Merge branch 'for-2.6.27' of git://linux-nfs.org/~bfields/linux
[linux-2.6] / 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 static struct netdev_queue noqueue_netdev_queue = {
345         .qdisc          =       &noqueue_qdisc,
346 };
347
348 static struct Qdisc noqueue_qdisc = {
349         .enqueue        =       NULL,
350         .dequeue        =       noop_dequeue,
351         .flags          =       TCQ_F_BUILTIN,
352         .ops            =       &noqueue_qdisc_ops,
353         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
354         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
355         .dev_queue      =       &noqueue_netdev_queue,
356 };
357
358
359 static int fifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
360 {
361         struct sk_buff_head *list = &qdisc->q;
362
363         if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len)
364                 return __qdisc_enqueue_tail(skb, qdisc, list);
365
366         return qdisc_drop(skb, qdisc);
367 }
368
369 static struct sk_buff *fifo_fast_dequeue(struct Qdisc* qdisc)
370 {
371         struct sk_buff_head *list = &qdisc->q;
372
373         if (!skb_queue_empty(list))
374                 return __qdisc_dequeue_head(qdisc, list);
375
376         return NULL;
377 }
378
379 static int fifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
380 {
381         return __qdisc_requeue(skb, qdisc, &qdisc->q);
382 }
383
384 static void fifo_fast_reset(struct Qdisc* qdisc)
385 {
386         __qdisc_reset_queue(qdisc, &qdisc->q);
387         qdisc->qstats.backlog = 0;
388 }
389
390 static struct Qdisc_ops fifo_fast_ops __read_mostly = {
391         .id             =       "fifo_fast",
392         .priv_size      =       0,
393         .enqueue        =       fifo_fast_enqueue,
394         .dequeue        =       fifo_fast_dequeue,
395         .requeue        =       fifo_fast_requeue,
396         .reset          =       fifo_fast_reset,
397         .owner          =       THIS_MODULE,
398 };
399
400 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
401                           struct Qdisc_ops *ops)
402 {
403         void *p;
404         struct Qdisc *sch;
405         unsigned int size;
406         int err = -ENOBUFS;
407
408         /* ensure that the Qdisc and the private data are 32-byte aligned */
409         size = QDISC_ALIGN(sizeof(*sch));
410         size += ops->priv_size + (QDISC_ALIGNTO - 1);
411
412         p = kzalloc(size, GFP_KERNEL);
413         if (!p)
414                 goto errout;
415         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
416         sch->padded = (char *) sch - (char *) p;
417
418         INIT_LIST_HEAD(&sch->list);
419         skb_queue_head_init(&sch->q);
420         sch->ops = ops;
421         sch->enqueue = ops->enqueue;
422         sch->dequeue = ops->dequeue;
423         sch->dev_queue = dev_queue;
424         dev_hold(qdisc_dev(sch));
425         atomic_set(&sch->refcnt, 1);
426
427         return sch;
428 errout:
429         return ERR_PTR(err);
430 }
431
432 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
433                                  struct netdev_queue *dev_queue,
434                                  struct Qdisc_ops *ops,
435                                  unsigned int parentid)
436 {
437         struct Qdisc *sch;
438
439         sch = qdisc_alloc(dev_queue, ops);
440         if (IS_ERR(sch))
441                 goto errout;
442         sch->parent = parentid;
443
444         if (!ops->init || ops->init(sch, NULL) == 0)
445                 return sch;
446
447         qdisc_destroy(sch);
448 errout:
449         return NULL;
450 }
451 EXPORT_SYMBOL(qdisc_create_dflt);
452
453 /* Under qdisc_root_lock(qdisc) and BH! */
454
455 void qdisc_reset(struct Qdisc *qdisc)
456 {
457         const struct Qdisc_ops *ops = qdisc->ops;
458
459         if (ops->reset)
460                 ops->reset(qdisc);
461 }
462 EXPORT_SYMBOL(qdisc_reset);
463
464 /* this is the rcu callback function to clean up a qdisc when there
465  * are no further references to it */
466
467 static void __qdisc_destroy(struct rcu_head *head)
468 {
469         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
470         const struct Qdisc_ops  *ops = qdisc->ops;
471
472 #ifdef CONFIG_NET_SCHED
473         qdisc_put_stab(qdisc->stab);
474 #endif
475         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
476         if (ops->reset)
477                 ops->reset(qdisc);
478         if (ops->destroy)
479                 ops->destroy(qdisc);
480
481         module_put(ops->owner);
482         dev_put(qdisc_dev(qdisc));
483
484         kfree_skb(qdisc->gso_skb);
485
486         kfree((char *) qdisc - qdisc->padded);
487 }
488
489 /* Under qdisc_root_lock(qdisc) and BH! */
490
491 void qdisc_destroy(struct Qdisc *qdisc)
492 {
493         if (qdisc->flags & TCQ_F_BUILTIN ||
494             !atomic_dec_and_test(&qdisc->refcnt))
495                 return;
496
497         if (qdisc->parent)
498                 list_del(&qdisc->list);
499
500         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
501 }
502 EXPORT_SYMBOL(qdisc_destroy);
503
504 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
505 {
506         unsigned int i;
507
508         for (i = 0; i < dev->num_tx_queues; i++) {
509                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
510
511                 if (txq->qdisc_sleeping != &noop_qdisc)
512                         return false;
513         }
514         return true;
515 }
516
517 static void attach_one_default_qdisc(struct net_device *dev,
518                                      struct netdev_queue *dev_queue,
519                                      void *_unused)
520 {
521         struct Qdisc *qdisc;
522
523         if (dev->tx_queue_len) {
524                 qdisc = qdisc_create_dflt(dev, dev_queue,
525                                           &fifo_fast_ops, TC_H_ROOT);
526                 if (!qdisc) {
527                         printk(KERN_INFO "%s: activation failed\n", dev->name);
528                         return;
529                 }
530         } else {
531                 qdisc =  &noqueue_qdisc;
532         }
533         dev_queue->qdisc_sleeping = qdisc;
534 }
535
536 static void transition_one_qdisc(struct net_device *dev,
537                                  struct netdev_queue *dev_queue,
538                                  void *_need_watchdog)
539 {
540         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
541         int *need_watchdog_p = _need_watchdog;
542
543         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
544         if (new_qdisc != &noqueue_qdisc)
545                 *need_watchdog_p = 1;
546 }
547
548 void dev_activate(struct net_device *dev)
549 {
550         int need_watchdog;
551
552         /* No queueing discipline is attached to device;
553          * create default one i.e. fifo_fast for devices,
554          * which need queueing and noqueue_qdisc for
555          * virtual interfaces.
556          */
557
558         if (dev_all_qdisc_sleeping_noop(dev))
559                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
560
561         if (!netif_carrier_ok(dev))
562                 /* Delay activation until next carrier-on event */
563                 return;
564
565         need_watchdog = 0;
566         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
567
568         if (need_watchdog) {
569                 dev->trans_start = jiffies;
570                 dev_watchdog_up(dev);
571         }
572 }
573
574 static void dev_deactivate_queue(struct net_device *dev,
575                                  struct netdev_queue *dev_queue,
576                                  void *_qdisc_default)
577 {
578         struct Qdisc *qdisc_default = _qdisc_default;
579         struct sk_buff *skb = NULL;
580         struct Qdisc *qdisc;
581
582         qdisc = dev_queue->qdisc;
583         if (qdisc) {
584                 spin_lock_bh(qdisc_lock(qdisc));
585
586                 dev_queue->qdisc = qdisc_default;
587                 qdisc_reset(qdisc);
588
589                 spin_unlock_bh(qdisc_lock(qdisc));
590         }
591
592         kfree_skb(skb);
593 }
594
595 static bool some_qdisc_is_running(struct net_device *dev, int lock)
596 {
597         unsigned int i;
598
599         for (i = 0; i < dev->num_tx_queues; i++) {
600                 struct netdev_queue *dev_queue;
601                 spinlock_t *root_lock;
602                 struct Qdisc *q;
603                 int val;
604
605                 dev_queue = netdev_get_tx_queue(dev, i);
606                 q = dev_queue->qdisc;
607                 root_lock = qdisc_root_lock(q);
608
609                 if (lock)
610                         spin_lock_bh(root_lock);
611
612                 val = test_bit(__QDISC_STATE_RUNNING, &q->state);
613
614                 if (lock)
615                         spin_unlock_bh(root_lock);
616
617                 if (val)
618                         return true;
619         }
620         return false;
621 }
622
623 void dev_deactivate(struct net_device *dev)
624 {
625         bool running;
626
627         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
628
629         dev_watchdog_down(dev);
630
631         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
632         synchronize_rcu();
633
634         /* Wait for outstanding qdisc_run calls. */
635         do {
636                 while (some_qdisc_is_running(dev, 0))
637                         yield();
638
639                 /*
640                  * Double-check inside queue lock to ensure that all effects
641                  * of the queue run are visible when we return.
642                  */
643                 running = some_qdisc_is_running(dev, 1);
644
645                 /*
646                  * The running flag should never be set at this point because
647                  * we've already set dev->qdisc to noop_qdisc *inside* the same
648                  * pair of spin locks.  That is, if any qdisc_run starts after
649                  * our initial test it should see the noop_qdisc and then
650                  * clear the RUNNING bit before dropping the queue lock.  So
651                  * if it is set here then we've found a bug.
652                  */
653         } while (WARN_ON_ONCE(running));
654 }
655
656 static void dev_init_scheduler_queue(struct net_device *dev,
657                                      struct netdev_queue *dev_queue,
658                                      void *_qdisc)
659 {
660         struct Qdisc *qdisc = _qdisc;
661
662         dev_queue->qdisc = qdisc;
663         dev_queue->qdisc_sleeping = qdisc;
664 }
665
666 void dev_init_scheduler(struct net_device *dev)
667 {
668         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
669         dev_init_scheduler_queue(dev, &dev->rx_queue, NULL);
670
671         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
672 }
673
674 static void shutdown_scheduler_queue(struct net_device *dev,
675                                      struct netdev_queue *dev_queue,
676                                      void *_qdisc_default)
677 {
678         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
679         struct Qdisc *qdisc_default = _qdisc_default;
680
681         if (qdisc) {
682                 spinlock_t *root_lock = qdisc_root_lock(qdisc);
683
684                 dev_queue->qdisc = qdisc_default;
685                 dev_queue->qdisc_sleeping = qdisc_default;
686
687                 spin_lock(root_lock);
688                 qdisc_destroy(qdisc);
689                 spin_unlock(root_lock);
690         }
691 }
692
693 void dev_shutdown(struct net_device *dev)
694 {
695         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
696         shutdown_scheduler_queue(dev, &dev->rx_queue, NULL);
697         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
698 }