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)) {
411 * Ordered by possible occurrence: Postpone processing if
412 * 1. we've exceeded packet quota
413 * 2. another process needs the CPU;
416 if (quota <= 0 || need_resched()) {
423 unsigned long dev_trans_start(struct net_device *dev)
425 unsigned long val, res;
428 if (is_vlan_dev(dev))
429 dev = vlan_dev_real_dev(dev);
430 else if (netif_is_macvlan(dev))
431 dev = macvlan_dev_real_dev(dev);
432 res = netdev_get_tx_queue(dev, 0)->trans_start;
433 for (i = 1; i < dev->num_tx_queues; i++) {
434 val = netdev_get_tx_queue(dev, i)->trans_start;
435 if (val && time_after(val, res))
441 EXPORT_SYMBOL(dev_trans_start);
443 static void dev_watchdog(struct timer_list *t)
445 struct net_device *dev = from_timer(dev, t, watchdog_timer);
448 if (!qdisc_tx_is_noop(dev)) {
449 if (netif_device_present(dev) &&
450 netif_running(dev) &&
451 netif_carrier_ok(dev)) {
452 int some_queue_timedout = 0;
454 unsigned long trans_start;
456 for (i = 0; i < dev->num_tx_queues; i++) {
457 struct netdev_queue *txq;
459 txq = netdev_get_tx_queue(dev, i);
460 trans_start = txq->trans_start;
461 if (netif_xmit_stopped(txq) &&
462 time_after(jiffies, (trans_start +
463 dev->watchdog_timeo))) {
464 some_queue_timedout = 1;
465 txq->trans_timeout++;
470 if (some_queue_timedout) {
471 trace_net_dev_xmit_timeout(dev, i);
472 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
473 dev->name, netdev_drivername(dev), i);
474 dev->netdev_ops->ndo_tx_timeout(dev);
476 if (!mod_timer(&dev->watchdog_timer,
477 round_jiffies(jiffies +
478 dev->watchdog_timeo)))
482 netif_tx_unlock(dev);
487 void __netdev_watchdog_up(struct net_device *dev)
489 if (dev->netdev_ops->ndo_tx_timeout) {
490 if (dev->watchdog_timeo <= 0)
491 dev->watchdog_timeo = 5*HZ;
492 if (!mod_timer(&dev->watchdog_timer,
493 round_jiffies(jiffies + dev->watchdog_timeo)))
497 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
499 static void dev_watchdog_up(struct net_device *dev)
501 __netdev_watchdog_up(dev);
504 static void dev_watchdog_down(struct net_device *dev)
506 netif_tx_lock_bh(dev);
507 if (del_timer(&dev->watchdog_timer))
509 netif_tx_unlock_bh(dev);
513 * netif_carrier_on - set carrier
514 * @dev: network device
516 * Device has detected acquisition of carrier.
518 void netif_carrier_on(struct net_device *dev)
520 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
521 if (dev->reg_state == NETREG_UNINITIALIZED)
523 atomic_inc(&dev->carrier_up_count);
524 linkwatch_fire_event(dev);
525 if (netif_running(dev))
526 __netdev_watchdog_up(dev);
529 EXPORT_SYMBOL(netif_carrier_on);
532 * netif_carrier_off - clear carrier
533 * @dev: network device
535 * Device has detected loss of carrier.
537 void netif_carrier_off(struct net_device *dev)
539 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
540 if (dev->reg_state == NETREG_UNINITIALIZED)
542 atomic_inc(&dev->carrier_down_count);
543 linkwatch_fire_event(dev);
546 EXPORT_SYMBOL(netif_carrier_off);
548 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
549 under all circumstances. It is difficult to invent anything faster or
553 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
554 struct sk_buff **to_free)
556 __qdisc_drop(skb, to_free);
560 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
565 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
568 .enqueue = noop_enqueue,
569 .dequeue = noop_dequeue,
570 .peek = noop_dequeue,
571 .owner = THIS_MODULE,
574 static struct netdev_queue noop_netdev_queue = {
575 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
576 .qdisc_sleeping = &noop_qdisc,
579 struct Qdisc noop_qdisc = {
580 .enqueue = noop_enqueue,
581 .dequeue = noop_dequeue,
582 .flags = TCQ_F_BUILTIN,
583 .ops = &noop_qdisc_ops,
584 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
585 .dev_queue = &noop_netdev_queue,
586 .running = SEQCNT_ZERO(noop_qdisc.running),
587 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
589 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
590 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
592 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
595 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
596 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
598 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
601 EXPORT_SYMBOL(noop_qdisc);
603 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
604 struct netlink_ext_ack *extack)
606 /* register_qdisc() assigns a default of noop_enqueue if unset,
607 * but __dev_queue_xmit() treats noqueue only as such
608 * if this is NULL - so clear it here. */
609 qdisc->enqueue = NULL;
613 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
616 .init = noqueue_init,
617 .enqueue = noop_enqueue,
618 .dequeue = noop_dequeue,
619 .peek = noop_dequeue,
620 .owner = THIS_MODULE,
623 static const u8 prio2band[TC_PRIO_MAX + 1] = {
624 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
627 /* 3-band FIFO queue: old style, but should be a bit faster than
628 generic prio+fifo combination.
631 #define PFIFO_FAST_BANDS 3
634 * Private data for a pfifo_fast scheduler containing:
635 * - rings for priority bands
637 struct pfifo_fast_priv {
638 struct skb_array q[PFIFO_FAST_BANDS];
641 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
644 return &priv->q[band];
647 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
648 struct sk_buff **to_free)
650 int band = prio2band[skb->priority & TC_PRIO_MAX];
651 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
652 struct skb_array *q = band2list(priv, band);
653 unsigned int pkt_len = qdisc_pkt_len(skb);
656 err = skb_array_produce(q, skb);
659 if (qdisc_is_percpu_stats(qdisc))
660 return qdisc_drop_cpu(skb, qdisc, to_free);
662 return qdisc_drop(skb, qdisc, to_free);
665 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
666 return NET_XMIT_SUCCESS;
669 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
671 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
672 struct sk_buff *skb = NULL;
673 bool need_retry = true;
677 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
678 struct skb_array *q = band2list(priv, band);
680 if (__skb_array_empty(q))
683 skb = __skb_array_consume(q);
686 qdisc_update_stats_at_dequeue(qdisc, skb);
687 } else if (need_retry &&
688 test_bit(__QDISC_STATE_MISSED, &qdisc->state)) {
689 /* Delay clearing the STATE_MISSED here to reduce
690 * the overhead of the second spin_trylock() in
691 * qdisc_run_begin() and __netif_schedule() calling
692 * in qdisc_run_end().
694 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
696 /* Make sure dequeuing happens after clearing
699 smp_mb__after_atomic();
705 WRITE_ONCE(qdisc->empty, true);
711 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
713 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
714 struct sk_buff *skb = NULL;
717 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
718 struct skb_array *q = band2list(priv, band);
720 skb = __skb_array_peek(q);
726 static void pfifo_fast_reset(struct Qdisc *qdisc)
729 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
731 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
732 struct skb_array *q = band2list(priv, band);
735 /* NULL ring is possible if destroy path is due to a failed
736 * skb_array_init() in pfifo_fast_init() case.
741 while ((skb = __skb_array_consume(q)) != NULL)
745 if (qdisc_is_percpu_stats(qdisc)) {
746 for_each_possible_cpu(i) {
747 struct gnet_stats_queue *q;
749 q = per_cpu_ptr(qdisc->cpu_qstats, i);
756 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
758 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
760 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
761 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
762 goto nla_put_failure;
769 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
770 struct netlink_ext_ack *extack)
772 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
773 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
776 /* guard against zero length rings */
780 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
781 struct skb_array *q = band2list(priv, prio);
784 err = skb_array_init(q, qlen, GFP_KERNEL);
789 /* Can by-pass the queue discipline */
790 qdisc->flags |= TCQ_F_CAN_BYPASS;
794 static void pfifo_fast_destroy(struct Qdisc *sch)
796 struct pfifo_fast_priv *priv = qdisc_priv(sch);
799 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
800 struct skb_array *q = band2list(priv, prio);
802 /* NULL ring is possible if destroy path is due to a failed
803 * skb_array_init() in pfifo_fast_init() case.
807 /* Destroy ring but no need to kfree_skb because a call to
808 * pfifo_fast_reset() has already done that work.
810 ptr_ring_cleanup(&q->ring, NULL);
814 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
815 unsigned int new_len)
817 struct pfifo_fast_priv *priv = qdisc_priv(sch);
818 struct skb_array *bands[PFIFO_FAST_BANDS];
821 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
822 struct skb_array *q = band2list(priv, prio);
827 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
831 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
833 .priv_size = sizeof(struct pfifo_fast_priv),
834 .enqueue = pfifo_fast_enqueue,
835 .dequeue = pfifo_fast_dequeue,
836 .peek = pfifo_fast_peek,
837 .init = pfifo_fast_init,
838 .destroy = pfifo_fast_destroy,
839 .reset = pfifo_fast_reset,
840 .dump = pfifo_fast_dump,
841 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
842 .owner = THIS_MODULE,
843 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
845 EXPORT_SYMBOL(pfifo_fast_ops);
847 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
848 const struct Qdisc_ops *ops,
849 struct netlink_ext_ack *extack)
853 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
855 struct net_device *dev;
858 NL_SET_ERR_MSG(extack, "No device queue given");
863 dev = dev_queue->dev;
864 p = kzalloc_node(size, GFP_KERNEL,
865 netdev_queue_numa_node_read(dev_queue));
869 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
870 /* if we got non aligned memory, ask more and do alignment ourself */
873 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
874 netdev_queue_numa_node_read(dev_queue));
877 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
878 sch->padded = (char *) sch - (char *) p;
880 __skb_queue_head_init(&sch->gso_skb);
881 __skb_queue_head_init(&sch->skb_bad_txq);
882 qdisc_skb_head_init(&sch->q);
883 spin_lock_init(&sch->q.lock);
885 if (ops->static_flags & TCQ_F_CPUSTATS) {
887 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
888 if (!sch->cpu_bstats)
891 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
892 if (!sch->cpu_qstats) {
893 free_percpu(sch->cpu_bstats);
898 spin_lock_init(&sch->busylock);
899 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
900 spin_lock_init(&sch->seqlock);
901 seqcount_init(&sch->running);
904 sch->flags = ops->static_flags;
905 sch->enqueue = ops->enqueue;
906 sch->dequeue = ops->dequeue;
907 sch->dev_queue = dev_queue;
910 refcount_set(&sch->refcnt, 1);
912 if (sch != &noop_qdisc) {
913 lockdep_set_class(&sch->busylock, &dev->qdisc_tx_busylock_key);
914 lockdep_set_class(&sch->seqlock, &dev->qdisc_tx_busylock_key);
915 lockdep_set_class(&sch->running, &dev->qdisc_running_key);
925 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
926 const struct Qdisc_ops *ops,
927 unsigned int parentid,
928 struct netlink_ext_ack *extack)
932 if (!try_module_get(ops->owner)) {
933 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
937 sch = qdisc_alloc(dev_queue, ops, extack);
939 module_put(ops->owner);
942 sch->parent = parentid;
944 if (!ops->init || ops->init(sch, NULL, extack) == 0)
950 EXPORT_SYMBOL(qdisc_create_dflt);
952 /* Under qdisc_lock(qdisc) and BH! */
954 void qdisc_reset(struct Qdisc *qdisc)
956 const struct Qdisc_ops *ops = qdisc->ops;
957 struct sk_buff *skb, *tmp;
962 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
963 __skb_unlink(skb, &qdisc->gso_skb);
967 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
968 __skb_unlink(skb, &qdisc->skb_bad_txq);
973 qdisc->qstats.backlog = 0;
975 EXPORT_SYMBOL(qdisc_reset);
977 void qdisc_free(struct Qdisc *qdisc)
979 if (qdisc_is_percpu_stats(qdisc)) {
980 free_percpu(qdisc->cpu_bstats);
981 free_percpu(qdisc->cpu_qstats);
984 kfree((char *) qdisc - qdisc->padded);
987 static void qdisc_free_cb(struct rcu_head *head)
989 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
994 static void qdisc_destroy(struct Qdisc *qdisc)
996 const struct Qdisc_ops *ops = qdisc->ops;
997 struct sk_buff *skb, *tmp;
999 #ifdef CONFIG_NET_SCHED
1000 qdisc_hash_del(qdisc);
1002 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1004 gen_kill_estimator(&qdisc->rate_est);
1008 ops->destroy(qdisc);
1010 module_put(ops->owner);
1011 dev_put(qdisc_dev(qdisc));
1013 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
1014 __skb_unlink(skb, &qdisc->gso_skb);
1015 kfree_skb_list(skb);
1018 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
1019 __skb_unlink(skb, &qdisc->skb_bad_txq);
1020 kfree_skb_list(skb);
1023 call_rcu(&qdisc->rcu, qdisc_free_cb);
1026 void qdisc_put(struct Qdisc *qdisc)
1031 if (qdisc->flags & TCQ_F_BUILTIN ||
1032 !refcount_dec_and_test(&qdisc->refcnt))
1035 qdisc_destroy(qdisc);
1037 EXPORT_SYMBOL(qdisc_put);
1039 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1040 * Intended to be used as optimization, this function only takes rtnl lock if
1041 * qdisc reference counter reached zero.
1044 void qdisc_put_unlocked(struct Qdisc *qdisc)
1046 if (qdisc->flags & TCQ_F_BUILTIN ||
1047 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1050 qdisc_destroy(qdisc);
1053 EXPORT_SYMBOL(qdisc_put_unlocked);
1055 /* Attach toplevel qdisc to device queue. */
1056 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1057 struct Qdisc *qdisc)
1059 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1060 spinlock_t *root_lock;
1062 root_lock = qdisc_lock(oqdisc);
1063 spin_lock_bh(root_lock);
1065 /* ... and graft new one */
1067 qdisc = &noop_qdisc;
1068 dev_queue->qdisc_sleeping = qdisc;
1069 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1071 spin_unlock_bh(root_lock);
1075 EXPORT_SYMBOL(dev_graft_qdisc);
1077 static void attach_one_default_qdisc(struct net_device *dev,
1078 struct netdev_queue *dev_queue,
1081 struct Qdisc *qdisc;
1082 const struct Qdisc_ops *ops = default_qdisc_ops;
1084 if (dev->priv_flags & IFF_NO_QUEUE)
1085 ops = &noqueue_qdisc_ops;
1086 else if(dev->type == ARPHRD_CAN)
1087 ops = &pfifo_fast_ops;
1089 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1091 netdev_info(dev, "activation failed\n");
1094 if (!netif_is_multiqueue(dev))
1095 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1096 dev_queue->qdisc_sleeping = qdisc;
1099 static void attach_default_qdiscs(struct net_device *dev)
1101 struct netdev_queue *txq;
1102 struct Qdisc *qdisc;
1104 txq = netdev_get_tx_queue(dev, 0);
1106 if (!netif_is_multiqueue(dev) ||
1107 dev->priv_flags & IFF_NO_QUEUE) {
1108 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1109 qdisc = txq->qdisc_sleeping;
1110 rcu_assign_pointer(dev->qdisc, qdisc);
1111 qdisc_refcount_inc(qdisc);
1113 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1115 rcu_assign_pointer(dev->qdisc, qdisc);
1116 qdisc->ops->attach(qdisc);
1119 qdisc = rtnl_dereference(dev->qdisc);
1121 #ifdef CONFIG_NET_SCHED
1122 if (qdisc != &noop_qdisc)
1123 qdisc_hash_add(qdisc, false);
1127 static void transition_one_qdisc(struct net_device *dev,
1128 struct netdev_queue *dev_queue,
1129 void *_need_watchdog)
1131 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1132 int *need_watchdog_p = _need_watchdog;
1134 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1135 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1137 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1138 if (need_watchdog_p) {
1139 dev_queue->trans_start = 0;
1140 *need_watchdog_p = 1;
1144 void dev_activate(struct net_device *dev)
1148 /* No queueing discipline is attached to device;
1149 * create default one for devices, which need queueing
1150 * and noqueue_qdisc for virtual interfaces
1153 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1154 attach_default_qdiscs(dev);
1156 if (!netif_carrier_ok(dev))
1157 /* Delay activation until next carrier-on event */
1161 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1162 if (dev_ingress_queue(dev))
1163 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1165 if (need_watchdog) {
1166 netif_trans_update(dev);
1167 dev_watchdog_up(dev);
1170 EXPORT_SYMBOL(dev_activate);
1172 static void dev_deactivate_queue(struct net_device *dev,
1173 struct netdev_queue *dev_queue,
1174 void *_qdisc_default)
1176 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc);
1177 struct Qdisc *qdisc_default = _qdisc_default;
1180 if (!(qdisc->flags & TCQ_F_BUILTIN))
1181 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1183 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1187 static void dev_reset_queue(struct net_device *dev,
1188 struct netdev_queue *dev_queue,
1191 struct Qdisc *qdisc;
1194 qdisc = dev_queue->qdisc_sleeping;
1198 nolock = qdisc->flags & TCQ_F_NOLOCK;
1201 spin_lock_bh(&qdisc->seqlock);
1202 spin_lock_bh(qdisc_lock(qdisc));
1206 spin_unlock_bh(qdisc_lock(qdisc));
1208 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1209 spin_unlock_bh(&qdisc->seqlock);
1213 static bool some_qdisc_is_busy(struct net_device *dev)
1217 for (i = 0; i < dev->num_tx_queues; i++) {
1218 struct netdev_queue *dev_queue;
1219 spinlock_t *root_lock;
1223 dev_queue = netdev_get_tx_queue(dev, i);
1224 q = dev_queue->qdisc_sleeping;
1226 root_lock = qdisc_lock(q);
1227 spin_lock_bh(root_lock);
1229 val = (qdisc_is_running(q) ||
1230 test_bit(__QDISC_STATE_SCHED, &q->state));
1232 spin_unlock_bh(root_lock);
1240 static void dev_qdisc_reset(struct net_device *dev,
1241 struct netdev_queue *dev_queue,
1244 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1251 * dev_deactivate_many - deactivate transmissions on several devices
1252 * @head: list of devices to deactivate
1254 * This function returns only when all outstanding transmissions
1255 * have completed, unless all devices are in dismantle phase.
1257 void dev_deactivate_many(struct list_head *head)
1259 struct net_device *dev;
1261 list_for_each_entry(dev, head, close_list) {
1262 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1264 if (dev_ingress_queue(dev))
1265 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1268 dev_watchdog_down(dev);
1271 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1272 * outstanding qdisc enqueuing calls.
1273 * This is avoided if all devices are in dismantle phase :
1274 * Caller will call synchronize_net() for us
1278 list_for_each_entry(dev, head, close_list) {
1279 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1281 if (dev_ingress_queue(dev))
1282 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1285 /* Wait for outstanding qdisc_run calls. */
1286 list_for_each_entry(dev, head, close_list) {
1287 while (some_qdisc_is_busy(dev))
1289 /* The new qdisc is assigned at this point so we can safely
1290 * unwind stale skb lists and qdisc statistics
1292 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1293 if (dev_ingress_queue(dev))
1294 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1298 void dev_deactivate(struct net_device *dev)
1302 list_add(&dev->close_list, &single);
1303 dev_deactivate_many(&single);
1306 EXPORT_SYMBOL(dev_deactivate);
1308 static int qdisc_change_tx_queue_len(struct net_device *dev,
1309 struct netdev_queue *dev_queue)
1311 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1312 const struct Qdisc_ops *ops = qdisc->ops;
1314 if (ops->change_tx_queue_len)
1315 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1319 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1320 unsigned int new_real_tx)
1322 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1324 if (qdisc->ops->change_real_num_tx)
1325 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1328 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1330 bool up = dev->flags & IFF_UP;
1335 dev_deactivate(dev);
1337 for (i = 0; i < dev->num_tx_queues; i++) {
1338 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1340 /* TODO: revert changes on a partial failure */
1350 static void dev_init_scheduler_queue(struct net_device *dev,
1351 struct netdev_queue *dev_queue,
1354 struct Qdisc *qdisc = _qdisc;
1356 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1357 dev_queue->qdisc_sleeping = qdisc;
1360 void dev_init_scheduler(struct net_device *dev)
1362 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1363 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1364 if (dev_ingress_queue(dev))
1365 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1367 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1370 static void shutdown_scheduler_queue(struct net_device *dev,
1371 struct netdev_queue *dev_queue,
1372 void *_qdisc_default)
1374 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1375 struct Qdisc *qdisc_default = _qdisc_default;
1378 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1379 dev_queue->qdisc_sleeping = qdisc_default;
1385 void dev_shutdown(struct net_device *dev)
1387 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1388 if (dev_ingress_queue(dev))
1389 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1390 qdisc_put(rtnl_dereference(dev->qdisc));
1391 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1393 WARN_ON(timer_pending(&dev->watchdog_timer));
1396 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1397 const struct tc_ratespec *conf,
1400 memset(r, 0, sizeof(*r));
1401 r->overhead = conf->overhead;
1403 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1404 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1407 * The deal here is to replace a divide by a reciprocal one
1408 * in fast path (a reciprocal divide is a multiply and a shift)
1410 * Normal formula would be :
1411 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1413 * We compute mult/shift to use instead :
1414 * time_in_ns = (len * mult) >> shift;
1416 * We try to get the highest possible mult value for accuracy,
1417 * but have to make sure no overflows will ever happen.
1419 if (r->rate_bytes_ps > 0) {
1420 u64 factor = NSEC_PER_SEC;
1423 r->mult = div64_u64(factor, r->rate_bytes_ps);
1424 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1431 EXPORT_SYMBOL(psched_ratecfg_precompute);
1433 static void mini_qdisc_rcu_func(struct rcu_head *head)
1437 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1438 struct tcf_proto *tp_head)
1440 /* Protected with chain0->filter_chain_lock.
1441 * Can't access chain directly because tp_head can be NULL.
1443 struct mini_Qdisc *miniq_old =
1444 rcu_dereference_protected(*miniqp->p_miniq, 1);
1445 struct mini_Qdisc *miniq;
1448 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1449 /* Wait for flying RCU callback before it is freed. */
1454 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1455 &miniqp->miniq1 : &miniqp->miniq2;
1457 /* We need to make sure that readers won't see the miniq
1458 * we are about to modify. So wait until previous call_rcu callback
1462 miniq->filter_list = tp_head;
1463 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1466 /* This is counterpart of the rcu barriers above. We need to
1467 * block potential new user of miniq_old until all readers
1468 * are not seeing it.
1470 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1472 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1474 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1475 struct mini_Qdisc __rcu **p_miniq)
1477 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1478 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1479 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1480 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1481 miniqp->p_miniq = p_miniq;
1483 EXPORT_SYMBOL(mini_qdisc_pair_init);