2 * net/sched/sch_generic.c Generic packet scheduler routines.
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.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
14 #include <asm/uaccess.h>
15 #include <asm/system.h>
16 #include <linux/bitops.h>
17 #include <linux/module.h>
18 #include <linux/types.h>
19 #include <linux/kernel.h>
20 #include <linux/sched.h>
21 #include <linux/string.h>
23 #include <linux/socket.h>
24 #include <linux/sockios.h>
26 #include <linux/errno.h>
27 #include <linux/interrupt.h>
28 #include <linux/netdevice.h>
29 #include <linux/skbuff.h>
30 #include <linux/rtnetlink.h>
31 #include <linux/init.h>
32 #include <linux/rcupdate.h>
33 #include <linux/list.h>
35 #include <net/pkt_sched.h>
37 /* Main transmission queue. */
39 /* Main qdisc structure lock.
41 However, modifications
42 to data, participating in scheduling must be additionally
43 protected with dev->queue_lock spinlock.
45 The idea is the following:
46 - enqueue, dequeue are serialized via top level device
47 spinlock dev->queue_lock.
48 - tree walking is protected by read_lock_bh(qdisc_tree_lock)
49 and this lock is used only in process context.
50 - updates to tree are made under rtnl semaphore or
51 from softirq context (__qdisc_destroy rcu-callback)
52 hence this lock needs local bh disabling.
54 qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
56 DEFINE_RWLOCK(qdisc_tree_lock
);
58 void qdisc_lock_tree(struct net_device
*dev
)
60 write_lock_bh(&qdisc_tree_lock
);
61 spin_lock_bh(&dev
->queue_lock
);
64 void qdisc_unlock_tree(struct net_device
*dev
)
66 spin_unlock_bh(&dev
->queue_lock
);
67 write_unlock_bh(&qdisc_tree_lock
);
71 dev->queue_lock serializes queue accesses for this device
72 AND dev->qdisc pointer itself.
74 netif_tx_lock serializes accesses to device driver.
76 dev->queue_lock and netif_tx_lock are mutually exclusive,
77 if one is grabbed, another must be free.
82 Note, that this procedure can be called by a watchdog timer, so that
83 we do not check dev->tbusy flag here.
85 Returns: 0 - queue is empty.
86 >0 - queue is not empty, but throttled.
87 <0 - queue is not empty. Device is throttled, if dev->tbusy != 0.
89 NOTE: Called under dev->queue_lock with locally disabled BH.
92 static inline int qdisc_restart(struct net_device
*dev
)
94 struct Qdisc
*q
= dev
->qdisc
;
98 if (((skb
= dev
->gso_skb
)) || ((skb
= q
->dequeue(q
)))) {
99 unsigned nolock
= (dev
->features
& NETIF_F_LLTX
);
104 * When the driver has LLTX set it does its own locking
105 * in start_xmit. No need to add additional overhead by
106 * locking again. These checks are worth it because
107 * even uncongested locks can be quite expensive.
108 * The driver can do trylock like here too, in case
109 * of lock congestion it should return -1 and the packet
113 if (!netif_tx_trylock(dev
)) {
115 /* So, someone grabbed the driver. */
117 /* It may be transient configuration error,
118 when hard_start_xmit() recurses. We detect
119 it by checking xmit owner and drop the
120 packet when deadloop is detected.
122 if (dev
->xmit_lock_owner
== smp_processor_id()) {
125 printk(KERN_DEBUG
"Dead loop on netdevice %s, fix it urgently!\n", dev
->name
);
128 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
134 /* And release queue */
135 spin_unlock(&dev
->queue_lock
);
137 if (!netif_queue_stopped(dev
)) {
140 ret
= dev_hard_start_xmit(skb
, dev
);
141 if (ret
== NETDEV_TX_OK
) {
143 netif_tx_unlock(dev
);
145 spin_lock(&dev
->queue_lock
);
148 if (ret
== NETDEV_TX_LOCKED
&& nolock
) {
149 spin_lock(&dev
->queue_lock
);
154 /* NETDEV_TX_BUSY - we need to requeue */
155 /* Release the driver */
157 netif_tx_unlock(dev
);
159 spin_lock(&dev
->queue_lock
);
163 /* Device kicked us out :(
164 This is possible in three cases:
167 1. fastroute is enabled
168 2. device cannot determine busy state
169 before start of transmission (f.e. dialout)
170 3. device is buggy (ppp)
177 q
->ops
->requeue(skb
, q
);
181 BUG_ON((int) q
->q
.qlen
< 0);
185 void __qdisc_run(struct net_device
*dev
)
187 if (unlikely(dev
->qdisc
== &noop_qdisc
))
190 while (qdisc_restart(dev
) < 0 && !netif_queue_stopped(dev
))
194 clear_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
197 static void dev_watchdog(unsigned long arg
)
199 struct net_device
*dev
= (struct net_device
*)arg
;
202 if (dev
->qdisc
!= &noop_qdisc
) {
203 if (netif_device_present(dev
) &&
204 netif_running(dev
) &&
205 netif_carrier_ok(dev
)) {
206 if (netif_queue_stopped(dev
) &&
207 time_after(jiffies
, dev
->trans_start
+ dev
->watchdog_timeo
)) {
209 printk(KERN_INFO
"NETDEV WATCHDOG: %s: transmit timed out\n",
211 dev
->tx_timeout(dev
);
213 if (!mod_timer(&dev
->watchdog_timer
, jiffies
+ dev
->watchdog_timeo
))
217 netif_tx_unlock(dev
);
222 static void dev_watchdog_init(struct net_device
*dev
)
224 init_timer(&dev
->watchdog_timer
);
225 dev
->watchdog_timer
.data
= (unsigned long)dev
;
226 dev
->watchdog_timer
.function
= dev_watchdog
;
229 void __netdev_watchdog_up(struct net_device
*dev
)
231 if (dev
->tx_timeout
) {
232 if (dev
->watchdog_timeo
<= 0)
233 dev
->watchdog_timeo
= 5*HZ
;
234 if (!mod_timer(&dev
->watchdog_timer
, jiffies
+ dev
->watchdog_timeo
))
239 static void dev_watchdog_up(struct net_device
*dev
)
241 netif_tx_lock_bh(dev
);
242 __netdev_watchdog_up(dev
);
243 netif_tx_unlock_bh(dev
);
246 static void dev_watchdog_down(struct net_device
*dev
)
248 netif_tx_lock_bh(dev
);
249 if (del_timer(&dev
->watchdog_timer
))
251 netif_tx_unlock_bh(dev
);
254 void netif_carrier_on(struct net_device
*dev
)
256 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
257 linkwatch_fire_event(dev
);
258 if (netif_running(dev
))
259 __netdev_watchdog_up(dev
);
262 void netif_carrier_off(struct net_device
*dev
)
264 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
265 linkwatch_fire_event(dev
);
268 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
269 under all circumstances. It is difficult to invent anything faster or
273 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
279 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
284 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
287 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
293 struct Qdisc_ops noop_qdisc_ops
= {
296 .enqueue
= noop_enqueue
,
297 .dequeue
= noop_dequeue
,
298 .requeue
= noop_requeue
,
299 .owner
= THIS_MODULE
,
302 struct Qdisc noop_qdisc
= {
303 .enqueue
= noop_enqueue
,
304 .dequeue
= noop_dequeue
,
305 .flags
= TCQ_F_BUILTIN
,
306 .ops
= &noop_qdisc_ops
,
307 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
310 static struct Qdisc_ops noqueue_qdisc_ops
= {
313 .enqueue
= noop_enqueue
,
314 .dequeue
= noop_dequeue
,
315 .requeue
= noop_requeue
,
316 .owner
= THIS_MODULE
,
319 static struct Qdisc noqueue_qdisc
= {
321 .dequeue
= noop_dequeue
,
322 .flags
= TCQ_F_BUILTIN
,
323 .ops
= &noqueue_qdisc_ops
,
324 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
328 static const u8 prio2band
[TC_PRIO_MAX
+1] =
329 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
331 /* 3-band FIFO queue: old style, but should be a bit faster than
332 generic prio+fifo combination.
335 #define PFIFO_FAST_BANDS 3
337 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
340 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
341 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
344 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
346 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
348 if (skb_queue_len(list
) < qdisc
->dev
->tx_queue_len
) {
350 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
353 return qdisc_drop(skb
, qdisc
);
356 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
359 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
361 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
362 if (!skb_queue_empty(list
+ prio
)) {
364 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
371 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
374 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
377 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
380 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
382 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
383 __qdisc_reset_queue(qdisc
, list
+ prio
);
385 qdisc
->qstats
.backlog
= 0;
389 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
391 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
393 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
394 RTA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
401 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct rtattr
*opt
)
404 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
406 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
407 skb_queue_head_init(list
+ prio
);
412 static struct Qdisc_ops pfifo_fast_ops
= {
414 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
415 .enqueue
= pfifo_fast_enqueue
,
416 .dequeue
= pfifo_fast_dequeue
,
417 .requeue
= pfifo_fast_requeue
,
418 .init
= pfifo_fast_init
,
419 .reset
= pfifo_fast_reset
,
420 .dump
= pfifo_fast_dump
,
421 .owner
= THIS_MODULE
,
424 struct Qdisc
*qdisc_alloc(struct net_device
*dev
, struct Qdisc_ops
*ops
)
431 /* ensure that the Qdisc and the private data are 32-byte aligned */
432 size
= QDISC_ALIGN(sizeof(*sch
));
433 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
435 p
= kmalloc(size
, GFP_KERNEL
);
439 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
440 sch
->padded
= (char *) sch
- (char *) p
;
442 INIT_LIST_HEAD(&sch
->list
);
443 skb_queue_head_init(&sch
->q
);
445 sch
->enqueue
= ops
->enqueue
;
446 sch
->dequeue
= ops
->dequeue
;
449 sch
->stats_lock
= &dev
->queue_lock
;
450 atomic_set(&sch
->refcnt
, 1);
454 return ERR_PTR(-err
);
457 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
, struct Qdisc_ops
*ops
)
461 sch
= qdisc_alloc(dev
, ops
);
465 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
473 /* Under dev->queue_lock and BH! */
475 void qdisc_reset(struct Qdisc
*qdisc
)
477 struct Qdisc_ops
*ops
= qdisc
->ops
;
483 /* this is the rcu callback function to clean up a qdisc when there
484 * are no further references to it */
486 static void __qdisc_destroy(struct rcu_head
*head
)
488 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
489 struct Qdisc_ops
*ops
= qdisc
->ops
;
491 #ifdef CONFIG_NET_ESTIMATOR
492 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
494 write_lock(&qdisc_tree_lock
);
499 write_unlock(&qdisc_tree_lock
);
500 module_put(ops
->owner
);
503 kfree((char *) qdisc
- qdisc
->padded
);
506 /* Under dev->queue_lock and BH! */
508 void qdisc_destroy(struct Qdisc
*qdisc
)
510 struct list_head cql
= LIST_HEAD_INIT(cql
);
511 struct Qdisc
*cq
, *q
, *n
;
513 if (qdisc
->flags
& TCQ_F_BUILTIN
||
514 !atomic_dec_and_test(&qdisc
->refcnt
))
517 if (!list_empty(&qdisc
->list
)) {
518 if (qdisc
->ops
->cl_ops
== NULL
)
519 list_del(&qdisc
->list
);
521 list_move(&qdisc
->list
, &cql
);
524 /* unlink inner qdiscs from dev->qdisc_list immediately */
525 list_for_each_entry(cq
, &cql
, list
)
526 list_for_each_entry_safe(q
, n
, &qdisc
->dev
->qdisc_list
, list
)
527 if (TC_H_MAJ(q
->parent
) == TC_H_MAJ(cq
->handle
)) {
528 if (q
->ops
->cl_ops
== NULL
)
529 list_del_init(&q
->list
);
531 list_move_tail(&q
->list
, &cql
);
533 list_for_each_entry_safe(cq
, n
, &cql
, list
)
534 list_del_init(&cq
->list
);
536 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
539 void dev_activate(struct net_device
*dev
)
541 /* No queueing discipline is attached to device;
542 create default one i.e. pfifo_fast for devices,
543 which need queueing and noqueue_qdisc for
547 if (dev
->qdisc_sleeping
== &noop_qdisc
) {
549 if (dev
->tx_queue_len
) {
550 qdisc
= qdisc_create_dflt(dev
, &pfifo_fast_ops
);
552 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
555 write_lock_bh(&qdisc_tree_lock
);
556 list_add_tail(&qdisc
->list
, &dev
->qdisc_list
);
557 write_unlock_bh(&qdisc_tree_lock
);
559 qdisc
= &noqueue_qdisc
;
561 write_lock_bh(&qdisc_tree_lock
);
562 dev
->qdisc_sleeping
= qdisc
;
563 write_unlock_bh(&qdisc_tree_lock
);
566 if (!netif_carrier_ok(dev
))
567 /* Delay activation until next carrier-on event */
570 spin_lock_bh(&dev
->queue_lock
);
571 rcu_assign_pointer(dev
->qdisc
, dev
->qdisc_sleeping
);
572 if (dev
->qdisc
!= &noqueue_qdisc
) {
573 dev
->trans_start
= jiffies
;
574 dev_watchdog_up(dev
);
576 spin_unlock_bh(&dev
->queue_lock
);
579 void dev_deactivate(struct net_device
*dev
)
583 spin_lock_bh(&dev
->queue_lock
);
585 dev
->qdisc
= &noop_qdisc
;
589 spin_unlock_bh(&dev
->queue_lock
);
591 dev_watchdog_down(dev
);
593 /* Wait for outstanding dev_queue_xmit calls. */
596 /* Wait for outstanding qdisc_run calls. */
597 while (test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
))
601 kfree_skb(dev
->gso_skb
);
606 void dev_init_scheduler(struct net_device
*dev
)
608 qdisc_lock_tree(dev
);
609 dev
->qdisc
= &noop_qdisc
;
610 dev
->qdisc_sleeping
= &noop_qdisc
;
611 INIT_LIST_HEAD(&dev
->qdisc_list
);
612 qdisc_unlock_tree(dev
);
614 dev_watchdog_init(dev
);
617 void dev_shutdown(struct net_device
*dev
)
621 qdisc_lock_tree(dev
);
622 qdisc
= dev
->qdisc_sleeping
;
623 dev
->qdisc
= &noop_qdisc
;
624 dev
->qdisc_sleeping
= &noop_qdisc
;
625 qdisc_destroy(qdisc
);
626 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
627 if ((qdisc
= dev
->qdisc_ingress
) != NULL
) {
628 dev
->qdisc_ingress
= NULL
;
629 qdisc_destroy(qdisc
);
632 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));
633 qdisc_unlock_tree(dev
);
636 EXPORT_SYMBOL(__netdev_watchdog_up
);
637 EXPORT_SYMBOL(netif_carrier_on
);
638 EXPORT_SYMBOL(netif_carrier_off
);
639 EXPORT_SYMBOL(noop_qdisc
);
640 EXPORT_SYMBOL(noop_qdisc_ops
);
641 EXPORT_SYMBOL(qdisc_create_dflt
);
642 EXPORT_SYMBOL(qdisc_alloc
);
643 EXPORT_SYMBOL(qdisc_destroy
);
644 EXPORT_SYMBOL(qdisc_reset
);
645 EXPORT_SYMBOL(qdisc_lock_tree
);
646 EXPORT_SYMBOL(qdisc_unlock_tree
);