Merge branch 'next' into for-linus
[deliverable/linux.git] / net / sched / sch_generic.c
1 /*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
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.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
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>
28
29 /* Main transmission queue. */
30
31 /* Modifications to data participating in scheduling must be protected with
32 * qdisc_lock(qdisc) spinlock.
33 *
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.
38 */
39
40 static inline int qdisc_qlen(struct Qdisc *q)
41 {
42 return q->q.qlen;
43 }
44
45 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
46 {
47 if (unlikely(skb->next))
48 q->gso_skb = skb;
49 else
50 q->ops->requeue(skb, q);
51
52 __netif_schedule(q);
53 return 0;
54 }
55
56 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
57 {
58 struct sk_buff *skb;
59
60 if ((skb = q->gso_skb))
61 q->gso_skb = NULL;
62 else
63 skb = q->dequeue(q);
64
65 return skb;
66 }
67
68 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
69 struct netdev_queue *dev_queue,
70 struct Qdisc *q)
71 {
72 int ret;
73
74 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
75 /*
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.
80 */
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 {
87 /*
88 * Another cpu is holding lock, requeue & delay xmits for
89 * some time.
90 */
91 __get_cpu_var(netdev_rx_stat).cpu_collision++;
92 ret = dev_requeue_skb(skb, q);
93 }
94
95 return ret;
96 }
97
98 /*
99 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
100 *
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.
104 *
105 * netif_tx_lock serializes accesses to device driver.
106 *
107 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
108 * if one is grabbed, another must be free.
109 *
110 * Note, that this procedure can be called by a watchdog timer
111 *
112 * Returns to the caller:
113 * 0 - queue is empty or throttled.
114 * >0 - queue is not empty.
115 *
116 */
117 static inline int qdisc_restart(struct Qdisc *q)
118 {
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;
124
125 /* Dequeue packet */
126 if (unlikely((skb = dequeue_skb(q)) == NULL))
127 return 0;
128
129 root_lock = qdisc_lock(q);
130
131 /* And release qdisc */
132 spin_unlock(root_lock);
133
134 dev = qdisc_dev(q);
135 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
136
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);
142
143 spin_lock(root_lock);
144
145 switch (ret) {
146 case NETDEV_TX_OK:
147 /* Driver sent out skb successfully */
148 ret = qdisc_qlen(q);
149 break;
150
151 case NETDEV_TX_LOCKED:
152 /* Driver try lock failed */
153 ret = handle_dev_cpu_collision(skb, txq, q);
154 break;
155
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);
161
162 ret = dev_requeue_skb(skb, q);
163 break;
164 }
165
166 if (ret && (netif_tx_queue_stopped(txq) ||
167 netif_tx_queue_frozen(txq)))
168 ret = 0;
169
170 return ret;
171 }
172
173 void __qdisc_run(struct Qdisc *q)
174 {
175 unsigned long start_time = jiffies;
176
177 while (qdisc_restart(q)) {
178 /*
179 * Postpone processing if
180 * 1. another process needs the CPU;
181 * 2. we've been doing it for too long.
182 */
183 if (need_resched() || jiffies != start_time) {
184 __netif_schedule(q);
185 break;
186 }
187 }
188
189 clear_bit(__QDISC_STATE_RUNNING, &q->state);
190 }
191
192 static void dev_watchdog(unsigned long arg)
193 {
194 struct net_device *dev = (struct net_device *)arg;
195
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;
203
204 for (i = 0; i < dev->num_tx_queues; i++) {
205 struct netdev_queue *txq;
206
207 txq = netdev_get_tx_queue(dev, i);
208 if (netif_tx_queue_stopped(txq)) {
209 some_queue_stopped = 1;
210 break;
211 }
212 }
213
214 if (some_queue_stopped &&
215 time_after(jiffies, (dev->trans_start +
216 dev->watchdog_timeo))) {
217 char drivername[64];
218 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit timed out\n",
219 dev->name, netdev_drivername(dev, drivername, 64));
220 dev->tx_timeout(dev);
221 }
222 if (!mod_timer(&dev->watchdog_timer,
223 round_jiffies(jiffies +
224 dev->watchdog_timeo)))
225 dev_hold(dev);
226 }
227 }
228 netif_tx_unlock(dev);
229
230 dev_put(dev);
231 }
232
233 void __netdev_watchdog_up(struct net_device *dev)
234 {
235 if (dev->tx_timeout) {
236 if (dev->watchdog_timeo <= 0)
237 dev->watchdog_timeo = 5*HZ;
238 if (!mod_timer(&dev->watchdog_timer,
239 round_jiffies(jiffies + dev->watchdog_timeo)))
240 dev_hold(dev);
241 }
242 }
243
244 static void dev_watchdog_up(struct net_device *dev)
245 {
246 __netdev_watchdog_up(dev);
247 }
248
249 static void dev_watchdog_down(struct net_device *dev)
250 {
251 netif_tx_lock_bh(dev);
252 if (del_timer(&dev->watchdog_timer))
253 dev_put(dev);
254 netif_tx_unlock_bh(dev);
255 }
256
257 /**
258 * netif_carrier_on - set carrier
259 * @dev: network device
260 *
261 * Device has detected that carrier.
262 */
263 void netif_carrier_on(struct net_device *dev)
264 {
265 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
266 linkwatch_fire_event(dev);
267 if (netif_running(dev))
268 __netdev_watchdog_up(dev);
269 }
270 }
271 EXPORT_SYMBOL(netif_carrier_on);
272
273 /**
274 * netif_carrier_off - clear carrier
275 * @dev: network device
276 *
277 * Device has detected loss of carrier.
278 */
279 void netif_carrier_off(struct net_device *dev)
280 {
281 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
282 linkwatch_fire_event(dev);
283 }
284 EXPORT_SYMBOL(netif_carrier_off);
285
286 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
287 under all circumstances. It is difficult to invent anything faster or
288 cheaper.
289 */
290
291 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
292 {
293 kfree_skb(skb);
294 return NET_XMIT_CN;
295 }
296
297 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
298 {
299 return NULL;
300 }
301
302 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
303 {
304 if (net_ratelimit())
305 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
306 skb->dev->name);
307 kfree_skb(skb);
308 return NET_XMIT_CN;
309 }
310
311 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
312 .id = "noop",
313 .priv_size = 0,
314 .enqueue = noop_enqueue,
315 .dequeue = noop_dequeue,
316 .requeue = noop_requeue,
317 .owner = THIS_MODULE,
318 };
319
320 static struct netdev_queue noop_netdev_queue = {
321 .qdisc = &noop_qdisc,
322 };
323
324 struct Qdisc noop_qdisc = {
325 .enqueue = noop_enqueue,
326 .dequeue = noop_dequeue,
327 .flags = TCQ_F_BUILTIN,
328 .ops = &noop_qdisc_ops,
329 .list = LIST_HEAD_INIT(noop_qdisc.list),
330 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
331 .dev_queue = &noop_netdev_queue,
332 };
333 EXPORT_SYMBOL(noop_qdisc);
334
335 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
336 .id = "noqueue",
337 .priv_size = 0,
338 .enqueue = noop_enqueue,
339 .dequeue = noop_dequeue,
340 .requeue = noop_requeue,
341 .owner = THIS_MODULE,
342 };
343
344 static struct Qdisc noqueue_qdisc;
345 static struct netdev_queue noqueue_netdev_queue = {
346 .qdisc = &noqueue_qdisc,
347 };
348
349 static struct Qdisc noqueue_qdisc = {
350 .enqueue = NULL,
351 .dequeue = noop_dequeue,
352 .flags = TCQ_F_BUILTIN,
353 .ops = &noqueue_qdisc_ops,
354 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
355 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
356 .dev_queue = &noqueue_netdev_queue,
357 };
358
359
360 static const u8 prio2band[TC_PRIO_MAX+1] =
361 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
362
363 /* 3-band FIFO queue: old style, but should be a bit faster than
364 generic prio+fifo combination.
365 */
366
367 #define PFIFO_FAST_BANDS 3
368
369 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
370 struct Qdisc *qdisc)
371 {
372 struct sk_buff_head *list = qdisc_priv(qdisc);
373 return list + prio2band[skb->priority & TC_PRIO_MAX];
374 }
375
376 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
377 {
378 struct sk_buff_head *list = prio2list(skb, qdisc);
379
380 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
381 qdisc->q.qlen++;
382 return __qdisc_enqueue_tail(skb, qdisc, list);
383 }
384
385 return qdisc_drop(skb, qdisc);
386 }
387
388 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
389 {
390 int prio;
391 struct sk_buff_head *list = qdisc_priv(qdisc);
392
393 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
394 if (!skb_queue_empty(list + prio)) {
395 qdisc->q.qlen--;
396 return __qdisc_dequeue_head(qdisc, list + prio);
397 }
398 }
399
400 return NULL;
401 }
402
403 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
404 {
405 qdisc->q.qlen++;
406 return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
407 }
408
409 static void pfifo_fast_reset(struct Qdisc* qdisc)
410 {
411 int prio;
412 struct sk_buff_head *list = qdisc_priv(qdisc);
413
414 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
415 __qdisc_reset_queue(qdisc, list + prio);
416
417 qdisc->qstats.backlog = 0;
418 qdisc->q.qlen = 0;
419 }
420
421 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
422 {
423 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
424
425 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
426 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
427 return skb->len;
428
429 nla_put_failure:
430 return -1;
431 }
432
433 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
434 {
435 int prio;
436 struct sk_buff_head *list = qdisc_priv(qdisc);
437
438 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
439 skb_queue_head_init(list + prio);
440
441 return 0;
442 }
443
444 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
445 .id = "pfifo_fast",
446 .priv_size = PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
447 .enqueue = pfifo_fast_enqueue,
448 .dequeue = pfifo_fast_dequeue,
449 .requeue = pfifo_fast_requeue,
450 .init = pfifo_fast_init,
451 .reset = pfifo_fast_reset,
452 .dump = pfifo_fast_dump,
453 .owner = THIS_MODULE,
454 };
455
456 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
457 struct Qdisc_ops *ops)
458 {
459 void *p;
460 struct Qdisc *sch;
461 unsigned int size;
462 int err = -ENOBUFS;
463
464 /* ensure that the Qdisc and the private data are 32-byte aligned */
465 size = QDISC_ALIGN(sizeof(*sch));
466 size += ops->priv_size + (QDISC_ALIGNTO - 1);
467
468 p = kzalloc(size, GFP_KERNEL);
469 if (!p)
470 goto errout;
471 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
472 sch->padded = (char *) sch - (char *) p;
473
474 INIT_LIST_HEAD(&sch->list);
475 skb_queue_head_init(&sch->q);
476 sch->ops = ops;
477 sch->enqueue = ops->enqueue;
478 sch->dequeue = ops->dequeue;
479 sch->dev_queue = dev_queue;
480 dev_hold(qdisc_dev(sch));
481 atomic_set(&sch->refcnt, 1);
482
483 return sch;
484 errout:
485 return ERR_PTR(err);
486 }
487
488 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
489 struct netdev_queue *dev_queue,
490 struct Qdisc_ops *ops,
491 unsigned int parentid)
492 {
493 struct Qdisc *sch;
494
495 sch = qdisc_alloc(dev_queue, ops);
496 if (IS_ERR(sch))
497 goto errout;
498 sch->parent = parentid;
499
500 if (!ops->init || ops->init(sch, NULL) == 0)
501 return sch;
502
503 qdisc_destroy(sch);
504 errout:
505 return NULL;
506 }
507 EXPORT_SYMBOL(qdisc_create_dflt);
508
509 /* Under qdisc_lock(qdisc) and BH! */
510
511 void qdisc_reset(struct Qdisc *qdisc)
512 {
513 const struct Qdisc_ops *ops = qdisc->ops;
514
515 if (ops->reset)
516 ops->reset(qdisc);
517 }
518 EXPORT_SYMBOL(qdisc_reset);
519
520 void qdisc_destroy(struct Qdisc *qdisc)
521 {
522 const struct Qdisc_ops *ops = qdisc->ops;
523
524 if (qdisc->flags & TCQ_F_BUILTIN ||
525 !atomic_dec_and_test(&qdisc->refcnt))
526 return;
527
528 #ifdef CONFIG_NET_SCHED
529 qdisc_list_del(qdisc);
530
531 qdisc_put_stab(qdisc->stab);
532 #endif
533 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
534 if (ops->reset)
535 ops->reset(qdisc);
536 if (ops->destroy)
537 ops->destroy(qdisc);
538
539 module_put(ops->owner);
540 dev_put(qdisc_dev(qdisc));
541
542 kfree_skb(qdisc->gso_skb);
543
544 kfree((char *) qdisc - qdisc->padded);
545 }
546 EXPORT_SYMBOL(qdisc_destroy);
547
548 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
549 {
550 unsigned int i;
551
552 for (i = 0; i < dev->num_tx_queues; i++) {
553 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
554
555 if (txq->qdisc_sleeping != &noop_qdisc)
556 return false;
557 }
558 return true;
559 }
560
561 static void attach_one_default_qdisc(struct net_device *dev,
562 struct netdev_queue *dev_queue,
563 void *_unused)
564 {
565 struct Qdisc *qdisc;
566
567 if (dev->tx_queue_len) {
568 qdisc = qdisc_create_dflt(dev, dev_queue,
569 &pfifo_fast_ops, TC_H_ROOT);
570 if (!qdisc) {
571 printk(KERN_INFO "%s: activation failed\n", dev->name);
572 return;
573 }
574 } else {
575 qdisc = &noqueue_qdisc;
576 }
577 dev_queue->qdisc_sleeping = qdisc;
578 }
579
580 static void transition_one_qdisc(struct net_device *dev,
581 struct netdev_queue *dev_queue,
582 void *_need_watchdog)
583 {
584 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
585 int *need_watchdog_p = _need_watchdog;
586
587 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
588 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
589
590 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
591 if (need_watchdog_p && new_qdisc != &noqueue_qdisc)
592 *need_watchdog_p = 1;
593 }
594
595 void dev_activate(struct net_device *dev)
596 {
597 int need_watchdog;
598
599 /* No queueing discipline is attached to device;
600 create default one i.e. pfifo_fast for devices,
601 which need queueing and noqueue_qdisc for
602 virtual interfaces
603 */
604
605 if (dev_all_qdisc_sleeping_noop(dev))
606 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
607
608 if (!netif_carrier_ok(dev))
609 /* Delay activation until next carrier-on event */
610 return;
611
612 need_watchdog = 0;
613 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
614 transition_one_qdisc(dev, &dev->rx_queue, NULL);
615
616 if (need_watchdog) {
617 dev->trans_start = jiffies;
618 dev_watchdog_up(dev);
619 }
620 }
621
622 static void dev_deactivate_queue(struct net_device *dev,
623 struct netdev_queue *dev_queue,
624 void *_qdisc_default)
625 {
626 struct Qdisc *qdisc_default = _qdisc_default;
627 struct Qdisc *qdisc;
628
629 qdisc = dev_queue->qdisc;
630 if (qdisc) {
631 spin_lock_bh(qdisc_lock(qdisc));
632
633 if (!(qdisc->flags & TCQ_F_BUILTIN))
634 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
635
636 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
637 qdisc_reset(qdisc);
638
639 spin_unlock_bh(qdisc_lock(qdisc));
640 }
641 }
642
643 static bool some_qdisc_is_busy(struct net_device *dev)
644 {
645 unsigned int i;
646
647 for (i = 0; i < dev->num_tx_queues; i++) {
648 struct netdev_queue *dev_queue;
649 spinlock_t *root_lock;
650 struct Qdisc *q;
651 int val;
652
653 dev_queue = netdev_get_tx_queue(dev, i);
654 q = dev_queue->qdisc_sleeping;
655 root_lock = qdisc_lock(q);
656
657 spin_lock_bh(root_lock);
658
659 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
660 test_bit(__QDISC_STATE_SCHED, &q->state));
661
662 spin_unlock_bh(root_lock);
663
664 if (val)
665 return true;
666 }
667 return false;
668 }
669
670 void dev_deactivate(struct net_device *dev)
671 {
672 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
673 dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
674
675 dev_watchdog_down(dev);
676
677 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
678 synchronize_rcu();
679
680 /* Wait for outstanding qdisc_run calls. */
681 while (some_qdisc_is_busy(dev))
682 yield();
683 }
684
685 static void dev_init_scheduler_queue(struct net_device *dev,
686 struct netdev_queue *dev_queue,
687 void *_qdisc)
688 {
689 struct Qdisc *qdisc = _qdisc;
690
691 dev_queue->qdisc = qdisc;
692 dev_queue->qdisc_sleeping = qdisc;
693 }
694
695 void dev_init_scheduler(struct net_device *dev)
696 {
697 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
698 dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
699
700 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
701 }
702
703 static void shutdown_scheduler_queue(struct net_device *dev,
704 struct netdev_queue *dev_queue,
705 void *_qdisc_default)
706 {
707 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
708 struct Qdisc *qdisc_default = _qdisc_default;
709
710 if (qdisc) {
711 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
712 dev_queue->qdisc_sleeping = qdisc_default;
713
714 qdisc_destroy(qdisc);
715 }
716 }
717
718 void dev_shutdown(struct net_device *dev)
719 {
720 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
721 shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
722 WARN_ON(timer_pending(&dev->watchdog_timer));
723 }
This page took 0.044526 seconds and 6 git commands to generate.