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);
57 /* Main transmission queue. */
59 /* Modifications to data participating in scheduling must be protected with
60 * qdisc_lock(qdisc) spinlock.
62 * The idea is the following:
63 * - enqueue, dequeue are serialized via qdisc root lock
64 * - ingress filtering is also serialized via qdisc root lock
65 * - updates to tree and tree walking are only done under the rtnl mutex.
68 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
70 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
72 const struct netdev_queue *txq = q->dev_queue;
73 spinlock_t *lock = NULL;
76 if (q->flags & TCQ_F_NOLOCK) {
81 skb = skb_peek(&q->skb_bad_txq);
83 /* check the reason of requeuing without tx lock first */
84 txq = skb_get_tx_queue(txq->dev, skb);
85 if (!netif_xmit_frozen_or_stopped(txq)) {
86 skb = __skb_dequeue(&q->skb_bad_txq);
87 if (qdisc_is_percpu_stats(q)) {
88 qdisc_qstats_cpu_backlog_dec(q, skb);
89 qdisc_qstats_cpu_qlen_dec(q);
91 qdisc_qstats_backlog_dec(q, skb);
96 qdisc_maybe_clear_missed(q, txq);
106 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
108 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
111 skb = __skb_dequeue_bad_txq(q);
116 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
119 spinlock_t *lock = NULL;
121 if (q->flags & TCQ_F_NOLOCK) {
122 lock = qdisc_lock(q);
126 __skb_queue_tail(&q->skb_bad_txq, skb);
128 if (qdisc_is_percpu_stats(q)) {
129 qdisc_qstats_cpu_backlog_inc(q, skb);
130 qdisc_qstats_cpu_qlen_inc(q);
132 qdisc_qstats_backlog_inc(q, skb);
140 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
142 spinlock_t *lock = NULL;
144 if (q->flags & TCQ_F_NOLOCK) {
145 lock = qdisc_lock(q);
150 struct sk_buff *next = skb->next;
152 __skb_queue_tail(&q->gso_skb, skb);
154 /* it's still part of the queue */
155 if (qdisc_is_percpu_stats(q)) {
156 qdisc_qstats_cpu_requeues_inc(q);
157 qdisc_qstats_cpu_backlog_inc(q, skb);
158 qdisc_qstats_cpu_qlen_inc(q);
160 q->qstats.requeues++;
161 qdisc_qstats_backlog_inc(q, skb);
172 static void try_bulk_dequeue_skb(struct Qdisc *q,
174 const struct netdev_queue *txq,
177 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
179 while (bytelimit > 0) {
180 struct sk_buff *nskb = q->dequeue(q);
185 bytelimit -= nskb->len; /* covers GSO len */
188 (*packets)++; /* GSO counts as one pkt */
190 skb_mark_not_on_list(skb);
193 /* This variant of try_bulk_dequeue_skb() makes sure
194 * all skbs in the chain are for the same txq
196 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
200 int mapping = skb_get_queue_mapping(skb);
201 struct sk_buff *nskb;
205 nskb = q->dequeue(q);
208 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
209 qdisc_enqueue_skb_bad_txq(q, nskb);
216 skb_mark_not_on_list(skb);
219 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
220 * A requeued skb (via q->gso_skb) can also be a SKB list.
222 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
225 const struct netdev_queue *txq = q->dev_queue;
226 struct sk_buff *skb = NULL;
229 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
230 spinlock_t *lock = NULL;
232 if (q->flags & TCQ_F_NOLOCK) {
233 lock = qdisc_lock(q);
237 skb = skb_peek(&q->gso_skb);
239 /* skb may be null if another cpu pulls gso_skb off in between
240 * empty check and lock.
248 /* skb in gso_skb were already validated */
250 if (xfrm_offload(skb))
252 /* check the reason of requeuing without tx lock first */
253 txq = skb_get_tx_queue(txq->dev, skb);
254 if (!netif_xmit_frozen_or_stopped(txq)) {
255 skb = __skb_dequeue(&q->gso_skb);
256 if (qdisc_is_percpu_stats(q)) {
257 qdisc_qstats_cpu_backlog_dec(q, skb);
258 qdisc_qstats_cpu_qlen_dec(q);
260 qdisc_qstats_backlog_dec(q, skb);
265 qdisc_maybe_clear_missed(q, txq);
274 if ((q->flags & TCQ_F_ONETXQUEUE) &&
275 netif_xmit_frozen_or_stopped(txq)) {
276 qdisc_maybe_clear_missed(q, txq);
280 skb = qdisc_dequeue_skb_bad_txq(q);
282 if (skb == SKB_XOFF_MAGIC)
289 if (qdisc_may_bulk(q))
290 try_bulk_dequeue_skb(q, skb, txq, packets);
292 try_bulk_dequeue_skb_slow(q, skb, packets);
295 trace_qdisc_dequeue(q, txq, *packets, skb);
300 * Transmit possibly several skbs, and handle the return status as
301 * required. Owning running seqcount bit guarantees that
302 * only one CPU can execute this function.
304 * Returns to the caller:
305 * false - hardware queue frozen backoff
306 * true - feel free to send more pkts
308 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
309 struct net_device *dev, struct netdev_queue *txq,
310 spinlock_t *root_lock, bool validate)
312 int ret = NETDEV_TX_BUSY;
315 /* And release qdisc */
317 spin_unlock(root_lock);
319 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
321 skb = validate_xmit_skb_list(skb, dev, &again);
323 #ifdef CONFIG_XFRM_OFFLOAD
324 if (unlikely(again)) {
326 spin_lock(root_lock);
328 dev_requeue_skb(skb, q);
334 HARD_TX_LOCK(dev, txq, smp_processor_id());
335 if (!netif_xmit_frozen_or_stopped(txq))
336 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
338 qdisc_maybe_clear_missed(q, txq);
340 HARD_TX_UNLOCK(dev, txq);
343 spin_lock(root_lock);
348 spin_lock(root_lock);
350 if (!dev_xmit_complete(ret)) {
351 /* Driver returned NETDEV_TX_BUSY - requeue skb */
352 if (unlikely(ret != NETDEV_TX_BUSY))
353 net_warn_ratelimited("BUG %s code %d qlen %d\n",
354 dev->name, ret, q->q.qlen);
356 dev_requeue_skb(skb, q);
364 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
366 * running seqcount guarantees only one CPU can process
367 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
370 * netif_tx_lock serializes accesses to device driver.
372 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
373 * if one is grabbed, another must be free.
375 * Note, that this procedure can be called by a watchdog timer
377 * Returns to the caller:
378 * 0 - queue is empty or throttled.
379 * >0 - queue is not empty.
382 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
384 spinlock_t *root_lock = NULL;
385 struct netdev_queue *txq;
386 struct net_device *dev;
391 skb = dequeue_skb(q, &validate, packets);
395 if (!(q->flags & TCQ_F_NOLOCK))
396 root_lock = qdisc_lock(q);
399 txq = skb_get_tx_queue(dev, skb);
401 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
404 void __qdisc_run(struct Qdisc *q)
406 int quota = READ_ONCE(dev_tx_weight);
409 while (qdisc_restart(q, &packets)) {
418 unsigned long dev_trans_start(struct net_device *dev)
420 unsigned long val, res;
423 if (is_vlan_dev(dev))
424 dev = vlan_dev_real_dev(dev);
425 else if (netif_is_macvlan(dev))
426 dev = macvlan_dev_real_dev(dev);
427 res = netdev_get_tx_queue(dev, 0)->trans_start;
428 for (i = 1; i < dev->num_tx_queues; i++) {
429 val = netdev_get_tx_queue(dev, i)->trans_start;
430 if (val && time_after(val, res))
436 EXPORT_SYMBOL(dev_trans_start);
438 static void dev_watchdog(struct timer_list *t)
440 struct net_device *dev = from_timer(dev, t, watchdog_timer);
443 if (!qdisc_tx_is_noop(dev)) {
444 if (netif_device_present(dev) &&
445 netif_running(dev) &&
446 netif_carrier_ok(dev)) {
447 int some_queue_timedout = 0;
449 unsigned long trans_start;
451 for (i = 0; i < dev->num_tx_queues; i++) {
452 struct netdev_queue *txq;
454 txq = netdev_get_tx_queue(dev, i);
455 trans_start = txq->trans_start;
456 if (netif_xmit_stopped(txq) &&
457 time_after(jiffies, (trans_start +
458 dev->watchdog_timeo))) {
459 some_queue_timedout = 1;
460 txq->trans_timeout++;
465 if (some_queue_timedout) {
466 trace_net_dev_xmit_timeout(dev, i);
467 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
468 dev->name, netdev_drivername(dev), i);
469 dev->netdev_ops->ndo_tx_timeout(dev, i);
471 if (!mod_timer(&dev->watchdog_timer,
472 round_jiffies(jiffies +
473 dev->watchdog_timeo)))
477 netif_tx_unlock(dev);
482 void __netdev_watchdog_up(struct net_device *dev)
484 if (dev->netdev_ops->ndo_tx_timeout) {
485 if (dev->watchdog_timeo <= 0)
486 dev->watchdog_timeo = 5*HZ;
487 if (!mod_timer(&dev->watchdog_timer,
488 round_jiffies(jiffies + dev->watchdog_timeo)))
492 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
494 static void dev_watchdog_up(struct net_device *dev)
496 __netdev_watchdog_up(dev);
499 static void dev_watchdog_down(struct net_device *dev)
501 netif_tx_lock_bh(dev);
502 if (del_timer(&dev->watchdog_timer))
504 netif_tx_unlock_bh(dev);
508 * netif_carrier_on - set carrier
509 * @dev: network device
511 * Device has detected acquisition of carrier.
513 void netif_carrier_on(struct net_device *dev)
515 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
516 if (dev->reg_state == NETREG_UNINITIALIZED)
518 atomic_inc(&dev->carrier_up_count);
519 linkwatch_fire_event(dev);
520 if (netif_running(dev))
521 __netdev_watchdog_up(dev);
524 EXPORT_SYMBOL(netif_carrier_on);
527 * netif_carrier_off - clear carrier
528 * @dev: network device
530 * Device has detected loss of carrier.
532 void netif_carrier_off(struct net_device *dev)
534 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
535 if (dev->reg_state == NETREG_UNINITIALIZED)
537 atomic_inc(&dev->carrier_down_count);
538 linkwatch_fire_event(dev);
541 EXPORT_SYMBOL(netif_carrier_off);
543 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
544 under all circumstances. It is difficult to invent anything faster or
548 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
549 struct sk_buff **to_free)
551 __qdisc_drop(skb, to_free);
555 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
560 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
563 .enqueue = noop_enqueue,
564 .dequeue = noop_dequeue,
565 .peek = noop_dequeue,
566 .owner = THIS_MODULE,
569 static struct netdev_queue noop_netdev_queue = {
570 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
571 .qdisc_sleeping = &noop_qdisc,
574 struct Qdisc noop_qdisc = {
575 .enqueue = noop_enqueue,
576 .dequeue = noop_dequeue,
577 .flags = TCQ_F_BUILTIN,
578 .ops = &noop_qdisc_ops,
579 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
580 .dev_queue = &noop_netdev_queue,
581 .running = SEQCNT_ZERO(noop_qdisc.running),
582 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
584 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
585 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
587 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
590 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
591 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
593 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
596 EXPORT_SYMBOL(noop_qdisc);
598 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
599 struct netlink_ext_ack *extack)
601 /* register_qdisc() assigns a default of noop_enqueue if unset,
602 * but __dev_queue_xmit() treats noqueue only as such
603 * if this is NULL - so clear it here. */
604 qdisc->enqueue = NULL;
608 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
611 .init = noqueue_init,
612 .enqueue = noop_enqueue,
613 .dequeue = noop_dequeue,
614 .peek = noop_dequeue,
615 .owner = THIS_MODULE,
618 static const u8 prio2band[TC_PRIO_MAX + 1] = {
619 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
622 /* 3-band FIFO queue: old style, but should be a bit faster than
623 generic prio+fifo combination.
626 #define PFIFO_FAST_BANDS 3
629 * Private data for a pfifo_fast scheduler containing:
630 * - rings for priority bands
632 struct pfifo_fast_priv {
633 struct skb_array q[PFIFO_FAST_BANDS];
636 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
639 return &priv->q[band];
642 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
643 struct sk_buff **to_free)
645 int band = prio2band[skb->priority & TC_PRIO_MAX];
646 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
647 struct skb_array *q = band2list(priv, band);
648 unsigned int pkt_len = qdisc_pkt_len(skb);
651 err = skb_array_produce(q, skb);
654 if (qdisc_is_percpu_stats(qdisc))
655 return qdisc_drop_cpu(skb, qdisc, to_free);
657 return qdisc_drop(skb, qdisc, to_free);
660 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
661 return NET_XMIT_SUCCESS;
664 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
666 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
667 struct sk_buff *skb = NULL;
668 bool need_retry = true;
672 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
673 struct skb_array *q = band2list(priv, band);
675 if (__skb_array_empty(q))
678 skb = __skb_array_consume(q);
681 qdisc_update_stats_at_dequeue(qdisc, skb);
682 } else if (need_retry &&
683 test_bit(__QDISC_STATE_MISSED, &qdisc->state)) {
684 /* Delay clearing the STATE_MISSED here to reduce
685 * the overhead of the second spin_trylock() in
686 * qdisc_run_begin() and __netif_schedule() calling
687 * in qdisc_run_end().
689 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
691 /* Make sure dequeuing happens after clearing
694 smp_mb__after_atomic();
700 WRITE_ONCE(qdisc->empty, true);
706 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
708 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
709 struct sk_buff *skb = NULL;
712 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
713 struct skb_array *q = band2list(priv, band);
715 skb = __skb_array_peek(q);
721 static void pfifo_fast_reset(struct Qdisc *qdisc)
724 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
726 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
727 struct skb_array *q = band2list(priv, band);
730 /* NULL ring is possible if destroy path is due to a failed
731 * skb_array_init() in pfifo_fast_init() case.
736 while ((skb = __skb_array_consume(q)) != NULL)
740 if (qdisc_is_percpu_stats(qdisc)) {
741 for_each_possible_cpu(i) {
742 struct gnet_stats_queue *q;
744 q = per_cpu_ptr(qdisc->cpu_qstats, i);
751 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
753 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
755 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
756 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
757 goto nla_put_failure;
764 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
765 struct netlink_ext_ack *extack)
767 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
768 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
771 /* guard against zero length rings */
775 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
776 struct skb_array *q = band2list(priv, prio);
779 err = skb_array_init(q, qlen, GFP_KERNEL);
784 /* Can by-pass the queue discipline */
785 qdisc->flags |= TCQ_F_CAN_BYPASS;
789 static void pfifo_fast_destroy(struct Qdisc *sch)
791 struct pfifo_fast_priv *priv = qdisc_priv(sch);
794 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
795 struct skb_array *q = band2list(priv, prio);
797 /* NULL ring is possible if destroy path is due to a failed
798 * skb_array_init() in pfifo_fast_init() case.
802 /* Destroy ring but no need to kfree_skb because a call to
803 * pfifo_fast_reset() has already done that work.
805 ptr_ring_cleanup(&q->ring, NULL);
809 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
810 unsigned int new_len)
812 struct pfifo_fast_priv *priv = qdisc_priv(sch);
813 struct skb_array *bands[PFIFO_FAST_BANDS];
816 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
817 struct skb_array *q = band2list(priv, prio);
822 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
826 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
828 .priv_size = sizeof(struct pfifo_fast_priv),
829 .enqueue = pfifo_fast_enqueue,
830 .dequeue = pfifo_fast_dequeue,
831 .peek = pfifo_fast_peek,
832 .init = pfifo_fast_init,
833 .destroy = pfifo_fast_destroy,
834 .reset = pfifo_fast_reset,
835 .dump = pfifo_fast_dump,
836 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
837 .owner = THIS_MODULE,
838 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
840 EXPORT_SYMBOL(pfifo_fast_ops);
842 static struct lock_class_key qdisc_tx_busylock;
843 static struct lock_class_key qdisc_running_key;
845 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
846 const struct Qdisc_ops *ops,
847 struct netlink_ext_ack *extack)
850 unsigned int size = sizeof(*sch) + ops->priv_size;
852 struct net_device *dev;
855 NL_SET_ERR_MSG(extack, "No device queue given");
860 dev = dev_queue->dev;
861 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
865 __skb_queue_head_init(&sch->gso_skb);
866 __skb_queue_head_init(&sch->skb_bad_txq);
867 qdisc_skb_head_init(&sch->q);
868 spin_lock_init(&sch->q.lock);
870 if (ops->static_flags & TCQ_F_CPUSTATS) {
872 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
873 if (!sch->cpu_bstats)
876 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
877 if (!sch->cpu_qstats) {
878 free_percpu(sch->cpu_bstats);
883 spin_lock_init(&sch->busylock);
884 lockdep_set_class(&sch->busylock,
885 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
887 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
888 spin_lock_init(&sch->seqlock);
889 lockdep_set_class(&sch->seqlock,
890 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
892 seqcount_init(&sch->running);
893 lockdep_set_class(&sch->running,
894 dev->qdisc_running_key ?: &qdisc_running_key);
897 sch->flags = ops->static_flags;
898 sch->enqueue = ops->enqueue;
899 sch->dequeue = ops->dequeue;
900 sch->dev_queue = dev_queue;
903 refcount_set(&sch->refcnt, 1);
912 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
913 const struct Qdisc_ops *ops,
914 unsigned int parentid,
915 struct netlink_ext_ack *extack)
919 if (!try_module_get(ops->owner)) {
920 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
924 sch = qdisc_alloc(dev_queue, ops, extack);
926 module_put(ops->owner);
929 sch->parent = parentid;
931 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
932 trace_qdisc_create(ops, dev_queue->dev, parentid);
939 EXPORT_SYMBOL(qdisc_create_dflt);
941 /* Under qdisc_lock(qdisc) and BH! */
943 void qdisc_reset(struct Qdisc *qdisc)
945 const struct Qdisc_ops *ops = qdisc->ops;
946 struct sk_buff *skb, *tmp;
948 trace_qdisc_reset(qdisc);
953 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
954 __skb_unlink(skb, &qdisc->gso_skb);
958 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
959 __skb_unlink(skb, &qdisc->skb_bad_txq);
964 qdisc->qstats.backlog = 0;
966 EXPORT_SYMBOL(qdisc_reset);
968 void qdisc_free(struct Qdisc *qdisc)
970 if (qdisc_is_percpu_stats(qdisc)) {
971 free_percpu(qdisc->cpu_bstats);
972 free_percpu(qdisc->cpu_qstats);
978 static void qdisc_free_cb(struct rcu_head *head)
980 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
985 static void qdisc_destroy(struct Qdisc *qdisc)
987 const struct Qdisc_ops *ops = qdisc->ops;
989 #ifdef CONFIG_NET_SCHED
990 qdisc_hash_del(qdisc);
992 qdisc_put_stab(rtnl_dereference(qdisc->stab));
994 gen_kill_estimator(&qdisc->rate_est);
1001 module_put(ops->owner);
1002 dev_put(qdisc_dev(qdisc));
1004 trace_qdisc_destroy(qdisc);
1006 call_rcu(&qdisc->rcu, qdisc_free_cb);
1009 void qdisc_put(struct Qdisc *qdisc)
1014 if (qdisc->flags & TCQ_F_BUILTIN ||
1015 !refcount_dec_and_test(&qdisc->refcnt))
1018 qdisc_destroy(qdisc);
1020 EXPORT_SYMBOL(qdisc_put);
1022 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1023 * Intended to be used as optimization, this function only takes rtnl lock if
1024 * qdisc reference counter reached zero.
1027 void qdisc_put_unlocked(struct Qdisc *qdisc)
1029 if (qdisc->flags & TCQ_F_BUILTIN ||
1030 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1033 qdisc_destroy(qdisc);
1036 EXPORT_SYMBOL(qdisc_put_unlocked);
1038 /* Attach toplevel qdisc to device queue. */
1039 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1040 struct Qdisc *qdisc)
1042 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1043 spinlock_t *root_lock;
1045 root_lock = qdisc_lock(oqdisc);
1046 spin_lock_bh(root_lock);
1048 /* ... and graft new one */
1050 qdisc = &noop_qdisc;
1051 dev_queue->qdisc_sleeping = qdisc;
1052 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1054 spin_unlock_bh(root_lock);
1058 EXPORT_SYMBOL(dev_graft_qdisc);
1060 static void shutdown_scheduler_queue(struct net_device *dev,
1061 struct netdev_queue *dev_queue,
1062 void *_qdisc_default)
1064 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1065 struct Qdisc *qdisc_default = _qdisc_default;
1068 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1069 dev_queue->qdisc_sleeping = qdisc_default;
1075 static void attach_one_default_qdisc(struct net_device *dev,
1076 struct netdev_queue *dev_queue,
1079 struct Qdisc *qdisc;
1080 const struct Qdisc_ops *ops = default_qdisc_ops;
1082 if (dev->priv_flags & IFF_NO_QUEUE)
1083 ops = &noqueue_qdisc_ops;
1084 else if(dev->type == ARPHRD_CAN)
1085 ops = &pfifo_fast_ops;
1087 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1091 if (!netif_is_multiqueue(dev))
1092 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1093 dev_queue->qdisc_sleeping = qdisc;
1096 static void attach_default_qdiscs(struct net_device *dev)
1098 struct netdev_queue *txq;
1099 struct Qdisc *qdisc;
1101 txq = netdev_get_tx_queue(dev, 0);
1103 if (!netif_is_multiqueue(dev) ||
1104 dev->priv_flags & IFF_NO_QUEUE) {
1105 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1106 qdisc = txq->qdisc_sleeping;
1107 rcu_assign_pointer(dev->qdisc, qdisc);
1108 qdisc_refcount_inc(qdisc);
1110 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1112 rcu_assign_pointer(dev->qdisc, qdisc);
1113 qdisc->ops->attach(qdisc);
1116 qdisc = rtnl_dereference(dev->qdisc);
1118 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1119 if (qdisc == &noop_qdisc) {
1120 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1121 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1122 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1123 dev->priv_flags |= IFF_NO_QUEUE;
1124 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1125 qdisc = txq->qdisc_sleeping;
1126 rcu_assign_pointer(dev->qdisc, qdisc);
1127 qdisc_refcount_inc(qdisc);
1128 dev->priv_flags ^= IFF_NO_QUEUE;
1131 #ifdef CONFIG_NET_SCHED
1132 if (qdisc != &noop_qdisc)
1133 qdisc_hash_add(qdisc, false);
1137 static void transition_one_qdisc(struct net_device *dev,
1138 struct netdev_queue *dev_queue,
1139 void *_need_watchdog)
1141 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1142 int *need_watchdog_p = _need_watchdog;
1144 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1145 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1147 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1148 if (need_watchdog_p) {
1149 dev_queue->trans_start = 0;
1150 *need_watchdog_p = 1;
1154 void dev_activate(struct net_device *dev)
1158 /* No queueing discipline is attached to device;
1159 * create default one for devices, which need queueing
1160 * and noqueue_qdisc for virtual interfaces
1163 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1164 attach_default_qdiscs(dev);
1166 if (!netif_carrier_ok(dev))
1167 /* Delay activation until next carrier-on event */
1171 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1172 if (dev_ingress_queue(dev))
1173 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1175 if (need_watchdog) {
1176 netif_trans_update(dev);
1177 dev_watchdog_up(dev);
1180 EXPORT_SYMBOL(dev_activate);
1182 static void qdisc_deactivate(struct Qdisc *qdisc)
1184 if (qdisc->flags & TCQ_F_BUILTIN)
1187 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1190 static void dev_deactivate_queue(struct net_device *dev,
1191 struct netdev_queue *dev_queue,
1192 void *_qdisc_default)
1194 struct Qdisc *qdisc_default = _qdisc_default;
1195 struct Qdisc *qdisc;
1197 qdisc = rtnl_dereference(dev_queue->qdisc);
1199 qdisc_deactivate(qdisc);
1200 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1204 static void dev_reset_queue(struct net_device *dev,
1205 struct netdev_queue *dev_queue,
1208 struct Qdisc *qdisc;
1211 qdisc = dev_queue->qdisc_sleeping;
1215 nolock = qdisc->flags & TCQ_F_NOLOCK;
1218 spin_lock_bh(&qdisc->seqlock);
1219 spin_lock_bh(qdisc_lock(qdisc));
1223 spin_unlock_bh(qdisc_lock(qdisc));
1225 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1226 spin_unlock_bh(&qdisc->seqlock);
1230 static bool some_qdisc_is_busy(struct net_device *dev)
1234 for (i = 0; i < dev->num_tx_queues; i++) {
1235 struct netdev_queue *dev_queue;
1236 spinlock_t *root_lock;
1240 dev_queue = netdev_get_tx_queue(dev, i);
1241 q = dev_queue->qdisc_sleeping;
1243 root_lock = qdisc_lock(q);
1244 spin_lock_bh(root_lock);
1246 val = (qdisc_is_running(q) ||
1247 test_bit(__QDISC_STATE_SCHED, &q->state));
1249 spin_unlock_bh(root_lock);
1258 * dev_deactivate_many - deactivate transmissions on several devices
1259 * @head: list of devices to deactivate
1261 * This function returns only when all outstanding transmissions
1262 * have completed, unless all devices are in dismantle phase.
1264 void dev_deactivate_many(struct list_head *head)
1266 struct net_device *dev;
1268 list_for_each_entry(dev, head, close_list) {
1269 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1271 if (dev_ingress_queue(dev))
1272 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1275 dev_watchdog_down(dev);
1278 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1279 * outstanding qdisc enqueuing calls.
1280 * This is avoided if all devices are in dismantle phase :
1281 * Caller will call synchronize_net() for us
1285 list_for_each_entry(dev, head, close_list) {
1286 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1288 if (dev_ingress_queue(dev))
1289 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1292 /* Wait for outstanding qdisc_run calls. */
1293 list_for_each_entry(dev, head, close_list) {
1294 while (some_qdisc_is_busy(dev)) {
1295 /* wait_event() would avoid this sleep-loop but would
1296 * require expensive checks in the fast paths of packet
1297 * processing which isn't worth it.
1299 schedule_timeout_uninterruptible(1);
1304 void dev_deactivate(struct net_device *dev)
1308 list_add(&dev->close_list, &single);
1309 dev_deactivate_many(&single);
1312 EXPORT_SYMBOL(dev_deactivate);
1314 static int qdisc_change_tx_queue_len(struct net_device *dev,
1315 struct netdev_queue *dev_queue)
1317 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1318 const struct Qdisc_ops *ops = qdisc->ops;
1320 if (ops->change_tx_queue_len)
1321 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1325 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1326 unsigned int new_real_tx)
1328 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1330 if (qdisc->ops->change_real_num_tx)
1331 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1334 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1336 bool up = dev->flags & IFF_UP;
1341 dev_deactivate(dev);
1343 for (i = 0; i < dev->num_tx_queues; i++) {
1344 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1346 /* TODO: revert changes on a partial failure */
1356 static void dev_init_scheduler_queue(struct net_device *dev,
1357 struct netdev_queue *dev_queue,
1360 struct Qdisc *qdisc = _qdisc;
1362 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1363 dev_queue->qdisc_sleeping = qdisc;
1366 void dev_init_scheduler(struct net_device *dev)
1368 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1369 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1370 if (dev_ingress_queue(dev))
1371 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1373 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1376 void dev_shutdown(struct net_device *dev)
1378 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1379 if (dev_ingress_queue(dev))
1380 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1381 qdisc_put(rtnl_dereference(dev->qdisc));
1382 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1384 WARN_ON(timer_pending(&dev->watchdog_timer));
1387 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1388 const struct tc_ratespec *conf,
1391 memset(r, 0, sizeof(*r));
1392 r->overhead = conf->overhead;
1394 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1395 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1398 * The deal here is to replace a divide by a reciprocal one
1399 * in fast path (a reciprocal divide is a multiply and a shift)
1401 * Normal formula would be :
1402 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1404 * We compute mult/shift to use instead :
1405 * time_in_ns = (len * mult) >> shift;
1407 * We try to get the highest possible mult value for accuracy,
1408 * but have to make sure no overflows will ever happen.
1410 if (r->rate_bytes_ps > 0) {
1411 u64 factor = NSEC_PER_SEC;
1414 r->mult = div64_u64(factor, r->rate_bytes_ps);
1415 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1422 EXPORT_SYMBOL(psched_ratecfg_precompute);
1424 static void mini_qdisc_rcu_func(struct rcu_head *head)
1428 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1429 struct tcf_proto *tp_head)
1431 /* Protected with chain0->filter_chain_lock.
1432 * Can't access chain directly because tp_head can be NULL.
1434 struct mini_Qdisc *miniq_old =
1435 rcu_dereference_protected(*miniqp->p_miniq, 1);
1436 struct mini_Qdisc *miniq;
1439 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1440 /* Wait for flying RCU callback before it is freed. */
1445 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1446 &miniqp->miniq1 : &miniqp->miniq2;
1448 /* We need to make sure that readers won't see the miniq
1449 * we are about to modify. So wait until previous call_rcu callback
1453 miniq->filter_list = tp_head;
1454 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1457 /* This is counterpart of the rcu barriers above. We need to
1458 * block potential new user of miniq_old until all readers
1459 * are not seeing it.
1461 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1463 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1465 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1466 struct tcf_block *block)
1468 miniqp->miniq1.block = block;
1469 miniqp->miniq2.block = block;
1471 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1473 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1474 struct mini_Qdisc __rcu **p_miniq)
1476 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1477 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1478 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1479 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1480 miniqp->p_miniq = p_miniq;
1482 EXPORT_SYMBOL(mini_qdisc_pair_init);