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 running seqcount bit guarantees that
308 * only one CPU 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 = netdev_get_tx_queue(dev, 0)->trans_start;
438 for (i = 1; i < dev->num_tx_queues; i++) {
439 val = netdev_get_tx_queue(dev, i)->trans_start;
440 if (val && time_after(val, res))
446 EXPORT_SYMBOL(dev_trans_start);
448 static void dev_watchdog(struct timer_list *t)
450 struct net_device *dev = from_timer(dev, t, watchdog_timer);
453 if (!qdisc_tx_is_noop(dev)) {
454 if (netif_device_present(dev) &&
455 netif_running(dev) &&
456 netif_carrier_ok(dev)) {
457 int some_queue_timedout = 0;
459 unsigned long trans_start;
461 for (i = 0; i < dev->num_tx_queues; i++) {
462 struct netdev_queue *txq;
464 txq = netdev_get_tx_queue(dev, i);
465 trans_start = txq->trans_start;
466 if (netif_xmit_stopped(txq) &&
467 time_after(jiffies, (trans_start +
468 dev->watchdog_timeo))) {
469 some_queue_timedout = 1;
470 txq->trans_timeout++;
475 if (some_queue_timedout) {
476 trace_net_dev_xmit_timeout(dev, i);
477 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
478 dev->name, netdev_drivername(dev), i);
479 dev->netdev_ops->ndo_tx_timeout(dev, i);
481 if (!mod_timer(&dev->watchdog_timer,
482 round_jiffies(jiffies +
483 dev->watchdog_timeo)))
487 netif_tx_unlock(dev);
492 void __netdev_watchdog_up(struct net_device *dev)
494 if (dev->netdev_ops->ndo_tx_timeout) {
495 if (dev->watchdog_timeo <= 0)
496 dev->watchdog_timeo = 5*HZ;
497 if (!mod_timer(&dev->watchdog_timer,
498 round_jiffies(jiffies + dev->watchdog_timeo)))
502 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
504 static void dev_watchdog_up(struct net_device *dev)
506 __netdev_watchdog_up(dev);
509 static void dev_watchdog_down(struct net_device *dev)
511 netif_tx_lock_bh(dev);
512 if (del_timer(&dev->watchdog_timer))
514 netif_tx_unlock_bh(dev);
518 * netif_carrier_on - set carrier
519 * @dev: network device
521 * Device has detected acquisition of carrier.
523 void netif_carrier_on(struct net_device *dev)
525 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
526 if (dev->reg_state == NETREG_UNINITIALIZED)
528 atomic_inc(&dev->carrier_up_count);
529 linkwatch_fire_event(dev);
530 if (netif_running(dev))
531 __netdev_watchdog_up(dev);
534 EXPORT_SYMBOL(netif_carrier_on);
537 * netif_carrier_off - clear carrier
538 * @dev: network device
540 * Device has detected loss of carrier.
542 void netif_carrier_off(struct net_device *dev)
544 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
545 if (dev->reg_state == NETREG_UNINITIALIZED)
547 atomic_inc(&dev->carrier_down_count);
548 linkwatch_fire_event(dev);
551 EXPORT_SYMBOL(netif_carrier_off);
554 * netif_carrier_event - report carrier state event
555 * @dev: network device
557 * Device has detected a carrier event but the carrier state wasn't changed.
558 * Use in drivers when querying carrier state asynchronously, to avoid missing
559 * events (link flaps) if link recovers before it's queried.
561 void netif_carrier_event(struct net_device *dev)
563 if (dev->reg_state == NETREG_UNINITIALIZED)
565 atomic_inc(&dev->carrier_up_count);
566 atomic_inc(&dev->carrier_down_count);
567 linkwatch_fire_event(dev);
569 EXPORT_SYMBOL_GPL(netif_carrier_event);
571 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
572 under all circumstances. It is difficult to invent anything faster or
576 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
577 struct sk_buff **to_free)
579 __qdisc_drop(skb, to_free);
583 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
588 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
591 .enqueue = noop_enqueue,
592 .dequeue = noop_dequeue,
593 .peek = noop_dequeue,
594 .owner = THIS_MODULE,
597 static struct netdev_queue noop_netdev_queue = {
598 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
599 .qdisc_sleeping = &noop_qdisc,
602 struct Qdisc noop_qdisc = {
603 .enqueue = noop_enqueue,
604 .dequeue = noop_dequeue,
605 .flags = TCQ_F_BUILTIN,
606 .ops = &noop_qdisc_ops,
607 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
608 .dev_queue = &noop_netdev_queue,
609 .running = SEQCNT_ZERO(noop_qdisc.running),
610 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
612 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
613 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
615 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
618 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
619 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
621 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
624 EXPORT_SYMBOL(noop_qdisc);
626 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
627 struct netlink_ext_ack *extack)
629 /* register_qdisc() assigns a default of noop_enqueue if unset,
630 * but __dev_queue_xmit() treats noqueue only as such
631 * if this is NULL - so clear it here. */
632 qdisc->enqueue = NULL;
636 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
639 .init = noqueue_init,
640 .enqueue = noop_enqueue,
641 .dequeue = noop_dequeue,
642 .peek = noop_dequeue,
643 .owner = THIS_MODULE,
646 static const u8 prio2band[TC_PRIO_MAX + 1] = {
647 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
650 /* 3-band FIFO queue: old style, but should be a bit faster than
651 generic prio+fifo combination.
654 #define PFIFO_FAST_BANDS 3
657 * Private data for a pfifo_fast scheduler containing:
658 * - rings for priority bands
660 struct pfifo_fast_priv {
661 struct skb_array q[PFIFO_FAST_BANDS];
664 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
667 return &priv->q[band];
670 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
671 struct sk_buff **to_free)
673 int band = prio2band[skb->priority & TC_PRIO_MAX];
674 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
675 struct skb_array *q = band2list(priv, band);
676 unsigned int pkt_len = qdisc_pkt_len(skb);
679 err = skb_array_produce(q, skb);
682 if (qdisc_is_percpu_stats(qdisc))
683 return qdisc_drop_cpu(skb, qdisc, to_free);
685 return qdisc_drop(skb, qdisc, to_free);
688 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
689 return NET_XMIT_SUCCESS;
692 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
694 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
695 struct sk_buff *skb = NULL;
696 bool need_retry = true;
700 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
701 struct skb_array *q = band2list(priv, band);
703 if (__skb_array_empty(q))
706 skb = __skb_array_consume(q);
709 qdisc_update_stats_at_dequeue(qdisc, skb);
710 } else if (need_retry &&
711 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
712 /* Delay clearing the STATE_MISSED here to reduce
713 * the overhead of the second spin_trylock() in
714 * qdisc_run_begin() and __netif_schedule() calling
715 * in qdisc_run_end().
717 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
718 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
720 /* Make sure dequeuing happens after clearing
723 smp_mb__after_atomic();
733 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
735 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
736 struct sk_buff *skb = NULL;
739 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
740 struct skb_array *q = band2list(priv, band);
742 skb = __skb_array_peek(q);
748 static void pfifo_fast_reset(struct Qdisc *qdisc)
751 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
753 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
754 struct skb_array *q = band2list(priv, band);
757 /* NULL ring is possible if destroy path is due to a failed
758 * skb_array_init() in pfifo_fast_init() case.
763 while ((skb = __skb_array_consume(q)) != NULL)
767 if (qdisc_is_percpu_stats(qdisc)) {
768 for_each_possible_cpu(i) {
769 struct gnet_stats_queue *q;
771 q = per_cpu_ptr(qdisc->cpu_qstats, i);
778 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
780 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
782 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
783 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
784 goto nla_put_failure;
791 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
792 struct netlink_ext_ack *extack)
794 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
795 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
798 /* guard against zero length rings */
802 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
803 struct skb_array *q = band2list(priv, prio);
806 err = skb_array_init(q, qlen, GFP_KERNEL);
811 /* Can by-pass the queue discipline */
812 qdisc->flags |= TCQ_F_CAN_BYPASS;
816 static void pfifo_fast_destroy(struct Qdisc *sch)
818 struct pfifo_fast_priv *priv = qdisc_priv(sch);
821 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
822 struct skb_array *q = band2list(priv, prio);
824 /* NULL ring is possible if destroy path is due to a failed
825 * skb_array_init() in pfifo_fast_init() case.
829 /* Destroy ring but no need to kfree_skb because a call to
830 * pfifo_fast_reset() has already done that work.
832 ptr_ring_cleanup(&q->ring, NULL);
836 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
837 unsigned int new_len)
839 struct pfifo_fast_priv *priv = qdisc_priv(sch);
840 struct skb_array *bands[PFIFO_FAST_BANDS];
843 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
844 struct skb_array *q = band2list(priv, prio);
849 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
853 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
855 .priv_size = sizeof(struct pfifo_fast_priv),
856 .enqueue = pfifo_fast_enqueue,
857 .dequeue = pfifo_fast_dequeue,
858 .peek = pfifo_fast_peek,
859 .init = pfifo_fast_init,
860 .destroy = pfifo_fast_destroy,
861 .reset = pfifo_fast_reset,
862 .dump = pfifo_fast_dump,
863 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
864 .owner = THIS_MODULE,
865 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
867 EXPORT_SYMBOL(pfifo_fast_ops);
869 static struct lock_class_key qdisc_tx_busylock;
870 static struct lock_class_key qdisc_running_key;
872 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
873 const struct Qdisc_ops *ops,
874 struct netlink_ext_ack *extack)
877 unsigned int size = sizeof(*sch) + ops->priv_size;
879 struct net_device *dev;
882 NL_SET_ERR_MSG(extack, "No device queue given");
887 dev = dev_queue->dev;
888 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
892 __skb_queue_head_init(&sch->gso_skb);
893 __skb_queue_head_init(&sch->skb_bad_txq);
894 qdisc_skb_head_init(&sch->q);
895 spin_lock_init(&sch->q.lock);
897 if (ops->static_flags & TCQ_F_CPUSTATS) {
899 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
900 if (!sch->cpu_bstats)
903 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
904 if (!sch->cpu_qstats) {
905 free_percpu(sch->cpu_bstats);
910 spin_lock_init(&sch->busylock);
911 lockdep_set_class(&sch->busylock,
912 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
914 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
915 spin_lock_init(&sch->seqlock);
916 lockdep_set_class(&sch->busylock,
917 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
919 seqcount_init(&sch->running);
920 lockdep_set_class(&sch->running,
921 dev->qdisc_running_key ?: &qdisc_running_key);
924 sch->flags = ops->static_flags;
925 sch->enqueue = ops->enqueue;
926 sch->dequeue = ops->dequeue;
927 sch->dev_queue = dev_queue;
929 refcount_set(&sch->refcnt, 1);
938 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
939 const struct Qdisc_ops *ops,
940 unsigned int parentid,
941 struct netlink_ext_ack *extack)
945 if (!try_module_get(ops->owner)) {
946 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
950 sch = qdisc_alloc(dev_queue, ops, extack);
952 module_put(ops->owner);
955 sch->parent = parentid;
957 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
958 trace_qdisc_create(ops, dev_queue->dev, parentid);
965 EXPORT_SYMBOL(qdisc_create_dflt);
967 /* Under qdisc_lock(qdisc) and BH! */
969 void qdisc_reset(struct Qdisc *qdisc)
971 const struct Qdisc_ops *ops = qdisc->ops;
972 struct sk_buff *skb, *tmp;
974 trace_qdisc_reset(qdisc);
979 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
980 __skb_unlink(skb, &qdisc->gso_skb);
984 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
985 __skb_unlink(skb, &qdisc->skb_bad_txq);
990 qdisc->qstats.backlog = 0;
992 EXPORT_SYMBOL(qdisc_reset);
994 void qdisc_free(struct Qdisc *qdisc)
996 if (qdisc_is_percpu_stats(qdisc)) {
997 free_percpu(qdisc->cpu_bstats);
998 free_percpu(qdisc->cpu_qstats);
1004 static void qdisc_free_cb(struct rcu_head *head)
1006 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1011 static void qdisc_destroy(struct Qdisc *qdisc)
1013 const struct Qdisc_ops *ops = qdisc->ops;
1015 #ifdef CONFIG_NET_SCHED
1016 qdisc_hash_del(qdisc);
1018 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1020 gen_kill_estimator(&qdisc->rate_est);
1025 ops->destroy(qdisc);
1027 module_put(ops->owner);
1028 dev_put(qdisc_dev(qdisc));
1030 trace_qdisc_destroy(qdisc);
1032 call_rcu(&qdisc->rcu, qdisc_free_cb);
1035 void qdisc_put(struct Qdisc *qdisc)
1040 if (qdisc->flags & TCQ_F_BUILTIN ||
1041 !refcount_dec_and_test(&qdisc->refcnt))
1044 qdisc_destroy(qdisc);
1046 EXPORT_SYMBOL(qdisc_put);
1048 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1049 * Intended to be used as optimization, this function only takes rtnl lock if
1050 * qdisc reference counter reached zero.
1053 void qdisc_put_unlocked(struct Qdisc *qdisc)
1055 if (qdisc->flags & TCQ_F_BUILTIN ||
1056 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1059 qdisc_destroy(qdisc);
1062 EXPORT_SYMBOL(qdisc_put_unlocked);
1064 /* Attach toplevel qdisc to device queue. */
1065 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1066 struct Qdisc *qdisc)
1068 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1069 spinlock_t *root_lock;
1071 root_lock = qdisc_lock(oqdisc);
1072 spin_lock_bh(root_lock);
1074 /* ... and graft new one */
1076 qdisc = &noop_qdisc;
1077 dev_queue->qdisc_sleeping = qdisc;
1078 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1080 spin_unlock_bh(root_lock);
1084 EXPORT_SYMBOL(dev_graft_qdisc);
1086 static void attach_one_default_qdisc(struct net_device *dev,
1087 struct netdev_queue *dev_queue,
1090 struct Qdisc *qdisc;
1091 const struct Qdisc_ops *ops = default_qdisc_ops;
1093 if (dev->priv_flags & IFF_NO_QUEUE)
1094 ops = &noqueue_qdisc_ops;
1095 else if(dev->type == ARPHRD_CAN)
1096 ops = &pfifo_fast_ops;
1098 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1102 if (!netif_is_multiqueue(dev))
1103 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1104 dev_queue->qdisc_sleeping = qdisc;
1107 static void attach_default_qdiscs(struct net_device *dev)
1109 struct netdev_queue *txq;
1110 struct Qdisc *qdisc;
1112 txq = netdev_get_tx_queue(dev, 0);
1114 if (!netif_is_multiqueue(dev) ||
1115 dev->priv_flags & IFF_NO_QUEUE) {
1116 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1117 dev->qdisc = txq->qdisc_sleeping;
1118 qdisc_refcount_inc(dev->qdisc);
1120 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1123 qdisc->ops->attach(qdisc);
1127 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1128 if (dev->qdisc == &noop_qdisc) {
1129 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1130 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1131 dev->priv_flags |= IFF_NO_QUEUE;
1132 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1133 dev->qdisc = txq->qdisc_sleeping;
1134 qdisc_refcount_inc(dev->qdisc);
1135 dev->priv_flags ^= IFF_NO_QUEUE;
1138 #ifdef CONFIG_NET_SCHED
1139 if (dev->qdisc != &noop_qdisc)
1140 qdisc_hash_add(dev->qdisc, false);
1144 static void transition_one_qdisc(struct net_device *dev,
1145 struct netdev_queue *dev_queue,
1146 void *_need_watchdog)
1148 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1149 int *need_watchdog_p = _need_watchdog;
1151 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1152 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1154 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1155 if (need_watchdog_p) {
1156 dev_queue->trans_start = 0;
1157 *need_watchdog_p = 1;
1161 void dev_activate(struct net_device *dev)
1165 /* No queueing discipline is attached to device;
1166 * create default one for devices, which need queueing
1167 * and noqueue_qdisc for virtual interfaces
1170 if (dev->qdisc == &noop_qdisc)
1171 attach_default_qdiscs(dev);
1173 if (!netif_carrier_ok(dev))
1174 /* Delay activation until next carrier-on event */
1178 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1179 if (dev_ingress_queue(dev))
1180 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1182 if (need_watchdog) {
1183 netif_trans_update(dev);
1184 dev_watchdog_up(dev);
1187 EXPORT_SYMBOL(dev_activate);
1189 static void qdisc_deactivate(struct Qdisc *qdisc)
1191 if (qdisc->flags & TCQ_F_BUILTIN)
1194 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1197 static void dev_deactivate_queue(struct net_device *dev,
1198 struct netdev_queue *dev_queue,
1199 void *_qdisc_default)
1201 struct Qdisc *qdisc_default = _qdisc_default;
1202 struct Qdisc *qdisc;
1204 qdisc = rtnl_dereference(dev_queue->qdisc);
1206 qdisc_deactivate(qdisc);
1207 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1211 static void dev_reset_queue(struct net_device *dev,
1212 struct netdev_queue *dev_queue,
1215 struct Qdisc *qdisc;
1218 qdisc = dev_queue->qdisc_sleeping;
1222 nolock = qdisc->flags & TCQ_F_NOLOCK;
1225 spin_lock_bh(&qdisc->seqlock);
1226 spin_lock_bh(qdisc_lock(qdisc));
1230 spin_unlock_bh(qdisc_lock(qdisc));
1232 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1233 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1234 spin_unlock_bh(&qdisc->seqlock);
1238 static bool some_qdisc_is_busy(struct net_device *dev)
1242 for (i = 0; i < dev->num_tx_queues; i++) {
1243 struct netdev_queue *dev_queue;
1244 spinlock_t *root_lock;
1248 dev_queue = netdev_get_tx_queue(dev, i);
1249 q = dev_queue->qdisc_sleeping;
1251 root_lock = qdisc_lock(q);
1252 spin_lock_bh(root_lock);
1254 val = (qdisc_is_running(q) ||
1255 test_bit(__QDISC_STATE_SCHED, &q->state));
1257 spin_unlock_bh(root_lock);
1266 * dev_deactivate_many - deactivate transmissions on several devices
1267 * @head: list of devices to deactivate
1269 * This function returns only when all outstanding transmissions
1270 * have completed, unless all devices are in dismantle phase.
1272 void dev_deactivate_many(struct list_head *head)
1274 struct net_device *dev;
1276 list_for_each_entry(dev, head, close_list) {
1277 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1279 if (dev_ingress_queue(dev))
1280 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1283 dev_watchdog_down(dev);
1286 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1287 * outstanding qdisc enqueuing calls.
1288 * This is avoided if all devices are in dismantle phase :
1289 * Caller will call synchronize_net() for us
1293 list_for_each_entry(dev, head, close_list) {
1294 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1296 if (dev_ingress_queue(dev))
1297 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1300 /* Wait for outstanding qdisc_run calls. */
1301 list_for_each_entry(dev, head, close_list) {
1302 while (some_qdisc_is_busy(dev)) {
1303 /* wait_event() would avoid this sleep-loop but would
1304 * require expensive checks in the fast paths of packet
1305 * processing which isn't worth it.
1307 schedule_timeout_uninterruptible(1);
1312 void dev_deactivate(struct net_device *dev)
1316 list_add(&dev->close_list, &single);
1317 dev_deactivate_many(&single);
1320 EXPORT_SYMBOL(dev_deactivate);
1322 static int qdisc_change_tx_queue_len(struct net_device *dev,
1323 struct netdev_queue *dev_queue)
1325 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1326 const struct Qdisc_ops *ops = qdisc->ops;
1328 if (ops->change_tx_queue_len)
1329 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1333 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1335 bool up = dev->flags & IFF_UP;
1340 dev_deactivate(dev);
1342 for (i = 0; i < dev->num_tx_queues; i++) {
1343 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1345 /* TODO: revert changes on a partial failure */
1355 static void dev_init_scheduler_queue(struct net_device *dev,
1356 struct netdev_queue *dev_queue,
1359 struct Qdisc *qdisc = _qdisc;
1361 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1362 dev_queue->qdisc_sleeping = qdisc;
1365 void dev_init_scheduler(struct net_device *dev)
1367 dev->qdisc = &noop_qdisc;
1368 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1369 if (dev_ingress_queue(dev))
1370 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1372 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1375 static void shutdown_scheduler_queue(struct net_device *dev,
1376 struct netdev_queue *dev_queue,
1377 void *_qdisc_default)
1379 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1380 struct Qdisc *qdisc_default = _qdisc_default;
1383 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1384 dev_queue->qdisc_sleeping = qdisc_default;
1390 void dev_shutdown(struct net_device *dev)
1392 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1393 if (dev_ingress_queue(dev))
1394 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1395 qdisc_put(dev->qdisc);
1396 dev->qdisc = &noop_qdisc;
1398 WARN_ON(timer_pending(&dev->watchdog_timer));
1402 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1403 * @rate: Rate to compute reciprocal division values of
1404 * @mult: Multiplier for reciprocal division
1405 * @shift: Shift for reciprocal division
1407 * The multiplier and shift for reciprocal division by rate are stored
1408 * in mult and shift.
1410 * The deal here is to replace a divide by a reciprocal one
1411 * in fast path (a reciprocal divide is a multiply and a shift)
1413 * Normal formula would be :
1414 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1416 * We compute mult/shift to use instead :
1417 * time_in_ns = (len * mult) >> shift;
1419 * We try to get the highest possible mult value for accuracy,
1420 * but have to make sure no overflows will ever happen.
1422 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1424 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1426 u64 factor = NSEC_PER_SEC;
1435 *mult = div64_u64(factor, rate);
1436 if (*mult & (1U << 31) || factor & (1ULL << 63))
1443 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1444 const struct tc_ratespec *conf,
1447 memset(r, 0, sizeof(*r));
1448 r->overhead = conf->overhead;
1449 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1450 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1451 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1453 EXPORT_SYMBOL(psched_ratecfg_precompute);
1455 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1457 r->rate_pkts_ps = pktrate64;
1458 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1460 EXPORT_SYMBOL(psched_ppscfg_precompute);
1462 static void mini_qdisc_rcu_func(struct rcu_head *head)
1466 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1467 struct tcf_proto *tp_head)
1469 /* Protected with chain0->filter_chain_lock.
1470 * Can't access chain directly because tp_head can be NULL.
1472 struct mini_Qdisc *miniq_old =
1473 rcu_dereference_protected(*miniqp->p_miniq, 1);
1474 struct mini_Qdisc *miniq;
1477 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1478 /* Wait for flying RCU callback before it is freed. */
1483 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1484 &miniqp->miniq1 : &miniqp->miniq2;
1486 /* We need to make sure that readers won't see the miniq
1487 * we are about to modify. So wait until previous call_rcu callback
1491 miniq->filter_list = tp_head;
1492 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1495 /* This is counterpart of the rcu barriers above. We need to
1496 * block potential new user of miniq_old until all readers
1497 * are not seeing it.
1499 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1501 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1503 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1504 struct tcf_block *block)
1506 miniqp->miniq1.block = block;
1507 miniqp->miniq2.block = block;
1509 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1511 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1512 struct mini_Qdisc __rcu **p_miniq)
1514 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1515 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1516 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1517 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1518 miniqp->p_miniq = p_miniq;
1520 EXPORT_SYMBOL(mini_qdisc_pair_init);