]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - net/sched/sch_generic.c
Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/torvalds/linux-2.6
[mirror_ubuntu-artful-kernel.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_root_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_root_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_subqueue_stopped(dev, skb))
139 ret = dev_hard_start_xmit(skb, dev, txq);
140 HARD_TX_UNLOCK(dev, txq);
141
142 spin_lock(root_lock);
143
144 switch (ret) {
145 case NETDEV_TX_OK:
146 /* Driver sent out skb successfully */
147 ret = qdisc_qlen(q);
148 break;
149
150 case NETDEV_TX_LOCKED:
151 /* Driver try lock failed */
152 ret = handle_dev_cpu_collision(skb, txq, q);
153 break;
154
155 default:
156 /* Driver returned NETDEV_TX_BUSY - requeue skb */
157 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
158 printk(KERN_WARNING "BUG %s code %d qlen %d\n",
159 dev->name, ret, q->q.qlen);
160
161 ret = dev_requeue_skb(skb, q);
162 break;
163 }
164
165 if (ret && netif_tx_queue_stopped(txq))
166 ret = 0;
167
168 return ret;
169 }
170
171 void __qdisc_run(struct Qdisc *q)
172 {
173 unsigned long start_time = jiffies;
174
175 while (qdisc_restart(q)) {
176 /*
177 * Postpone processing if
178 * 1. another process needs the CPU;
179 * 2. we've been doing it for too long.
180 */
181 if (need_resched() || jiffies != start_time) {
182 __netif_schedule(q);
183 break;
184 }
185 }
186
187 clear_bit(__QDISC_STATE_RUNNING, &q->state);
188 }
189
190 static void dev_watchdog(unsigned long arg)
191 {
192 struct net_device *dev = (struct net_device *)arg;
193
194 netif_tx_lock(dev);
195 if (!qdisc_tx_is_noop(dev)) {
196 if (netif_device_present(dev) &&
197 netif_running(dev) &&
198 netif_carrier_ok(dev)) {
199 int some_queue_stopped = 0;
200 unsigned int i;
201
202 for (i = 0; i < dev->num_tx_queues; i++) {
203 struct netdev_queue *txq;
204
205 txq = netdev_get_tx_queue(dev, i);
206 if (netif_tx_queue_stopped(txq)) {
207 some_queue_stopped = 1;
208 break;
209 }
210 }
211
212 if (some_queue_stopped &&
213 time_after(jiffies, (dev->trans_start +
214 dev->watchdog_timeo))) {
215 printk(KERN_INFO "NETDEV WATCHDOG: %s: "
216 "transmit timed out\n",
217 dev->name);
218 dev->tx_timeout(dev);
219 WARN_ON_ONCE(1);
220 }
221 if (!mod_timer(&dev->watchdog_timer,
222 round_jiffies(jiffies +
223 dev->watchdog_timeo)))
224 dev_hold(dev);
225 }
226 }
227 netif_tx_unlock(dev);
228
229 dev_put(dev);
230 }
231
232 void __netdev_watchdog_up(struct net_device *dev)
233 {
234 if (dev->tx_timeout) {
235 if (dev->watchdog_timeo <= 0)
236 dev->watchdog_timeo = 5*HZ;
237 if (!mod_timer(&dev->watchdog_timer,
238 round_jiffies(jiffies + dev->watchdog_timeo)))
239 dev_hold(dev);
240 }
241 }
242
243 static void dev_watchdog_up(struct net_device *dev)
244 {
245 __netdev_watchdog_up(dev);
246 }
247
248 static void dev_watchdog_down(struct net_device *dev)
249 {
250 netif_tx_lock_bh(dev);
251 if (del_timer(&dev->watchdog_timer))
252 dev_put(dev);
253 netif_tx_unlock_bh(dev);
254 }
255
256 /**
257 * netif_carrier_on - set carrier
258 * @dev: network device
259 *
260 * Device has detected that carrier.
261 */
262 void netif_carrier_on(struct net_device *dev)
263 {
264 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
265 linkwatch_fire_event(dev);
266 if (netif_running(dev))
267 __netdev_watchdog_up(dev);
268 }
269 }
270 EXPORT_SYMBOL(netif_carrier_on);
271
272 /**
273 * netif_carrier_off - clear carrier
274 * @dev: network device
275 *
276 * Device has detected loss of carrier.
277 */
278 void netif_carrier_off(struct net_device *dev)
279 {
280 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
281 linkwatch_fire_event(dev);
282 }
283 EXPORT_SYMBOL(netif_carrier_off);
284
285 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
286 under all circumstances. It is difficult to invent anything faster or
287 cheaper.
288 */
289
290 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
291 {
292 kfree_skb(skb);
293 return NET_XMIT_CN;
294 }
295
296 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
297 {
298 return NULL;
299 }
300
301 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
302 {
303 if (net_ratelimit())
304 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
305 skb->dev->name);
306 kfree_skb(skb);
307 return NET_XMIT_CN;
308 }
309
310 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
311 .id = "noop",
312 .priv_size = 0,
313 .enqueue = noop_enqueue,
314 .dequeue = noop_dequeue,
315 .requeue = noop_requeue,
316 .owner = THIS_MODULE,
317 };
318
319 static struct netdev_queue noop_netdev_queue = {
320 .qdisc = &noop_qdisc,
321 };
322
323 struct Qdisc noop_qdisc = {
324 .enqueue = noop_enqueue,
325 .dequeue = noop_dequeue,
326 .flags = TCQ_F_BUILTIN,
327 .ops = &noop_qdisc_ops,
328 .list = LIST_HEAD_INIT(noop_qdisc.list),
329 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
330 .dev_queue = &noop_netdev_queue,
331 };
332 EXPORT_SYMBOL(noop_qdisc);
333
334 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
335 .id = "noqueue",
336 .priv_size = 0,
337 .enqueue = noop_enqueue,
338 .dequeue = noop_dequeue,
339 .requeue = noop_requeue,
340 .owner = THIS_MODULE,
341 };
342
343 static struct Qdisc noqueue_qdisc = {
344 .enqueue = NULL,
345 .dequeue = noop_dequeue,
346 .flags = TCQ_F_BUILTIN,
347 .ops = &noqueue_qdisc_ops,
348 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
349 };
350
351
352 static int fifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
353 {
354 struct sk_buff_head *list = &qdisc->q;
355
356 if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len)
357 return __qdisc_enqueue_tail(skb, qdisc, list);
358
359 return qdisc_drop(skb, qdisc);
360 }
361
362 static struct sk_buff *fifo_fast_dequeue(struct Qdisc* qdisc)
363 {
364 struct sk_buff_head *list = &qdisc->q;
365
366 if (!skb_queue_empty(list))
367 return __qdisc_dequeue_head(qdisc, list);
368
369 return NULL;
370 }
371
372 static int fifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
373 {
374 return __qdisc_requeue(skb, qdisc, &qdisc->q);
375 }
376
377 static void fifo_fast_reset(struct Qdisc* qdisc)
378 {
379 __qdisc_reset_queue(qdisc, &qdisc->q);
380 qdisc->qstats.backlog = 0;
381 }
382
383 static struct Qdisc_ops fifo_fast_ops __read_mostly = {
384 .id = "fifo_fast",
385 .priv_size = 0,
386 .enqueue = fifo_fast_enqueue,
387 .dequeue = fifo_fast_dequeue,
388 .requeue = fifo_fast_requeue,
389 .reset = fifo_fast_reset,
390 .owner = THIS_MODULE,
391 };
392
393 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
394 struct Qdisc_ops *ops)
395 {
396 void *p;
397 struct Qdisc *sch;
398 unsigned int size;
399 int err = -ENOBUFS;
400
401 /* ensure that the Qdisc and the private data are 32-byte aligned */
402 size = QDISC_ALIGN(sizeof(*sch));
403 size += ops->priv_size + (QDISC_ALIGNTO - 1);
404
405 p = kzalloc(size, GFP_KERNEL);
406 if (!p)
407 goto errout;
408 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
409 sch->padded = (char *) sch - (char *) p;
410
411 INIT_LIST_HEAD(&sch->list);
412 skb_queue_head_init(&sch->q);
413 sch->ops = ops;
414 sch->enqueue = ops->enqueue;
415 sch->dequeue = ops->dequeue;
416 sch->dev_queue = dev_queue;
417 dev_hold(qdisc_dev(sch));
418 atomic_set(&sch->refcnt, 1);
419
420 return sch;
421 errout:
422 return ERR_PTR(err);
423 }
424
425 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
426 struct netdev_queue *dev_queue,
427 struct Qdisc_ops *ops,
428 unsigned int parentid)
429 {
430 struct Qdisc *sch;
431
432 sch = qdisc_alloc(dev_queue, ops);
433 if (IS_ERR(sch))
434 goto errout;
435 sch->parent = parentid;
436
437 if (!ops->init || ops->init(sch, NULL) == 0)
438 return sch;
439
440 qdisc_destroy(sch);
441 errout:
442 return NULL;
443 }
444 EXPORT_SYMBOL(qdisc_create_dflt);
445
446 /* Under qdisc_root_lock(qdisc) and BH! */
447
448 void qdisc_reset(struct Qdisc *qdisc)
449 {
450 const struct Qdisc_ops *ops = qdisc->ops;
451
452 if (ops->reset)
453 ops->reset(qdisc);
454 }
455 EXPORT_SYMBOL(qdisc_reset);
456
457 /* this is the rcu callback function to clean up a qdisc when there
458 * are no further references to it */
459
460 static void __qdisc_destroy(struct rcu_head *head)
461 {
462 struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
463 const struct Qdisc_ops *ops = qdisc->ops;
464
465 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
466 if (ops->reset)
467 ops->reset(qdisc);
468 if (ops->destroy)
469 ops->destroy(qdisc);
470
471 module_put(ops->owner);
472 dev_put(qdisc_dev(qdisc));
473
474 kfree_skb(qdisc->gso_skb);
475
476 kfree((char *) qdisc - qdisc->padded);
477 }
478
479 /* Under qdisc_root_lock(qdisc) and BH! */
480
481 void qdisc_destroy(struct Qdisc *qdisc)
482 {
483 struct net_device *dev = qdisc_dev(qdisc);
484
485 if (qdisc->flags & TCQ_F_BUILTIN ||
486 !atomic_dec_and_test(&qdisc->refcnt))
487 return;
488
489 spin_lock_bh(&dev->qdisc_list_lock);
490 list_del(&qdisc->list);
491 spin_unlock_bh(&dev->qdisc_list_lock);
492
493 call_rcu(&qdisc->q_rcu, __qdisc_destroy);
494 }
495 EXPORT_SYMBOL(qdisc_destroy);
496
497 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
498 {
499 unsigned int i;
500
501 for (i = 0; i < dev->num_tx_queues; i++) {
502 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
503
504 if (txq->qdisc_sleeping != &noop_qdisc)
505 return false;
506 }
507 return true;
508 }
509
510 static void attach_one_default_qdisc(struct net_device *dev,
511 struct netdev_queue *dev_queue,
512 void *_unused)
513 {
514 struct Qdisc *qdisc;
515
516 if (dev->tx_queue_len) {
517 qdisc = qdisc_create_dflt(dev, dev_queue,
518 &fifo_fast_ops, TC_H_ROOT);
519 if (!qdisc) {
520 printk(KERN_INFO "%s: activation failed\n", dev->name);
521 return;
522 }
523 spin_lock_bh(&dev->qdisc_list_lock);
524 list_add_tail(&qdisc->list, &dev->qdisc_list);
525 spin_unlock_bh(&dev->qdisc_list_lock);
526 } else {
527 qdisc = &noqueue_qdisc;
528 }
529 dev_queue->qdisc_sleeping = qdisc;
530 }
531
532 static void transition_one_qdisc(struct net_device *dev,
533 struct netdev_queue *dev_queue,
534 void *_need_watchdog)
535 {
536 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
537 int *need_watchdog_p = _need_watchdog;
538
539 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
540 if (new_qdisc != &noqueue_qdisc)
541 *need_watchdog_p = 1;
542 }
543
544 void dev_activate(struct net_device *dev)
545 {
546 int need_watchdog;
547
548 /* No queueing discipline is attached to device;
549 * create default one i.e. fifo_fast for devices,
550 * which need queueing and noqueue_qdisc for
551 * virtual interfaces.
552 */
553
554 if (dev_all_qdisc_sleeping_noop(dev))
555 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
556
557 if (!netif_carrier_ok(dev))
558 /* Delay activation until next carrier-on event */
559 return;
560
561 need_watchdog = 0;
562 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
563
564 if (need_watchdog) {
565 dev->trans_start = jiffies;
566 dev_watchdog_up(dev);
567 }
568 }
569
570 static void dev_deactivate_queue(struct net_device *dev,
571 struct netdev_queue *dev_queue,
572 void *_qdisc_default)
573 {
574 struct Qdisc *qdisc_default = _qdisc_default;
575 struct sk_buff *skb = NULL;
576 struct Qdisc *qdisc;
577
578 qdisc = dev_queue->qdisc;
579 if (qdisc) {
580 spin_lock_bh(qdisc_lock(qdisc));
581
582 dev_queue->qdisc = qdisc_default;
583 qdisc_reset(qdisc);
584
585 spin_unlock_bh(qdisc_lock(qdisc));
586 }
587
588 kfree_skb(skb);
589 }
590
591 static bool some_qdisc_is_running(struct net_device *dev, int lock)
592 {
593 unsigned int i;
594
595 for (i = 0; i < dev->num_tx_queues; i++) {
596 struct netdev_queue *dev_queue;
597 spinlock_t *root_lock;
598 struct Qdisc *q;
599 int val;
600
601 dev_queue = netdev_get_tx_queue(dev, i);
602 q = dev_queue->qdisc;
603 root_lock = qdisc_root_lock(q);
604
605 if (lock)
606 spin_lock_bh(root_lock);
607
608 val = test_bit(__QDISC_STATE_RUNNING, &q->state);
609
610 if (lock)
611 spin_unlock_bh(root_lock);
612
613 if (val)
614 return true;
615 }
616 return false;
617 }
618
619 void dev_deactivate(struct net_device *dev)
620 {
621 bool running;
622
623 netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
624
625 dev_watchdog_down(dev);
626
627 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
628 synchronize_rcu();
629
630 /* Wait for outstanding qdisc_run calls. */
631 do {
632 while (some_qdisc_is_running(dev, 0))
633 yield();
634
635 /*
636 * Double-check inside queue lock to ensure that all effects
637 * of the queue run are visible when we return.
638 */
639 running = some_qdisc_is_running(dev, 1);
640
641 /*
642 * The running flag should never be set at this point because
643 * we've already set dev->qdisc to noop_qdisc *inside* the same
644 * pair of spin locks. That is, if any qdisc_run starts after
645 * our initial test it should see the noop_qdisc and then
646 * clear the RUNNING bit before dropping the queue lock. So
647 * if it is set here then we've found a bug.
648 */
649 } while (WARN_ON_ONCE(running));
650 }
651
652 static void dev_init_scheduler_queue(struct net_device *dev,
653 struct netdev_queue *dev_queue,
654 void *_qdisc)
655 {
656 struct Qdisc *qdisc = _qdisc;
657
658 dev_queue->qdisc = qdisc;
659 dev_queue->qdisc_sleeping = qdisc;
660 }
661
662 void dev_init_scheduler(struct net_device *dev)
663 {
664 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
665 dev_init_scheduler_queue(dev, &dev->rx_queue, NULL);
666
667 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
668 }
669
670 static void shutdown_scheduler_queue(struct net_device *dev,
671 struct netdev_queue *dev_queue,
672 void *_qdisc_default)
673 {
674 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
675 struct Qdisc *qdisc_default = _qdisc_default;
676
677 if (qdisc) {
678 spinlock_t *root_lock = qdisc_root_lock(qdisc);
679
680 dev_queue->qdisc = qdisc_default;
681 dev_queue->qdisc_sleeping = qdisc_default;
682
683 spin_lock(root_lock);
684 qdisc_destroy(qdisc);
685 spin_unlock(root_lock);
686 }
687 }
688
689 void dev_shutdown(struct net_device *dev)
690 {
691 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
692 shutdown_scheduler_queue(dev, &dev->rx_queue, NULL);
693 BUG_TRAP(!timer_pending(&dev->watchdog_timer));
694 }