Merge master.kernel.org:/home/rmk/linux-2.6-serial
[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 <asm/uaccess.h>
15 #include <asm/system.h>
16 #include <linux/bitops.h>
17 #include <linux/config.h>
18 #include <linux/module.h>
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/string.h>
23 #include <linux/mm.h>
24 #include <linux/socket.h>
25 #include <linux/sockios.h>
26 #include <linux/in.h>
27 #include <linux/errno.h>
28 #include <linux/interrupt.h>
29 #include <linux/netdevice.h>
30 #include <linux/skbuff.h>
31 #include <linux/rtnetlink.h>
32 #include <linux/init.h>
33 #include <linux/rcupdate.h>
34 #include <linux/list.h>
35 #include <net/sock.h>
36 #include <net/pkt_sched.h>
37
38 /* Main transmission queue. */
39
40 /* Main qdisc structure lock. 
41
42    However, modifications
43    to data, participating in scheduling must be additionally
44    protected with dev->queue_lock spinlock.
45
46    The idea is the following:
47    - enqueue, dequeue are serialized via top level device
48      spinlock dev->queue_lock.
49    - tree walking is protected by read_lock_bh(qdisc_tree_lock)
50      and this lock is used only in process context.
51    - updates to tree are made under rtnl semaphore or
52      from softirq context (__qdisc_destroy rcu-callback)
53      hence this lock needs local bh disabling.
54
55    qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
56  */
57 DEFINE_RWLOCK(qdisc_tree_lock);
58
59 void qdisc_lock_tree(struct net_device *dev)
60 {
61         write_lock_bh(&qdisc_tree_lock);
62         spin_lock_bh(&dev->queue_lock);
63 }
64
65 void qdisc_unlock_tree(struct net_device *dev)
66 {
67         spin_unlock_bh(&dev->queue_lock);
68         write_unlock_bh(&qdisc_tree_lock);
69 }
70
71 /* 
72    dev->queue_lock serializes queue accesses for this device
73    AND dev->qdisc pointer itself.
74
75    dev->xmit_lock serializes accesses to device driver.
76
77    dev->queue_lock and dev->xmit_lock are mutually exclusive,
78    if one is grabbed, another must be free.
79  */
80
81
82 /* Kick device.
83    Note, that this procedure can be called by a watchdog timer, so that
84    we do not check dev->tbusy flag here.
85
86    Returns:  0  - queue is empty.
87             >0  - queue is not empty, but throttled.
88             <0  - queue is not empty. Device is throttled, if dev->tbusy != 0.
89
90    NOTE: Called under dev->queue_lock with locally disabled BH.
91 */
92
93 int qdisc_restart(struct net_device *dev)
94 {
95         struct Qdisc *q = dev->qdisc;
96         struct sk_buff *skb;
97
98         /* Dequeue packet */
99         if ((skb = q->dequeue(q)) != NULL) {
100                 unsigned nolock = (dev->features & NETIF_F_LLTX);
101                 /*
102                  * When the driver has LLTX set it does its own locking
103                  * in start_xmit. No need to add additional overhead by
104                  * locking again. These checks are worth it because
105                  * even uncongested locks can be quite expensive.
106                  * The driver can do trylock like here too, in case
107                  * of lock congestion it should return -1 and the packet
108                  * will be requeued.
109                  */
110                 if (!nolock) {
111                         if (!spin_trylock(&dev->xmit_lock)) {
112                         collision:
113                                 /* So, someone grabbed the driver. */
114                                 
115                                 /* It may be transient configuration error,
116                                    when hard_start_xmit() recurses. We detect
117                                    it by checking xmit owner and drop the
118                                    packet when deadloop is detected.
119                                 */
120                                 if (dev->xmit_lock_owner == smp_processor_id()) {
121                                         kfree_skb(skb);
122                                         if (net_ratelimit())
123                                                 printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
124                                         return -1;
125                                 }
126                                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
127                                 goto requeue;
128                         }
129                         /* Remember that the driver is grabbed by us. */
130                         dev->xmit_lock_owner = smp_processor_id();
131                 }
132                 
133                 {
134                         /* And release queue */
135                         spin_unlock(&dev->queue_lock);
136
137                         if (!netif_queue_stopped(dev)) {
138                                 int ret;
139                                 if (netdev_nit)
140                                         dev_queue_xmit_nit(skb, dev);
141
142                                 ret = dev->hard_start_xmit(skb, dev);
143                                 if (ret == NETDEV_TX_OK) { 
144                                         if (!nolock) {
145                                                 dev->xmit_lock_owner = -1;
146                                                 spin_unlock(&dev->xmit_lock);
147                                         }
148                                         spin_lock(&dev->queue_lock);
149                                         return -1;
150                                 }
151                                 if (ret == NETDEV_TX_LOCKED && nolock) {
152                                         spin_lock(&dev->queue_lock);
153                                         goto collision; 
154                                 }
155                         }
156
157                         /* NETDEV_TX_BUSY - we need to requeue */
158                         /* Release the driver */
159                         if (!nolock) { 
160                                 dev->xmit_lock_owner = -1;
161                                 spin_unlock(&dev->xmit_lock);
162                         } 
163                         spin_lock(&dev->queue_lock);
164                         q = dev->qdisc;
165                 }
166
167                 /* Device kicked us out :(
168                    This is possible in three cases:
169
170                    0. driver is locked
171                    1. fastroute is enabled
172                    2. device cannot determine busy state
173                       before start of transmission (f.e. dialout)
174                    3. device is buggy (ppp)
175                  */
176
177 requeue:
178                 q->ops->requeue(skb, q);
179                 netif_schedule(dev);
180                 return 1;
181         }
182         BUG_ON((int) q->q.qlen < 0);
183         return q->q.qlen;
184 }
185
186 static void dev_watchdog(unsigned long arg)
187 {
188         struct net_device *dev = (struct net_device *)arg;
189
190         spin_lock(&dev->xmit_lock);
191         if (dev->qdisc != &noop_qdisc) {
192                 if (netif_device_present(dev) &&
193                     netif_running(dev) &&
194                     netif_carrier_ok(dev)) {
195                         if (netif_queue_stopped(dev) &&
196                             (jiffies - dev->trans_start) > dev->watchdog_timeo) {
197                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n", dev->name);
198                                 dev->tx_timeout(dev);
199                         }
200                         if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
201                                 dev_hold(dev);
202                 }
203         }
204         spin_unlock(&dev->xmit_lock);
205
206         dev_put(dev);
207 }
208
209 static void dev_watchdog_init(struct net_device *dev)
210 {
211         init_timer(&dev->watchdog_timer);
212         dev->watchdog_timer.data = (unsigned long)dev;
213         dev->watchdog_timer.function = dev_watchdog;
214 }
215
216 void __netdev_watchdog_up(struct net_device *dev)
217 {
218         if (dev->tx_timeout) {
219                 if (dev->watchdog_timeo <= 0)
220                         dev->watchdog_timeo = 5*HZ;
221                 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
222                         dev_hold(dev);
223         }
224 }
225
226 static void dev_watchdog_up(struct net_device *dev)
227 {
228         spin_lock_bh(&dev->xmit_lock);
229         __netdev_watchdog_up(dev);
230         spin_unlock_bh(&dev->xmit_lock);
231 }
232
233 static void dev_watchdog_down(struct net_device *dev)
234 {
235         spin_lock_bh(&dev->xmit_lock);
236         if (del_timer(&dev->watchdog_timer))
237                 __dev_put(dev);
238         spin_unlock_bh(&dev->xmit_lock);
239 }
240
241 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
242    under all circumstances. It is difficult to invent anything faster or
243    cheaper.
244  */
245
246 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
247 {
248         kfree_skb(skb);
249         return NET_XMIT_CN;
250 }
251
252 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
253 {
254         return NULL;
255 }
256
257 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
258 {
259         if (net_ratelimit())
260                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
261                        skb->dev->name);
262         kfree_skb(skb);
263         return NET_XMIT_CN;
264 }
265
266 struct Qdisc_ops noop_qdisc_ops = {
267         .id             =       "noop",
268         .priv_size      =       0,
269         .enqueue        =       noop_enqueue,
270         .dequeue        =       noop_dequeue,
271         .requeue        =       noop_requeue,
272         .owner          =       THIS_MODULE,
273 };
274
275 struct Qdisc noop_qdisc = {
276         .enqueue        =       noop_enqueue,
277         .dequeue        =       noop_dequeue,
278         .flags          =       TCQ_F_BUILTIN,
279         .ops            =       &noop_qdisc_ops,        
280         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
281 };
282
283 static struct Qdisc_ops noqueue_qdisc_ops = {
284         .id             =       "noqueue",
285         .priv_size      =       0,
286         .enqueue        =       noop_enqueue,
287         .dequeue        =       noop_dequeue,
288         .requeue        =       noop_requeue,
289         .owner          =       THIS_MODULE,
290 };
291
292 static struct Qdisc noqueue_qdisc = {
293         .enqueue        =       NULL,
294         .dequeue        =       noop_dequeue,
295         .flags          =       TCQ_F_BUILTIN,
296         .ops            =       &noqueue_qdisc_ops,
297         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
298 };
299
300
301 static const u8 prio2band[TC_PRIO_MAX+1] =
302         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
303
304 /* 3-band FIFO queue: old style, but should be a bit faster than
305    generic prio+fifo combination.
306  */
307
308 #define PFIFO_FAST_BANDS 3
309
310 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
311                                              struct Qdisc *qdisc)
312 {
313         struct sk_buff_head *list = qdisc_priv(qdisc);
314         return list + prio2band[skb->priority & TC_PRIO_MAX];
315 }
316
317 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
318 {
319         struct sk_buff_head *list = prio2list(skb, qdisc);
320
321         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
322                 qdisc->q.qlen++;
323                 return __qdisc_enqueue_tail(skb, qdisc, list);
324         }
325
326         return qdisc_drop(skb, qdisc);
327 }
328
329 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
330 {
331         int prio;
332         struct sk_buff_head *list = qdisc_priv(qdisc);
333
334         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
335                 if (!skb_queue_empty(list + prio)) {
336                         qdisc->q.qlen--;
337                         return __qdisc_dequeue_head(qdisc, list + prio);
338                 }
339         }
340
341         return NULL;
342 }
343
344 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
345 {
346         qdisc->q.qlen++;
347         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
348 }
349
350 static void pfifo_fast_reset(struct Qdisc* qdisc)
351 {
352         int prio;
353         struct sk_buff_head *list = qdisc_priv(qdisc);
354
355         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
356                 __qdisc_reset_queue(qdisc, list + prio);
357
358         qdisc->qstats.backlog = 0;
359         qdisc->q.qlen = 0;
360 }
361
362 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
363 {
364         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
365
366         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
367         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
368         return skb->len;
369
370 rtattr_failure:
371         return -1;
372 }
373
374 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
375 {
376         int prio;
377         struct sk_buff_head *list = qdisc_priv(qdisc);
378
379         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
380                 skb_queue_head_init(list + prio);
381
382         return 0;
383 }
384
385 static struct Qdisc_ops pfifo_fast_ops = {
386         .id             =       "pfifo_fast",
387         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
388         .enqueue        =       pfifo_fast_enqueue,
389         .dequeue        =       pfifo_fast_dequeue,
390         .requeue        =       pfifo_fast_requeue,
391         .init           =       pfifo_fast_init,
392         .reset          =       pfifo_fast_reset,
393         .dump           =       pfifo_fast_dump,
394         .owner          =       THIS_MODULE,
395 };
396
397 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
398 {
399         void *p;
400         struct Qdisc *sch;
401         unsigned int size;
402         int err = -ENOBUFS;
403
404         /* ensure that the Qdisc and the private data are 32-byte aligned */
405         size = QDISC_ALIGN(sizeof(*sch));
406         size += ops->priv_size + (QDISC_ALIGNTO - 1);
407
408         p = kmalloc(size, GFP_KERNEL);
409         if (!p)
410                 goto errout;
411         memset(p, 0, size);
412         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
413         sch->padded = (char *) sch - (char *) p;
414
415         INIT_LIST_HEAD(&sch->list);
416         skb_queue_head_init(&sch->q);
417         sch->ops = ops;
418         sch->enqueue = ops->enqueue;
419         sch->dequeue = ops->dequeue;
420         sch->dev = dev;
421         dev_hold(dev);
422         sch->stats_lock = &dev->queue_lock;
423         atomic_set(&sch->refcnt, 1);
424
425         return sch;
426 errout:
427         return ERR_PTR(-err);
428 }
429
430 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
431 {
432         struct Qdisc *sch;
433         
434         sch = qdisc_alloc(dev, ops);
435         if (IS_ERR(sch))
436                 goto errout;
437
438         if (!ops->init || ops->init(sch, NULL) == 0)
439                 return sch;
440
441 errout:
442         return NULL;
443 }
444
445 /* Under dev->queue_lock and BH! */
446
447 void qdisc_reset(struct Qdisc *qdisc)
448 {
449         struct Qdisc_ops *ops = qdisc->ops;
450
451         if (ops->reset)
452                 ops->reset(qdisc);
453 }
454
455 /* this is the rcu callback function to clean up a qdisc when there 
456  * are no further references to it */
457
458 static void __qdisc_destroy(struct rcu_head *head)
459 {
460         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
461         struct Qdisc_ops  *ops = qdisc->ops;
462
463 #ifdef CONFIG_NET_ESTIMATOR
464         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
465 #endif
466         write_lock(&qdisc_tree_lock);
467         if (ops->reset)
468                 ops->reset(qdisc);
469         if (ops->destroy)
470                 ops->destroy(qdisc);
471         write_unlock(&qdisc_tree_lock);
472         module_put(ops->owner);
473
474         dev_put(qdisc->dev);
475         kfree((char *) qdisc - qdisc->padded);
476 }
477
478 /* Under dev->queue_lock and BH! */
479
480 void qdisc_destroy(struct Qdisc *qdisc)
481 {
482         struct list_head cql = LIST_HEAD_INIT(cql);
483         struct Qdisc *cq, *q, *n;
484
485         if (qdisc->flags & TCQ_F_BUILTIN ||
486                 !atomic_dec_and_test(&qdisc->refcnt))
487                 return;
488
489         if (!list_empty(&qdisc->list)) {
490                 if (qdisc->ops->cl_ops == NULL)
491                         list_del(&qdisc->list);
492                 else
493                         list_move(&qdisc->list, &cql);
494         }
495
496         /* unlink inner qdiscs from dev->qdisc_list immediately */
497         list_for_each_entry(cq, &cql, list)
498                 list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
499                         if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
500                                 if (q->ops->cl_ops == NULL)
501                                         list_del_init(&q->list);
502                                 else
503                                         list_move_tail(&q->list, &cql);
504                         }
505         list_for_each_entry_safe(cq, n, &cql, list)
506                 list_del_init(&cq->list);
507
508         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
509 }
510
511 void dev_activate(struct net_device *dev)
512 {
513         /* No queueing discipline is attached to device;
514            create default one i.e. pfifo_fast for devices,
515            which need queueing and noqueue_qdisc for
516            virtual interfaces
517          */
518
519         if (dev->qdisc_sleeping == &noop_qdisc) {
520                 struct Qdisc *qdisc;
521                 if (dev->tx_queue_len) {
522                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
523                         if (qdisc == NULL) {
524                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
525                                 return;
526                         }
527                         write_lock_bh(&qdisc_tree_lock);
528                         list_add_tail(&qdisc->list, &dev->qdisc_list);
529                         write_unlock_bh(&qdisc_tree_lock);
530                 } else {
531                         qdisc =  &noqueue_qdisc;
532                 }
533                 write_lock_bh(&qdisc_tree_lock);
534                 dev->qdisc_sleeping = qdisc;
535                 write_unlock_bh(&qdisc_tree_lock);
536         }
537
538         if (!netif_carrier_ok(dev))
539                 /* Delay activation until next carrier-on event */
540                 return;
541
542         spin_lock_bh(&dev->queue_lock);
543         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
544         if (dev->qdisc != &noqueue_qdisc) {
545                 dev->trans_start = jiffies;
546                 dev_watchdog_up(dev);
547         }
548         spin_unlock_bh(&dev->queue_lock);
549 }
550
551 void dev_deactivate(struct net_device *dev)
552 {
553         struct Qdisc *qdisc;
554
555         spin_lock_bh(&dev->queue_lock);
556         qdisc = dev->qdisc;
557         dev->qdisc = &noop_qdisc;
558
559         qdisc_reset(qdisc);
560
561         spin_unlock_bh(&dev->queue_lock);
562
563         dev_watchdog_down(dev);
564
565         while (test_bit(__LINK_STATE_SCHED, &dev->state))
566                 yield();
567
568         spin_unlock_wait(&dev->xmit_lock);
569 }
570
571 void dev_init_scheduler(struct net_device *dev)
572 {
573         qdisc_lock_tree(dev);
574         dev->qdisc = &noop_qdisc;
575         dev->qdisc_sleeping = &noop_qdisc;
576         INIT_LIST_HEAD(&dev->qdisc_list);
577         qdisc_unlock_tree(dev);
578
579         dev_watchdog_init(dev);
580 }
581
582 void dev_shutdown(struct net_device *dev)
583 {
584         struct Qdisc *qdisc;
585
586         qdisc_lock_tree(dev);
587         qdisc = dev->qdisc_sleeping;
588         dev->qdisc = &noop_qdisc;
589         dev->qdisc_sleeping = &noop_qdisc;
590         qdisc_destroy(qdisc);
591 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
592         if ((qdisc = dev->qdisc_ingress) != NULL) {
593                 dev->qdisc_ingress = NULL;
594                 qdisc_destroy(qdisc);
595         }
596 #endif
597         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
598         qdisc_unlock_tree(dev);
599 }
600
601 EXPORT_SYMBOL(__netdev_watchdog_up);
602 EXPORT_SYMBOL(noop_qdisc);
603 EXPORT_SYMBOL(noop_qdisc_ops);
604 EXPORT_SYMBOL(qdisc_create_dflt);
605 EXPORT_SYMBOL(qdisc_alloc);
606 EXPORT_SYMBOL(qdisc_destroy);
607 EXPORT_SYMBOL(qdisc_reset);
608 EXPORT_SYMBOL(qdisc_restart);
609 EXPORT_SYMBOL(qdisc_lock_tree);
610 EXPORT_SYMBOL(qdisc_unlock_tree);