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 <net/pkt_sched.h>
29 /* Main transmission queue. */
31 /* Modifications to data participating in scheduling must be protected with
32 * qdisc_lock(qdisc) spinlock.
34 * The idea is the following:
35 * - enqueue, dequeue are serialized via qdisc root lock
36 * - ingress filtering is also serialized via qdisc root lock
37 * - updates to tree and tree walking are only done under the rtnl mutex.
40 static inline int qdisc_qlen(struct Qdisc *q)
45 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
47 if (unlikely(skb->next))
48 __skb_queue_head(&q->requeue, skb);
50 q->ops->requeue(skb, q);
56 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
60 skb = __skb_dequeue(&q->requeue);
67 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
68 struct netdev_queue *dev_queue,
73 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
75 * Same CPU holding the lock. It may be a transient
76 * configuration error, when hard_start_xmit() recurses. We
77 * detect it by checking xmit owner and drop the packet when
78 * deadloop is detected. Return OK to try the next skb.
82 printk(KERN_WARNING "Dead loop on netdevice %s, "
83 "fix it urgently!\n", dev_queue->dev->name);
87 * Another cpu is holding lock, requeue & delay xmits for
90 __get_cpu_var(netdev_rx_stat).cpu_collision++;
91 ret = dev_requeue_skb(skb, q);
98 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
100 * __QDISC_STATE_RUNNING guarantees only one CPU can process
101 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
104 * netif_tx_lock serializes accesses to device driver.
106 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
107 * if one is grabbed, another must be free.
109 * Note, that this procedure can be called by a watchdog timer
111 * Returns to the caller:
112 * 0 - queue is empty or throttled.
113 * >0 - queue is not empty.
116 static inline int qdisc_restart(struct Qdisc *q)
118 struct netdev_queue *txq;
119 int ret = NETDEV_TX_BUSY;
120 struct net_device *dev;
121 spinlock_t *root_lock;
125 if (unlikely((skb = dequeue_skb(q)) == NULL))
128 root_lock = qdisc_lock(q);
130 /* And release qdisc */
131 spin_unlock(root_lock);
134 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
136 HARD_TX_LOCK(dev, txq, smp_processor_id());
137 if (!netif_tx_queue_stopped(txq) &&
138 !netif_tx_queue_frozen(txq))
139 ret = dev_hard_start_xmit(skb, dev, txq);
140 HARD_TX_UNLOCK(dev, txq);
142 spin_lock(root_lock);
146 /* Driver sent out skb successfully */
150 case NETDEV_TX_LOCKED:
151 /* Driver try lock failed */
152 ret = handle_dev_cpu_collision(skb, txq, q);
156 /* Driver returned NETDEV_TX_BUSY - requeue skb */
157 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
158 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
159 dev->name, ret, q->q.qlen);
161 ret = dev_requeue_skb(skb, q);
165 if (ret && (netif_tx_queue_stopped(txq) ||
166 netif_tx_queue_frozen(txq)))
172 void __qdisc_run(struct Qdisc *q)
174 unsigned long start_time = jiffies;
176 while (qdisc_restart(q)) {
178 * Postpone processing if
179 * 1. another process needs the CPU;
180 * 2. we've been doing it for too long.
182 if (need_resched() || jiffies != start_time) {
188 clear_bit(__QDISC_STATE_RUNNING, &q->state);
191 static void dev_watchdog(unsigned long arg)
193 struct net_device *dev = (struct net_device *)arg;
196 if (!qdisc_tx_is_noop(dev)) {
197 if (netif_device_present(dev) &&
198 netif_running(dev) &&
199 netif_carrier_ok(dev)) {
200 int some_queue_stopped = 0;
203 for (i = 0; i < dev->num_tx_queues; i++) {
204 struct netdev_queue *txq;
206 txq = netdev_get_tx_queue(dev, i);
207 if (netif_tx_queue_stopped(txq)) {
208 some_queue_stopped = 1;
213 if (some_queue_stopped &&
214 time_after(jiffies, (dev->trans_start +
215 dev->watchdog_timeo))) {
217 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
218 dev->name, netdev_drivername(dev, drivername, 64));
219 dev->tx_timeout(dev);
221 if (!mod_timer(&dev->watchdog_timer,
222 round_jiffies(jiffies +
223 dev->watchdog_timeo)))
227 netif_tx_unlock(dev);
232 void __netdev_watchdog_up(struct net_device *dev)
234 if (dev->tx_timeout) {
235 if (dev->watchdog_timeo <= 0)
236 dev->watchdog_timeo = 5*HZ;
237 if (!mod_timer(&dev->watchdog_timer,
238 round_jiffies(jiffies + dev->watchdog_timeo)))
243 static void dev_watchdog_up(struct net_device *dev)
245 __netdev_watchdog_up(dev);
248 static void dev_watchdog_down(struct net_device *dev)
250 netif_tx_lock_bh(dev);
251 if (del_timer(&dev->watchdog_timer))
253 netif_tx_unlock_bh(dev);
257 * netif_carrier_on - set carrier
258 * @dev: network device
260 * Device has detected that carrier.
262 void netif_carrier_on(struct net_device *dev)
264 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
265 linkwatch_fire_event(dev);
266 if (netif_running(dev))
267 __netdev_watchdog_up(dev);
270 EXPORT_SYMBOL(netif_carrier_on);
273 * netif_carrier_off - clear carrier
274 * @dev: network device
276 * Device has detected loss of carrier.
278 void netif_carrier_off(struct net_device *dev)
280 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
281 linkwatch_fire_event(dev);
283 EXPORT_SYMBOL(netif_carrier_off);
285 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
286 under all circumstances. It is difficult to invent anything faster or
290 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
296 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
301 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
304 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
310 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
313 .enqueue = noop_enqueue,
314 .dequeue = noop_dequeue,
315 .requeue = noop_requeue,
316 .owner = THIS_MODULE,
319 static struct netdev_queue noop_netdev_queue = {
320 .qdisc = &noop_qdisc,
323 struct Qdisc noop_qdisc = {
324 .enqueue = noop_enqueue,
325 .dequeue = noop_dequeue,
326 .flags = TCQ_F_BUILTIN,
327 .ops = &noop_qdisc_ops,
328 .list = LIST_HEAD_INIT(noop_qdisc.list),
329 .requeue.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
330 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
331 .dev_queue = &noop_netdev_queue,
333 EXPORT_SYMBOL(noop_qdisc);
335 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
338 .enqueue = noop_enqueue,
339 .dequeue = noop_dequeue,
340 .requeue = noop_requeue,
341 .owner = THIS_MODULE,
344 static struct Qdisc noqueue_qdisc;
345 static struct netdev_queue noqueue_netdev_queue = {
346 .qdisc = &noqueue_qdisc,
349 static struct Qdisc noqueue_qdisc = {
351 .dequeue = noop_dequeue,
352 .flags = TCQ_F_BUILTIN,
353 .ops = &noqueue_qdisc_ops,
354 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
355 .requeue.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
356 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
357 .dev_queue = &noqueue_netdev_queue,
361 static const u8 prio2band[TC_PRIO_MAX+1] =
362 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
364 /* 3-band FIFO queue: old style, but should be a bit faster than
365 generic prio+fifo combination.
368 #define PFIFO_FAST_BANDS 3
370 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
373 struct sk_buff_head *list = qdisc_priv(qdisc);
374 return list + prio2band[skb->priority & TC_PRIO_MAX];
377 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
379 struct sk_buff_head *list = prio2list(skb, qdisc);
381 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
383 return __qdisc_enqueue_tail(skb, qdisc, list);
386 return qdisc_drop(skb, qdisc);
389 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
392 struct sk_buff_head *list = qdisc_priv(qdisc);
394 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
395 if (!skb_queue_empty(list + prio)) {
397 return __qdisc_dequeue_head(qdisc, list + prio);
404 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
407 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
410 static void pfifo_fast_reset(struct Qdisc* qdisc)
413 struct sk_buff_head *list = qdisc_priv(qdisc);
415 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
416 __qdisc_reset_queue(qdisc, list + prio);
418 qdisc->qstats.backlog = 0;
422 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
424 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
426 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
427 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
434 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
437 struct sk_buff_head *list = qdisc_priv(qdisc);
439 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
440 skb_queue_head_init(list + prio);
445 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
447 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
448 .enqueue = pfifo_fast_enqueue,
449 .dequeue = pfifo_fast_dequeue,
450 .requeue = pfifo_fast_requeue,
451 .init = pfifo_fast_init,
452 .reset = pfifo_fast_reset,
453 .dump = pfifo_fast_dump,
454 .owner = THIS_MODULE,
457 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
458 struct Qdisc_ops *ops)
465 /* ensure that the Qdisc and the private data are 32-byte aligned */
466 size = QDISC_ALIGN(sizeof(*sch));
467 size += ops->priv_size + (QDISC_ALIGNTO - 1);
469 p = kzalloc(size, GFP_KERNEL);
472 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
473 sch->padded = (char *) sch - (char *) p;
475 INIT_LIST_HEAD(&sch->list);
476 skb_queue_head_init(&sch->requeue);
477 skb_queue_head_init(&sch->q);
479 sch->enqueue = ops->enqueue;
480 sch->dequeue = ops->dequeue;
481 sch->dev_queue = dev_queue;
482 dev_hold(qdisc_dev(sch));
483 atomic_set(&sch->refcnt, 1);
490 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
491 struct netdev_queue *dev_queue,
492 struct Qdisc_ops *ops,
493 unsigned int parentid)
497 sch = qdisc_alloc(dev_queue, ops);
500 sch->parent = parentid;
502 if (!ops->init || ops->init(sch, NULL) == 0)
509 EXPORT_SYMBOL(qdisc_create_dflt);
511 /* Under qdisc_lock(qdisc) and BH! */
513 void qdisc_reset(struct Qdisc *qdisc)
515 const struct Qdisc_ops *ops = qdisc->ops;
520 EXPORT_SYMBOL(qdisc_reset);
522 void qdisc_destroy(struct Qdisc *qdisc)
524 const struct Qdisc_ops *ops = qdisc->ops;
526 if (qdisc->flags & TCQ_F_BUILTIN ||
527 !atomic_dec_and_test(&qdisc->refcnt))
530 #ifdef CONFIG_NET_SCHED
531 qdisc_list_del(qdisc);
533 qdisc_put_stab(qdisc->stab);
535 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
541 module_put(ops->owner);
542 dev_put(qdisc_dev(qdisc));
544 __skb_queue_purge(&qdisc->requeue);
546 kfree((char *) qdisc - qdisc->padded);
548 EXPORT_SYMBOL(qdisc_destroy);
550 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
554 for (i = 0; i < dev->num_tx_queues; i++) {
555 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
557 if (txq->qdisc_sleeping != &noop_qdisc)
563 static void attach_one_default_qdisc(struct net_device *dev,
564 struct netdev_queue *dev_queue,
569 if (dev->tx_queue_len) {
570 qdisc = qdisc_create_dflt(dev, dev_queue,
571 &pfifo_fast_ops, TC_H_ROOT);
573 printk(KERN_INFO "%s: activation failed\n", dev->name);
577 qdisc = &noqueue_qdisc;
579 dev_queue->qdisc_sleeping = qdisc;
582 static void transition_one_qdisc(struct net_device *dev,
583 struct netdev_queue *dev_queue,
584 void *_need_watchdog)
586 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
587 int *need_watchdog_p = _need_watchdog;
589 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
590 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
592 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
593 if (need_watchdog_p && new_qdisc != &noqueue_qdisc)
594 *need_watchdog_p = 1;
597 void dev_activate(struct net_device *dev)
601 /* No queueing discipline is attached to device;
602 create default one i.e. pfifo_fast for devices,
603 which need queueing and noqueue_qdisc for
607 if (dev_all_qdisc_sleeping_noop(dev))
608 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
610 if (!netif_carrier_ok(dev))
611 /* Delay activation until next carrier-on event */
615 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
616 transition_one_qdisc(dev, &dev->rx_queue, NULL);
619 dev->trans_start = jiffies;
620 dev_watchdog_up(dev);
624 static void dev_deactivate_queue(struct net_device *dev,
625 struct netdev_queue *dev_queue,
626 void *_qdisc_default)
628 struct Qdisc *qdisc_default = _qdisc_default;
631 qdisc = dev_queue->qdisc;
633 spin_lock_bh(qdisc_lock(qdisc));
635 if (!(qdisc->flags & TCQ_F_BUILTIN))
636 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
638 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
641 spin_unlock_bh(qdisc_lock(qdisc));
645 static bool some_qdisc_is_busy(struct net_device *dev)
649 for (i = 0; i < dev->num_tx_queues; i++) {
650 struct netdev_queue *dev_queue;
651 spinlock_t *root_lock;
655 dev_queue = netdev_get_tx_queue(dev, i);
656 q = dev_queue->qdisc_sleeping;
657 root_lock = qdisc_lock(q);
659 spin_lock_bh(root_lock);
661 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
662 test_bit(__QDISC_STATE_SCHED, &q->state));
664 spin_unlock_bh(root_lock);
672 void dev_deactivate(struct net_device *dev)
674 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
675 dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
677 dev_watchdog_down(dev);
679 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
682 /* Wait for outstanding qdisc_run calls. */
683 while (some_qdisc_is_busy(dev))
687 static void dev_init_scheduler_queue(struct net_device *dev,
688 struct netdev_queue *dev_queue,
691 struct Qdisc *qdisc = _qdisc;
693 dev_queue->qdisc = qdisc;
694 dev_queue->qdisc_sleeping = qdisc;
697 void dev_init_scheduler(struct net_device *dev)
699 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
700 dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
702 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
705 static void shutdown_scheduler_queue(struct net_device *dev,
706 struct netdev_queue *dev_queue,
707 void *_qdisc_default)
709 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
710 struct Qdisc *qdisc_default = _qdisc_default;
713 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
714 dev_queue->qdisc_sleeping = qdisc_default;
716 qdisc_destroy(qdisc);
720 void dev_shutdown(struct net_device *dev)
722 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
723 shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
724 WARN_ON(timer_pending(&dev->watchdog_timer));