Loading...
1/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/sched.h>
19#include <linux/string.h>
20#include <linux/errno.h>
21#include <linux/netdevice.h>
22#include <linux/skbuff.h>
23#include <linux/rtnetlink.h>
24#include <linux/init.h>
25#include <linux/rcupdate.h>
26#include <linux/list.h>
27#include <linux/slab.h>
28#include <linux/if_vlan.h>
29#include <net/sch_generic.h>
30#include <net/pkt_sched.h>
31#include <net/dst.h>
32
33/* Qdisc to use by default */
34const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35EXPORT_SYMBOL(default_qdisc_ops);
36
37/* Main transmission queue. */
38
39/* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
41 *
42 * The idea is the following:
43 * - enqueue, dequeue are serialized via qdisc root lock
44 * - ingress filtering is also serialized via qdisc root lock
45 * - updates to tree and tree walking are only done under the rtnl mutex.
46 */
47
48static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
49{
50 q->gso_skb = skb;
51 q->qstats.requeues++;
52 q->q.qlen++; /* it's still part of the queue */
53 __netif_schedule(q);
54
55 return 0;
56}
57
58static void try_bulk_dequeue_skb(struct Qdisc *q,
59 struct sk_buff *skb,
60 const struct netdev_queue *txq,
61 int *packets)
62{
63 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
64
65 while (bytelimit > 0) {
66 struct sk_buff *nskb = q->dequeue(q);
67
68 if (!nskb)
69 break;
70
71 bytelimit -= nskb->len; /* covers GSO len */
72 skb->next = nskb;
73 skb = nskb;
74 (*packets)++; /* GSO counts as one pkt */
75 }
76 skb->next = NULL;
77}
78
79/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
80 * A requeued skb (via q->gso_skb) can also be a SKB list.
81 */
82static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
83 int *packets)
84{
85 struct sk_buff *skb = q->gso_skb;
86 const struct netdev_queue *txq = q->dev_queue;
87
88 *packets = 1;
89 *validate = true;
90 if (unlikely(skb)) {
91 /* check the reason of requeuing without tx lock first */
92 txq = skb_get_tx_queue(txq->dev, skb);
93 if (!netif_xmit_frozen_or_stopped(txq)) {
94 q->gso_skb = NULL;
95 q->q.qlen--;
96 } else
97 skb = NULL;
98 /* skb in gso_skb were already validated */
99 *validate = false;
100 } else {
101 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
102 !netif_xmit_frozen_or_stopped(txq)) {
103 skb = q->dequeue(q);
104 if (skb && qdisc_may_bulk(q))
105 try_bulk_dequeue_skb(q, skb, txq, packets);
106 }
107 }
108 return skb;
109}
110
111static inline int handle_dev_cpu_collision(struct sk_buff *skb,
112 struct netdev_queue *dev_queue,
113 struct Qdisc *q)
114{
115 int ret;
116
117 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
118 /*
119 * Same CPU holding the lock. It may be a transient
120 * configuration error, when hard_start_xmit() recurses. We
121 * detect it by checking xmit owner and drop the packet when
122 * deadloop is detected. Return OK to try the next skb.
123 */
124 kfree_skb_list(skb);
125 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
126 dev_queue->dev->name);
127 ret = qdisc_qlen(q);
128 } else {
129 /*
130 * Another cpu is holding lock, requeue & delay xmits for
131 * some time.
132 */
133 __this_cpu_inc(softnet_data.cpu_collision);
134 ret = dev_requeue_skb(skb, q);
135 }
136
137 return ret;
138}
139
140/*
141 * Transmit possibly several skbs, and handle the return status as
142 * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
143 * only one CPU can execute this function.
144 *
145 * Returns to the caller:
146 * 0 - queue is empty or throttled.
147 * >0 - queue is not empty.
148 */
149int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
150 struct net_device *dev, struct netdev_queue *txq,
151 spinlock_t *root_lock, bool validate)
152{
153 int ret = NETDEV_TX_BUSY;
154
155 /* And release qdisc */
156 spin_unlock(root_lock);
157
158 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
159 if (validate)
160 skb = validate_xmit_skb_list(skb, dev);
161
162 if (likely(skb)) {
163 HARD_TX_LOCK(dev, txq, smp_processor_id());
164 if (!netif_xmit_frozen_or_stopped(txq))
165 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
166
167 HARD_TX_UNLOCK(dev, txq);
168 } else {
169 spin_lock(root_lock);
170 return qdisc_qlen(q);
171 }
172 spin_lock(root_lock);
173
174 if (dev_xmit_complete(ret)) {
175 /* Driver sent out skb successfully or skb was consumed */
176 ret = qdisc_qlen(q);
177 } else if (ret == NETDEV_TX_LOCKED) {
178 /* Driver try lock failed */
179 ret = handle_dev_cpu_collision(skb, txq, q);
180 } else {
181 /* Driver returned NETDEV_TX_BUSY - requeue skb */
182 if (unlikely(ret != NETDEV_TX_BUSY))
183 net_warn_ratelimited("BUG %s code %d qlen %d\n",
184 dev->name, ret, q->q.qlen);
185
186 ret = dev_requeue_skb(skb, q);
187 }
188
189 if (ret && netif_xmit_frozen_or_stopped(txq))
190 ret = 0;
191
192 return ret;
193}
194
195/*
196 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
197 *
198 * __QDISC___STATE_RUNNING guarantees only one CPU can process
199 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
200 * this queue.
201 *
202 * netif_tx_lock serializes accesses to device driver.
203 *
204 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
205 * if one is grabbed, another must be free.
206 *
207 * Note, that this procedure can be called by a watchdog timer
208 *
209 * Returns to the caller:
210 * 0 - queue is empty or throttled.
211 * >0 - queue is not empty.
212 *
213 */
214static inline int qdisc_restart(struct Qdisc *q, int *packets)
215{
216 struct netdev_queue *txq;
217 struct net_device *dev;
218 spinlock_t *root_lock;
219 struct sk_buff *skb;
220 bool validate;
221
222 /* Dequeue packet */
223 skb = dequeue_skb(q, &validate, packets);
224 if (unlikely(!skb))
225 return 0;
226
227 root_lock = qdisc_lock(q);
228 dev = qdisc_dev(q);
229 txq = skb_get_tx_queue(dev, skb);
230
231 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
232}
233
234void __qdisc_run(struct Qdisc *q)
235{
236 int quota = weight_p;
237 int packets;
238
239 while (qdisc_restart(q, &packets)) {
240 /*
241 * Ordered by possible occurrence: Postpone processing if
242 * 1. we've exceeded packet quota
243 * 2. another process needs the CPU;
244 */
245 quota -= packets;
246 if (quota <= 0 || need_resched()) {
247 __netif_schedule(q);
248 break;
249 }
250 }
251
252 qdisc_run_end(q);
253}
254
255unsigned long dev_trans_start(struct net_device *dev)
256{
257 unsigned long val, res;
258 unsigned int i;
259
260 if (is_vlan_dev(dev))
261 dev = vlan_dev_real_dev(dev);
262 res = dev->trans_start;
263 for (i = 0; i < dev->num_tx_queues; i++) {
264 val = netdev_get_tx_queue(dev, i)->trans_start;
265 if (val && time_after(val, res))
266 res = val;
267 }
268 dev->trans_start = res;
269
270 return res;
271}
272EXPORT_SYMBOL(dev_trans_start);
273
274static void dev_watchdog(unsigned long arg)
275{
276 struct net_device *dev = (struct net_device *)arg;
277
278 netif_tx_lock(dev);
279 if (!qdisc_tx_is_noop(dev)) {
280 if (netif_device_present(dev) &&
281 netif_running(dev) &&
282 netif_carrier_ok(dev)) {
283 int some_queue_timedout = 0;
284 unsigned int i;
285 unsigned long trans_start;
286
287 for (i = 0; i < dev->num_tx_queues; i++) {
288 struct netdev_queue *txq;
289
290 txq = netdev_get_tx_queue(dev, i);
291 /*
292 * old device drivers set dev->trans_start
293 */
294 trans_start = txq->trans_start ? : dev->trans_start;
295 if (netif_xmit_stopped(txq) &&
296 time_after(jiffies, (trans_start +
297 dev->watchdog_timeo))) {
298 some_queue_timedout = 1;
299 txq->trans_timeout++;
300 break;
301 }
302 }
303
304 if (some_queue_timedout) {
305 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
306 dev->name, netdev_drivername(dev), i);
307 dev->netdev_ops->ndo_tx_timeout(dev);
308 }
309 if (!mod_timer(&dev->watchdog_timer,
310 round_jiffies(jiffies +
311 dev->watchdog_timeo)))
312 dev_hold(dev);
313 }
314 }
315 netif_tx_unlock(dev);
316
317 dev_put(dev);
318}
319
320void __netdev_watchdog_up(struct net_device *dev)
321{
322 if (dev->netdev_ops->ndo_tx_timeout) {
323 if (dev->watchdog_timeo <= 0)
324 dev->watchdog_timeo = 5*HZ;
325 if (!mod_timer(&dev->watchdog_timer,
326 round_jiffies(jiffies + dev->watchdog_timeo)))
327 dev_hold(dev);
328 }
329}
330
331static void dev_watchdog_up(struct net_device *dev)
332{
333 __netdev_watchdog_up(dev);
334}
335
336static void dev_watchdog_down(struct net_device *dev)
337{
338 netif_tx_lock_bh(dev);
339 if (del_timer(&dev->watchdog_timer))
340 dev_put(dev);
341 netif_tx_unlock_bh(dev);
342}
343
344/**
345 * netif_carrier_on - set carrier
346 * @dev: network device
347 *
348 * Device has detected that carrier.
349 */
350void netif_carrier_on(struct net_device *dev)
351{
352 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
353 if (dev->reg_state == NETREG_UNINITIALIZED)
354 return;
355 atomic_inc(&dev->carrier_changes);
356 linkwatch_fire_event(dev);
357 if (netif_running(dev))
358 __netdev_watchdog_up(dev);
359 }
360}
361EXPORT_SYMBOL(netif_carrier_on);
362
363/**
364 * netif_carrier_off - clear carrier
365 * @dev: network device
366 *
367 * Device has detected loss of carrier.
368 */
369void netif_carrier_off(struct net_device *dev)
370{
371 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
372 if (dev->reg_state == NETREG_UNINITIALIZED)
373 return;
374 atomic_inc(&dev->carrier_changes);
375 linkwatch_fire_event(dev);
376 }
377}
378EXPORT_SYMBOL(netif_carrier_off);
379
380/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
381 under all circumstances. It is difficult to invent anything faster or
382 cheaper.
383 */
384
385static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
386{
387 kfree_skb(skb);
388 return NET_XMIT_CN;
389}
390
391static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
392{
393 return NULL;
394}
395
396struct Qdisc_ops noop_qdisc_ops __read_mostly = {
397 .id = "noop",
398 .priv_size = 0,
399 .enqueue = noop_enqueue,
400 .dequeue = noop_dequeue,
401 .peek = noop_dequeue,
402 .owner = THIS_MODULE,
403};
404
405static struct netdev_queue noop_netdev_queue = {
406 .qdisc = &noop_qdisc,
407 .qdisc_sleeping = &noop_qdisc,
408};
409
410struct Qdisc noop_qdisc = {
411 .enqueue = noop_enqueue,
412 .dequeue = noop_dequeue,
413 .flags = TCQ_F_BUILTIN,
414 .ops = &noop_qdisc_ops,
415 .list = LIST_HEAD_INIT(noop_qdisc.list),
416 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
417 .dev_queue = &noop_netdev_queue,
418 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
419};
420EXPORT_SYMBOL(noop_qdisc);
421
422static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
423{
424 /* register_qdisc() assigns a default of noop_enqueue if unset,
425 * but __dev_queue_xmit() treats noqueue only as such
426 * if this is NULL - so clear it here. */
427 qdisc->enqueue = NULL;
428 return 0;
429}
430
431struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
432 .id = "noqueue",
433 .priv_size = 0,
434 .init = noqueue_init,
435 .enqueue = noop_enqueue,
436 .dequeue = noop_dequeue,
437 .peek = noop_dequeue,
438 .owner = THIS_MODULE,
439};
440
441static const u8 prio2band[TC_PRIO_MAX + 1] = {
442 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
443};
444
445/* 3-band FIFO queue: old style, but should be a bit faster than
446 generic prio+fifo combination.
447 */
448
449#define PFIFO_FAST_BANDS 3
450
451/*
452 * Private data for a pfifo_fast scheduler containing:
453 * - queues for the three band
454 * - bitmap indicating which of the bands contain skbs
455 */
456struct pfifo_fast_priv {
457 u32 bitmap;
458 struct sk_buff_head q[PFIFO_FAST_BANDS];
459};
460
461/*
462 * Convert a bitmap to the first band number where an skb is queued, where:
463 * bitmap=0 means there are no skbs on any band.
464 * bitmap=1 means there is an skb on band 0.
465 * bitmap=7 means there are skbs on all 3 bands, etc.
466 */
467static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
468
469static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
470 int band)
471{
472 return priv->q + band;
473}
474
475static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
476{
477 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
478 int band = prio2band[skb->priority & TC_PRIO_MAX];
479 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
480 struct sk_buff_head *list = band2list(priv, band);
481
482 priv->bitmap |= (1 << band);
483 qdisc->q.qlen++;
484 return __qdisc_enqueue_tail(skb, qdisc, list);
485 }
486
487 return qdisc_drop(skb, qdisc);
488}
489
490static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
491{
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493 int band = bitmap2band[priv->bitmap];
494
495 if (likely(band >= 0)) {
496 struct sk_buff_head *list = band2list(priv, band);
497 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
498
499 qdisc->q.qlen--;
500 if (skb_queue_empty(list))
501 priv->bitmap &= ~(1 << band);
502
503 return skb;
504 }
505
506 return NULL;
507}
508
509static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
510{
511 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
512 int band = bitmap2band[priv->bitmap];
513
514 if (band >= 0) {
515 struct sk_buff_head *list = band2list(priv, band);
516
517 return skb_peek(list);
518 }
519
520 return NULL;
521}
522
523static void pfifo_fast_reset(struct Qdisc *qdisc)
524{
525 int prio;
526 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
527
528 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
529 __qdisc_reset_queue(qdisc, band2list(priv, prio));
530
531 priv->bitmap = 0;
532 qdisc->qstats.backlog = 0;
533 qdisc->q.qlen = 0;
534}
535
536static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
537{
538 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
539
540 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
541 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
542 goto nla_put_failure;
543 return skb->len;
544
545nla_put_failure:
546 return -1;
547}
548
549static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
550{
551 int prio;
552 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
553
554 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
555 __skb_queue_head_init(band2list(priv, prio));
556
557 /* Can by-pass the queue discipline */
558 qdisc->flags |= TCQ_F_CAN_BYPASS;
559 return 0;
560}
561
562struct Qdisc_ops pfifo_fast_ops __read_mostly = {
563 .id = "pfifo_fast",
564 .priv_size = sizeof(struct pfifo_fast_priv),
565 .enqueue = pfifo_fast_enqueue,
566 .dequeue = pfifo_fast_dequeue,
567 .peek = pfifo_fast_peek,
568 .init = pfifo_fast_init,
569 .reset = pfifo_fast_reset,
570 .dump = pfifo_fast_dump,
571 .owner = THIS_MODULE,
572};
573EXPORT_SYMBOL(pfifo_fast_ops);
574
575static struct lock_class_key qdisc_tx_busylock;
576
577struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
578 const struct Qdisc_ops *ops)
579{
580 void *p;
581 struct Qdisc *sch;
582 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
583 int err = -ENOBUFS;
584 struct net_device *dev = dev_queue->dev;
585
586 p = kzalloc_node(size, GFP_KERNEL,
587 netdev_queue_numa_node_read(dev_queue));
588
589 if (!p)
590 goto errout;
591 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
592 /* if we got non aligned memory, ask more and do alignment ourself */
593 if (sch != p) {
594 kfree(p);
595 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
596 netdev_queue_numa_node_read(dev_queue));
597 if (!p)
598 goto errout;
599 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
600 sch->padded = (char *) sch - (char *) p;
601 }
602 INIT_LIST_HEAD(&sch->list);
603 skb_queue_head_init(&sch->q);
604
605 spin_lock_init(&sch->busylock);
606 lockdep_set_class(&sch->busylock,
607 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
608
609 sch->ops = ops;
610 sch->enqueue = ops->enqueue;
611 sch->dequeue = ops->dequeue;
612 sch->dev_queue = dev_queue;
613 dev_hold(dev);
614 atomic_set(&sch->refcnt, 1);
615
616 return sch;
617errout:
618 return ERR_PTR(err);
619}
620
621struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
622 const struct Qdisc_ops *ops,
623 unsigned int parentid)
624{
625 struct Qdisc *sch;
626
627 if (!try_module_get(ops->owner))
628 goto errout;
629
630 sch = qdisc_alloc(dev_queue, ops);
631 if (IS_ERR(sch))
632 goto errout;
633 sch->parent = parentid;
634
635 if (!ops->init || ops->init(sch, NULL) == 0)
636 return sch;
637
638 qdisc_destroy(sch);
639errout:
640 return NULL;
641}
642EXPORT_SYMBOL(qdisc_create_dflt);
643
644/* Under qdisc_lock(qdisc) and BH! */
645
646void qdisc_reset(struct Qdisc *qdisc)
647{
648 const struct Qdisc_ops *ops = qdisc->ops;
649
650 if (ops->reset)
651 ops->reset(qdisc);
652
653 if (qdisc->gso_skb) {
654 kfree_skb_list(qdisc->gso_skb);
655 qdisc->gso_skb = NULL;
656 qdisc->q.qlen = 0;
657 }
658}
659EXPORT_SYMBOL(qdisc_reset);
660
661static void qdisc_rcu_free(struct rcu_head *head)
662{
663 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
664
665 if (qdisc_is_percpu_stats(qdisc)) {
666 free_percpu(qdisc->cpu_bstats);
667 free_percpu(qdisc->cpu_qstats);
668 }
669
670 kfree((char *) qdisc - qdisc->padded);
671}
672
673void qdisc_destroy(struct Qdisc *qdisc)
674{
675 const struct Qdisc_ops *ops = qdisc->ops;
676
677 if (qdisc->flags & TCQ_F_BUILTIN ||
678 !atomic_dec_and_test(&qdisc->refcnt))
679 return;
680
681#ifdef CONFIG_NET_SCHED
682 qdisc_list_del(qdisc);
683
684 qdisc_put_stab(rtnl_dereference(qdisc->stab));
685#endif
686 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
687 if (ops->reset)
688 ops->reset(qdisc);
689 if (ops->destroy)
690 ops->destroy(qdisc);
691
692 module_put(ops->owner);
693 dev_put(qdisc_dev(qdisc));
694
695 kfree_skb_list(qdisc->gso_skb);
696 /*
697 * gen_estimator est_timer() might access qdisc->q.lock,
698 * wait a RCU grace period before freeing qdisc.
699 */
700 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
701}
702EXPORT_SYMBOL(qdisc_destroy);
703
704/* Attach toplevel qdisc to device queue. */
705struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
706 struct Qdisc *qdisc)
707{
708 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
709 spinlock_t *root_lock;
710
711 root_lock = qdisc_lock(oqdisc);
712 spin_lock_bh(root_lock);
713
714 /* Prune old scheduler */
715 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
716 qdisc_reset(oqdisc);
717
718 /* ... and graft new one */
719 if (qdisc == NULL)
720 qdisc = &noop_qdisc;
721 dev_queue->qdisc_sleeping = qdisc;
722 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
723
724 spin_unlock_bh(root_lock);
725
726 return oqdisc;
727}
728EXPORT_SYMBOL(dev_graft_qdisc);
729
730static void attach_one_default_qdisc(struct net_device *dev,
731 struct netdev_queue *dev_queue,
732 void *_unused)
733{
734 struct Qdisc *qdisc;
735 const struct Qdisc_ops *ops = default_qdisc_ops;
736
737 if (dev->priv_flags & IFF_NO_QUEUE)
738 ops = &noqueue_qdisc_ops;
739
740 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
741 if (!qdisc) {
742 netdev_info(dev, "activation failed\n");
743 return;
744 }
745 if (!netif_is_multiqueue(dev))
746 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
747 dev_queue->qdisc_sleeping = qdisc;
748}
749
750static void attach_default_qdiscs(struct net_device *dev)
751{
752 struct netdev_queue *txq;
753 struct Qdisc *qdisc;
754
755 txq = netdev_get_tx_queue(dev, 0);
756
757 if (!netif_is_multiqueue(dev) ||
758 dev->priv_flags & IFF_NO_QUEUE) {
759 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
760 dev->qdisc = txq->qdisc_sleeping;
761 atomic_inc(&dev->qdisc->refcnt);
762 } else {
763 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
764 if (qdisc) {
765 dev->qdisc = qdisc;
766 qdisc->ops->attach(qdisc);
767 }
768 }
769}
770
771static void transition_one_qdisc(struct net_device *dev,
772 struct netdev_queue *dev_queue,
773 void *_need_watchdog)
774{
775 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
776 int *need_watchdog_p = _need_watchdog;
777
778 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
779 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
780
781 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
782 if (need_watchdog_p) {
783 dev_queue->trans_start = 0;
784 *need_watchdog_p = 1;
785 }
786}
787
788void dev_activate(struct net_device *dev)
789{
790 int need_watchdog;
791
792 /* No queueing discipline is attached to device;
793 * create default one for devices, which need queueing
794 * and noqueue_qdisc for virtual interfaces
795 */
796
797 if (dev->qdisc == &noop_qdisc)
798 attach_default_qdiscs(dev);
799
800 if (!netif_carrier_ok(dev))
801 /* Delay activation until next carrier-on event */
802 return;
803
804 need_watchdog = 0;
805 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
806 if (dev_ingress_queue(dev))
807 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
808
809 if (need_watchdog) {
810 dev->trans_start = jiffies;
811 dev_watchdog_up(dev);
812 }
813}
814EXPORT_SYMBOL(dev_activate);
815
816static void dev_deactivate_queue(struct net_device *dev,
817 struct netdev_queue *dev_queue,
818 void *_qdisc_default)
819{
820 struct Qdisc *qdisc_default = _qdisc_default;
821 struct Qdisc *qdisc;
822
823 qdisc = rtnl_dereference(dev_queue->qdisc);
824 if (qdisc) {
825 spin_lock_bh(qdisc_lock(qdisc));
826
827 if (!(qdisc->flags & TCQ_F_BUILTIN))
828 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
829
830 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
831 qdisc_reset(qdisc);
832
833 spin_unlock_bh(qdisc_lock(qdisc));
834 }
835}
836
837static bool some_qdisc_is_busy(struct net_device *dev)
838{
839 unsigned int i;
840
841 for (i = 0; i < dev->num_tx_queues; i++) {
842 struct netdev_queue *dev_queue;
843 spinlock_t *root_lock;
844 struct Qdisc *q;
845 int val;
846
847 dev_queue = netdev_get_tx_queue(dev, i);
848 q = dev_queue->qdisc_sleeping;
849 root_lock = qdisc_lock(q);
850
851 spin_lock_bh(root_lock);
852
853 val = (qdisc_is_running(q) ||
854 test_bit(__QDISC_STATE_SCHED, &q->state));
855
856 spin_unlock_bh(root_lock);
857
858 if (val)
859 return true;
860 }
861 return false;
862}
863
864/**
865 * dev_deactivate_many - deactivate transmissions on several devices
866 * @head: list of devices to deactivate
867 *
868 * This function returns only when all outstanding transmissions
869 * have completed, unless all devices are in dismantle phase.
870 */
871void dev_deactivate_many(struct list_head *head)
872{
873 struct net_device *dev;
874 bool sync_needed = false;
875
876 list_for_each_entry(dev, head, close_list) {
877 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
878 &noop_qdisc);
879 if (dev_ingress_queue(dev))
880 dev_deactivate_queue(dev, dev_ingress_queue(dev),
881 &noop_qdisc);
882
883 dev_watchdog_down(dev);
884 sync_needed |= !dev->dismantle;
885 }
886
887 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
888 * This is avoided if all devices are in dismantle phase :
889 * Caller will call synchronize_net() for us
890 */
891 if (sync_needed)
892 synchronize_net();
893
894 /* Wait for outstanding qdisc_run calls. */
895 list_for_each_entry(dev, head, close_list)
896 while (some_qdisc_is_busy(dev))
897 yield();
898}
899
900void dev_deactivate(struct net_device *dev)
901{
902 LIST_HEAD(single);
903
904 list_add(&dev->close_list, &single);
905 dev_deactivate_many(&single);
906 list_del(&single);
907}
908EXPORT_SYMBOL(dev_deactivate);
909
910static void dev_init_scheduler_queue(struct net_device *dev,
911 struct netdev_queue *dev_queue,
912 void *_qdisc)
913{
914 struct Qdisc *qdisc = _qdisc;
915
916 rcu_assign_pointer(dev_queue->qdisc, qdisc);
917 dev_queue->qdisc_sleeping = qdisc;
918}
919
920void dev_init_scheduler(struct net_device *dev)
921{
922 dev->qdisc = &noop_qdisc;
923 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
924 if (dev_ingress_queue(dev))
925 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
926
927 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
928}
929
930static void shutdown_scheduler_queue(struct net_device *dev,
931 struct netdev_queue *dev_queue,
932 void *_qdisc_default)
933{
934 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
935 struct Qdisc *qdisc_default = _qdisc_default;
936
937 if (qdisc) {
938 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
939 dev_queue->qdisc_sleeping = qdisc_default;
940
941 qdisc_destroy(qdisc);
942 }
943}
944
945void dev_shutdown(struct net_device *dev)
946{
947 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
948 if (dev_ingress_queue(dev))
949 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
950 qdisc_destroy(dev->qdisc);
951 dev->qdisc = &noop_qdisc;
952
953 WARN_ON(timer_pending(&dev->watchdog_timer));
954}
955
956void psched_ratecfg_precompute(struct psched_ratecfg *r,
957 const struct tc_ratespec *conf,
958 u64 rate64)
959{
960 memset(r, 0, sizeof(*r));
961 r->overhead = conf->overhead;
962 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
963 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
964 r->mult = 1;
965 /*
966 * The deal here is to replace a divide by a reciprocal one
967 * in fast path (a reciprocal divide is a multiply and a shift)
968 *
969 * Normal formula would be :
970 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
971 *
972 * We compute mult/shift to use instead :
973 * time_in_ns = (len * mult) >> shift;
974 *
975 * We try to get the highest possible mult value for accuracy,
976 * but have to make sure no overflows will ever happen.
977 */
978 if (r->rate_bytes_ps > 0) {
979 u64 factor = NSEC_PER_SEC;
980
981 for (;;) {
982 r->mult = div64_u64(factor, r->rate_bytes_ps);
983 if (r->mult & (1U << 31) || factor & (1ULL << 63))
984 break;
985 factor <<= 1;
986 r->shift++;
987 }
988 }
989}
990EXPORT_SYMBOL(psched_ratecfg_precompute);
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 */
36const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
37EXPORT_SYMBOL(default_qdisc_ops);
38
39static 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
73static 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
109static 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
119static 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
143static 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
179static 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 */
203static 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 */
229static 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 }
278validate:
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) {
295bulk:
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 }
301trace:
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 */
315bool 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 */
389static 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
411void __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
429unsigned 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}
443EXPORT_SYMBOL(dev_trans_start);
444
445static 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
466void netif_tx_lock(struct net_device *dev)
467{
468 spin_lock(&dev->tx_global_lock);
469 netif_freeze_queues(dev);
470}
471EXPORT_SYMBOL(netif_tx_lock);
472
473static 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
489void netif_tx_unlock(struct net_device *dev)
490{
491 netif_unfreeze_queues(dev);
492 spin_unlock(&dev->tx_global_lock);
493}
494EXPORT_SYMBOL(netif_tx_unlock);
495
496static 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
510 for (i = 0; i < dev->num_tx_queues; i++) {
511 struct netdev_queue *txq;
512
513 txq = netdev_get_tx_queue(dev, i);
514 trans_start = READ_ONCE(txq->trans_start);
515 if (netif_xmit_stopped(txq) &&
516 time_after(jiffies, (trans_start +
517 dev->watchdog_timeo))) {
518 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
519 atomic_long_inc(&txq->trans_timeout);
520 break;
521 }
522 }
523
524 if (unlikely(timedout_ms)) {
525 trace_net_dev_xmit_timeout(dev, i);
526 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
527 raw_smp_processor_id(),
528 i, timedout_ms);
529 netif_freeze_queues(dev);
530 dev->netdev_ops->ndo_tx_timeout(dev, i);
531 netif_unfreeze_queues(dev);
532 }
533 if (!mod_timer(&dev->watchdog_timer,
534 round_jiffies(jiffies +
535 dev->watchdog_timeo)))
536 release = false;
537 }
538 }
539 spin_unlock(&dev->tx_global_lock);
540
541 if (release)
542 netdev_put(dev, &dev->watchdog_dev_tracker);
543}
544
545void __netdev_watchdog_up(struct net_device *dev)
546{
547 if (dev->netdev_ops->ndo_tx_timeout) {
548 if (dev->watchdog_timeo <= 0)
549 dev->watchdog_timeo = 5*HZ;
550 if (!mod_timer(&dev->watchdog_timer,
551 round_jiffies(jiffies + dev->watchdog_timeo)))
552 netdev_hold(dev, &dev->watchdog_dev_tracker,
553 GFP_ATOMIC);
554 }
555}
556EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
557
558static void dev_watchdog_up(struct net_device *dev)
559{
560 __netdev_watchdog_up(dev);
561}
562
563static void dev_watchdog_down(struct net_device *dev)
564{
565 netif_tx_lock_bh(dev);
566 if (del_timer(&dev->watchdog_timer))
567 netdev_put(dev, &dev->watchdog_dev_tracker);
568 netif_tx_unlock_bh(dev);
569}
570
571/**
572 * netif_carrier_on - set carrier
573 * @dev: network device
574 *
575 * Device has detected acquisition of carrier.
576 */
577void netif_carrier_on(struct net_device *dev)
578{
579 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
580 if (dev->reg_state == NETREG_UNINITIALIZED)
581 return;
582 atomic_inc(&dev->carrier_up_count);
583 linkwatch_fire_event(dev);
584 if (netif_running(dev))
585 __netdev_watchdog_up(dev);
586 }
587}
588EXPORT_SYMBOL(netif_carrier_on);
589
590/**
591 * netif_carrier_off - clear carrier
592 * @dev: network device
593 *
594 * Device has detected loss of carrier.
595 */
596void netif_carrier_off(struct net_device *dev)
597{
598 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
599 if (dev->reg_state == NETREG_UNINITIALIZED)
600 return;
601 atomic_inc(&dev->carrier_down_count);
602 linkwatch_fire_event(dev);
603 }
604}
605EXPORT_SYMBOL(netif_carrier_off);
606
607/**
608 * netif_carrier_event - report carrier state event
609 * @dev: network device
610 *
611 * Device has detected a carrier event but the carrier state wasn't changed.
612 * Use in drivers when querying carrier state asynchronously, to avoid missing
613 * events (link flaps) if link recovers before it's queried.
614 */
615void netif_carrier_event(struct net_device *dev)
616{
617 if (dev->reg_state == NETREG_UNINITIALIZED)
618 return;
619 atomic_inc(&dev->carrier_up_count);
620 atomic_inc(&dev->carrier_down_count);
621 linkwatch_fire_event(dev);
622}
623EXPORT_SYMBOL_GPL(netif_carrier_event);
624
625/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
626 under all circumstances. It is difficult to invent anything faster or
627 cheaper.
628 */
629
630static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
631 struct sk_buff **to_free)
632{
633 __qdisc_drop(skb, to_free);
634 return NET_XMIT_CN;
635}
636
637static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
638{
639 return NULL;
640}
641
642struct Qdisc_ops noop_qdisc_ops __read_mostly = {
643 .id = "noop",
644 .priv_size = 0,
645 .enqueue = noop_enqueue,
646 .dequeue = noop_dequeue,
647 .peek = noop_dequeue,
648 .owner = THIS_MODULE,
649};
650
651static struct netdev_queue noop_netdev_queue = {
652 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
653 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
654};
655
656struct Qdisc noop_qdisc = {
657 .enqueue = noop_enqueue,
658 .dequeue = noop_dequeue,
659 .flags = TCQ_F_BUILTIN,
660 .ops = &noop_qdisc_ops,
661 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
662 .dev_queue = &noop_netdev_queue,
663 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
664 .gso_skb = {
665 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
666 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
667 .qlen = 0,
668 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
669 },
670 .skb_bad_txq = {
671 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
672 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
673 .qlen = 0,
674 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
675 },
676};
677EXPORT_SYMBOL(noop_qdisc);
678
679static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
680 struct netlink_ext_ack *extack)
681{
682 /* register_qdisc() assigns a default of noop_enqueue if unset,
683 * but __dev_queue_xmit() treats noqueue only as such
684 * if this is NULL - so clear it here. */
685 qdisc->enqueue = NULL;
686 return 0;
687}
688
689struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
690 .id = "noqueue",
691 .priv_size = 0,
692 .init = noqueue_init,
693 .enqueue = noop_enqueue,
694 .dequeue = noop_dequeue,
695 .peek = noop_dequeue,
696 .owner = THIS_MODULE,
697};
698
699const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
700 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
701};
702EXPORT_SYMBOL(sch_default_prio2band);
703
704/* 3-band FIFO queue: old style, but should be a bit faster than
705 generic prio+fifo combination.
706 */
707
708#define PFIFO_FAST_BANDS 3
709
710/*
711 * Private data for a pfifo_fast scheduler containing:
712 * - rings for priority bands
713 */
714struct pfifo_fast_priv {
715 struct skb_array q[PFIFO_FAST_BANDS];
716};
717
718static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
719 int band)
720{
721 return &priv->q[band];
722}
723
724static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
725 struct sk_buff **to_free)
726{
727 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
728 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
729 struct skb_array *q = band2list(priv, band);
730 unsigned int pkt_len = qdisc_pkt_len(skb);
731 int err;
732
733 err = skb_array_produce(q, skb);
734
735 if (unlikely(err)) {
736 if (qdisc_is_percpu_stats(qdisc))
737 return qdisc_drop_cpu(skb, qdisc, to_free);
738 else
739 return qdisc_drop(skb, qdisc, to_free);
740 }
741
742 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
743 return NET_XMIT_SUCCESS;
744}
745
746static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
747{
748 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
749 struct sk_buff *skb = NULL;
750 bool need_retry = true;
751 int band;
752
753retry:
754 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
755 struct skb_array *q = band2list(priv, band);
756
757 if (__skb_array_empty(q))
758 continue;
759
760 skb = __skb_array_consume(q);
761 }
762 if (likely(skb)) {
763 qdisc_update_stats_at_dequeue(qdisc, skb);
764 } else if (need_retry &&
765 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
766 /* Delay clearing the STATE_MISSED here to reduce
767 * the overhead of the second spin_trylock() in
768 * qdisc_run_begin() and __netif_schedule() calling
769 * in qdisc_run_end().
770 */
771 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
772 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
773
774 /* Make sure dequeuing happens after clearing
775 * STATE_MISSED.
776 */
777 smp_mb__after_atomic();
778
779 need_retry = false;
780
781 goto retry;
782 }
783
784 return skb;
785}
786
787static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
788{
789 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
790 struct sk_buff *skb = NULL;
791 int band;
792
793 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
794 struct skb_array *q = band2list(priv, band);
795
796 skb = __skb_array_peek(q);
797 }
798
799 return skb;
800}
801
802static void pfifo_fast_reset(struct Qdisc *qdisc)
803{
804 int i, band;
805 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
806
807 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
808 struct skb_array *q = band2list(priv, band);
809 struct sk_buff *skb;
810
811 /* NULL ring is possible if destroy path is due to a failed
812 * skb_array_init() in pfifo_fast_init() case.
813 */
814 if (!q->ring.queue)
815 continue;
816
817 while ((skb = __skb_array_consume(q)) != NULL)
818 kfree_skb(skb);
819 }
820
821 if (qdisc_is_percpu_stats(qdisc)) {
822 for_each_possible_cpu(i) {
823 struct gnet_stats_queue *q;
824
825 q = per_cpu_ptr(qdisc->cpu_qstats, i);
826 q->backlog = 0;
827 q->qlen = 0;
828 }
829 }
830}
831
832static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
833{
834 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
835
836 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
837 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
838 goto nla_put_failure;
839 return skb->len;
840
841nla_put_failure:
842 return -1;
843}
844
845static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
846 struct netlink_ext_ack *extack)
847{
848 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
849 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
850 int prio;
851
852 /* guard against zero length rings */
853 if (!qlen)
854 return -EINVAL;
855
856 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
857 struct skb_array *q = band2list(priv, prio);
858 int err;
859
860 err = skb_array_init(q, qlen, GFP_KERNEL);
861 if (err)
862 return -ENOMEM;
863 }
864
865 /* Can by-pass the queue discipline */
866 qdisc->flags |= TCQ_F_CAN_BYPASS;
867 return 0;
868}
869
870static void pfifo_fast_destroy(struct Qdisc *sch)
871{
872 struct pfifo_fast_priv *priv = qdisc_priv(sch);
873 int prio;
874
875 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
876 struct skb_array *q = band2list(priv, prio);
877
878 /* NULL ring is possible if destroy path is due to a failed
879 * skb_array_init() in pfifo_fast_init() case.
880 */
881 if (!q->ring.queue)
882 continue;
883 /* Destroy ring but no need to kfree_skb because a call to
884 * pfifo_fast_reset() has already done that work.
885 */
886 ptr_ring_cleanup(&q->ring, NULL);
887 }
888}
889
890static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
891 unsigned int new_len)
892{
893 struct pfifo_fast_priv *priv = qdisc_priv(sch);
894 struct skb_array *bands[PFIFO_FAST_BANDS];
895 int prio;
896
897 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
898 struct skb_array *q = band2list(priv, prio);
899
900 bands[prio] = q;
901 }
902
903 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
904 GFP_KERNEL);
905}
906
907struct Qdisc_ops pfifo_fast_ops __read_mostly = {
908 .id = "pfifo_fast",
909 .priv_size = sizeof(struct pfifo_fast_priv),
910 .enqueue = pfifo_fast_enqueue,
911 .dequeue = pfifo_fast_dequeue,
912 .peek = pfifo_fast_peek,
913 .init = pfifo_fast_init,
914 .destroy = pfifo_fast_destroy,
915 .reset = pfifo_fast_reset,
916 .dump = pfifo_fast_dump,
917 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
918 .owner = THIS_MODULE,
919 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
920};
921EXPORT_SYMBOL(pfifo_fast_ops);
922
923static struct lock_class_key qdisc_tx_busylock;
924
925struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
926 const struct Qdisc_ops *ops,
927 struct netlink_ext_ack *extack)
928{
929 struct Qdisc *sch;
930 unsigned int size = sizeof(*sch) + ops->priv_size;
931 int err = -ENOBUFS;
932 struct net_device *dev;
933
934 if (!dev_queue) {
935 NL_SET_ERR_MSG(extack, "No device queue given");
936 err = -EINVAL;
937 goto errout;
938 }
939
940 dev = dev_queue->dev;
941 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
942
943 if (!sch)
944 goto errout;
945 __skb_queue_head_init(&sch->gso_skb);
946 __skb_queue_head_init(&sch->skb_bad_txq);
947 gnet_stats_basic_sync_init(&sch->bstats);
948 spin_lock_init(&sch->q.lock);
949
950 if (ops->static_flags & TCQ_F_CPUSTATS) {
951 sch->cpu_bstats =
952 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
953 if (!sch->cpu_bstats)
954 goto errout1;
955
956 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
957 if (!sch->cpu_qstats) {
958 free_percpu(sch->cpu_bstats);
959 goto errout1;
960 }
961 }
962
963 spin_lock_init(&sch->busylock);
964 lockdep_set_class(&sch->busylock,
965 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
966
967 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
968 spin_lock_init(&sch->seqlock);
969 lockdep_set_class(&sch->seqlock,
970 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
971
972 sch->ops = ops;
973 sch->flags = ops->static_flags;
974 sch->enqueue = ops->enqueue;
975 sch->dequeue = ops->dequeue;
976 sch->dev_queue = dev_queue;
977 sch->owner = -1;
978 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
979 refcount_set(&sch->refcnt, 1);
980
981 return sch;
982errout1:
983 kfree(sch);
984errout:
985 return ERR_PTR(err);
986}
987
988struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
989 const struct Qdisc_ops *ops,
990 unsigned int parentid,
991 struct netlink_ext_ack *extack)
992{
993 struct Qdisc *sch;
994
995 if (!try_module_get(ops->owner)) {
996 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
997 return NULL;
998 }
999
1000 sch = qdisc_alloc(dev_queue, ops, extack);
1001 if (IS_ERR(sch)) {
1002 module_put(ops->owner);
1003 return NULL;
1004 }
1005 sch->parent = parentid;
1006
1007 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1008 trace_qdisc_create(ops, dev_queue->dev, parentid);
1009 return sch;
1010 }
1011
1012 qdisc_put(sch);
1013 return NULL;
1014}
1015EXPORT_SYMBOL(qdisc_create_dflt);
1016
1017/* Under qdisc_lock(qdisc) and BH! */
1018
1019void qdisc_reset(struct Qdisc *qdisc)
1020{
1021 const struct Qdisc_ops *ops = qdisc->ops;
1022
1023 trace_qdisc_reset(qdisc);
1024
1025 if (ops->reset)
1026 ops->reset(qdisc);
1027
1028 __skb_queue_purge(&qdisc->gso_skb);
1029 __skb_queue_purge(&qdisc->skb_bad_txq);
1030
1031 qdisc->q.qlen = 0;
1032 qdisc->qstats.backlog = 0;
1033}
1034EXPORT_SYMBOL(qdisc_reset);
1035
1036void qdisc_free(struct Qdisc *qdisc)
1037{
1038 if (qdisc_is_percpu_stats(qdisc)) {
1039 free_percpu(qdisc->cpu_bstats);
1040 free_percpu(qdisc->cpu_qstats);
1041 }
1042
1043 kfree(qdisc);
1044}
1045
1046static void qdisc_free_cb(struct rcu_head *head)
1047{
1048 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1049
1050 qdisc_free(q);
1051}
1052
1053static void __qdisc_destroy(struct Qdisc *qdisc)
1054{
1055 const struct Qdisc_ops *ops = qdisc->ops;
1056 struct net_device *dev = qdisc_dev(qdisc);
1057
1058#ifdef CONFIG_NET_SCHED
1059 qdisc_hash_del(qdisc);
1060
1061 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1062#endif
1063 gen_kill_estimator(&qdisc->rate_est);
1064
1065 qdisc_reset(qdisc);
1066
1067
1068 if (ops->destroy)
1069 ops->destroy(qdisc);
1070
1071 module_put(ops->owner);
1072 netdev_put(dev, &qdisc->dev_tracker);
1073
1074 trace_qdisc_destroy(qdisc);
1075
1076 call_rcu(&qdisc->rcu, qdisc_free_cb);
1077}
1078
1079void qdisc_destroy(struct Qdisc *qdisc)
1080{
1081 if (qdisc->flags & TCQ_F_BUILTIN)
1082 return;
1083
1084 __qdisc_destroy(qdisc);
1085}
1086
1087void qdisc_put(struct Qdisc *qdisc)
1088{
1089 if (!qdisc)
1090 return;
1091
1092 if (qdisc->flags & TCQ_F_BUILTIN ||
1093 !refcount_dec_and_test(&qdisc->refcnt))
1094 return;
1095
1096 __qdisc_destroy(qdisc);
1097}
1098EXPORT_SYMBOL(qdisc_put);
1099
1100/* Version of qdisc_put() that is called with rtnl mutex unlocked.
1101 * Intended to be used as optimization, this function only takes rtnl lock if
1102 * qdisc reference counter reached zero.
1103 */
1104
1105void qdisc_put_unlocked(struct Qdisc *qdisc)
1106{
1107 if (qdisc->flags & TCQ_F_BUILTIN ||
1108 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1109 return;
1110
1111 __qdisc_destroy(qdisc);
1112 rtnl_unlock();
1113}
1114EXPORT_SYMBOL(qdisc_put_unlocked);
1115
1116/* Attach toplevel qdisc to device queue. */
1117struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1118 struct Qdisc *qdisc)
1119{
1120 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1121 spinlock_t *root_lock;
1122
1123 root_lock = qdisc_lock(oqdisc);
1124 spin_lock_bh(root_lock);
1125
1126 /* ... and graft new one */
1127 if (qdisc == NULL)
1128 qdisc = &noop_qdisc;
1129 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1130 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1131
1132 spin_unlock_bh(root_lock);
1133
1134 return oqdisc;
1135}
1136EXPORT_SYMBOL(dev_graft_qdisc);
1137
1138static void shutdown_scheduler_queue(struct net_device *dev,
1139 struct netdev_queue *dev_queue,
1140 void *_qdisc_default)
1141{
1142 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1143 struct Qdisc *qdisc_default = _qdisc_default;
1144
1145 if (qdisc) {
1146 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1147 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1148
1149 qdisc_put(qdisc);
1150 }
1151}
1152
1153static void attach_one_default_qdisc(struct net_device *dev,
1154 struct netdev_queue *dev_queue,
1155 void *_unused)
1156{
1157 struct Qdisc *qdisc;
1158 const struct Qdisc_ops *ops = default_qdisc_ops;
1159
1160 if (dev->priv_flags & IFF_NO_QUEUE)
1161 ops = &noqueue_qdisc_ops;
1162 else if(dev->type == ARPHRD_CAN)
1163 ops = &pfifo_fast_ops;
1164
1165 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1166 if (!qdisc)
1167 return;
1168
1169 if (!netif_is_multiqueue(dev))
1170 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1171 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1172}
1173
1174static void attach_default_qdiscs(struct net_device *dev)
1175{
1176 struct netdev_queue *txq;
1177 struct Qdisc *qdisc;
1178
1179 txq = netdev_get_tx_queue(dev, 0);
1180
1181 if (!netif_is_multiqueue(dev) ||
1182 dev->priv_flags & IFF_NO_QUEUE) {
1183 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1184 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1185 rcu_assign_pointer(dev->qdisc, qdisc);
1186 qdisc_refcount_inc(qdisc);
1187 } else {
1188 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1189 if (qdisc) {
1190 rcu_assign_pointer(dev->qdisc, qdisc);
1191 qdisc->ops->attach(qdisc);
1192 }
1193 }
1194 qdisc = rtnl_dereference(dev->qdisc);
1195
1196 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1197 if (qdisc == &noop_qdisc) {
1198 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1199 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1200 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1201 dev->priv_flags |= IFF_NO_QUEUE;
1202 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1203 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1204 rcu_assign_pointer(dev->qdisc, qdisc);
1205 qdisc_refcount_inc(qdisc);
1206 dev->priv_flags ^= IFF_NO_QUEUE;
1207 }
1208
1209#ifdef CONFIG_NET_SCHED
1210 if (qdisc != &noop_qdisc)
1211 qdisc_hash_add(qdisc, false);
1212#endif
1213}
1214
1215static void transition_one_qdisc(struct net_device *dev,
1216 struct netdev_queue *dev_queue,
1217 void *_need_watchdog)
1218{
1219 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1220 int *need_watchdog_p = _need_watchdog;
1221
1222 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1223 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1224
1225 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1226 if (need_watchdog_p) {
1227 WRITE_ONCE(dev_queue->trans_start, 0);
1228 *need_watchdog_p = 1;
1229 }
1230}
1231
1232void dev_activate(struct net_device *dev)
1233{
1234 int need_watchdog;
1235
1236 /* No queueing discipline is attached to device;
1237 * create default one for devices, which need queueing
1238 * and noqueue_qdisc for virtual interfaces
1239 */
1240
1241 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1242 attach_default_qdiscs(dev);
1243
1244 if (!netif_carrier_ok(dev))
1245 /* Delay activation until next carrier-on event */
1246 return;
1247
1248 need_watchdog = 0;
1249 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1250 if (dev_ingress_queue(dev))
1251 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1252
1253 if (need_watchdog) {
1254 netif_trans_update(dev);
1255 dev_watchdog_up(dev);
1256 }
1257}
1258EXPORT_SYMBOL(dev_activate);
1259
1260static void qdisc_deactivate(struct Qdisc *qdisc)
1261{
1262 if (qdisc->flags & TCQ_F_BUILTIN)
1263 return;
1264
1265 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1266}
1267
1268static void dev_deactivate_queue(struct net_device *dev,
1269 struct netdev_queue *dev_queue,
1270 void *_qdisc_default)
1271{
1272 struct Qdisc *qdisc_default = _qdisc_default;
1273 struct Qdisc *qdisc;
1274
1275 qdisc = rtnl_dereference(dev_queue->qdisc);
1276 if (qdisc) {
1277 qdisc_deactivate(qdisc);
1278 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1279 }
1280}
1281
1282static void dev_reset_queue(struct net_device *dev,
1283 struct netdev_queue *dev_queue,
1284 void *_unused)
1285{
1286 struct Qdisc *qdisc;
1287 bool nolock;
1288
1289 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1290 if (!qdisc)
1291 return;
1292
1293 nolock = qdisc->flags & TCQ_F_NOLOCK;
1294
1295 if (nolock)
1296 spin_lock_bh(&qdisc->seqlock);
1297 spin_lock_bh(qdisc_lock(qdisc));
1298
1299 qdisc_reset(qdisc);
1300
1301 spin_unlock_bh(qdisc_lock(qdisc));
1302 if (nolock) {
1303 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1304 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1305 spin_unlock_bh(&qdisc->seqlock);
1306 }
1307}
1308
1309static bool some_qdisc_is_busy(struct net_device *dev)
1310{
1311 unsigned int i;
1312
1313 for (i = 0; i < dev->num_tx_queues; i++) {
1314 struct netdev_queue *dev_queue;
1315 spinlock_t *root_lock;
1316 struct Qdisc *q;
1317 int val;
1318
1319 dev_queue = netdev_get_tx_queue(dev, i);
1320 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1321
1322 root_lock = qdisc_lock(q);
1323 spin_lock_bh(root_lock);
1324
1325 val = (qdisc_is_running(q) ||
1326 test_bit(__QDISC_STATE_SCHED, &q->state));
1327
1328 spin_unlock_bh(root_lock);
1329
1330 if (val)
1331 return true;
1332 }
1333 return false;
1334}
1335
1336/**
1337 * dev_deactivate_many - deactivate transmissions on several devices
1338 * @head: list of devices to deactivate
1339 *
1340 * This function returns only when all outstanding transmissions
1341 * have completed, unless all devices are in dismantle phase.
1342 */
1343void dev_deactivate_many(struct list_head *head)
1344{
1345 struct net_device *dev;
1346
1347 list_for_each_entry(dev, head, close_list) {
1348 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1349 &noop_qdisc);
1350 if (dev_ingress_queue(dev))
1351 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1352 &noop_qdisc);
1353
1354 dev_watchdog_down(dev);
1355 }
1356
1357 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1358 * outstanding qdisc enqueuing calls.
1359 * This is avoided if all devices are in dismantle phase :
1360 * Caller will call synchronize_net() for us
1361 */
1362 synchronize_net();
1363
1364 list_for_each_entry(dev, head, close_list) {
1365 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1366
1367 if (dev_ingress_queue(dev))
1368 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1369 }
1370
1371 /* Wait for outstanding qdisc_run calls. */
1372 list_for_each_entry(dev, head, close_list) {
1373 while (some_qdisc_is_busy(dev)) {
1374 /* wait_event() would avoid this sleep-loop but would
1375 * require expensive checks in the fast paths of packet
1376 * processing which isn't worth it.
1377 */
1378 schedule_timeout_uninterruptible(1);
1379 }
1380 }
1381}
1382
1383void dev_deactivate(struct net_device *dev)
1384{
1385 LIST_HEAD(single);
1386
1387 list_add(&dev->close_list, &single);
1388 dev_deactivate_many(&single);
1389 list_del(&single);
1390}
1391EXPORT_SYMBOL(dev_deactivate);
1392
1393static int qdisc_change_tx_queue_len(struct net_device *dev,
1394 struct netdev_queue *dev_queue)
1395{
1396 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1397 const struct Qdisc_ops *ops = qdisc->ops;
1398
1399 if (ops->change_tx_queue_len)
1400 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1401 return 0;
1402}
1403
1404void dev_qdisc_change_real_num_tx(struct net_device *dev,
1405 unsigned int new_real_tx)
1406{
1407 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1408
1409 if (qdisc->ops->change_real_num_tx)
1410 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1411}
1412
1413void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1414{
1415#ifdef CONFIG_NET_SCHED
1416 struct net_device *dev = qdisc_dev(sch);
1417 struct Qdisc *qdisc;
1418 unsigned int i;
1419
1420 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1421 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1422 /* Only update the default qdiscs we created,
1423 * qdiscs with handles are always hashed.
1424 */
1425 if (qdisc != &noop_qdisc && !qdisc->handle)
1426 qdisc_hash_del(qdisc);
1427 }
1428 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1429 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1430 if (qdisc != &noop_qdisc && !qdisc->handle)
1431 qdisc_hash_add(qdisc, false);
1432 }
1433#endif
1434}
1435EXPORT_SYMBOL(mq_change_real_num_tx);
1436
1437int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1438{
1439 bool up = dev->flags & IFF_UP;
1440 unsigned int i;
1441 int ret = 0;
1442
1443 if (up)
1444 dev_deactivate(dev);
1445
1446 for (i = 0; i < dev->num_tx_queues; i++) {
1447 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1448
1449 /* TODO: revert changes on a partial failure */
1450 if (ret)
1451 break;
1452 }
1453
1454 if (up)
1455 dev_activate(dev);
1456 return ret;
1457}
1458
1459static void dev_init_scheduler_queue(struct net_device *dev,
1460 struct netdev_queue *dev_queue,
1461 void *_qdisc)
1462{
1463 struct Qdisc *qdisc = _qdisc;
1464
1465 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1466 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1467}
1468
1469void dev_init_scheduler(struct net_device *dev)
1470{
1471 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1472 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1473 if (dev_ingress_queue(dev))
1474 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1475
1476 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1477}
1478
1479void dev_shutdown(struct net_device *dev)
1480{
1481 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1482 if (dev_ingress_queue(dev))
1483 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1484 qdisc_put(rtnl_dereference(dev->qdisc));
1485 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1486
1487 WARN_ON(timer_pending(&dev->watchdog_timer));
1488}
1489
1490/**
1491 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1492 * @rate: Rate to compute reciprocal division values of
1493 * @mult: Multiplier for reciprocal division
1494 * @shift: Shift for reciprocal division
1495 *
1496 * The multiplier and shift for reciprocal division by rate are stored
1497 * in mult and shift.
1498 *
1499 * The deal here is to replace a divide by a reciprocal one
1500 * in fast path (a reciprocal divide is a multiply and a shift)
1501 *
1502 * Normal formula would be :
1503 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1504 *
1505 * We compute mult/shift to use instead :
1506 * time_in_ns = (len * mult) >> shift;
1507 *
1508 * We try to get the highest possible mult value for accuracy,
1509 * but have to make sure no overflows will ever happen.
1510 *
1511 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1512 */
1513static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1514{
1515 u64 factor = NSEC_PER_SEC;
1516
1517 *mult = 1;
1518 *shift = 0;
1519
1520 if (rate <= 0)
1521 return;
1522
1523 for (;;) {
1524 *mult = div64_u64(factor, rate);
1525 if (*mult & (1U << 31) || factor & (1ULL << 63))
1526 break;
1527 factor <<= 1;
1528 (*shift)++;
1529 }
1530}
1531
1532void psched_ratecfg_precompute(struct psched_ratecfg *r,
1533 const struct tc_ratespec *conf,
1534 u64 rate64)
1535{
1536 memset(r, 0, sizeof(*r));
1537 r->overhead = conf->overhead;
1538 r->mpu = conf->mpu;
1539 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1540 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1541 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1542}
1543EXPORT_SYMBOL(psched_ratecfg_precompute);
1544
1545void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1546{
1547 r->rate_pkts_ps = pktrate64;
1548 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1549}
1550EXPORT_SYMBOL(psched_ppscfg_precompute);
1551
1552void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1553 struct tcf_proto *tp_head)
1554{
1555 /* Protected with chain0->filter_chain_lock.
1556 * Can't access chain directly because tp_head can be NULL.
1557 */
1558 struct mini_Qdisc *miniq_old =
1559 rcu_dereference_protected(*miniqp->p_miniq, 1);
1560 struct mini_Qdisc *miniq;
1561
1562 if (!tp_head) {
1563 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1564 } else {
1565 miniq = miniq_old != &miniqp->miniq1 ?
1566 &miniqp->miniq1 : &miniqp->miniq2;
1567
1568 /* We need to make sure that readers won't see the miniq
1569 * we are about to modify. So ensure that at least one RCU
1570 * grace period has elapsed since the miniq was made
1571 * inactive.
1572 */
1573 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1574 cond_synchronize_rcu(miniq->rcu_state);
1575 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1576 synchronize_rcu_expedited();
1577
1578 miniq->filter_list = tp_head;
1579 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1580 }
1581
1582 if (miniq_old)
1583 /* This is counterpart of the rcu sync above. We need to
1584 * block potential new user of miniq_old until all readers
1585 * are not seeing it.
1586 */
1587 miniq_old->rcu_state = start_poll_synchronize_rcu();
1588}
1589EXPORT_SYMBOL(mini_qdisc_pair_swap);
1590
1591void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1592 struct tcf_block *block)
1593{
1594 miniqp->miniq1.block = block;
1595 miniqp->miniq2.block = block;
1596}
1597EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1598
1599void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1600 struct mini_Qdisc __rcu **p_miniq)
1601{
1602 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1603 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1604 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1605 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1606 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1607 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1608 miniqp->p_miniq = p_miniq;
1609}
1610EXPORT_SYMBOL(mini_qdisc_pair_init);