Merge branch 'for-2.6.24' of master.kernel.org:/pub/scm/linux/kernel/git/jwboyer...
[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  * dev->queue_lock spinlock.
33  *
34  * The idea is the following:
35  * - enqueue, dequeue are serialized via top level device
36  *   spinlock dev->queue_lock.
37  * - ingress filtering is serialized via top level device
38  *   spinlock dev->ingress_lock.
39  * - updates to tree and tree walking are only done under the rtnl mutex.
40  */
41
42 void qdisc_lock_tree(struct net_device *dev)
43 {
44         spin_lock_bh(&dev->queue_lock);
45         spin_lock(&dev->ingress_lock);
46 }
47
48 void qdisc_unlock_tree(struct net_device *dev)
49 {
50         spin_unlock(&dev->ingress_lock);
51         spin_unlock_bh(&dev->queue_lock);
52 }
53
54 static inline int qdisc_qlen(struct Qdisc *q)
55 {
56         return q->q.qlen;
57 }
58
59 static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
60                                   struct Qdisc *q)
61 {
62         if (unlikely(skb->next))
63                 dev->gso_skb = skb;
64         else
65                 q->ops->requeue(skb, q);
66
67         netif_schedule(dev);
68         return 0;
69 }
70
71 static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
72                                               struct Qdisc *q)
73 {
74         struct sk_buff *skb;
75
76         if ((skb = dev->gso_skb))
77                 dev->gso_skb = NULL;
78         else
79                 skb = q->dequeue(q);
80
81         return skb;
82 }
83
84 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
85                                            struct net_device *dev,
86                                            struct Qdisc *q)
87 {
88         int ret;
89
90         if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
91                 /*
92                  * Same CPU holding the lock. It may be a transient
93                  * configuration error, when hard_start_xmit() recurses. We
94                  * detect it by checking xmit owner and drop the packet when
95                  * deadloop is detected. Return OK to try the next skb.
96                  */
97                 kfree_skb(skb);
98                 if (net_ratelimit())
99                         printk(KERN_WARNING "Dead loop on netdevice %s, "
100                                "fix it urgently!\n", dev->name);
101                 ret = qdisc_qlen(q);
102         } else {
103                 /*
104                  * Another cpu is holding lock, requeue & delay xmits for
105                  * some time.
106                  */
107                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
108                 ret = dev_requeue_skb(skb, dev, q);
109         }
110
111         return ret;
112 }
113
114 /*
115  * NOTE: Called under dev->queue_lock with locally disabled BH.
116  *
117  * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
118  * device at a time. dev->queue_lock serializes queue accesses for
119  * this device AND dev->qdisc pointer itself.
120  *
121  *  netif_tx_lock serializes accesses to device driver.
122  *
123  *  dev->queue_lock and netif_tx_lock are mutually exclusive,
124  *  if one is grabbed, another must be free.
125  *
126  * Note, that this procedure can be called by a watchdog timer
127  *
128  * Returns to the caller:
129  *                              0  - queue is empty or throttled.
130  *                              >0 - queue is not empty.
131  *
132  */
133 static inline int qdisc_restart(struct net_device *dev)
134 {
135         struct Qdisc *q = dev->qdisc;
136         struct sk_buff *skb;
137         int ret = NETDEV_TX_BUSY;
138
139         /* Dequeue packet */
140         if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
141                 return 0;
142
143
144         /* And release queue */
145         spin_unlock(&dev->queue_lock);
146
147         HARD_TX_LOCK(dev, smp_processor_id());
148         if (!netif_subqueue_stopped(dev, skb))
149                 ret = dev_hard_start_xmit(skb, dev);
150         HARD_TX_UNLOCK(dev);
151
152         spin_lock(&dev->queue_lock);
153         q = dev->qdisc;
154
155         switch (ret) {
156         case NETDEV_TX_OK:
157                 /* Driver sent out skb successfully */
158                 ret = qdisc_qlen(q);
159                 break;
160
161         case NETDEV_TX_LOCKED:
162                 /* Driver try lock failed */
163                 ret = handle_dev_cpu_collision(skb, dev, q);
164                 break;
165
166         default:
167                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
168                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
169                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
170                                dev->name, ret, q->q.qlen);
171
172                 ret = dev_requeue_skb(skb, dev, q);
173                 break;
174         }
175
176         return ret;
177 }
178
179 void __qdisc_run(struct net_device *dev)
180 {
181         do {
182                 if (!qdisc_restart(dev))
183                         break;
184         } while (!netif_queue_stopped(dev));
185
186         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
187 }
188
189 static void dev_watchdog(unsigned long arg)
190 {
191         struct net_device *dev = (struct net_device *)arg;
192
193         netif_tx_lock(dev);
194         if (dev->qdisc != &noop_qdisc) {
195                 if (netif_device_present(dev) &&
196                     netif_running(dev) &&
197                     netif_carrier_ok(dev)) {
198                         if (netif_queue_stopped(dev) &&
199                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
200
201                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
202                                        dev->name);
203                                 dev->tx_timeout(dev);
204                         }
205                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
206                                 dev_hold(dev);
207                 }
208         }
209         netif_tx_unlock(dev);
210
211         dev_put(dev);
212 }
213
214 static void dev_watchdog_init(struct net_device *dev)
215 {
216         init_timer(&dev->watchdog_timer);
217         dev->watchdog_timer.data = (unsigned long)dev;
218         dev->watchdog_timer.function = dev_watchdog;
219 }
220
221 void __netdev_watchdog_up(struct net_device *dev)
222 {
223         if (dev->tx_timeout) {
224                 if (dev->watchdog_timeo <= 0)
225                         dev->watchdog_timeo = 5*HZ;
226                 if (!mod_timer(&dev->watchdog_timer,
227                                round_jiffies(jiffies + dev->watchdog_timeo)))
228                         dev_hold(dev);
229         }
230 }
231
232 static void dev_watchdog_up(struct net_device *dev)
233 {
234         __netdev_watchdog_up(dev);
235 }
236
237 static void dev_watchdog_down(struct net_device *dev)
238 {
239         netif_tx_lock_bh(dev);
240         if (del_timer(&dev->watchdog_timer))
241                 dev_put(dev);
242         netif_tx_unlock_bh(dev);
243 }
244
245 /**
246  *      netif_carrier_on - set carrier
247  *      @dev: network device
248  *
249  * Device has detected that carrier.
250  */
251 void netif_carrier_on(struct net_device *dev)
252 {
253         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
254                 linkwatch_fire_event(dev);
255                 if (netif_running(dev))
256                         __netdev_watchdog_up(dev);
257         }
258 }
259
260 /**
261  *      netif_carrier_off - clear carrier
262  *      @dev: network device
263  *
264  * Device has detected loss of carrier.
265  */
266 void netif_carrier_off(struct net_device *dev)
267 {
268         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
269                 linkwatch_fire_event(dev);
270 }
271
272 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
273    under all circumstances. It is difficult to invent anything faster or
274    cheaper.
275  */
276
277 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
278 {
279         kfree_skb(skb);
280         return NET_XMIT_CN;
281 }
282
283 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
284 {
285         return NULL;
286 }
287
288 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
289 {
290         if (net_ratelimit())
291                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
292                        skb->dev->name);
293         kfree_skb(skb);
294         return NET_XMIT_CN;
295 }
296
297 struct Qdisc_ops noop_qdisc_ops = {
298         .id             =       "noop",
299         .priv_size      =       0,
300         .enqueue        =       noop_enqueue,
301         .dequeue        =       noop_dequeue,
302         .requeue        =       noop_requeue,
303         .owner          =       THIS_MODULE,
304 };
305
306 struct Qdisc noop_qdisc = {
307         .enqueue        =       noop_enqueue,
308         .dequeue        =       noop_dequeue,
309         .flags          =       TCQ_F_BUILTIN,
310         .ops            =       &noop_qdisc_ops,
311         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
312 };
313
314 static struct Qdisc_ops noqueue_qdisc_ops = {
315         .id             =       "noqueue",
316         .priv_size      =       0,
317         .enqueue        =       noop_enqueue,
318         .dequeue        =       noop_dequeue,
319         .requeue        =       noop_requeue,
320         .owner          =       THIS_MODULE,
321 };
322
323 static struct Qdisc noqueue_qdisc = {
324         .enqueue        =       NULL,
325         .dequeue        =       noop_dequeue,
326         .flags          =       TCQ_F_BUILTIN,
327         .ops            =       &noqueue_qdisc_ops,
328         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
329 };
330
331
332 static const u8 prio2band[TC_PRIO_MAX+1] =
333         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
334
335 /* 3-band FIFO queue: old style, but should be a bit faster than
336    generic prio+fifo combination.
337  */
338
339 #define PFIFO_FAST_BANDS 3
340
341 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
342                                              struct Qdisc *qdisc)
343 {
344         struct sk_buff_head *list = qdisc_priv(qdisc);
345         return list + prio2band[skb->priority & TC_PRIO_MAX];
346 }
347
348 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
349 {
350         struct sk_buff_head *list = prio2list(skb, qdisc);
351
352         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
353                 qdisc->q.qlen++;
354                 return __qdisc_enqueue_tail(skb, qdisc, list);
355         }
356
357         return qdisc_drop(skb, qdisc);
358 }
359
360 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
361 {
362         int prio;
363         struct sk_buff_head *list = qdisc_priv(qdisc);
364
365         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
366                 if (!skb_queue_empty(list + prio)) {
367                         qdisc->q.qlen--;
368                         return __qdisc_dequeue_head(qdisc, list + prio);
369                 }
370         }
371
372         return NULL;
373 }
374
375 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
376 {
377         qdisc->q.qlen++;
378         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
379 }
380
381 static void pfifo_fast_reset(struct Qdisc* qdisc)
382 {
383         int prio;
384         struct sk_buff_head *list = qdisc_priv(qdisc);
385
386         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
387                 __qdisc_reset_queue(qdisc, list + prio);
388
389         qdisc->qstats.backlog = 0;
390         qdisc->q.qlen = 0;
391 }
392
393 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
394 {
395         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
396
397         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
398         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
399         return skb->len;
400
401 rtattr_failure:
402         return -1;
403 }
404
405 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
406 {
407         int prio;
408         struct sk_buff_head *list = qdisc_priv(qdisc);
409
410         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
411                 skb_queue_head_init(list + prio);
412
413         return 0;
414 }
415
416 static struct Qdisc_ops pfifo_fast_ops = {
417         .id             =       "pfifo_fast",
418         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
419         .enqueue        =       pfifo_fast_enqueue,
420         .dequeue        =       pfifo_fast_dequeue,
421         .requeue        =       pfifo_fast_requeue,
422         .init           =       pfifo_fast_init,
423         .reset          =       pfifo_fast_reset,
424         .dump           =       pfifo_fast_dump,
425         .owner          =       THIS_MODULE,
426 };
427
428 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
429 {
430         void *p;
431         struct Qdisc *sch;
432         unsigned int size;
433         int err = -ENOBUFS;
434
435         /* ensure that the Qdisc and the private data are 32-byte aligned */
436         size = QDISC_ALIGN(sizeof(*sch));
437         size += ops->priv_size + (QDISC_ALIGNTO - 1);
438
439         p = kzalloc(size, GFP_KERNEL);
440         if (!p)
441                 goto errout;
442         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
443         sch->padded = (char *) sch - (char *) p;
444
445         INIT_LIST_HEAD(&sch->list);
446         skb_queue_head_init(&sch->q);
447         sch->ops = ops;
448         sch->enqueue = ops->enqueue;
449         sch->dequeue = ops->dequeue;
450         sch->dev = dev;
451         dev_hold(dev);
452         atomic_set(&sch->refcnt, 1);
453
454         return sch;
455 errout:
456         return ERR_PTR(-err);
457 }
458
459 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
460                                  unsigned int parentid)
461 {
462         struct Qdisc *sch;
463
464         sch = qdisc_alloc(dev, ops);
465         if (IS_ERR(sch))
466                 goto errout;
467         sch->stats_lock = &dev->queue_lock;
468         sch->parent = parentid;
469
470         if (!ops->init || ops->init(sch, NULL) == 0)
471                 return sch;
472
473         qdisc_destroy(sch);
474 errout:
475         return NULL;
476 }
477
478 /* Under dev->queue_lock and BH! */
479
480 void qdisc_reset(struct Qdisc *qdisc)
481 {
482         struct Qdisc_ops *ops = qdisc->ops;
483
484         if (ops->reset)
485                 ops->reset(qdisc);
486 }
487
488 /* this is the rcu callback function to clean up a qdisc when there
489  * are no further references to it */
490
491 static void __qdisc_destroy(struct rcu_head *head)
492 {
493         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
494         kfree((char *) qdisc - qdisc->padded);
495 }
496
497 /* Under dev->queue_lock and BH! */
498
499 void qdisc_destroy(struct Qdisc *qdisc)
500 {
501         struct Qdisc_ops  *ops = qdisc->ops;
502
503         if (qdisc->flags & TCQ_F_BUILTIN ||
504             !atomic_dec_and_test(&qdisc->refcnt))
505                 return;
506
507         list_del(&qdisc->list);
508         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
509         if (ops->reset)
510                 ops->reset(qdisc);
511         if (ops->destroy)
512                 ops->destroy(qdisc);
513
514         module_put(ops->owner);
515         dev_put(qdisc->dev);
516         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
517 }
518
519 void dev_activate(struct net_device *dev)
520 {
521         /* No queueing discipline is attached to device;
522            create default one i.e. pfifo_fast for devices,
523            which need queueing and noqueue_qdisc for
524            virtual interfaces
525          */
526
527         if (dev->qdisc_sleeping == &noop_qdisc) {
528                 struct Qdisc *qdisc;
529                 if (dev->tx_queue_len) {
530                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
531                                                   TC_H_ROOT);
532                         if (qdisc == NULL) {
533                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
534                                 return;
535                         }
536                         list_add_tail(&qdisc->list, &dev->qdisc_list);
537                 } else {
538                         qdisc =  &noqueue_qdisc;
539                 }
540                 dev->qdisc_sleeping = qdisc;
541         }
542
543         if (!netif_carrier_ok(dev))
544                 /* Delay activation until next carrier-on event */
545                 return;
546
547         spin_lock_bh(&dev->queue_lock);
548         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
549         if (dev->qdisc != &noqueue_qdisc) {
550                 dev->trans_start = jiffies;
551                 dev_watchdog_up(dev);
552         }
553         spin_unlock_bh(&dev->queue_lock);
554 }
555
556 void dev_deactivate(struct net_device *dev)
557 {
558         struct Qdisc *qdisc;
559         struct sk_buff *skb;
560         int running;
561
562         spin_lock_bh(&dev->queue_lock);
563         qdisc = dev->qdisc;
564         dev->qdisc = &noop_qdisc;
565
566         qdisc_reset(qdisc);
567
568         skb = dev->gso_skb;
569         dev->gso_skb = NULL;
570         spin_unlock_bh(&dev->queue_lock);
571
572         kfree_skb(skb);
573
574         dev_watchdog_down(dev);
575
576         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
577         synchronize_rcu();
578
579         /* Wait for outstanding qdisc_run calls. */
580         do {
581                 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
582                         yield();
583
584                 /*
585                  * Double-check inside queue lock to ensure that all effects
586                  * of the queue run are visible when we return.
587                  */
588                 spin_lock_bh(&dev->queue_lock);
589                 running = test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
590                 spin_unlock_bh(&dev->queue_lock);
591
592                 /*
593                  * The running flag should never be set at this point because
594                  * we've already set dev->qdisc to noop_qdisc *inside* the same
595                  * pair of spin locks.  That is, if any qdisc_run starts after
596                  * our initial test it should see the noop_qdisc and then
597                  * clear the RUNNING bit before dropping the queue lock.  So
598                  * if it is set here then we've found a bug.
599                  */
600         } while (WARN_ON_ONCE(running));
601 }
602
603 void dev_init_scheduler(struct net_device *dev)
604 {
605         qdisc_lock_tree(dev);
606         dev->qdisc = &noop_qdisc;
607         dev->qdisc_sleeping = &noop_qdisc;
608         INIT_LIST_HEAD(&dev->qdisc_list);
609         qdisc_unlock_tree(dev);
610
611         dev_watchdog_init(dev);
612 }
613
614 void dev_shutdown(struct net_device *dev)
615 {
616         struct Qdisc *qdisc;
617
618         qdisc_lock_tree(dev);
619         qdisc = dev->qdisc_sleeping;
620         dev->qdisc = &noop_qdisc;
621         dev->qdisc_sleeping = &noop_qdisc;
622         qdisc_destroy(qdisc);
623 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
624         if ((qdisc = dev->qdisc_ingress) != NULL) {
625                 dev->qdisc_ingress = NULL;
626                 qdisc_destroy(qdisc);
627         }
628 #endif
629         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
630         qdisc_unlock_tree(dev);
631 }
632
633 EXPORT_SYMBOL(netif_carrier_on);
634 EXPORT_SYMBOL(netif_carrier_off);
635 EXPORT_SYMBOL(noop_qdisc);
636 EXPORT_SYMBOL(qdisc_create_dflt);
637 EXPORT_SYMBOL(qdisc_destroy);
638 EXPORT_SYMBOL(qdisc_reset);
639 EXPORT_SYMBOL(qdisc_lock_tree);
640 EXPORT_SYMBOL(qdisc_unlock_tree);