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 <linux/skb_array.h>
30 #include <linux/if_macvlan.h>
31 #include <net/sch_generic.h>
32 #include <net/pkt_sched.h>
34 #include <trace/events/qdisc.h>
37 /* Qdisc to use by default */
38 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
39 EXPORT_SYMBOL(default_qdisc_ops);
41 /* Main transmission queue. */
43 /* Modifications to data participating in scheduling must be protected with
44 * qdisc_lock(qdisc) spinlock.
46 * The idea is the following:
47 * - enqueue, dequeue are serialized via qdisc root lock
48 * - ingress filtering is also serialized via qdisc root lock
49 * - updates to tree and tree walking are only done under the rtnl mutex.
52 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
54 const struct netdev_queue *txq = q->dev_queue;
55 spinlock_t *lock = NULL;
58 if (q->flags & TCQ_F_NOLOCK) {
63 skb = skb_peek(&q->skb_bad_txq);
65 /* check the reason of requeuing without tx lock first */
66 txq = skb_get_tx_queue(txq->dev, skb);
67 if (!netif_xmit_frozen_or_stopped(txq)) {
68 skb = __skb_dequeue(&q->skb_bad_txq);
69 if (qdisc_is_percpu_stats(q)) {
70 qdisc_qstats_cpu_backlog_dec(q, skb);
71 qdisc_qstats_cpu_qlen_dec(q);
73 qdisc_qstats_backlog_dec(q, skb);
87 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
89 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
92 skb = __skb_dequeue_bad_txq(q);
97 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
100 spinlock_t *lock = NULL;
102 if (q->flags & TCQ_F_NOLOCK) {
103 lock = qdisc_lock(q);
107 __skb_queue_tail(&q->skb_bad_txq, skb);
109 if (qdisc_is_percpu_stats(q)) {
110 qdisc_qstats_cpu_backlog_inc(q, skb);
111 qdisc_qstats_cpu_qlen_inc(q);
113 qdisc_qstats_backlog_inc(q, skb);
121 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
123 spinlock_t *lock = NULL;
125 if (q->flags & TCQ_F_NOLOCK) {
126 lock = qdisc_lock(q);
131 struct sk_buff *next = skb->next;
133 __skb_queue_tail(&q->gso_skb, skb);
135 /* it's still part of the queue */
136 if (qdisc_is_percpu_stats(q)) {
137 qdisc_qstats_cpu_requeues_inc(q);
138 qdisc_qstats_cpu_backlog_inc(q, skb);
139 qdisc_qstats_cpu_qlen_inc(q);
141 q->qstats.requeues++;
142 qdisc_qstats_backlog_inc(q, skb);
153 static void try_bulk_dequeue_skb(struct Qdisc *q,
155 const struct netdev_queue *txq,
158 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
160 while (bytelimit > 0) {
161 struct sk_buff *nskb = q->dequeue(q);
166 bytelimit -= nskb->len; /* covers GSO len */
169 (*packets)++; /* GSO counts as one pkt */
171 skb_mark_not_on_list(skb);
174 /* This variant of try_bulk_dequeue_skb() makes sure
175 * all skbs in the chain are for the same txq
177 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
181 int mapping = skb_get_queue_mapping(skb);
182 struct sk_buff *nskb;
186 nskb = q->dequeue(q);
189 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
190 qdisc_enqueue_skb_bad_txq(q, nskb);
197 skb_mark_not_on_list(skb);
200 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
201 * A requeued skb (via q->gso_skb) can also be a SKB list.
203 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
206 const struct netdev_queue *txq = q->dev_queue;
207 struct sk_buff *skb = NULL;
210 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
211 spinlock_t *lock = NULL;
213 if (q->flags & TCQ_F_NOLOCK) {
214 lock = qdisc_lock(q);
218 skb = skb_peek(&q->gso_skb);
220 /* skb may be null if another cpu pulls gso_skb off in between
221 * empty check and lock.
229 /* skb in gso_skb were already validated */
231 if (xfrm_offload(skb))
233 /* check the reason of requeuing without tx lock first */
234 txq = skb_get_tx_queue(txq->dev, skb);
235 if (!netif_xmit_frozen_or_stopped(txq)) {
236 skb = __skb_dequeue(&q->gso_skb);
237 if (qdisc_is_percpu_stats(q)) {
238 qdisc_qstats_cpu_backlog_dec(q, skb);
239 qdisc_qstats_cpu_qlen_dec(q);
241 qdisc_qstats_backlog_dec(q, skb);
254 if ((q->flags & TCQ_F_ONETXQUEUE) &&
255 netif_xmit_frozen_or_stopped(txq))
258 skb = qdisc_dequeue_skb_bad_txq(q);
264 if (qdisc_may_bulk(q))
265 try_bulk_dequeue_skb(q, skb, txq, packets);
267 try_bulk_dequeue_skb_slow(q, skb, packets);
270 trace_qdisc_dequeue(q, txq, *packets, skb);
275 * Transmit possibly several skbs, and handle the return status as
276 * required. Owning running seqcount bit guarantees that
277 * only one CPU can execute this function.
279 * Returns to the caller:
280 * false - hardware queue frozen backoff
281 * true - feel free to send more pkts
283 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
284 struct net_device *dev, struct netdev_queue *txq,
285 spinlock_t *root_lock, bool validate)
287 int ret = NETDEV_TX_BUSY;
290 /* And release qdisc */
292 spin_unlock(root_lock);
294 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
296 skb = validate_xmit_skb_list(skb, dev, &again);
298 #ifdef CONFIG_XFRM_OFFLOAD
299 if (unlikely(again)) {
301 spin_lock(root_lock);
303 dev_requeue_skb(skb, q);
309 HARD_TX_LOCK(dev, txq, smp_processor_id());
310 if (!netif_xmit_frozen_or_stopped(txq))
311 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
313 HARD_TX_UNLOCK(dev, txq);
316 spin_lock(root_lock);
321 spin_lock(root_lock);
323 if (!dev_xmit_complete(ret)) {
324 /* Driver returned NETDEV_TX_BUSY - requeue skb */
325 if (unlikely(ret != NETDEV_TX_BUSY))
326 net_warn_ratelimited("BUG %s code %d qlen %d\n",
327 dev->name, ret, q->q.qlen);
329 dev_requeue_skb(skb, q);
337 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
339 * running seqcount guarantees only one CPU can process
340 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
343 * netif_tx_lock serializes accesses to device driver.
345 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
346 * if one is grabbed, another must be free.
348 * Note, that this procedure can be called by a watchdog timer
350 * Returns to the caller:
351 * 0 - queue is empty or throttled.
352 * >0 - queue is not empty.
355 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
357 spinlock_t *root_lock = NULL;
358 struct netdev_queue *txq;
359 struct net_device *dev;
364 skb = dequeue_skb(q, &validate, packets);
368 if (!(q->flags & TCQ_F_NOLOCK))
369 root_lock = qdisc_lock(q);
372 txq = skb_get_tx_queue(dev, skb);
374 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
377 void __qdisc_run(struct Qdisc *q)
379 int quota = dev_tx_weight;
382 while (qdisc_restart(q, &packets)) {
384 * Ordered by possible occurrence: Postpone processing if
385 * 1. we've exceeded packet quota
386 * 2. another process needs the CPU;
389 if (quota <= 0 || need_resched()) {
396 unsigned long dev_trans_start(struct net_device *dev)
398 unsigned long val, res;
401 if (is_vlan_dev(dev))
402 dev = vlan_dev_real_dev(dev);
403 else if (netif_is_macvlan(dev))
404 dev = macvlan_dev_real_dev(dev);
405 res = netdev_get_tx_queue(dev, 0)->trans_start;
406 for (i = 1; i < dev->num_tx_queues; i++) {
407 val = netdev_get_tx_queue(dev, i)->trans_start;
408 if (val && time_after(val, res))
414 EXPORT_SYMBOL(dev_trans_start);
416 static void dev_watchdog(struct timer_list *t)
418 struct net_device *dev = from_timer(dev, t, watchdog_timer);
421 if (!qdisc_tx_is_noop(dev)) {
422 if (netif_device_present(dev) &&
423 netif_running(dev) &&
424 netif_carrier_ok(dev)) {
425 int some_queue_timedout = 0;
427 unsigned long trans_start;
429 for (i = 0; i < dev->num_tx_queues; i++) {
430 struct netdev_queue *txq;
432 txq = netdev_get_tx_queue(dev, i);
433 trans_start = txq->trans_start;
434 if (netif_xmit_stopped(txq) &&
435 time_after(jiffies, (trans_start +
436 dev->watchdog_timeo))) {
437 some_queue_timedout = 1;
438 txq->trans_timeout++;
443 if (some_queue_timedout) {
444 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
445 dev->name, netdev_drivername(dev), i);
446 dev->netdev_ops->ndo_tx_timeout(dev);
448 if (!mod_timer(&dev->watchdog_timer,
449 round_jiffies(jiffies +
450 dev->watchdog_timeo)))
454 netif_tx_unlock(dev);
459 void __netdev_watchdog_up(struct net_device *dev)
461 if (dev->netdev_ops->ndo_tx_timeout) {
462 if (dev->watchdog_timeo <= 0)
463 dev->watchdog_timeo = 5*HZ;
464 if (!mod_timer(&dev->watchdog_timer,
465 round_jiffies(jiffies + dev->watchdog_timeo)))
470 static void dev_watchdog_up(struct net_device *dev)
472 __netdev_watchdog_up(dev);
475 static void dev_watchdog_down(struct net_device *dev)
477 netif_tx_lock_bh(dev);
478 if (del_timer(&dev->watchdog_timer))
480 netif_tx_unlock_bh(dev);
484 * netif_carrier_on - set carrier
485 * @dev: network device
487 * Device has detected acquisition of carrier.
489 void netif_carrier_on(struct net_device *dev)
491 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
492 if (dev->reg_state == NETREG_UNINITIALIZED)
494 atomic_inc(&dev->carrier_up_count);
495 linkwatch_fire_event(dev);
496 if (netif_running(dev))
497 __netdev_watchdog_up(dev);
500 EXPORT_SYMBOL(netif_carrier_on);
503 * netif_carrier_off - clear carrier
504 * @dev: network device
506 * Device has detected loss of carrier.
508 void netif_carrier_off(struct net_device *dev)
510 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
511 if (dev->reg_state == NETREG_UNINITIALIZED)
513 atomic_inc(&dev->carrier_down_count);
514 linkwatch_fire_event(dev);
517 EXPORT_SYMBOL(netif_carrier_off);
519 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
520 under all circumstances. It is difficult to invent anything faster or
524 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
525 struct sk_buff **to_free)
527 __qdisc_drop(skb, to_free);
531 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
536 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
539 .enqueue = noop_enqueue,
540 .dequeue = noop_dequeue,
541 .peek = noop_dequeue,
542 .owner = THIS_MODULE,
545 static struct netdev_queue noop_netdev_queue = {
546 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
547 .qdisc_sleeping = &noop_qdisc,
550 struct Qdisc noop_qdisc = {
551 .enqueue = noop_enqueue,
552 .dequeue = noop_dequeue,
553 .flags = TCQ_F_BUILTIN,
554 .ops = &noop_qdisc_ops,
555 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
556 .dev_queue = &noop_netdev_queue,
557 .running = SEQCNT_ZERO(noop_qdisc.running),
558 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
560 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
561 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
563 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
566 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
567 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
569 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
572 EXPORT_SYMBOL(noop_qdisc);
574 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
575 struct netlink_ext_ack *extack)
577 /* register_qdisc() assigns a default of noop_enqueue if unset,
578 * but __dev_queue_xmit() treats noqueue only as such
579 * if this is NULL - so clear it here. */
580 qdisc->enqueue = NULL;
584 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
587 .init = noqueue_init,
588 .enqueue = noop_enqueue,
589 .dequeue = noop_dequeue,
590 .peek = noop_dequeue,
591 .owner = THIS_MODULE,
594 static const u8 prio2band[TC_PRIO_MAX + 1] = {
595 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
598 /* 3-band FIFO queue: old style, but should be a bit faster than
599 generic prio+fifo combination.
602 #define PFIFO_FAST_BANDS 3
605 * Private data for a pfifo_fast scheduler containing:
606 * - rings for priority bands
608 struct pfifo_fast_priv {
609 struct skb_array q[PFIFO_FAST_BANDS];
612 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
615 return &priv->q[band];
618 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
619 struct sk_buff **to_free)
621 int band = prio2band[skb->priority & TC_PRIO_MAX];
622 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
623 struct skb_array *q = band2list(priv, band);
624 unsigned int pkt_len = qdisc_pkt_len(skb);
627 err = skb_array_produce(q, skb);
630 return qdisc_drop_cpu(skb, qdisc, to_free);
632 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
633 return NET_XMIT_SUCCESS;
636 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
638 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
639 struct sk_buff *skb = NULL;
642 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
643 struct skb_array *q = band2list(priv, band);
645 if (__skb_array_empty(q))
648 skb = __skb_array_consume(q);
651 qdisc_update_stats_at_dequeue(qdisc, skb);
659 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
661 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
662 struct sk_buff *skb = NULL;
665 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
666 struct skb_array *q = band2list(priv, band);
668 skb = __skb_array_peek(q);
674 static void pfifo_fast_reset(struct Qdisc *qdisc)
677 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
679 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
680 struct skb_array *q = band2list(priv, band);
683 /* NULL ring is possible if destroy path is due to a failed
684 * skb_array_init() in pfifo_fast_init() case.
689 while ((skb = __skb_array_consume(q)) != NULL)
693 for_each_possible_cpu(i) {
694 struct gnet_stats_queue *q = per_cpu_ptr(qdisc->cpu_qstats, i);
701 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
703 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
705 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
706 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
707 goto nla_put_failure;
714 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
715 struct netlink_ext_ack *extack)
717 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
718 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
721 /* guard against zero length rings */
725 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
726 struct skb_array *q = band2list(priv, prio);
729 err = skb_array_init(q, qlen, GFP_KERNEL);
734 /* Can by-pass the queue discipline */
735 qdisc->flags |= TCQ_F_CAN_BYPASS;
739 static void pfifo_fast_destroy(struct Qdisc *sch)
741 struct pfifo_fast_priv *priv = qdisc_priv(sch);
744 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
745 struct skb_array *q = band2list(priv, prio);
747 /* NULL ring is possible if destroy path is due to a failed
748 * skb_array_init() in pfifo_fast_init() case.
752 /* Destroy ring but no need to kfree_skb because a call to
753 * pfifo_fast_reset() has already done that work.
755 ptr_ring_cleanup(&q->ring, NULL);
759 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
760 unsigned int new_len)
762 struct pfifo_fast_priv *priv = qdisc_priv(sch);
763 struct skb_array *bands[PFIFO_FAST_BANDS];
766 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
767 struct skb_array *q = band2list(priv, prio);
772 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
776 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
778 .priv_size = sizeof(struct pfifo_fast_priv),
779 .enqueue = pfifo_fast_enqueue,
780 .dequeue = pfifo_fast_dequeue,
781 .peek = pfifo_fast_peek,
782 .init = pfifo_fast_init,
783 .destroy = pfifo_fast_destroy,
784 .reset = pfifo_fast_reset,
785 .dump = pfifo_fast_dump,
786 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
787 .owner = THIS_MODULE,
788 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
790 EXPORT_SYMBOL(pfifo_fast_ops);
792 static struct lock_class_key qdisc_tx_busylock;
793 static struct lock_class_key qdisc_running_key;
795 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
796 const struct Qdisc_ops *ops,
797 struct netlink_ext_ack *extack)
801 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
803 struct net_device *dev;
806 NL_SET_ERR_MSG(extack, "No device queue given");
811 dev = dev_queue->dev;
812 p = kzalloc_node(size, GFP_KERNEL,
813 netdev_queue_numa_node_read(dev_queue));
817 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
818 /* if we got non aligned memory, ask more and do alignment ourself */
821 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
822 netdev_queue_numa_node_read(dev_queue));
825 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
826 sch->padded = (char *) sch - (char *) p;
828 __skb_queue_head_init(&sch->gso_skb);
829 __skb_queue_head_init(&sch->skb_bad_txq);
830 qdisc_skb_head_init(&sch->q);
831 spin_lock_init(&sch->q.lock);
833 if (ops->static_flags & TCQ_F_CPUSTATS) {
835 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
836 if (!sch->cpu_bstats)
839 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
840 if (!sch->cpu_qstats) {
841 free_percpu(sch->cpu_bstats);
846 spin_lock_init(&sch->busylock);
847 lockdep_set_class(&sch->busylock,
848 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
850 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
851 spin_lock_init(&sch->seqlock);
852 lockdep_set_class(&sch->busylock,
853 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
855 seqcount_init(&sch->running);
856 lockdep_set_class(&sch->running,
857 dev->qdisc_running_key ?: &qdisc_running_key);
860 sch->flags = ops->static_flags;
861 sch->enqueue = ops->enqueue;
862 sch->dequeue = ops->dequeue;
863 sch->dev_queue = dev_queue;
866 refcount_set(&sch->refcnt, 1);
875 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
876 const struct Qdisc_ops *ops,
877 unsigned int parentid,
878 struct netlink_ext_ack *extack)
882 if (!try_module_get(ops->owner)) {
883 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
887 sch = qdisc_alloc(dev_queue, ops, extack);
889 module_put(ops->owner);
892 sch->parent = parentid;
894 if (!ops->init || ops->init(sch, NULL, extack) == 0)
900 EXPORT_SYMBOL(qdisc_create_dflt);
902 /* Under qdisc_lock(qdisc) and BH! */
904 void qdisc_reset(struct Qdisc *qdisc)
906 const struct Qdisc_ops *ops = qdisc->ops;
907 struct sk_buff *skb, *tmp;
912 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
913 __skb_unlink(skb, &qdisc->gso_skb);
917 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
918 __skb_unlink(skb, &qdisc->skb_bad_txq);
923 qdisc->qstats.backlog = 0;
925 EXPORT_SYMBOL(qdisc_reset);
927 void qdisc_free(struct Qdisc *qdisc)
929 if (qdisc_is_percpu_stats(qdisc)) {
930 free_percpu(qdisc->cpu_bstats);
931 free_percpu(qdisc->cpu_qstats);
934 kfree((char *) qdisc - qdisc->padded);
937 static void qdisc_free_cb(struct rcu_head *head)
939 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
944 static void qdisc_destroy(struct Qdisc *qdisc)
946 const struct Qdisc_ops *ops = qdisc->ops;
947 struct sk_buff *skb, *tmp;
949 #ifdef CONFIG_NET_SCHED
950 qdisc_hash_del(qdisc);
952 qdisc_put_stab(rtnl_dereference(qdisc->stab));
954 gen_kill_estimator(&qdisc->rate_est);
960 module_put(ops->owner);
961 dev_put(qdisc_dev(qdisc));
963 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
964 __skb_unlink(skb, &qdisc->gso_skb);
968 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
969 __skb_unlink(skb, &qdisc->skb_bad_txq);
973 call_rcu(&qdisc->rcu, qdisc_free_cb);
976 void qdisc_put(struct Qdisc *qdisc)
978 if (qdisc->flags & TCQ_F_BUILTIN ||
979 !refcount_dec_and_test(&qdisc->refcnt))
982 qdisc_destroy(qdisc);
984 EXPORT_SYMBOL(qdisc_put);
986 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
987 * Intended to be used as optimization, this function only takes rtnl lock if
988 * qdisc reference counter reached zero.
991 void qdisc_put_unlocked(struct Qdisc *qdisc)
993 if (qdisc->flags & TCQ_F_BUILTIN ||
994 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
997 qdisc_destroy(qdisc);
1000 EXPORT_SYMBOL(qdisc_put_unlocked);
1002 /* Attach toplevel qdisc to device queue. */
1003 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1004 struct Qdisc *qdisc)
1006 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1007 spinlock_t *root_lock;
1009 root_lock = qdisc_lock(oqdisc);
1010 spin_lock_bh(root_lock);
1012 /* ... and graft new one */
1014 qdisc = &noop_qdisc;
1015 dev_queue->qdisc_sleeping = qdisc;
1016 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1018 spin_unlock_bh(root_lock);
1022 EXPORT_SYMBOL(dev_graft_qdisc);
1024 static void attach_one_default_qdisc(struct net_device *dev,
1025 struct netdev_queue *dev_queue,
1028 struct Qdisc *qdisc;
1029 const struct Qdisc_ops *ops = default_qdisc_ops;
1031 if (dev->priv_flags & IFF_NO_QUEUE)
1032 ops = &noqueue_qdisc_ops;
1034 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1036 netdev_info(dev, "activation failed\n");
1039 if (!netif_is_multiqueue(dev))
1040 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1041 dev_queue->qdisc_sleeping = qdisc;
1044 static void attach_default_qdiscs(struct net_device *dev)
1046 struct netdev_queue *txq;
1047 struct Qdisc *qdisc;
1049 txq = netdev_get_tx_queue(dev, 0);
1051 if (!netif_is_multiqueue(dev) ||
1052 dev->priv_flags & IFF_NO_QUEUE) {
1053 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1054 dev->qdisc = txq->qdisc_sleeping;
1055 qdisc_refcount_inc(dev->qdisc);
1057 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1060 qdisc->ops->attach(qdisc);
1063 #ifdef CONFIG_NET_SCHED
1064 if (dev->qdisc != &noop_qdisc)
1065 qdisc_hash_add(dev->qdisc, false);
1069 static void transition_one_qdisc(struct net_device *dev,
1070 struct netdev_queue *dev_queue,
1071 void *_need_watchdog)
1073 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1074 int *need_watchdog_p = _need_watchdog;
1076 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1077 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1079 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1080 if (need_watchdog_p) {
1081 dev_queue->trans_start = 0;
1082 *need_watchdog_p = 1;
1086 void dev_activate(struct net_device *dev)
1090 /* No queueing discipline is attached to device;
1091 * create default one for devices, which need queueing
1092 * and noqueue_qdisc for virtual interfaces
1095 if (dev->qdisc == &noop_qdisc)
1096 attach_default_qdiscs(dev);
1098 if (!netif_carrier_ok(dev))
1099 /* Delay activation until next carrier-on event */
1103 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1104 if (dev_ingress_queue(dev))
1105 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1107 if (need_watchdog) {
1108 netif_trans_update(dev);
1109 dev_watchdog_up(dev);
1112 EXPORT_SYMBOL(dev_activate);
1114 static void dev_deactivate_queue(struct net_device *dev,
1115 struct netdev_queue *dev_queue,
1116 void *_qdisc_default)
1118 struct Qdisc *qdisc_default = _qdisc_default;
1119 struct Qdisc *qdisc;
1121 qdisc = rtnl_dereference(dev_queue->qdisc);
1123 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1126 spin_lock_bh(&qdisc->seqlock);
1127 spin_lock_bh(qdisc_lock(qdisc));
1129 if (!(qdisc->flags & TCQ_F_BUILTIN))
1130 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1132 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1135 spin_unlock_bh(qdisc_lock(qdisc));
1137 spin_unlock_bh(&qdisc->seqlock);
1141 static bool some_qdisc_is_busy(struct net_device *dev)
1145 for (i = 0; i < dev->num_tx_queues; i++) {
1146 struct netdev_queue *dev_queue;
1147 spinlock_t *root_lock;
1151 dev_queue = netdev_get_tx_queue(dev, i);
1152 q = dev_queue->qdisc_sleeping;
1154 root_lock = qdisc_lock(q);
1155 spin_lock_bh(root_lock);
1157 val = (qdisc_is_running(q) ||
1158 test_bit(__QDISC_STATE_SCHED, &q->state));
1160 spin_unlock_bh(root_lock);
1168 static void dev_qdisc_reset(struct net_device *dev,
1169 struct netdev_queue *dev_queue,
1172 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1179 * dev_deactivate_many - deactivate transmissions on several devices
1180 * @head: list of devices to deactivate
1182 * This function returns only when all outstanding transmissions
1183 * have completed, unless all devices are in dismantle phase.
1185 void dev_deactivate_many(struct list_head *head)
1187 struct net_device *dev;
1189 list_for_each_entry(dev, head, close_list) {
1190 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1192 if (dev_ingress_queue(dev))
1193 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1196 dev_watchdog_down(dev);
1199 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1200 * This is avoided if all devices are in dismantle phase :
1201 * Caller will call synchronize_net() for us
1205 /* Wait for outstanding qdisc_run calls. */
1206 list_for_each_entry(dev, head, close_list) {
1207 while (some_qdisc_is_busy(dev))
1209 /* The new qdisc is assigned at this point so we can safely
1210 * unwind stale skb lists and qdisc statistics
1212 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1213 if (dev_ingress_queue(dev))
1214 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1218 void dev_deactivate(struct net_device *dev)
1222 list_add(&dev->close_list, &single);
1223 dev_deactivate_many(&single);
1226 EXPORT_SYMBOL(dev_deactivate);
1228 static int qdisc_change_tx_queue_len(struct net_device *dev,
1229 struct netdev_queue *dev_queue)
1231 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1232 const struct Qdisc_ops *ops = qdisc->ops;
1234 if (ops->change_tx_queue_len)
1235 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1239 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1241 bool up = dev->flags & IFF_UP;
1246 dev_deactivate(dev);
1248 for (i = 0; i < dev->num_tx_queues; i++) {
1249 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1251 /* TODO: revert changes on a partial failure */
1261 static void dev_init_scheduler_queue(struct net_device *dev,
1262 struct netdev_queue *dev_queue,
1265 struct Qdisc *qdisc = _qdisc;
1267 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1268 dev_queue->qdisc_sleeping = qdisc;
1271 void dev_init_scheduler(struct net_device *dev)
1273 dev->qdisc = &noop_qdisc;
1274 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1275 if (dev_ingress_queue(dev))
1276 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1278 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1281 static void shutdown_scheduler_queue(struct net_device *dev,
1282 struct netdev_queue *dev_queue,
1283 void *_qdisc_default)
1285 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1286 struct Qdisc *qdisc_default = _qdisc_default;
1289 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1290 dev_queue->qdisc_sleeping = qdisc_default;
1296 void dev_shutdown(struct net_device *dev)
1298 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1299 if (dev_ingress_queue(dev))
1300 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1301 qdisc_put(dev->qdisc);
1302 dev->qdisc = &noop_qdisc;
1304 WARN_ON(timer_pending(&dev->watchdog_timer));
1307 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1308 const struct tc_ratespec *conf,
1311 memset(r, 0, sizeof(*r));
1312 r->overhead = conf->overhead;
1313 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1314 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1317 * The deal here is to replace a divide by a reciprocal one
1318 * in fast path (a reciprocal divide is a multiply and a shift)
1320 * Normal formula would be :
1321 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1323 * We compute mult/shift to use instead :
1324 * time_in_ns = (len * mult) >> shift;
1326 * We try to get the highest possible mult value for accuracy,
1327 * but have to make sure no overflows will ever happen.
1329 if (r->rate_bytes_ps > 0) {
1330 u64 factor = NSEC_PER_SEC;
1333 r->mult = div64_u64(factor, r->rate_bytes_ps);
1334 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1341 EXPORT_SYMBOL(psched_ratecfg_precompute);
1343 static void mini_qdisc_rcu_func(struct rcu_head *head)
1347 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1348 struct tcf_proto *tp_head)
1350 /* Protected with chain0->filter_chain_lock.
1351 * Can't access chain directly because tp_head can be NULL.
1353 struct mini_Qdisc *miniq_old =
1354 rcu_dereference_protected(*miniqp->p_miniq, 1);
1355 struct mini_Qdisc *miniq;
1358 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1359 /* Wait for flying RCU callback before it is freed. */
1364 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1365 &miniqp->miniq1 : &miniqp->miniq2;
1367 /* We need to make sure that readers won't see the miniq
1368 * we are about to modify. So wait until previous call_rcu callback
1372 miniq->filter_list = tp_head;
1373 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1376 /* This is counterpart of the rcu barriers above. We need to
1377 * block potential new user of miniq_old until all readers
1378 * are not seeing it.
1380 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1382 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1384 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1385 struct mini_Qdisc __rcu **p_miniq)
1387 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1388 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1389 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1390 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1391 miniqp->p_miniq = p_miniq;
1393 EXPORT_SYMBOL(mini_qdisc_pair_init);