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
))) {
216 printk(KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit timed out\n",
217 dev
->name
, netdev_drivername(dev
, drivername
, 64));
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
;
344 static struct netdev_queue noqueue_netdev_queue
= {
345 .qdisc
= &noqueue_qdisc
,
348 static struct Qdisc noqueue_qdisc
= {
350 .dequeue
= noop_dequeue
,
351 .flags
= TCQ_F_BUILTIN
,
352 .ops
= &noqueue_qdisc_ops
,
353 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
354 .q
.lock
= __SPIN_LOCK_UNLOCKED(noqueue_qdisc
.q
.lock
),
355 .dev_queue
= &noqueue_netdev_queue
,
359 static const u8 prio2band
[TC_PRIO_MAX
+1] =
360 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
362 /* 3-band FIFO queue: old style, but should be a bit faster than
363 generic prio+fifo combination.
366 #define PFIFO_FAST_BANDS 3
368 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
371 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
372 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
375 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
377 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
379 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
381 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
384 return qdisc_drop(skb
, qdisc
);
387 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
390 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
392 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
393 if (!skb_queue_empty(list
+ prio
)) {
395 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
402 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
405 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
408 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
411 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
413 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
414 __qdisc_reset_queue(qdisc
, list
+ prio
);
416 qdisc
->qstats
.backlog
= 0;
420 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
422 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
424 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
425 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
432 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
435 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
437 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
438 skb_queue_head_init(list
+ prio
);
443 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
445 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
446 .enqueue
= pfifo_fast_enqueue
,
447 .dequeue
= pfifo_fast_dequeue
,
448 .requeue
= pfifo_fast_requeue
,
449 .init
= pfifo_fast_init
,
450 .reset
= pfifo_fast_reset
,
451 .dump
= pfifo_fast_dump
,
452 .owner
= THIS_MODULE
,
455 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
456 struct Qdisc_ops
*ops
)
463 /* ensure that the Qdisc and the private data are 32-byte aligned */
464 size
= QDISC_ALIGN(sizeof(*sch
));
465 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
467 p
= kzalloc(size
, GFP_KERNEL
);
470 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
471 sch
->padded
= (char *) sch
- (char *) p
;
473 INIT_LIST_HEAD(&sch
->list
);
474 skb_queue_head_init(&sch
->q
);
476 sch
->enqueue
= ops
->enqueue
;
477 sch
->dequeue
= ops
->dequeue
;
478 sch
->dev_queue
= dev_queue
;
479 dev_hold(qdisc_dev(sch
));
480 atomic_set(&sch
->refcnt
, 1);
487 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
488 struct netdev_queue
*dev_queue
,
489 struct Qdisc_ops
*ops
,
490 unsigned int parentid
)
494 sch
= qdisc_alloc(dev_queue
, ops
);
497 sch
->parent
= parentid
;
499 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
506 EXPORT_SYMBOL(qdisc_create_dflt
);
508 /* Under qdisc_root_lock(qdisc) and BH! */
510 void qdisc_reset(struct Qdisc
*qdisc
)
512 const struct Qdisc_ops
*ops
= qdisc
->ops
;
517 EXPORT_SYMBOL(qdisc_reset
);
519 /* this is the rcu callback function to clean up a qdisc when there
520 * are no further references to it */
522 static void __qdisc_destroy(struct rcu_head
*head
)
524 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
525 const struct Qdisc_ops
*ops
= qdisc
->ops
;
527 #ifdef CONFIG_NET_SCHED
528 qdisc_put_stab(qdisc
->stab
);
530 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
536 module_put(ops
->owner
);
537 dev_put(qdisc_dev(qdisc
));
539 kfree_skb(qdisc
->gso_skb
);
541 kfree((char *) qdisc
- qdisc
->padded
);
544 /* Under qdisc_root_lock(qdisc) and BH! */
546 void qdisc_destroy(struct Qdisc
*qdisc
)
548 if (qdisc
->flags
& TCQ_F_BUILTIN
||
549 !atomic_dec_and_test(&qdisc
->refcnt
))
553 list_del(&qdisc
->list
);
555 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
557 EXPORT_SYMBOL(qdisc_destroy
);
559 static bool dev_all_qdisc_sleeping_noop(struct net_device
*dev
)
563 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
564 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, i
);
566 if (txq
->qdisc_sleeping
!= &noop_qdisc
)
572 static void attach_one_default_qdisc(struct net_device
*dev
,
573 struct netdev_queue
*dev_queue
,
578 if (dev
->tx_queue_len
) {
579 qdisc
= qdisc_create_dflt(dev
, dev_queue
,
580 &pfifo_fast_ops
, TC_H_ROOT
);
582 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
586 qdisc
= &noqueue_qdisc
;
588 dev_queue
->qdisc_sleeping
= qdisc
;
591 static void transition_one_qdisc(struct net_device
*dev
,
592 struct netdev_queue
*dev_queue
,
593 void *_need_watchdog
)
595 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
596 int *need_watchdog_p
= _need_watchdog
;
598 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
599 if (new_qdisc
!= &noqueue_qdisc
)
600 *need_watchdog_p
= 1;
603 void dev_activate(struct net_device
*dev
)
607 /* No queueing discipline is attached to device;
608 create default one i.e. pfifo_fast for devices,
609 which need queueing and noqueue_qdisc for
613 if (dev_all_qdisc_sleeping_noop(dev
))
614 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
616 if (!netif_carrier_ok(dev
))
617 /* Delay activation until next carrier-on event */
621 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
624 dev
->trans_start
= jiffies
;
625 dev_watchdog_up(dev
);
629 static void dev_deactivate_queue(struct net_device
*dev
,
630 struct netdev_queue
*dev_queue
,
631 void *_qdisc_default
)
633 struct Qdisc
*qdisc_default
= _qdisc_default
;
636 qdisc
= dev_queue
->qdisc
;
638 spin_lock_bh(qdisc_lock(qdisc
));
640 dev_queue
->qdisc
= qdisc_default
;
643 spin_unlock_bh(qdisc_lock(qdisc
));
647 static bool some_qdisc_is_running(struct net_device
*dev
, int lock
)
651 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
652 struct netdev_queue
*dev_queue
;
653 spinlock_t
*root_lock
;
657 dev_queue
= netdev_get_tx_queue(dev
, i
);
658 q
= dev_queue
->qdisc
;
659 root_lock
= qdisc_root_lock(q
);
662 spin_lock_bh(root_lock
);
664 val
= test_bit(__QDISC_STATE_RUNNING
, &q
->state
);
667 spin_unlock_bh(root_lock
);
675 void dev_deactivate(struct net_device
*dev
)
679 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
, &noop_qdisc
);
681 dev_watchdog_down(dev
);
683 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
686 /* Wait for outstanding qdisc_run calls. */
688 while (some_qdisc_is_running(dev
, 0))
692 * Double-check inside queue lock to ensure that all effects
693 * of the queue run are visible when we return.
695 running
= some_qdisc_is_running(dev
, 1);
698 * The running flag should never be set at this point because
699 * we've already set dev->qdisc to noop_qdisc *inside* the same
700 * pair of spin locks. That is, if any qdisc_run starts after
701 * our initial test it should see the noop_qdisc and then
702 * clear the RUNNING bit before dropping the queue lock. So
703 * if it is set here then we've found a bug.
705 } while (WARN_ON_ONCE(running
));
708 static void dev_init_scheduler_queue(struct net_device
*dev
,
709 struct netdev_queue
*dev_queue
,
712 struct Qdisc
*qdisc
= _qdisc
;
714 dev_queue
->qdisc
= qdisc
;
715 dev_queue
->qdisc_sleeping
= qdisc
;
718 void dev_init_scheduler(struct net_device
*dev
)
720 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
721 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
723 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
726 static void shutdown_scheduler_queue(struct net_device
*dev
,
727 struct netdev_queue
*dev_queue
,
728 void *_qdisc_default
)
730 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
731 struct Qdisc
*qdisc_default
= _qdisc_default
;
734 spinlock_t
*root_lock
= qdisc_root_lock(qdisc
);
736 dev_queue
->qdisc
= qdisc_default
;
737 dev_queue
->qdisc_sleeping
= qdisc_default
;
739 spin_lock(root_lock
);
740 qdisc_destroy(qdisc
);
741 spin_unlock(root_lock
);
745 void dev_shutdown(struct net_device
*dev
)
747 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
748 shutdown_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
749 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));