QDISC_STATE_SCHED, &q->state));
1251 
1252 		spin_unlock_bh(root_lock);
1253 
1254 		if (val)
1255 			return true;
1256 	}
1257 	return false;
1258 }
1259 
1260 /**
1261  * 	dev_deactivate_many - deactivate transmissions on several devices
1262  * 	@head: list of devices to deactivate
1263  *
1264  *	This function returns only when all outstanding transmissions
1265  *	have completed, unless all devices are in dismantle phase.
1266  */
dev_deactivate_many(struct list_head * head)1267 void dev_deactivate_many(struct list_head *head)
1268 {
1269 	struct net_device *dev;
1270 
1271 	list_for_each_entry(dev, head, close_list) {
1272 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1273 					 &noop_qdisc);
1274 		if (dev_ingress_queue(dev))
1275 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1276 					     &noop_qdisc);
1277 
1278 		dev_watchdog_down(dev);
1279 	}
1280 
1281 	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
1282 	 * outstanding qdisc enqueuing calls.
1283 	 * This is avoided if all devices are in dismantle phase :
1284 	 * Caller will call synchronize_net() for us
1285 	 */
1286 	synchronize_net();
1287 
1288 	list_for_each_entry(dev, head, close_list) {
1289 		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1290 
1291 		if (dev_ingress_queue(dev))
1292 			dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1293 	}
1294 
1295 	/* Wait for outstanding qdisc_run calls. */
1296 	list_for_each_entry(dev, head, close_list) {
1297 		while (some_qdisc_is_busy(dev)) {
1298 			/* wait_event() would avoid this sleep-loop but would
1299 			 * require expensive checks in the fast paths of packet
1300 			 * processing which isn't worth it.
1301 			 */
1302 			schedule_timeout_uninterruptible(1);
1303 		}
1304 	}
1305 }
1306 
dev_deactivate(struct net_device * dev)1307 void dev_deactivate(struct net_device *dev)
1308 {
1309 	LIST_HEAD(single);
1310 
1311 	list_add(&dev->close_list, &single);
1312 	dev_deactivate_many(&single);
1313 	list_del(&single);
1314 }
1315 EXPORT_SYMBOL(dev_deactivate);
1316 
qdisc_change_tx_queue_len(struct net_device * dev,struct netdev_queue * dev_queue)1317 static int qdisc_change_tx_queue_len(struct net_device *dev,
1318 				     struct netdev_queue *dev_queue)
1319 {
1320 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1321 	const struct Qdisc_ops *ops = qdisc->ops;
1322 
1323 	if (ops->change_tx_queue_len)
1324 		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1325 	return 0;
1326 }
1327 
dev_qdisc_change_real_num_tx(struct net_device * dev,unsigned int new_real_tx)1328 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1329 				  unsigned int new_real_tx)
1330 {
1331 	struct Qdisc *qdisc = dev->qdisc;
1332 
1333 	if (qdisc->ops->change_real_num_tx)
1334 		qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1335 }
1336 
mq_change_real_num_tx(struct Qdisc * sch,unsigned int new_real_tx)1337 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1338 {
1339 #ifdef CONFIG_NET_SCHED
1340 	struct net_device *dev = qdisc_dev(sch);
1341 	struct Qdisc *qdisc;
1342 	unsigned int i;
1343 
1344 	for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1345 		qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1346 		/* Only update the default qdiscs we created,
1347 		 * qdiscs with handles are always hashed.
1348 		 */
1349 		if (qdisc != &noop_qdisc && !qdisc->handle)
1350 			qdisc_hash_del(qdisc);
1351 	}
1352 	for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1353 		qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1354 		if (qdisc != &noop_qdisc && !qdisc->handle)
1355 			qdisc_hash_add(qdisc, false);
1356 	}
1357 #endif
1358 }
1359 EXPORT_SYMBOL(mq_change_real_num_tx);
1360 
dev_qdisc_change_tx_queue_len(struct net_device * dev)1361 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1362 {
1363 	bool up = dev->flags & IFF_UP;
1364 	unsigned int i;
1365 	int ret = 0;
1366 
1367 	if (up)
1368 		dev_deactivate(dev);
1369 
1370 	for (i = 0; i < dev->num_tx_queues; i++) {
1371 		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1372 
1373 		/* TODO: revert changes on a partial failure */
1374 		if (ret)
1375 			break;
1376 	}
1377 
1378 	if (up)
1379 		dev_activate(dev);
1380 	return ret;
1381 }
1382 
dev_init_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc)1383 static void dev_init_scheduler_queue(struct net_device *dev,
1384 				     struct netdev_queue *dev_queue,
1385 				     void *_qdisc)
1386 {
1387 	struct Qdisc *qdisc = _qdisc;
1388 
1389 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1390 	dev_queue->qdisc_sleeping = qdisc;
1391 }
1392 
dev_init_scheduler(struct net_device * dev)1393 void dev_init_scheduler(struct net_device *dev)
1394 {
1395 	dev->qdisc = &noop_qdisc;
1396 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1397 	if (dev_ingress_queue(dev))
1398 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1399 
1400 	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1401 }
1402 
shutdown_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1403 static void shutdown_scheduler_queue(struct net_device *dev,
1404 				     struct netdev_queue *dev_queue,
1405 				     void *_qdisc_default)
1406 {
1407 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1408 	struct Qdisc *qdisc_default = _qdisc_default;
1409 
1410 	if (qdisc) {
1411 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1412 		dev_queue->qdisc_sleeping = qdisc_default;
1413 
1414 		qdisc_put(qdisc);
1415 	}
1416 }
1417 
dev_shutdown(struct net_device * dev)1418 void dev_shutdown(struct net_device *dev)
1419 {
1420 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1421 	if (dev_ingress_queue(dev))
1422 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1423 	qdisc_put(dev->qdisc);
1424 	dev->qdisc = &noop_qdisc;
1425 
1426 	WARN_ON(timer_pending(&dev->watchdog_timer));
1427 }
1428 
1429 /**
1430  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1431  * @rate:   Rate to compute reciprocal division values of
1432  * @mult:   Multiplier for reciprocal division
1433  * @shift:  Shift for reciprocal division
1434  *
1435  * The multiplier and shift for reciprocal division by rate are stored
1436  * in mult and shift.
1437  *
1438  * The deal here is to replace a divide by a reciprocal one
1439  * in fast path (a reciprocal divide is a multiply and a shift)
1440  *
1441  * Normal formula would be :
1442  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1443  *
1444  * We compute mult/shift to use instead :
1445  *  time_in_ns = (len * mult) >> shift;
1446  *
1447  * We try to get the highest possible mult value for accuracy,
1448  * but have to make sure no overflows will ever happen.
1449  *
1450  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1451  */
psched_ratecfg_precompute__(u64 rate,u32 * mult,u8 * shift)1452 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1453 {
1454 	u64 factor = NSEC_PER_SEC;
1455 
1456 	*mult = 1;
1457 	*shift = 0;
1458 
1459 	if (rate <= 0)
1460 		return;
1461 
1462 	for (;;) {
1463 		*mult = div64_u64(factor, rate);
1464 		if (*mult & (1U << 31) || factor & (1ULL << 63))
1465 			break;
1466 		factor <<= 1;
1467 		(*shift)++;
1468 	}
1469 }
1470 
psched_ratecfg_precompute(struct psched_ratecfg * r,const struct tc_ratespec * conf,u64 rate64)1471 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1472 			       const struct tc_ratespec *conf,
1473 			       u64 rate64)
1474 {
1475 	memset(r, 0, sizeof(*r));
1476 	r->overhead = conf->overhead;
1477 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1478 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1479 	psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1480 }
1481 EXPORT_SYMBOL(psched_ratecfg_precompute);
1482 
psched_ppscfg_precompute(struct psched_pktrate * r,u64 pktrate64)1483 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1484 {
1485 	r->rate_pkts_ps = pktrate64;
1486 	psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1487 }
1488 EXPORT_SYMBOL(psched_ppscfg_precompute);
1489 
mini_qdisc_pair_swap(struct mini_Qdisc_pair * miniqp,struct tcf_proto * tp_head)1490 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1491 			  struct tcf_proto *tp_head)
1492 {
1493 	/* Protected with chain0->filter_chain_lock.
1494 	 * Can't access chain directly because tp_head can be NULL.
1495 	 */
1496 	struct mini_Qdisc *miniq_old =
1497 		rcu_dereference_protected(*miniqp->p_miniq, 1);
1498 	struct mini_Qdisc *miniq;
1499 
1500 	if (!tp_head) {
1501 		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1502 	} else {
1503 		miniq = miniq_old != &miniqp->miniq1 ?
1504 			&miniqp->miniq1 : &miniqp->miniq2;
1505 
1506 		/* We need to make sure that readers won't see the miniq
1507 		 * we are about to modify. So ensure that at least one RCU
1508 		 * grace period has elapsed since the miniq was made
1509 		 * inactive.
1510 		 */
1511 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
1512 			cond_synchronize_rcu(miniq->rcu_state);
1513 		else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1514 			synchronize_rcu_expedited();
1515 
1516 		miniq->filter_list = tp_head;
1517 		rcu_assign_pointer(*miniqp->p_miniq, miniq);
1518 	}
1519 
1520 	if (miniq_old)
1521 		/* This is counterpart of the rcu sync above. We need to
1522 		 * block potential new user of miniq_old until all readers
1523 		 * are not seeing it.
1524 		 */
1525 		miniq_old->rcu_state = start_poll_synchronize_rcu();
1526 }
1527 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1528 
mini_qdisc_pair_block_init(struct mini_Qdisc_pair * miniqp,struct tcf_block * block)1529 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1530 				struct tcf_block *block)
1531 {
1532 	miniqp->miniq1.block = block;
1533 	miniqp->miniq2.block = block;
1534 }
1535 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1536 
mini_qdisc_pair_init(struct mini_Qdisc_pair * miniqp,struct Qdisc * qdisc,struct mini_Qdisc __rcu ** p_miniq)1537 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1538 			  struct mini_Qdisc __rcu **p_miniq)
1539 {
1540 	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1541 	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1542 	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1543 	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1544 	miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1545 	miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1546 	miniqp->p_miniq = p_miniq;
1547 }
1548 EXPORT_SYMBOL(mini_qdisc_pair_init);
1549