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 <asm/uaccess.h>
15 #include <asm/system.h>
16 #include <linux/bitops.h>
17 #include <linux/module.h>
18 #include <linux/types.h>
19 #include <linux/kernel.h>
20 #include <linux/sched.h>
21 #include <linux/string.h>
23 #include <linux/socket.h>
24 #include <linux/sockios.h>
26 #include <linux/errno.h>
27 #include <linux/interrupt.h>
28 #include <linux/netdevice.h>
29 #include <linux/skbuff.h>
30 #include <linux/rtnetlink.h>
31 #include <linux/init.h>
32 #include <linux/rcupdate.h>
33 #include <linux/list.h>
35 #include <net/pkt_sched.h>
37 /* Main transmission queue. */
39 /* Main qdisc structure lock.
41 However, modifications
42 to data, participating in scheduling must be additionally
43 protected with dev->queue_lock spinlock.
45 The idea is the following:
46 - enqueue, dequeue are serialized via top level device
47 spinlock dev->queue_lock.
48 - tree walking is protected by read_lock_bh(qdisc_tree_lock)
49 and this lock is used only in process context.
50 - updates to tree are made under rtnl semaphore or
51 from softirq context (__qdisc_destroy rcu-callback)
52 hence this lock needs local bh disabling.
54 qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
56 DEFINE_RWLOCK(qdisc_tree_lock);
58 void qdisc_lock_tree(struct net_device *dev)
60 write_lock_bh(&qdisc_tree_lock);
61 spin_lock_bh(&dev->queue_lock);
64 void qdisc_unlock_tree(struct net_device *dev)
66 spin_unlock_bh(&dev->queue_lock);
67 write_unlock_bh(&qdisc_tree_lock);
71 dev->queue_lock serializes queue accesses for this device
72 AND dev->qdisc pointer itself.
74 netif_tx_lock serializes accesses to device driver.
76 dev->queue_lock and netif_tx_lock are mutually exclusive,
77 if one is grabbed, another must be free.
82 Note, that this procedure can be called by a watchdog timer, so that
83 we do not check dev->tbusy flag here.
85 Returns: 0 - queue is empty.
86 >0 - queue is not empty, but throttled.
87 <0 - queue is not empty. Device is throttled, if dev->tbusy != 0.
89 NOTE: Called under dev->queue_lock with locally disabled BH.
92 static inline int qdisc_restart(struct net_device *dev)
94 struct Qdisc *q = dev->qdisc;
98 if (((skb = dev->gso_skb)) || ((skb = q->dequeue(q)))) {
99 unsigned nolock = (dev->features & NETIF_F_LLTX);
104 * When the driver has LLTX set it does its own locking
105 * in start_xmit. No need to add additional overhead by
106 * locking again. These checks are worth it because
107 * even uncongested locks can be quite expensive.
108 * The driver can do trylock like here too, in case
109 * of lock congestion it should return -1 and the packet
113 if (!netif_tx_trylock(dev)) {
115 /* So, someone grabbed the driver. */
117 /* It may be transient configuration error,
118 when hard_start_xmit() recurses. We detect
119 it by checking xmit owner and drop the
120 packet when deadloop is detected.
122 if (dev->xmit_lock_owner == smp_processor_id()) {
125 printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
128 __get_cpu_var(netdev_rx_stat).cpu_collision++;
134 /* And release queue */
135 spin_unlock(&dev->queue_lock);
137 if (!netif_queue_stopped(dev)) {
140 ret = dev_hard_start_xmit(skb, dev);
141 if (ret == NETDEV_TX_OK) {
143 netif_tx_unlock(dev);
145 spin_lock(&dev->queue_lock);
148 if (ret == NETDEV_TX_LOCKED && nolock) {
149 spin_lock(&dev->queue_lock);
154 /* NETDEV_TX_BUSY - we need to requeue */
155 /* Release the driver */
157 netif_tx_unlock(dev);
159 spin_lock(&dev->queue_lock);
163 /* Device kicked us out :(
164 This is possible in three cases:
167 1. fastroute is enabled
168 2. device cannot determine busy state
169 before start of transmission (f.e. dialout)
170 3. device is buggy (ppp)
177 q->ops->requeue(skb, q);
181 BUG_ON((int) q->q.qlen < 0);
185 void __qdisc_run(struct net_device *dev)
187 if (unlikely(dev->qdisc == &noop_qdisc))
190 while (qdisc_restart(dev) < 0 && !netif_queue_stopped(dev))
194 clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
197 static void dev_watchdog(unsigned long arg)
199 struct net_device *dev = (struct net_device *)arg;
202 if (dev->qdisc != &noop_qdisc) {
203 if (netif_device_present(dev) &&
204 netif_running(dev) &&
205 netif_carrier_ok(dev)) {
206 if (netif_queue_stopped(dev) &&
207 time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
209 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
211 dev->tx_timeout(dev);
213 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
217 netif_tx_unlock(dev);
222 static void dev_watchdog_init(struct net_device *dev)
224 init_timer(&dev->watchdog_timer);
225 dev->watchdog_timer.data = (unsigned long)dev;
226 dev->watchdog_timer.function = dev_watchdog;
229 void __netdev_watchdog_up(struct net_device *dev)
231 if (dev->tx_timeout) {
232 if (dev->watchdog_timeo <= 0)
233 dev->watchdog_timeo = 5*HZ;
234 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
239 static void dev_watchdog_up(struct net_device *dev)
241 __netdev_watchdog_up(dev);
244 static void dev_watchdog_down(struct net_device *dev)
246 netif_tx_lock_bh(dev);
247 if (del_timer(&dev->watchdog_timer))
249 netif_tx_unlock_bh(dev);
252 void netif_carrier_on(struct net_device *dev)
254 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
255 linkwatch_fire_event(dev);
256 if (netif_running(dev))
257 __netdev_watchdog_up(dev);
260 void netif_carrier_off(struct net_device *dev)
262 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
263 linkwatch_fire_event(dev);
266 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
267 under all circumstances. It is difficult to invent anything faster or
271 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
277 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
282 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
285 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
291 struct Qdisc_ops noop_qdisc_ops = {
294 .enqueue = noop_enqueue,
295 .dequeue = noop_dequeue,
296 .requeue = noop_requeue,
297 .owner = THIS_MODULE,
300 struct Qdisc noop_qdisc = {
301 .enqueue = noop_enqueue,
302 .dequeue = noop_dequeue,
303 .flags = TCQ_F_BUILTIN,
304 .ops = &noop_qdisc_ops,
305 .list = LIST_HEAD_INIT(noop_qdisc.list),
308 static struct Qdisc_ops noqueue_qdisc_ops = {
311 .enqueue = noop_enqueue,
312 .dequeue = noop_dequeue,
313 .requeue = noop_requeue,
314 .owner = THIS_MODULE,
317 static struct Qdisc noqueue_qdisc = {
319 .dequeue = noop_dequeue,
320 .flags = TCQ_F_BUILTIN,
321 .ops = &noqueue_qdisc_ops,
322 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
326 static const u8 prio2band[TC_PRIO_MAX+1] =
327 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
329 /* 3-band FIFO queue: old style, but should be a bit faster than
330 generic prio+fifo combination.
333 #define PFIFO_FAST_BANDS 3
335 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
338 struct sk_buff_head *list = qdisc_priv(qdisc);
339 return list + prio2band[skb->priority & TC_PRIO_MAX];
342 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
344 struct sk_buff_head *list = prio2list(skb, qdisc);
346 if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
348 return __qdisc_enqueue_tail(skb, qdisc, list);
351 return qdisc_drop(skb, qdisc);
354 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
357 struct sk_buff_head *list = qdisc_priv(qdisc);
359 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
360 if (!skb_queue_empty(list + prio)) {
362 return __qdisc_dequeue_head(qdisc, list + prio);
369 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
372 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
375 static void pfifo_fast_reset(struct Qdisc* qdisc)
378 struct sk_buff_head *list = qdisc_priv(qdisc);
380 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
381 __qdisc_reset_queue(qdisc, list + prio);
383 qdisc->qstats.backlog = 0;
387 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
389 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
391 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
392 RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
399 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
402 struct sk_buff_head *list = qdisc_priv(qdisc);
404 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
405 skb_queue_head_init(list + prio);
410 static struct Qdisc_ops pfifo_fast_ops = {
412 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
413 .enqueue = pfifo_fast_enqueue,
414 .dequeue = pfifo_fast_dequeue,
415 .requeue = pfifo_fast_requeue,
416 .init = pfifo_fast_init,
417 .reset = pfifo_fast_reset,
418 .dump = pfifo_fast_dump,
419 .owner = THIS_MODULE,
422 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
429 /* ensure that the Qdisc and the private data are 32-byte aligned */
430 size = QDISC_ALIGN(sizeof(*sch));
431 size += ops->priv_size + (QDISC_ALIGNTO - 1);
433 p = kzalloc(size, GFP_KERNEL);
436 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
437 sch->padded = (char *) sch - (char *) p;
439 INIT_LIST_HEAD(&sch->list);
440 skb_queue_head_init(&sch->q);
442 sch->enqueue = ops->enqueue;
443 sch->dequeue = ops->dequeue;
446 sch->stats_lock = &dev->queue_lock;
447 atomic_set(&sch->refcnt, 1);
451 return ERR_PTR(-err);
454 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
458 sch = qdisc_alloc(dev, ops);
462 if (!ops->init || ops->init(sch, NULL) == 0)
470 /* Under dev->queue_lock and BH! */
472 void qdisc_reset(struct Qdisc *qdisc)
474 struct Qdisc_ops *ops = qdisc->ops;
480 /* this is the rcu callback function to clean up a qdisc when there
481 * are no further references to it */
483 static void __qdisc_destroy(struct rcu_head *head)
485 struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
486 struct Qdisc_ops *ops = qdisc->ops;
488 #ifdef CONFIG_NET_ESTIMATOR
489 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
491 write_lock(&qdisc_tree_lock);
496 write_unlock(&qdisc_tree_lock);
497 module_put(ops->owner);
500 kfree((char *) qdisc - qdisc->padded);
503 /* Under dev->queue_lock and BH! */
505 void qdisc_destroy(struct Qdisc *qdisc)
507 struct list_head cql = LIST_HEAD_INIT(cql);
508 struct Qdisc *cq, *q, *n;
510 if (qdisc->flags & TCQ_F_BUILTIN ||
511 !atomic_dec_and_test(&qdisc->refcnt))
514 if (!list_empty(&qdisc->list)) {
515 if (qdisc->ops->cl_ops == NULL)
516 list_del(&qdisc->list);
518 list_move(&qdisc->list, &cql);
521 /* unlink inner qdiscs from dev->qdisc_list immediately */
522 list_for_each_entry(cq, &cql, list)
523 list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
524 if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
525 if (q->ops->cl_ops == NULL)
526 list_del_init(&q->list);
528 list_move_tail(&q->list, &cql);
530 list_for_each_entry_safe(cq, n, &cql, list)
531 list_del_init(&cq->list);
533 call_rcu(&qdisc->q_rcu, __qdisc_destroy);
536 void dev_activate(struct net_device *dev)
538 /* No queueing discipline is attached to device;
539 create default one i.e. pfifo_fast for devices,
540 which need queueing and noqueue_qdisc for
544 if (dev->qdisc_sleeping == &noop_qdisc) {
546 if (dev->tx_queue_len) {
547 qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
549 printk(KERN_INFO "%s: activation failed\n", dev->name);
552 write_lock_bh(&qdisc_tree_lock);
553 list_add_tail(&qdisc->list, &dev->qdisc_list);
554 write_unlock_bh(&qdisc_tree_lock);
556 qdisc = &noqueue_qdisc;
558 write_lock_bh(&qdisc_tree_lock);
559 dev->qdisc_sleeping = qdisc;
560 write_unlock_bh(&qdisc_tree_lock);
563 if (!netif_carrier_ok(dev))
564 /* Delay activation until next carrier-on event */
567 spin_lock_bh(&dev->queue_lock);
568 rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
569 if (dev->qdisc != &noqueue_qdisc) {
570 dev->trans_start = jiffies;
571 dev_watchdog_up(dev);
573 spin_unlock_bh(&dev->queue_lock);
576 void dev_deactivate(struct net_device *dev)
580 spin_lock_bh(&dev->queue_lock);
582 dev->qdisc = &noop_qdisc;
586 spin_unlock_bh(&dev->queue_lock);
588 dev_watchdog_down(dev);
590 /* Wait for outstanding dev_queue_xmit calls. */
593 /* Wait for outstanding qdisc_run calls. */
594 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
598 kfree_skb(dev->gso_skb);
603 void dev_init_scheduler(struct net_device *dev)
605 qdisc_lock_tree(dev);
606 dev->qdisc = &noop_qdisc;
607 dev->qdisc_sleeping = &noop_qdisc;
608 INIT_LIST_HEAD(&dev->qdisc_list);
609 qdisc_unlock_tree(dev);
611 dev_watchdog_init(dev);
614 void dev_shutdown(struct net_device *dev)
618 qdisc_lock_tree(dev);
619 qdisc = dev->qdisc_sleeping;
620 dev->qdisc = &noop_qdisc;
621 dev->qdisc_sleeping = &noop_qdisc;
622 qdisc_destroy(qdisc);
623 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
624 if ((qdisc = dev->qdisc_ingress) != NULL) {
625 dev->qdisc_ingress = NULL;
626 qdisc_destroy(qdisc);
629 BUG_TRAP(!timer_pending(&dev->watchdog_timer));
630 qdisc_unlock_tree(dev);
633 EXPORT_SYMBOL(__netdev_watchdog_up);
634 EXPORT_SYMBOL(netif_carrier_on);
635 EXPORT_SYMBOL(netif_carrier_off);
636 EXPORT_SYMBOL(noop_qdisc);
637 EXPORT_SYMBOL(noop_qdisc_ops);
638 EXPORT_SYMBOL(qdisc_create_dflt);
639 EXPORT_SYMBOL(qdisc_alloc);
640 EXPORT_SYMBOL(qdisc_destroy);
641 EXPORT_SYMBOL(qdisc_reset);
642 EXPORT_SYMBOL(qdisc_lock_tree);
643 EXPORT_SYMBOL(qdisc_unlock_tree);