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 netdev_put(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 netdev_hold(dev, &dev->watchdog_dev_tracker,
558 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
560 static void dev_watchdog_up(struct net_device *dev)
562 __netdev_watchdog_up(dev);
565 static void dev_watchdog_down(struct net_device *dev)
567 netif_tx_lock_bh(dev);
568 if (del_timer(&dev->watchdog_timer))
569 netdev_put(dev, &dev->watchdog_dev_tracker);
570 netif_tx_unlock_bh(dev);
574 * netif_carrier_on - set carrier
575 * @dev: network device
577 * Device has detected acquisition of carrier.
579 void netif_carrier_on(struct net_device *dev)
581 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
582 if (dev->reg_state == NETREG_UNINITIALIZED)
584 atomic_inc(&dev->carrier_up_count);
585 linkwatch_fire_event(dev);
586 if (netif_running(dev))
587 __netdev_watchdog_up(dev);
590 EXPORT_SYMBOL(netif_carrier_on);
593 * netif_carrier_off - clear carrier
594 * @dev: network device
596 * Device has detected loss of carrier.
598 void netif_carrier_off(struct net_device *dev)
600 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
601 if (dev->reg_state == NETREG_UNINITIALIZED)
603 atomic_inc(&dev->carrier_down_count);
604 linkwatch_fire_event(dev);
607 EXPORT_SYMBOL(netif_carrier_off);
610 * netif_carrier_event - report carrier state event
611 * @dev: network device
613 * Device has detected a carrier event but the carrier state wasn't changed.
614 * Use in drivers when querying carrier state asynchronously, to avoid missing
615 * events (link flaps) if link recovers before it's queried.
617 void netif_carrier_event(struct net_device *dev)
619 if (dev->reg_state == NETREG_UNINITIALIZED)
621 atomic_inc(&dev->carrier_up_count);
622 atomic_inc(&dev->carrier_down_count);
623 linkwatch_fire_event(dev);
625 EXPORT_SYMBOL_GPL(netif_carrier_event);
627 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
628 under all circumstances. It is difficult to invent anything faster or
632 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
633 struct sk_buff **to_free)
635 __qdisc_drop(skb, to_free);
639 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
644 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
647 .enqueue = noop_enqueue,
648 .dequeue = noop_dequeue,
649 .peek = noop_dequeue,
650 .owner = THIS_MODULE,
653 static struct netdev_queue noop_netdev_queue = {
654 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
655 .qdisc_sleeping = &noop_qdisc,
658 struct Qdisc noop_qdisc = {
659 .enqueue = noop_enqueue,
660 .dequeue = noop_dequeue,
661 .flags = TCQ_F_BUILTIN,
662 .ops = &noop_qdisc_ops,
663 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
664 .dev_queue = &noop_netdev_queue,
665 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
667 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
668 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
670 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
673 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
674 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
676 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
679 EXPORT_SYMBOL(noop_qdisc);
681 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
682 struct netlink_ext_ack *extack)
684 /* register_qdisc() assigns a default of noop_enqueue if unset,
685 * but __dev_queue_xmit() treats noqueue only as such
686 * if this is NULL - so clear it here. */
687 qdisc->enqueue = NULL;
691 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
694 .init = noqueue_init,
695 .enqueue = noop_enqueue,
696 .dequeue = noop_dequeue,
697 .peek = noop_dequeue,
698 .owner = THIS_MODULE,
701 static const u8 prio2band[TC_PRIO_MAX + 1] = {
702 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
705 /* 3-band FIFO queue: old style, but should be a bit faster than
706 generic prio+fifo combination.
709 #define PFIFO_FAST_BANDS 3
712 * Private data for a pfifo_fast scheduler containing:
713 * - rings for priority bands
715 struct pfifo_fast_priv {
716 struct skb_array q[PFIFO_FAST_BANDS];
719 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
722 return &priv->q[band];
725 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
726 struct sk_buff **to_free)
728 int band = prio2band[skb->priority & TC_PRIO_MAX];
729 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
730 struct skb_array *q = band2list(priv, band);
731 unsigned int pkt_len = qdisc_pkt_len(skb);
734 err = skb_array_produce(q, skb);
737 if (qdisc_is_percpu_stats(qdisc))
738 return qdisc_drop_cpu(skb, qdisc, to_free);
740 return qdisc_drop(skb, qdisc, to_free);
743 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
744 return NET_XMIT_SUCCESS;
747 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
749 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
750 struct sk_buff *skb = NULL;
751 bool need_retry = true;
755 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
756 struct skb_array *q = band2list(priv, band);
758 if (__skb_array_empty(q))
761 skb = __skb_array_consume(q);
764 qdisc_update_stats_at_dequeue(qdisc, skb);
765 } else if (need_retry &&
766 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
767 /* Delay clearing the STATE_MISSED here to reduce
768 * the overhead of the second spin_trylock() in
769 * qdisc_run_begin() and __netif_schedule() calling
770 * in qdisc_run_end().
772 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
773 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
775 /* Make sure dequeuing happens after clearing
778 smp_mb__after_atomic();
788 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
790 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
791 struct sk_buff *skb = NULL;
794 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
795 struct skb_array *q = band2list(priv, band);
797 skb = __skb_array_peek(q);
803 static void pfifo_fast_reset(struct Qdisc *qdisc)
806 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
808 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
809 struct skb_array *q = band2list(priv, band);
812 /* NULL ring is possible if destroy path is due to a failed
813 * skb_array_init() in pfifo_fast_init() case.
818 while ((skb = __skb_array_consume(q)) != NULL)
822 if (qdisc_is_percpu_stats(qdisc)) {
823 for_each_possible_cpu(i) {
824 struct gnet_stats_queue *q;
826 q = per_cpu_ptr(qdisc->cpu_qstats, i);
833 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
835 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
837 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
838 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
839 goto nla_put_failure;
846 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
847 struct netlink_ext_ack *extack)
849 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
850 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
853 /* guard against zero length rings */
857 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
858 struct skb_array *q = band2list(priv, prio);
861 err = skb_array_init(q, qlen, GFP_KERNEL);
866 /* Can by-pass the queue discipline */
867 qdisc->flags |= TCQ_F_CAN_BYPASS;
871 static void pfifo_fast_destroy(struct Qdisc *sch)
873 struct pfifo_fast_priv *priv = qdisc_priv(sch);
876 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
877 struct skb_array *q = band2list(priv, prio);
879 /* NULL ring is possible if destroy path is due to a failed
880 * skb_array_init() in pfifo_fast_init() case.
884 /* Destroy ring but no need to kfree_skb because a call to
885 * pfifo_fast_reset() has already done that work.
887 ptr_ring_cleanup(&q->ring, NULL);
891 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
892 unsigned int new_len)
894 struct pfifo_fast_priv *priv = qdisc_priv(sch);
895 struct skb_array *bands[PFIFO_FAST_BANDS];
898 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
899 struct skb_array *q = band2list(priv, prio);
904 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
908 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
910 .priv_size = sizeof(struct pfifo_fast_priv),
911 .enqueue = pfifo_fast_enqueue,
912 .dequeue = pfifo_fast_dequeue,
913 .peek = pfifo_fast_peek,
914 .init = pfifo_fast_init,
915 .destroy = pfifo_fast_destroy,
916 .reset = pfifo_fast_reset,
917 .dump = pfifo_fast_dump,
918 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
919 .owner = THIS_MODULE,
920 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
922 EXPORT_SYMBOL(pfifo_fast_ops);
924 static struct lock_class_key qdisc_tx_busylock;
926 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
927 const struct Qdisc_ops *ops,
928 struct netlink_ext_ack *extack)
931 unsigned int size = sizeof(*sch) + ops->priv_size;
933 struct net_device *dev;
936 NL_SET_ERR_MSG(extack, "No device queue given");
941 dev = dev_queue->dev;
942 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
946 __skb_queue_head_init(&sch->gso_skb);
947 __skb_queue_head_init(&sch->skb_bad_txq);
948 qdisc_skb_head_init(&sch->q);
949 gnet_stats_basic_sync_init(&sch->bstats);
950 spin_lock_init(&sch->q.lock);
952 if (ops->static_flags & TCQ_F_CPUSTATS) {
954 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
955 if (!sch->cpu_bstats)
958 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
959 if (!sch->cpu_qstats) {
960 free_percpu(sch->cpu_bstats);
965 spin_lock_init(&sch->busylock);
966 lockdep_set_class(&sch->busylock,
967 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
969 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
970 spin_lock_init(&sch->seqlock);
971 lockdep_set_class(&sch->seqlock,
972 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
975 sch->flags = ops->static_flags;
976 sch->enqueue = ops->enqueue;
977 sch->dequeue = ops->dequeue;
978 sch->dev_queue = dev_queue;
979 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
980 refcount_set(&sch->refcnt, 1);
989 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
990 const struct Qdisc_ops *ops,
991 unsigned int parentid,
992 struct netlink_ext_ack *extack)
996 if (!try_module_get(ops->owner)) {
997 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1001 sch = qdisc_alloc(dev_queue, ops, extack);
1003 module_put(ops->owner);
1006 sch->parent = parentid;
1008 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1009 trace_qdisc_create(ops, dev_queue->dev, parentid);
1016 EXPORT_SYMBOL(qdisc_create_dflt);
1018 /* Under qdisc_lock(qdisc) and BH! */
1020 void qdisc_reset(struct Qdisc *qdisc)
1022 const struct Qdisc_ops *ops = qdisc->ops;
1024 trace_qdisc_reset(qdisc);
1029 __skb_queue_purge(&qdisc->gso_skb);
1030 __skb_queue_purge(&qdisc->skb_bad_txq);
1033 qdisc->qstats.backlog = 0;
1035 EXPORT_SYMBOL(qdisc_reset);
1037 void qdisc_free(struct Qdisc *qdisc)
1039 if (qdisc_is_percpu_stats(qdisc)) {
1040 free_percpu(qdisc->cpu_bstats);
1041 free_percpu(qdisc->cpu_qstats);
1047 static void qdisc_free_cb(struct rcu_head *head)
1049 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1054 static void qdisc_destroy(struct Qdisc *qdisc)
1056 const struct Qdisc_ops *ops = qdisc->ops;
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);
1068 ops->destroy(qdisc);
1070 module_put(ops->owner);
1071 netdev_put(qdisc_dev(qdisc), &qdisc->dev_tracker);
1073 trace_qdisc_destroy(qdisc);
1075 call_rcu(&qdisc->rcu, qdisc_free_cb);
1078 void qdisc_put(struct Qdisc *qdisc)
1083 if (qdisc->flags & TCQ_F_BUILTIN ||
1084 !refcount_dec_and_test(&qdisc->refcnt))
1087 qdisc_destroy(qdisc);
1089 EXPORT_SYMBOL(qdisc_put);
1091 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1092 * Intended to be used as optimization, this function only takes rtnl lock if
1093 * qdisc reference counter reached zero.
1096 void qdisc_put_unlocked(struct Qdisc *qdisc)
1098 if (qdisc->flags & TCQ_F_BUILTIN ||
1099 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1102 qdisc_destroy(qdisc);
1105 EXPORT_SYMBOL(qdisc_put_unlocked);
1107 /* Attach toplevel qdisc to device queue. */
1108 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1109 struct Qdisc *qdisc)
1111 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1112 spinlock_t *root_lock;
1114 root_lock = qdisc_lock(oqdisc);
1115 spin_lock_bh(root_lock);
1117 /* ... and graft new one */
1119 qdisc = &noop_qdisc;
1120 dev_queue->qdisc_sleeping = qdisc;
1121 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1123 spin_unlock_bh(root_lock);
1127 EXPORT_SYMBOL(dev_graft_qdisc);
1129 static void attach_one_default_qdisc(struct net_device *dev,
1130 struct netdev_queue *dev_queue,
1133 struct Qdisc *qdisc;
1134 const struct Qdisc_ops *ops = default_qdisc_ops;
1136 if (dev->priv_flags & IFF_NO_QUEUE)
1137 ops = &noqueue_qdisc_ops;
1138 else if(dev->type == ARPHRD_CAN)
1139 ops = &pfifo_fast_ops;
1141 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1145 if (!netif_is_multiqueue(dev))
1146 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1147 dev_queue->qdisc_sleeping = qdisc;
1150 static void attach_default_qdiscs(struct net_device *dev)
1152 struct netdev_queue *txq;
1153 struct Qdisc *qdisc;
1155 txq = netdev_get_tx_queue(dev, 0);
1157 if (!netif_is_multiqueue(dev) ||
1158 dev->priv_flags & IFF_NO_QUEUE) {
1159 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1160 qdisc = txq->qdisc_sleeping;
1161 rcu_assign_pointer(dev->qdisc, qdisc);
1162 qdisc_refcount_inc(qdisc);
1164 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1166 rcu_assign_pointer(dev->qdisc, qdisc);
1167 qdisc->ops->attach(qdisc);
1170 qdisc = rtnl_dereference(dev->qdisc);
1172 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1173 if (qdisc == &noop_qdisc) {
1174 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1175 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1176 dev->priv_flags |= IFF_NO_QUEUE;
1177 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1178 qdisc = txq->qdisc_sleeping;
1179 rcu_assign_pointer(dev->qdisc, qdisc);
1180 qdisc_refcount_inc(qdisc);
1181 dev->priv_flags ^= IFF_NO_QUEUE;
1184 #ifdef CONFIG_NET_SCHED
1185 if (qdisc != &noop_qdisc)
1186 qdisc_hash_add(qdisc, false);
1190 static void transition_one_qdisc(struct net_device *dev,
1191 struct netdev_queue *dev_queue,
1192 void *_need_watchdog)
1194 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1195 int *need_watchdog_p = _need_watchdog;
1197 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1198 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1200 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1201 if (need_watchdog_p) {
1202 WRITE_ONCE(dev_queue->trans_start, 0);
1203 *need_watchdog_p = 1;
1207 void dev_activate(struct net_device *dev)
1211 /* No queueing discipline is attached to device;
1212 * create default one for devices, which need queueing
1213 * and noqueue_qdisc for virtual interfaces
1216 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1217 attach_default_qdiscs(dev);
1219 if (!netif_carrier_ok(dev))
1220 /* Delay activation until next carrier-on event */
1224 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1225 if (dev_ingress_queue(dev))
1226 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1228 if (need_watchdog) {
1229 netif_trans_update(dev);
1230 dev_watchdog_up(dev);
1233 EXPORT_SYMBOL(dev_activate);
1235 static void qdisc_deactivate(struct Qdisc *qdisc)
1237 if (qdisc->flags & TCQ_F_BUILTIN)
1240 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1243 static void dev_deactivate_queue(struct net_device *dev,
1244 struct netdev_queue *dev_queue,
1245 void *_qdisc_default)
1247 struct Qdisc *qdisc_default = _qdisc_default;
1248 struct Qdisc *qdisc;
1250 qdisc = rtnl_dereference(dev_queue->qdisc);
1252 qdisc_deactivate(qdisc);
1253 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1257 static void dev_reset_queue(struct net_device *dev,
1258 struct netdev_queue *dev_queue,
1261 struct Qdisc *qdisc;
1264 qdisc = dev_queue->qdisc_sleeping;
1268 nolock = qdisc->flags & TCQ_F_NOLOCK;
1271 spin_lock_bh(&qdisc->seqlock);
1272 spin_lock_bh(qdisc_lock(qdisc));
1276 spin_unlock_bh(qdisc_lock(qdisc));
1278 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1279 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1280 spin_unlock_bh(&qdisc->seqlock);
1284 static bool some_qdisc_is_busy(struct net_device *dev)
1288 for (i = 0; i < dev->num_tx_queues; i++) {
1289 struct netdev_queue *dev_queue;
1290 spinlock_t *root_lock;
1294 dev_queue = netdev_get_tx_queue(dev, i);
1295 q = dev_queue->qdisc_sleeping;
1297 root_lock = qdisc_lock(q);
1298 spin_lock_bh(root_lock);
1300 val = (qdisc_is_running(q) ||
1301 test_bit(__QDISC_STATE_SCHED, &q->state));
1303 spin_unlock_bh(root_lock);
1312 * dev_deactivate_many - deactivate transmissions on several devices
1313 * @head: list of devices to deactivate
1315 * This function returns only when all outstanding transmissions
1316 * have completed, unless all devices are in dismantle phase.
1318 void dev_deactivate_many(struct list_head *head)
1320 struct net_device *dev;
1322 list_for_each_entry(dev, head, close_list) {
1323 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1325 if (dev_ingress_queue(dev))
1326 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1329 dev_watchdog_down(dev);
1332 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1333 * outstanding qdisc enqueuing calls.
1334 * This is avoided if all devices are in dismantle phase :
1335 * Caller will call synchronize_net() for us
1339 list_for_each_entry(dev, head, close_list) {
1340 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1342 if (dev_ingress_queue(dev))
1343 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1346 /* Wait for outstanding qdisc_run calls. */
1347 list_for_each_entry(dev, head, close_list) {
1348 while (some_qdisc_is_busy(dev)) {
1349 /* wait_event() would avoid this sleep-loop but would
1350 * require expensive checks in the fast paths of packet
1351 * processing which isn't worth it.
1353 schedule_timeout_uninterruptible(1);
1358 void dev_deactivate(struct net_device *dev)
1362 list_add(&dev->close_list, &single);
1363 dev_deactivate_many(&single);
1366 EXPORT_SYMBOL(dev_deactivate);
1368 static int qdisc_change_tx_queue_len(struct net_device *dev,
1369 struct netdev_queue *dev_queue)
1371 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1372 const struct Qdisc_ops *ops = qdisc->ops;
1374 if (ops->change_tx_queue_len)
1375 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1379 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1380 unsigned int new_real_tx)
1382 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1384 if (qdisc->ops->change_real_num_tx)
1385 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1388 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1390 #ifdef CONFIG_NET_SCHED
1391 struct net_device *dev = qdisc_dev(sch);
1392 struct Qdisc *qdisc;
1395 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1396 qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1397 /* Only update the default qdiscs we created,
1398 * qdiscs with handles are always hashed.
1400 if (qdisc != &noop_qdisc && !qdisc->handle)
1401 qdisc_hash_del(qdisc);
1403 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1404 qdisc = netdev_get_tx_queue(dev, i)->qdisc_sleeping;
1405 if (qdisc != &noop_qdisc && !qdisc->handle)
1406 qdisc_hash_add(qdisc, false);
1410 EXPORT_SYMBOL(mq_change_real_num_tx);
1412 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1414 bool up = dev->flags & IFF_UP;
1419 dev_deactivate(dev);
1421 for (i = 0; i < dev->num_tx_queues; i++) {
1422 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1424 /* TODO: revert changes on a partial failure */
1434 static void dev_init_scheduler_queue(struct net_device *dev,
1435 struct netdev_queue *dev_queue,
1438 struct Qdisc *qdisc = _qdisc;
1440 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1441 dev_queue->qdisc_sleeping = qdisc;
1444 void dev_init_scheduler(struct net_device *dev)
1446 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1447 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1448 if (dev_ingress_queue(dev))
1449 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1451 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1454 static void shutdown_scheduler_queue(struct net_device *dev,
1455 struct netdev_queue *dev_queue,
1456 void *_qdisc_default)
1458 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1459 struct Qdisc *qdisc_default = _qdisc_default;
1462 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1463 dev_queue->qdisc_sleeping = qdisc_default;
1469 void dev_shutdown(struct net_device *dev)
1471 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1472 if (dev_ingress_queue(dev))
1473 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1474 qdisc_put(rtnl_dereference(dev->qdisc));
1475 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1477 WARN_ON(timer_pending(&dev->watchdog_timer));
1481 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1482 * @rate: Rate to compute reciprocal division values of
1483 * @mult: Multiplier for reciprocal division
1484 * @shift: Shift for reciprocal division
1486 * The multiplier and shift for reciprocal division by rate are stored
1487 * in mult and shift.
1489 * The deal here is to replace a divide by a reciprocal one
1490 * in fast path (a reciprocal divide is a multiply and a shift)
1492 * Normal formula would be :
1493 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1495 * We compute mult/shift to use instead :
1496 * time_in_ns = (len * mult) >> shift;
1498 * We try to get the highest possible mult value for accuracy,
1499 * but have to make sure no overflows will ever happen.
1501 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1503 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1505 u64 factor = NSEC_PER_SEC;
1514 *mult = div64_u64(factor, rate);
1515 if (*mult & (1U << 31) || factor & (1ULL << 63))
1522 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1523 const struct tc_ratespec *conf,
1526 memset(r, 0, sizeof(*r));
1527 r->overhead = conf->overhead;
1529 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1530 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1531 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1533 EXPORT_SYMBOL(psched_ratecfg_precompute);
1535 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1537 r->rate_pkts_ps = pktrate64;
1538 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1540 EXPORT_SYMBOL(psched_ppscfg_precompute);
1542 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1543 struct tcf_proto *tp_head)
1545 /* Protected with chain0->filter_chain_lock.
1546 * Can't access chain directly because tp_head can be NULL.
1548 struct mini_Qdisc *miniq_old =
1549 rcu_dereference_protected(*miniqp->p_miniq, 1);
1550 struct mini_Qdisc *miniq;
1553 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1555 miniq = miniq_old != &miniqp->miniq1 ?
1556 &miniqp->miniq1 : &miniqp->miniq2;
1558 /* We need to make sure that readers won't see the miniq
1559 * we are about to modify. So ensure that at least one RCU
1560 * grace period has elapsed since the miniq was made
1563 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1564 cond_synchronize_rcu(miniq->rcu_state);
1565 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1566 synchronize_rcu_expedited();
1568 miniq->filter_list = tp_head;
1569 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1573 /* This is counterpart of the rcu sync above. We need to
1574 * block potential new user of miniq_old until all readers
1575 * are not seeing it.
1577 miniq_old->rcu_state = start_poll_synchronize_rcu();
1579 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1581 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1582 struct tcf_block *block)
1584 miniqp->miniq1.block = block;
1585 miniqp->miniq2.block = block;
1587 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1589 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1590 struct mini_Qdisc __rcu **p_miniq)
1592 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1593 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1594 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1595 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1596 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1597 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1598 miniqp->p_miniq = p_miniq;
1600 EXPORT_SYMBOL(mini_qdisc_pair_init);