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 qdisc_qstats_backlog_inc(q, skb);
53 q->q.qlen++; /* it's still part of the queue */
59 static void try_bulk_dequeue_skb(struct Qdisc *q,
61 const struct netdev_queue *txq,
64 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
66 while (bytelimit > 0) {
67 struct sk_buff *nskb = q->dequeue(q);
72 bytelimit -= nskb->len; /* covers GSO len */
75 (*packets)++; /* GSO counts as one pkt */
80 /* This variant of try_bulk_dequeue_skb() makes sure
81 * all skbs in the chain are for the same txq
83 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
87 int mapping = skb_get_queue_mapping(skb);
95 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
96 q->skb_bad_txq = nskb;
97 qdisc_qstats_backlog_inc(q, nskb);
108 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
109 * A requeued skb (via q->gso_skb) can also be a SKB list.
111 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
114 struct sk_buff *skb = q->gso_skb;
115 const struct netdev_queue *txq = q->dev_queue;
119 /* skb in gso_skb were already validated */
121 /* check the reason of requeuing without tx lock first */
122 txq = skb_get_tx_queue(txq->dev, skb);
123 if (!netif_xmit_frozen_or_stopped(txq)) {
125 qdisc_qstats_backlog_dec(q, skb);
132 skb = q->skb_bad_txq;
134 /* check the reason of requeuing without tx lock first */
135 txq = skb_get_tx_queue(txq->dev, skb);
136 if (!netif_xmit_frozen_or_stopped(txq)) {
137 q->skb_bad_txq = NULL;
138 qdisc_qstats_backlog_dec(q, skb);
144 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
145 !netif_xmit_frozen_or_stopped(txq))
149 if (qdisc_may_bulk(q))
150 try_bulk_dequeue_skb(q, skb, txq, packets);
152 try_bulk_dequeue_skb_slow(q, skb, packets);
158 * Transmit possibly several skbs, and handle the return status as
159 * required. Owning running seqcount bit guarantees that
160 * only one CPU can execute this function.
162 * Returns to the caller:
163 * 0 - queue is empty or throttled.
164 * >0 - queue is not empty.
166 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
167 struct net_device *dev, struct netdev_queue *txq,
168 spinlock_t *root_lock, bool validate)
170 int ret = NETDEV_TX_BUSY;
172 /* And release qdisc */
173 spin_unlock(root_lock);
175 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
177 skb = validate_xmit_skb_list(skb, dev);
180 HARD_TX_LOCK(dev, txq, smp_processor_id());
181 if (!netif_xmit_frozen_or_stopped(txq))
182 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
184 HARD_TX_UNLOCK(dev, txq);
186 spin_lock(root_lock);
187 return qdisc_qlen(q);
189 spin_lock(root_lock);
191 if (dev_xmit_complete(ret)) {
192 /* Driver sent out skb successfully or skb was consumed */
195 /* Driver returned NETDEV_TX_BUSY - requeue skb */
196 if (unlikely(ret != NETDEV_TX_BUSY))
197 net_warn_ratelimited("BUG %s code %d qlen %d\n",
198 dev->name, ret, q->q.qlen);
200 ret = dev_requeue_skb(skb, q);
203 if (ret && netif_xmit_frozen_or_stopped(txq))
210 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
212 * running seqcount guarantees only one CPU can process
213 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
216 * netif_tx_lock serializes accesses to device driver.
218 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
219 * if one is grabbed, another must be free.
221 * Note, that this procedure can be called by a watchdog timer
223 * Returns to the caller:
224 * 0 - queue is empty or throttled.
225 * >0 - queue is not empty.
228 static inline int qdisc_restart(struct Qdisc *q, int *packets)
230 struct netdev_queue *txq;
231 struct net_device *dev;
232 spinlock_t *root_lock;
237 skb = dequeue_skb(q, &validate, packets);
241 root_lock = qdisc_lock(q);
243 txq = skb_get_tx_queue(dev, skb);
245 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
248 void __qdisc_run(struct Qdisc *q)
250 int quota = weight_p;
253 while (qdisc_restart(q, &packets)) {
255 * Ordered by possible occurrence: Postpone processing if
256 * 1. we've exceeded packet quota
257 * 2. another process needs the CPU;
260 if (quota <= 0 || need_resched()) {
269 unsigned long dev_trans_start(struct net_device *dev)
271 unsigned long val, res;
274 if (is_vlan_dev(dev))
275 dev = vlan_dev_real_dev(dev);
276 res = netdev_get_tx_queue(dev, 0)->trans_start;
277 for (i = 1; i < dev->num_tx_queues; i++) {
278 val = netdev_get_tx_queue(dev, i)->trans_start;
279 if (val && time_after(val, res))
285 EXPORT_SYMBOL(dev_trans_start);
287 static void dev_watchdog(unsigned long arg)
289 struct net_device *dev = (struct net_device *)arg;
292 if (!qdisc_tx_is_noop(dev)) {
293 if (netif_device_present(dev) &&
294 netif_running(dev) &&
295 netif_carrier_ok(dev)) {
296 int some_queue_timedout = 0;
298 unsigned long trans_start;
300 for (i = 0; i < dev->num_tx_queues; i++) {
301 struct netdev_queue *txq;
303 txq = netdev_get_tx_queue(dev, i);
304 trans_start = txq->trans_start;
305 if (netif_xmit_stopped(txq) &&
306 time_after(jiffies, (trans_start +
307 dev->watchdog_timeo))) {
308 some_queue_timedout = 1;
309 txq->trans_timeout++;
314 if (some_queue_timedout) {
315 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
316 dev->name, netdev_drivername(dev), i);
317 dev->netdev_ops->ndo_tx_timeout(dev);
319 if (!mod_timer(&dev->watchdog_timer,
320 round_jiffies(jiffies +
321 dev->watchdog_timeo)))
325 netif_tx_unlock(dev);
330 void __netdev_watchdog_up(struct net_device *dev)
332 if (dev->netdev_ops->ndo_tx_timeout) {
333 if (dev->watchdog_timeo <= 0)
334 dev->watchdog_timeo = 5*HZ;
335 if (!mod_timer(&dev->watchdog_timer,
336 round_jiffies(jiffies + dev->watchdog_timeo)))
340 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
342 static void dev_watchdog_up(struct net_device *dev)
344 __netdev_watchdog_up(dev);
347 static void dev_watchdog_down(struct net_device *dev)
349 netif_tx_lock_bh(dev);
350 if (del_timer(&dev->watchdog_timer))
352 netif_tx_unlock_bh(dev);
356 * netif_carrier_on - set carrier
357 * @dev: network device
359 * Device has detected that carrier.
361 void netif_carrier_on(struct net_device *dev)
363 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
364 if (dev->reg_state == NETREG_UNINITIALIZED)
366 atomic_inc(&dev->carrier_changes);
367 linkwatch_fire_event(dev);
368 if (netif_running(dev))
369 __netdev_watchdog_up(dev);
372 EXPORT_SYMBOL(netif_carrier_on);
375 * netif_carrier_off - clear carrier
376 * @dev: network device
378 * Device has detected loss of carrier.
380 void netif_carrier_off(struct net_device *dev)
382 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
383 if (dev->reg_state == NETREG_UNINITIALIZED)
385 atomic_inc(&dev->carrier_changes);
386 linkwatch_fire_event(dev);
389 EXPORT_SYMBOL(netif_carrier_off);
391 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
392 under all circumstances. It is difficult to invent anything faster or
396 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
397 struct sk_buff **to_free)
399 __qdisc_drop(skb, to_free);
403 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
408 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
411 .enqueue = noop_enqueue,
412 .dequeue = noop_dequeue,
413 .peek = noop_dequeue,
414 .owner = THIS_MODULE,
417 static struct netdev_queue noop_netdev_queue = {
418 .qdisc = &noop_qdisc,
419 .qdisc_sleeping = &noop_qdisc,
422 struct Qdisc noop_qdisc = {
423 .enqueue = noop_enqueue,
424 .dequeue = noop_dequeue,
425 .flags = TCQ_F_BUILTIN,
426 .ops = &noop_qdisc_ops,
427 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
428 .dev_queue = &noop_netdev_queue,
429 .running = SEQCNT_ZERO(noop_qdisc.running),
430 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
432 EXPORT_SYMBOL(noop_qdisc);
434 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
436 /* register_qdisc() assigns a default of noop_enqueue if unset,
437 * but __dev_queue_xmit() treats noqueue only as such
438 * if this is NULL - so clear it here. */
439 qdisc->enqueue = NULL;
443 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
446 .init = noqueue_init,
447 .enqueue = noop_enqueue,
448 .dequeue = noop_dequeue,
449 .peek = noop_dequeue,
450 .owner = THIS_MODULE,
453 static const u8 prio2band[TC_PRIO_MAX + 1] = {
454 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
457 /* 3-band FIFO queue: old style, but should be a bit faster than
458 generic prio+fifo combination.
461 #define PFIFO_FAST_BANDS 3
464 * Private data for a pfifo_fast scheduler containing:
465 * - queues for the three band
466 * - bitmap indicating which of the bands contain skbs
468 struct pfifo_fast_priv {
470 struct qdisc_skb_head q[PFIFO_FAST_BANDS];
474 * Convert a bitmap to the first band number where an skb is queued, where:
475 * bitmap=0 means there are no skbs on any band.
476 * bitmap=1 means there is an skb on band 0.
477 * bitmap=7 means there are skbs on all 3 bands, etc.
479 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
481 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
484 return priv->q + band;
487 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
488 struct sk_buff **to_free)
490 if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
491 int band = prio2band[skb->priority & TC_PRIO_MAX];
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493 struct qdisc_skb_head *list = band2list(priv, band);
495 priv->bitmap |= (1 << band);
497 return __qdisc_enqueue_tail(skb, qdisc, list);
500 return qdisc_drop(skb, qdisc, to_free);
503 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
505 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
506 int band = bitmap2band[priv->bitmap];
508 if (likely(band >= 0)) {
509 struct qdisc_skb_head *qh = band2list(priv, band);
510 struct sk_buff *skb = __qdisc_dequeue_head(qh);
512 if (likely(skb != NULL)) {
513 qdisc_qstats_backlog_dec(qdisc, skb);
514 qdisc_bstats_update(qdisc, skb);
519 priv->bitmap &= ~(1 << band);
527 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
529 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
530 int band = bitmap2band[priv->bitmap];
533 struct qdisc_skb_head *qh = band2list(priv, band);
541 static void pfifo_fast_reset(struct Qdisc *qdisc)
544 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
546 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
547 __qdisc_reset_queue(band2list(priv, prio));
550 qdisc->qstats.backlog = 0;
554 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
556 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
558 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
559 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
560 goto nla_put_failure;
567 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
570 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
572 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
573 qdisc_skb_head_init(band2list(priv, prio));
575 /* Can by-pass the queue discipline */
576 qdisc->flags |= TCQ_F_CAN_BYPASS;
580 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
582 .priv_size = sizeof(struct pfifo_fast_priv),
583 .enqueue = pfifo_fast_enqueue,
584 .dequeue = pfifo_fast_dequeue,
585 .peek = pfifo_fast_peek,
586 .init = pfifo_fast_init,
587 .reset = pfifo_fast_reset,
588 .dump = pfifo_fast_dump,
589 .owner = THIS_MODULE,
591 EXPORT_SYMBOL(pfifo_fast_ops);
593 static struct lock_class_key qdisc_tx_busylock;
594 static struct lock_class_key qdisc_running_key;
596 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
597 const struct Qdisc_ops *ops)
601 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
603 struct net_device *dev = dev_queue->dev;
605 p = kzalloc_node(size, GFP_KERNEL,
606 netdev_queue_numa_node_read(dev_queue));
610 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
611 /* if we got non aligned memory, ask more and do alignment ourself */
614 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
615 netdev_queue_numa_node_read(dev_queue));
618 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
619 sch->padded = (char *) sch - (char *) p;
621 qdisc_skb_head_init(&sch->q);
622 spin_lock_init(&sch->q.lock);
624 spin_lock_init(&sch->busylock);
625 lockdep_set_class(&sch->busylock,
626 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
628 seqcount_init(&sch->running);
629 lockdep_set_class(&sch->running,
630 dev->qdisc_running_key ?: &qdisc_running_key);
633 sch->enqueue = ops->enqueue;
634 sch->dequeue = ops->dequeue;
635 sch->dev_queue = dev_queue;
637 atomic_set(&sch->refcnt, 1);
644 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
645 const struct Qdisc_ops *ops,
646 unsigned int parentid)
650 if (!try_module_get(ops->owner))
653 sch = qdisc_alloc(dev_queue, ops);
655 module_put(ops->owner);
658 sch->parent = parentid;
660 if (!ops->init || ops->init(sch, NULL) == 0)
666 EXPORT_SYMBOL(qdisc_create_dflt);
668 /* Under qdisc_lock(qdisc) and BH! */
670 void qdisc_reset(struct Qdisc *qdisc)
672 const struct Qdisc_ops *ops = qdisc->ops;
677 kfree_skb(qdisc->skb_bad_txq);
678 qdisc->skb_bad_txq = NULL;
680 if (qdisc->gso_skb) {
681 kfree_skb_list(qdisc->gso_skb);
682 qdisc->gso_skb = NULL;
685 qdisc->qstats.backlog = 0;
687 EXPORT_SYMBOL(qdisc_reset);
689 static void qdisc_rcu_free(struct rcu_head *head)
691 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
693 if (qdisc_is_percpu_stats(qdisc)) {
694 free_percpu(qdisc->cpu_bstats);
695 free_percpu(qdisc->cpu_qstats);
698 kfree((char *) qdisc - qdisc->padded);
701 void qdisc_destroy(struct Qdisc *qdisc)
703 const struct Qdisc_ops *ops;
709 if (qdisc->flags & TCQ_F_BUILTIN ||
710 !atomic_dec_and_test(&qdisc->refcnt))
713 #ifdef CONFIG_NET_SCHED
714 qdisc_hash_del(qdisc);
716 qdisc_put_stab(rtnl_dereference(qdisc->stab));
718 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
724 module_put(ops->owner);
725 dev_put(qdisc_dev(qdisc));
727 kfree_skb_list(qdisc->gso_skb);
728 kfree_skb(qdisc->skb_bad_txq);
730 * gen_estimator est_timer() might access qdisc->q.lock,
731 * wait a RCU grace period before freeing qdisc.
733 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
735 EXPORT_SYMBOL(qdisc_destroy);
737 /* Attach toplevel qdisc to device queue. */
738 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
741 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
742 spinlock_t *root_lock;
744 root_lock = qdisc_lock(oqdisc);
745 spin_lock_bh(root_lock);
747 /* Prune old scheduler */
748 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
751 /* ... and graft new one */
754 dev_queue->qdisc_sleeping = qdisc;
755 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
757 spin_unlock_bh(root_lock);
761 EXPORT_SYMBOL(dev_graft_qdisc);
763 static void attach_one_default_qdisc(struct net_device *dev,
764 struct netdev_queue *dev_queue,
768 const struct Qdisc_ops *ops = default_qdisc_ops;
770 if (dev->priv_flags & IFF_NO_QUEUE)
771 ops = &noqueue_qdisc_ops;
773 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
775 netdev_info(dev, "activation failed\n");
778 if (!netif_is_multiqueue(dev))
779 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
780 dev_queue->qdisc_sleeping = qdisc;
783 static void attach_default_qdiscs(struct net_device *dev)
785 struct netdev_queue *txq;
788 txq = netdev_get_tx_queue(dev, 0);
790 if (!netif_is_multiqueue(dev) ||
791 dev->priv_flags & IFF_NO_QUEUE) {
792 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
793 dev->qdisc = txq->qdisc_sleeping;
794 atomic_inc(&dev->qdisc->refcnt);
796 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
799 qdisc->ops->attach(qdisc);
802 #ifdef CONFIG_NET_SCHED
804 qdisc_hash_add(dev->qdisc);
808 static void transition_one_qdisc(struct net_device *dev,
809 struct netdev_queue *dev_queue,
810 void *_need_watchdog)
812 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
813 int *need_watchdog_p = _need_watchdog;
815 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
816 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
818 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
819 if (need_watchdog_p) {
820 dev_queue->trans_start = 0;
821 *need_watchdog_p = 1;
825 void dev_activate(struct net_device *dev)
829 /* No queueing discipline is attached to device;
830 * create default one for devices, which need queueing
831 * and noqueue_qdisc for virtual interfaces
834 if (dev->qdisc == &noop_qdisc)
835 attach_default_qdiscs(dev);
837 if (!netif_carrier_ok(dev))
838 /* Delay activation until next carrier-on event */
842 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
843 if (dev_ingress_queue(dev))
844 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
847 netif_trans_update(dev);
848 dev_watchdog_up(dev);
851 EXPORT_SYMBOL(dev_activate);
853 static void dev_deactivate_queue(struct net_device *dev,
854 struct netdev_queue *dev_queue,
855 void *_qdisc_default)
857 struct Qdisc *qdisc_default = _qdisc_default;
860 qdisc = rtnl_dereference(dev_queue->qdisc);
862 spin_lock_bh(qdisc_lock(qdisc));
864 if (!(qdisc->flags & TCQ_F_BUILTIN))
865 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
867 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
870 spin_unlock_bh(qdisc_lock(qdisc));
874 static bool some_qdisc_is_busy(struct net_device *dev)
878 for (i = 0; i < dev->num_tx_queues; i++) {
879 struct netdev_queue *dev_queue;
880 spinlock_t *root_lock;
884 dev_queue = netdev_get_tx_queue(dev, i);
885 q = dev_queue->qdisc_sleeping;
886 root_lock = qdisc_lock(q);
888 spin_lock_bh(root_lock);
890 val = (qdisc_is_running(q) ||
891 test_bit(__QDISC_STATE_SCHED, &q->state));
893 spin_unlock_bh(root_lock);
902 * dev_deactivate_many - deactivate transmissions on several devices
903 * @head: list of devices to deactivate
905 * This function returns only when all outstanding transmissions
906 * have completed, unless all devices are in dismantle phase.
908 void dev_deactivate_many(struct list_head *head)
910 struct net_device *dev;
911 bool sync_needed = false;
913 list_for_each_entry(dev, head, close_list) {
914 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
916 if (dev_ingress_queue(dev))
917 dev_deactivate_queue(dev, dev_ingress_queue(dev),
920 dev_watchdog_down(dev);
921 sync_needed |= !dev->dismantle;
924 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
925 * This is avoided if all devices are in dismantle phase :
926 * Caller will call synchronize_net() for us
931 /* Wait for outstanding qdisc_run calls. */
932 list_for_each_entry(dev, head, close_list)
933 while (some_qdisc_is_busy(dev))
937 void dev_deactivate(struct net_device *dev)
941 list_add(&dev->close_list, &single);
942 dev_deactivate_many(&single);
945 EXPORT_SYMBOL(dev_deactivate);
947 static void dev_init_scheduler_queue(struct net_device *dev,
948 struct netdev_queue *dev_queue,
951 struct Qdisc *qdisc = _qdisc;
953 rcu_assign_pointer(dev_queue->qdisc, qdisc);
954 dev_queue->qdisc_sleeping = qdisc;
957 void dev_init_scheduler(struct net_device *dev)
959 dev->qdisc = &noop_qdisc;
960 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
961 if (dev_ingress_queue(dev))
962 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
964 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
967 static void shutdown_scheduler_queue(struct net_device *dev,
968 struct netdev_queue *dev_queue,
969 void *_qdisc_default)
971 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
972 struct Qdisc *qdisc_default = _qdisc_default;
975 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
976 dev_queue->qdisc_sleeping = qdisc_default;
978 qdisc_destroy(qdisc);
982 void dev_shutdown(struct net_device *dev)
984 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
985 if (dev_ingress_queue(dev))
986 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
987 qdisc_destroy(dev->qdisc);
988 dev->qdisc = &noop_qdisc;
990 WARN_ON(timer_pending(&dev->watchdog_timer));
993 void psched_ratecfg_precompute(struct psched_ratecfg *r,
994 const struct tc_ratespec *conf,
997 memset(r, 0, sizeof(*r));
998 r->overhead = conf->overhead;
1000 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1001 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1004 * The deal here is to replace a divide by a reciprocal one
1005 * in fast path (a reciprocal divide is a multiply and a shift)
1007 * Normal formula would be :
1008 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1010 * We compute mult/shift to use instead :
1011 * time_in_ns = (len * mult) >> shift;
1013 * We try to get the highest possible mult value for accuracy,
1014 * but have to make sure no overflows will ever happen.
1016 if (r->rate_bytes_ps > 0) {
1017 u64 factor = NSEC_PER_SEC;
1020 r->mult = div64_u64(factor, r->rate_bytes_ps);
1021 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1028 EXPORT_SYMBOL(psched_ratecfg_precompute);