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 static void qdisc_maybe_clear_missed(struct Qdisc *q,
39 const struct netdev_queue *txq)
41 clear_bit(__QDISC_STATE_MISSED, &q->state);
43 /* Make sure the below netif_xmit_frozen_or_stopped()
44 * checking happens after clearing STATE_MISSED.
46 smp_mb__after_atomic();
48 /* Checking netif_xmit_frozen_or_stopped() again to
49 * make sure STATE_MISSED is set if the STATE_MISSED
50 * set by netif_tx_wake_queue()'s rescheduling of
51 * net_tx_action() is cleared by the above clear_bit().
53 if (!netif_xmit_frozen_or_stopped(txq))
54 set_bit(__QDISC_STATE_MISSED, &q->state);
56 set_bit(__QDISC_STATE_DRAINING, &q->state);
59 /* Main transmission queue. */
61 /* Modifications to data participating in scheduling must be protected with
62 * qdisc_lock(qdisc) spinlock.
64 * The idea is the following:
65 * - enqueue, dequeue are serialized via qdisc root lock
66 * - ingress filtering is also serialized via qdisc root lock
67 * - updates to tree and tree walking are only done under the rtnl mutex.
70 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
72 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
74 const struct netdev_queue *txq = q->dev_queue;
75 spinlock_t *lock = NULL;
78 if (q->flags & TCQ_F_NOLOCK) {
83 skb = skb_peek(&q->skb_bad_txq);
85 /* check the reason of requeuing without tx lock first */
86 txq = skb_get_tx_queue(txq->dev, skb);
87 if (!netif_xmit_frozen_or_stopped(txq)) {
88 skb = __skb_dequeue(&q->skb_bad_txq);
89 if (qdisc_is_percpu_stats(q)) {
90 qdisc_qstats_cpu_backlog_dec(q, skb);
91 qdisc_qstats_cpu_qlen_dec(q);
93 qdisc_qstats_backlog_dec(q, skb);
98 qdisc_maybe_clear_missed(q, txq);
108 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
110 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
113 skb = __skb_dequeue_bad_txq(q);
118 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
121 spinlock_t *lock = NULL;
123 if (q->flags & TCQ_F_NOLOCK) {
124 lock = qdisc_lock(q);
128 __skb_queue_tail(&q->skb_bad_txq, skb);
130 if (qdisc_is_percpu_stats(q)) {
131 qdisc_qstats_cpu_backlog_inc(q, skb);
132 qdisc_qstats_cpu_qlen_inc(q);
134 qdisc_qstats_backlog_inc(q, skb);
142 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
144 spinlock_t *lock = NULL;
146 if (q->flags & TCQ_F_NOLOCK) {
147 lock = qdisc_lock(q);
152 struct sk_buff *next = skb->next;
154 __skb_queue_tail(&q->gso_skb, skb);
156 /* it's still part of the queue */
157 if (qdisc_is_percpu_stats(q)) {
158 qdisc_qstats_cpu_requeues_inc(q);
159 qdisc_qstats_cpu_backlog_inc(q, skb);
160 qdisc_qstats_cpu_qlen_inc(q);
162 q->qstats.requeues++;
163 qdisc_qstats_backlog_inc(q, skb);
172 set_bit(__QDISC_STATE_MISSED, &q->state);
178 static void try_bulk_dequeue_skb(struct Qdisc *q,
180 const struct netdev_queue *txq,
183 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
185 while (bytelimit > 0) {
186 struct sk_buff *nskb = q->dequeue(q);
191 bytelimit -= nskb->len; /* covers GSO len */
194 (*packets)++; /* GSO counts as one pkt */
196 skb_mark_not_on_list(skb);
199 /* This variant of try_bulk_dequeue_skb() makes sure
200 * all skbs in the chain are for the same txq
202 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
206 int mapping = skb_get_queue_mapping(skb);
207 struct sk_buff *nskb;
211 nskb = q->dequeue(q);
214 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
215 qdisc_enqueue_skb_bad_txq(q, nskb);
222 skb_mark_not_on_list(skb);
225 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
226 * A requeued skb (via q->gso_skb) can also be a SKB list.
228 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
231 const struct netdev_queue *txq = q->dev_queue;
232 struct sk_buff *skb = NULL;
235 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
236 spinlock_t *lock = NULL;
238 if (q->flags & TCQ_F_NOLOCK) {
239 lock = qdisc_lock(q);
243 skb = skb_peek(&q->gso_skb);
245 /* skb may be null if another cpu pulls gso_skb off in between
246 * empty check and lock.
254 /* skb in gso_skb were already validated */
256 if (xfrm_offload(skb))
258 /* check the reason of requeuing without tx lock first */
259 txq = skb_get_tx_queue(txq->dev, skb);
260 if (!netif_xmit_frozen_or_stopped(txq)) {
261 skb = __skb_dequeue(&q->gso_skb);
262 if (qdisc_is_percpu_stats(q)) {
263 qdisc_qstats_cpu_backlog_dec(q, skb);
264 qdisc_qstats_cpu_qlen_dec(q);
266 qdisc_qstats_backlog_dec(q, skb);
271 qdisc_maybe_clear_missed(q, txq);
280 if ((q->flags & TCQ_F_ONETXQUEUE) &&
281 netif_xmit_frozen_or_stopped(txq)) {
282 qdisc_maybe_clear_missed(q, txq);
286 skb = qdisc_dequeue_skb_bad_txq(q);
288 if (skb == SKB_XOFF_MAGIC)
295 if (qdisc_may_bulk(q))
296 try_bulk_dequeue_skb(q, skb, txq, packets);
298 try_bulk_dequeue_skb_slow(q, skb, packets);
301 trace_qdisc_dequeue(q, txq, *packets, skb);
306 * Transmit possibly several skbs, and handle the return status as
307 * required. Owning qdisc running bit guarantees that only one CPU
308 * can execute this function.
310 * Returns to the caller:
311 * false - hardware queue frozen backoff
312 * true - feel free to send more pkts
314 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
315 struct net_device *dev, struct netdev_queue *txq,
316 spinlock_t *root_lock, bool validate)
318 int ret = NETDEV_TX_BUSY;
321 /* And release qdisc */
323 spin_unlock(root_lock);
325 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
327 skb = validate_xmit_skb_list(skb, dev, &again);
329 #ifdef CONFIG_XFRM_OFFLOAD
330 if (unlikely(again)) {
332 spin_lock(root_lock);
334 dev_requeue_skb(skb, q);
340 HARD_TX_LOCK(dev, txq, smp_processor_id());
341 if (!netif_xmit_frozen_or_stopped(txq))
342 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
344 qdisc_maybe_clear_missed(q, txq);
346 HARD_TX_UNLOCK(dev, txq);
349 spin_lock(root_lock);
354 spin_lock(root_lock);
356 if (!dev_xmit_complete(ret)) {
357 /* Driver returned NETDEV_TX_BUSY - requeue skb */
358 if (unlikely(ret != NETDEV_TX_BUSY))
359 net_warn_ratelimited("BUG %s code %d qlen %d\n",
360 dev->name, ret, q->q.qlen);
362 dev_requeue_skb(skb, q);
370 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
372 * running seqcount guarantees only one CPU can process
373 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
376 * netif_tx_lock serializes accesses to device driver.
378 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
379 * if one is grabbed, another must be free.
381 * Note, that this procedure can be called by a watchdog timer
383 * Returns to the caller:
384 * 0 - queue is empty or throttled.
385 * >0 - queue is not empty.
388 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
390 spinlock_t *root_lock = NULL;
391 struct netdev_queue *txq;
392 struct net_device *dev;
397 skb = dequeue_skb(q, &validate, packets);
401 if (!(q->flags & TCQ_F_NOLOCK))
402 root_lock = qdisc_lock(q);
405 txq = skb_get_tx_queue(dev, skb);
407 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
410 void __qdisc_run(struct Qdisc *q)
412 int quota = dev_tx_weight;
415 while (qdisc_restart(q, &packets)) {
418 if (q->flags & TCQ_F_NOLOCK)
419 set_bit(__QDISC_STATE_MISSED, &q->state);
428 unsigned long dev_trans_start(struct net_device *dev)
430 unsigned long val, res;
433 if (is_vlan_dev(dev))
434 dev = vlan_dev_real_dev(dev);
435 else if (netif_is_macvlan(dev))
436 dev = macvlan_dev_real_dev(dev);
437 res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
438 for (i = 1; i < dev->num_tx_queues; i++) {
439 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
440 if (val && time_after(val, res))
446 EXPORT_SYMBOL(dev_trans_start);
448 static void netif_freeze_queues(struct net_device *dev)
453 cpu = smp_processor_id();
454 for (i = 0; i < dev->num_tx_queues; i++) {
455 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
457 /* We are the only thread of execution doing a
458 * freeze, but we have to grab the _xmit_lock in
459 * order to synchronize with threads which are in
460 * the ->hard_start_xmit() handler and already
461 * checked the frozen bit.
463 __netif_tx_lock(txq, cpu);
464 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
465 __netif_tx_unlock(txq);
469 void netif_tx_lock(struct net_device *dev)
471 spin_lock(&dev->tx_global_lock);
472 netif_freeze_queues(dev);
474 EXPORT_SYMBOL(netif_tx_lock);
476 static void netif_unfreeze_queues(struct net_device *dev)
480 for (i = 0; i < dev->num_tx_queues; i++) {
481 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
483 /* No need to grab the _xmit_lock here. If the
484 * queue is not stopped for another reason, we
487 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
488 netif_schedule_queue(txq);
492 void netif_tx_unlock(struct net_device *dev)
494 netif_unfreeze_queues(dev);
495 spin_unlock(&dev->tx_global_lock);
497 EXPORT_SYMBOL(netif_tx_unlock);
499 static void dev_watchdog(struct timer_list *t)
501 struct net_device *dev = from_timer(dev, t, watchdog_timer);
504 spin_lock(&dev->tx_global_lock);
505 if (!qdisc_tx_is_noop(dev)) {
506 if (netif_device_present(dev) &&
507 netif_running(dev) &&
508 netif_carrier_ok(dev)) {
509 int some_queue_timedout = 0;
511 unsigned long trans_start;
513 for (i = 0; i < dev->num_tx_queues; i++) {
514 struct netdev_queue *txq;
516 txq = netdev_get_tx_queue(dev, i);
517 trans_start = READ_ONCE(txq->trans_start);
518 if (netif_xmit_stopped(txq) &&
519 time_after(jiffies, (trans_start +
520 dev->watchdog_timeo))) {
521 some_queue_timedout = 1;
522 atomic_long_inc(&txq->trans_timeout);
527 if (unlikely(some_queue_timedout)) {
528 trace_net_dev_xmit_timeout(dev, i);
529 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
530 dev->name, netdev_drivername(dev), i);
531 netif_freeze_queues(dev);
532 dev->netdev_ops->ndo_tx_timeout(dev, i);
533 netif_unfreeze_queues(dev);
535 if (!mod_timer(&dev->watchdog_timer,
536 round_jiffies(jiffies +
537 dev->watchdog_timeo)))
541 spin_unlock(&dev->tx_global_lock);
544 dev_put_track(dev, &dev->watchdog_dev_tracker);
547 void __netdev_watchdog_up(struct net_device *dev)
549 if (dev->netdev_ops->ndo_tx_timeout) {
550 if (dev->watchdog_timeo <= 0)
551 dev->watchdog_timeo = 5*HZ;
552 if (!mod_timer(&dev->watchdog_timer,
553 round_jiffies(jiffies + dev->watchdog_timeo)))
554 dev_hold_track(dev, &dev->watchdog_dev_tracker, GFP_ATOMIC);
557 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
559 static void dev_watchdog_up(struct net_device *dev)
561 __netdev_watchdog_up(dev);
564 static void dev_watchdog_down(struct net_device *dev)
566 netif_tx_lock_bh(dev);
567 if (del_timer(&dev->watchdog_timer))
568 dev_put_track(dev, &dev->watchdog_dev_tracker);
569 netif_tx_unlock_bh(dev);
573 * netif_carrier_on - set carrier
574 * @dev: network device
576 * Device has detected acquisition of carrier.
578 void netif_carrier_on(struct net_device *dev)
580 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
581 if (dev->reg_state == NETREG_UNINITIALIZED)
583 atomic_inc(&dev->carrier_up_count);
584 linkwatch_fire_event(dev);
585 if (netif_running(dev))
586 __netdev_watchdog_up(dev);
589 EXPORT_SYMBOL(netif_carrier_on);
592 * netif_carrier_off - clear carrier
593 * @dev: network device
595 * Device has detected loss of carrier.
597 void netif_carrier_off(struct net_device *dev)
599 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
600 if (dev->reg_state == NETREG_UNINITIALIZED)
602 atomic_inc(&dev->carrier_down_count);
603 linkwatch_fire_event(dev);
606 EXPORT_SYMBOL(netif_carrier_off);
609 * netif_carrier_event - report carrier state event
610 * @dev: network device
612 * Device has detected a carrier event but the carrier state wasn't changed.
613 * Use in drivers when querying carrier state asynchronously, to avoid missing
614 * events (link flaps) if link recovers before it's queried.
616 void netif_carrier_event(struct net_device *dev)
618 if (dev->reg_state == NETREG_UNINITIALIZED)
620 atomic_inc(&dev->carrier_up_count);
621 atomic_inc(&dev->carrier_down_count);
622 linkwatch_fire_event(dev);
624 EXPORT_SYMBOL_GPL(netif_carrier_event);
626 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
627 under all circumstances. It is difficult to invent anything faster or
631 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
632 struct sk_buff **to_free)
634 __qdisc_drop(skb, to_free);
638 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
643 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
646 .enqueue = noop_enqueue,
647 .dequeue = noop_dequeue,
648 .peek = noop_dequeue,
649 .owner = THIS_MODULE,
652 static struct netdev_queue noop_netdev_queue = {
653 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
654 .qdisc_sleeping = &noop_qdisc,
657 struct Qdisc noop_qdisc = {
658 .enqueue = noop_enqueue,
659 .dequeue = noop_dequeue,
660 .flags = TCQ_F_BUILTIN,
661 .ops = &noop_qdisc_ops,
662 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
663 .dev_queue = &noop_netdev_queue,
664 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
666 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
667 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
669 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
672 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
673 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
675 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
678 EXPORT_SYMBOL(noop_qdisc);
680 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
681 struct netlink_ext_ack *extack)
683 /* register_qdisc() assigns a default of noop_enqueue if unset,
684 * but __dev_queue_xmit() treats noqueue only as such
685 * if this is NULL - so clear it here. */
686 qdisc->enqueue = NULL;
690 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
693 .init = noqueue_init,
694 .enqueue = noop_enqueue,
695 .dequeue = noop_dequeue,
696 .peek = noop_dequeue,
697 .owner = THIS_MODULE,
700 static const u8 prio2band[TC_PRIO_MAX + 1] = {
701 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
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 = 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, 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 qdisc_skb_head_init(&sch->q);
948 gnet_stats_basic_sync_init(&sch->bstats);
949 spin_lock_init(&sch->q.lock);
951 if (ops->static_flags & TCQ_F_CPUSTATS) {
953 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
954 if (!sch->cpu_bstats)
957 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
958 if (!sch->cpu_qstats) {
959 free_percpu(sch->cpu_bstats);
964 spin_lock_init(&sch->busylock);
965 lockdep_set_class(&sch->busylock,
966 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
968 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
969 spin_lock_init(&sch->seqlock);
970 lockdep_set_class(&sch->seqlock,
971 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
974 sch->flags = ops->static_flags;
975 sch->enqueue = ops->enqueue;
976 sch->dequeue = ops->dequeue;
977 sch->dev_queue = dev_queue;
978 dev_hold_track(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;
1057 #ifdef CONFIG_NET_SCHED
1058 qdisc_hash_del(qdisc);
1060 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1062 gen_kill_estimator(&qdisc->rate_est);
1067 ops->destroy(qdisc);
1069 module_put(ops->owner);
1070 dev_put_track(qdisc_dev(qdisc), &qdisc->dev_tracker);
1072 trace_qdisc_destroy(qdisc);
1074 call_rcu(&qdisc->rcu, qdisc_free_cb);
1077 void qdisc_put(struct Qdisc *qdisc)
1082 if (qdisc->flags & TCQ_F_BUILTIN ||
1083 !refcount_dec_and_test(&qdisc->refcnt))
1086 qdisc_destroy(qdisc);
1088 EXPORT_SYMBOL(qdisc_put);
1090 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1091 * Intended to be used as optimization, this function only takes rtnl lock if
1092 * qdisc reference counter reached zero.
1095 void qdisc_put_unlocked(struct Qdisc *qdisc)
1097 if (qdisc->flags & TCQ_F_BUILTIN ||
1098 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1101 qdisc_destroy(qdisc);
1104 EXPORT_SYMBOL(qdisc_put_unlocked);
1106 /* Attach toplevel qdisc to device queue. */
1107 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1108 struct Qdisc *qdisc)
1110 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1111 spinlock_t *root_lock;
1113 root_lock = qdisc_lock(oqdisc);
1114 spin_lock_bh(root_lock);
1116 /* ... and graft new one */
1118 qdisc = &noop_qdisc;
1119 dev_queue->qdisc_sleeping = qdisc;
1120 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1122 spin_unlock_bh(root_lock);
1126 EXPORT_SYMBOL(dev_graft_qdisc);
1128 static void attach_one_default_qdisc(struct net_device *dev,
1129 struct netdev_queue *dev_queue,
1132 struct Qdisc *qdisc;
1133 const struct Qdisc_ops *ops = default_qdisc_ops;
1135 if (dev->priv_flags & IFF_NO_QUEUE)
1136 ops = &noqueue_qdisc_ops;
1137 else if(dev->type == ARPHRD_CAN)
1138 ops = &pfifo_fast_ops;
1140 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1144 if (!netif_is_multiqueue(dev))
1145 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1146 dev_queue->qdisc_sleeping = qdisc;
1149 static void attach_default_qdiscs(struct net_device *dev)
1151 struct netdev_queue *txq;
1152 struct Qdisc *qdisc;
1154 txq = netdev_get_tx_queue(dev, 0);
1156 if (!netif_is_multiqueue(dev) ||
1157 dev->priv_flags & IFF_NO_QUEUE) {
1158 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1159 qdisc = txq->qdisc_sleeping;
1160 rcu_assign_pointer(dev->qdisc, qdisc);
1161 qdisc_refcount_inc(qdisc);
1163 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1165 rcu_assign_pointer(dev->qdisc, qdisc);
1166 qdisc->ops->attach(qdisc);
1169 qdisc = rtnl_dereference(dev->qdisc);
1171 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1172 if (qdisc == &noop_qdisc) {
1173 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1174 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1175 dev->priv_flags |= IFF_NO_QUEUE;
1176 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1177 qdisc = txq->qdisc_sleeping;
1178 rcu_assign_pointer(dev->qdisc, qdisc);
1179 qdisc_refcount_inc(qdisc);
1180 dev->priv_flags ^= IFF_NO_QUEUE;
1183 #ifdef CONFIG_NET_SCHED
1184 if (qdisc != &noop_qdisc)
1185 qdisc_hash_add(qdisc, false);
1189 static void transition_one_qdisc(struct net_device *dev,
1190 struct netdev_queue *dev_queue,
1191 void *_need_watchdog)
1193 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1194 int *need_watchdog_p = _need_watchdog;
1196 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1197 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1199 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1200 if (need_watchdog_p) {
1201 WRITE_ONCE(dev_queue->trans_start, 0);
1202 *need_watchdog_p = 1;
1206 void dev_activate(struct net_device *dev)
1210 /* No queueing discipline is attached to device;
1211 * create default one for devices, which need queueing
1212 * and noqueue_qdisc for virtual interfaces
1215 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1216 attach_default_qdiscs(dev);
1218 if (!netif_carrier_ok(dev))
1219 /* Delay activation until next carrier-on event */
1223 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1224 if (dev_ingress_queue(dev))
1225 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1227 if (need_watchdog) {
1228 netif_trans_update(dev);
1229 dev_watchdog_up(dev);
1232 EXPORT_SYMBOL(dev_activate);
1234 static void qdisc_deactivate(struct Qdisc *qdisc)
1236 if (qdisc->flags & TCQ_F_BUILTIN)
1239 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1242 static void dev_deactivate_queue(struct net_device *dev,
1243 struct netdev_queue *dev_queue,
1244 void *_qdisc_default)
1246 struct Qdisc *qdisc_default = _qdisc_default;
1247 struct Qdisc *qdisc;
1249 qdisc = rtnl_dereference(dev_queue->qdisc);
1251 qdisc_deactivate(qdisc);
1252 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1256 static void dev_reset_queue(struct net_device *dev,
1257 struct netdev_queue *dev_queue,
1260 struct Qdisc *qdisc;
1263 qdisc = dev_queue->qdisc_sleeping;
1267 nolock = qdisc->flags & TCQ_F_NOLOCK;
1270 spin_lock_bh(&qdisc->seqlock);
1271 spin_lock_bh(qdisc_lock(qdisc));
1275 spin_unlock_bh(qdisc_lock(qdisc));
1277 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1278 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1279 spin_unlock_bh(&qdisc->seqlock);
1283 static bool some_qdisc_is_busy(struct net_device *dev)
1287 for (i = 0; i < dev->num_tx_queues; i++) {
1288 struct netdev_queue *dev_queue;
1289 spinlock_t *root_lock;
1293 dev_queue = netdev_get_tx_queue(dev, i);
1294 q = dev_queue->qdisc_sleeping;
1296 root_lock = qdisc_lock(q);
1297 spin_lock_bh(root_lock);
1299 val = (qdisc_is_running(q) ||
1300 test_bit(__QDISC_STATE_SCHED, &q->state));
1302 spin_unlock_bh(root_lock);
1311 * dev_deactivate_many - deactivate transmissions on several devices
1312 * @head: list of devices to deactivate
1314 * This function returns only when all outstanding transmissions
1315 * have completed, unless all devices are in dismantle phase.
1317 void dev_deactivate_many(struct list_head *head)
1319 struct net_device *dev;
1321 list_for_each_entry(dev, head, close_list) {
1322 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1324 if (dev_ingress_queue(dev))
1325 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1328 dev_watchdog_down(dev);
1331 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1332 * outstanding qdisc enqueuing calls.
1333 * This is avoided if all devices are in dismantle phase :
1334 * Caller will call synchronize_net() for us
1338 list_for_each_entry(dev, head, close_list) {
1339 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1341 if (dev_ingress_queue(dev))
1342 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1345 /* Wait for outstanding qdisc_run calls. */
1346 list_for_each_entry(dev, head, close_list) {
1347 while (some_qdisc_is_busy(dev)) {
1348 /* wait_event() would avoid this sleep-loop but would
1349 * require expensive checks in the fast paths of packet
1350 * processing which isn't worth it.
1352 schedule_timeout_uninterruptible(1);
1357 void dev_deactivate(struct net_device *dev)
1361 list_add(&dev->close_list, &single);
1362 dev_deactivate_many(&single);
1365 EXPORT_SYMBOL(dev_deactivate);
1367 static int qdisc_change_tx_queue_len(struct net_device *dev,
1368 struct netdev_queue *dev_queue)
1370 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1371 const struct Qdisc_ops *ops = qdisc->ops;
1373 if (ops->change_tx_queue_len)
1374 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1378 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1379 unsigned int new_real_tx)
1381 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1383 if (qdisc->ops->change_real_num_tx)
1384 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1387 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1389 #ifdef CONFIG_NET_SCHED
1390 struct net_device *dev = qdisc_dev(sch);
1391 struct Qdisc *qdisc;
1394 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1395 qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1396 /* Only update the default qdiscs we created,
1397 * qdiscs with handles are always hashed.
1399 if (qdisc != &noop_qdisc && !qdisc->handle)
1400 qdisc_hash_del(qdisc);
1402 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1403 qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1404 if (qdisc != &noop_qdisc && !qdisc->handle)
1405 qdisc_hash_add(qdisc, false);
1409 EXPORT_SYMBOL(mq_change_real_num_tx);
1411 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1413 bool up = dev->flags & IFF_UP;
1418 dev_deactivate(dev);
1420 for (i = 0; i < dev->num_tx_queues; i++) {
1421 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1423 /* TODO: revert changes on a partial failure */
1433 static void dev_init_scheduler_queue(struct net_device *dev,
1434 struct netdev_queue *dev_queue,
1437 struct Qdisc *qdisc = _qdisc;
1439 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1440 dev_queue->qdisc_sleeping = qdisc;
1443 void dev_init_scheduler(struct net_device *dev)
1445 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1446 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1447 if (dev_ingress_queue(dev))
1448 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1450 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1453 static void shutdown_scheduler_queue(struct net_device *dev,
1454 struct netdev_queue *dev_queue,
1455 void *_qdisc_default)
1457 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1458 struct Qdisc *qdisc_default = _qdisc_default;
1461 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1462 dev_queue->qdisc_sleeping = qdisc_default;
1468 void dev_shutdown(struct net_device *dev)
1470 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1471 if (dev_ingress_queue(dev))
1472 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1473 qdisc_put(rtnl_dereference(dev->qdisc));
1474 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1476 WARN_ON(timer_pending(&dev->watchdog_timer));
1480 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1481 * @rate: Rate to compute reciprocal division values of
1482 * @mult: Multiplier for reciprocal division
1483 * @shift: Shift for reciprocal division
1485 * The multiplier and shift for reciprocal division by rate are stored
1486 * in mult and shift.
1488 * The deal here is to replace a divide by a reciprocal one
1489 * in fast path (a reciprocal divide is a multiply and a shift)
1491 * Normal formula would be :
1492 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1494 * We compute mult/shift to use instead :
1495 * time_in_ns = (len * mult) >> shift;
1497 * We try to get the highest possible mult value for accuracy,
1498 * but have to make sure no overflows will ever happen.
1500 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1502 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1504 u64 factor = NSEC_PER_SEC;
1513 *mult = div64_u64(factor, rate);
1514 if (*mult & (1U << 31) || factor & (1ULL << 63))
1521 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1522 const struct tc_ratespec *conf,
1525 memset(r, 0, sizeof(*r));
1526 r->overhead = conf->overhead;
1528 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1529 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1530 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1532 EXPORT_SYMBOL(psched_ratecfg_precompute);
1534 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1536 r->rate_pkts_ps = pktrate64;
1537 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1539 EXPORT_SYMBOL(psched_ppscfg_precompute);
1541 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1542 struct tcf_proto *tp_head)
1544 /* Protected with chain0->filter_chain_lock.
1545 * Can't access chain directly because tp_head can be NULL.
1547 struct mini_Qdisc *miniq_old =
1548 rcu_dereference_protected(*miniqp->p_miniq, 1);
1549 struct mini_Qdisc *miniq;
1552 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1554 miniq = miniq_old != &miniqp->miniq1 ?
1555 &miniqp->miniq1 : &miniqp->miniq2;
1557 /* We need to make sure that readers won't see the miniq
1558 * we are about to modify. So ensure that at least one RCU
1559 * grace period has elapsed since the miniq was made
1562 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1563 cond_synchronize_rcu(miniq->rcu_state);
1564 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1565 synchronize_rcu_expedited();
1567 miniq->filter_list = tp_head;
1568 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1572 /* This is counterpart of the rcu sync above. We need to
1573 * block potential new user of miniq_old until all readers
1574 * are not seeing it.
1576 miniq_old->rcu_state = start_poll_synchronize_rcu();
1578 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1580 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1581 struct tcf_block *block)
1583 miniqp->miniq1.block = block;
1584 miniqp->miniq2.block = block;
1586 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1588 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1589 struct mini_Qdisc __rcu **p_miniq)
1591 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1592 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1593 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1594 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1595 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1596 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1597 miniqp->p_miniq = p_miniq;
1599 EXPORT_SYMBOL(mini_qdisc_pair_init);