1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/sched/sch_generic.c Generic packet scheduler routines.
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
10 #include <linux/bitops.h>
11 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/netdevice.h>
18 #include <linux/skbuff.h>
19 #include <linux/rtnetlink.h>
20 #include <linux/init.h>
21 #include <linux/rcupdate.h>
22 #include <linux/list.h>
23 #include <linux/slab.h>
24 #include <linux/if_vlan.h>
25 #include <linux/skb_array.h>
26 #include <linux/if_macvlan.h>
27 #include <net/sch_generic.h>
28 #include <net/pkt_sched.h>
30 #include <trace/events/qdisc.h>
31 #include <trace/events/net.h>
34 /* Qdisc to use by default */
35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36 EXPORT_SYMBOL(default_qdisc_ops);
38 /* Main transmission queue. */
40 /* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
49 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
51 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53 const struct netdev_queue *txq = q->dev_queue;
54 spinlock_t *lock = NULL;
57 if (q->flags & TCQ_F_NOLOCK) {
62 skb = skb_peek(&q->skb_bad_txq);
64 /* check the reason of requeuing without tx lock first */
65 txq = skb_get_tx_queue(txq->dev, skb);
66 if (!netif_xmit_frozen_or_stopped(txq)) {
67 skb = __skb_dequeue(&q->skb_bad_txq);
68 if (qdisc_is_percpu_stats(q)) {
69 qdisc_qstats_cpu_backlog_dec(q, skb);
70 qdisc_qstats_cpu_qlen_dec(q);
72 qdisc_qstats_backlog_dec(q, skb);
86 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
91 skb = __skb_dequeue_bad_txq(q);
96 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
99 spinlock_t *lock = NULL;
101 if (q->flags & TCQ_F_NOLOCK) {
102 lock = qdisc_lock(q);
106 __skb_queue_tail(&q->skb_bad_txq, skb);
108 if (qdisc_is_percpu_stats(q)) {
109 qdisc_qstats_cpu_backlog_inc(q, skb);
110 qdisc_qstats_cpu_qlen_inc(q);
112 qdisc_qstats_backlog_inc(q, skb);
120 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
122 spinlock_t *lock = NULL;
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
130 struct sk_buff *next = skb->next;
132 __skb_queue_tail(&q->gso_skb, skb);
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q)) {
136 qdisc_qstats_cpu_requeues_inc(q);
137 qdisc_qstats_cpu_backlog_inc(q, skb);
138 qdisc_qstats_cpu_qlen_inc(q);
140 q->qstats.requeues++;
141 qdisc_qstats_backlog_inc(q, skb);
152 static void try_bulk_dequeue_skb(struct Qdisc *q,
154 const struct netdev_queue *txq,
157 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
159 while (bytelimit > 0) {
160 struct sk_buff *nskb = q->dequeue(q);
165 bytelimit -= nskb->len; /* covers GSO len */
168 (*packets)++; /* GSO counts as one pkt */
170 skb_mark_not_on_list(skb);
173 /* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
176 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
180 int mapping = skb_get_queue_mapping(skb);
181 struct sk_buff *nskb;
185 nskb = q->dequeue(q);
188 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
189 qdisc_enqueue_skb_bad_txq(q, nskb);
196 skb_mark_not_on_list(skb);
199 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
202 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
205 const struct netdev_queue *txq = q->dev_queue;
206 struct sk_buff *skb = NULL;
209 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
210 spinlock_t *lock = NULL;
212 if (q->flags & TCQ_F_NOLOCK) {
213 lock = qdisc_lock(q);
217 skb = skb_peek(&q->gso_skb);
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
228 /* skb in gso_skb were already validated */
230 if (xfrm_offload(skb))
232 /* check the reason of requeuing without tx lock first */
233 txq = skb_get_tx_queue(txq->dev, skb);
234 if (!netif_xmit_frozen_or_stopped(txq)) {
235 skb = __skb_dequeue(&q->gso_skb);
236 if (qdisc_is_percpu_stats(q)) {
237 qdisc_qstats_cpu_backlog_dec(q, skb);
238 qdisc_qstats_cpu_qlen_dec(q);
240 qdisc_qstats_backlog_dec(q, skb);
253 if ((q->flags & TCQ_F_ONETXQUEUE) &&
254 netif_xmit_frozen_or_stopped(txq))
257 skb = qdisc_dequeue_skb_bad_txq(q);
259 if (skb == SKB_XOFF_MAGIC)
266 if (qdisc_may_bulk(q))
267 try_bulk_dequeue_skb(q, skb, txq, packets);
269 try_bulk_dequeue_skb_slow(q, skb, packets);
272 trace_qdisc_dequeue(q, txq, *packets, skb);
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
285 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
286 struct net_device *dev, struct netdev_queue *txq,
287 spinlock_t *root_lock, bool validate)
289 int ret = NETDEV_TX_BUSY;
292 /* And release qdisc */
294 spin_unlock(root_lock);
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
298 skb = validate_xmit_skb_list(skb, dev, &again);
300 #ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again)) {
303 spin_lock(root_lock);
305 dev_requeue_skb(skb, q);
311 HARD_TX_LOCK(dev, txq, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq))
313 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
315 HARD_TX_UNLOCK(dev, txq);
318 spin_lock(root_lock);
323 spin_lock(root_lock);
325 if (!dev_xmit_complete(ret)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret != NETDEV_TX_BUSY))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev->name, ret, q->q.qlen);
331 dev_requeue_skb(skb, q);
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
345 * netif_tx_lock serializes accesses to device driver.
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
350 * Note, that this procedure can be called by a watchdog timer
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
357 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
359 spinlock_t *root_lock = NULL;
360 struct netdev_queue *txq;
361 struct net_device *dev;
366 skb = dequeue_skb(q, &validate, packets);
370 if (!(q->flags & TCQ_F_NOLOCK))
371 root_lock = qdisc_lock(q);
374 txq = skb_get_tx_queue(dev, skb);
376 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
379 void __qdisc_run(struct Qdisc *q)
381 int quota = dev_tx_weight;
384 while (qdisc_restart(q, &packets)) {
386 * Ordered by possible occurrence: Postpone processing if
387 * 1. we've exceeded packet quota
388 * 2. another process needs the CPU;
391 if (quota <= 0 || need_resched()) {
398 unsigned long dev_trans_start(struct net_device *dev)
400 unsigned long val, res;
403 if (is_vlan_dev(dev))
404 dev = vlan_dev_real_dev(dev);
405 else if (netif_is_macvlan(dev))
406 dev = macvlan_dev_real_dev(dev);
407 res = netdev_get_tx_queue(dev, 0)->trans_start;
408 for (i = 1; i < dev->num_tx_queues; i++) {
409 val = netdev_get_tx_queue(dev, i)->trans_start;
410 if (val && time_after(val, res))
416 EXPORT_SYMBOL(dev_trans_start);
418 static void dev_watchdog(struct timer_list *t)
420 struct net_device *dev = from_timer(dev, t, watchdog_timer);
423 if (!qdisc_tx_is_noop(dev)) {
424 if (netif_device_present(dev) &&
425 netif_running(dev) &&
426 netif_carrier_ok(dev)) {
427 int some_queue_timedout = 0;
429 unsigned long trans_start;
431 for (i = 0; i < dev->num_tx_queues; i++) {
432 struct netdev_queue *txq;
434 txq = netdev_get_tx_queue(dev, i);
435 trans_start = txq->trans_start;
436 if (netif_xmit_stopped(txq) &&
437 time_after(jiffies, (trans_start +
438 dev->watchdog_timeo))) {
439 some_queue_timedout = 1;
440 txq->trans_timeout++;
445 if (some_queue_timedout) {
446 trace_net_dev_xmit_timeout(dev, i);
447 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
448 dev->name, netdev_drivername(dev), i);
449 dev->netdev_ops->ndo_tx_timeout(dev);
451 if (!mod_timer(&dev->watchdog_timer,
452 round_jiffies(jiffies +
453 dev->watchdog_timeo)))
457 netif_tx_unlock(dev);
462 void __netdev_watchdog_up(struct net_device *dev)
464 if (dev->netdev_ops->ndo_tx_timeout) {
465 if (dev->watchdog_timeo <= 0)
466 dev->watchdog_timeo = 5*HZ;
467 if (!mod_timer(&dev->watchdog_timer,
468 round_jiffies(jiffies + dev->watchdog_timeo)))
473 static void dev_watchdog_up(struct net_device *dev)
475 __netdev_watchdog_up(dev);
478 static void dev_watchdog_down(struct net_device *dev)
480 netif_tx_lock_bh(dev);
481 if (del_timer(&dev->watchdog_timer))
483 netif_tx_unlock_bh(dev);
487 * netif_carrier_on - set carrier
488 * @dev: network device
490 * Device has detected acquisition of carrier.
492 void netif_carrier_on(struct net_device *dev)
494 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
495 if (dev->reg_state == NETREG_UNINITIALIZED)
497 atomic_inc(&dev->carrier_up_count);
498 linkwatch_fire_event(dev);
499 if (netif_running(dev))
500 __netdev_watchdog_up(dev);
503 EXPORT_SYMBOL(netif_carrier_on);
506 * netif_carrier_off - clear carrier
507 * @dev: network device
509 * Device has detected loss of carrier.
511 void netif_carrier_off(struct net_device *dev)
513 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
514 if (dev->reg_state == NETREG_UNINITIALIZED)
516 atomic_inc(&dev->carrier_down_count);
517 linkwatch_fire_event(dev);
520 EXPORT_SYMBOL(netif_carrier_off);
522 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
523 under all circumstances. It is difficult to invent anything faster or
527 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
528 struct sk_buff **to_free)
530 __qdisc_drop(skb, to_free);
534 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
539 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
542 .enqueue = noop_enqueue,
543 .dequeue = noop_dequeue,
544 .peek = noop_dequeue,
545 .owner = THIS_MODULE,
548 static struct netdev_queue noop_netdev_queue = {
549 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
550 .qdisc_sleeping = &noop_qdisc,
553 struct Qdisc noop_qdisc = {
554 .enqueue = noop_enqueue,
555 .dequeue = noop_dequeue,
556 .flags = TCQ_F_BUILTIN,
557 .ops = &noop_qdisc_ops,
558 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
559 .dev_queue = &noop_netdev_queue,
560 .running = SEQCNT_ZERO(noop_qdisc.running),
561 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
563 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
564 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
566 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
569 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
570 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
572 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
575 EXPORT_SYMBOL(noop_qdisc);
577 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
578 struct netlink_ext_ack *extack)
580 /* register_qdisc() assigns a default of noop_enqueue if unset,
581 * but __dev_queue_xmit() treats noqueue only as such
582 * if this is NULL - so clear it here. */
583 qdisc->enqueue = NULL;
587 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
590 .init = noqueue_init,
591 .enqueue = noop_enqueue,
592 .dequeue = noop_dequeue,
593 .peek = noop_dequeue,
594 .owner = THIS_MODULE,
597 static const u8 prio2band[TC_PRIO_MAX + 1] = {
598 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
601 /* 3-band FIFO queue: old style, but should be a bit faster than
602 generic prio+fifo combination.
605 #define PFIFO_FAST_BANDS 3
608 * Private data for a pfifo_fast scheduler containing:
609 * - rings for priority bands
611 struct pfifo_fast_priv {
612 struct skb_array q[PFIFO_FAST_BANDS];
615 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
618 return &priv->q[band];
621 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
622 struct sk_buff **to_free)
624 int band = prio2band[skb->priority & TC_PRIO_MAX];
625 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
626 struct skb_array *q = band2list(priv, band);
627 unsigned int pkt_len = qdisc_pkt_len(skb);
630 err = skb_array_produce(q, skb);
633 if (qdisc_is_percpu_stats(qdisc))
634 return qdisc_drop_cpu(skb, qdisc, to_free);
636 return qdisc_drop(skb, qdisc, to_free);
639 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
640 return NET_XMIT_SUCCESS;
643 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
645 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
646 struct sk_buff *skb = NULL;
649 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
650 struct skb_array *q = band2list(priv, band);
652 if (__skb_array_empty(q))
655 skb = __skb_array_consume(q);
658 qdisc_update_stats_at_dequeue(qdisc, skb);
666 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
668 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
669 struct sk_buff *skb = NULL;
672 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
673 struct skb_array *q = band2list(priv, band);
675 skb = __skb_array_peek(q);
681 static void pfifo_fast_reset(struct Qdisc *qdisc)
684 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
686 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
687 struct skb_array *q = band2list(priv, band);
690 /* NULL ring is possible if destroy path is due to a failed
691 * skb_array_init() in pfifo_fast_init() case.
696 while ((skb = __skb_array_consume(q)) != NULL)
700 if (qdisc_is_percpu_stats(qdisc)) {
701 for_each_possible_cpu(i) {
702 struct gnet_stats_queue *q;
704 q = per_cpu_ptr(qdisc->cpu_qstats, i);
711 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
713 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
715 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
716 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
717 goto nla_put_failure;
724 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
725 struct netlink_ext_ack *extack)
727 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
728 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
731 /* guard against zero length rings */
735 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
736 struct skb_array *q = band2list(priv, prio);
739 err = skb_array_init(q, qlen, GFP_KERNEL);
744 /* Can by-pass the queue discipline */
745 qdisc->flags |= TCQ_F_CAN_BYPASS;
749 static void pfifo_fast_destroy(struct Qdisc *sch)
751 struct pfifo_fast_priv *priv = qdisc_priv(sch);
754 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
755 struct skb_array *q = band2list(priv, prio);
757 /* NULL ring is possible if destroy path is due to a failed
758 * skb_array_init() in pfifo_fast_init() case.
762 /* Destroy ring but no need to kfree_skb because a call to
763 * pfifo_fast_reset() has already done that work.
765 ptr_ring_cleanup(&q->ring, NULL);
769 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
770 unsigned int new_len)
772 struct pfifo_fast_priv *priv = qdisc_priv(sch);
773 struct skb_array *bands[PFIFO_FAST_BANDS];
776 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
777 struct skb_array *q = band2list(priv, prio);
782 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
786 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
788 .priv_size = sizeof(struct pfifo_fast_priv),
789 .enqueue = pfifo_fast_enqueue,
790 .dequeue = pfifo_fast_dequeue,
791 .peek = pfifo_fast_peek,
792 .init = pfifo_fast_init,
793 .destroy = pfifo_fast_destroy,
794 .reset = pfifo_fast_reset,
795 .dump = pfifo_fast_dump,
796 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
797 .owner = THIS_MODULE,
798 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
800 EXPORT_SYMBOL(pfifo_fast_ops);
802 static struct lock_class_key qdisc_tx_busylock;
803 static struct lock_class_key qdisc_running_key;
805 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
806 const struct Qdisc_ops *ops,
807 struct netlink_ext_ack *extack)
811 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
813 struct net_device *dev;
816 NL_SET_ERR_MSG(extack, "No device queue given");
821 dev = dev_queue->dev;
822 p = kzalloc_node(size, GFP_KERNEL,
823 netdev_queue_numa_node_read(dev_queue));
827 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
828 /* if we got non aligned memory, ask more and do alignment ourself */
831 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
832 netdev_queue_numa_node_read(dev_queue));
835 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
836 sch->padded = (char *) sch - (char *) p;
838 __skb_queue_head_init(&sch->gso_skb);
839 __skb_queue_head_init(&sch->skb_bad_txq);
840 qdisc_skb_head_init(&sch->q);
841 spin_lock_init(&sch->q.lock);
843 if (ops->static_flags & TCQ_F_CPUSTATS) {
845 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
846 if (!sch->cpu_bstats)
849 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
850 if (!sch->cpu_qstats) {
851 free_percpu(sch->cpu_bstats);
856 spin_lock_init(&sch->busylock);
857 lockdep_set_class(&sch->busylock,
858 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
860 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
861 spin_lock_init(&sch->seqlock);
862 lockdep_set_class(&sch->busylock,
863 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
865 seqcount_init(&sch->running);
866 lockdep_set_class(&sch->running,
867 dev->qdisc_running_key ?: &qdisc_running_key);
870 sch->flags = ops->static_flags;
871 sch->enqueue = ops->enqueue;
872 sch->dequeue = ops->dequeue;
873 sch->dev_queue = dev_queue;
876 refcount_set(&sch->refcnt, 1);
885 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
886 const struct Qdisc_ops *ops,
887 unsigned int parentid,
888 struct netlink_ext_ack *extack)
892 if (!try_module_get(ops->owner)) {
893 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
897 sch = qdisc_alloc(dev_queue, ops, extack);
899 module_put(ops->owner);
902 sch->parent = parentid;
904 if (!ops->init || ops->init(sch, NULL, extack) == 0)
910 EXPORT_SYMBOL(qdisc_create_dflt);
912 /* Under qdisc_lock(qdisc) and BH! */
914 void qdisc_reset(struct Qdisc *qdisc)
916 const struct Qdisc_ops *ops = qdisc->ops;
917 struct sk_buff *skb, *tmp;
922 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
923 __skb_unlink(skb, &qdisc->gso_skb);
927 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
928 __skb_unlink(skb, &qdisc->skb_bad_txq);
933 qdisc->qstats.backlog = 0;
935 EXPORT_SYMBOL(qdisc_reset);
937 void qdisc_free(struct Qdisc *qdisc)
939 if (qdisc_is_percpu_stats(qdisc)) {
940 free_percpu(qdisc->cpu_bstats);
941 free_percpu(qdisc->cpu_qstats);
944 kfree((char *) qdisc - qdisc->padded);
947 static void qdisc_free_cb(struct rcu_head *head)
949 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
954 static void qdisc_destroy(struct Qdisc *qdisc)
956 const struct Qdisc_ops *ops = qdisc->ops;
957 struct sk_buff *skb, *tmp;
959 #ifdef CONFIG_NET_SCHED
960 qdisc_hash_del(qdisc);
962 qdisc_put_stab(rtnl_dereference(qdisc->stab));
964 gen_kill_estimator(&qdisc->rate_est);
970 module_put(ops->owner);
971 dev_put(qdisc_dev(qdisc));
973 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
974 __skb_unlink(skb, &qdisc->gso_skb);
978 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
979 __skb_unlink(skb, &qdisc->skb_bad_txq);
983 call_rcu(&qdisc->rcu, qdisc_free_cb);
986 void qdisc_put(struct Qdisc *qdisc)
988 if (qdisc->flags & TCQ_F_BUILTIN ||
989 !refcount_dec_and_test(&qdisc->refcnt))
992 qdisc_destroy(qdisc);
994 EXPORT_SYMBOL(qdisc_put);
996 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
997 * Intended to be used as optimization, this function only takes rtnl lock if
998 * qdisc reference counter reached zero.
1001 void qdisc_put_unlocked(struct Qdisc *qdisc)
1003 if (qdisc->flags & TCQ_F_BUILTIN ||
1004 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1007 qdisc_destroy(qdisc);
1010 EXPORT_SYMBOL(qdisc_put_unlocked);
1012 /* Attach toplevel qdisc to device queue. */
1013 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1014 struct Qdisc *qdisc)
1016 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1017 spinlock_t *root_lock;
1019 root_lock = qdisc_lock(oqdisc);
1020 spin_lock_bh(root_lock);
1022 /* ... and graft new one */
1024 qdisc = &noop_qdisc;
1025 dev_queue->qdisc_sleeping = qdisc;
1026 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1028 spin_unlock_bh(root_lock);
1032 EXPORT_SYMBOL(dev_graft_qdisc);
1034 static void attach_one_default_qdisc(struct net_device *dev,
1035 struct netdev_queue *dev_queue,
1038 struct Qdisc *qdisc;
1039 const struct Qdisc_ops *ops = default_qdisc_ops;
1041 if (dev->priv_flags & IFF_NO_QUEUE)
1042 ops = &noqueue_qdisc_ops;
1044 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1046 netdev_info(dev, "activation failed\n");
1049 if (!netif_is_multiqueue(dev))
1050 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1051 dev_queue->qdisc_sleeping = qdisc;
1054 static void attach_default_qdiscs(struct net_device *dev)
1056 struct netdev_queue *txq;
1057 struct Qdisc *qdisc;
1059 txq = netdev_get_tx_queue(dev, 0);
1061 if (!netif_is_multiqueue(dev) ||
1062 dev->priv_flags & IFF_NO_QUEUE) {
1063 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1064 dev->qdisc = txq->qdisc_sleeping;
1065 qdisc_refcount_inc(dev->qdisc);
1067 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1070 qdisc->ops->attach(qdisc);
1073 #ifdef CONFIG_NET_SCHED
1074 if (dev->qdisc != &noop_qdisc)
1075 qdisc_hash_add(dev->qdisc, false);
1079 static void transition_one_qdisc(struct net_device *dev,
1080 struct netdev_queue *dev_queue,
1081 void *_need_watchdog)
1083 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1084 int *need_watchdog_p = _need_watchdog;
1086 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1087 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1089 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1090 if (need_watchdog_p) {
1091 dev_queue->trans_start = 0;
1092 *need_watchdog_p = 1;
1096 void dev_activate(struct net_device *dev)
1100 /* No queueing discipline is attached to device;
1101 * create default one for devices, which need queueing
1102 * and noqueue_qdisc for virtual interfaces
1105 if (dev->qdisc == &noop_qdisc)
1106 attach_default_qdiscs(dev);
1108 if (!netif_carrier_ok(dev))
1109 /* Delay activation until next carrier-on event */
1113 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1114 if (dev_ingress_queue(dev))
1115 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1117 if (need_watchdog) {
1118 netif_trans_update(dev);
1119 dev_watchdog_up(dev);
1122 EXPORT_SYMBOL(dev_activate);
1124 static void dev_deactivate_queue(struct net_device *dev,
1125 struct netdev_queue *dev_queue,
1126 void *_qdisc_default)
1128 struct Qdisc *qdisc_default = _qdisc_default;
1129 struct Qdisc *qdisc;
1131 qdisc = rtnl_dereference(dev_queue->qdisc);
1133 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1136 spin_lock_bh(&qdisc->seqlock);
1137 spin_lock_bh(qdisc_lock(qdisc));
1139 if (!(qdisc->flags & TCQ_F_BUILTIN))
1140 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1142 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1145 spin_unlock_bh(qdisc_lock(qdisc));
1147 spin_unlock_bh(&qdisc->seqlock);
1151 static bool some_qdisc_is_busy(struct net_device *dev)
1155 for (i = 0; i < dev->num_tx_queues; i++) {
1156 struct netdev_queue *dev_queue;
1157 spinlock_t *root_lock;
1161 dev_queue = netdev_get_tx_queue(dev, i);
1162 q = dev_queue->qdisc_sleeping;
1164 root_lock = qdisc_lock(q);
1165 spin_lock_bh(root_lock);
1167 val = (qdisc_is_running(q) ||
1168 test_bit(__QDISC_STATE_SCHED, &q->state));
1170 spin_unlock_bh(root_lock);
1178 static void dev_qdisc_reset(struct net_device *dev,
1179 struct netdev_queue *dev_queue,
1182 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1189 * dev_deactivate_many - deactivate transmissions on several devices
1190 * @head: list of devices to deactivate
1192 * This function returns only when all outstanding transmissions
1193 * have completed, unless all devices are in dismantle phase.
1195 void dev_deactivate_many(struct list_head *head)
1197 struct net_device *dev;
1199 list_for_each_entry(dev, head, close_list) {
1200 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1202 if (dev_ingress_queue(dev))
1203 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1206 dev_watchdog_down(dev);
1209 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1210 * This is avoided if all devices are in dismantle phase :
1211 * Caller will call synchronize_net() for us
1215 /* Wait for outstanding qdisc_run calls. */
1216 list_for_each_entry(dev, head, close_list) {
1217 while (some_qdisc_is_busy(dev))
1219 /* The new qdisc is assigned at this point so we can safely
1220 * unwind stale skb lists and qdisc statistics
1222 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1223 if (dev_ingress_queue(dev))
1224 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1228 void dev_deactivate(struct net_device *dev)
1232 list_add(&dev->close_list, &single);
1233 dev_deactivate_many(&single);
1236 EXPORT_SYMBOL(dev_deactivate);
1238 static int qdisc_change_tx_queue_len(struct net_device *dev,
1239 struct netdev_queue *dev_queue)
1241 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1242 const struct Qdisc_ops *ops = qdisc->ops;
1244 if (ops->change_tx_queue_len)
1245 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1249 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1251 bool up = dev->flags & IFF_UP;
1256 dev_deactivate(dev);
1258 for (i = 0; i < dev->num_tx_queues; i++) {
1259 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1261 /* TODO: revert changes on a partial failure */
1271 static void dev_init_scheduler_queue(struct net_device *dev,
1272 struct netdev_queue *dev_queue,
1275 struct Qdisc *qdisc = _qdisc;
1277 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1278 dev_queue->qdisc_sleeping = qdisc;
1281 void dev_init_scheduler(struct net_device *dev)
1283 dev->qdisc = &noop_qdisc;
1284 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1288 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1291 static void shutdown_scheduler_queue(struct net_device *dev,
1292 struct netdev_queue *dev_queue,
1293 void *_qdisc_default)
1295 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1296 struct Qdisc *qdisc_default = _qdisc_default;
1299 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1300 dev_queue->qdisc_sleeping = qdisc_default;
1306 void dev_shutdown(struct net_device *dev)
1308 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1309 if (dev_ingress_queue(dev))
1310 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311 qdisc_put(dev->qdisc);
1312 dev->qdisc = &noop_qdisc;
1314 WARN_ON(timer_pending(&dev->watchdog_timer));
1317 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1318 const struct tc_ratespec *conf,
1321 memset(r, 0, sizeof(*r));
1322 r->overhead = conf->overhead;
1323 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1324 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1327 * The deal here is to replace a divide by a reciprocal one
1328 * in fast path (a reciprocal divide is a multiply and a shift)
1330 * Normal formula would be :
1331 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1333 * We compute mult/shift to use instead :
1334 * time_in_ns = (len * mult) >> shift;
1336 * We try to get the highest possible mult value for accuracy,
1337 * but have to make sure no overflows will ever happen.
1339 if (r->rate_bytes_ps > 0) {
1340 u64 factor = NSEC_PER_SEC;
1343 r->mult = div64_u64(factor, r->rate_bytes_ps);
1344 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1351 EXPORT_SYMBOL(psched_ratecfg_precompute);
1353 static void mini_qdisc_rcu_func(struct rcu_head *head)
1357 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1358 struct tcf_proto *tp_head)
1360 /* Protected with chain0->filter_chain_lock.
1361 * Can't access chain directly because tp_head can be NULL.
1363 struct mini_Qdisc *miniq_old =
1364 rcu_dereference_protected(*miniqp->p_miniq, 1);
1365 struct mini_Qdisc *miniq;
1368 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1369 /* Wait for flying RCU callback before it is freed. */
1374 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1375 &miniqp->miniq1 : &miniqp->miniq2;
1377 /* We need to make sure that readers won't see the miniq
1378 * we are about to modify. So wait until previous call_rcu callback
1382 miniq->filter_list = tp_head;
1383 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1386 /* This is counterpart of the rcu barriers above. We need to
1387 * block potential new user of miniq_old until all readers
1388 * are not seeing it.
1390 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1392 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1394 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1395 struct mini_Qdisc __rcu **p_miniq)
1397 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1398 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1399 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1400 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1401 miniqp->p_miniq = p_miniq;
1403 EXPORT_SYMBOL(mini_qdisc_pair_init);