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 = READ_ONCE(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 res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
434 for (i = 1; i < dev->num_tx_queues; i++) {
435 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
436 if (val && time_after(val, res))
442 EXPORT_SYMBOL(dev_trans_start);
444 static void netif_freeze_queues(struct net_device *dev)
449 cpu = smp_processor_id();
450 for (i = 0; i < dev->num_tx_queues; i++) {
451 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
453 /* We are the only thread of execution doing a
454 * freeze, but we have to grab the _xmit_lock in
455 * order to synchronize with threads which are in
456 * the ->hard_start_xmit() handler and already
457 * checked the frozen bit.
459 __netif_tx_lock(txq, cpu);
460 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
461 __netif_tx_unlock(txq);
465 void netif_tx_lock(struct net_device *dev)
467 spin_lock(&dev->tx_global_lock);
468 netif_freeze_queues(dev);
470 EXPORT_SYMBOL(netif_tx_lock);
472 static void netif_unfreeze_queues(struct net_device *dev)
476 for (i = 0; i < dev->num_tx_queues; i++) {
477 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
479 /* No need to grab the _xmit_lock here. If the
480 * queue is not stopped for another reason, we
483 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
484 netif_schedule_queue(txq);
488 void netif_tx_unlock(struct net_device *dev)
490 netif_unfreeze_queues(dev);
491 spin_unlock(&dev->tx_global_lock);
493 EXPORT_SYMBOL(netif_tx_unlock);
495 static void dev_watchdog(struct timer_list *t)
497 struct net_device *dev = from_timer(dev, t, watchdog_timer);
500 spin_lock(&dev->tx_global_lock);
501 if (!qdisc_tx_is_noop(dev)) {
502 if (netif_device_present(dev) &&
503 netif_running(dev) &&
504 netif_carrier_ok(dev)) {
505 unsigned int timedout_ms = 0;
507 unsigned long trans_start;
509 for (i = 0; i < dev->num_tx_queues; i++) {
510 struct netdev_queue *txq;
512 txq = netdev_get_tx_queue(dev, i);
513 trans_start = READ_ONCE(txq->trans_start);
514 if (netif_xmit_stopped(txq) &&
515 time_after(jiffies, (trans_start +
516 dev->watchdog_timeo))) {
517 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
518 atomic_long_inc(&txq->trans_timeout);
523 if (unlikely(timedout_ms)) {
524 trace_net_dev_xmit_timeout(dev, i);
525 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
526 raw_smp_processor_id(),
528 netif_freeze_queues(dev);
529 dev->netdev_ops->ndo_tx_timeout(dev, i);
530 netif_unfreeze_queues(dev);
532 if (!mod_timer(&dev->watchdog_timer,
533 round_jiffies(jiffies +
534 dev->watchdog_timeo)))
538 spin_unlock(&dev->tx_global_lock);
541 netdev_put(dev, &dev->watchdog_dev_tracker);
544 void __netdev_watchdog_up(struct net_device *dev)
546 if (dev->netdev_ops->ndo_tx_timeout) {
547 if (dev->watchdog_timeo <= 0)
548 dev->watchdog_timeo = 5*HZ;
549 if (!mod_timer(&dev->watchdog_timer,
550 round_jiffies(jiffies + dev->watchdog_timeo)))
551 netdev_hold(dev, &dev->watchdog_dev_tracker,
555 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
557 static void dev_watchdog_up(struct net_device *dev)
559 __netdev_watchdog_up(dev);
562 static void dev_watchdog_down(struct net_device *dev)
564 netif_tx_lock_bh(dev);
565 if (del_timer(&dev->watchdog_timer))
566 netdev_put(dev, &dev->watchdog_dev_tracker);
567 netif_tx_unlock_bh(dev);
571 * netif_carrier_on - set carrier
572 * @dev: network device
574 * Device has detected acquisition of carrier.
576 void netif_carrier_on(struct net_device *dev)
578 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
579 if (dev->reg_state == NETREG_UNINITIALIZED)
581 atomic_inc(&dev->carrier_up_count);
582 linkwatch_fire_event(dev);
583 if (netif_running(dev))
584 __netdev_watchdog_up(dev);
587 EXPORT_SYMBOL(netif_carrier_on);
590 * netif_carrier_off - clear carrier
591 * @dev: network device
593 * Device has detected loss of carrier.
595 void netif_carrier_off(struct net_device *dev)
597 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
598 if (dev->reg_state == NETREG_UNINITIALIZED)
600 atomic_inc(&dev->carrier_down_count);
601 linkwatch_fire_event(dev);
604 EXPORT_SYMBOL(netif_carrier_off);
607 * netif_carrier_event - report carrier state event
608 * @dev: network device
610 * Device has detected a carrier event but the carrier state wasn't changed.
611 * Use in drivers when querying carrier state asynchronously, to avoid missing
612 * events (link flaps) if link recovers before it's queried.
614 void netif_carrier_event(struct net_device *dev)
616 if (dev->reg_state == NETREG_UNINITIALIZED)
618 atomic_inc(&dev->carrier_up_count);
619 atomic_inc(&dev->carrier_down_count);
620 linkwatch_fire_event(dev);
622 EXPORT_SYMBOL_GPL(netif_carrier_event);
624 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
625 under all circumstances. It is difficult to invent anything faster or
629 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
630 struct sk_buff **to_free)
632 __qdisc_drop(skb, to_free);
636 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
641 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
644 .enqueue = noop_enqueue,
645 .dequeue = noop_dequeue,
646 .peek = noop_dequeue,
647 .owner = THIS_MODULE,
650 static struct netdev_queue noop_netdev_queue = {
651 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
652 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
655 struct Qdisc noop_qdisc = {
656 .enqueue = noop_enqueue,
657 .dequeue = noop_dequeue,
658 .flags = TCQ_F_BUILTIN,
659 .ops = &noop_qdisc_ops,
660 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
661 .dev_queue = &noop_netdev_queue,
662 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
664 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
665 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
667 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
670 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
671 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
673 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
676 EXPORT_SYMBOL(noop_qdisc);
678 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
679 struct netlink_ext_ack *extack)
681 /* register_qdisc() assigns a default of noop_enqueue if unset,
682 * but __dev_queue_xmit() treats noqueue only as such
683 * if this is NULL - so clear it here. */
684 qdisc->enqueue = NULL;
688 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
691 .init = noqueue_init,
692 .enqueue = noop_enqueue,
693 .dequeue = noop_dequeue,
694 .peek = noop_dequeue,
695 .owner = THIS_MODULE,
698 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
699 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
701 EXPORT_SYMBOL(sch_default_prio2band);
703 /* 3-band FIFO queue: old style, but should be a bit faster than
704 generic prio+fifo combination.
707 #define PFIFO_FAST_BANDS 3
710 * Private data for a pfifo_fast scheduler containing:
711 * - rings for priority bands
713 struct pfifo_fast_priv {
714 struct skb_array q[PFIFO_FAST_BANDS];
717 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
720 return &priv->q[band];
723 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
724 struct sk_buff **to_free)
726 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
727 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
728 struct skb_array *q = band2list(priv, band);
729 unsigned int pkt_len = qdisc_pkt_len(skb);
732 err = skb_array_produce(q, skb);
735 if (qdisc_is_percpu_stats(qdisc))
736 return qdisc_drop_cpu(skb, qdisc, to_free);
738 return qdisc_drop(skb, qdisc, to_free);
741 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
742 return NET_XMIT_SUCCESS;
745 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
747 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
748 struct sk_buff *skb = NULL;
749 bool need_retry = true;
753 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
754 struct skb_array *q = band2list(priv, band);
756 if (__skb_array_empty(q))
759 skb = __skb_array_consume(q);
762 qdisc_update_stats_at_dequeue(qdisc, skb);
763 } else if (need_retry &&
764 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
765 /* Delay clearing the STATE_MISSED here to reduce
766 * the overhead of the second spin_trylock() in
767 * qdisc_run_begin() and __netif_schedule() calling
768 * in qdisc_run_end().
770 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
771 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
773 /* Make sure dequeuing happens after clearing
776 smp_mb__after_atomic();
786 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
788 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
789 struct sk_buff *skb = NULL;
792 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
793 struct skb_array *q = band2list(priv, band);
795 skb = __skb_array_peek(q);
801 static void pfifo_fast_reset(struct Qdisc *qdisc)
804 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
806 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
807 struct skb_array *q = band2list(priv, band);
810 /* NULL ring is possible if destroy path is due to a failed
811 * skb_array_init() in pfifo_fast_init() case.
816 while ((skb = __skb_array_consume(q)) != NULL)
820 if (qdisc_is_percpu_stats(qdisc)) {
821 for_each_possible_cpu(i) {
822 struct gnet_stats_queue *q;
824 q = per_cpu_ptr(qdisc->cpu_qstats, i);
831 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
833 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
835 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
836 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
837 goto nla_put_failure;
844 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
845 struct netlink_ext_ack *extack)
847 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
848 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
851 /* guard against zero length rings */
855 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
856 struct skb_array *q = band2list(priv, prio);
859 err = skb_array_init(q, qlen, GFP_KERNEL);
864 /* Can by-pass the queue discipline */
865 qdisc->flags |= TCQ_F_CAN_BYPASS;
869 static void pfifo_fast_destroy(struct Qdisc *sch)
871 struct pfifo_fast_priv *priv = qdisc_priv(sch);
874 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
875 struct skb_array *q = band2list(priv, prio);
877 /* NULL ring is possible if destroy path is due to a failed
878 * skb_array_init() in pfifo_fast_init() case.
882 /* Destroy ring but no need to kfree_skb because a call to
883 * pfifo_fast_reset() has already done that work.
885 ptr_ring_cleanup(&q->ring, NULL);
889 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
890 unsigned int new_len)
892 struct pfifo_fast_priv *priv = qdisc_priv(sch);
893 struct skb_array *bands[PFIFO_FAST_BANDS];
896 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
897 struct skb_array *q = band2list(priv, prio);
902 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
906 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
908 .priv_size = sizeof(struct pfifo_fast_priv),
909 .enqueue = pfifo_fast_enqueue,
910 .dequeue = pfifo_fast_dequeue,
911 .peek = pfifo_fast_peek,
912 .init = pfifo_fast_init,
913 .destroy = pfifo_fast_destroy,
914 .reset = pfifo_fast_reset,
915 .dump = pfifo_fast_dump,
916 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
917 .owner = THIS_MODULE,
918 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
920 EXPORT_SYMBOL(pfifo_fast_ops);
922 static struct lock_class_key qdisc_tx_busylock;
924 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
925 const struct Qdisc_ops *ops,
926 struct netlink_ext_ack *extack)
929 unsigned int size = sizeof(*sch) + ops->priv_size;
931 struct net_device *dev;
934 NL_SET_ERR_MSG(extack, "No device queue given");
939 dev = dev_queue->dev;
940 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
944 __skb_queue_head_init(&sch->gso_skb);
945 __skb_queue_head_init(&sch->skb_bad_txq);
946 gnet_stats_basic_sync_init(&sch->bstats);
947 spin_lock_init(&sch->q.lock);
949 if (ops->static_flags & TCQ_F_CPUSTATS) {
951 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
952 if (!sch->cpu_bstats)
955 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
956 if (!sch->cpu_qstats) {
957 free_percpu(sch->cpu_bstats);
962 spin_lock_init(&sch->busylock);
963 lockdep_set_class(&sch->busylock,
964 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
966 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
967 spin_lock_init(&sch->seqlock);
968 lockdep_set_class(&sch->seqlock,
969 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
972 sch->flags = ops->static_flags;
973 sch->enqueue = ops->enqueue;
974 sch->dequeue = ops->dequeue;
975 sch->dev_queue = dev_queue;
977 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
978 refcount_set(&sch->refcnt, 1);
987 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
988 const struct Qdisc_ops *ops,
989 unsigned int parentid,
990 struct netlink_ext_ack *extack)
994 if (!try_module_get(ops->owner)) {
995 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
999 sch = qdisc_alloc(dev_queue, ops, extack);
1001 module_put(ops->owner);
1004 sch->parent = parentid;
1006 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1007 trace_qdisc_create(ops, dev_queue->dev, parentid);
1014 EXPORT_SYMBOL(qdisc_create_dflt);
1016 /* Under qdisc_lock(qdisc) and BH! */
1018 void qdisc_reset(struct Qdisc *qdisc)
1020 const struct Qdisc_ops *ops = qdisc->ops;
1022 trace_qdisc_reset(qdisc);
1027 __skb_queue_purge(&qdisc->gso_skb);
1028 __skb_queue_purge(&qdisc->skb_bad_txq);
1031 qdisc->qstats.backlog = 0;
1033 EXPORT_SYMBOL(qdisc_reset);
1035 void qdisc_free(struct Qdisc *qdisc)
1037 if (qdisc_is_percpu_stats(qdisc)) {
1038 free_percpu(qdisc->cpu_bstats);
1039 free_percpu(qdisc->cpu_qstats);
1045 static void qdisc_free_cb(struct rcu_head *head)
1047 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1052 static void __qdisc_destroy(struct Qdisc *qdisc)
1054 const struct Qdisc_ops *ops = qdisc->ops;
1055 struct net_device *dev = qdisc_dev(qdisc);
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);
1068 ops->destroy(qdisc);
1070 module_put(ops->owner);
1071 netdev_put(dev, &qdisc->dev_tracker);
1073 trace_qdisc_destroy(qdisc);
1075 call_rcu(&qdisc->rcu, qdisc_free_cb);
1078 void qdisc_destroy(struct Qdisc *qdisc)
1080 if (qdisc->flags & TCQ_F_BUILTIN)
1083 __qdisc_destroy(qdisc);
1086 void qdisc_put(struct Qdisc *qdisc)
1091 if (qdisc->flags & TCQ_F_BUILTIN ||
1092 !refcount_dec_and_test(&qdisc->refcnt))
1095 __qdisc_destroy(qdisc);
1097 EXPORT_SYMBOL(qdisc_put);
1099 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1100 * Intended to be used as optimization, this function only takes rtnl lock if
1101 * qdisc reference counter reached zero.
1104 void qdisc_put_unlocked(struct Qdisc *qdisc)
1106 if (qdisc->flags & TCQ_F_BUILTIN ||
1107 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1110 __qdisc_destroy(qdisc);
1113 EXPORT_SYMBOL(qdisc_put_unlocked);
1115 /* Attach toplevel qdisc to device queue. */
1116 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1117 struct Qdisc *qdisc)
1119 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1120 spinlock_t *root_lock;
1122 root_lock = qdisc_lock(oqdisc);
1123 spin_lock_bh(root_lock);
1125 /* ... and graft new one */
1127 qdisc = &noop_qdisc;
1128 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1129 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1131 spin_unlock_bh(root_lock);
1135 EXPORT_SYMBOL(dev_graft_qdisc);
1137 static void shutdown_scheduler_queue(struct net_device *dev,
1138 struct netdev_queue *dev_queue,
1139 void *_qdisc_default)
1141 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1142 struct Qdisc *qdisc_default = _qdisc_default;
1145 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1146 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1152 static void attach_one_default_qdisc(struct net_device *dev,
1153 struct netdev_queue *dev_queue,
1156 struct Qdisc *qdisc;
1157 const struct Qdisc_ops *ops = default_qdisc_ops;
1159 if (dev->priv_flags & IFF_NO_QUEUE)
1160 ops = &noqueue_qdisc_ops;
1161 else if(dev->type == ARPHRD_CAN)
1162 ops = &pfifo_fast_ops;
1164 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1168 if (!netif_is_multiqueue(dev))
1169 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1170 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1173 static void attach_default_qdiscs(struct net_device *dev)
1175 struct netdev_queue *txq;
1176 struct Qdisc *qdisc;
1178 txq = netdev_get_tx_queue(dev, 0);
1180 if (!netif_is_multiqueue(dev) ||
1181 dev->priv_flags & IFF_NO_QUEUE) {
1182 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1183 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1184 rcu_assign_pointer(dev->qdisc, qdisc);
1185 qdisc_refcount_inc(qdisc);
1187 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1189 rcu_assign_pointer(dev->qdisc, qdisc);
1190 qdisc->ops->attach(qdisc);
1193 qdisc = rtnl_dereference(dev->qdisc);
1195 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1196 if (qdisc == &noop_qdisc) {
1197 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1198 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1199 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1200 dev->priv_flags |= IFF_NO_QUEUE;
1201 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1202 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1203 rcu_assign_pointer(dev->qdisc, qdisc);
1204 qdisc_refcount_inc(qdisc);
1205 dev->priv_flags ^= IFF_NO_QUEUE;
1208 #ifdef CONFIG_NET_SCHED
1209 if (qdisc != &noop_qdisc)
1210 qdisc_hash_add(qdisc, false);
1214 static void transition_one_qdisc(struct net_device *dev,
1215 struct netdev_queue *dev_queue,
1216 void *_need_watchdog)
1218 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1219 int *need_watchdog_p = _need_watchdog;
1221 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1222 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1224 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1225 if (need_watchdog_p) {
1226 WRITE_ONCE(dev_queue->trans_start, 0);
1227 *need_watchdog_p = 1;
1231 void dev_activate(struct net_device *dev)
1235 /* No queueing discipline is attached to device;
1236 * create default one for devices, which need queueing
1237 * and noqueue_qdisc for virtual interfaces
1240 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1241 attach_default_qdiscs(dev);
1243 if (!netif_carrier_ok(dev))
1244 /* Delay activation until next carrier-on event */
1248 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1249 if (dev_ingress_queue(dev))
1250 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1252 if (need_watchdog) {
1253 netif_trans_update(dev);
1254 dev_watchdog_up(dev);
1257 EXPORT_SYMBOL(dev_activate);
1259 static void qdisc_deactivate(struct Qdisc *qdisc)
1261 if (qdisc->flags & TCQ_F_BUILTIN)
1264 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1267 static void dev_deactivate_queue(struct net_device *dev,
1268 struct netdev_queue *dev_queue,
1269 void *_qdisc_default)
1271 struct Qdisc *qdisc_default = _qdisc_default;
1272 struct Qdisc *qdisc;
1274 qdisc = rtnl_dereference(dev_queue->qdisc);
1276 qdisc_deactivate(qdisc);
1277 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1281 static void dev_reset_queue(struct net_device *dev,
1282 struct netdev_queue *dev_queue,
1285 struct Qdisc *qdisc;
1288 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1292 nolock = qdisc->flags & TCQ_F_NOLOCK;
1295 spin_lock_bh(&qdisc->seqlock);
1296 spin_lock_bh(qdisc_lock(qdisc));
1300 spin_unlock_bh(qdisc_lock(qdisc));
1302 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1303 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1304 spin_unlock_bh(&qdisc->seqlock);
1308 static bool some_qdisc_is_busy(struct net_device *dev)
1312 for (i = 0; i < dev->num_tx_queues; i++) {
1313 struct netdev_queue *dev_queue;
1314 spinlock_t *root_lock;
1318 dev_queue = netdev_get_tx_queue(dev, i);
1319 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1321 root_lock = qdisc_lock(q);
1322 spin_lock_bh(root_lock);
1324 val = (qdisc_is_running(q) ||
1325 test_bit(__QDISC_STATE_SCHED, &q->state));
1327 spin_unlock_bh(root_lock);
1336 * dev_deactivate_many - deactivate transmissions on several devices
1337 * @head: list of devices to deactivate
1339 * This function returns only when all outstanding transmissions
1340 * have completed, unless all devices are in dismantle phase.
1342 void dev_deactivate_many(struct list_head *head)
1344 struct net_device *dev;
1346 list_for_each_entry(dev, head, close_list) {
1347 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1349 if (dev_ingress_queue(dev))
1350 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1353 dev_watchdog_down(dev);
1356 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1357 * outstanding qdisc enqueuing calls.
1358 * This is avoided if all devices are in dismantle phase :
1359 * Caller will call synchronize_net() for us
1363 list_for_each_entry(dev, head, close_list) {
1364 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1366 if (dev_ingress_queue(dev))
1367 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1370 /* Wait for outstanding qdisc_run calls. */
1371 list_for_each_entry(dev, head, close_list) {
1372 while (some_qdisc_is_busy(dev)) {
1373 /* wait_event() would avoid this sleep-loop but would
1374 * require expensive checks in the fast paths of packet
1375 * processing which isn't worth it.
1377 schedule_timeout_uninterruptible(1);
1382 void dev_deactivate(struct net_device *dev)
1386 list_add(&dev->close_list, &single);
1387 dev_deactivate_many(&single);
1390 EXPORT_SYMBOL(dev_deactivate);
1392 static int qdisc_change_tx_queue_len(struct net_device *dev,
1393 struct netdev_queue *dev_queue)
1395 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1396 const struct Qdisc_ops *ops = qdisc->ops;
1398 if (ops->change_tx_queue_len)
1399 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1403 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1404 unsigned int new_real_tx)
1406 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1408 if (qdisc->ops->change_real_num_tx)
1409 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1412 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1414 #ifdef CONFIG_NET_SCHED
1415 struct net_device *dev = qdisc_dev(sch);
1416 struct Qdisc *qdisc;
1419 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1420 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1421 /* Only update the default qdiscs we created,
1422 * qdiscs with handles are always hashed.
1424 if (qdisc != &noop_qdisc && !qdisc->handle)
1425 qdisc_hash_del(qdisc);
1427 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1428 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1429 if (qdisc != &noop_qdisc && !qdisc->handle)
1430 qdisc_hash_add(qdisc, false);
1434 EXPORT_SYMBOL(mq_change_real_num_tx);
1436 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1438 bool up = dev->flags & IFF_UP;
1443 dev_deactivate(dev);
1445 for (i = 0; i < dev->num_tx_queues; i++) {
1446 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1448 /* TODO: revert changes on a partial failure */
1458 static void dev_init_scheduler_queue(struct net_device *dev,
1459 struct netdev_queue *dev_queue,
1462 struct Qdisc *qdisc = _qdisc;
1464 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1465 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1468 void dev_init_scheduler(struct net_device *dev)
1470 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1471 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1472 if (dev_ingress_queue(dev))
1473 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1475 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1478 void dev_shutdown(struct net_device *dev)
1480 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1481 if (dev_ingress_queue(dev))
1482 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1483 qdisc_put(rtnl_dereference(dev->qdisc));
1484 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1486 WARN_ON(timer_pending(&dev->watchdog_timer));
1490 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1491 * @rate: Rate to compute reciprocal division values of
1492 * @mult: Multiplier for reciprocal division
1493 * @shift: Shift for reciprocal division
1495 * The multiplier and shift for reciprocal division by rate are stored
1496 * in mult and shift.
1498 * The deal here is to replace a divide by a reciprocal one
1499 * in fast path (a reciprocal divide is a multiply and a shift)
1501 * Normal formula would be :
1502 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1504 * We compute mult/shift to use instead :
1505 * time_in_ns = (len * mult) >> shift;
1507 * We try to get the highest possible mult value for accuracy,
1508 * but have to make sure no overflows will ever happen.
1510 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1512 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1514 u64 factor = NSEC_PER_SEC;
1523 *mult = div64_u64(factor, rate);
1524 if (*mult & (1U << 31) || factor & (1ULL << 63))
1531 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1532 const struct tc_ratespec *conf,
1535 memset(r, 0, sizeof(*r));
1536 r->overhead = conf->overhead;
1538 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1539 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1540 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1542 EXPORT_SYMBOL(psched_ratecfg_precompute);
1544 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1546 r->rate_pkts_ps = pktrate64;
1547 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1549 EXPORT_SYMBOL(psched_ppscfg_precompute);
1551 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1552 struct tcf_proto *tp_head)
1554 /* Protected with chain0->filter_chain_lock.
1555 * Can't access chain directly because tp_head can be NULL.
1557 struct mini_Qdisc *miniq_old =
1558 rcu_dereference_protected(*miniqp->p_miniq, 1);
1559 struct mini_Qdisc *miniq;
1562 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1564 miniq = miniq_old != &miniqp->miniq1 ?
1565 &miniqp->miniq1 : &miniqp->miniq2;
1567 /* We need to make sure that readers won't see the miniq
1568 * we are about to modify. So ensure that at least one RCU
1569 * grace period has elapsed since the miniq was made
1572 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1573 cond_synchronize_rcu(miniq->rcu_state);
1574 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1575 synchronize_rcu_expedited();
1577 miniq->filter_list = tp_head;
1578 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1582 /* This is counterpart of the rcu sync above. We need to
1583 * block potential new user of miniq_old until all readers
1584 * are not seeing it.
1586 miniq_old->rcu_state = start_poll_synchronize_rcu();
1588 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1590 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1591 struct tcf_block *block)
1593 miniqp->miniq1.block = block;
1594 miniqp->miniq2.block = block;
1596 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1598 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1599 struct mini_Qdisc __rcu **p_miniq)
1601 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1602 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1603 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1604 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1605 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1606 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1607 miniqp->p_miniq = p_miniq;
1609 EXPORT_SYMBOL(mini_qdisc_pair_init);