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