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 /* Main transmission queue. */
40 /* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
49 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
51 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53 const struct netdev_queue *txq = q->dev_queue;
54 spinlock_t *lock = NULL;
57 if (q->flags & TCQ_F_NOLOCK) {
62 skb = skb_peek(&q->skb_bad_txq);
64 /* check the reason of requeuing without tx lock first */
65 txq = skb_get_tx_queue(txq->dev, skb);
66 if (!netif_xmit_frozen_or_stopped(txq)) {
67 skb = __skb_dequeue(&q->skb_bad_txq);
68 if (qdisc_is_percpu_stats(q)) {
69 qdisc_qstats_cpu_backlog_dec(q, skb);
70 qdisc_qstats_cpu_qlen_dec(q);
72 qdisc_qstats_backlog_dec(q, skb);
86 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
91 skb = __skb_dequeue_bad_txq(q);
96 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
99 spinlock_t *lock = NULL;
101 if (q->flags & TCQ_F_NOLOCK) {
102 lock = qdisc_lock(q);
106 __skb_queue_tail(&q->skb_bad_txq, skb);
108 if (qdisc_is_percpu_stats(q)) {
109 qdisc_qstats_cpu_backlog_inc(q, skb);
110 qdisc_qstats_cpu_qlen_inc(q);
112 qdisc_qstats_backlog_inc(q, skb);
120 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
122 spinlock_t *lock = NULL;
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
130 struct sk_buff *next = skb->next;
132 __skb_queue_tail(&q->gso_skb, skb);
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q)) {
136 qdisc_qstats_cpu_requeues_inc(q);
137 qdisc_qstats_cpu_backlog_inc(q, skb);
138 qdisc_qstats_cpu_qlen_inc(q);
140 q->qstats.requeues++;
141 qdisc_qstats_backlog_inc(q, skb);
152 static void try_bulk_dequeue_skb(struct Qdisc *q,
154 const struct netdev_queue *txq,
157 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
159 while (bytelimit > 0) {
160 struct sk_buff *nskb = q->dequeue(q);
165 bytelimit -= nskb->len; /* covers GSO len */
168 (*packets)++; /* GSO counts as one pkt */
170 skb_mark_not_on_list(skb);
173 /* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
176 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
180 int mapping = skb_get_queue_mapping(skb);
181 struct sk_buff *nskb;
185 nskb = q->dequeue(q);
188 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
189 qdisc_enqueue_skb_bad_txq(q, nskb);
196 skb_mark_not_on_list(skb);
199 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
202 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
205 const struct netdev_queue *txq = q->dev_queue;
206 struct sk_buff *skb = NULL;
209 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
210 spinlock_t *lock = NULL;
212 if (q->flags & TCQ_F_NOLOCK) {
213 lock = qdisc_lock(q);
217 skb = skb_peek(&q->gso_skb);
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
228 /* skb in gso_skb were already validated */
230 if (xfrm_offload(skb))
232 /* check the reason of requeuing without tx lock first */
233 txq = skb_get_tx_queue(txq->dev, skb);
234 if (!netif_xmit_frozen_or_stopped(txq)) {
235 skb = __skb_dequeue(&q->gso_skb);
236 if (qdisc_is_percpu_stats(q)) {
237 qdisc_qstats_cpu_backlog_dec(q, skb);
238 qdisc_qstats_cpu_qlen_dec(q);
240 qdisc_qstats_backlog_dec(q, skb);
253 if ((q->flags & TCQ_F_ONETXQUEUE) &&
254 netif_xmit_frozen_or_stopped(txq))
257 skb = qdisc_dequeue_skb_bad_txq(q);
259 if (skb == SKB_XOFF_MAGIC)
266 if (qdisc_may_bulk(q))
267 try_bulk_dequeue_skb(q, skb, txq, packets);
269 try_bulk_dequeue_skb_slow(q, skb, packets);
272 trace_qdisc_dequeue(q, txq, *packets, skb);
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
285 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
286 struct net_device *dev, struct netdev_queue *txq,
287 spinlock_t *root_lock, bool validate)
289 int ret = NETDEV_TX_BUSY;
292 /* And release qdisc */
294 spin_unlock(root_lock);
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
298 skb = validate_xmit_skb_list(skb, dev, &again);
300 #ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again)) {
303 spin_lock(root_lock);
305 dev_requeue_skb(skb, q);
311 HARD_TX_LOCK(dev, txq, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq))
313 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
315 HARD_TX_UNLOCK(dev, txq);
318 spin_lock(root_lock);
323 spin_lock(root_lock);
325 if (!dev_xmit_complete(ret)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret != NETDEV_TX_BUSY))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev->name, ret, q->q.qlen);
331 dev_requeue_skb(skb, q);
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
345 * netif_tx_lock serializes accesses to device driver.
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
350 * Note, that this procedure can be called by a watchdog timer
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
357 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
359 spinlock_t *root_lock = NULL;
360 struct netdev_queue *txq;
361 struct net_device *dev;
366 skb = dequeue_skb(q, &validate, packets);
370 if (!(q->flags & TCQ_F_NOLOCK))
371 root_lock = qdisc_lock(q);
374 txq = skb_get_tx_queue(dev, skb);
376 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
379 void __qdisc_run(struct Qdisc *q)
381 int quota = dev_tx_weight;
384 while (qdisc_restart(q, &packets)) {
393 unsigned long dev_trans_start(struct net_device *dev)
395 unsigned long val, res;
398 if (is_vlan_dev(dev))
399 dev = vlan_dev_real_dev(dev);
400 else if (netif_is_macvlan(dev))
401 dev = macvlan_dev_real_dev(dev);
402 res = netdev_get_tx_queue(dev, 0)->trans_start;
403 for (i = 1; i < dev->num_tx_queues; i++) {
404 val = netdev_get_tx_queue(dev, i)->trans_start;
405 if (val && time_after(val, res))
411 EXPORT_SYMBOL(dev_trans_start);
413 static void dev_watchdog(struct timer_list *t)
415 struct net_device *dev = from_timer(dev, t, watchdog_timer);
418 if (!qdisc_tx_is_noop(dev)) {
419 if (netif_device_present(dev) &&
420 netif_running(dev) &&
421 netif_carrier_ok(dev)) {
422 int some_queue_timedout = 0;
424 unsigned long trans_start;
426 for (i = 0; i < dev->num_tx_queues; i++) {
427 struct netdev_queue *txq;
429 txq = netdev_get_tx_queue(dev, i);
430 trans_start = txq->trans_start;
431 if (netif_xmit_stopped(txq) &&
432 time_after(jiffies, (trans_start +
433 dev->watchdog_timeo))) {
434 some_queue_timedout = 1;
435 txq->trans_timeout++;
440 if (some_queue_timedout) {
441 trace_net_dev_xmit_timeout(dev, i);
442 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
443 dev->name, netdev_drivername(dev), i);
444 dev->netdev_ops->ndo_tx_timeout(dev, i);
446 if (!mod_timer(&dev->watchdog_timer,
447 round_jiffies(jiffies +
448 dev->watchdog_timeo)))
452 netif_tx_unlock(dev);
457 void __netdev_watchdog_up(struct net_device *dev)
459 if (dev->netdev_ops->ndo_tx_timeout) {
460 if (dev->watchdog_timeo <= 0)
461 dev->watchdog_timeo = 5*HZ;
462 if (!mod_timer(&dev->watchdog_timer,
463 round_jiffies(jiffies + dev->watchdog_timeo)))
467 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
469 static void dev_watchdog_up(struct net_device *dev)
471 __netdev_watchdog_up(dev);
474 static void dev_watchdog_down(struct net_device *dev)
476 netif_tx_lock_bh(dev);
477 if (del_timer(&dev->watchdog_timer))
479 netif_tx_unlock_bh(dev);
483 * netif_carrier_on - set carrier
484 * @dev: network device
486 * Device has detected acquisition of carrier.
488 void netif_carrier_on(struct net_device *dev)
490 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
491 if (dev->reg_state == NETREG_UNINITIALIZED)
493 atomic_inc(&dev->carrier_up_count);
494 linkwatch_fire_event(dev);
495 if (netif_running(dev))
496 __netdev_watchdog_up(dev);
499 EXPORT_SYMBOL(netif_carrier_on);
502 * netif_carrier_off - clear carrier
503 * @dev: network device
505 * Device has detected loss of carrier.
507 void netif_carrier_off(struct net_device *dev)
509 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
510 if (dev->reg_state == NETREG_UNINITIALIZED)
512 atomic_inc(&dev->carrier_down_count);
513 linkwatch_fire_event(dev);
516 EXPORT_SYMBOL(netif_carrier_off);
518 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
519 under all circumstances. It is difficult to invent anything faster or
523 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
524 struct sk_buff **to_free)
526 __qdisc_drop(skb, to_free);
530 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
535 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
538 .enqueue = noop_enqueue,
539 .dequeue = noop_dequeue,
540 .peek = noop_dequeue,
541 .owner = THIS_MODULE,
544 static struct netdev_queue noop_netdev_queue = {
545 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
546 .qdisc_sleeping = &noop_qdisc,
549 struct Qdisc noop_qdisc = {
550 .enqueue = noop_enqueue,
551 .dequeue = noop_dequeue,
552 .flags = TCQ_F_BUILTIN,
553 .ops = &noop_qdisc_ops,
554 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
555 .dev_queue = &noop_netdev_queue,
556 .running = SEQCNT_ZERO(noop_qdisc.running),
557 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
559 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
560 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
562 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
565 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
566 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
568 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
571 EXPORT_SYMBOL(noop_qdisc);
573 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
574 struct netlink_ext_ack *extack)
576 /* register_qdisc() assigns a default of noop_enqueue if unset,
577 * but __dev_queue_xmit() treats noqueue only as such
578 * if this is NULL - so clear it here. */
579 qdisc->enqueue = NULL;
583 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
586 .init = noqueue_init,
587 .enqueue = noop_enqueue,
588 .dequeue = noop_dequeue,
589 .peek = noop_dequeue,
590 .owner = THIS_MODULE,
593 static const u8 prio2band[TC_PRIO_MAX + 1] = {
594 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
597 /* 3-band FIFO queue: old style, but should be a bit faster than
598 generic prio+fifo combination.
601 #define PFIFO_FAST_BANDS 3
604 * Private data for a pfifo_fast scheduler containing:
605 * - rings for priority bands
607 struct pfifo_fast_priv {
608 struct skb_array q[PFIFO_FAST_BANDS];
611 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
614 return &priv->q[band];
617 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
618 struct sk_buff **to_free)
620 int band = prio2band[skb->priority & TC_PRIO_MAX];
621 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
622 struct skb_array *q = band2list(priv, band);
623 unsigned int pkt_len = qdisc_pkt_len(skb);
626 err = skb_array_produce(q, skb);
629 if (qdisc_is_percpu_stats(qdisc))
630 return qdisc_drop_cpu(skb, qdisc, to_free);
632 return qdisc_drop(skb, qdisc, to_free);
635 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
636 return NET_XMIT_SUCCESS;
639 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
641 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
642 struct sk_buff *skb = NULL;
645 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
646 struct skb_array *q = band2list(priv, band);
648 if (__skb_array_empty(q))
651 skb = __skb_array_consume(q);
654 qdisc_update_stats_at_dequeue(qdisc, skb);
656 WRITE_ONCE(qdisc->empty, true);
662 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
664 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
665 struct sk_buff *skb = NULL;
668 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
669 struct skb_array *q = band2list(priv, band);
671 skb = __skb_array_peek(q);
677 static void pfifo_fast_reset(struct Qdisc *qdisc)
680 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
682 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
683 struct skb_array *q = band2list(priv, band);
686 /* NULL ring is possible if destroy path is due to a failed
687 * skb_array_init() in pfifo_fast_init() case.
692 while ((skb = __skb_array_consume(q)) != NULL)
696 if (qdisc_is_percpu_stats(qdisc)) {
697 for_each_possible_cpu(i) {
698 struct gnet_stats_queue *q;
700 q = per_cpu_ptr(qdisc->cpu_qstats, i);
707 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
709 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
711 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
712 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
713 goto nla_put_failure;
720 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
721 struct netlink_ext_ack *extack)
723 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
724 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
727 /* guard against zero length rings */
731 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
732 struct skb_array *q = band2list(priv, prio);
735 err = skb_array_init(q, qlen, GFP_KERNEL);
740 /* Can by-pass the queue discipline */
741 qdisc->flags |= TCQ_F_CAN_BYPASS;
745 static void pfifo_fast_destroy(struct Qdisc *sch)
747 struct pfifo_fast_priv *priv = qdisc_priv(sch);
750 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
751 struct skb_array *q = band2list(priv, prio);
753 /* NULL ring is possible if destroy path is due to a failed
754 * skb_array_init() in pfifo_fast_init() case.
758 /* Destroy ring but no need to kfree_skb because a call to
759 * pfifo_fast_reset() has already done that work.
761 ptr_ring_cleanup(&q->ring, NULL);
765 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
766 unsigned int new_len)
768 struct pfifo_fast_priv *priv = qdisc_priv(sch);
769 struct skb_array *bands[PFIFO_FAST_BANDS];
772 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
773 struct skb_array *q = band2list(priv, prio);
778 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
782 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
784 .priv_size = sizeof(struct pfifo_fast_priv),
785 .enqueue = pfifo_fast_enqueue,
786 .dequeue = pfifo_fast_dequeue,
787 .peek = pfifo_fast_peek,
788 .init = pfifo_fast_init,
789 .destroy = pfifo_fast_destroy,
790 .reset = pfifo_fast_reset,
791 .dump = pfifo_fast_dump,
792 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
793 .owner = THIS_MODULE,
794 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
796 EXPORT_SYMBOL(pfifo_fast_ops);
798 static struct lock_class_key qdisc_tx_busylock;
799 static struct lock_class_key qdisc_running_key;
801 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
802 const struct Qdisc_ops *ops,
803 struct netlink_ext_ack *extack)
806 unsigned int size = sizeof(*sch) + ops->priv_size;
808 struct net_device *dev;
811 NL_SET_ERR_MSG(extack, "No device queue given");
816 dev = dev_queue->dev;
817 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
821 __skb_queue_head_init(&sch->gso_skb);
822 __skb_queue_head_init(&sch->skb_bad_txq);
823 qdisc_skb_head_init(&sch->q);
824 spin_lock_init(&sch->q.lock);
826 if (ops->static_flags & TCQ_F_CPUSTATS) {
828 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
829 if (!sch->cpu_bstats)
832 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
833 if (!sch->cpu_qstats) {
834 free_percpu(sch->cpu_bstats);
839 spin_lock_init(&sch->busylock);
840 lockdep_set_class(&sch->busylock,
841 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
843 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
844 spin_lock_init(&sch->seqlock);
845 lockdep_set_class(&sch->busylock,
846 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
848 seqcount_init(&sch->running);
849 lockdep_set_class(&sch->running,
850 dev->qdisc_running_key ?: &qdisc_running_key);
853 sch->flags = ops->static_flags;
854 sch->enqueue = ops->enqueue;
855 sch->dequeue = ops->dequeue;
856 sch->dev_queue = dev_queue;
859 refcount_set(&sch->refcnt, 1);
868 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
869 const struct Qdisc_ops *ops,
870 unsigned int parentid,
871 struct netlink_ext_ack *extack)
875 if (!try_module_get(ops->owner)) {
876 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
880 sch = qdisc_alloc(dev_queue, ops, extack);
882 module_put(ops->owner);
885 sch->parent = parentid;
887 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
888 trace_qdisc_create(ops, dev_queue->dev, parentid);
895 EXPORT_SYMBOL(qdisc_create_dflt);
897 /* Under qdisc_lock(qdisc) and BH! */
899 void qdisc_reset(struct Qdisc *qdisc)
901 const struct Qdisc_ops *ops = qdisc->ops;
902 struct sk_buff *skb, *tmp;
904 trace_qdisc_reset(qdisc);
909 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
910 __skb_unlink(skb, &qdisc->gso_skb);
914 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
915 __skb_unlink(skb, &qdisc->skb_bad_txq);
920 qdisc->qstats.backlog = 0;
922 EXPORT_SYMBOL(qdisc_reset);
924 void qdisc_free(struct Qdisc *qdisc)
926 if (qdisc_is_percpu_stats(qdisc)) {
927 free_percpu(qdisc->cpu_bstats);
928 free_percpu(qdisc->cpu_qstats);
934 static void qdisc_free_cb(struct rcu_head *head)
936 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
941 static void qdisc_destroy(struct Qdisc *qdisc)
943 const struct Qdisc_ops *ops = qdisc->ops;
945 #ifdef CONFIG_NET_SCHED
946 qdisc_hash_del(qdisc);
948 qdisc_put_stab(rtnl_dereference(qdisc->stab));
950 gen_kill_estimator(&qdisc->rate_est);
957 module_put(ops->owner);
958 dev_put(qdisc_dev(qdisc));
960 trace_qdisc_destroy(qdisc);
962 call_rcu(&qdisc->rcu, qdisc_free_cb);
965 void qdisc_put(struct Qdisc *qdisc)
970 if (qdisc->flags & TCQ_F_BUILTIN ||
971 !refcount_dec_and_test(&qdisc->refcnt))
974 qdisc_destroy(qdisc);
976 EXPORT_SYMBOL(qdisc_put);
978 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
979 * Intended to be used as optimization, this function only takes rtnl lock if
980 * qdisc reference counter reached zero.
983 void qdisc_put_unlocked(struct Qdisc *qdisc)
985 if (qdisc->flags & TCQ_F_BUILTIN ||
986 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
989 qdisc_destroy(qdisc);
992 EXPORT_SYMBOL(qdisc_put_unlocked);
994 /* Attach toplevel qdisc to device queue. */
995 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
998 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
999 spinlock_t *root_lock;
1001 root_lock = qdisc_lock(oqdisc);
1002 spin_lock_bh(root_lock);
1004 /* ... and graft new one */
1006 qdisc = &noop_qdisc;
1007 dev_queue->qdisc_sleeping = qdisc;
1008 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1010 spin_unlock_bh(root_lock);
1014 EXPORT_SYMBOL(dev_graft_qdisc);
1016 static void attach_one_default_qdisc(struct net_device *dev,
1017 struct netdev_queue *dev_queue,
1020 struct Qdisc *qdisc;
1021 const struct Qdisc_ops *ops = default_qdisc_ops;
1023 if (dev->priv_flags & IFF_NO_QUEUE)
1024 ops = &noqueue_qdisc_ops;
1025 else if(dev->type == ARPHRD_CAN)
1026 ops = &pfifo_fast_ops;
1028 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1032 if (!netif_is_multiqueue(dev))
1033 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1034 dev_queue->qdisc_sleeping = qdisc;
1037 static void attach_default_qdiscs(struct net_device *dev)
1039 struct netdev_queue *txq;
1040 struct Qdisc *qdisc;
1042 txq = netdev_get_tx_queue(dev, 0);
1044 if (!netif_is_multiqueue(dev) ||
1045 dev->priv_flags & IFF_NO_QUEUE) {
1046 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1047 dev->qdisc = txq->qdisc_sleeping;
1048 qdisc_refcount_inc(dev->qdisc);
1050 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1053 qdisc->ops->attach(qdisc);
1057 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1058 if (dev->qdisc == &noop_qdisc) {
1059 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1060 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1061 dev->priv_flags |= IFF_NO_QUEUE;
1062 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1063 dev->qdisc = txq->qdisc_sleeping;
1064 qdisc_refcount_inc(dev->qdisc);
1065 dev->priv_flags ^= IFF_NO_QUEUE;
1068 #ifdef CONFIG_NET_SCHED
1069 if (dev->qdisc != &noop_qdisc)
1070 qdisc_hash_add(dev->qdisc, false);
1074 static void transition_one_qdisc(struct net_device *dev,
1075 struct netdev_queue *dev_queue,
1076 void *_need_watchdog)
1078 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1079 int *need_watchdog_p = _need_watchdog;
1081 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1082 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1084 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1085 if (need_watchdog_p) {
1086 dev_queue->trans_start = 0;
1087 *need_watchdog_p = 1;
1091 void dev_activate(struct net_device *dev)
1095 /* No queueing discipline is attached to device;
1096 * create default one for devices, which need queueing
1097 * and noqueue_qdisc for virtual interfaces
1100 if (dev->qdisc == &noop_qdisc)
1101 attach_default_qdiscs(dev);
1103 if (!netif_carrier_ok(dev))
1104 /* Delay activation until next carrier-on event */
1108 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1109 if (dev_ingress_queue(dev))
1110 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1112 if (need_watchdog) {
1113 netif_trans_update(dev);
1114 dev_watchdog_up(dev);
1117 EXPORT_SYMBOL(dev_activate);
1119 static void qdisc_deactivate(struct Qdisc *qdisc)
1121 if (qdisc->flags & TCQ_F_BUILTIN)
1124 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1127 static void dev_deactivate_queue(struct net_device *dev,
1128 struct netdev_queue *dev_queue,
1129 void *_qdisc_default)
1131 struct Qdisc *qdisc_default = _qdisc_default;
1132 struct Qdisc *qdisc;
1134 qdisc = rtnl_dereference(dev_queue->qdisc);
1136 qdisc_deactivate(qdisc);
1137 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1141 static void dev_reset_queue(struct net_device *dev,
1142 struct netdev_queue *dev_queue,
1145 struct Qdisc *qdisc;
1148 qdisc = dev_queue->qdisc_sleeping;
1152 nolock = qdisc->flags & TCQ_F_NOLOCK;
1155 spin_lock_bh(&qdisc->seqlock);
1156 spin_lock_bh(qdisc_lock(qdisc));
1160 spin_unlock_bh(qdisc_lock(qdisc));
1162 spin_unlock_bh(&qdisc->seqlock);
1165 static bool some_qdisc_is_busy(struct net_device *dev)
1169 for (i = 0; i < dev->num_tx_queues; i++) {
1170 struct netdev_queue *dev_queue;
1171 spinlock_t *root_lock;
1175 dev_queue = netdev_get_tx_queue(dev, i);
1176 q = dev_queue->qdisc_sleeping;
1178 root_lock = qdisc_lock(q);
1179 spin_lock_bh(root_lock);
1181 val = (qdisc_is_running(q) ||
1182 test_bit(__QDISC_STATE_SCHED, &q->state));
1184 spin_unlock_bh(root_lock);
1193 * dev_deactivate_many - deactivate transmissions on several devices
1194 * @head: list of devices to deactivate
1196 * This function returns only when all outstanding transmissions
1197 * have completed, unless all devices are in dismantle phase.
1199 void dev_deactivate_many(struct list_head *head)
1201 struct net_device *dev;
1203 list_for_each_entry(dev, head, close_list) {
1204 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1206 if (dev_ingress_queue(dev))
1207 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1210 dev_watchdog_down(dev);
1213 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1214 * outstanding qdisc enqueuing calls.
1215 * This is avoided if all devices are in dismantle phase :
1216 * Caller will call synchronize_net() for us
1220 list_for_each_entry(dev, head, close_list) {
1221 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1223 if (dev_ingress_queue(dev))
1224 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1227 /* Wait for outstanding qdisc_run calls. */
1228 list_for_each_entry(dev, head, close_list) {
1229 while (some_qdisc_is_busy(dev)) {
1230 /* wait_event() would avoid this sleep-loop but would
1231 * require expensive checks in the fast paths of packet
1232 * processing which isn't worth it.
1234 schedule_timeout_uninterruptible(1);
1239 void dev_deactivate(struct net_device *dev)
1243 list_add(&dev->close_list, &single);
1244 dev_deactivate_many(&single);
1247 EXPORT_SYMBOL(dev_deactivate);
1249 static int qdisc_change_tx_queue_len(struct net_device *dev,
1250 struct netdev_queue *dev_queue)
1252 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1253 const struct Qdisc_ops *ops = qdisc->ops;
1255 if (ops->change_tx_queue_len)
1256 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1260 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1262 bool up = dev->flags & IFF_UP;
1267 dev_deactivate(dev);
1269 for (i = 0; i < dev->num_tx_queues; i++) {
1270 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1272 /* TODO: revert changes on a partial failure */
1282 static void dev_init_scheduler_queue(struct net_device *dev,
1283 struct netdev_queue *dev_queue,
1286 struct Qdisc *qdisc = _qdisc;
1288 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1289 dev_queue->qdisc_sleeping = qdisc;
1292 void dev_init_scheduler(struct net_device *dev)
1294 dev->qdisc = &noop_qdisc;
1295 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1296 if (dev_ingress_queue(dev))
1297 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1299 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1302 static void shutdown_scheduler_queue(struct net_device *dev,
1303 struct netdev_queue *dev_queue,
1304 void *_qdisc_default)
1306 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1307 struct Qdisc *qdisc_default = _qdisc_default;
1310 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1311 dev_queue->qdisc_sleeping = qdisc_default;
1317 void dev_shutdown(struct net_device *dev)
1319 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1320 if (dev_ingress_queue(dev))
1321 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1322 qdisc_put(dev->qdisc);
1323 dev->qdisc = &noop_qdisc;
1325 WARN_ON(timer_pending(&dev->watchdog_timer));
1329 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1330 * @rate: Rate to compute reciprocal division values of
1331 * @mult: Multiplier for reciprocal division
1332 * @shift: Shift for reciprocal division
1334 * The multiplier and shift for reciprocal division by rate are stored
1335 * in mult and shift.
1337 * The deal here is to replace a divide by a reciprocal one
1338 * in fast path (a reciprocal divide is a multiply and a shift)
1340 * Normal formula would be :
1341 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1343 * We compute mult/shift to use instead :
1344 * time_in_ns = (len * mult) >> shift;
1346 * We try to get the highest possible mult value for accuracy,
1347 * but have to make sure no overflows will ever happen.
1349 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1351 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1353 u64 factor = NSEC_PER_SEC;
1362 *mult = div64_u64(factor, rate);
1363 if (*mult & (1U << 31) || factor & (1ULL << 63))
1370 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1371 const struct tc_ratespec *conf,
1374 memset(r, 0, sizeof(*r));
1375 r->overhead = conf->overhead;
1376 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1377 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1378 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1380 EXPORT_SYMBOL(psched_ratecfg_precompute);
1382 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1384 r->rate_pkts_ps = pktrate64;
1385 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1387 EXPORT_SYMBOL(psched_ppscfg_precompute);
1389 static void mini_qdisc_rcu_func(struct rcu_head *head)
1393 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1394 struct tcf_proto *tp_head)
1396 /* Protected with chain0->filter_chain_lock.
1397 * Can't access chain directly because tp_head can be NULL.
1399 struct mini_Qdisc *miniq_old =
1400 rcu_dereference_protected(*miniqp->p_miniq, 1);
1401 struct mini_Qdisc *miniq;
1404 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1405 /* Wait for flying RCU callback before it is freed. */
1410 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1411 &miniqp->miniq1 : &miniqp->miniq2;
1413 /* We need to make sure that readers won't see the miniq
1414 * we are about to modify. So wait until previous call_rcu callback
1418 miniq->filter_list = tp_head;
1419 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1422 /* This is counterpart of the rcu barriers above. We need to
1423 * block potential new user of miniq_old until all readers
1424 * are not seeing it.
1426 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1428 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1430 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1431 struct tcf_block *block)
1433 miniqp->miniq1.block = block;
1434 miniqp->miniq2.block = block;
1436 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1438 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1439 struct mini_Qdisc __rcu **p_miniq)
1441 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1442 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1443 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1444 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1445 miniqp->p_miniq = p_miniq;
1447 EXPORT_SYMBOL(mini_qdisc_pair_init);