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))
50 __skb_queue_head(&q->requeue, skb);
56 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
58 struct sk_buff *skb = q->gso_skb;
61 skb = skb_peek(&q->requeue);
64 struct net_device *dev = qdisc_dev(q);
65 struct netdev_queue *txq;
67 /* check the reason of requeuing without tx lock first */
68 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
69 if (!netif_tx_queue_stopped(txq) &&
70 !netif_tx_queue_frozen(txq)) {
74 __skb_unlink(skb, &q->requeue);
85 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
86 struct netdev_queue *dev_queue,
91 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
93 * Same CPU holding the lock. It may be a transient
94 * configuration error, when hard_start_xmit() recurses. We
95 * detect it by checking xmit owner and drop the packet when
96 * deadloop is detected. Return OK to try the next skb.
100 printk(KERN_WARNING "Dead loop on netdevice %s, "
101 "fix it urgently!\n", dev_queue->dev->name);
105 * Another cpu is holding lock, requeue & delay xmits for
108 __get_cpu_var(netdev_rx_stat).cpu_collision++;
109 ret = dev_requeue_skb(skb, q);
116 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
118 * __QDISC_STATE_RUNNING guarantees only one CPU can process
119 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
122 * netif_tx_lock serializes accesses to device driver.
124 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
125 * if one is grabbed, another must be free.
127 * Note, that this procedure can be called by a watchdog timer
129 * Returns to the caller:
130 * 0 - queue is empty or throttled.
131 * >0 - queue is not empty.
134 static inline int qdisc_restart(struct Qdisc *q)
136 struct netdev_queue *txq;
137 int ret = NETDEV_TX_BUSY;
138 struct net_device *dev;
139 spinlock_t *root_lock;
143 if (unlikely((skb = dequeue_skb(q)) == NULL))
146 root_lock = qdisc_lock(q);
148 /* And release qdisc */
149 spin_unlock(root_lock);
152 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
154 HARD_TX_LOCK(dev, txq, smp_processor_id());
155 if (!netif_tx_queue_stopped(txq) &&
156 !netif_tx_queue_frozen(txq))
157 ret = dev_hard_start_xmit(skb, dev, txq);
158 HARD_TX_UNLOCK(dev, txq);
160 spin_lock(root_lock);
164 /* Driver sent out skb successfully */
168 case NETDEV_TX_LOCKED:
169 /* Driver try lock failed */
170 ret = handle_dev_cpu_collision(skb, txq, q);
174 /* Driver returned NETDEV_TX_BUSY - requeue skb */
175 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
176 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
177 dev->name, ret, q->q.qlen);
179 ret = dev_requeue_skb(skb, q);
183 if (ret && (netif_tx_queue_stopped(txq) ||
184 netif_tx_queue_frozen(txq)))
190 void __qdisc_run(struct Qdisc *q)
192 unsigned long start_time = jiffies;
194 while (qdisc_restart(q)) {
196 * Postpone processing if
197 * 1. another process needs the CPU;
198 * 2. we've been doing it for too long.
200 if (need_resched() || jiffies != start_time) {
206 clear_bit(__QDISC_STATE_RUNNING, &q->state);
209 static void dev_watchdog(unsigned long arg)
211 struct net_device *dev = (struct net_device *)arg;
214 if (!qdisc_tx_is_noop(dev)) {
215 if (netif_device_present(dev) &&
216 netif_running(dev) &&
217 netif_carrier_ok(dev)) {
218 int some_queue_stopped = 0;
221 for (i = 0; i < dev->num_tx_queues; i++) {
222 struct netdev_queue *txq;
224 txq = netdev_get_tx_queue(dev, i);
225 if (netif_tx_queue_stopped(txq)) {
226 some_queue_stopped = 1;
231 if (some_queue_stopped &&
232 time_after(jiffies, (dev->trans_start +
233 dev->watchdog_timeo))) {
235 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
236 dev->name, netdev_drivername(dev, drivername, 64));
237 dev->tx_timeout(dev);
239 if (!mod_timer(&dev->watchdog_timer,
240 round_jiffies(jiffies +
241 dev->watchdog_timeo)))
245 netif_tx_unlock(dev);
250 void __netdev_watchdog_up(struct net_device *dev)
252 if (dev->tx_timeout) {
253 if (dev->watchdog_timeo <= 0)
254 dev->watchdog_timeo = 5*HZ;
255 if (!mod_timer(&dev->watchdog_timer,
256 round_jiffies(jiffies + dev->watchdog_timeo)))
261 static void dev_watchdog_up(struct net_device *dev)
263 __netdev_watchdog_up(dev);
266 static void dev_watchdog_down(struct net_device *dev)
268 netif_tx_lock_bh(dev);
269 if (del_timer(&dev->watchdog_timer))
271 netif_tx_unlock_bh(dev);
275 * netif_carrier_on - set carrier
276 * @dev: network device
278 * Device has detected that carrier.
280 void netif_carrier_on(struct net_device *dev)
282 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
283 linkwatch_fire_event(dev);
284 if (netif_running(dev))
285 __netdev_watchdog_up(dev);
288 EXPORT_SYMBOL(netif_carrier_on);
291 * netif_carrier_off - clear carrier
292 * @dev: network device
294 * Device has detected loss of carrier.
296 void netif_carrier_off(struct net_device *dev)
298 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
299 linkwatch_fire_event(dev);
301 EXPORT_SYMBOL(netif_carrier_off);
303 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
304 under all circumstances. It is difficult to invent anything faster or
308 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
314 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
319 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
322 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
328 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
331 .enqueue = noop_enqueue,
332 .dequeue = noop_dequeue,
333 .requeue = noop_requeue,
334 .owner = THIS_MODULE,
337 static struct netdev_queue noop_netdev_queue = {
338 .qdisc = &noop_qdisc,
341 struct Qdisc noop_qdisc = {
342 .enqueue = noop_enqueue,
343 .dequeue = noop_dequeue,
344 .flags = TCQ_F_BUILTIN,
345 .ops = &noop_qdisc_ops,
346 .list = LIST_HEAD_INIT(noop_qdisc.list),
347 .requeue.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
348 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
349 .dev_queue = &noop_netdev_queue,
351 EXPORT_SYMBOL(noop_qdisc);
353 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
356 .enqueue = noop_enqueue,
357 .dequeue = noop_dequeue,
358 .requeue = noop_requeue,
359 .owner = THIS_MODULE,
362 static struct Qdisc noqueue_qdisc;
363 static struct netdev_queue noqueue_netdev_queue = {
364 .qdisc = &noqueue_qdisc,
367 static struct Qdisc noqueue_qdisc = {
369 .dequeue = noop_dequeue,
370 .flags = TCQ_F_BUILTIN,
371 .ops = &noqueue_qdisc_ops,
372 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
373 .requeue.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
374 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
375 .dev_queue = &noqueue_netdev_queue,
379 static const u8 prio2band[TC_PRIO_MAX+1] =
380 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
382 /* 3-band FIFO queue: old style, but should be a bit faster than
383 generic prio+fifo combination.
386 #define PFIFO_FAST_BANDS 3
388 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
391 struct sk_buff_head *list = qdisc_priv(qdisc);
392 return list + prio2band[skb->priority & TC_PRIO_MAX];
395 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
397 struct sk_buff_head *list = prio2list(skb, qdisc);
399 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
401 return __qdisc_enqueue_tail(skb, qdisc, list);
404 return qdisc_drop(skb, qdisc);
407 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
410 struct sk_buff_head *list = qdisc_priv(qdisc);
412 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
413 if (!skb_queue_empty(list + prio)) {
415 return __qdisc_dequeue_head(qdisc, list + prio);
422 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
425 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
428 static void pfifo_fast_reset(struct Qdisc* qdisc)
431 struct sk_buff_head *list = qdisc_priv(qdisc);
433 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
434 __qdisc_reset_queue(qdisc, list + prio);
436 qdisc->qstats.backlog = 0;
440 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
442 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
444 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
445 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
452 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
455 struct sk_buff_head *list = qdisc_priv(qdisc);
457 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
458 skb_queue_head_init(list + prio);
463 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
465 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
466 .enqueue = pfifo_fast_enqueue,
467 .dequeue = pfifo_fast_dequeue,
468 .requeue = pfifo_fast_requeue,
469 .init = pfifo_fast_init,
470 .reset = pfifo_fast_reset,
471 .dump = pfifo_fast_dump,
472 .owner = THIS_MODULE,
475 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
476 struct Qdisc_ops *ops)
483 /* ensure that the Qdisc and the private data are 32-byte aligned */
484 size = QDISC_ALIGN(sizeof(*sch));
485 size += ops->priv_size + (QDISC_ALIGNTO - 1);
487 p = kzalloc(size, GFP_KERNEL);
490 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
491 sch->padded = (char *) sch - (char *) p;
493 INIT_LIST_HEAD(&sch->list);
494 skb_queue_head_init(&sch->requeue);
495 skb_queue_head_init(&sch->q);
497 sch->enqueue = ops->enqueue;
498 sch->dequeue = ops->dequeue;
499 sch->dev_queue = dev_queue;
500 dev_hold(qdisc_dev(sch));
501 atomic_set(&sch->refcnt, 1);
508 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
509 struct netdev_queue *dev_queue,
510 struct Qdisc_ops *ops,
511 unsigned int parentid)
515 sch = qdisc_alloc(dev_queue, ops);
518 sch->parent = parentid;
520 if (!ops->init || ops->init(sch, NULL) == 0)
527 EXPORT_SYMBOL(qdisc_create_dflt);
529 /* Under qdisc_lock(qdisc) and BH! */
531 void qdisc_reset(struct Qdisc *qdisc)
533 const struct Qdisc_ops *ops = qdisc->ops;
538 EXPORT_SYMBOL(qdisc_reset);
540 void qdisc_destroy(struct Qdisc *qdisc)
542 const struct Qdisc_ops *ops = qdisc->ops;
544 if (qdisc->flags & TCQ_F_BUILTIN ||
545 !atomic_dec_and_test(&qdisc->refcnt))
548 #ifdef CONFIG_NET_SCHED
549 qdisc_list_del(qdisc);
551 qdisc_put_stab(qdisc->stab);
553 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
559 module_put(ops->owner);
560 dev_put(qdisc_dev(qdisc));
562 kfree_skb(qdisc->gso_skb);
563 __skb_queue_purge(&qdisc->requeue);
565 kfree((char *) qdisc - qdisc->padded);
567 EXPORT_SYMBOL(qdisc_destroy);
569 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
573 for (i = 0; i < dev->num_tx_queues; i++) {
574 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
576 if (txq->qdisc_sleeping != &noop_qdisc)
582 static void attach_one_default_qdisc(struct net_device *dev,
583 struct netdev_queue *dev_queue,
588 if (dev->tx_queue_len) {
589 qdisc = qdisc_create_dflt(dev, dev_queue,
590 &pfifo_fast_ops, TC_H_ROOT);
592 printk(KERN_INFO "%s: activation failed\n", dev->name);
596 qdisc = &noqueue_qdisc;
598 dev_queue->qdisc_sleeping = qdisc;
601 static void transition_one_qdisc(struct net_device *dev,
602 struct netdev_queue *dev_queue,
603 void *_need_watchdog)
605 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
606 int *need_watchdog_p = _need_watchdog;
608 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
609 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
611 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
612 if (need_watchdog_p && new_qdisc != &noqueue_qdisc)
613 *need_watchdog_p = 1;
616 void dev_activate(struct net_device *dev)
620 /* No queueing discipline is attached to device;
621 create default one i.e. pfifo_fast for devices,
622 which need queueing and noqueue_qdisc for
626 if (dev_all_qdisc_sleeping_noop(dev))
627 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
629 if (!netif_carrier_ok(dev))
630 /* Delay activation until next carrier-on event */
634 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
635 transition_one_qdisc(dev, &dev->rx_queue, NULL);
638 dev->trans_start = jiffies;
639 dev_watchdog_up(dev);
643 static void dev_deactivate_queue(struct net_device *dev,
644 struct netdev_queue *dev_queue,
645 void *_qdisc_default)
647 struct Qdisc *qdisc_default = _qdisc_default;
650 qdisc = dev_queue->qdisc;
652 spin_lock_bh(qdisc_lock(qdisc));
654 if (!(qdisc->flags & TCQ_F_BUILTIN))
655 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
657 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
660 spin_unlock_bh(qdisc_lock(qdisc));
664 static bool some_qdisc_is_busy(struct net_device *dev)
668 for (i = 0; i < dev->num_tx_queues; i++) {
669 struct netdev_queue *dev_queue;
670 spinlock_t *root_lock;
674 dev_queue = netdev_get_tx_queue(dev, i);
675 q = dev_queue->qdisc_sleeping;
676 root_lock = qdisc_lock(q);
678 spin_lock_bh(root_lock);
680 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
681 test_bit(__QDISC_STATE_SCHED, &q->state));
683 spin_unlock_bh(root_lock);
691 void dev_deactivate(struct net_device *dev)
693 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
694 dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
696 dev_watchdog_down(dev);
698 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
701 /* Wait for outstanding qdisc_run calls. */
702 while (some_qdisc_is_busy(dev))
706 static void dev_init_scheduler_queue(struct net_device *dev,
707 struct netdev_queue *dev_queue,
710 struct Qdisc *qdisc = _qdisc;
712 dev_queue->qdisc = qdisc;
713 dev_queue->qdisc_sleeping = qdisc;
716 void dev_init_scheduler(struct net_device *dev)
718 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
719 dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
721 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
724 static void shutdown_scheduler_queue(struct net_device *dev,
725 struct netdev_queue *dev_queue,
726 void *_qdisc_default)
728 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
729 struct Qdisc *qdisc_default = _qdisc_default;
732 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
733 dev_queue->qdisc_sleeping = qdisc_default;
735 qdisc_destroy(qdisc);
739 void dev_shutdown(struct net_device *dev)
741 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
742 shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
743 WARN_ON(timer_pending(&dev->watchdog_timer));