Ath5k: unify resets
[linux-2.6/mini2440.git] / net / sched / sch_generic.c
blobc3ed4d44fc14fe8c5db80ea95af9546deb5da407
1 /*
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
11 * - Ingress support
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)
42 return q->q.qlen;
45 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
47 if (unlikely(skb->next))
48 q->gso_skb = skb;
49 else
50 q->ops->requeue(skb, q);
52 __netif_schedule(q);
53 return 0;
56 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
58 struct sk_buff *skb;
60 if ((skb = q->gso_skb))
61 q->gso_skb = NULL;
62 else
63 skb = q->dequeue(q);
65 return skb;
68 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
69 struct netdev_queue *dev_queue,
70 struct Qdisc *q)
72 int ret;
74 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
76 * Same CPU holding the lock. It may be a transient
77 * configuration error, when hard_start_xmit() recurses. We
78 * detect it by checking xmit owner and drop the packet when
79 * deadloop is detected. Return OK to try the next skb.
81 kfree_skb(skb);
82 if (net_ratelimit())
83 printk(KERN_WARNING "Dead loop on netdevice %s, "
84 "fix it urgently!\n", dev_queue->dev->name);
85 ret = qdisc_qlen(q);
86 } else {
88 * Another cpu is holding lock, requeue & delay xmits for
89 * some time.
91 __get_cpu_var(netdev_rx_stat).cpu_collision++;
92 ret = dev_requeue_skb(skb, q);
95 return ret;
99 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
101 * __QDISC_STATE_RUNNING guarantees only one CPU can process
102 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
103 * this queue.
105 * netif_tx_lock serializes accesses to device driver.
107 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
108 * if one is grabbed, another must be free.
110 * Note, that this procedure can be called by a watchdog timer
112 * Returns to the caller:
113 * 0 - queue is empty or throttled.
114 * >0 - queue is not empty.
117 static inline int qdisc_restart(struct Qdisc *q)
119 struct netdev_queue *txq;
120 int ret = NETDEV_TX_BUSY;
121 struct net_device *dev;
122 spinlock_t *root_lock;
123 struct sk_buff *skb;
125 /* Dequeue packet */
126 if (unlikely((skb = dequeue_skb(q)) == NULL))
127 return 0;
129 root_lock = qdisc_lock(q);
131 /* And release qdisc */
132 spin_unlock(root_lock);
134 dev = qdisc_dev(q);
135 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
137 HARD_TX_LOCK(dev, txq, smp_processor_id());
138 if (!netif_tx_queue_stopped(txq) &&
139 !netif_tx_queue_frozen(txq))
140 ret = dev_hard_start_xmit(skb, dev, txq);
141 HARD_TX_UNLOCK(dev, txq);
143 spin_lock(root_lock);
145 switch (ret) {
146 case NETDEV_TX_OK:
147 /* Driver sent out skb successfully */
148 ret = qdisc_qlen(q);
149 break;
151 case NETDEV_TX_LOCKED:
152 /* Driver try lock failed */
153 ret = handle_dev_cpu_collision(skb, txq, q);
154 break;
156 default:
157 /* Driver returned NETDEV_TX_BUSY - requeue skb */
158 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
159 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
160 dev->name, ret, q->q.qlen);
162 ret = dev_requeue_skb(skb, q);
163 break;
166 if (ret && (netif_tx_queue_stopped(txq) ||
167 netif_tx_queue_frozen(txq)))
168 ret = 0;
170 return ret;
173 void __qdisc_run(struct Qdisc *q)
175 unsigned long start_time = jiffies;
177 while (qdisc_restart(q)) {
179 * Postpone processing if
180 * 1. another process needs the CPU;
181 * 2. we've been doing it for too long.
183 if (need_resched() || jiffies != start_time) {
184 __netif_schedule(q);
185 break;
189 clear_bit(__QDISC_STATE_RUNNING, &q->state);
192 static void dev_watchdog(unsigned long arg)
194 struct net_device *dev = (struct net_device *)arg;
196 netif_tx_lock(dev);
197 if (!qdisc_tx_is_noop(dev)) {
198 if (netif_device_present(dev) &&
199 netif_running(dev) &&
200 netif_carrier_ok(dev)) {
201 int some_queue_stopped = 0;
202 unsigned int i;
204 for (i = 0; i < dev->num_tx_queues; i++) {
205 struct netdev_queue *txq;
207 txq = netdev_get_tx_queue(dev, i);
208 if (netif_tx_queue_stopped(txq)) {
209 some_queue_stopped = 1;
210 break;
214 if (some_queue_stopped &&
215 time_after(jiffies, (dev->trans_start +
216 dev->watchdog_timeo))) {
217 char drivername[64];
218 printk(KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
219 dev->name, netdev_drivername(dev, drivername, 64));
220 dev->tx_timeout(dev);
221 WARN_ON_ONCE(1);
223 if (!mod_timer(&dev->watchdog_timer,
224 round_jiffies(jiffies +
225 dev->watchdog_timeo)))
226 dev_hold(dev);
229 netif_tx_unlock(dev);
231 dev_put(dev);
234 void __netdev_watchdog_up(struct net_device *dev)
236 if (dev->tx_timeout) {
237 if (dev->watchdog_timeo <= 0)
238 dev->watchdog_timeo = 5*HZ;
239 if (!mod_timer(&dev->watchdog_timer,
240 round_jiffies(jiffies + dev->watchdog_timeo)))
241 dev_hold(dev);
245 static void dev_watchdog_up(struct net_device *dev)
247 __netdev_watchdog_up(dev);
250 static void dev_watchdog_down(struct net_device *dev)
252 netif_tx_lock_bh(dev);
253 if (del_timer(&dev->watchdog_timer))
254 dev_put(dev);
255 netif_tx_unlock_bh(dev);
259 * netif_carrier_on - set carrier
260 * @dev: network device
262 * Device has detected that carrier.
264 void netif_carrier_on(struct net_device *dev)
266 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
267 linkwatch_fire_event(dev);
268 if (netif_running(dev))
269 __netdev_watchdog_up(dev);
272 EXPORT_SYMBOL(netif_carrier_on);
275 * netif_carrier_off - clear carrier
276 * @dev: network device
278 * Device has detected loss of carrier.
280 void netif_carrier_off(struct net_device *dev)
282 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
283 linkwatch_fire_event(dev);
285 EXPORT_SYMBOL(netif_carrier_off);
287 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
288 under all circumstances. It is difficult to invent anything faster or
289 cheaper.
292 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
294 kfree_skb(skb);
295 return NET_XMIT_CN;
298 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
300 return NULL;
303 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
305 if (net_ratelimit())
306 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
307 skb->dev->name);
308 kfree_skb(skb);
309 return NET_XMIT_CN;
312 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
313 .id = "noop",
314 .priv_size = 0,
315 .enqueue = noop_enqueue,
316 .dequeue = noop_dequeue,
317 .requeue = noop_requeue,
318 .owner = THIS_MODULE,
321 static struct netdev_queue noop_netdev_queue = {
322 .qdisc = &noop_qdisc,
325 struct Qdisc noop_qdisc = {
326 .enqueue = noop_enqueue,
327 .dequeue = noop_dequeue,
328 .flags = TCQ_F_BUILTIN,
329 .ops = &noop_qdisc_ops,
330 .list = LIST_HEAD_INIT(noop_qdisc.list),
331 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
332 .dev_queue = &noop_netdev_queue,
334 EXPORT_SYMBOL(noop_qdisc);
336 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
337 .id = "noqueue",
338 .priv_size = 0,
339 .enqueue = noop_enqueue,
340 .dequeue = noop_dequeue,
341 .requeue = noop_requeue,
342 .owner = THIS_MODULE,
345 static struct Qdisc noqueue_qdisc;
346 static struct netdev_queue noqueue_netdev_queue = {
347 .qdisc = &noqueue_qdisc,
350 static struct Qdisc noqueue_qdisc = {
351 .enqueue = NULL,
352 .dequeue = noop_dequeue,
353 .flags = TCQ_F_BUILTIN,
354 .ops = &noqueue_qdisc_ops,
355 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
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,
371 struct Qdisc *qdisc)
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) {
382 qdisc->q.qlen++;
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)
391 int prio;
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)) {
396 qdisc->q.qlen--;
397 return __qdisc_dequeue_head(qdisc, list + prio);
401 return NULL;
404 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
406 qdisc->q.qlen++;
407 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
410 static void pfifo_fast_reset(struct Qdisc* qdisc)
412 int prio;
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;
419 qdisc->q.qlen = 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);
428 return skb->len;
430 nla_put_failure:
431 return -1;
434 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
436 int prio;
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);
442 return 0;
445 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
446 .id = "pfifo_fast",
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)
460 void *p;
461 struct Qdisc *sch;
462 unsigned int size;
463 int err = -ENOBUFS;
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);
470 if (!p)
471 goto errout;
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->q);
477 sch->ops = ops;
478 sch->enqueue = ops->enqueue;
479 sch->dequeue = ops->dequeue;
480 sch->dev_queue = dev_queue;
481 dev_hold(qdisc_dev(sch));
482 atomic_set(&sch->refcnt, 1);
484 return sch;
485 errout:
486 return ERR_PTR(err);
489 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
490 struct netdev_queue *dev_queue,
491 struct Qdisc_ops *ops,
492 unsigned int parentid)
494 struct Qdisc *sch;
496 sch = qdisc_alloc(dev_queue, ops);
497 if (IS_ERR(sch))
498 goto errout;
499 sch->parent = parentid;
501 if (!ops->init || ops->init(sch, NULL) == 0)
502 return sch;
504 qdisc_destroy(sch);
505 errout:
506 return NULL;
508 EXPORT_SYMBOL(qdisc_create_dflt);
510 /* Under qdisc_lock(qdisc) and BH! */
512 void qdisc_reset(struct Qdisc *qdisc)
514 const struct Qdisc_ops *ops = qdisc->ops;
516 if (ops->reset)
517 ops->reset(qdisc);
519 EXPORT_SYMBOL(qdisc_reset);
521 void qdisc_destroy(struct Qdisc *qdisc)
523 const struct Qdisc_ops *ops = qdisc->ops;
525 if (qdisc->flags & TCQ_F_BUILTIN ||
526 !atomic_dec_and_test(&qdisc->refcnt))
527 return;
529 if (qdisc->parent)
530 list_del(&qdisc->list);
532 #ifdef CONFIG_NET_SCHED
533 qdisc_put_stab(qdisc->stab);
534 #endif
535 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
536 if (ops->reset)
537 ops->reset(qdisc);
538 if (ops->destroy)
539 ops->destroy(qdisc);
541 module_put(ops->owner);
542 dev_put(qdisc_dev(qdisc));
544 kfree_skb(qdisc->gso_skb);
546 kfree((char *) qdisc - qdisc->padded);
548 EXPORT_SYMBOL(qdisc_destroy);
550 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
552 unsigned int i;
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)
558 return false;
560 return true;
563 static void attach_one_default_qdisc(struct net_device *dev,
564 struct netdev_queue *dev_queue,
565 void *_unused)
567 struct Qdisc *qdisc;
569 if (dev->tx_queue_len) {
570 qdisc = qdisc_create_dflt(dev, dev_queue,
571 &pfifo_fast_ops, TC_H_ROOT);
572 if (!qdisc) {
573 printk(KERN_INFO "%s: activation failed\n", dev->name);
574 return;
576 } else {
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)
599 int need_watchdog;
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
604 virtual interfaces
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 */
612 return;
614 need_watchdog = 0;
615 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
616 transition_one_qdisc(dev, &dev->rx_queue, NULL);
618 if (need_watchdog) {
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;
629 struct Qdisc *qdisc;
631 qdisc = dev_queue->qdisc;
632 if (qdisc) {
633 spin_lock_bh(qdisc_lock(qdisc));
635 if (!(qdisc->flags & TCQ_F_BUILTIN))
636 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
638 dev_queue->qdisc = qdisc_default;
639 qdisc_reset(qdisc);
641 spin_unlock_bh(qdisc_lock(qdisc));
645 static bool some_qdisc_is_busy(struct net_device *dev)
647 unsigned int i;
649 for (i = 0; i < dev->num_tx_queues; i++) {
650 struct netdev_queue *dev_queue;
651 spinlock_t *root_lock;
652 struct Qdisc *q;
653 int val;
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);
666 if (val)
667 return true;
669 return false;
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. */
680 synchronize_rcu();
682 /* Wait for outstanding qdisc_run calls. */
683 while (some_qdisc_is_busy(dev))
684 yield();
687 static void dev_init_scheduler_queue(struct net_device *dev,
688 struct netdev_queue *dev_queue,
689 void *_qdisc)
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;
712 if (qdisc) {
713 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));