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 <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>
29 /* Main transmission queue. */
31 /* Modifications to data participating in scheduling must be protected with
32 * queue->lock spinlock.
34 * The idea is the following:
35 * - enqueue, dequeue are serialized via top level device
36 * spinlock queue->lock.
37 * - ingress filtering is serialized via top level device
38 * spinlock dev->rx_queue.lock.
39 * - updates to tree and tree walking are only done under the rtnl mutex.
42 void qdisc_lock_tree(struct net_device
*dev
)
43 __acquires(dev
->tx_queue
.lock
)
44 __acquires(dev
->rx_queue
.lock
)
46 spin_lock_bh(&dev
->tx_queue
.lock
);
47 spin_lock(&dev
->rx_queue
.lock
);
49 EXPORT_SYMBOL(qdisc_lock_tree
);
51 void qdisc_unlock_tree(struct net_device
*dev
)
52 __releases(dev
->rx_queue
.lock
)
53 __releases(dev
->tx_queue
.lock
)
55 spin_unlock(&dev
->rx_queue
.lock
);
56 spin_unlock_bh(&dev
->tx_queue
.lock
);
58 EXPORT_SYMBOL(qdisc_unlock_tree
);
60 static inline int qdisc_qlen(struct Qdisc
*q
)
65 static inline int dev_requeue_skb(struct sk_buff
*skb
,
66 struct netdev_queue
*dev_queue
,
69 if (unlikely(skb
->next
))
70 dev_queue
->gso_skb
= skb
;
72 q
->ops
->requeue(skb
, q
);
74 netif_schedule_queue(dev_queue
);
78 static inline struct sk_buff
*dequeue_skb(struct netdev_queue
*dev_queue
,
83 if ((skb
= dev_queue
->gso_skb
))
84 dev_queue
->gso_skb
= NULL
;
91 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
92 struct netdev_queue
*dev_queue
,
97 if (unlikely(dev_queue
->xmit_lock_owner
== smp_processor_id())) {
99 * Same CPU holding the lock. It may be a transient
100 * configuration error, when hard_start_xmit() recurses. We
101 * detect it by checking xmit owner and drop the packet when
102 * deadloop is detected. Return OK to try the next skb.
106 printk(KERN_WARNING
"Dead loop on netdevice %s, "
107 "fix it urgently!\n", dev_queue
->dev
->name
);
111 * Another cpu is holding lock, requeue & delay xmits for
114 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
115 ret
= dev_requeue_skb(skb
, dev_queue
, q
);
122 * NOTE: Called under queue->lock with locally disabled BH.
124 * __QUEUE_STATE_QDISC_RUNNING guarantees only one CPU can process
125 * this queue at a time. queue->lock serializes queue accesses for
126 * this queue AND txq->qdisc pointer itself.
128 * netif_tx_lock serializes accesses to device driver.
130 * queue->lock and netif_tx_lock are mutually exclusive,
131 * if one is grabbed, another must be free.
133 * Note, that this procedure can be called by a watchdog timer
135 * Returns to the caller:
136 * 0 - queue is empty or throttled.
137 * >0 - queue is not empty.
140 static inline int qdisc_restart(struct netdev_queue
*txq
)
142 struct Qdisc
*q
= txq
->qdisc
;
143 int ret
= NETDEV_TX_BUSY
;
144 struct net_device
*dev
;
148 if (unlikely((skb
= dequeue_skb(txq
, q
)) == NULL
))
152 /* And release queue */
153 spin_unlock(&txq
->lock
);
157 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
158 if (!netif_subqueue_stopped(dev
, skb
))
159 ret
= dev_hard_start_xmit(skb
, dev
);
160 HARD_TX_UNLOCK(dev
, txq
);
162 spin_lock(&txq
->lock
);
167 /* Driver sent out skb successfully */
171 case NETDEV_TX_LOCKED
:
172 /* Driver try lock failed */
173 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
177 /* Driver returned NETDEV_TX_BUSY - requeue skb */
178 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
179 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
180 dev
->name
, ret
, q
->q
.qlen
);
182 ret
= dev_requeue_skb(skb
, txq
, q
);
189 void __qdisc_run(struct netdev_queue
*txq
)
191 struct net_device
*dev
= txq
->dev
;
192 unsigned long start_time
= jiffies
;
194 while (qdisc_restart(txq
)) {
195 if (netif_queue_stopped(dev
))
199 * Postpone processing if
200 * 1. another process needs the CPU;
201 * 2. we've been doing it for too long.
203 if (need_resched() || jiffies
!= start_time
) {
204 netif_schedule_queue(txq
);
209 clear_bit(__QUEUE_STATE_QDISC_RUNNING
, &txq
->state
);
212 static void dev_watchdog(unsigned long arg
)
214 struct net_device
*dev
= (struct net_device
*)arg
;
215 struct netdev_queue
*txq
= &dev
->tx_queue
;
218 if (txq
->qdisc
!= &noop_qdisc
) {
219 if (netif_device_present(dev
) &&
220 netif_running(dev
) &&
221 netif_carrier_ok(dev
)) {
222 if (netif_queue_stopped(dev
) &&
223 time_after(jiffies
, dev
->trans_start
+ dev
->watchdog_timeo
)) {
225 printk(KERN_INFO
"NETDEV WATCHDOG: %s: transmit timed out\n",
227 dev
->tx_timeout(dev
);
230 if (!mod_timer(&dev
->watchdog_timer
, round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
234 netif_tx_unlock(dev
);
239 void __netdev_watchdog_up(struct net_device
*dev
)
241 if (dev
->tx_timeout
) {
242 if (dev
->watchdog_timeo
<= 0)
243 dev
->watchdog_timeo
= 5*HZ
;
244 if (!mod_timer(&dev
->watchdog_timer
,
245 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
250 static void dev_watchdog_up(struct net_device
*dev
)
252 __netdev_watchdog_up(dev
);
255 static void dev_watchdog_down(struct net_device
*dev
)
257 netif_tx_lock_bh(dev
);
258 if (del_timer(&dev
->watchdog_timer
))
260 netif_tx_unlock_bh(dev
);
264 * netif_carrier_on - set carrier
265 * @dev: network device
267 * Device has detected that carrier.
269 void netif_carrier_on(struct net_device
*dev
)
271 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
272 linkwatch_fire_event(dev
);
273 if (netif_running(dev
))
274 __netdev_watchdog_up(dev
);
277 EXPORT_SYMBOL(netif_carrier_on
);
280 * netif_carrier_off - clear carrier
281 * @dev: network device
283 * Device has detected loss of carrier.
285 void netif_carrier_off(struct net_device
*dev
)
287 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
288 linkwatch_fire_event(dev
);
290 EXPORT_SYMBOL(netif_carrier_off
);
292 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
293 under all circumstances. It is difficult to invent anything faster or
297 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
303 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
308 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
311 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
317 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
320 .enqueue
= noop_enqueue
,
321 .dequeue
= noop_dequeue
,
322 .requeue
= noop_requeue
,
323 .owner
= THIS_MODULE
,
326 struct Qdisc noop_qdisc
= {
327 .enqueue
= noop_enqueue
,
328 .dequeue
= noop_dequeue
,
329 .flags
= TCQ_F_BUILTIN
,
330 .ops
= &noop_qdisc_ops
,
331 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
333 EXPORT_SYMBOL(noop_qdisc
);
335 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
338 .enqueue
= noop_enqueue
,
339 .dequeue
= noop_dequeue
,
340 .requeue
= noop_requeue
,
341 .owner
= THIS_MODULE
,
344 static struct Qdisc noqueue_qdisc
= {
346 .dequeue
= noop_dequeue
,
347 .flags
= TCQ_F_BUILTIN
,
348 .ops
= &noqueue_qdisc_ops
,
349 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
353 static const u8 prio2band
[TC_PRIO_MAX
+1] =
354 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
356 /* 3-band FIFO queue: old style, but should be a bit faster than
357 generic prio+fifo combination.
360 #define PFIFO_FAST_BANDS 3
362 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
365 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
366 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
369 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
371 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
373 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
375 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
378 return qdisc_drop(skb
, qdisc
);
381 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
384 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
386 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
387 if (!skb_queue_empty(list
+ prio
)) {
389 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
396 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
399 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
402 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
405 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
407 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
408 __qdisc_reset_queue(qdisc
, list
+ prio
);
410 qdisc
->qstats
.backlog
= 0;
414 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
416 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
418 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
419 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
426 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
429 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
431 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
432 skb_queue_head_init(list
+ prio
);
437 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
439 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
440 .enqueue
= pfifo_fast_enqueue
,
441 .dequeue
= pfifo_fast_dequeue
,
442 .requeue
= pfifo_fast_requeue
,
443 .init
= pfifo_fast_init
,
444 .reset
= pfifo_fast_reset
,
445 .dump
= pfifo_fast_dump
,
446 .owner
= THIS_MODULE
,
449 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
450 struct Qdisc_ops
*ops
)
457 /* ensure that the Qdisc and the private data are 32-byte aligned */
458 size
= QDISC_ALIGN(sizeof(*sch
));
459 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
461 p
= kzalloc(size
, GFP_KERNEL
);
464 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
465 sch
->padded
= (char *) sch
- (char *) p
;
467 INIT_LIST_HEAD(&sch
->list
);
468 skb_queue_head_init(&sch
->q
);
470 sch
->enqueue
= ops
->enqueue
;
471 sch
->dequeue
= ops
->dequeue
;
472 sch
->dev_queue
= dev_queue
;
473 dev_hold(qdisc_dev(sch
));
474 atomic_set(&sch
->refcnt
, 1);
481 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
482 struct netdev_queue
*dev_queue
,
483 struct Qdisc_ops
*ops
,
484 unsigned int parentid
)
488 sch
= qdisc_alloc(dev_queue
, ops
);
491 sch
->parent
= parentid
;
493 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
500 EXPORT_SYMBOL(qdisc_create_dflt
);
502 /* Under queue->lock and BH! */
504 void qdisc_reset(struct Qdisc
*qdisc
)
506 const struct Qdisc_ops
*ops
= qdisc
->ops
;
511 EXPORT_SYMBOL(qdisc_reset
);
513 /* this is the rcu callback function to clean up a qdisc when there
514 * are no further references to it */
516 static void __qdisc_destroy(struct rcu_head
*head
)
518 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
519 kfree((char *) qdisc
- qdisc
->padded
);
522 /* Under queue->lock and BH! */
524 void qdisc_destroy(struct Qdisc
*qdisc
)
526 const struct Qdisc_ops
*ops
= qdisc
->ops
;
528 if (qdisc
->flags
& TCQ_F_BUILTIN
||
529 !atomic_dec_and_test(&qdisc
->refcnt
))
532 list_del(&qdisc
->list
);
533 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
539 module_put(ops
->owner
);
540 dev_put(qdisc_dev(qdisc
));
541 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
543 EXPORT_SYMBOL(qdisc_destroy
);
545 void dev_activate(struct net_device
*dev
)
547 struct netdev_queue
*txq
= &dev
->tx_queue
;
549 /* No queueing discipline is attached to device;
550 create default one i.e. pfifo_fast for devices,
551 which need queueing and noqueue_qdisc for
555 if (txq
->qdisc_sleeping
== &noop_qdisc
) {
557 if (dev
->tx_queue_len
) {
558 qdisc
= qdisc_create_dflt(dev
, txq
,
562 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
565 list_add_tail(&qdisc
->list
, &txq
->qdisc_list
);
567 qdisc
= &noqueue_qdisc
;
569 txq
->qdisc_sleeping
= qdisc
;
572 if (!netif_carrier_ok(dev
))
573 /* Delay activation until next carrier-on event */
576 spin_lock_bh(&txq
->lock
);
577 rcu_assign_pointer(txq
->qdisc
, txq
->qdisc_sleeping
);
578 if (txq
->qdisc
!= &noqueue_qdisc
) {
579 dev
->trans_start
= jiffies
;
580 dev_watchdog_up(dev
);
582 spin_unlock_bh(&txq
->lock
);
585 static void dev_deactivate_queue(struct netdev_queue
*dev_queue
,
586 struct Qdisc
*qdisc_default
)
591 spin_lock_bh(&dev_queue
->lock
);
593 qdisc
= dev_queue
->qdisc
;
595 dev_queue
->qdisc
= qdisc_default
;
598 skb
= dev_queue
->gso_skb
;
599 dev_queue
->gso_skb
= NULL
;
601 spin_unlock_bh(&dev_queue
->lock
);
606 void dev_deactivate(struct net_device
*dev
)
608 struct netdev_queue
*dev_queue
= &dev
->tx_queue
;
611 dev_deactivate_queue(dev_queue
, &noop_qdisc
);
613 dev_watchdog_down(dev
);
615 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
618 /* Wait for outstanding qdisc_run calls. */
620 while (test_bit(__QUEUE_STATE_QDISC_RUNNING
, &dev_queue
->state
))
624 * Double-check inside queue lock to ensure that all effects
625 * of the queue run are visible when we return.
627 spin_lock_bh(&dev_queue
->lock
);
628 running
= test_bit(__QUEUE_STATE_QDISC_RUNNING
,
630 spin_unlock_bh(&dev_queue
->lock
);
633 * The running flag should never be set at this point because
634 * we've already set dev->qdisc to noop_qdisc *inside* the same
635 * pair of spin locks. That is, if any qdisc_run starts after
636 * our initial test it should see the noop_qdisc and then
637 * clear the RUNNING bit before dropping the queue lock. So
638 * if it is set here then we've found a bug.
640 } while (WARN_ON_ONCE(running
));
643 static void dev_init_scheduler_queue(struct net_device
*dev
,
644 struct netdev_queue
*dev_queue
,
647 dev_queue
->qdisc
= qdisc
;
648 dev_queue
->qdisc_sleeping
= qdisc
;
649 INIT_LIST_HEAD(&dev_queue
->qdisc_list
);
652 void dev_init_scheduler(struct net_device
*dev
)
654 qdisc_lock_tree(dev
);
655 dev_init_scheduler_queue(dev
, &dev
->tx_queue
, &noop_qdisc
);
656 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
657 qdisc_unlock_tree(dev
);
659 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
662 static void dev_shutdown_scheduler_queue(struct net_device
*dev
,
663 struct netdev_queue
*dev_queue
,
664 struct Qdisc
*qdisc_default
)
666 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
669 dev_queue
->qdisc
= qdisc_default
;
670 dev_queue
->qdisc_sleeping
= qdisc_default
;
672 qdisc_destroy(qdisc
);
676 void dev_shutdown(struct net_device
*dev
)
678 qdisc_lock_tree(dev
);
679 dev_shutdown_scheduler_queue(dev
, &dev
->tx_queue
, &noop_qdisc
);
680 dev_shutdown_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
681 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));
682 qdisc_unlock_tree(dev
);