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 int __dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
124 struct sk_buff *next = skb->next;
126 __skb_queue_tail(&q->gso_skb, skb);
127 q->qstats.requeues++;
128 qdisc_qstats_backlog_inc(q, skb);
129 q->q.qlen++; /* it's still part of the queue */
138 static inline int dev_requeue_skb_locked(struct sk_buff *skb, struct Qdisc *q)
140 spinlock_t *lock = qdisc_lock(q);
144 struct sk_buff *next = skb->next;
146 __skb_queue_tail(&q->gso_skb, skb);
148 qdisc_qstats_cpu_requeues_inc(q);
149 qdisc_qstats_cpu_backlog_inc(q, skb);
150 qdisc_qstats_cpu_qlen_inc(q);
161 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
163 if (q->flags & TCQ_F_NOLOCK)
164 return dev_requeue_skb_locked(skb, q);
166 return __dev_requeue_skb(skb, q);
169 static void try_bulk_dequeue_skb(struct Qdisc *q,
171 const struct netdev_queue *txq,
174 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
176 while (bytelimit > 0) {
177 struct sk_buff *nskb = q->dequeue(q);
182 bytelimit -= nskb->len; /* covers GSO len */
185 (*packets)++; /* GSO counts as one pkt */
190 /* This variant of try_bulk_dequeue_skb() makes sure
191 * all skbs in the chain are for the same txq
193 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
197 int mapping = skb_get_queue_mapping(skb);
198 struct sk_buff *nskb;
202 nskb = q->dequeue(q);
205 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
206 qdisc_enqueue_skb_bad_txq(q, nskb);
216 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
217 * A requeued skb (via q->gso_skb) can also be a SKB list.
219 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
222 const struct netdev_queue *txq = q->dev_queue;
223 struct sk_buff *skb = NULL;
226 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
227 spinlock_t *lock = NULL;
229 if (q->flags & TCQ_F_NOLOCK) {
230 lock = qdisc_lock(q);
234 skb = skb_peek(&q->gso_skb);
236 /* skb may be null if another cpu pulls gso_skb off in between
237 * empty check and lock.
245 /* skb in gso_skb were already validated */
247 if (xfrm_offload(skb))
249 /* check the reason of requeuing without tx lock first */
250 txq = skb_get_tx_queue(txq->dev, skb);
251 if (!netif_xmit_frozen_or_stopped(txq)) {
252 skb = __skb_dequeue(&q->gso_skb);
253 if (qdisc_is_percpu_stats(q)) {
254 qdisc_qstats_cpu_backlog_dec(q, skb);
255 qdisc_qstats_cpu_qlen_dec(q);
257 qdisc_qstats_backlog_dec(q, skb);
270 if ((q->flags & TCQ_F_ONETXQUEUE) &&
271 netif_xmit_frozen_or_stopped(txq))
274 skb = qdisc_dequeue_skb_bad_txq(q);
280 if (qdisc_may_bulk(q))
281 try_bulk_dequeue_skb(q, skb, txq, packets);
283 try_bulk_dequeue_skb_slow(q, skb, packets);
286 trace_qdisc_dequeue(q, txq, *packets, skb);
291 * Transmit possibly several skbs, and handle the return status as
292 * required. Owning running seqcount bit guarantees that
293 * only one CPU can execute this function.
295 * Returns to the caller:
296 * false - hardware queue frozen backoff
297 * true - feel free to send more pkts
299 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
300 struct net_device *dev, struct netdev_queue *txq,
301 spinlock_t *root_lock, bool validate)
303 int ret = NETDEV_TX_BUSY;
306 /* And release qdisc */
308 spin_unlock(root_lock);
310 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
312 skb = validate_xmit_skb_list(skb, dev, &again);
314 #ifdef CONFIG_XFRM_OFFLOAD
315 if (unlikely(again)) {
317 spin_lock(root_lock);
319 dev_requeue_skb(skb, q);
325 HARD_TX_LOCK(dev, txq, smp_processor_id());
326 if (!netif_xmit_frozen_or_stopped(txq))
327 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
329 HARD_TX_UNLOCK(dev, txq);
332 spin_lock(root_lock);
337 spin_lock(root_lock);
339 if (!dev_xmit_complete(ret)) {
340 /* Driver returned NETDEV_TX_BUSY - requeue skb */
341 if (unlikely(ret != NETDEV_TX_BUSY))
342 net_warn_ratelimited("BUG %s code %d qlen %d\n",
343 dev->name, ret, q->q.qlen);
345 dev_requeue_skb(skb, q);
349 if (ret && netif_xmit_frozen_or_stopped(txq))
356 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
358 * running seqcount guarantees only one CPU can process
359 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
362 * netif_tx_lock serializes accesses to device driver.
364 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
365 * if one is grabbed, another must be free.
367 * Note, that this procedure can be called by a watchdog timer
369 * Returns to the caller:
370 * 0 - queue is empty or throttled.
371 * >0 - queue is not empty.
374 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
376 bool more, validate, nolock = q->flags & TCQ_F_NOLOCK;
377 spinlock_t *root_lock = NULL;
378 struct netdev_queue *txq;
379 struct net_device *dev;
383 if (nolock && test_and_set_bit(__QDISC_STATE_RUNNING, &q->state))
386 skb = dequeue_skb(q, &validate, packets);
387 if (unlikely(!skb)) {
389 clear_bit(__QDISC_STATE_RUNNING, &q->state);
394 root_lock = qdisc_lock(q);
397 txq = skb_get_tx_queue(dev, skb);
399 more = sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
401 clear_bit(__QDISC_STATE_RUNNING, &q->state);
405 void __qdisc_run(struct Qdisc *q)
407 int quota = dev_tx_weight;
410 while (qdisc_restart(q, &packets)) {
412 * Ordered by possible occurrence: Postpone processing if
413 * 1. we've exceeded packet quota
414 * 2. another process needs the CPU;
417 if (quota <= 0 || need_resched()) {
424 unsigned long dev_trans_start(struct net_device *dev)
426 unsigned long val, res;
429 if (is_vlan_dev(dev))
430 dev = vlan_dev_real_dev(dev);
431 else if (netif_is_macvlan(dev))
432 dev = macvlan_dev_real_dev(dev);
433 res = netdev_get_tx_queue(dev, 0)->trans_start;
434 for (i = 1; i < dev->num_tx_queues; i++) {
435 val = netdev_get_tx_queue(dev, i)->trans_start;
436 if (val && time_after(val, res))
442 EXPORT_SYMBOL(dev_trans_start);
444 static void dev_watchdog(struct timer_list *t)
446 struct net_device *dev = from_timer(dev, t, watchdog_timer);
449 if (!qdisc_tx_is_noop(dev)) {
450 if (netif_device_present(dev) &&
451 netif_running(dev) &&
452 netif_carrier_ok(dev)) {
453 int some_queue_timedout = 0;
455 unsigned long trans_start;
457 for (i = 0; i < dev->num_tx_queues; i++) {
458 struct netdev_queue *txq;
460 txq = netdev_get_tx_queue(dev, i);
461 trans_start = txq->trans_start;
462 if (netif_xmit_stopped(txq) &&
463 time_after(jiffies, (trans_start +
464 dev->watchdog_timeo))) {
465 some_queue_timedout = 1;
466 txq->trans_timeout++;
471 if (some_queue_timedout) {
472 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
473 dev->name, netdev_drivername(dev), i);
474 dev->netdev_ops->ndo_tx_timeout(dev);
476 if (!mod_timer(&dev->watchdog_timer,
477 round_jiffies(jiffies +
478 dev->watchdog_timeo)))
482 netif_tx_unlock(dev);
487 void __netdev_watchdog_up(struct net_device *dev)
489 if (dev->netdev_ops->ndo_tx_timeout) {
490 if (dev->watchdog_timeo <= 0)
491 dev->watchdog_timeo = 5*HZ;
492 if (!mod_timer(&dev->watchdog_timer,
493 round_jiffies(jiffies + dev->watchdog_timeo)))
498 static void dev_watchdog_up(struct net_device *dev)
500 __netdev_watchdog_up(dev);
503 static void dev_watchdog_down(struct net_device *dev)
505 netif_tx_lock_bh(dev);
506 if (del_timer(&dev->watchdog_timer))
508 netif_tx_unlock_bh(dev);
512 * netif_carrier_on - set carrier
513 * @dev: network device
515 * Device has detected that carrier.
517 void netif_carrier_on(struct net_device *dev)
519 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
520 if (dev->reg_state == NETREG_UNINITIALIZED)
522 atomic_inc(&dev->carrier_up_count);
523 linkwatch_fire_event(dev);
524 if (netif_running(dev))
525 __netdev_watchdog_up(dev);
528 EXPORT_SYMBOL(netif_carrier_on);
531 * netif_carrier_off - clear carrier
532 * @dev: network device
534 * Device has detected loss of carrier.
536 void netif_carrier_off(struct net_device *dev)
538 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
539 if (dev->reg_state == NETREG_UNINITIALIZED)
541 atomic_inc(&dev->carrier_down_count);
542 linkwatch_fire_event(dev);
545 EXPORT_SYMBOL(netif_carrier_off);
547 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
548 under all circumstances. It is difficult to invent anything faster or
552 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
553 struct sk_buff **to_free)
555 __qdisc_drop(skb, to_free);
559 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
564 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
567 .enqueue = noop_enqueue,
568 .dequeue = noop_dequeue,
569 .peek = noop_dequeue,
570 .owner = THIS_MODULE,
573 static struct netdev_queue noop_netdev_queue = {
574 .qdisc = &noop_qdisc,
575 .qdisc_sleeping = &noop_qdisc,
578 struct Qdisc noop_qdisc = {
579 .enqueue = noop_enqueue,
580 .dequeue = noop_dequeue,
581 .flags = TCQ_F_BUILTIN,
582 .ops = &noop_qdisc_ops,
583 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
584 .dev_queue = &noop_netdev_queue,
585 .running = SEQCNT_ZERO(noop_qdisc.running),
586 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
588 EXPORT_SYMBOL(noop_qdisc);
590 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
591 struct netlink_ext_ack *extack)
593 /* register_qdisc() assigns a default of noop_enqueue if unset,
594 * but __dev_queue_xmit() treats noqueue only as such
595 * if this is NULL - so clear it here. */
596 qdisc->enqueue = NULL;
600 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
603 .init = noqueue_init,
604 .enqueue = noop_enqueue,
605 .dequeue = noop_dequeue,
606 .peek = noop_dequeue,
607 .owner = THIS_MODULE,
610 static const u8 prio2band[TC_PRIO_MAX + 1] = {
611 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
614 /* 3-band FIFO queue: old style, but should be a bit faster than
615 generic prio+fifo combination.
618 #define PFIFO_FAST_BANDS 3
621 * Private data for a pfifo_fast scheduler containing:
622 * - rings for priority bands
624 struct pfifo_fast_priv {
625 struct skb_array q[PFIFO_FAST_BANDS];
628 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
631 return &priv->q[band];
634 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
635 struct sk_buff **to_free)
637 int band = prio2band[skb->priority & TC_PRIO_MAX];
638 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
639 struct skb_array *q = band2list(priv, band);
640 unsigned int pkt_len = qdisc_pkt_len(skb);
643 err = skb_array_produce(q, skb);
646 return qdisc_drop_cpu(skb, qdisc, to_free);
648 qdisc_qstats_cpu_qlen_inc(qdisc);
649 /* Note: skb can not be used after skb_array_produce(),
650 * so we better not use qdisc_qstats_cpu_backlog_inc()
652 this_cpu_add(qdisc->cpu_qstats->backlog, pkt_len);
653 return NET_XMIT_SUCCESS;
656 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
658 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
659 struct sk_buff *skb = NULL;
662 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
663 struct skb_array *q = band2list(priv, band);
665 if (__skb_array_empty(q))
668 skb = skb_array_consume_bh(q);
671 qdisc_qstats_cpu_backlog_dec(qdisc, skb);
672 qdisc_bstats_cpu_update(qdisc, skb);
673 qdisc_qstats_cpu_qlen_dec(qdisc);
679 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
681 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
682 struct sk_buff *skb = NULL;
685 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
686 struct skb_array *q = band2list(priv, band);
688 skb = __skb_array_peek(q);
694 static void pfifo_fast_reset(struct Qdisc *qdisc)
697 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
699 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
700 struct skb_array *q = band2list(priv, band);
703 /* NULL ring is possible if destroy path is due to a failed
704 * skb_array_init() in pfifo_fast_init() case.
709 while ((skb = skb_array_consume_bh(q)) != NULL)
713 for_each_possible_cpu(i) {
714 struct gnet_stats_queue *q = per_cpu_ptr(qdisc->cpu_qstats, i);
721 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
723 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
725 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
726 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
727 goto nla_put_failure;
734 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
735 struct netlink_ext_ack *extack)
737 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
738 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
741 /* guard against zero length rings */
745 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
746 struct skb_array *q = band2list(priv, prio);
749 err = skb_array_init(q, qlen, GFP_KERNEL);
754 /* Can by-pass the queue discipline */
755 qdisc->flags |= TCQ_F_CAN_BYPASS;
759 static void pfifo_fast_destroy(struct Qdisc *sch)
761 struct pfifo_fast_priv *priv = qdisc_priv(sch);
764 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
765 struct skb_array *q = band2list(priv, prio);
767 /* NULL ring is possible if destroy path is due to a failed
768 * skb_array_init() in pfifo_fast_init() case.
772 /* Destroy ring but no need to kfree_skb because a call to
773 * pfifo_fast_reset() has already done that work.
775 ptr_ring_cleanup(&q->ring, NULL);
779 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
780 unsigned int new_len)
782 struct pfifo_fast_priv *priv = qdisc_priv(sch);
783 struct skb_array *bands[PFIFO_FAST_BANDS];
786 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
787 struct skb_array *q = band2list(priv, prio);
792 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
796 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
798 .priv_size = sizeof(struct pfifo_fast_priv),
799 .enqueue = pfifo_fast_enqueue,
800 .dequeue = pfifo_fast_dequeue,
801 .peek = pfifo_fast_peek,
802 .init = pfifo_fast_init,
803 .destroy = pfifo_fast_destroy,
804 .reset = pfifo_fast_reset,
805 .dump = pfifo_fast_dump,
806 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
807 .owner = THIS_MODULE,
808 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
810 EXPORT_SYMBOL(pfifo_fast_ops);
812 static struct lock_class_key qdisc_tx_busylock;
813 static struct lock_class_key qdisc_running_key;
815 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
816 const struct Qdisc_ops *ops,
817 struct netlink_ext_ack *extack)
821 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
823 struct net_device *dev;
826 NL_SET_ERR_MSG(extack, "No device queue given");
831 dev = dev_queue->dev;
832 p = kzalloc_node(size, GFP_KERNEL,
833 netdev_queue_numa_node_read(dev_queue));
837 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
838 /* if we got non aligned memory, ask more and do alignment ourself */
841 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
842 netdev_queue_numa_node_read(dev_queue));
845 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
846 sch->padded = (char *) sch - (char *) p;
848 __skb_queue_head_init(&sch->gso_skb);
849 __skb_queue_head_init(&sch->skb_bad_txq);
850 qdisc_skb_head_init(&sch->q);
851 spin_lock_init(&sch->q.lock);
853 if (ops->static_flags & TCQ_F_CPUSTATS) {
855 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
856 if (!sch->cpu_bstats)
859 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
860 if (!sch->cpu_qstats) {
861 free_percpu(sch->cpu_bstats);
866 spin_lock_init(&sch->busylock);
867 lockdep_set_class(&sch->busylock,
868 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
870 seqcount_init(&sch->running);
871 lockdep_set_class(&sch->running,
872 dev->qdisc_running_key ?: &qdisc_running_key);
875 sch->flags = ops->static_flags;
876 sch->enqueue = ops->enqueue;
877 sch->dequeue = ops->dequeue;
878 sch->dev_queue = dev_queue;
880 refcount_set(&sch->refcnt, 1);
889 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
890 const struct Qdisc_ops *ops,
891 unsigned int parentid,
892 struct netlink_ext_ack *extack)
896 if (!try_module_get(ops->owner)) {
897 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
901 sch = qdisc_alloc(dev_queue, ops, extack);
903 module_put(ops->owner);
906 sch->parent = parentid;
908 if (!ops->init || ops->init(sch, NULL, extack) == 0)
914 EXPORT_SYMBOL(qdisc_create_dflt);
916 /* Under qdisc_lock(qdisc) and BH! */
918 void qdisc_reset(struct Qdisc *qdisc)
920 const struct Qdisc_ops *ops = qdisc->ops;
921 struct sk_buff *skb, *tmp;
926 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
927 __skb_unlink(skb, &qdisc->gso_skb);
931 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
932 __skb_unlink(skb, &qdisc->skb_bad_txq);
937 qdisc->qstats.backlog = 0;
939 EXPORT_SYMBOL(qdisc_reset);
941 void qdisc_free(struct Qdisc *qdisc)
943 if (qdisc_is_percpu_stats(qdisc)) {
944 free_percpu(qdisc->cpu_bstats);
945 free_percpu(qdisc->cpu_qstats);
948 kfree((char *) qdisc - qdisc->padded);
951 void qdisc_destroy(struct Qdisc *qdisc)
953 const struct Qdisc_ops *ops = qdisc->ops;
954 struct sk_buff *skb, *tmp;
956 if (qdisc->flags & TCQ_F_BUILTIN ||
957 !refcount_dec_and_test(&qdisc->refcnt))
960 #ifdef CONFIG_NET_SCHED
961 qdisc_hash_del(qdisc);
963 qdisc_put_stab(rtnl_dereference(qdisc->stab));
965 gen_kill_estimator(&qdisc->rate_est);
971 module_put(ops->owner);
972 dev_put(qdisc_dev(qdisc));
974 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
975 __skb_unlink(skb, &qdisc->gso_skb);
979 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
980 __skb_unlink(skb, &qdisc->skb_bad_txq);
986 EXPORT_SYMBOL(qdisc_destroy);
988 /* Attach toplevel qdisc to device queue. */
989 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
992 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
993 spinlock_t *root_lock;
995 root_lock = qdisc_lock(oqdisc);
996 spin_lock_bh(root_lock);
998 /* ... and graft new one */
1000 qdisc = &noop_qdisc;
1001 dev_queue->qdisc_sleeping = qdisc;
1002 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1004 spin_unlock_bh(root_lock);
1008 EXPORT_SYMBOL(dev_graft_qdisc);
1010 static void attach_one_default_qdisc(struct net_device *dev,
1011 struct netdev_queue *dev_queue,
1014 struct Qdisc *qdisc;
1015 const struct Qdisc_ops *ops = default_qdisc_ops;
1017 if (dev->priv_flags & IFF_NO_QUEUE)
1018 ops = &noqueue_qdisc_ops;
1020 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1022 netdev_info(dev, "activation failed\n");
1025 if (!netif_is_multiqueue(dev))
1026 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1027 dev_queue->qdisc_sleeping = qdisc;
1030 static void attach_default_qdiscs(struct net_device *dev)
1032 struct netdev_queue *txq;
1033 struct Qdisc *qdisc;
1035 txq = netdev_get_tx_queue(dev, 0);
1037 if (!netif_is_multiqueue(dev) ||
1038 dev->priv_flags & IFF_NO_QUEUE) {
1039 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1040 dev->qdisc = txq->qdisc_sleeping;
1041 qdisc_refcount_inc(dev->qdisc);
1043 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1046 qdisc->ops->attach(qdisc);
1049 #ifdef CONFIG_NET_SCHED
1050 if (dev->qdisc != &noop_qdisc)
1051 qdisc_hash_add(dev->qdisc, false);
1055 static void transition_one_qdisc(struct net_device *dev,
1056 struct netdev_queue *dev_queue,
1057 void *_need_watchdog)
1059 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1060 int *need_watchdog_p = _need_watchdog;
1062 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1063 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1065 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1066 if (need_watchdog_p) {
1067 dev_queue->trans_start = 0;
1068 *need_watchdog_p = 1;
1072 void dev_activate(struct net_device *dev)
1076 /* No queueing discipline is attached to device;
1077 * create default one for devices, which need queueing
1078 * and noqueue_qdisc for virtual interfaces
1081 if (dev->qdisc == &noop_qdisc)
1082 attach_default_qdiscs(dev);
1084 if (!netif_carrier_ok(dev))
1085 /* Delay activation until next carrier-on event */
1089 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1090 if (dev_ingress_queue(dev))
1091 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1093 if (need_watchdog) {
1094 netif_trans_update(dev);
1095 dev_watchdog_up(dev);
1098 EXPORT_SYMBOL(dev_activate);
1100 static void dev_deactivate_queue(struct net_device *dev,
1101 struct netdev_queue *dev_queue,
1102 void *_qdisc_default)
1104 struct Qdisc *qdisc_default = _qdisc_default;
1105 struct Qdisc *qdisc;
1107 qdisc = rtnl_dereference(dev_queue->qdisc);
1109 spin_lock_bh(qdisc_lock(qdisc));
1111 if (!(qdisc->flags & TCQ_F_BUILTIN))
1112 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1114 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1117 spin_unlock_bh(qdisc_lock(qdisc));
1121 static bool some_qdisc_is_busy(struct net_device *dev)
1125 for (i = 0; i < dev->num_tx_queues; i++) {
1126 struct netdev_queue *dev_queue;
1127 spinlock_t *root_lock;
1131 dev_queue = netdev_get_tx_queue(dev, i);
1132 q = dev_queue->qdisc_sleeping;
1134 if (q->flags & TCQ_F_NOLOCK) {
1135 val = test_bit(__QDISC_STATE_SCHED, &q->state);
1137 root_lock = qdisc_lock(q);
1138 spin_lock_bh(root_lock);
1140 val = (qdisc_is_running(q) ||
1141 test_bit(__QDISC_STATE_SCHED, &q->state));
1143 spin_unlock_bh(root_lock);
1152 static void dev_qdisc_reset(struct net_device *dev,
1153 struct netdev_queue *dev_queue,
1156 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1163 * dev_deactivate_many - deactivate transmissions on several devices
1164 * @head: list of devices to deactivate
1166 * This function returns only when all outstanding transmissions
1167 * have completed, unless all devices are in dismantle phase.
1169 void dev_deactivate_many(struct list_head *head)
1171 struct net_device *dev;
1173 list_for_each_entry(dev, head, close_list) {
1174 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1176 if (dev_ingress_queue(dev))
1177 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1180 dev_watchdog_down(dev);
1183 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1184 * This is avoided if all devices are in dismantle phase :
1185 * Caller will call synchronize_net() for us
1189 /* Wait for outstanding qdisc_run calls. */
1190 list_for_each_entry(dev, head, close_list) {
1191 while (some_qdisc_is_busy(dev))
1193 /* The new qdisc is assigned at this point so we can safely
1194 * unwind stale skb lists and qdisc statistics
1196 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1197 if (dev_ingress_queue(dev))
1198 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1202 void dev_deactivate(struct net_device *dev)
1206 list_add(&dev->close_list, &single);
1207 dev_deactivate_many(&single);
1210 EXPORT_SYMBOL(dev_deactivate);
1212 static int qdisc_change_tx_queue_len(struct net_device *dev,
1213 struct netdev_queue *dev_queue)
1215 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1216 const struct Qdisc_ops *ops = qdisc->ops;
1218 if (ops->change_tx_queue_len)
1219 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1223 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1225 bool up = dev->flags & IFF_UP;
1230 dev_deactivate(dev);
1232 for (i = 0; i < dev->num_tx_queues; i++) {
1233 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1235 /* TODO: revert changes on a partial failure */
1245 static void dev_init_scheduler_queue(struct net_device *dev,
1246 struct netdev_queue *dev_queue,
1249 struct Qdisc *qdisc = _qdisc;
1251 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1252 dev_queue->qdisc_sleeping = qdisc;
1253 __skb_queue_head_init(&qdisc->gso_skb);
1254 __skb_queue_head_init(&qdisc->skb_bad_txq);
1257 void dev_init_scheduler(struct net_device *dev)
1259 dev->qdisc = &noop_qdisc;
1260 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1261 if (dev_ingress_queue(dev))
1262 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1264 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1267 static void shutdown_scheduler_queue(struct net_device *dev,
1268 struct netdev_queue *dev_queue,
1269 void *_qdisc_default)
1271 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1272 struct Qdisc *qdisc_default = _qdisc_default;
1275 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1276 dev_queue->qdisc_sleeping = qdisc_default;
1278 qdisc_destroy(qdisc);
1282 void dev_shutdown(struct net_device *dev)
1284 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1287 qdisc_destroy(dev->qdisc);
1288 dev->qdisc = &noop_qdisc;
1290 WARN_ON(timer_pending(&dev->watchdog_timer));
1293 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1294 const struct tc_ratespec *conf,
1297 memset(r, 0, sizeof(*r));
1298 r->overhead = conf->overhead;
1299 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1300 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1303 * The deal here is to replace a divide by a reciprocal one
1304 * in fast path (a reciprocal divide is a multiply and a shift)
1306 * Normal formula would be :
1307 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1309 * We compute mult/shift to use instead :
1310 * time_in_ns = (len * mult) >> shift;
1312 * We try to get the highest possible mult value for accuracy,
1313 * but have to make sure no overflows will ever happen.
1315 if (r->rate_bytes_ps > 0) {
1316 u64 factor = NSEC_PER_SEC;
1319 r->mult = div64_u64(factor, r->rate_bytes_ps);
1320 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1327 EXPORT_SYMBOL(psched_ratecfg_precompute);
1329 static void mini_qdisc_rcu_func(struct rcu_head *head)
1333 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1334 struct tcf_proto *tp_head)
1336 struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1337 struct mini_Qdisc *miniq;
1340 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1341 /* Wait for flying RCU callback before it is freed. */
1346 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1347 &miniqp->miniq1 : &miniqp->miniq2;
1349 /* We need to make sure that readers won't see the miniq
1350 * we are about to modify. So wait until previous call_rcu_bh callback
1354 miniq->filter_list = tp_head;
1355 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1358 /* This is counterpart of the rcu barriers above. We need to
1359 * block potential new user of miniq_old until all readers
1360 * are not seeing it.
1362 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1364 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1366 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1367 struct mini_Qdisc __rcu **p_miniq)
1369 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1370 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1371 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1372 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1373 miniqp->p_miniq = p_miniq;
1375 EXPORT_SYMBOL(mini_qdisc_pair_init);