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_root_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_root_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_subqueue_stopped(dev
, skb
))
139 ret
= dev_hard_start_xmit(skb
, dev
, txq
);
140 HARD_TX_UNLOCK(dev
, txq
);
142 spin_lock(root_lock
);
146 /* Driver sent out skb successfully */
150 case NETDEV_TX_LOCKED
:
151 /* Driver try lock failed */
152 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
156 /* Driver returned NETDEV_TX_BUSY - requeue skb */
157 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
158 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
159 dev
->name
, ret
, q
->q
.qlen
);
161 ret
= dev_requeue_skb(skb
, q
);
165 if (ret
&& netif_tx_queue_stopped(txq
))
171 void __qdisc_run(struct Qdisc
*q
)
173 unsigned long start_time
= jiffies
;
175 while (qdisc_restart(q
)) {
177 * Postpone processing if
178 * 1. another process needs the CPU;
179 * 2. we've been doing it for too long.
181 if (need_resched() || jiffies
!= start_time
) {
187 clear_bit(__QDISC_STATE_RUNNING
, &q
->state
);
190 static void dev_watchdog(unsigned long arg
)
192 struct net_device
*dev
= (struct net_device
*)arg
;
195 if (!qdisc_tx_is_noop(dev
)) {
196 if (netif_device_present(dev
) &&
197 netif_running(dev
) &&
198 netif_carrier_ok(dev
)) {
199 int some_queue_stopped
= 0;
202 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
203 struct netdev_queue
*txq
;
205 txq
= netdev_get_tx_queue(dev
, i
);
206 if (netif_tx_queue_stopped(txq
)) {
207 some_queue_stopped
= 1;
212 if (some_queue_stopped
&&
213 time_after(jiffies
, (dev
->trans_start
+
214 dev
->watchdog_timeo
))) {
215 printk(KERN_INFO
"NETDEV WATCHDOG: %s: "
216 "transmit timed out\n",
218 dev
->tx_timeout(dev
);
221 if (!mod_timer(&dev
->watchdog_timer
,
222 round_jiffies(jiffies
+
223 dev
->watchdog_timeo
)))
227 netif_tx_unlock(dev
);
232 void __netdev_watchdog_up(struct net_device
*dev
)
234 if (dev
->tx_timeout
) {
235 if (dev
->watchdog_timeo
<= 0)
236 dev
->watchdog_timeo
= 5*HZ
;
237 if (!mod_timer(&dev
->watchdog_timer
,
238 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
243 static void dev_watchdog_up(struct net_device
*dev
)
245 __netdev_watchdog_up(dev
);
248 static void dev_watchdog_down(struct net_device
*dev
)
250 netif_tx_lock_bh(dev
);
251 if (del_timer(&dev
->watchdog_timer
))
253 netif_tx_unlock_bh(dev
);
257 * netif_carrier_on - set carrier
258 * @dev: network device
260 * Device has detected that carrier.
262 void netif_carrier_on(struct net_device
*dev
)
264 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
265 linkwatch_fire_event(dev
);
266 if (netif_running(dev
))
267 __netdev_watchdog_up(dev
);
270 EXPORT_SYMBOL(netif_carrier_on
);
273 * netif_carrier_off - clear carrier
274 * @dev: network device
276 * Device has detected loss of carrier.
278 void netif_carrier_off(struct net_device
*dev
)
280 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
281 linkwatch_fire_event(dev
);
283 EXPORT_SYMBOL(netif_carrier_off
);
285 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
286 under all circumstances. It is difficult to invent anything faster or
290 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
296 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
301 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
304 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
310 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
313 .enqueue
= noop_enqueue
,
314 .dequeue
= noop_dequeue
,
315 .requeue
= noop_requeue
,
316 .owner
= THIS_MODULE
,
319 static struct netdev_queue noop_netdev_queue
= {
320 .qdisc
= &noop_qdisc
,
323 struct Qdisc noop_qdisc
= {
324 .enqueue
= noop_enqueue
,
325 .dequeue
= noop_dequeue
,
326 .flags
= TCQ_F_BUILTIN
,
327 .ops
= &noop_qdisc_ops
,
328 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
329 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
330 .dev_queue
= &noop_netdev_queue
,
332 EXPORT_SYMBOL(noop_qdisc
);
334 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
337 .enqueue
= noop_enqueue
,
338 .dequeue
= noop_dequeue
,
339 .requeue
= noop_requeue
,
340 .owner
= THIS_MODULE
,
343 static struct Qdisc noqueue_qdisc
= {
345 .dequeue
= noop_dequeue
,
346 .flags
= TCQ_F_BUILTIN
,
347 .ops
= &noqueue_qdisc_ops
,
348 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
352 static const u8 prio2band
[TC_PRIO_MAX
+1] =
353 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
355 /* 3-band FIFO queue: old style, but should be a bit faster than
356 generic prio+fifo combination.
359 #define PFIFO_FAST_BANDS 3
361 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
364 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
365 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
368 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
370 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
372 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
374 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
377 return qdisc_drop(skb
, qdisc
);
380 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
383 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
385 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
386 if (!skb_queue_empty(list
+ prio
)) {
388 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
395 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
398 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
401 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
404 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
406 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
407 __qdisc_reset_queue(qdisc
, list
+ prio
);
409 qdisc
->qstats
.backlog
= 0;
413 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
415 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
417 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
418 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
425 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
428 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
430 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
431 skb_queue_head_init(list
+ prio
);
436 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
438 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
439 .enqueue
= pfifo_fast_enqueue
,
440 .dequeue
= pfifo_fast_dequeue
,
441 .requeue
= pfifo_fast_requeue
,
442 .init
= pfifo_fast_init
,
443 .reset
= pfifo_fast_reset
,
444 .dump
= pfifo_fast_dump
,
445 .owner
= THIS_MODULE
,
448 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
449 struct Qdisc_ops
*ops
)
456 /* ensure that the Qdisc and the private data are 32-byte aligned */
457 size
= QDISC_ALIGN(sizeof(*sch
));
458 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
460 p
= kzalloc(size
, GFP_KERNEL
);
463 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
464 sch
->padded
= (char *) sch
- (char *) p
;
466 INIT_LIST_HEAD(&sch
->list
);
467 skb_queue_head_init(&sch
->q
);
469 sch
->enqueue
= ops
->enqueue
;
470 sch
->dequeue
= ops
->dequeue
;
471 sch
->dev_queue
= dev_queue
;
472 dev_hold(qdisc_dev(sch
));
473 atomic_set(&sch
->refcnt
, 1);
480 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
481 struct netdev_queue
*dev_queue
,
482 struct Qdisc_ops
*ops
,
483 unsigned int parentid
)
487 sch
= qdisc_alloc(dev_queue
, ops
);
490 sch
->parent
= parentid
;
492 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
499 EXPORT_SYMBOL(qdisc_create_dflt
);
501 /* Under qdisc_root_lock(qdisc) and BH! */
503 void qdisc_reset(struct Qdisc
*qdisc
)
505 const struct Qdisc_ops
*ops
= qdisc
->ops
;
510 EXPORT_SYMBOL(qdisc_reset
);
512 /* this is the rcu callback function to clean up a qdisc when there
513 * are no further references to it */
515 static void __qdisc_destroy(struct rcu_head
*head
)
517 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
518 const struct Qdisc_ops
*ops
= qdisc
->ops
;
520 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
526 module_put(ops
->owner
);
527 dev_put(qdisc_dev(qdisc
));
529 kfree_skb(qdisc
->gso_skb
);
531 kfree((char *) qdisc
- qdisc
->padded
);
534 /* Under qdisc_root_lock(qdisc) and BH! */
536 void qdisc_destroy(struct Qdisc
*qdisc
)
538 struct net_device
*dev
= qdisc_dev(qdisc
);
540 if (qdisc
->flags
& TCQ_F_BUILTIN
||
541 !atomic_dec_and_test(&qdisc
->refcnt
))
544 spin_lock_bh(&dev
->qdisc_list_lock
);
545 list_del(&qdisc
->list
);
546 spin_unlock_bh(&dev
->qdisc_list_lock
);
548 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
550 EXPORT_SYMBOL(qdisc_destroy
);
552 static bool dev_all_qdisc_sleeping_noop(struct net_device
*dev
)
556 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
557 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, i
);
559 if (txq
->qdisc_sleeping
!= &noop_qdisc
)
565 static void attach_one_default_qdisc(struct net_device
*dev
,
566 struct netdev_queue
*dev_queue
,
571 if (dev
->tx_queue_len
) {
572 qdisc
= qdisc_create_dflt(dev
, dev_queue
,
573 &pfifo_fast_ops
, TC_H_ROOT
);
575 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
578 spin_lock_bh(&dev
->qdisc_list_lock
);
579 list_add_tail(&qdisc
->list
, &dev
->qdisc_list
);
580 spin_unlock_bh(&dev
->qdisc_list_lock
);
582 qdisc
= &noqueue_qdisc
;
584 dev_queue
->qdisc_sleeping
= qdisc
;
587 static void transition_one_qdisc(struct net_device
*dev
,
588 struct netdev_queue
*dev_queue
,
589 void *_need_watchdog
)
591 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
592 int *need_watchdog_p
= _need_watchdog
;
594 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
595 if (new_qdisc
!= &noqueue_qdisc
)
596 *need_watchdog_p
= 1;
599 void dev_activate(struct net_device
*dev
)
603 /* No queueing discipline is attached to device;
604 create default one i.e. pfifo_fast for devices,
605 which need queueing and noqueue_qdisc for
609 if (dev_all_qdisc_sleeping_noop(dev
))
610 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
612 if (!netif_carrier_ok(dev
))
613 /* Delay activation until next carrier-on event */
617 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
620 dev
->trans_start
= jiffies
;
621 dev_watchdog_up(dev
);
625 static void dev_deactivate_queue(struct net_device
*dev
,
626 struct netdev_queue
*dev_queue
,
627 void *_qdisc_default
)
629 struct Qdisc
*qdisc_default
= _qdisc_default
;
630 struct sk_buff
*skb
= NULL
;
633 qdisc
= dev_queue
->qdisc
;
635 spin_lock_bh(qdisc_lock(qdisc
));
637 dev_queue
->qdisc
= qdisc_default
;
640 spin_unlock_bh(qdisc_lock(qdisc
));
646 static bool some_qdisc_is_running(struct net_device
*dev
, int lock
)
650 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
651 struct netdev_queue
*dev_queue
;
652 spinlock_t
*root_lock
;
656 dev_queue
= netdev_get_tx_queue(dev
, i
);
657 q
= dev_queue
->qdisc
;
658 root_lock
= qdisc_root_lock(q
);
661 spin_lock_bh(root_lock
);
663 val
= test_bit(__QDISC_STATE_RUNNING
, &q
->state
);
666 spin_unlock_bh(root_lock
);
674 void dev_deactivate(struct net_device
*dev
)
678 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
, &noop_qdisc
);
680 dev_watchdog_down(dev
);
682 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
685 /* Wait for outstanding qdisc_run calls. */
687 while (some_qdisc_is_running(dev
, 0))
691 * Double-check inside queue lock to ensure that all effects
692 * of the queue run are visible when we return.
694 running
= some_qdisc_is_running(dev
, 1);
697 * The running flag should never be set at this point because
698 * we've already set dev->qdisc to noop_qdisc *inside* the same
699 * pair of spin locks. That is, if any qdisc_run starts after
700 * our initial test it should see the noop_qdisc and then
701 * clear the RUNNING bit before dropping the queue lock. So
702 * if it is set here then we've found a bug.
704 } while (WARN_ON_ONCE(running
));
707 static void dev_init_scheduler_queue(struct net_device
*dev
,
708 struct netdev_queue
*dev_queue
,
711 struct Qdisc
*qdisc
= _qdisc
;
713 dev_queue
->qdisc
= qdisc
;
714 dev_queue
->qdisc_sleeping
= qdisc
;
717 void dev_init_scheduler(struct net_device
*dev
)
719 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
720 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
722 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
725 static void shutdown_scheduler_queue(struct net_device
*dev
,
726 struct netdev_queue
*dev_queue
,
727 void *_qdisc_default
)
729 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
730 struct Qdisc
*qdisc_default
= _qdisc_default
;
733 spinlock_t
*root_lock
= qdisc_root_lock(qdisc
);
735 dev_queue
->qdisc
= qdisc_default
;
736 dev_queue
->qdisc_sleeping
= qdisc_default
;
738 spin_lock(root_lock
);
739 qdisc_destroy(qdisc
);
740 spin_unlock(root_lock
);
744 void dev_shutdown(struct net_device
*dev
)
746 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
747 shutdown_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
748 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));