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 <net/hotdata.h>
31 #include <trace/events/qdisc.h>
32 #include <trace/events/net.h>
35 /* Qdisc to use by default */
36 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
37 EXPORT_SYMBOL(default_qdisc_ops);
39 static void qdisc_maybe_clear_missed(struct Qdisc *q,
40 const struct netdev_queue *txq)
42 clear_bit(__QDISC_STATE_MISSED, &q->state);
44 /* Make sure the below netif_xmit_frozen_or_stopped()
45 * checking happens after clearing STATE_MISSED.
47 smp_mb__after_atomic();
49 /* Checking netif_xmit_frozen_or_stopped() again to
50 * make sure STATE_MISSED is set if the STATE_MISSED
51 * set by netif_tx_wake_queue()'s rescheduling of
52 * net_tx_action() is cleared by the above clear_bit().
54 if (!netif_xmit_frozen_or_stopped(txq))
55 set_bit(__QDISC_STATE_MISSED, &q->state);
57 set_bit(__QDISC_STATE_DRAINING, &q->state);
60 /* Main transmission queue. */
62 /* Modifications to data participating in scheduling must be protected with
63 * qdisc_lock(qdisc) spinlock.
65 * The idea is the following:
66 * - enqueue, dequeue are serialized via qdisc root lock
67 * - ingress filtering is also serialized via qdisc root lock
68 * - updates to tree and tree walking are only done under the rtnl mutex.
71 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
73 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
75 const struct netdev_queue *txq = q->dev_queue;
76 spinlock_t *lock = NULL;
79 if (q->flags & TCQ_F_NOLOCK) {
84 skb = skb_peek(&q->skb_bad_txq);
86 /* check the reason of requeuing without tx lock first */
87 txq = skb_get_tx_queue(txq->dev, skb);
88 if (!netif_xmit_frozen_or_stopped(txq)) {
89 skb = __skb_dequeue(&q->skb_bad_txq);
90 if (qdisc_is_percpu_stats(q)) {
91 qdisc_qstats_cpu_backlog_dec(q, skb);
92 qdisc_qstats_cpu_qlen_dec(q);
94 qdisc_qstats_backlog_dec(q, skb);
99 qdisc_maybe_clear_missed(q, txq);
109 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
111 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
114 skb = __skb_dequeue_bad_txq(q);
119 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
122 spinlock_t *lock = NULL;
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
129 __skb_queue_tail(&q->skb_bad_txq, skb);
131 if (qdisc_is_percpu_stats(q)) {
132 qdisc_qstats_cpu_backlog_inc(q, skb);
133 qdisc_qstats_cpu_qlen_inc(q);
135 qdisc_qstats_backlog_inc(q, skb);
143 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
145 spinlock_t *lock = NULL;
147 if (q->flags & TCQ_F_NOLOCK) {
148 lock = qdisc_lock(q);
153 struct sk_buff *next = skb->next;
155 __skb_queue_tail(&q->gso_skb, skb);
157 /* it's still part of the queue */
158 if (qdisc_is_percpu_stats(q)) {
159 qdisc_qstats_cpu_requeues_inc(q);
160 qdisc_qstats_cpu_backlog_inc(q, skb);
161 qdisc_qstats_cpu_qlen_inc(q);
163 q->qstats.requeues++;
164 qdisc_qstats_backlog_inc(q, skb);
173 set_bit(__QDISC_STATE_MISSED, &q->state);
179 static void try_bulk_dequeue_skb(struct Qdisc *q,
181 const struct netdev_queue *txq,
184 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
186 while (bytelimit > 0) {
187 struct sk_buff *nskb = q->dequeue(q);
192 bytelimit -= nskb->len; /* covers GSO len */
195 (*packets)++; /* GSO counts as one pkt */
197 skb_mark_not_on_list(skb);
200 /* This variant of try_bulk_dequeue_skb() makes sure
201 * all skbs in the chain are for the same txq
203 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
207 int mapping = skb_get_queue_mapping(skb);
208 struct sk_buff *nskb;
212 nskb = q->dequeue(q);
215 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
216 qdisc_enqueue_skb_bad_txq(q, nskb);
223 skb_mark_not_on_list(skb);
226 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
227 * A requeued skb (via q->gso_skb) can also be a SKB list.
229 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
232 const struct netdev_queue *txq = q->dev_queue;
233 struct sk_buff *skb = NULL;
236 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
237 spinlock_t *lock = NULL;
239 if (q->flags & TCQ_F_NOLOCK) {
240 lock = qdisc_lock(q);
244 skb = skb_peek(&q->gso_skb);
246 /* skb may be null if another cpu pulls gso_skb off in between
247 * empty check and lock.
255 /* skb in gso_skb were already validated */
257 if (xfrm_offload(skb))
259 /* check the reason of requeuing without tx lock first */
260 txq = skb_get_tx_queue(txq->dev, skb);
261 if (!netif_xmit_frozen_or_stopped(txq)) {
262 skb = __skb_dequeue(&q->gso_skb);
263 if (qdisc_is_percpu_stats(q)) {
264 qdisc_qstats_cpu_backlog_dec(q, skb);
265 qdisc_qstats_cpu_qlen_dec(q);
267 qdisc_qstats_backlog_dec(q, skb);
272 qdisc_maybe_clear_missed(q, txq);
281 if ((q->flags & TCQ_F_ONETXQUEUE) &&
282 netif_xmit_frozen_or_stopped(txq)) {
283 qdisc_maybe_clear_missed(q, txq);
287 skb = qdisc_dequeue_skb_bad_txq(q);
289 if (skb == SKB_XOFF_MAGIC)
296 if (qdisc_may_bulk(q))
297 try_bulk_dequeue_skb(q, skb, txq, packets);
299 try_bulk_dequeue_skb_slow(q, skb, packets);
302 trace_qdisc_dequeue(q, txq, *packets, skb);
307 * Transmit possibly several skbs, and handle the return status as
308 * required. Owning qdisc running bit guarantees that only one CPU
309 * can execute this function.
311 * Returns to the caller:
312 * false - hardware queue frozen backoff
313 * true - feel free to send more pkts
315 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
316 struct net_device *dev, struct netdev_queue *txq,
317 spinlock_t *root_lock, bool validate)
319 int ret = NETDEV_TX_BUSY;
322 /* And release qdisc */
324 spin_unlock(root_lock);
326 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
328 skb = validate_xmit_skb_list(skb, dev, &again);
330 #ifdef CONFIG_XFRM_OFFLOAD
331 if (unlikely(again)) {
333 spin_lock(root_lock);
335 dev_requeue_skb(skb, q);
341 HARD_TX_LOCK(dev, txq, smp_processor_id());
342 if (!netif_xmit_frozen_or_stopped(txq))
343 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
345 qdisc_maybe_clear_missed(q, txq);
347 HARD_TX_UNLOCK(dev, txq);
350 spin_lock(root_lock);
355 spin_lock(root_lock);
357 if (!dev_xmit_complete(ret)) {
358 /* Driver returned NETDEV_TX_BUSY - requeue skb */
359 if (unlikely(ret != NETDEV_TX_BUSY))
360 net_warn_ratelimited("BUG %s code %d qlen %d\n",
361 dev->name, ret, q->q.qlen);
363 dev_requeue_skb(skb, q);
371 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
373 * running seqcount guarantees only one CPU can process
374 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
377 * netif_tx_lock serializes accesses to device driver.
379 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
380 * if one is grabbed, another must be free.
382 * Note, that this procedure can be called by a watchdog timer
384 * Returns to the caller:
385 * 0 - queue is empty or throttled.
386 * >0 - queue is not empty.
389 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
391 spinlock_t *root_lock = NULL;
392 struct netdev_queue *txq;
393 struct net_device *dev;
398 skb = dequeue_skb(q, &validate, packets);
402 if (!(q->flags & TCQ_F_NOLOCK))
403 root_lock = qdisc_lock(q);
406 txq = skb_get_tx_queue(dev, skb);
408 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
411 void __qdisc_run(struct Qdisc *q)
413 int quota = READ_ONCE(net_hotdata.dev_tx_weight);
416 while (qdisc_restart(q, &packets)) {
419 if (q->flags & TCQ_F_NOLOCK)
420 set_bit(__QDISC_STATE_MISSED, &q->state);
429 unsigned long dev_trans_start(struct net_device *dev)
431 unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
435 for (i = 1; i < dev->num_tx_queues; i++) {
436 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
437 if (val && time_after(val, res))
443 EXPORT_SYMBOL(dev_trans_start);
445 static void netif_freeze_queues(struct net_device *dev)
450 cpu = smp_processor_id();
451 for (i = 0; i < dev->num_tx_queues; i++) {
452 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
454 /* We are the only thread of execution doing a
455 * freeze, but we have to grab the _xmit_lock in
456 * order to synchronize with threads which are in
457 * the ->hard_start_xmit() handler and already
458 * checked the frozen bit.
460 __netif_tx_lock(txq, cpu);
461 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
462 __netif_tx_unlock(txq);
466 void netif_tx_lock(struct net_device *dev)
468 spin_lock(&dev->tx_global_lock);
469 netif_freeze_queues(dev);
471 EXPORT_SYMBOL(netif_tx_lock);
473 static void netif_unfreeze_queues(struct net_device *dev)
477 for (i = 0; i < dev->num_tx_queues; i++) {
478 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
480 /* No need to grab the _xmit_lock here. If the
481 * queue is not stopped for another reason, we
484 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
485 netif_schedule_queue(txq);
489 void netif_tx_unlock(struct net_device *dev)
491 netif_unfreeze_queues(dev);
492 spin_unlock(&dev->tx_global_lock);
494 EXPORT_SYMBOL(netif_tx_unlock);
496 static void dev_watchdog(struct timer_list *t)
498 struct net_device *dev = from_timer(dev, t, watchdog_timer);
501 spin_lock(&dev->tx_global_lock);
502 if (!qdisc_tx_is_noop(dev)) {
503 if (netif_device_present(dev) &&
504 netif_running(dev) &&
505 netif_carrier_ok(dev)) {
506 unsigned int timedout_ms = 0;
508 unsigned long trans_start;
510 for (i = 0; i < dev->num_tx_queues; i++) {
511 struct netdev_queue *txq;
513 txq = netdev_get_tx_queue(dev, i);
514 trans_start = READ_ONCE(txq->trans_start);
515 if (netif_xmit_stopped(txq) &&
516 time_after(jiffies, (trans_start +
517 dev->watchdog_timeo))) {
518 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
519 atomic_long_inc(&txq->trans_timeout);
524 if (unlikely(timedout_ms)) {
525 trace_net_dev_xmit_timeout(dev, i);
526 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
527 raw_smp_processor_id(),
529 netif_freeze_queues(dev);
530 dev->netdev_ops->ndo_tx_timeout(dev, i);
531 netif_unfreeze_queues(dev);
533 if (!mod_timer(&dev->watchdog_timer,
534 round_jiffies(jiffies +
535 dev->watchdog_timeo)))
539 spin_unlock(&dev->tx_global_lock);
542 netdev_put(dev, &dev->watchdog_dev_tracker);
545 void __netdev_watchdog_up(struct net_device *dev)
547 if (dev->netdev_ops->ndo_tx_timeout) {
548 if (dev->watchdog_timeo <= 0)
549 dev->watchdog_timeo = 5*HZ;
550 if (!mod_timer(&dev->watchdog_timer,
551 round_jiffies(jiffies + dev->watchdog_timeo)))
552 netdev_hold(dev, &dev->watchdog_dev_tracker,
556 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
558 static void dev_watchdog_up(struct net_device *dev)
560 __netdev_watchdog_up(dev);
563 static void dev_watchdog_down(struct net_device *dev)
565 netif_tx_lock_bh(dev);
566 if (del_timer(&dev->watchdog_timer))
567 netdev_put(dev, &dev->watchdog_dev_tracker);
568 netif_tx_unlock_bh(dev);
572 * netif_carrier_on - set carrier
573 * @dev: network device
575 * Device has detected acquisition of carrier.
577 void netif_carrier_on(struct net_device *dev)
579 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
580 if (dev->reg_state == NETREG_UNINITIALIZED)
582 atomic_inc(&dev->carrier_up_count);
583 linkwatch_fire_event(dev);
584 if (netif_running(dev))
585 __netdev_watchdog_up(dev);
588 EXPORT_SYMBOL(netif_carrier_on);
591 * netif_carrier_off - clear carrier
592 * @dev: network device
594 * Device has detected loss of carrier.
596 void netif_carrier_off(struct net_device *dev)
598 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
599 if (dev->reg_state == NETREG_UNINITIALIZED)
601 atomic_inc(&dev->carrier_down_count);
602 linkwatch_fire_event(dev);
605 EXPORT_SYMBOL(netif_carrier_off);
608 * netif_carrier_event - report carrier state event
609 * @dev: network device
611 * Device has detected a carrier event but the carrier state wasn't changed.
612 * Use in drivers when querying carrier state asynchronously, to avoid missing
613 * events (link flaps) if link recovers before it's queried.
615 void netif_carrier_event(struct net_device *dev)
617 if (dev->reg_state == NETREG_UNINITIALIZED)
619 atomic_inc(&dev->carrier_up_count);
620 atomic_inc(&dev->carrier_down_count);
621 linkwatch_fire_event(dev);
623 EXPORT_SYMBOL_GPL(netif_carrier_event);
625 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
626 under all circumstances. It is difficult to invent anything faster or
630 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
631 struct sk_buff **to_free)
633 __qdisc_drop(skb, to_free);
637 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
642 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
645 .enqueue = noop_enqueue,
646 .dequeue = noop_dequeue,
647 .peek = noop_dequeue,
648 .owner = THIS_MODULE,
651 static struct netdev_queue noop_netdev_queue = {
652 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
653 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
656 struct Qdisc noop_qdisc = {
657 .enqueue = noop_enqueue,
658 .dequeue = noop_dequeue,
659 .flags = TCQ_F_BUILTIN,
660 .ops = &noop_qdisc_ops,
661 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
662 .dev_queue = &noop_netdev_queue,
663 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
665 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
666 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
668 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
671 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
672 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
674 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
677 EXPORT_SYMBOL(noop_qdisc);
679 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
680 struct netlink_ext_ack *extack)
682 /* register_qdisc() assigns a default of noop_enqueue if unset,
683 * but __dev_queue_xmit() treats noqueue only as such
684 * if this is NULL - so clear it here. */
685 qdisc->enqueue = NULL;
689 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
692 .init = noqueue_init,
693 .enqueue = noop_enqueue,
694 .dequeue = noop_dequeue,
695 .peek = noop_dequeue,
696 .owner = THIS_MODULE,
699 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
700 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
702 EXPORT_SYMBOL(sch_default_prio2band);
704 /* 3-band FIFO queue: old style, but should be a bit faster than
705 generic prio+fifo combination.
708 #define PFIFO_FAST_BANDS 3
711 * Private data for a pfifo_fast scheduler containing:
712 * - rings for priority bands
714 struct pfifo_fast_priv {
715 struct skb_array q[PFIFO_FAST_BANDS];
718 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
721 return &priv->q[band];
724 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
725 struct sk_buff **to_free)
727 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
728 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
729 struct skb_array *q = band2list(priv, band);
730 unsigned int pkt_len = qdisc_pkt_len(skb);
733 err = skb_array_produce(q, skb);
736 if (qdisc_is_percpu_stats(qdisc))
737 return qdisc_drop_cpu(skb, qdisc, to_free);
739 return qdisc_drop(skb, qdisc, to_free);
742 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
743 return NET_XMIT_SUCCESS;
746 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
748 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
749 struct sk_buff *skb = NULL;
750 bool need_retry = true;
754 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
755 struct skb_array *q = band2list(priv, band);
757 if (__skb_array_empty(q))
760 skb = __skb_array_consume(q);
763 qdisc_update_stats_at_dequeue(qdisc, skb);
764 } else if (need_retry &&
765 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
766 /* Delay clearing the STATE_MISSED here to reduce
767 * the overhead of the second spin_trylock() in
768 * qdisc_run_begin() and __netif_schedule() calling
769 * in qdisc_run_end().
771 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
772 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
774 /* Make sure dequeuing happens after clearing
777 smp_mb__after_atomic();
787 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
789 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
790 struct sk_buff *skb = NULL;
793 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
794 struct skb_array *q = band2list(priv, band);
796 skb = __skb_array_peek(q);
802 static void pfifo_fast_reset(struct Qdisc *qdisc)
805 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
807 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
808 struct skb_array *q = band2list(priv, band);
811 /* NULL ring is possible if destroy path is due to a failed
812 * skb_array_init() in pfifo_fast_init() case.
817 while ((skb = __skb_array_consume(q)) != NULL)
821 if (qdisc_is_percpu_stats(qdisc)) {
822 for_each_possible_cpu(i) {
823 struct gnet_stats_queue *q;
825 q = per_cpu_ptr(qdisc->cpu_qstats, i);
832 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
834 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
836 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
837 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
838 goto nla_put_failure;
845 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
846 struct netlink_ext_ack *extack)
848 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
849 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
852 /* guard against zero length rings */
856 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
857 struct skb_array *q = band2list(priv, prio);
860 err = skb_array_init(q, qlen, GFP_KERNEL);
865 /* Can by-pass the queue discipline */
866 qdisc->flags |= TCQ_F_CAN_BYPASS;
870 static void pfifo_fast_destroy(struct Qdisc *sch)
872 struct pfifo_fast_priv *priv = qdisc_priv(sch);
875 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
876 struct skb_array *q = band2list(priv, prio);
878 /* NULL ring is possible if destroy path is due to a failed
879 * skb_array_init() in pfifo_fast_init() case.
883 /* Destroy ring but no need to kfree_skb because a call to
884 * pfifo_fast_reset() has already done that work.
886 ptr_ring_cleanup(&q->ring, NULL);
890 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
891 unsigned int new_len)
893 struct pfifo_fast_priv *priv = qdisc_priv(sch);
894 struct skb_array *bands[PFIFO_FAST_BANDS];
897 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
898 struct skb_array *q = band2list(priv, prio);
903 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
907 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
909 .priv_size = sizeof(struct pfifo_fast_priv),
910 .enqueue = pfifo_fast_enqueue,
911 .dequeue = pfifo_fast_dequeue,
912 .peek = pfifo_fast_peek,
913 .init = pfifo_fast_init,
914 .destroy = pfifo_fast_destroy,
915 .reset = pfifo_fast_reset,
916 .dump = pfifo_fast_dump,
917 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
918 .owner = THIS_MODULE,
919 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
921 EXPORT_SYMBOL(pfifo_fast_ops);
923 static struct lock_class_key qdisc_tx_busylock;
925 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
926 const struct Qdisc_ops *ops,
927 struct netlink_ext_ack *extack)
930 unsigned int size = sizeof(*sch) + ops->priv_size;
932 struct net_device *dev;
935 NL_SET_ERR_MSG(extack, "No device queue given");
940 dev = dev_queue->dev;
941 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
945 __skb_queue_head_init(&sch->gso_skb);
946 __skb_queue_head_init(&sch->skb_bad_txq);
947 gnet_stats_basic_sync_init(&sch->bstats);
948 spin_lock_init(&sch->q.lock);
950 if (ops->static_flags & TCQ_F_CPUSTATS) {
952 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
953 if (!sch->cpu_bstats)
956 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
957 if (!sch->cpu_qstats) {
958 free_percpu(sch->cpu_bstats);
963 spin_lock_init(&sch->busylock);
964 lockdep_set_class(&sch->busylock,
965 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
967 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
968 spin_lock_init(&sch->seqlock);
969 lockdep_set_class(&sch->seqlock,
970 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
973 sch->flags = ops->static_flags;
974 sch->enqueue = ops->enqueue;
975 sch->dequeue = ops->dequeue;
976 sch->dev_queue = dev_queue;
978 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
979 refcount_set(&sch->refcnt, 1);
988 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
989 const struct Qdisc_ops *ops,
990 unsigned int parentid,
991 struct netlink_ext_ack *extack)
995 if (!try_module_get(ops->owner)) {
996 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1000 sch = qdisc_alloc(dev_queue, ops, extack);
1002 module_put(ops->owner);
1005 sch->parent = parentid;
1007 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1008 trace_qdisc_create(ops, dev_queue->dev, parentid);
1015 EXPORT_SYMBOL(qdisc_create_dflt);
1017 /* Under qdisc_lock(qdisc) and BH! */
1019 void qdisc_reset(struct Qdisc *qdisc)
1021 const struct Qdisc_ops *ops = qdisc->ops;
1023 trace_qdisc_reset(qdisc);
1028 __skb_queue_purge(&qdisc->gso_skb);
1029 __skb_queue_purge(&qdisc->skb_bad_txq);
1032 qdisc->qstats.backlog = 0;
1034 EXPORT_SYMBOL(qdisc_reset);
1036 void qdisc_free(struct Qdisc *qdisc)
1038 if (qdisc_is_percpu_stats(qdisc)) {
1039 free_percpu(qdisc->cpu_bstats);
1040 free_percpu(qdisc->cpu_qstats);
1046 static void qdisc_free_cb(struct rcu_head *head)
1048 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1053 static void __qdisc_destroy(struct Qdisc *qdisc)
1055 const struct Qdisc_ops *ops = qdisc->ops;
1056 struct net_device *dev = qdisc_dev(qdisc);
1058 #ifdef CONFIG_NET_SCHED
1059 qdisc_hash_del(qdisc);
1061 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1063 gen_kill_estimator(&qdisc->rate_est);
1069 ops->destroy(qdisc);
1071 module_put(ops->owner);
1072 netdev_put(dev, &qdisc->dev_tracker);
1074 trace_qdisc_destroy(qdisc);
1076 call_rcu(&qdisc->rcu, qdisc_free_cb);
1079 void qdisc_destroy(struct Qdisc *qdisc)
1081 if (qdisc->flags & TCQ_F_BUILTIN)
1084 __qdisc_destroy(qdisc);
1087 void qdisc_put(struct Qdisc *qdisc)
1092 if (qdisc->flags & TCQ_F_BUILTIN ||
1093 !refcount_dec_and_test(&qdisc->refcnt))
1096 __qdisc_destroy(qdisc);
1098 EXPORT_SYMBOL(qdisc_put);
1100 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1101 * Intended to be used as optimization, this function only takes rtnl lock if
1102 * qdisc reference counter reached zero.
1105 void qdisc_put_unlocked(struct Qdisc *qdisc)
1107 if (qdisc->flags & TCQ_F_BUILTIN ||
1108 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1111 __qdisc_destroy(qdisc);
1114 EXPORT_SYMBOL(qdisc_put_unlocked);
1116 /* Attach toplevel qdisc to device queue. */
1117 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1118 struct Qdisc *qdisc)
1120 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1121 spinlock_t *root_lock;
1123 root_lock = qdisc_lock(oqdisc);
1124 spin_lock_bh(root_lock);
1126 /* ... and graft new one */
1128 qdisc = &noop_qdisc;
1129 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1130 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1132 spin_unlock_bh(root_lock);
1136 EXPORT_SYMBOL(dev_graft_qdisc);
1138 static void shutdown_scheduler_queue(struct net_device *dev,
1139 struct netdev_queue *dev_queue,
1140 void *_qdisc_default)
1142 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1143 struct Qdisc *qdisc_default = _qdisc_default;
1146 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1147 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1153 static void attach_one_default_qdisc(struct net_device *dev,
1154 struct netdev_queue *dev_queue,
1157 struct Qdisc *qdisc;
1158 const struct Qdisc_ops *ops = default_qdisc_ops;
1160 if (dev->priv_flags & IFF_NO_QUEUE)
1161 ops = &noqueue_qdisc_ops;
1162 else if(dev->type == ARPHRD_CAN)
1163 ops = &pfifo_fast_ops;
1165 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1169 if (!netif_is_multiqueue(dev))
1170 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1171 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1174 static void attach_default_qdiscs(struct net_device *dev)
1176 struct netdev_queue *txq;
1177 struct Qdisc *qdisc;
1179 txq = netdev_get_tx_queue(dev, 0);
1181 if (!netif_is_multiqueue(dev) ||
1182 dev->priv_flags & IFF_NO_QUEUE) {
1183 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1184 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1185 rcu_assign_pointer(dev->qdisc, qdisc);
1186 qdisc_refcount_inc(qdisc);
1188 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1190 rcu_assign_pointer(dev->qdisc, qdisc);
1191 qdisc->ops->attach(qdisc);
1194 qdisc = rtnl_dereference(dev->qdisc);
1196 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1197 if (qdisc == &noop_qdisc) {
1198 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1199 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1200 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1201 dev->priv_flags |= IFF_NO_QUEUE;
1202 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1203 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1204 rcu_assign_pointer(dev->qdisc, qdisc);
1205 qdisc_refcount_inc(qdisc);
1206 dev->priv_flags ^= IFF_NO_QUEUE;
1209 #ifdef CONFIG_NET_SCHED
1210 if (qdisc != &noop_qdisc)
1211 qdisc_hash_add(qdisc, false);
1215 static void transition_one_qdisc(struct net_device *dev,
1216 struct netdev_queue *dev_queue,
1217 void *_need_watchdog)
1219 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1220 int *need_watchdog_p = _need_watchdog;
1222 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1223 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1225 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1226 if (need_watchdog_p) {
1227 WRITE_ONCE(dev_queue->trans_start, 0);
1228 *need_watchdog_p = 1;
1232 void dev_activate(struct net_device *dev)
1236 /* No queueing discipline is attached to device;
1237 * create default one for devices, which need queueing
1238 * and noqueue_qdisc for virtual interfaces
1241 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1242 attach_default_qdiscs(dev);
1244 if (!netif_carrier_ok(dev))
1245 /* Delay activation until next carrier-on event */
1249 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1250 if (dev_ingress_queue(dev))
1251 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1253 if (need_watchdog) {
1254 netif_trans_update(dev);
1255 dev_watchdog_up(dev);
1258 EXPORT_SYMBOL(dev_activate);
1260 static void qdisc_deactivate(struct Qdisc *qdisc)
1262 if (qdisc->flags & TCQ_F_BUILTIN)
1265 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1268 static void dev_deactivate_queue(struct net_device *dev,
1269 struct netdev_queue *dev_queue,
1270 void *_qdisc_default)
1272 struct Qdisc *qdisc_default = _qdisc_default;
1273 struct Qdisc *qdisc;
1275 qdisc = rtnl_dereference(dev_queue->qdisc);
1277 qdisc_deactivate(qdisc);
1278 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1282 static void dev_reset_queue(struct net_device *dev,
1283 struct netdev_queue *dev_queue,
1286 struct Qdisc *qdisc;
1289 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1293 nolock = qdisc->flags & TCQ_F_NOLOCK;
1296 spin_lock_bh(&qdisc->seqlock);
1297 spin_lock_bh(qdisc_lock(qdisc));
1301 spin_unlock_bh(qdisc_lock(qdisc));
1303 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1304 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1305 spin_unlock_bh(&qdisc->seqlock);
1309 static bool some_qdisc_is_busy(struct net_device *dev)
1313 for (i = 0; i < dev->num_tx_queues; i++) {
1314 struct netdev_queue *dev_queue;
1315 spinlock_t *root_lock;
1319 dev_queue = netdev_get_tx_queue(dev, i);
1320 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1322 root_lock = qdisc_lock(q);
1323 spin_lock_bh(root_lock);
1325 val = (qdisc_is_running(q) ||
1326 test_bit(__QDISC_STATE_SCHED, &q->state));
1328 spin_unlock_bh(root_lock);
1337 * dev_deactivate_many - deactivate transmissions on several devices
1338 * @head: list of devices to deactivate
1340 * This function returns only when all outstanding transmissions
1341 * have completed, unless all devices are in dismantle phase.
1343 void dev_deactivate_many(struct list_head *head)
1345 struct net_device *dev;
1347 list_for_each_entry(dev, head, close_list) {
1348 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1350 if (dev_ingress_queue(dev))
1351 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1354 dev_watchdog_down(dev);
1357 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1358 * outstanding qdisc enqueuing calls.
1359 * This is avoided if all devices are in dismantle phase :
1360 * Caller will call synchronize_net() for us
1364 list_for_each_entry(dev, head, close_list) {
1365 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1367 if (dev_ingress_queue(dev))
1368 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1371 /* Wait for outstanding qdisc_run calls. */
1372 list_for_each_entry(dev, head, close_list) {
1373 while (some_qdisc_is_busy(dev)) {
1374 /* wait_event() would avoid this sleep-loop but would
1375 * require expensive checks in the fast paths of packet
1376 * processing which isn't worth it.
1378 schedule_timeout_uninterruptible(1);
1383 void dev_deactivate(struct net_device *dev)
1387 list_add(&dev->close_list, &single);
1388 dev_deactivate_many(&single);
1391 EXPORT_SYMBOL(dev_deactivate);
1393 static int qdisc_change_tx_queue_len(struct net_device *dev,
1394 struct netdev_queue *dev_queue)
1396 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1397 const struct Qdisc_ops *ops = qdisc->ops;
1399 if (ops->change_tx_queue_len)
1400 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1404 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1405 unsigned int new_real_tx)
1407 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1409 if (qdisc->ops->change_real_num_tx)
1410 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1413 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1415 #ifdef CONFIG_NET_SCHED
1416 struct net_device *dev = qdisc_dev(sch);
1417 struct Qdisc *qdisc;
1420 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1421 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1422 /* Only update the default qdiscs we created,
1423 * qdiscs with handles are always hashed.
1425 if (qdisc != &noop_qdisc && !qdisc->handle)
1426 qdisc_hash_del(qdisc);
1428 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1429 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1430 if (qdisc != &noop_qdisc && !qdisc->handle)
1431 qdisc_hash_add(qdisc, false);
1435 EXPORT_SYMBOL(mq_change_real_num_tx);
1437 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1439 bool up = dev->flags & IFF_UP;
1444 dev_deactivate(dev);
1446 for (i = 0; i < dev->num_tx_queues; i++) {
1447 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1449 /* TODO: revert changes on a partial failure */
1459 static void dev_init_scheduler_queue(struct net_device *dev,
1460 struct netdev_queue *dev_queue,
1463 struct Qdisc *qdisc = _qdisc;
1465 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1466 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1469 void dev_init_scheduler(struct net_device *dev)
1471 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1472 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1473 if (dev_ingress_queue(dev))
1474 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1476 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1479 void dev_shutdown(struct net_device *dev)
1481 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1482 if (dev_ingress_queue(dev))
1483 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1484 qdisc_put(rtnl_dereference(dev->qdisc));
1485 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1487 WARN_ON(timer_pending(&dev->watchdog_timer));
1491 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1492 * @rate: Rate to compute reciprocal division values of
1493 * @mult: Multiplier for reciprocal division
1494 * @shift: Shift for reciprocal division
1496 * The multiplier and shift for reciprocal division by rate are stored
1497 * in mult and shift.
1499 * The deal here is to replace a divide by a reciprocal one
1500 * in fast path (a reciprocal divide is a multiply and a shift)
1502 * Normal formula would be :
1503 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1505 * We compute mult/shift to use instead :
1506 * time_in_ns = (len * mult) >> shift;
1508 * We try to get the highest possible mult value for accuracy,
1509 * but have to make sure no overflows will ever happen.
1511 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1513 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1515 u64 factor = NSEC_PER_SEC;
1524 *mult = div64_u64(factor, rate);
1525 if (*mult & (1U << 31) || factor & (1ULL << 63))
1532 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1533 const struct tc_ratespec *conf,
1536 memset(r, 0, sizeof(*r));
1537 r->overhead = conf->overhead;
1539 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1540 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1541 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1543 EXPORT_SYMBOL(psched_ratecfg_precompute);
1545 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1547 r->rate_pkts_ps = pktrate64;
1548 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1550 EXPORT_SYMBOL(psched_ppscfg_precompute);
1552 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1553 struct tcf_proto *tp_head)
1555 /* Protected with chain0->filter_chain_lock.
1556 * Can't access chain directly because tp_head can be NULL.
1558 struct mini_Qdisc *miniq_old =
1559 rcu_dereference_protected(*miniqp->p_miniq, 1);
1560 struct mini_Qdisc *miniq;
1563 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1565 miniq = miniq_old != &miniqp->miniq1 ?
1566 &miniqp->miniq1 : &miniqp->miniq2;
1568 /* We need to make sure that readers won't see the miniq
1569 * we are about to modify. So ensure that at least one RCU
1570 * grace period has elapsed since the miniq was made
1573 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1574 cond_synchronize_rcu(miniq->rcu_state);
1575 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1576 synchronize_rcu_expedited();
1578 miniq->filter_list = tp_head;
1579 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1583 /* This is counterpart of the rcu sync above. We need to
1584 * block potential new user of miniq_old until all readers
1585 * are not seeing it.
1587 miniq_old->rcu_state = start_poll_synchronize_rcu();
1589 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1591 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1592 struct tcf_block *block)
1594 miniqp->miniq1.block = block;
1595 miniqp->miniq2.block = block;
1597 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1599 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1600 struct mini_Qdisc __rcu **p_miniq)
1602 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1603 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1604 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1605 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1606 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1607 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1608 miniqp->p_miniq = p_miniq;
1610 EXPORT_SYMBOL(mini_qdisc_pair_init);