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)
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) && !netif_tx_queue_frozen(txq))
74 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
75 struct netdev_queue *dev_queue,
80 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
82 * Same CPU holding the lock. It may be a transient
83 * configuration error, when hard_start_xmit() recurses. We
84 * detect it by checking xmit owner and drop the packet when
85 * deadloop is detected. Return OK to try the next skb.
89 printk(KERN_WARNING "Dead loop on netdevice %s, "
90 "fix it urgently!\n", dev_queue->dev->name);
94 * Another cpu is holding lock, requeue & delay xmits for
97 __get_cpu_var(netdev_rx_stat).cpu_collision++;
98 ret = dev_requeue_skb(skb, q);
105 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
107 * __QDISC_STATE_RUNNING guarantees only one CPU can process
108 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
111 * netif_tx_lock serializes accesses to device driver.
113 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
114 * if one is grabbed, another must be free.
116 * Note, that this procedure can be called by a watchdog timer
118 * Returns to the caller:
119 * 0 - queue is empty or throttled.
120 * >0 - queue is not empty.
123 static inline int qdisc_restart(struct Qdisc *q)
125 struct netdev_queue *txq;
126 int ret = NETDEV_TX_BUSY;
127 struct net_device *dev;
128 spinlock_t *root_lock;
132 if (unlikely((skb = dequeue_skb(q)) == NULL))
135 root_lock = qdisc_lock(q);
137 /* And release qdisc */
138 spin_unlock(root_lock);
141 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
143 HARD_TX_LOCK(dev, txq, smp_processor_id());
144 if (!netif_tx_queue_stopped(txq) &&
145 !netif_tx_queue_frozen(txq))
146 ret = dev_hard_start_xmit(skb, dev, txq);
147 HARD_TX_UNLOCK(dev, txq);
149 spin_lock(root_lock);
153 /* Driver sent out skb successfully */
157 case NETDEV_TX_LOCKED:
158 /* Driver try lock failed */
159 ret = handle_dev_cpu_collision(skb, txq, q);
163 /* Driver returned NETDEV_TX_BUSY - requeue skb */
164 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
165 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
166 dev->name, ret, q->q.qlen);
168 ret = dev_requeue_skb(skb, q);
172 if (ret && (netif_tx_queue_stopped(txq) ||
173 netif_tx_queue_frozen(txq)))
179 void __qdisc_run(struct Qdisc *q)
181 unsigned long start_time = jiffies;
183 while (qdisc_restart(q)) {
185 * Postpone processing if
186 * 1. another process needs the CPU;
187 * 2. we've been doing it for too long.
189 if (need_resched() || jiffies != start_time) {
195 clear_bit(__QDISC_STATE_RUNNING, &q->state);
198 static void dev_watchdog(unsigned long arg)
200 struct net_device *dev = (struct net_device *)arg;
203 if (!qdisc_tx_is_noop(dev)) {
204 if (netif_device_present(dev) &&
205 netif_running(dev) &&
206 netif_carrier_ok(dev)) {
207 int some_queue_stopped = 0;
210 for (i = 0; i < dev->num_tx_queues; i++) {
211 struct netdev_queue *txq;
213 txq = netdev_get_tx_queue(dev, i);
214 if (netif_tx_queue_stopped(txq)) {
215 some_queue_stopped = 1;
220 if (some_queue_stopped &&
221 time_after(jiffies, (dev->trans_start +
222 dev->watchdog_timeo))) {
224 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
225 dev->name, netdev_drivername(dev, drivername, 64));
226 dev->tx_timeout(dev);
228 if (!mod_timer(&dev->watchdog_timer,
229 round_jiffies(jiffies +
230 dev->watchdog_timeo)))
234 netif_tx_unlock(dev);
239 void __netdev_watchdog_up(struct net_device *dev)
241 if (dev->tx_timeout) {
242 if (dev->watchdog_timeo <= 0)
243 dev->watchdog_timeo = 5*HZ;
244 if (!mod_timer(&dev->watchdog_timer,
245 round_jiffies(jiffies + dev->watchdog_timeo)))
250 static void dev_watchdog_up(struct net_device *dev)
252 __netdev_watchdog_up(dev);
255 static void dev_watchdog_down(struct net_device *dev)
257 netif_tx_lock_bh(dev);
258 if (del_timer(&dev->watchdog_timer))
260 netif_tx_unlock_bh(dev);
264 * netif_carrier_on - set carrier
265 * @dev: network device
267 * Device has detected that carrier.
269 void netif_carrier_on(struct net_device *dev)
271 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
272 linkwatch_fire_event(dev);
273 if (netif_running(dev))
274 __netdev_watchdog_up(dev);
277 EXPORT_SYMBOL(netif_carrier_on);
280 * netif_carrier_off - clear carrier
281 * @dev: network device
283 * Device has detected loss of carrier.
285 void netif_carrier_off(struct net_device *dev)
287 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
288 linkwatch_fire_event(dev);
290 EXPORT_SYMBOL(netif_carrier_off);
292 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
293 under all circumstances. It is difficult to invent anything faster or
297 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
303 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
308 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
311 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
317 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
320 .enqueue = noop_enqueue,
321 .dequeue = noop_dequeue,
322 .requeue = noop_requeue,
323 .owner = THIS_MODULE,
326 static struct netdev_queue noop_netdev_queue = {
327 .qdisc = &noop_qdisc,
330 struct Qdisc noop_qdisc = {
331 .enqueue = noop_enqueue,
332 .dequeue = noop_dequeue,
333 .flags = TCQ_F_BUILTIN,
334 .ops = &noop_qdisc_ops,
335 .list = LIST_HEAD_INIT(noop_qdisc.list),
336 .requeue.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
337 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
338 .dev_queue = &noop_netdev_queue,
340 EXPORT_SYMBOL(noop_qdisc);
342 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
345 .enqueue = noop_enqueue,
346 .dequeue = noop_dequeue,
347 .requeue = noop_requeue,
348 .owner = THIS_MODULE,
351 static struct Qdisc noqueue_qdisc;
352 static struct netdev_queue noqueue_netdev_queue = {
353 .qdisc = &noqueue_qdisc,
356 static struct Qdisc noqueue_qdisc = {
358 .dequeue = noop_dequeue,
359 .flags = TCQ_F_BUILTIN,
360 .ops = &noqueue_qdisc_ops,
361 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
362 .requeue.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
363 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
364 .dev_queue = &noqueue_netdev_queue,
368 static const u8 prio2band[TC_PRIO_MAX+1] =
369 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
371 /* 3-band FIFO queue: old style, but should be a bit faster than
372 generic prio+fifo combination.
375 #define PFIFO_FAST_BANDS 3
377 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
380 struct sk_buff_head *list = qdisc_priv(qdisc);
381 return list + prio2band[skb->priority & TC_PRIO_MAX];
384 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
386 struct sk_buff_head *list = prio2list(skb, qdisc);
388 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
390 return __qdisc_enqueue_tail(skb, qdisc, list);
393 return qdisc_drop(skb, qdisc);
396 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
399 struct sk_buff_head *list = qdisc_priv(qdisc);
401 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
402 if (!skb_queue_empty(list + prio)) {
404 return __qdisc_dequeue_head(qdisc, list + prio);
411 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
414 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
417 static void pfifo_fast_reset(struct Qdisc* qdisc)
420 struct sk_buff_head *list = qdisc_priv(qdisc);
422 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
423 __qdisc_reset_queue(qdisc, list + prio);
425 qdisc->qstats.backlog = 0;
429 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
431 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
433 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
434 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
441 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
444 struct sk_buff_head *list = qdisc_priv(qdisc);
446 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
447 skb_queue_head_init(list + prio);
452 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
454 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
455 .enqueue = pfifo_fast_enqueue,
456 .dequeue = pfifo_fast_dequeue,
457 .requeue = pfifo_fast_requeue,
458 .init = pfifo_fast_init,
459 .reset = pfifo_fast_reset,
460 .dump = pfifo_fast_dump,
461 .owner = THIS_MODULE,
464 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
465 struct Qdisc_ops *ops)
472 /* ensure that the Qdisc and the private data are 32-byte aligned */
473 size = QDISC_ALIGN(sizeof(*sch));
474 size += ops->priv_size + (QDISC_ALIGNTO - 1);
476 p = kzalloc(size, GFP_KERNEL);
479 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
480 sch->padded = (char *) sch - (char *) p;
482 INIT_LIST_HEAD(&sch->list);
483 skb_queue_head_init(&sch->requeue);
484 skb_queue_head_init(&sch->q);
486 sch->enqueue = ops->enqueue;
487 sch->dequeue = ops->dequeue;
488 sch->dev_queue = dev_queue;
489 dev_hold(qdisc_dev(sch));
490 atomic_set(&sch->refcnt, 1);
497 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
498 struct netdev_queue *dev_queue,
499 struct Qdisc_ops *ops,
500 unsigned int parentid)
504 sch = qdisc_alloc(dev_queue, ops);
507 sch->parent = parentid;
509 if (!ops->init || ops->init(sch, NULL) == 0)
516 EXPORT_SYMBOL(qdisc_create_dflt);
518 /* Under qdisc_lock(qdisc) and BH! */
520 void qdisc_reset(struct Qdisc *qdisc)
522 const struct Qdisc_ops *ops = qdisc->ops;
527 EXPORT_SYMBOL(qdisc_reset);
529 void qdisc_destroy(struct Qdisc *qdisc)
531 const struct Qdisc_ops *ops = qdisc->ops;
533 if (qdisc->flags & TCQ_F_BUILTIN ||
534 !atomic_dec_and_test(&qdisc->refcnt))
537 #ifdef CONFIG_NET_SCHED
538 qdisc_list_del(qdisc);
540 qdisc_put_stab(qdisc->stab);
542 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
548 module_put(ops->owner);
549 dev_put(qdisc_dev(qdisc));
551 kfree_skb(qdisc->gso_skb);
552 __skb_queue_purge(&qdisc->requeue);
554 kfree((char *) qdisc - qdisc->padded);
556 EXPORT_SYMBOL(qdisc_destroy);
558 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
562 for (i = 0; i < dev->num_tx_queues; i++) {
563 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
565 if (txq->qdisc_sleeping != &noop_qdisc)
571 static void attach_one_default_qdisc(struct net_device *dev,
572 struct netdev_queue *dev_queue,
577 if (dev->tx_queue_len) {
578 qdisc = qdisc_create_dflt(dev, dev_queue,
579 &pfifo_fast_ops, TC_H_ROOT);
581 printk(KERN_INFO "%s: activation failed\n", dev->name);
585 qdisc = &noqueue_qdisc;
587 dev_queue->qdisc_sleeping = qdisc;
590 static void transition_one_qdisc(struct net_device *dev,
591 struct netdev_queue *dev_queue,
592 void *_need_watchdog)
594 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
595 int *need_watchdog_p = _need_watchdog;
597 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
598 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
600 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
601 if (need_watchdog_p && new_qdisc != &noqueue_qdisc)
602 *need_watchdog_p = 1;
605 void dev_activate(struct net_device *dev)
609 /* No queueing discipline is attached to device;
610 create default one i.e. pfifo_fast for devices,
611 which need queueing and noqueue_qdisc for
615 if (dev_all_qdisc_sleeping_noop(dev))
616 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
618 if (!netif_carrier_ok(dev))
619 /* Delay activation until next carrier-on event */
623 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
624 transition_one_qdisc(dev, &dev->rx_queue, NULL);
627 dev->trans_start = jiffies;
628 dev_watchdog_up(dev);
632 static void dev_deactivate_queue(struct net_device *dev,
633 struct netdev_queue *dev_queue,
634 void *_qdisc_default)
636 struct Qdisc *qdisc_default = _qdisc_default;
639 qdisc = dev_queue->qdisc;
641 spin_lock_bh(qdisc_lock(qdisc));
643 if (!(qdisc->flags & TCQ_F_BUILTIN))
644 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
646 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
649 spin_unlock_bh(qdisc_lock(qdisc));
653 static bool some_qdisc_is_busy(struct net_device *dev)
657 for (i = 0; i < dev->num_tx_queues; i++) {
658 struct netdev_queue *dev_queue;
659 spinlock_t *root_lock;
663 dev_queue = netdev_get_tx_queue(dev, i);
664 q = dev_queue->qdisc_sleeping;
665 root_lock = qdisc_lock(q);
667 spin_lock_bh(root_lock);
669 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
670 test_bit(__QDISC_STATE_SCHED, &q->state));
672 spin_unlock_bh(root_lock);
680 void dev_deactivate(struct net_device *dev)
682 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
683 dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
685 dev_watchdog_down(dev);
687 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
690 /* Wait for outstanding qdisc_run calls. */
691 while (some_qdisc_is_busy(dev))
695 static void dev_init_scheduler_queue(struct net_device *dev,
696 struct netdev_queue *dev_queue,
699 struct Qdisc *qdisc = _qdisc;
701 dev_queue->qdisc = qdisc;
702 dev_queue->qdisc_sleeping = qdisc;
705 void dev_init_scheduler(struct net_device *dev)
707 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
708 dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
710 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
713 static void shutdown_scheduler_queue(struct net_device *dev,
714 struct netdev_queue *dev_queue,
715 void *_qdisc_default)
717 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
718 struct Qdisc *qdisc_default = _qdisc_default;
721 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
722 dev_queue->qdisc_sleeping = qdisc_default;
724 qdisc_destroy(qdisc);
728 void dev_shutdown(struct net_device *dev)
730 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
731 shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
732 WARN_ON(timer_pending(&dev->watchdog_timer));