2 * net/sched/sch_generic.c Generic packet scheduler routines.
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.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
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>
33 /* Qdisc to use by default */
34 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35 EXPORT_SYMBOL(default_qdisc_ops);
37 /* Main transmission queue. */
39 /* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
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.
48 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
52 q->q.qlen++; /* it's still part of the queue */
58 static void try_bulk_dequeue_skb(struct Qdisc *q,
60 const struct netdev_queue *txq,
63 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
65 while (bytelimit > 0) {
66 struct sk_buff *nskb = q->dequeue(q);
71 bytelimit -= nskb->len; /* covers GSO len */
74 (*packets)++; /* GSO counts as one pkt */
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.
82 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
85 struct sk_buff *skb = q->gso_skb;
86 const struct netdev_queue *txq = q->dev_queue;
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)) {
98 /* skb in gso_skb were already validated */
101 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
102 !netif_xmit_frozen_or_stopped(txq)) {
104 if (skb && qdisc_may_bulk(q))
105 try_bulk_dequeue_skb(q, skb, txq, packets);
111 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
112 struct netdev_queue *dev_queue,
117 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
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.
125 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
126 dev_queue->dev->name);
130 * Another cpu is holding lock, requeue & delay xmits for
133 __this_cpu_inc(softnet_data.cpu_collision);
134 ret = dev_requeue_skb(skb, q);
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.
145 * Returns to the caller:
146 * 0 - queue is empty or throttled.
147 * >0 - queue is not empty.
149 int 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)
153 int ret = NETDEV_TX_BUSY;
155 /* And release qdisc */
156 spin_unlock(root_lock);
158 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
160 skb = validate_xmit_skb_list(skb, dev);
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);
167 HARD_TX_UNLOCK(dev, txq);
169 spin_lock(root_lock);
170 return qdisc_qlen(q);
172 spin_lock(root_lock);
174 if (dev_xmit_complete(ret)) {
175 /* Driver sent out skb successfully or skb was consumed */
177 } else if (ret == NETDEV_TX_LOCKED) {
178 /* Driver try lock failed */
179 ret = handle_dev_cpu_collision(skb, txq, q);
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);
186 ret = dev_requeue_skb(skb, q);
189 if (ret && netif_xmit_frozen_or_stopped(txq))
196 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
198 * __QDISC___STATE_RUNNING guarantees only one CPU can process
199 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
202 * netif_tx_lock serializes accesses to device driver.
204 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
205 * if one is grabbed, another must be free.
207 * Note, that this procedure can be called by a watchdog timer
209 * Returns to the caller:
210 * 0 - queue is empty or throttled.
211 * >0 - queue is not empty.
214 static inline int qdisc_restart(struct Qdisc *q, int *packets)
216 struct netdev_queue *txq;
217 struct net_device *dev;
218 spinlock_t *root_lock;
223 skb = dequeue_skb(q, &validate, packets);
227 root_lock = qdisc_lock(q);
229 txq = skb_get_tx_queue(dev, skb);
231 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
234 void __qdisc_run(struct Qdisc *q)
236 int quota = weight_p;
239 while (qdisc_restart(q, &packets)) {
241 * Ordered by possible occurrence: Postpone processing if
242 * 1. we've exceeded packet quota
243 * 2. another process needs the CPU;
246 if (quota <= 0 || need_resched()) {
255 unsigned long dev_trans_start(struct net_device *dev)
257 unsigned long val, res;
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))
268 dev->trans_start = res;
272 EXPORT_SYMBOL(dev_trans_start);
274 static void dev_watchdog(unsigned long arg)
276 struct net_device *dev = (struct net_device *)arg;
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;
285 unsigned long trans_start;
287 for (i = 0; i < dev->num_tx_queues; i++) {
288 struct netdev_queue *txq;
290 txq = netdev_get_tx_queue(dev, i);
292 * old device drivers set dev->trans_start
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++;
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);
309 if (!mod_timer(&dev->watchdog_timer,
310 round_jiffies(jiffies +
311 dev->watchdog_timeo)))
315 netif_tx_unlock(dev);
320 void __netdev_watchdog_up(struct net_device *dev)
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)))
331 static void dev_watchdog_up(struct net_device *dev)
333 __netdev_watchdog_up(dev);
336 static void dev_watchdog_down(struct net_device *dev)
338 netif_tx_lock_bh(dev);
339 if (del_timer(&dev->watchdog_timer))
341 netif_tx_unlock_bh(dev);
345 * netif_carrier_on - set carrier
346 * @dev: network device
348 * Device has detected that carrier.
350 void netif_carrier_on(struct net_device *dev)
352 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
353 if (dev->reg_state == NETREG_UNINITIALIZED)
355 atomic_inc(&dev->carrier_changes);
356 linkwatch_fire_event(dev);
357 if (netif_running(dev))
358 __netdev_watchdog_up(dev);
361 EXPORT_SYMBOL(netif_carrier_on);
364 * netif_carrier_off - clear carrier
365 * @dev: network device
367 * Device has detected loss of carrier.
369 void netif_carrier_off(struct net_device *dev)
371 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
372 if (dev->reg_state == NETREG_UNINITIALIZED)
374 atomic_inc(&dev->carrier_changes);
375 linkwatch_fire_event(dev);
378 EXPORT_SYMBOL(netif_carrier_off);
380 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
381 under all circumstances. It is difficult to invent anything faster or
385 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
391 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
396 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
399 .enqueue = noop_enqueue,
400 .dequeue = noop_dequeue,
401 .peek = noop_dequeue,
402 .owner = THIS_MODULE,
405 static struct netdev_queue noop_netdev_queue = {
406 .qdisc = &noop_qdisc,
407 .qdisc_sleeping = &noop_qdisc,
410 struct 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),
420 EXPORT_SYMBOL(noop_qdisc);
422 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
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;
431 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
434 .init = noqueue_init,
435 .enqueue = noop_enqueue,
436 .dequeue = noop_dequeue,
437 .peek = noop_dequeue,
438 .owner = THIS_MODULE,
441 static const u8 prio2band[TC_PRIO_MAX + 1] = {
442 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
445 /* 3-band FIFO queue: old style, but should be a bit faster than
446 generic prio+fifo combination.
449 #define PFIFO_FAST_BANDS 3
452 * Private data for a pfifo_fast scheduler containing:
453 * - queues for the three band
454 * - bitmap indicating which of the bands contain skbs
456 struct pfifo_fast_priv {
458 struct sk_buff_head q[PFIFO_FAST_BANDS];
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.
467 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
469 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
472 return priv->q + band;
475 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
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);
482 priv->bitmap |= (1 << band);
484 return __qdisc_enqueue_tail(skb, qdisc, list);
487 return qdisc_drop(skb, qdisc);
490 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493 int band = bitmap2band[priv->bitmap];
495 if (likely(band >= 0)) {
496 struct sk_buff_head *list = band2list(priv, band);
497 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
500 if (skb_queue_empty(list))
501 priv->bitmap &= ~(1 << band);
509 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
511 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
512 int band = bitmap2band[priv->bitmap];
515 struct sk_buff_head *list = band2list(priv, band);
517 return skb_peek(list);
523 static void pfifo_fast_reset(struct Qdisc *qdisc)
526 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
528 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
529 __qdisc_reset_queue(qdisc, band2list(priv, prio));
532 qdisc->qstats.backlog = 0;
536 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
538 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
540 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
541 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
542 goto nla_put_failure;
549 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
552 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
554 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
555 __skb_queue_head_init(band2list(priv, prio));
557 /* Can by-pass the queue discipline */
558 qdisc->flags |= TCQ_F_CAN_BYPASS;
562 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
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,
573 EXPORT_SYMBOL(pfifo_fast_ops);
575 static struct lock_class_key qdisc_tx_busylock;
577 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
578 const struct Qdisc_ops *ops)
582 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
584 struct net_device *dev = dev_queue->dev;
586 p = kzalloc_node(size, GFP_KERNEL,
587 netdev_queue_numa_node_read(dev_queue));
591 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
592 /* if we got non aligned memory, ask more and do alignment ourself */
595 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
596 netdev_queue_numa_node_read(dev_queue));
599 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
600 sch->padded = (char *) sch - (char *) p;
602 INIT_LIST_HEAD(&sch->list);
603 skb_queue_head_init(&sch->q);
605 spin_lock_init(&sch->busylock);
606 lockdep_set_class(&sch->busylock,
607 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
610 sch->enqueue = ops->enqueue;
611 sch->dequeue = ops->dequeue;
612 sch->dev_queue = dev_queue;
614 atomic_set(&sch->refcnt, 1);
621 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
622 const struct Qdisc_ops *ops,
623 unsigned int parentid)
627 if (!try_module_get(ops->owner))
630 sch = qdisc_alloc(dev_queue, ops);
633 sch->parent = parentid;
635 if (!ops->init || ops->init(sch, NULL) == 0)
642 EXPORT_SYMBOL(qdisc_create_dflt);
644 /* Under qdisc_lock(qdisc) and BH! */
646 void qdisc_reset(struct Qdisc *qdisc)
648 const struct Qdisc_ops *ops = qdisc->ops;
653 if (qdisc->gso_skb) {
654 kfree_skb_list(qdisc->gso_skb);
655 qdisc->gso_skb = NULL;
659 EXPORT_SYMBOL(qdisc_reset);
661 static void qdisc_rcu_free(struct rcu_head *head)
663 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
665 if (qdisc_is_percpu_stats(qdisc)) {
666 free_percpu(qdisc->cpu_bstats);
667 free_percpu(qdisc->cpu_qstats);
670 kfree((char *) qdisc - qdisc->padded);
673 void qdisc_destroy(struct Qdisc *qdisc)
675 const struct Qdisc_ops *ops = qdisc->ops;
677 if (qdisc->flags & TCQ_F_BUILTIN ||
678 !atomic_dec_and_test(&qdisc->refcnt))
681 #ifdef CONFIG_NET_SCHED
682 qdisc_list_del(qdisc);
684 qdisc_put_stab(rtnl_dereference(qdisc->stab));
686 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
692 module_put(ops->owner);
693 dev_put(qdisc_dev(qdisc));
695 kfree_skb_list(qdisc->gso_skb);
697 * gen_estimator est_timer() might access qdisc->q.lock,
698 * wait a RCU grace period before freeing qdisc.
700 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
702 EXPORT_SYMBOL(qdisc_destroy);
704 /* Attach toplevel qdisc to device queue. */
705 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
708 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
709 spinlock_t *root_lock;
711 root_lock = qdisc_lock(oqdisc);
712 spin_lock_bh(root_lock);
714 /* Prune old scheduler */
715 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
718 /* ... and graft new one */
721 dev_queue->qdisc_sleeping = qdisc;
722 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
724 spin_unlock_bh(root_lock);
728 EXPORT_SYMBOL(dev_graft_qdisc);
730 static void attach_one_default_qdisc(struct net_device *dev,
731 struct netdev_queue *dev_queue,
735 const struct Qdisc_ops *ops = default_qdisc_ops;
737 if (dev->priv_flags & IFF_NO_QUEUE)
738 ops = &noqueue_qdisc_ops;
740 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
742 netdev_info(dev, "activation failed\n");
745 if (!netif_is_multiqueue(dev))
746 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
747 dev_queue->qdisc_sleeping = qdisc;
750 static void attach_default_qdiscs(struct net_device *dev)
752 struct netdev_queue *txq;
755 txq = netdev_get_tx_queue(dev, 0);
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);
763 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
766 qdisc->ops->attach(qdisc);
771 static void transition_one_qdisc(struct net_device *dev,
772 struct netdev_queue *dev_queue,
773 void *_need_watchdog)
775 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
776 int *need_watchdog_p = _need_watchdog;
778 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
779 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
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;
788 void dev_activate(struct net_device *dev)
792 /* No queueing discipline is attached to device;
793 * create default one for devices, which need queueing
794 * and noqueue_qdisc for virtual interfaces
797 if (dev->qdisc == &noop_qdisc)
798 attach_default_qdiscs(dev);
800 if (!netif_carrier_ok(dev))
801 /* Delay activation until next carrier-on event */
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);
810 dev->trans_start = jiffies;
811 dev_watchdog_up(dev);
814 EXPORT_SYMBOL(dev_activate);
816 static void dev_deactivate_queue(struct net_device *dev,
817 struct netdev_queue *dev_queue,
818 void *_qdisc_default)
820 struct Qdisc *qdisc_default = _qdisc_default;
823 qdisc = rtnl_dereference(dev_queue->qdisc);
825 spin_lock_bh(qdisc_lock(qdisc));
827 if (!(qdisc->flags & TCQ_F_BUILTIN))
828 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
830 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
833 spin_unlock_bh(qdisc_lock(qdisc));
837 static bool some_qdisc_is_busy(struct net_device *dev)
841 for (i = 0; i < dev->num_tx_queues; i++) {
842 struct netdev_queue *dev_queue;
843 spinlock_t *root_lock;
847 dev_queue = netdev_get_tx_queue(dev, i);
848 q = dev_queue->qdisc_sleeping;
849 root_lock = qdisc_lock(q);
851 spin_lock_bh(root_lock);
853 val = (qdisc_is_running(q) ||
854 test_bit(__QDISC_STATE_SCHED, &q->state));
856 spin_unlock_bh(root_lock);
865 * dev_deactivate_many - deactivate transmissions on several devices
866 * @head: list of devices to deactivate
868 * This function returns only when all outstanding transmissions
869 * have completed, unless all devices are in dismantle phase.
871 void dev_deactivate_many(struct list_head *head)
873 struct net_device *dev;
874 bool sync_needed = false;
876 list_for_each_entry(dev, head, close_list) {
877 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
879 if (dev_ingress_queue(dev))
880 dev_deactivate_queue(dev, dev_ingress_queue(dev),
883 dev_watchdog_down(dev);
884 sync_needed |= !dev->dismantle;
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
894 /* Wait for outstanding qdisc_run calls. */
895 list_for_each_entry(dev, head, close_list)
896 while (some_qdisc_is_busy(dev))
900 void dev_deactivate(struct net_device *dev)
904 list_add(&dev->close_list, &single);
905 dev_deactivate_many(&single);
908 EXPORT_SYMBOL(dev_deactivate);
910 static void dev_init_scheduler_queue(struct net_device *dev,
911 struct netdev_queue *dev_queue,
914 struct Qdisc *qdisc = _qdisc;
916 rcu_assign_pointer(dev_queue->qdisc, qdisc);
917 dev_queue->qdisc_sleeping = qdisc;
920 void dev_init_scheduler(struct net_device *dev)
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);
927 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
930 static void shutdown_scheduler_queue(struct net_device *dev,
931 struct netdev_queue *dev_queue,
932 void *_qdisc_default)
934 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
935 struct Qdisc *qdisc_default = _qdisc_default;
938 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
939 dev_queue->qdisc_sleeping = qdisc_default;
941 qdisc_destroy(qdisc);
945 void dev_shutdown(struct net_device *dev)
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;
953 WARN_ON(timer_pending(&dev->watchdog_timer));
956 void psched_ratecfg_precompute(struct psched_ratecfg *r,
957 const struct tc_ratespec *conf,
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);
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)
969 * Normal formula would be :
970 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
972 * We compute mult/shift to use instead :
973 * time_in_ns = (len * mult) >> shift;
975 * We try to get the highest possible mult value for accuracy,
976 * but have to make sure no overflows will ever happen.
978 if (r->rate_bytes_ps > 0) {
979 u64 factor = NSEC_PER_SEC;
982 r->mult = div64_u64(factor, r->rate_bytes_ps);
983 if (r->mult & (1U << 31) || factor & (1ULL << 63))
990 EXPORT_SYMBOL(psched_ratecfg_precompute);