2 * net/sched/sch_generic.c Generic packet scheduler routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
14 #include <linux/bitops.h>
15 #include <linux/module.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/sched.h>
19 #include <linux/string.h>
20 #include <linux/errno.h>
21 #include <linux/netdevice.h>
22 #include <linux/skbuff.h>
23 #include <linux/rtnetlink.h>
24 #include <linux/init.h>
25 #include <linux/rcupdate.h>
26 #include <linux/list.h>
27 #include <linux/slab.h>
28 #include <net/pkt_sched.h>
31 /* Main transmission queue. */
33 /* Modifications to data participating in scheduling must be protected with
34 * qdisc_lock(qdisc) spinlock.
36 * The idea is the following:
37 * - enqueue, dequeue are serialized via qdisc root lock
38 * - ingress filtering is also serialized via qdisc root lock
39 * - updates to tree and tree walking are only done under the rtnl mutex.
42 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
47 q->q.qlen++; /* it's still part of the queue */
53 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
55 struct sk_buff *skb = q->gso_skb;
58 struct net_device *dev = qdisc_dev(q);
59 struct netdev_queue *txq;
61 /* check the reason of requeuing without tx lock first */
62 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
63 if (!netif_tx_queue_stopped(txq) &&
64 !netif_tx_queue_frozen(txq)) {
76 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
77 struct netdev_queue *dev_queue,
82 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
84 * Same CPU holding the lock. It may be a transient
85 * configuration error, when hard_start_xmit() recurses. We
86 * detect it by checking xmit owner and drop the packet when
87 * deadloop is detected. Return OK to try the next skb.
91 printk(KERN_WARNING "Dead loop on netdevice %s, "
92 "fix it urgently!\n", dev_queue->dev->name);
96 * Another cpu is holding lock, requeue & delay xmits for
99 __this_cpu_inc(softnet_data.cpu_collision);
100 ret = dev_requeue_skb(skb, q);
107 * Transmit one skb, and handle the return status as required. Holding the
108 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
111 * Returns to the caller:
112 * 0 - queue is empty or throttled.
113 * >0 - queue is not empty.
115 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
116 struct net_device *dev, struct netdev_queue *txq,
117 spinlock_t *root_lock)
119 int ret = NETDEV_TX_BUSY;
121 /* And release qdisc */
122 spin_unlock(root_lock);
124 HARD_TX_LOCK(dev, txq, smp_processor_id());
125 if (!netif_tx_queue_stopped(txq) && !netif_tx_queue_frozen(txq))
126 ret = dev_hard_start_xmit(skb, dev, txq);
128 HARD_TX_UNLOCK(dev, txq);
130 spin_lock(root_lock);
132 if (dev_xmit_complete(ret)) {
133 /* Driver sent out skb successfully or skb was consumed */
135 } else if (ret == NETDEV_TX_LOCKED) {
136 /* Driver try lock failed */
137 ret = handle_dev_cpu_collision(skb, txq, q);
139 /* Driver returned NETDEV_TX_BUSY - requeue skb */
140 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
141 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
142 dev->name, ret, q->q.qlen);
144 ret = dev_requeue_skb(skb, q);
147 if (ret && (netif_tx_queue_stopped(txq) ||
148 netif_tx_queue_frozen(txq)))
155 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
157 * __QDISC_STATE_RUNNING guarantees only one CPU can process
158 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
161 * netif_tx_lock serializes accesses to device driver.
163 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
164 * if one is grabbed, another must be free.
166 * Note, that this procedure can be called by a watchdog timer
168 * Returns to the caller:
169 * 0 - queue is empty or throttled.
170 * >0 - queue is not empty.
173 static inline int qdisc_restart(struct Qdisc *q)
175 struct netdev_queue *txq;
176 struct net_device *dev;
177 spinlock_t *root_lock;
181 skb = dequeue_skb(q);
184 WARN_ON_ONCE(skb_dst_is_noref(skb));
185 root_lock = qdisc_lock(q);
187 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
189 return sch_direct_xmit(skb, q, dev, txq, root_lock);
192 void __qdisc_run(struct Qdisc *q)
194 unsigned long start_time = jiffies;
196 while (qdisc_restart(q)) {
198 * Postpone processing if
199 * 1. another process needs the CPU;
200 * 2. we've been doing it for too long.
202 if (need_resched() || jiffies != start_time) {
211 unsigned long dev_trans_start(struct net_device *dev)
213 unsigned long val, res = dev->trans_start;
216 for (i = 0; i < dev->num_tx_queues; i++) {
217 val = netdev_get_tx_queue(dev, i)->trans_start;
218 if (val && time_after(val, res))
221 dev->trans_start = res;
224 EXPORT_SYMBOL(dev_trans_start);
226 static void dev_watchdog(unsigned long arg)
228 struct net_device *dev = (struct net_device *)arg;
231 if (!qdisc_tx_is_noop(dev)) {
232 if (netif_device_present(dev) &&
233 netif_running(dev) &&
234 netif_carrier_ok(dev)) {
235 int some_queue_timedout = 0;
237 unsigned long trans_start;
239 for (i = 0; i < dev->num_tx_queues; i++) {
240 struct netdev_queue *txq;
242 txq = netdev_get_tx_queue(dev, i);
244 * old device drivers set dev->trans_start
246 trans_start = txq->trans_start ? : dev->trans_start;
247 if (netif_tx_queue_stopped(txq) &&
248 time_after(jiffies, (trans_start +
249 dev->watchdog_timeo))) {
250 some_queue_timedout = 1;
255 if (some_queue_timedout) {
257 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
258 dev->name, netdev_drivername(dev, drivername, 64), i);
259 dev->netdev_ops->ndo_tx_timeout(dev);
261 if (!mod_timer(&dev->watchdog_timer,
262 round_jiffies(jiffies +
263 dev->watchdog_timeo)))
267 netif_tx_unlock(dev);
272 void __netdev_watchdog_up(struct net_device *dev)
274 if (dev->netdev_ops->ndo_tx_timeout) {
275 if (dev->watchdog_timeo <= 0)
276 dev->watchdog_timeo = 5*HZ;
277 if (!mod_timer(&dev->watchdog_timer,
278 round_jiffies(jiffies + dev->watchdog_timeo)))
283 static void dev_watchdog_up(struct net_device *dev)
285 __netdev_watchdog_up(dev);
288 static void dev_watchdog_down(struct net_device *dev)
290 netif_tx_lock_bh(dev);
291 if (del_timer(&dev->watchdog_timer))
293 netif_tx_unlock_bh(dev);
297 * netif_carrier_on - set carrier
298 * @dev: network device
300 * Device has detected that carrier.
302 void netif_carrier_on(struct net_device *dev)
304 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
305 if (dev->reg_state == NETREG_UNINITIALIZED)
307 linkwatch_fire_event(dev);
308 if (netif_running(dev))
309 __netdev_watchdog_up(dev);
312 EXPORT_SYMBOL(netif_carrier_on);
315 * netif_carrier_off - clear carrier
316 * @dev: network device
318 * Device has detected loss of carrier.
320 void netif_carrier_off(struct net_device *dev)
322 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
323 if (dev->reg_state == NETREG_UNINITIALIZED)
325 linkwatch_fire_event(dev);
328 EXPORT_SYMBOL(netif_carrier_off);
331 * netif_notify_peers - notify network peers about existence of @dev
332 * @dev: network device
334 * Generate traffic such that interested network peers are aware of
335 * @dev, such as by generating a gratuitous ARP. This may be used when
336 * a device wants to inform the rest of the network about some sort of
337 * reconfiguration such as a failover event or virtual machine
340 void netif_notify_peers(struct net_device *dev)
343 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
346 EXPORT_SYMBOL(netif_notify_peers);
348 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
349 under all circumstances. It is difficult to invent anything faster or
353 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
359 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
364 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
367 .enqueue = noop_enqueue,
368 .dequeue = noop_dequeue,
369 .peek = noop_dequeue,
370 .owner = THIS_MODULE,
373 static struct netdev_queue noop_netdev_queue = {
374 .qdisc = &noop_qdisc,
375 .qdisc_sleeping = &noop_qdisc,
378 struct Qdisc noop_qdisc = {
379 .enqueue = noop_enqueue,
380 .dequeue = noop_dequeue,
381 .flags = TCQ_F_BUILTIN,
382 .ops = &noop_qdisc_ops,
383 .list = LIST_HEAD_INIT(noop_qdisc.list),
384 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
385 .dev_queue = &noop_netdev_queue,
386 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
388 EXPORT_SYMBOL(noop_qdisc);
390 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
393 .enqueue = noop_enqueue,
394 .dequeue = noop_dequeue,
395 .peek = noop_dequeue,
396 .owner = THIS_MODULE,
399 static struct Qdisc noqueue_qdisc;
400 static struct netdev_queue noqueue_netdev_queue = {
401 .qdisc = &noqueue_qdisc,
402 .qdisc_sleeping = &noqueue_qdisc,
405 static struct Qdisc noqueue_qdisc = {
407 .dequeue = noop_dequeue,
408 .flags = TCQ_F_BUILTIN,
409 .ops = &noqueue_qdisc_ops,
410 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
411 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
412 .dev_queue = &noqueue_netdev_queue,
413 .busylock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
417 static const u8 prio2band[TC_PRIO_MAX+1] =
418 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
420 /* 3-band FIFO queue: old style, but should be a bit faster than
421 generic prio+fifo combination.
424 #define PFIFO_FAST_BANDS 3
427 * Private data for a pfifo_fast scheduler containing:
428 * - queues for the three band
429 * - bitmap indicating which of the bands contain skbs
431 struct pfifo_fast_priv {
433 struct sk_buff_head q[PFIFO_FAST_BANDS];
437 * Convert a bitmap to the first band number where an skb is queued, where:
438 * bitmap=0 means there are no skbs on any band.
439 * bitmap=1 means there is an skb on band 0.
440 * bitmap=7 means there are skbs on all 3 bands, etc.
442 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
444 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
447 return priv->q + band;
450 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
452 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
453 int band = prio2band[skb->priority & TC_PRIO_MAX];
454 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
455 struct sk_buff_head *list = band2list(priv, band);
457 priv->bitmap |= (1 << band);
459 return __qdisc_enqueue_tail(skb, qdisc, list);
462 return qdisc_drop(skb, qdisc);
465 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
467 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
468 int band = bitmap2band[priv->bitmap];
470 if (likely(band >= 0)) {
471 struct sk_buff_head *list = band2list(priv, band);
472 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
475 if (skb_queue_empty(list))
476 priv->bitmap &= ~(1 << band);
484 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
486 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
487 int band = bitmap2band[priv->bitmap];
490 struct sk_buff_head *list = band2list(priv, band);
492 return skb_peek(list);
498 static void pfifo_fast_reset(struct Qdisc* qdisc)
501 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
503 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
504 __qdisc_reset_queue(qdisc, band2list(priv, prio));
507 qdisc->qstats.backlog = 0;
511 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
513 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
515 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
516 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
523 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
526 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
528 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
529 skb_queue_head_init(band2list(priv, prio));
534 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
536 .priv_size = sizeof(struct pfifo_fast_priv),
537 .enqueue = pfifo_fast_enqueue,
538 .dequeue = pfifo_fast_dequeue,
539 .peek = pfifo_fast_peek,
540 .init = pfifo_fast_init,
541 .reset = pfifo_fast_reset,
542 .dump = pfifo_fast_dump,
543 .owner = THIS_MODULE,
546 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
547 struct Qdisc_ops *ops)
554 /* ensure that the Qdisc and the private data are 64-byte aligned */
555 size = QDISC_ALIGN(sizeof(*sch));
556 size += ops->priv_size + (QDISC_ALIGNTO - 1);
558 p = kzalloc(size, GFP_KERNEL);
561 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
562 sch->padded = (char *) sch - (char *) p;
564 INIT_LIST_HEAD(&sch->list);
565 skb_queue_head_init(&sch->q);
566 spin_lock_init(&sch->busylock);
568 sch->enqueue = ops->enqueue;
569 sch->dequeue = ops->dequeue;
570 sch->dev_queue = dev_queue;
571 dev_hold(qdisc_dev(sch));
572 atomic_set(&sch->refcnt, 1);
579 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
580 struct netdev_queue *dev_queue,
581 struct Qdisc_ops *ops,
582 unsigned int parentid)
586 sch = qdisc_alloc(dev_queue, ops);
589 sch->parent = parentid;
591 if (!ops->init || ops->init(sch, NULL) == 0)
598 EXPORT_SYMBOL(qdisc_create_dflt);
600 /* Under qdisc_lock(qdisc) and BH! */
602 void qdisc_reset(struct Qdisc *qdisc)
604 const struct Qdisc_ops *ops = qdisc->ops;
609 if (qdisc->gso_skb) {
610 kfree_skb(qdisc->gso_skb);
611 qdisc->gso_skb = NULL;
615 EXPORT_SYMBOL(qdisc_reset);
617 static void qdisc_rcu_free(struct rcu_head *head)
619 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
621 kfree((char *) qdisc - qdisc->padded);
624 void qdisc_destroy(struct Qdisc *qdisc)
626 const struct Qdisc_ops *ops = qdisc->ops;
628 if (qdisc->flags & TCQ_F_BUILTIN ||
629 !atomic_dec_and_test(&qdisc->refcnt))
632 #ifdef CONFIG_NET_SCHED
633 qdisc_list_del(qdisc);
635 qdisc_put_stab(qdisc->stab);
637 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
643 module_put(ops->owner);
644 dev_put(qdisc_dev(qdisc));
646 kfree_skb(qdisc->gso_skb);
648 * gen_estimator est_timer() might access qdisc->q.lock,
649 * wait a RCU grace period before freeing qdisc.
651 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
653 EXPORT_SYMBOL(qdisc_destroy);
655 /* Attach toplevel qdisc to device queue. */
656 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
659 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
660 spinlock_t *root_lock;
662 root_lock = qdisc_lock(oqdisc);
663 spin_lock_bh(root_lock);
665 /* Prune old scheduler */
666 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
669 /* ... and graft new one */
672 dev_queue->qdisc_sleeping = qdisc;
673 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
675 spin_unlock_bh(root_lock);
680 static void attach_one_default_qdisc(struct net_device *dev,
681 struct netdev_queue *dev_queue,
686 if (dev->tx_queue_len) {
687 qdisc = qdisc_create_dflt(dev, dev_queue,
688 &pfifo_fast_ops, TC_H_ROOT);
690 printk(KERN_INFO "%s: activation failed\n", dev->name);
694 /* Can by-pass the queue discipline for default qdisc */
695 qdisc->flags |= TCQ_F_CAN_BYPASS;
697 qdisc = &noqueue_qdisc;
699 dev_queue->qdisc_sleeping = qdisc;
702 static void attach_default_qdiscs(struct net_device *dev)
704 struct netdev_queue *txq;
707 txq = netdev_get_tx_queue(dev, 0);
709 if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
710 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
711 dev->qdisc = txq->qdisc_sleeping;
712 atomic_inc(&dev->qdisc->refcnt);
714 qdisc = qdisc_create_dflt(dev, txq, &mq_qdisc_ops, TC_H_ROOT);
716 qdisc->ops->attach(qdisc);
722 static void transition_one_qdisc(struct net_device *dev,
723 struct netdev_queue *dev_queue,
724 void *_need_watchdog)
726 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
727 int *need_watchdog_p = _need_watchdog;
729 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
730 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
732 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
733 if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
734 dev_queue->trans_start = 0;
735 *need_watchdog_p = 1;
739 void dev_activate(struct net_device *dev)
743 /* No queueing discipline is attached to device;
744 create default one i.e. pfifo_fast for devices,
745 which need queueing and noqueue_qdisc for
749 if (dev->qdisc == &noop_qdisc)
750 attach_default_qdiscs(dev);
752 if (!netif_carrier_ok(dev))
753 /* Delay activation until next carrier-on event */
757 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
758 if (dev_ingress_queue(dev))
759 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
762 dev->trans_start = jiffies;
763 dev_watchdog_up(dev);
767 static void dev_deactivate_queue(struct net_device *dev,
768 struct netdev_queue *dev_queue,
769 void *_qdisc_default)
771 struct Qdisc *qdisc_default = _qdisc_default;
774 qdisc = dev_queue->qdisc;
776 spin_lock_bh(qdisc_lock(qdisc));
778 if (!(qdisc->flags & TCQ_F_BUILTIN))
779 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
781 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
784 spin_unlock_bh(qdisc_lock(qdisc));
788 static bool some_qdisc_is_busy(struct net_device *dev)
792 for (i = 0; i < dev->num_tx_queues; i++) {
793 struct netdev_queue *dev_queue;
794 spinlock_t *root_lock;
798 dev_queue = netdev_get_tx_queue(dev, i);
799 q = dev_queue->qdisc_sleeping;
800 root_lock = qdisc_lock(q);
802 spin_lock_bh(root_lock);
804 val = (qdisc_is_running(q) ||
805 test_bit(__QDISC_STATE_SCHED, &q->state));
807 spin_unlock_bh(root_lock);
815 void dev_deactivate(struct net_device *dev)
817 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
818 if (dev_ingress_queue(dev))
819 dev_deactivate_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
821 dev_watchdog_down(dev);
823 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
826 /* Wait for outstanding qdisc_run calls. */
827 while (some_qdisc_is_busy(dev))
831 static void dev_init_scheduler_queue(struct net_device *dev,
832 struct netdev_queue *dev_queue,
835 struct Qdisc *qdisc = _qdisc;
837 dev_queue->qdisc = qdisc;
838 dev_queue->qdisc_sleeping = qdisc;
841 void dev_init_scheduler(struct net_device *dev)
843 dev->qdisc = &noop_qdisc;
844 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
845 if (dev_ingress_queue(dev))
846 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
848 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
851 static void shutdown_scheduler_queue(struct net_device *dev,
852 struct netdev_queue *dev_queue,
853 void *_qdisc_default)
855 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
856 struct Qdisc *qdisc_default = _qdisc_default;
859 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
860 dev_queue->qdisc_sleeping = qdisc_default;
862 qdisc_destroy(qdisc);
866 void dev_shutdown(struct net_device *dev)
868 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
869 if (dev_ingress_queue(dev))
870 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
871 qdisc_destroy(dev->qdisc);
872 dev->qdisc = &noop_qdisc;
874 WARN_ON(timer_pending(&dev->watchdog_timer));