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/config.h>
18 #include <linux/module.h>
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/string.h>
24 #include <linux/socket.h>
25 #include <linux/sockios.h>
27 #include <linux/errno.h>
28 #include <linux/interrupt.h>
29 #include <linux/netdevice.h>
30 #include <linux/skbuff.h>
31 #include <linux/rtnetlink.h>
32 #include <linux/init.h>
33 #include <linux/rcupdate.h>
34 #include <linux/list.h>
36 #include <net/pkt_sched.h>
38 /* Main transmission queue. */
40 /* Main qdisc structure lock.
42 However, modifications
43 to data, participating in scheduling must be additionally
44 protected with dev->queue_lock spinlock.
46 The idea is the following:
47 - enqueue, dequeue are serialized via top level device
48 spinlock dev->queue_lock.
49 - tree walking is protected by read_lock_bh(qdisc_tree_lock)
50 and this lock is used only in process context.
51 - updates to tree are made under rtnl semaphore or
52 from softirq context (__qdisc_destroy rcu-callback)
53 hence this lock needs local bh disabling.
55 qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
57 DEFINE_RWLOCK(qdisc_tree_lock);
59 void qdisc_lock_tree(struct net_device *dev)
61 write_lock_bh(&qdisc_tree_lock);
62 spin_lock_bh(&dev->queue_lock);
65 void qdisc_unlock_tree(struct net_device *dev)
67 spin_unlock_bh(&dev->queue_lock);
68 write_unlock_bh(&qdisc_tree_lock);
72 dev->queue_lock serializes queue accesses for this device
73 AND dev->qdisc pointer itself.
75 netif_tx_lock serializes accesses to device driver.
77 dev->queue_lock and netif_tx_lock are mutually exclusive,
78 if one is grabbed, another must be free.
83 Note, that this procedure can be called by a watchdog timer, so that
84 we do not check dev->tbusy flag here.
86 Returns: 0 - queue is empty.
87 >0 - queue is not empty, but throttled.
88 <0 - queue is not empty. Device is throttled, if dev->tbusy != 0.
90 NOTE: Called under dev->queue_lock with locally disabled BH.
93 static inline int qdisc_restart(struct net_device *dev)
95 struct Qdisc *q = dev->qdisc;
99 if (((skb = dev->gso_skb)) || ((skb = q->dequeue(q)))) {
100 unsigned nolock = (dev->features & NETIF_F_LLTX);
105 * When the driver has LLTX set it does its own locking
106 * in start_xmit. No need to add additional overhead by
107 * locking again. These checks are worth it because
108 * even uncongested locks can be quite expensive.
109 * The driver can do trylock like here too, in case
110 * of lock congestion it should return -1 and the packet
114 if (!netif_tx_trylock(dev)) {
116 /* So, someone grabbed the driver. */
118 /* It may be transient configuration error,
119 when hard_start_xmit() recurses. We detect
120 it by checking xmit owner and drop the
121 packet when deadloop is detected.
123 if (dev->xmit_lock_owner == smp_processor_id()) {
126 printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
129 __get_cpu_var(netdev_rx_stat).cpu_collision++;
135 /* And release queue */
136 spin_unlock(&dev->queue_lock);
138 if (!netif_queue_stopped(dev)) {
141 ret = dev_hard_start_xmit(skb, dev);
142 if (ret == NETDEV_TX_OK) {
144 netif_tx_unlock(dev);
146 spin_lock(&dev->queue_lock);
149 if (ret == NETDEV_TX_LOCKED && nolock) {
150 spin_lock(&dev->queue_lock);
155 /* NETDEV_TX_BUSY - we need to requeue */
156 /* Release the driver */
158 netif_tx_unlock(dev);
160 spin_lock(&dev->queue_lock);
164 /* Device kicked us out :(
165 This is possible in three cases:
168 1. fastroute is enabled
169 2. device cannot determine busy state
170 before start of transmission (f.e. dialout)
171 3. device is buggy (ppp)
178 q->ops->requeue(skb, q);
182 BUG_ON((int) q->q.qlen < 0);
186 void __qdisc_run(struct net_device *dev)
188 if (unlikely(dev->qdisc == &noop_qdisc))
191 while (qdisc_restart(dev) < 0 && !netif_queue_stopped(dev))
195 clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
198 static void dev_watchdog(unsigned long arg)
200 struct net_device *dev = (struct net_device *)arg;
203 if (dev->qdisc != &noop_qdisc) {
204 if (netif_device_present(dev) &&
205 netif_running(dev) &&
206 netif_carrier_ok(dev)) {
207 if (netif_queue_stopped(dev) &&
208 time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
210 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
212 dev->tx_timeout(dev);
214 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
218 netif_tx_unlock(dev);
223 static void dev_watchdog_init(struct net_device *dev)
225 init_timer(&dev->watchdog_timer);
226 dev->watchdog_timer.data = (unsigned long)dev;
227 dev->watchdog_timer.function = dev_watchdog;
230 void __netdev_watchdog_up(struct net_device *dev)
232 if (dev->tx_timeout) {
233 if (dev->watchdog_timeo <= 0)
234 dev->watchdog_timeo = 5*HZ;
235 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
240 static void dev_watchdog_up(struct net_device *dev)
242 netif_tx_lock_bh(dev);
243 __netdev_watchdog_up(dev);
244 netif_tx_unlock_bh(dev);
247 static void dev_watchdog_down(struct net_device *dev)
249 netif_tx_lock_bh(dev);
250 if (del_timer(&dev->watchdog_timer))
252 netif_tx_unlock_bh(dev);
255 void netif_carrier_on(struct net_device *dev)
257 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
258 linkwatch_fire_event(dev);
259 if (netif_running(dev))
260 __netdev_watchdog_up(dev);
263 void netif_carrier_off(struct net_device *dev)
265 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
266 linkwatch_fire_event(dev);
269 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
270 under all circumstances. It is difficult to invent anything faster or
274 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
280 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
285 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
288 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
294 struct Qdisc_ops noop_qdisc_ops = {
297 .enqueue = noop_enqueue,
298 .dequeue = noop_dequeue,
299 .requeue = noop_requeue,
300 .owner = THIS_MODULE,
303 struct Qdisc noop_qdisc = {
304 .enqueue = noop_enqueue,
305 .dequeue = noop_dequeue,
306 .flags = TCQ_F_BUILTIN,
307 .ops = &noop_qdisc_ops,
308 .list = LIST_HEAD_INIT(noop_qdisc.list),
311 static struct Qdisc_ops noqueue_qdisc_ops = {
314 .enqueue = noop_enqueue,
315 .dequeue = noop_dequeue,
316 .requeue = noop_requeue,
317 .owner = THIS_MODULE,
320 static struct Qdisc noqueue_qdisc = {
322 .dequeue = noop_dequeue,
323 .flags = TCQ_F_BUILTIN,
324 .ops = &noqueue_qdisc_ops,
325 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
329 static const u8 prio2band[TC_PRIO_MAX+1] =
330 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
332 /* 3-band FIFO queue: old style, but should be a bit faster than
333 generic prio+fifo combination.
336 #define PFIFO_FAST_BANDS 3
338 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
341 struct sk_buff_head *list = qdisc_priv(qdisc);
342 return list + prio2band[skb->priority & TC_PRIO_MAX];
345 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
347 struct sk_buff_head *list = prio2list(skb, qdisc);
349 if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
351 return __qdisc_enqueue_tail(skb, qdisc, list);
354 return qdisc_drop(skb, qdisc);
357 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
360 struct sk_buff_head *list = qdisc_priv(qdisc);
362 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
363 if (!skb_queue_empty(list + prio)) {
365 return __qdisc_dequeue_head(qdisc, list + prio);
372 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
375 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
378 static void pfifo_fast_reset(struct Qdisc* qdisc)
381 struct sk_buff_head *list = qdisc_priv(qdisc);
383 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
384 __qdisc_reset_queue(qdisc, list + prio);
386 qdisc->qstats.backlog = 0;
390 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
392 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
394 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
395 RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
402 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
405 struct sk_buff_head *list = qdisc_priv(qdisc);
407 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
408 skb_queue_head_init(list + prio);
413 static struct Qdisc_ops pfifo_fast_ops = {
415 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
416 .enqueue = pfifo_fast_enqueue,
417 .dequeue = pfifo_fast_dequeue,
418 .requeue = pfifo_fast_requeue,
419 .init = pfifo_fast_init,
420 .reset = pfifo_fast_reset,
421 .dump = pfifo_fast_dump,
422 .owner = THIS_MODULE,
425 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
432 /* ensure that the Qdisc and the private data are 32-byte aligned */
433 size = QDISC_ALIGN(sizeof(*sch));
434 size += ops->priv_size + (QDISC_ALIGNTO - 1);
436 p = kmalloc(size, GFP_KERNEL);
440 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
441 sch->padded = (char *) sch - (char *) p;
443 INIT_LIST_HEAD(&sch->list);
444 skb_queue_head_init(&sch->q);
446 sch->enqueue = ops->enqueue;
447 sch->dequeue = ops->dequeue;
450 sch->stats_lock = &dev->queue_lock;
451 atomic_set(&sch->refcnt, 1);
455 return ERR_PTR(-err);
458 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
462 sch = qdisc_alloc(dev, ops);
466 if (!ops->init || ops->init(sch, NULL) == 0)
474 /* Under dev->queue_lock and BH! */
476 void qdisc_reset(struct Qdisc *qdisc)
478 struct Qdisc_ops *ops = qdisc->ops;
484 /* this is the rcu callback function to clean up a qdisc when there
485 * are no further references to it */
487 static void __qdisc_destroy(struct rcu_head *head)
489 struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
490 struct Qdisc_ops *ops = qdisc->ops;
492 #ifdef CONFIG_NET_ESTIMATOR
493 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
495 write_lock(&qdisc_tree_lock);
500 write_unlock(&qdisc_tree_lock);
501 module_put(ops->owner);
504 kfree((char *) qdisc - qdisc->padded);
507 /* Under dev->queue_lock and BH! */
509 void qdisc_destroy(struct Qdisc *qdisc)
511 struct list_head cql = LIST_HEAD_INIT(cql);
512 struct Qdisc *cq, *q, *n;
514 if (qdisc->flags & TCQ_F_BUILTIN ||
515 !atomic_dec_and_test(&qdisc->refcnt))
518 if (!list_empty(&qdisc->list)) {
519 if (qdisc->ops->cl_ops == NULL)
520 list_del(&qdisc->list);
522 list_move(&qdisc->list, &cql);
525 /* unlink inner qdiscs from dev->qdisc_list immediately */
526 list_for_each_entry(cq, &cql, list)
527 list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
528 if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
529 if (q->ops->cl_ops == NULL)
530 list_del_init(&q->list);
532 list_move_tail(&q->list, &cql);
534 list_for_each_entry_safe(cq, n, &cql, list)
535 list_del_init(&cq->list);
537 call_rcu(&qdisc->q_rcu, __qdisc_destroy);
540 void dev_activate(struct net_device *dev)
542 /* No queueing discipline is attached to device;
543 create default one i.e. pfifo_fast for devices,
544 which need queueing and noqueue_qdisc for
548 if (dev->qdisc_sleeping == &noop_qdisc) {
550 if (dev->tx_queue_len) {
551 qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
553 printk(KERN_INFO "%s: activation failed\n", dev->name);
556 write_lock_bh(&qdisc_tree_lock);
557 list_add_tail(&qdisc->list, &dev->qdisc_list);
558 write_unlock_bh(&qdisc_tree_lock);
560 qdisc = &noqueue_qdisc;
562 write_lock_bh(&qdisc_tree_lock);
563 dev->qdisc_sleeping = qdisc;
564 write_unlock_bh(&qdisc_tree_lock);
567 if (!netif_carrier_ok(dev))
568 /* Delay activation until next carrier-on event */
571 spin_lock_bh(&dev->queue_lock);
572 rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
573 if (dev->qdisc != &noqueue_qdisc) {
574 dev->trans_start = jiffies;
575 dev_watchdog_up(dev);
577 spin_unlock_bh(&dev->queue_lock);
580 void dev_deactivate(struct net_device *dev)
584 spin_lock_bh(&dev->queue_lock);
586 dev->qdisc = &noop_qdisc;
590 spin_unlock_bh(&dev->queue_lock);
592 dev_watchdog_down(dev);
594 /* Wait for outstanding dev_queue_xmit calls. */
597 /* Wait for outstanding qdisc_run calls. */
598 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
602 kfree_skb(dev->gso_skb);
607 void dev_init_scheduler(struct net_device *dev)
609 qdisc_lock_tree(dev);
610 dev->qdisc = &noop_qdisc;
611 dev->qdisc_sleeping = &noop_qdisc;
612 INIT_LIST_HEAD(&dev->qdisc_list);
613 qdisc_unlock_tree(dev);
615 dev_watchdog_init(dev);
618 void dev_shutdown(struct net_device *dev)
622 qdisc_lock_tree(dev);
623 qdisc = dev->qdisc_sleeping;
624 dev->qdisc = &noop_qdisc;
625 dev->qdisc_sleeping = &noop_qdisc;
626 qdisc_destroy(qdisc);
627 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
628 if ((qdisc = dev->qdisc_ingress) != NULL) {
629 dev->qdisc_ingress = NULL;
630 qdisc_destroy(qdisc);
633 BUG_TRAP(!timer_pending(&dev->watchdog_timer));
634 qdisc_unlock_tree(dev);
637 EXPORT_SYMBOL(__netdev_watchdog_up);
638 EXPORT_SYMBOL(netif_carrier_on);
639 EXPORT_SYMBOL(netif_carrier_off);
640 EXPORT_SYMBOL(noop_qdisc);
641 EXPORT_SYMBOL(noop_qdisc_ops);
642 EXPORT_SYMBOL(qdisc_create_dflt);
643 EXPORT_SYMBOL(qdisc_alloc);
644 EXPORT_SYMBOL(qdisc_destroy);
645 EXPORT_SYMBOL(qdisc_reset);
646 EXPORT_SYMBOL(qdisc_lock_tree);
647 EXPORT_SYMBOL(qdisc_unlock_tree);