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 * dev->queue_lock spinlock.
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.
42 void qdisc_lock_tree(struct net_device
*dev
)
43 __acquires(dev
->queue_lock
)
44 __acquires(dev
->ingress_lock
)
46 spin_lock_bh(&dev
->queue_lock
);
47 spin_lock(&dev
->ingress_lock
);
50 void qdisc_unlock_tree(struct net_device
*dev
)
51 __releases(dev
->ingress_lock
)
52 __releases(dev
->queue_lock
)
54 spin_unlock(&dev
->ingress_lock
);
55 spin_unlock_bh(&dev
->queue_lock
);
58 static inline int qdisc_qlen(struct Qdisc
*q
)
63 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct net_device
*dev
,
66 if (unlikely(skb
->next
))
69 q
->ops
->requeue(skb
, q
);
75 static inline struct sk_buff
*dev_dequeue_skb(struct net_device
*dev
,
80 if ((skb
= dev
->gso_skb
))
88 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
89 struct net_device
*dev
,
94 if (unlikely(dev
->xmit_lock_owner
== smp_processor_id())) {
96 * Same CPU holding the lock. It may be a transient
97 * configuration error, when hard_start_xmit() recurses. We
98 * detect it by checking xmit owner and drop the packet when
99 * deadloop is detected. Return OK to try the next skb.
103 printk(KERN_WARNING
"Dead loop on netdevice %s, "
104 "fix it urgently!\n", dev
->name
);
108 * Another cpu is holding lock, requeue & delay xmits for
111 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
112 ret
= dev_requeue_skb(skb
, dev
, q
);
119 * NOTE: Called under dev->queue_lock with locally disabled BH.
121 * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
122 * device at a time. dev->queue_lock serializes queue accesses for
123 * this device AND dev->qdisc pointer itself.
125 * netif_tx_lock serializes accesses to device driver.
127 * dev->queue_lock and netif_tx_lock are mutually exclusive,
128 * if one is grabbed, another must be free.
130 * Note, that this procedure can be called by a watchdog timer
132 * Returns to the caller:
133 * 0 - queue is empty or throttled.
134 * >0 - queue is not empty.
137 static inline int qdisc_restart(struct net_device
*dev
)
139 struct Qdisc
*q
= dev
->qdisc
;
141 int ret
= NETDEV_TX_BUSY
;
144 if (unlikely((skb
= dev_dequeue_skb(dev
, q
)) == NULL
))
148 /* And release queue */
149 spin_unlock(&dev
->queue_lock
);
151 HARD_TX_LOCK(dev
, smp_processor_id());
152 if (!netif_subqueue_stopped(dev
, skb
))
153 ret
= dev_hard_start_xmit(skb
, dev
);
156 spin_lock(&dev
->queue_lock
);
161 /* Driver sent out skb successfully */
165 case NETDEV_TX_LOCKED
:
166 /* Driver try lock failed */
167 ret
= handle_dev_cpu_collision(skb
, dev
, q
);
171 /* Driver returned NETDEV_TX_BUSY - requeue skb */
172 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
173 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
174 dev
->name
, ret
, q
->q
.qlen
);
176 ret
= dev_requeue_skb(skb
, dev
, q
);
183 void __qdisc_run(struct net_device
*dev
)
186 if (!qdisc_restart(dev
))
188 } while (!netif_queue_stopped(dev
));
190 clear_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
193 static void dev_watchdog(unsigned long arg
)
195 struct net_device
*dev
= (struct net_device
*)arg
;
198 if (dev
->qdisc
!= &noop_qdisc
) {
199 if (netif_device_present(dev
) &&
200 netif_running(dev
) &&
201 netif_carrier_ok(dev
)) {
202 if (netif_queue_stopped(dev
) &&
203 time_after(jiffies
, dev
->trans_start
+ dev
->watchdog_timeo
)) {
205 printk(KERN_INFO
"NETDEV WATCHDOG: %s: transmit timed out\n",
207 dev
->tx_timeout(dev
);
209 if (!mod_timer(&dev
->watchdog_timer
, round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
213 netif_tx_unlock(dev
);
218 void __netdev_watchdog_up(struct net_device
*dev
)
220 if (dev
->tx_timeout
) {
221 if (dev
->watchdog_timeo
<= 0)
222 dev
->watchdog_timeo
= 5*HZ
;
223 if (!mod_timer(&dev
->watchdog_timer
,
224 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
229 static void dev_watchdog_up(struct net_device
*dev
)
231 __netdev_watchdog_up(dev
);
234 static void dev_watchdog_down(struct net_device
*dev
)
236 netif_tx_lock_bh(dev
);
237 if (del_timer(&dev
->watchdog_timer
))
239 netif_tx_unlock_bh(dev
);
243 * netif_carrier_on - set carrier
244 * @dev: network device
246 * Device has detected that carrier.
248 void netif_carrier_on(struct net_device
*dev
)
250 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
251 linkwatch_fire_event(dev
);
252 if (netif_running(dev
))
253 __netdev_watchdog_up(dev
);
258 * netif_carrier_off - clear carrier
259 * @dev: network device
261 * Device has detected loss of carrier.
263 void netif_carrier_off(struct net_device
*dev
)
265 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
266 linkwatch_fire_event(dev
);
269 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
270 under all circumstances. It is difficult to invent anything faster or
274 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
280 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
285 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
288 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
294 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
297 .enqueue
= noop_enqueue
,
298 .dequeue
= noop_dequeue
,
299 .requeue
= noop_requeue
,
300 .owner
= THIS_MODULE
,
303 struct Qdisc noop_qdisc
= {
304 .enqueue
= noop_enqueue
,
305 .dequeue
= noop_dequeue
,
306 .flags
= TCQ_F_BUILTIN
,
307 .ops
= &noop_qdisc_ops
,
308 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
311 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
314 .enqueue
= noop_enqueue
,
315 .dequeue
= noop_dequeue
,
316 .requeue
= noop_requeue
,
317 .owner
= THIS_MODULE
,
320 static struct Qdisc noqueue_qdisc
= {
322 .dequeue
= noop_dequeue
,
323 .flags
= TCQ_F_BUILTIN
,
324 .ops
= &noqueue_qdisc_ops
,
325 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
329 static const u8 prio2band
[TC_PRIO_MAX
+1] =
330 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
332 /* 3-band FIFO queue: old style, but should be a bit faster than
333 generic prio+fifo combination.
336 #define PFIFO_FAST_BANDS 3
338 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
341 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
342 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
345 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
347 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
349 if (skb_queue_len(list
) < qdisc
->dev
->tx_queue_len
) {
351 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
354 return qdisc_drop(skb
, qdisc
);
357 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
360 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
362 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
363 if (!skb_queue_empty(list
+ prio
)) {
365 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
372 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
375 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
378 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
381 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
383 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
384 __qdisc_reset_queue(qdisc
, list
+ prio
);
386 qdisc
->qstats
.backlog
= 0;
390 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
392 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
394 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
395 RTA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
402 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct rtattr
*opt
)
405 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
407 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
408 skb_queue_head_init(list
+ prio
);
413 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
415 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
416 .enqueue
= pfifo_fast_enqueue
,
417 .dequeue
= pfifo_fast_dequeue
,
418 .requeue
= pfifo_fast_requeue
,
419 .init
= pfifo_fast_init
,
420 .reset
= pfifo_fast_reset
,
421 .dump
= pfifo_fast_dump
,
422 .owner
= THIS_MODULE
,
425 struct Qdisc
*qdisc_alloc(struct net_device
*dev
, struct Qdisc_ops
*ops
)
432 /* ensure that the Qdisc and the private data are 32-byte aligned */
433 size
= QDISC_ALIGN(sizeof(*sch
));
434 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
436 p
= kzalloc(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 atomic_set(&sch
->refcnt
, 1);
453 return ERR_PTR(-err
);
456 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
, struct Qdisc_ops
*ops
,
457 unsigned int parentid
)
461 sch
= qdisc_alloc(dev
, ops
);
464 sch
->stats_lock
= &dev
->queue_lock
;
465 sch
->parent
= parentid
;
467 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
475 /* Under dev->queue_lock and BH! */
477 void qdisc_reset(struct Qdisc
*qdisc
)
479 const struct Qdisc_ops
*ops
= qdisc
->ops
;
485 /* this is the rcu callback function to clean up a qdisc when there
486 * are no further references to it */
488 static void __qdisc_destroy(struct rcu_head
*head
)
490 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
491 kfree((char *) qdisc
- qdisc
->padded
);
494 /* Under dev->queue_lock and BH! */
496 void qdisc_destroy(struct Qdisc
*qdisc
)
498 const struct Qdisc_ops
*ops
= qdisc
->ops
;
500 if (qdisc
->flags
& TCQ_F_BUILTIN
||
501 !atomic_dec_and_test(&qdisc
->refcnt
))
504 list_del(&qdisc
->list
);
505 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
511 module_put(ops
->owner
);
513 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
516 void dev_activate(struct net_device
*dev
)
518 /* No queueing discipline is attached to device;
519 create default one i.e. pfifo_fast for devices,
520 which need queueing and noqueue_qdisc for
524 if (dev
->qdisc_sleeping
== &noop_qdisc
) {
526 if (dev
->tx_queue_len
) {
527 qdisc
= qdisc_create_dflt(dev
, &pfifo_fast_ops
,
530 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
533 list_add_tail(&qdisc
->list
, &dev
->qdisc_list
);
535 qdisc
= &noqueue_qdisc
;
537 dev
->qdisc_sleeping
= qdisc
;
540 if (!netif_carrier_ok(dev
))
541 /* Delay activation until next carrier-on event */
544 spin_lock_bh(&dev
->queue_lock
);
545 rcu_assign_pointer(dev
->qdisc
, dev
->qdisc_sleeping
);
546 if (dev
->qdisc
!= &noqueue_qdisc
) {
547 dev
->trans_start
= jiffies
;
548 dev_watchdog_up(dev
);
550 spin_unlock_bh(&dev
->queue_lock
);
553 void dev_deactivate(struct net_device
*dev
)
559 spin_lock_bh(&dev
->queue_lock
);
561 dev
->qdisc
= &noop_qdisc
;
567 spin_unlock_bh(&dev
->queue_lock
);
571 dev_watchdog_down(dev
);
573 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
576 /* Wait for outstanding qdisc_run calls. */
578 while (test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
))
582 * Double-check inside queue lock to ensure that all effects
583 * of the queue run are visible when we return.
585 spin_lock_bh(&dev
->queue_lock
);
586 running
= test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
587 spin_unlock_bh(&dev
->queue_lock
);
590 * The running flag should never be set at this point because
591 * we've already set dev->qdisc to noop_qdisc *inside* the same
592 * pair of spin locks. That is, if any qdisc_run starts after
593 * our initial test it should see the noop_qdisc and then
594 * clear the RUNNING bit before dropping the queue lock. So
595 * if it is set here then we've found a bug.
597 } while (WARN_ON_ONCE(running
));
600 void dev_init_scheduler(struct net_device
*dev
)
602 qdisc_lock_tree(dev
);
603 dev
->qdisc
= &noop_qdisc
;
604 dev
->qdisc_sleeping
= &noop_qdisc
;
605 INIT_LIST_HEAD(&dev
->qdisc_list
);
606 qdisc_unlock_tree(dev
);
608 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
611 void dev_shutdown(struct net_device
*dev
)
615 qdisc_lock_tree(dev
);
616 qdisc
= dev
->qdisc_sleeping
;
617 dev
->qdisc
= &noop_qdisc
;
618 dev
->qdisc_sleeping
= &noop_qdisc
;
619 qdisc_destroy(qdisc
);
620 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
621 if ((qdisc
= dev
->qdisc_ingress
) != NULL
) {
622 dev
->qdisc_ingress
= NULL
;
623 qdisc_destroy(qdisc
);
626 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));
627 qdisc_unlock_tree(dev
);
630 EXPORT_SYMBOL(netif_carrier_on
);
631 EXPORT_SYMBOL(netif_carrier_off
);
632 EXPORT_SYMBOL(noop_qdisc
);
633 EXPORT_SYMBOL(qdisc_create_dflt
);
634 EXPORT_SYMBOL(qdisc_destroy
);
635 EXPORT_SYMBOL(qdisc_reset
);
636 EXPORT_SYMBOL(qdisc_lock_tree
);
637 EXPORT_SYMBOL(qdisc_unlock_tree
);