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 * qdisc_lock(qdisc) spinlock.
34 * The idea is the following:
35 * - enqueue, dequeue are serialized via qdisc root lock
36 * - ingress filtering is also serialized via qdisc root lock
37 * - updates to tree and tree walking are only done under the rtnl mutex.
40 static inline int qdisc_qlen(struct Qdisc
*q
)
45 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
47 if (unlikely(skb
->next
))
50 q
->ops
->requeue(skb
, q
);
56 static inline struct sk_buff
*dequeue_skb(struct Qdisc
*q
)
60 if ((skb
= q
->gso_skb
))
68 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
69 struct netdev_queue
*dev_queue
,
74 if (unlikely(dev_queue
->xmit_lock_owner
== smp_processor_id())) {
76 * Same CPU holding the lock. It may be a transient
77 * configuration error, when hard_start_xmit() recurses. We
78 * detect it by checking xmit owner and drop the packet when
79 * deadloop is detected. Return OK to try the next skb.
83 printk(KERN_WARNING
"Dead loop on netdevice %s, "
84 "fix it urgently!\n", dev_queue
->dev
->name
);
88 * Another cpu is holding lock, requeue & delay xmits for
91 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
92 ret
= dev_requeue_skb(skb
, q
);
99 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
101 * __QDISC_STATE_RUNNING guarantees only one CPU can process
102 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
105 * netif_tx_lock serializes accesses to device driver.
107 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
108 * if one is grabbed, another must be free.
110 * Note, that this procedure can be called by a watchdog timer
112 * Returns to the caller:
113 * 0 - queue is empty or throttled.
114 * >0 - queue is not empty.
117 static inline int qdisc_restart(struct Qdisc
*q
)
119 struct netdev_queue
*txq
;
120 int ret
= NETDEV_TX_BUSY
;
121 struct net_device
*dev
;
122 spinlock_t
*root_lock
;
126 if (unlikely((skb
= dequeue_skb(q
)) == NULL
))
129 root_lock
= qdisc_lock(q
);
131 /* And release qdisc */
132 spin_unlock(root_lock
);
135 txq
= netdev_get_tx_queue(dev
, skb_get_queue_mapping(skb
));
137 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
138 if (!netif_tx_queue_stopped(txq
) &&
139 !netif_tx_queue_frozen(txq
))
140 ret
= dev_hard_start_xmit(skb
, dev
, txq
);
141 HARD_TX_UNLOCK(dev
, txq
);
143 spin_lock(root_lock
);
147 /* Driver sent out skb successfully */
151 case NETDEV_TX_LOCKED
:
152 /* Driver try lock failed */
153 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
157 /* Driver returned NETDEV_TX_BUSY - requeue skb */
158 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
159 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
160 dev
->name
, ret
, q
->q
.qlen
);
162 ret
= dev_requeue_skb(skb
, q
);
166 if (ret
&& (netif_tx_queue_stopped(txq
) ||
167 netif_tx_queue_frozen(txq
)))
173 void __qdisc_run(struct Qdisc
*q
)
175 unsigned long start_time
= jiffies
;
177 while (qdisc_restart(q
)) {
179 * Postpone processing if
180 * 1. another process needs the CPU;
181 * 2. we've been doing it for too long.
183 if (need_resched() || jiffies
!= start_time
) {
189 clear_bit(__QDISC_STATE_RUNNING
, &q
->state
);
192 static void dev_watchdog(unsigned long arg
)
194 struct net_device
*dev
= (struct net_device
*)arg
;
197 if (!qdisc_tx_is_noop(dev
)) {
198 if (netif_device_present(dev
) &&
199 netif_running(dev
) &&
200 netif_carrier_ok(dev
)) {
201 int some_queue_stopped
= 0;
204 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
205 struct netdev_queue
*txq
;
207 txq
= netdev_get_tx_queue(dev
, i
);
208 if (netif_tx_queue_stopped(txq
)) {
209 some_queue_stopped
= 1;
214 if (some_queue_stopped
&&
215 time_after(jiffies
, (dev
->trans_start
+
216 dev
->watchdog_timeo
))) {
218 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit timed out\n",
219 dev
->name
, netdev_drivername(dev
, drivername
, 64));
220 dev
->tx_timeout(dev
);
222 if (!mod_timer(&dev
->watchdog_timer
,
223 round_jiffies(jiffies
+
224 dev
->watchdog_timeo
)))
228 netif_tx_unlock(dev
);
233 void __netdev_watchdog_up(struct net_device
*dev
)
235 if (dev
->tx_timeout
) {
236 if (dev
->watchdog_timeo
<= 0)
237 dev
->watchdog_timeo
= 5*HZ
;
238 if (!mod_timer(&dev
->watchdog_timer
,
239 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
244 static void dev_watchdog_up(struct net_device
*dev
)
246 __netdev_watchdog_up(dev
);
249 static void dev_watchdog_down(struct net_device
*dev
)
251 netif_tx_lock_bh(dev
);
252 if (del_timer(&dev
->watchdog_timer
))
254 netif_tx_unlock_bh(dev
);
258 * netif_carrier_on - set carrier
259 * @dev: network device
261 * Device has detected that carrier.
263 void netif_carrier_on(struct net_device
*dev
)
265 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
266 linkwatch_fire_event(dev
);
267 if (netif_running(dev
))
268 __netdev_watchdog_up(dev
);
271 EXPORT_SYMBOL(netif_carrier_on
);
274 * netif_carrier_off - clear carrier
275 * @dev: network device
277 * Device has detected loss of carrier.
279 void netif_carrier_off(struct net_device
*dev
)
281 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
282 linkwatch_fire_event(dev
);
284 EXPORT_SYMBOL(netif_carrier_off
);
286 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
287 under all circumstances. It is difficult to invent anything faster or
291 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
297 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
302 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
305 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
311 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
314 .enqueue
= noop_enqueue
,
315 .dequeue
= noop_dequeue
,
316 .requeue
= noop_requeue
,
317 .owner
= THIS_MODULE
,
320 static struct netdev_queue noop_netdev_queue
= {
321 .qdisc
= &noop_qdisc
,
322 .qdisc_sleeping
= &noop_qdisc
,
325 struct Qdisc noop_qdisc
= {
326 .enqueue
= noop_enqueue
,
327 .dequeue
= noop_dequeue
,
328 .flags
= TCQ_F_BUILTIN
,
329 .ops
= &noop_qdisc_ops
,
330 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
331 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
332 .dev_queue
= &noop_netdev_queue
,
334 EXPORT_SYMBOL(noop_qdisc
);
336 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
339 .enqueue
= noop_enqueue
,
340 .dequeue
= noop_dequeue
,
341 .requeue
= noop_requeue
,
342 .owner
= THIS_MODULE
,
345 static struct Qdisc noqueue_qdisc
;
346 static struct netdev_queue noqueue_netdev_queue
= {
347 .qdisc
= &noqueue_qdisc
,
348 .qdisc_sleeping
= &noqueue_qdisc
,
351 static struct Qdisc noqueue_qdisc
= {
353 .dequeue
= noop_dequeue
,
354 .flags
= TCQ_F_BUILTIN
,
355 .ops
= &noqueue_qdisc_ops
,
356 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
357 .q
.lock
= __SPIN_LOCK_UNLOCKED(noqueue_qdisc
.q
.lock
),
358 .dev_queue
= &noqueue_netdev_queue
,
362 static const u8 prio2band
[TC_PRIO_MAX
+1] =
363 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
365 /* 3-band FIFO queue: old style, but should be a bit faster than
366 generic prio+fifo combination.
369 #define PFIFO_FAST_BANDS 3
371 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
374 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
375 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
378 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
380 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
382 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
384 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
387 return qdisc_drop(skb
, qdisc
);
390 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
393 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
395 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
396 if (!skb_queue_empty(list
+ prio
)) {
398 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
405 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
408 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
411 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
414 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
416 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
417 __qdisc_reset_queue(qdisc
, list
+ prio
);
419 qdisc
->qstats
.backlog
= 0;
423 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
425 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
427 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
428 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
435 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
438 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
440 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
441 skb_queue_head_init(list
+ prio
);
446 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
448 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
449 .enqueue
= pfifo_fast_enqueue
,
450 .dequeue
= pfifo_fast_dequeue
,
451 .requeue
= pfifo_fast_requeue
,
452 .init
= pfifo_fast_init
,
453 .reset
= pfifo_fast_reset
,
454 .dump
= pfifo_fast_dump
,
455 .owner
= THIS_MODULE
,
458 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
459 struct Qdisc_ops
*ops
)
466 /* ensure that the Qdisc and the private data are 32-byte aligned */
467 size
= QDISC_ALIGN(sizeof(*sch
));
468 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
470 p
= kzalloc(size
, GFP_KERNEL
);
473 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
474 sch
->padded
= (char *) sch
- (char *) p
;
476 INIT_LIST_HEAD(&sch
->list
);
477 skb_queue_head_init(&sch
->q
);
479 sch
->enqueue
= ops
->enqueue
;
480 sch
->dequeue
= ops
->dequeue
;
481 sch
->dev_queue
= dev_queue
;
482 dev_hold(qdisc_dev(sch
));
483 atomic_set(&sch
->refcnt
, 1);
490 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
491 struct netdev_queue
*dev_queue
,
492 struct Qdisc_ops
*ops
,
493 unsigned int parentid
)
497 sch
= qdisc_alloc(dev_queue
, ops
);
500 sch
->parent
= parentid
;
502 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
509 EXPORT_SYMBOL(qdisc_create_dflt
);
511 /* Under qdisc_lock(qdisc) and BH! */
513 void qdisc_reset(struct Qdisc
*qdisc
)
515 const struct Qdisc_ops
*ops
= qdisc
->ops
;
520 EXPORT_SYMBOL(qdisc_reset
);
522 void qdisc_destroy(struct Qdisc
*qdisc
)
524 const struct Qdisc_ops
*ops
= qdisc
->ops
;
526 if (qdisc
->flags
& TCQ_F_BUILTIN
||
527 !atomic_dec_and_test(&qdisc
->refcnt
))
530 #ifdef CONFIG_NET_SCHED
531 qdisc_list_del(qdisc
);
533 qdisc_put_stab(qdisc
->stab
);
535 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
541 module_put(ops
->owner
);
542 dev_put(qdisc_dev(qdisc
));
544 kfree_skb(qdisc
->gso_skb
);
546 kfree((char *) qdisc
- qdisc
->padded
);
548 EXPORT_SYMBOL(qdisc_destroy
);
550 static bool dev_all_qdisc_sleeping_noop(struct net_device
*dev
)
554 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
555 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, i
);
557 if (txq
->qdisc_sleeping
!= &noop_qdisc
)
563 static void attach_one_default_qdisc(struct net_device
*dev
,
564 struct netdev_queue
*dev_queue
,
569 if (dev
->tx_queue_len
) {
570 qdisc
= qdisc_create_dflt(dev
, dev_queue
,
571 &pfifo_fast_ops
, TC_H_ROOT
);
573 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
577 qdisc
= &noqueue_qdisc
;
579 dev_queue
->qdisc_sleeping
= qdisc
;
582 static void transition_one_qdisc(struct net_device
*dev
,
583 struct netdev_queue
*dev_queue
,
584 void *_need_watchdog
)
586 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
587 int *need_watchdog_p
= _need_watchdog
;
589 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
590 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
592 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
593 if (need_watchdog_p
&& new_qdisc
!= &noqueue_qdisc
)
594 *need_watchdog_p
= 1;
597 void dev_activate(struct net_device
*dev
)
601 /* No queueing discipline is attached to device;
602 create default one i.e. pfifo_fast for devices,
603 which need queueing and noqueue_qdisc for
607 if (dev_all_qdisc_sleeping_noop(dev
))
608 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
610 if (!netif_carrier_ok(dev
))
611 /* Delay activation until next carrier-on event */
615 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
616 transition_one_qdisc(dev
, &dev
->rx_queue
, NULL
);
619 dev
->trans_start
= jiffies
;
620 dev_watchdog_up(dev
);
624 static void dev_deactivate_queue(struct net_device
*dev
,
625 struct netdev_queue
*dev_queue
,
626 void *_qdisc_default
)
628 struct Qdisc
*qdisc_default
= _qdisc_default
;
631 qdisc
= dev_queue
->qdisc
;
633 spin_lock_bh(qdisc_lock(qdisc
));
635 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
636 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
638 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
641 spin_unlock_bh(qdisc_lock(qdisc
));
645 static bool some_qdisc_is_busy(struct net_device
*dev
)
649 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
650 struct netdev_queue
*dev_queue
;
651 spinlock_t
*root_lock
;
655 dev_queue
= netdev_get_tx_queue(dev
, i
);
656 q
= dev_queue
->qdisc_sleeping
;
657 root_lock
= qdisc_lock(q
);
659 spin_lock_bh(root_lock
);
661 val
= (test_bit(__QDISC_STATE_RUNNING
, &q
->state
) ||
662 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
664 spin_unlock_bh(root_lock
);
672 void dev_deactivate(struct net_device
*dev
)
674 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
, &noop_qdisc
);
675 dev_deactivate_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
677 dev_watchdog_down(dev
);
679 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
682 /* Wait for outstanding qdisc_run calls. */
683 while (some_qdisc_is_busy(dev
))
687 static void dev_init_scheduler_queue(struct net_device
*dev
,
688 struct netdev_queue
*dev_queue
,
691 struct Qdisc
*qdisc
= _qdisc
;
693 dev_queue
->qdisc
= qdisc
;
694 dev_queue
->qdisc_sleeping
= qdisc
;
697 void dev_init_scheduler(struct net_device
*dev
)
699 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
700 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
702 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
705 static void shutdown_scheduler_queue(struct net_device
*dev
,
706 struct netdev_queue
*dev_queue
,
707 void *_qdisc_default
)
709 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
710 struct Qdisc
*qdisc_default
= _qdisc_default
;
713 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
714 dev_queue
->qdisc_sleeping
= qdisc_default
;
716 qdisc_destroy(qdisc
);
720 void dev_shutdown(struct net_device
*dev
)
722 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
723 shutdown_scheduler_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
724 WARN_ON(timer_pending(&dev
->watchdog_timer
));