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
)
44 spin_lock_bh(&dev
->queue_lock
);
45 spin_lock(&dev
->ingress_lock
);
48 void qdisc_unlock_tree(struct net_device
*dev
)
50 spin_unlock(&dev
->ingress_lock
);
51 spin_unlock_bh(&dev
->queue_lock
);
54 static inline int qdisc_qlen(struct Qdisc
*q
)
59 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct net_device
*dev
,
62 if (unlikely(skb
->next
))
65 q
->ops
->requeue(skb
, q
);
71 static inline struct sk_buff
*dev_dequeue_skb(struct net_device
*dev
,
76 if ((skb
= dev
->gso_skb
))
84 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
85 struct net_device
*dev
,
90 if (unlikely(dev
->xmit_lock_owner
== smp_processor_id())) {
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.
99 printk(KERN_WARNING
"Dead loop on netdevice %s, "
100 "fix it urgently!\n", dev
->name
);
104 * Another cpu is holding lock, requeue & delay xmits for
107 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
108 ret
= dev_requeue_skb(skb
, dev
, q
);
115 * NOTE: Called under dev->queue_lock with locally disabled BH.
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.
121 * netif_tx_lock serializes accesses to device driver.
123 * dev->queue_lock and netif_tx_lock are mutually exclusive,
124 * if one is grabbed, another must be free.
126 * Note, that this procedure can be called by a watchdog timer
128 * Returns to the caller:
129 * 0 - queue is empty or throttled.
130 * >0 - queue is not empty.
133 static inline int qdisc_restart(struct net_device
*dev
)
135 struct Qdisc
*q
= dev
->qdisc
;
137 int ret
= NETDEV_TX_BUSY
;
140 if (unlikely((skb
= dev_dequeue_skb(dev
, q
)) == NULL
))
144 /* And release queue */
145 spin_unlock(&dev
->queue_lock
);
147 HARD_TX_LOCK(dev
, smp_processor_id());
148 if (!netif_subqueue_stopped(dev
, skb
))
149 ret
= dev_hard_start_xmit(skb
, dev
);
152 spin_lock(&dev
->queue_lock
);
157 /* Driver sent out skb successfully */
161 case NETDEV_TX_LOCKED
:
162 /* Driver try lock failed */
163 ret
= handle_dev_cpu_collision(skb
, dev
, q
);
167 /* Driver returned NETDEV_TX_BUSY - requeue skb */
168 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
169 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
170 dev
->name
, ret
, q
->q
.qlen
);
172 ret
= dev_requeue_skb(skb
, dev
, q
);
179 void __qdisc_run(struct net_device
*dev
)
182 if (!qdisc_restart(dev
))
184 } while (!netif_queue_stopped(dev
));
186 clear_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
189 static void dev_watchdog(unsigned long arg
)
191 struct net_device
*dev
= (struct net_device
*)arg
;
194 if (dev
->qdisc
!= &noop_qdisc
) {
195 if (netif_device_present(dev
) &&
196 netif_running(dev
) &&
197 netif_carrier_ok(dev
)) {
198 if (netif_queue_stopped(dev
) &&
199 time_after(jiffies
, dev
->trans_start
+ dev
->watchdog_timeo
)) {
201 printk(KERN_INFO
"NETDEV WATCHDOG: %s: transmit timed out\n",
203 dev
->tx_timeout(dev
);
205 if (!mod_timer(&dev
->watchdog_timer
, round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
209 netif_tx_unlock(dev
);
214 void __netdev_watchdog_up(struct net_device
*dev
)
216 if (dev
->tx_timeout
) {
217 if (dev
->watchdog_timeo
<= 0)
218 dev
->watchdog_timeo
= 5*HZ
;
219 if (!mod_timer(&dev
->watchdog_timer
,
220 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
225 static void dev_watchdog_up(struct net_device
*dev
)
227 __netdev_watchdog_up(dev
);
230 static void dev_watchdog_down(struct net_device
*dev
)
232 netif_tx_lock_bh(dev
);
233 if (del_timer(&dev
->watchdog_timer
))
235 netif_tx_unlock_bh(dev
);
239 * netif_carrier_on - set carrier
240 * @dev: network device
242 * Device has detected that carrier.
244 void netif_carrier_on(struct net_device
*dev
)
246 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
247 linkwatch_fire_event(dev
);
248 if (netif_running(dev
))
249 __netdev_watchdog_up(dev
);
254 * netif_carrier_off - clear carrier
255 * @dev: network device
257 * Device has detected loss of carrier.
259 void netif_carrier_off(struct net_device
*dev
)
261 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
262 linkwatch_fire_event(dev
);
265 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
266 under all circumstances. It is difficult to invent anything faster or
270 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
276 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
281 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
284 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
290 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
293 .enqueue
= noop_enqueue
,
294 .dequeue
= noop_dequeue
,
295 .requeue
= noop_requeue
,
296 .owner
= THIS_MODULE
,
299 struct Qdisc noop_qdisc
= {
300 .enqueue
= noop_enqueue
,
301 .dequeue
= noop_dequeue
,
302 .flags
= TCQ_F_BUILTIN
,
303 .ops
= &noop_qdisc_ops
,
304 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
307 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
310 .enqueue
= noop_enqueue
,
311 .dequeue
= noop_dequeue
,
312 .requeue
= noop_requeue
,
313 .owner
= THIS_MODULE
,
316 static struct Qdisc noqueue_qdisc
= {
318 .dequeue
= noop_dequeue
,
319 .flags
= TCQ_F_BUILTIN
,
320 .ops
= &noqueue_qdisc_ops
,
321 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
325 static const u8 prio2band
[TC_PRIO_MAX
+1] =
326 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
328 /* 3-band FIFO queue: old style, but should be a bit faster than
329 generic prio+fifo combination.
332 #define PFIFO_FAST_BANDS 3
334 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
337 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
338 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
341 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
343 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
345 if (skb_queue_len(list
) < qdisc
->dev
->tx_queue_len
) {
347 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
350 return qdisc_drop(skb
, qdisc
);
353 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
356 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
358 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
359 if (!skb_queue_empty(list
+ prio
)) {
361 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
368 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
371 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
374 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
377 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
379 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
380 __qdisc_reset_queue(qdisc
, list
+ prio
);
382 qdisc
->qstats
.backlog
= 0;
386 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
388 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
390 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
391 RTA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
398 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct rtattr
*opt
)
401 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
403 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
404 skb_queue_head_init(list
+ prio
);
409 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
411 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
412 .enqueue
= pfifo_fast_enqueue
,
413 .dequeue
= pfifo_fast_dequeue
,
414 .requeue
= pfifo_fast_requeue
,
415 .init
= pfifo_fast_init
,
416 .reset
= pfifo_fast_reset
,
417 .dump
= pfifo_fast_dump
,
418 .owner
= THIS_MODULE
,
421 struct Qdisc
*qdisc_alloc(struct net_device
*dev
, struct Qdisc_ops
*ops
)
428 /* ensure that the Qdisc and the private data are 32-byte aligned */
429 size
= QDISC_ALIGN(sizeof(*sch
));
430 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
432 p
= kzalloc(size
, GFP_KERNEL
);
435 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
436 sch
->padded
= (char *) sch
- (char *) p
;
438 INIT_LIST_HEAD(&sch
->list
);
439 skb_queue_head_init(&sch
->q
);
441 sch
->enqueue
= ops
->enqueue
;
442 sch
->dequeue
= ops
->dequeue
;
445 atomic_set(&sch
->refcnt
, 1);
449 return ERR_PTR(-err
);
452 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
, struct Qdisc_ops
*ops
,
453 unsigned int parentid
)
457 sch
= qdisc_alloc(dev
, ops
);
460 sch
->stats_lock
= &dev
->queue_lock
;
461 sch
->parent
= parentid
;
463 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
471 /* Under dev->queue_lock and BH! */
473 void qdisc_reset(struct Qdisc
*qdisc
)
475 const struct Qdisc_ops
*ops
= qdisc
->ops
;
481 /* this is the rcu callback function to clean up a qdisc when there
482 * are no further references to it */
484 static void __qdisc_destroy(struct rcu_head
*head
)
486 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
487 kfree((char *) qdisc
- qdisc
->padded
);
490 /* Under dev->queue_lock and BH! */
492 void qdisc_destroy(struct Qdisc
*qdisc
)
494 const struct Qdisc_ops
*ops
= qdisc
->ops
;
496 if (qdisc
->flags
& TCQ_F_BUILTIN
||
497 !atomic_dec_and_test(&qdisc
->refcnt
))
500 list_del(&qdisc
->list
);
501 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
507 module_put(ops
->owner
);
509 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
512 void dev_activate(struct net_device
*dev
)
514 /* No queueing discipline is attached to device;
515 create default one i.e. pfifo_fast for devices,
516 which need queueing and noqueue_qdisc for
520 if (dev
->qdisc_sleeping
== &noop_qdisc
) {
522 if (dev
->tx_queue_len
) {
523 qdisc
= qdisc_create_dflt(dev
, &pfifo_fast_ops
,
526 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
529 list_add_tail(&qdisc
->list
, &dev
->qdisc_list
);
531 qdisc
= &noqueue_qdisc
;
533 dev
->qdisc_sleeping
= qdisc
;
536 if (!netif_carrier_ok(dev
))
537 /* Delay activation until next carrier-on event */
540 spin_lock_bh(&dev
->queue_lock
);
541 rcu_assign_pointer(dev
->qdisc
, dev
->qdisc_sleeping
);
542 if (dev
->qdisc
!= &noqueue_qdisc
) {
543 dev
->trans_start
= jiffies
;
544 dev_watchdog_up(dev
);
546 spin_unlock_bh(&dev
->queue_lock
);
549 void dev_deactivate(struct net_device
*dev
)
555 spin_lock_bh(&dev
->queue_lock
);
557 dev
->qdisc
= &noop_qdisc
;
563 spin_unlock_bh(&dev
->queue_lock
);
567 dev_watchdog_down(dev
);
569 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
572 /* Wait for outstanding qdisc_run calls. */
574 while (test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
))
578 * Double-check inside queue lock to ensure that all effects
579 * of the queue run are visible when we return.
581 spin_lock_bh(&dev
->queue_lock
);
582 running
= test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
583 spin_unlock_bh(&dev
->queue_lock
);
586 * The running flag should never be set at this point because
587 * we've already set dev->qdisc to noop_qdisc *inside* the same
588 * pair of spin locks. That is, if any qdisc_run starts after
589 * our initial test it should see the noop_qdisc and then
590 * clear the RUNNING bit before dropping the queue lock. So
591 * if it is set here then we've found a bug.
593 } while (WARN_ON_ONCE(running
));
596 void dev_init_scheduler(struct net_device
*dev
)
598 qdisc_lock_tree(dev
);
599 dev
->qdisc
= &noop_qdisc
;
600 dev
->qdisc_sleeping
= &noop_qdisc
;
601 INIT_LIST_HEAD(&dev
->qdisc_list
);
602 qdisc_unlock_tree(dev
);
604 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
607 void dev_shutdown(struct net_device
*dev
)
611 qdisc_lock_tree(dev
);
612 qdisc
= dev
->qdisc_sleeping
;
613 dev
->qdisc
= &noop_qdisc
;
614 dev
->qdisc_sleeping
= &noop_qdisc
;
615 qdisc_destroy(qdisc
);
616 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
617 if ((qdisc
= dev
->qdisc_ingress
) != NULL
) {
618 dev
->qdisc_ingress
= NULL
;
619 qdisc_destroy(qdisc
);
622 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));
623 qdisc_unlock_tree(dev
);
626 EXPORT_SYMBOL(netif_carrier_on
);
627 EXPORT_SYMBOL(netif_carrier_off
);
628 EXPORT_SYMBOL(noop_qdisc
);
629 EXPORT_SYMBOL(qdisc_create_dflt
);
630 EXPORT_SYMBOL(qdisc_destroy
);
631 EXPORT_SYMBOL(qdisc_reset
);
632 EXPORT_SYMBOL(qdisc_lock_tree
);
633 EXPORT_SYMBOL(qdisc_unlock_tree
);