Merge branch 'master' of /pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / net / sched / sch_generic.c
blob5f5efe4e6072e2959d6f6802a268cc7c5afca66b
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 q->gso_skb = skb;
48 q->qstats.requeues++;
49 __netif_schedule(q);
51 return 0;
54 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
56 struct sk_buff *skb = q->gso_skb;
58 if (unlikely(skb)) {
59 struct net_device *dev = qdisc_dev(q);
60 struct netdev_queue *txq;
62 /* check the reason of requeuing without tx lock first */
63 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
64 if (!netif_tx_queue_stopped(txq) && !netif_tx_queue_frozen(txq))
65 q->gso_skb = NULL;
66 else
67 skb = NULL;
68 } else {
69 skb = q->dequeue(q);
72 return skb;
75 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
76 struct netdev_queue *dev_queue,
77 struct Qdisc *q)
79 int ret;
81 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
83 * Same CPU holding the lock. It may be a transient
84 * configuration error, when hard_start_xmit() recurses. We
85 * detect it by checking xmit owner and drop the packet when
86 * deadloop is detected. Return OK to try the next skb.
88 kfree_skb(skb);
89 if (net_ratelimit())
90 printk(KERN_WARNING "Dead loop on netdevice %s, "
91 "fix it urgently!\n", dev_queue->dev->name);
92 ret = qdisc_qlen(q);
93 } else {
95 * Another cpu is holding lock, requeue & delay xmits for
96 * some time.
98 __get_cpu_var(netdev_rx_stat).cpu_collision++;
99 ret = dev_requeue_skb(skb, q);
102 return ret;
106 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
108 * __QDISC_STATE_RUNNING guarantees only one CPU can process
109 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
110 * this queue.
112 * netif_tx_lock serializes accesses to device driver.
114 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
115 * if one is grabbed, another must be free.
117 * Note, that this procedure can be called by a watchdog timer
119 * Returns to the caller:
120 * 0 - queue is empty or throttled.
121 * >0 - queue is not empty.
124 static inline int qdisc_restart(struct Qdisc *q)
126 struct netdev_queue *txq;
127 int ret = NETDEV_TX_BUSY;
128 struct net_device *dev;
129 spinlock_t *root_lock;
130 struct sk_buff *skb;
132 /* Dequeue packet */
133 if (unlikely((skb = dequeue_skb(q)) == NULL))
134 return 0;
136 root_lock = qdisc_lock(q);
138 /* And release qdisc */
139 spin_unlock(root_lock);
141 dev = qdisc_dev(q);
142 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
144 HARD_TX_LOCK(dev, txq, smp_processor_id());
145 if (!netif_tx_queue_stopped(txq) &&
146 !netif_tx_queue_frozen(txq))
147 ret = dev_hard_start_xmit(skb, dev, txq);
148 HARD_TX_UNLOCK(dev, txq);
150 spin_lock(root_lock);
152 switch (ret) {
153 case NETDEV_TX_OK:
154 /* Driver sent out skb successfully */
155 ret = qdisc_qlen(q);
156 break;
158 case NETDEV_TX_LOCKED:
159 /* Driver try lock failed */
160 ret = handle_dev_cpu_collision(skb, txq, q);
161 break;
163 default:
164 /* Driver returned NETDEV_TX_BUSY - requeue skb */
165 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
166 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
167 dev->name, ret, q->q.qlen);
169 ret = dev_requeue_skb(skb, q);
170 break;
173 if (ret && (netif_tx_queue_stopped(txq) ||
174 netif_tx_queue_frozen(txq)))
175 ret = 0;
177 return ret;
180 void __qdisc_run(struct Qdisc *q)
182 unsigned long start_time = jiffies;
184 while (qdisc_restart(q)) {
186 * Postpone processing if
187 * 1. another process needs the CPU;
188 * 2. we've been doing it for too long.
190 if (need_resched() || jiffies != start_time) {
191 __netif_schedule(q);
192 break;
196 clear_bit(__QDISC_STATE_RUNNING, &q->state);
199 static void dev_watchdog(unsigned long arg)
201 struct net_device *dev = (struct net_device *)arg;
203 netif_tx_lock(dev);
204 if (!qdisc_tx_is_noop(dev)) {
205 if (netif_device_present(dev) &&
206 netif_running(dev) &&
207 netif_carrier_ok(dev)) {
208 int some_queue_stopped = 0;
209 unsigned int i;
211 for (i = 0; i < dev->num_tx_queues; i++) {
212 struct netdev_queue *txq;
214 txq = netdev_get_tx_queue(dev, i);
215 if (netif_tx_queue_stopped(txq)) {
216 some_queue_stopped = 1;
217 break;
221 if (some_queue_stopped &&
222 time_after(jiffies, (dev->trans_start +
223 dev->watchdog_timeo))) {
224 char drivername[64];
225 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
226 dev->name, netdev_drivername(dev, drivername, 64));
227 dev->netdev_ops->ndo_tx_timeout(dev);
229 if (!mod_timer(&dev->watchdog_timer,
230 round_jiffies(jiffies +
231 dev->watchdog_timeo)))
232 dev_hold(dev);
235 netif_tx_unlock(dev);
237 dev_put(dev);
240 void __netdev_watchdog_up(struct net_device *dev)
242 if (dev->netdev_ops->ndo_tx_timeout) {
243 if (dev->watchdog_timeo <= 0)
244 dev->watchdog_timeo = 5*HZ;
245 if (!mod_timer(&dev->watchdog_timer,
246 round_jiffies(jiffies + dev->watchdog_timeo)))
247 dev_hold(dev);
251 static void dev_watchdog_up(struct net_device *dev)
253 __netdev_watchdog_up(dev);
256 static void dev_watchdog_down(struct net_device *dev)
258 netif_tx_lock_bh(dev);
259 if (del_timer(&dev->watchdog_timer))
260 dev_put(dev);
261 netif_tx_unlock_bh(dev);
265 * netif_carrier_on - set carrier
266 * @dev: network device
268 * Device has detected that carrier.
270 void netif_carrier_on(struct net_device *dev)
272 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
273 if (dev->reg_state == NETREG_UNINITIALIZED)
274 return;
275 linkwatch_fire_event(dev);
276 if (netif_running(dev))
277 __netdev_watchdog_up(dev);
280 EXPORT_SYMBOL(netif_carrier_on);
283 * netif_carrier_off - clear carrier
284 * @dev: network device
286 * Device has detected loss of carrier.
288 void netif_carrier_off(struct net_device *dev)
290 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
291 if (dev->reg_state == NETREG_UNINITIALIZED)
292 return;
293 linkwatch_fire_event(dev);
296 EXPORT_SYMBOL(netif_carrier_off);
298 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
299 under all circumstances. It is difficult to invent anything faster or
300 cheaper.
303 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
305 kfree_skb(skb);
306 return NET_XMIT_CN;
309 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
311 return NULL;
314 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
315 .id = "noop",
316 .priv_size = 0,
317 .enqueue = noop_enqueue,
318 .dequeue = noop_dequeue,
319 .peek = noop_dequeue,
320 .owner = THIS_MODULE,
323 static struct netdev_queue noop_netdev_queue = {
324 .qdisc = &noop_qdisc,
325 .qdisc_sleeping = &noop_qdisc,
328 struct Qdisc noop_qdisc = {
329 .enqueue = noop_enqueue,
330 .dequeue = noop_dequeue,
331 .flags = TCQ_F_BUILTIN,
332 .ops = &noop_qdisc_ops,
333 .list = LIST_HEAD_INIT(noop_qdisc.list),
334 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
335 .dev_queue = &noop_netdev_queue,
337 EXPORT_SYMBOL(noop_qdisc);
339 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
340 .id = "noqueue",
341 .priv_size = 0,
342 .enqueue = noop_enqueue,
343 .dequeue = noop_dequeue,
344 .peek = noop_dequeue,
345 .owner = THIS_MODULE,
348 static struct Qdisc noqueue_qdisc;
349 static struct netdev_queue noqueue_netdev_queue = {
350 .qdisc = &noqueue_qdisc,
351 .qdisc_sleeping = &noqueue_qdisc,
354 static struct Qdisc noqueue_qdisc = {
355 .enqueue = NULL,
356 .dequeue = noop_dequeue,
357 .flags = TCQ_F_BUILTIN,
358 .ops = &noqueue_qdisc_ops,
359 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
360 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
361 .dev_queue = &noqueue_netdev_queue,
365 static const u8 prio2band[TC_PRIO_MAX+1] =
366 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
368 /* 3-band FIFO queue: old style, but should be a bit faster than
369 generic prio+fifo combination.
372 #define PFIFO_FAST_BANDS 3
374 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
375 struct Qdisc *qdisc)
377 struct sk_buff_head *list = qdisc_priv(qdisc);
378 return list + prio2band[skb->priority & TC_PRIO_MAX];
381 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
383 struct sk_buff_head *list = prio2list(skb, qdisc);
385 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
386 qdisc->q.qlen++;
387 return __qdisc_enqueue_tail(skb, qdisc, list);
390 return qdisc_drop(skb, qdisc);
393 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
395 int prio;
396 struct sk_buff_head *list = qdisc_priv(qdisc);
398 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
399 if (!skb_queue_empty(list + prio)) {
400 qdisc->q.qlen--;
401 return __qdisc_dequeue_head(qdisc, list + prio);
405 return NULL;
408 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
410 int prio;
411 struct sk_buff_head *list = qdisc_priv(qdisc);
413 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
414 if (!skb_queue_empty(list + prio))
415 return skb_peek(list + prio);
418 return NULL;
421 static void pfifo_fast_reset(struct Qdisc* qdisc)
423 int prio;
424 struct sk_buff_head *list = qdisc_priv(qdisc);
426 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
427 __qdisc_reset_queue(qdisc, list + prio);
429 qdisc->qstats.backlog = 0;
430 qdisc->q.qlen = 0;
433 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
435 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
437 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
438 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
439 return skb->len;
441 nla_put_failure:
442 return -1;
445 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
447 int prio;
448 struct sk_buff_head *list = qdisc_priv(qdisc);
450 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
451 skb_queue_head_init(list + prio);
453 return 0;
456 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
457 .id = "pfifo_fast",
458 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
459 .enqueue = pfifo_fast_enqueue,
460 .dequeue = pfifo_fast_dequeue,
461 .peek = pfifo_fast_peek,
462 .init = pfifo_fast_init,
463 .reset = pfifo_fast_reset,
464 .dump = pfifo_fast_dump,
465 .owner = THIS_MODULE,
468 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
469 struct Qdisc_ops *ops)
471 void *p;
472 struct Qdisc *sch;
473 unsigned int size;
474 int err = -ENOBUFS;
476 /* ensure that the Qdisc and the private data are 32-byte aligned */
477 size = QDISC_ALIGN(sizeof(*sch));
478 size += ops->priv_size + (QDISC_ALIGNTO - 1);
480 p = kzalloc(size, GFP_KERNEL);
481 if (!p)
482 goto errout;
483 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
484 sch->padded = (char *) sch - (char *) p;
486 INIT_LIST_HEAD(&sch->list);
487 skb_queue_head_init(&sch->q);
488 sch->ops = ops;
489 sch->enqueue = ops->enqueue;
490 sch->dequeue = ops->dequeue;
491 sch->dev_queue = dev_queue;
492 dev_hold(qdisc_dev(sch));
493 atomic_set(&sch->refcnt, 1);
495 return sch;
496 errout:
497 return ERR_PTR(err);
500 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
501 struct netdev_queue *dev_queue,
502 struct Qdisc_ops *ops,
503 unsigned int parentid)
505 struct Qdisc *sch;
507 sch = qdisc_alloc(dev_queue, ops);
508 if (IS_ERR(sch))
509 goto errout;
510 sch->parent = parentid;
512 if (!ops->init || ops->init(sch, NULL) == 0)
513 return sch;
515 qdisc_destroy(sch);
516 errout:
517 return NULL;
519 EXPORT_SYMBOL(qdisc_create_dflt);
521 /* Under qdisc_lock(qdisc) and BH! */
523 void qdisc_reset(struct Qdisc *qdisc)
525 const struct Qdisc_ops *ops = qdisc->ops;
527 if (ops->reset)
528 ops->reset(qdisc);
530 kfree_skb(qdisc->gso_skb);
531 qdisc->gso_skb = NULL;
533 EXPORT_SYMBOL(qdisc_reset);
535 void qdisc_destroy(struct Qdisc *qdisc)
537 const struct Qdisc_ops *ops = qdisc->ops;
539 if (qdisc->flags & TCQ_F_BUILTIN ||
540 !atomic_dec_and_test(&qdisc->refcnt))
541 return;
543 #ifdef CONFIG_NET_SCHED
544 qdisc_list_del(qdisc);
546 qdisc_put_stab(qdisc->stab);
547 #endif
548 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
549 if (ops->reset)
550 ops->reset(qdisc);
551 if (ops->destroy)
552 ops->destroy(qdisc);
554 module_put(ops->owner);
555 dev_put(qdisc_dev(qdisc));
557 kfree_skb(qdisc->gso_skb);
558 kfree((char *) qdisc - qdisc->padded);
560 EXPORT_SYMBOL(qdisc_destroy);
562 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
564 unsigned int i;
566 for (i = 0; i < dev->num_tx_queues; i++) {
567 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
569 if (txq->qdisc_sleeping != &noop_qdisc)
570 return false;
572 return true;
575 static void attach_one_default_qdisc(struct net_device *dev,
576 struct netdev_queue *dev_queue,
577 void *_unused)
579 struct Qdisc *qdisc;
581 if (dev->tx_queue_len) {
582 qdisc = qdisc_create_dflt(dev, dev_queue,
583 &pfifo_fast_ops, TC_H_ROOT);
584 if (!qdisc) {
585 printk(KERN_INFO "%s: activation failed\n", dev->name);
586 return;
588 } else {
589 qdisc = &noqueue_qdisc;
591 dev_queue->qdisc_sleeping = qdisc;
594 static void transition_one_qdisc(struct net_device *dev,
595 struct netdev_queue *dev_queue,
596 void *_need_watchdog)
598 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
599 int *need_watchdog_p = _need_watchdog;
601 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
602 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
604 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
605 if (need_watchdog_p && new_qdisc != &noqueue_qdisc)
606 *need_watchdog_p = 1;
609 void dev_activate(struct net_device *dev)
611 int need_watchdog;
613 /* No queueing discipline is attached to device;
614 create default one i.e. pfifo_fast for devices,
615 which need queueing and noqueue_qdisc for
616 virtual interfaces
619 if (dev_all_qdisc_sleeping_noop(dev))
620 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
622 if (!netif_carrier_ok(dev))
623 /* Delay activation until next carrier-on event */
624 return;
626 need_watchdog = 0;
627 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
628 transition_one_qdisc(dev, &dev->rx_queue, NULL);
630 if (need_watchdog) {
631 dev->trans_start = jiffies;
632 dev_watchdog_up(dev);
636 static void dev_deactivate_queue(struct net_device *dev,
637 struct netdev_queue *dev_queue,
638 void *_qdisc_default)
640 struct Qdisc *qdisc_default = _qdisc_default;
641 struct Qdisc *qdisc;
643 qdisc = dev_queue->qdisc;
644 if (qdisc) {
645 spin_lock_bh(qdisc_lock(qdisc));
647 if (!(qdisc->flags & TCQ_F_BUILTIN))
648 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
650 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
651 qdisc_reset(qdisc);
653 spin_unlock_bh(qdisc_lock(qdisc));
657 static bool some_qdisc_is_busy(struct net_device *dev)
659 unsigned int i;
661 for (i = 0; i < dev->num_tx_queues; i++) {
662 struct netdev_queue *dev_queue;
663 spinlock_t *root_lock;
664 struct Qdisc *q;
665 int val;
667 dev_queue = netdev_get_tx_queue(dev, i);
668 q = dev_queue->qdisc_sleeping;
669 root_lock = qdisc_lock(q);
671 spin_lock_bh(root_lock);
673 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
674 test_bit(__QDISC_STATE_SCHED, &q->state));
676 spin_unlock_bh(root_lock);
678 if (val)
679 return true;
681 return false;
684 void dev_deactivate(struct net_device *dev)
686 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
687 dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
689 dev_watchdog_down(dev);
691 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
692 synchronize_rcu();
694 /* Wait for outstanding qdisc_run calls. */
695 while (some_qdisc_is_busy(dev))
696 yield();
699 static void dev_init_scheduler_queue(struct net_device *dev,
700 struct netdev_queue *dev_queue,
701 void *_qdisc)
703 struct Qdisc *qdisc = _qdisc;
705 dev_queue->qdisc = qdisc;
706 dev_queue->qdisc_sleeping = qdisc;
709 void dev_init_scheduler(struct net_device *dev)
711 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
712 dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
714 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
717 static void shutdown_scheduler_queue(struct net_device *dev,
718 struct netdev_queue *dev_queue,
719 void *_qdisc_default)
721 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
722 struct Qdisc *qdisc_default = _qdisc_default;
724 if (qdisc) {
725 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
726 dev_queue->qdisc_sleeping = qdisc_default;
728 qdisc_destroy(qdisc);
732 void dev_shutdown(struct net_device *dev)
734 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
735 shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
736 WARN_ON(timer_pending(&dev->watchdog_timer));