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 <linux/slab.h>
28 #include <linux/if_vlan.h>
29 #include <net/sch_generic.h>
30 #include <net/pkt_sched.h>
33 /* Qdisc to use by default */
34 const struct Qdisc_ops
*default_qdisc_ops
= &pfifo_fast_ops
;
35 EXPORT_SYMBOL(default_qdisc_ops
);
37 /* Main transmission queue. */
39 /* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
42 * The idea is the following:
43 * - enqueue, dequeue are serialized via qdisc root lock
44 * - ingress filtering is also serialized via qdisc root lock
45 * - updates to tree and tree walking are only done under the rtnl mutex.
48 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
52 q
->q
.qlen
++; /* it's still part of the queue */
58 static void try_bulk_dequeue_skb(struct Qdisc
*q
,
60 const struct netdev_queue
*txq
,
63 int bytelimit
= qdisc_avail_bulklimit(txq
) - skb
->len
;
65 while (bytelimit
> 0) {
66 struct sk_buff
*nskb
= q
->dequeue(q
);
71 bytelimit
-= nskb
->len
; /* covers GSO len */
74 (*packets
)++; /* GSO counts as one pkt */
79 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
80 * A requeued skb (via q->gso_skb) can also be a SKB list.
82 static struct sk_buff
*dequeue_skb(struct Qdisc
*q
, bool *validate
,
85 struct sk_buff
*skb
= q
->gso_skb
;
86 const struct netdev_queue
*txq
= q
->dev_queue
;
91 /* check the reason of requeuing without tx lock first */
92 txq
= skb_get_tx_queue(txq
->dev
, skb
);
93 if (!netif_xmit_frozen_or_stopped(txq
)) {
98 /* skb in gso_skb were already validated */
101 if (!(q
->flags
& TCQ_F_ONETXQUEUE
) ||
102 !netif_xmit_frozen_or_stopped(txq
)) {
104 if (skb
&& qdisc_may_bulk(q
))
105 try_bulk_dequeue_skb(q
, skb
, txq
, packets
);
111 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
112 struct netdev_queue
*dev_queue
,
117 if (unlikely(dev_queue
->xmit_lock_owner
== smp_processor_id())) {
119 * Same CPU holding the lock. It may be a transient
120 * configuration error, when hard_start_xmit() recurses. We
121 * detect it by checking xmit owner and drop the packet when
122 * deadloop is detected. Return OK to try the next skb.
125 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
126 dev_queue
->dev
->name
);
130 * Another cpu is holding lock, requeue & delay xmits for
133 __this_cpu_inc(softnet_data
.cpu_collision
);
134 ret
= dev_requeue_skb(skb
, q
);
141 * Transmit possibly several skbs, and handle the return status as
142 * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
143 * only one CPU can execute this function.
145 * Returns to the caller:
146 * 0 - queue is empty or throttled.
147 * >0 - queue is not empty.
149 int sch_direct_xmit(struct sk_buff
*skb
, struct Qdisc
*q
,
150 struct net_device
*dev
, struct netdev_queue
*txq
,
151 spinlock_t
*root_lock
, bool validate
)
153 int ret
= NETDEV_TX_BUSY
;
155 /* And release qdisc */
156 spin_unlock(root_lock
);
158 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
160 skb
= validate_xmit_skb_list(skb
, dev
);
163 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
164 if (!netif_xmit_frozen_or_stopped(txq
))
165 skb
= dev_hard_start_xmit(skb
, dev
, txq
, &ret
);
167 HARD_TX_UNLOCK(dev
, txq
);
169 spin_lock(root_lock
);
171 if (dev_xmit_complete(ret
)) {
172 /* Driver sent out skb successfully or skb was consumed */
174 } else if (ret
== NETDEV_TX_LOCKED
) {
175 /* Driver try lock failed */
176 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
178 /* Driver returned NETDEV_TX_BUSY - requeue skb */
179 if (unlikely(ret
!= NETDEV_TX_BUSY
))
180 net_warn_ratelimited("BUG %s code %d qlen %d\n",
181 dev
->name
, ret
, q
->q
.qlen
);
183 ret
= dev_requeue_skb(skb
, q
);
186 if (ret
&& netif_xmit_frozen_or_stopped(txq
))
193 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
195 * __QDISC___STATE_RUNNING guarantees only one CPU can process
196 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
199 * netif_tx_lock serializes accesses to device driver.
201 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
202 * if one is grabbed, another must be free.
204 * Note, that this procedure can be called by a watchdog timer
206 * Returns to the caller:
207 * 0 - queue is empty or throttled.
208 * >0 - queue is not empty.
211 static inline int qdisc_restart(struct Qdisc
*q
, int *packets
)
213 struct netdev_queue
*txq
;
214 struct net_device
*dev
;
215 spinlock_t
*root_lock
;
220 skb
= dequeue_skb(q
, &validate
, packets
);
224 root_lock
= qdisc_lock(q
);
226 txq
= skb_get_tx_queue(dev
, skb
);
228 return sch_direct_xmit(skb
, q
, dev
, txq
, root_lock
, validate
);
231 void __qdisc_run(struct Qdisc
*q
)
233 int quota
= weight_p
;
236 while (qdisc_restart(q
, &packets
)) {
238 * Ordered by possible occurrence: Postpone processing if
239 * 1. we've exceeded packet quota
240 * 2. another process needs the CPU;
243 if (quota
<= 0 || need_resched()) {
252 unsigned long dev_trans_start(struct net_device
*dev
)
254 unsigned long val
, res
;
257 if (is_vlan_dev(dev
))
258 dev
= vlan_dev_real_dev(dev
);
259 res
= dev
->trans_start
;
260 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
261 val
= netdev_get_tx_queue(dev
, i
)->trans_start
;
262 if (val
&& time_after(val
, res
))
265 dev
->trans_start
= res
;
269 EXPORT_SYMBOL(dev_trans_start
);
271 static void dev_watchdog(unsigned long arg
)
273 struct net_device
*dev
= (struct net_device
*)arg
;
276 if (!qdisc_tx_is_noop(dev
)) {
277 if (netif_device_present(dev
) &&
278 netif_running(dev
) &&
279 netif_carrier_ok(dev
)) {
280 int some_queue_timedout
= 0;
282 unsigned long trans_start
;
284 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
285 struct netdev_queue
*txq
;
287 txq
= netdev_get_tx_queue(dev
, i
);
289 * old device drivers set dev->trans_start
291 trans_start
= txq
->trans_start
? : dev
->trans_start
;
292 if (netif_xmit_stopped(txq
) &&
293 time_after(jiffies
, (trans_start
+
294 dev
->watchdog_timeo
))) {
295 some_queue_timedout
= 1;
296 txq
->trans_timeout
++;
301 if (some_queue_timedout
) {
302 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
303 dev
->name
, netdev_drivername(dev
), i
);
304 dev
->netdev_ops
->ndo_tx_timeout(dev
);
306 if (!mod_timer(&dev
->watchdog_timer
,
307 round_jiffies(jiffies
+
308 dev
->watchdog_timeo
)))
312 netif_tx_unlock(dev
);
317 void __netdev_watchdog_up(struct net_device
*dev
)
319 if (dev
->netdev_ops
->ndo_tx_timeout
) {
320 if (dev
->watchdog_timeo
<= 0)
321 dev
->watchdog_timeo
= 5*HZ
;
322 if (!mod_timer(&dev
->watchdog_timer
,
323 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
328 static void dev_watchdog_up(struct net_device
*dev
)
330 __netdev_watchdog_up(dev
);
333 static void dev_watchdog_down(struct net_device
*dev
)
335 netif_tx_lock_bh(dev
);
336 if (del_timer(&dev
->watchdog_timer
))
338 netif_tx_unlock_bh(dev
);
342 * netif_carrier_on - set carrier
343 * @dev: network device
345 * Device has detected that carrier.
347 void netif_carrier_on(struct net_device
*dev
)
349 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
350 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
352 atomic_inc(&dev
->carrier_changes
);
353 linkwatch_fire_event(dev
);
354 if (netif_running(dev
))
355 __netdev_watchdog_up(dev
);
358 EXPORT_SYMBOL(netif_carrier_on
);
361 * netif_carrier_off - clear carrier
362 * @dev: network device
364 * Device has detected loss of carrier.
366 void netif_carrier_off(struct net_device
*dev
)
368 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
369 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
371 atomic_inc(&dev
->carrier_changes
);
372 linkwatch_fire_event(dev
);
375 EXPORT_SYMBOL(netif_carrier_off
);
377 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
378 under all circumstances. It is difficult to invent anything faster or
382 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
)
388 static struct sk_buff
*noop_dequeue(struct Qdisc
*qdisc
)
393 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
396 .enqueue
= noop_enqueue
,
397 .dequeue
= noop_dequeue
,
398 .peek
= noop_dequeue
,
399 .owner
= THIS_MODULE
,
402 static struct netdev_queue noop_netdev_queue
= {
403 .qdisc
= &noop_qdisc
,
404 .qdisc_sleeping
= &noop_qdisc
,
407 struct Qdisc noop_qdisc
= {
408 .enqueue
= noop_enqueue
,
409 .dequeue
= noop_dequeue
,
410 .flags
= TCQ_F_BUILTIN
,
411 .ops
= &noop_qdisc_ops
,
412 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
413 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
414 .dev_queue
= &noop_netdev_queue
,
415 .busylock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.busylock
),
417 EXPORT_SYMBOL(noop_qdisc
);
419 static int noqueue_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
421 /* register_qdisc() assigns a default of noop_enqueue if unset,
422 * but __dev_queue_xmit() treats noqueue only as such
423 * if this is NULL - so clear it here. */
424 qdisc
->enqueue
= NULL
;
428 struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
431 .init
= noqueue_init
,
432 .enqueue
= noop_enqueue
,
433 .dequeue
= noop_dequeue
,
434 .peek
= noop_dequeue
,
435 .owner
= THIS_MODULE
,
438 static const u8 prio2band
[TC_PRIO_MAX
+ 1] = {
439 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
442 /* 3-band FIFO queue: old style, but should be a bit faster than
443 generic prio+fifo combination.
446 #define PFIFO_FAST_BANDS 3
449 * Private data for a pfifo_fast scheduler containing:
450 * - queues for the three band
451 * - bitmap indicating which of the bands contain skbs
453 struct pfifo_fast_priv
{
455 struct sk_buff_head q
[PFIFO_FAST_BANDS
];
459 * Convert a bitmap to the first band number where an skb is queued, where:
460 * bitmap=0 means there are no skbs on any band.
461 * bitmap=1 means there is an skb on band 0.
462 * bitmap=7 means there are skbs on all 3 bands, etc.
464 static const int bitmap2band
[] = {-1, 0, 1, 0, 2, 0, 1, 0};
466 static inline struct sk_buff_head
*band2list(struct pfifo_fast_priv
*priv
,
469 return priv
->q
+ band
;
472 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
)
474 if (skb_queue_len(&qdisc
->q
) < qdisc_dev(qdisc
)->tx_queue_len
) {
475 int band
= prio2band
[skb
->priority
& TC_PRIO_MAX
];
476 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
477 struct sk_buff_head
*list
= band2list(priv
, band
);
479 priv
->bitmap
|= (1 << band
);
481 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
484 return qdisc_drop(skb
, qdisc
);
487 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
*qdisc
)
489 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
490 int band
= bitmap2band
[priv
->bitmap
];
492 if (likely(band
>= 0)) {
493 struct sk_buff_head
*list
= band2list(priv
, band
);
494 struct sk_buff
*skb
= __qdisc_dequeue_head(qdisc
, list
);
497 if (skb_queue_empty(list
))
498 priv
->bitmap
&= ~(1 << band
);
506 static struct sk_buff
*pfifo_fast_peek(struct Qdisc
*qdisc
)
508 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
509 int band
= bitmap2band
[priv
->bitmap
];
512 struct sk_buff_head
*list
= band2list(priv
, band
);
514 return skb_peek(list
);
520 static void pfifo_fast_reset(struct Qdisc
*qdisc
)
523 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
525 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
526 __qdisc_reset_queue(qdisc
, band2list(priv
, prio
));
529 qdisc
->qstats
.backlog
= 0;
533 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
535 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
537 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+ 1);
538 if (nla_put(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
))
539 goto nla_put_failure
;
546 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
549 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
551 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
552 __skb_queue_head_init(band2list(priv
, prio
));
554 /* Can by-pass the queue discipline */
555 qdisc
->flags
|= TCQ_F_CAN_BYPASS
;
559 struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
561 .priv_size
= sizeof(struct pfifo_fast_priv
),
562 .enqueue
= pfifo_fast_enqueue
,
563 .dequeue
= pfifo_fast_dequeue
,
564 .peek
= pfifo_fast_peek
,
565 .init
= pfifo_fast_init
,
566 .reset
= pfifo_fast_reset
,
567 .dump
= pfifo_fast_dump
,
568 .owner
= THIS_MODULE
,
570 EXPORT_SYMBOL(pfifo_fast_ops
);
572 static struct lock_class_key qdisc_tx_busylock
;
574 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
575 const struct Qdisc_ops
*ops
)
579 unsigned int size
= QDISC_ALIGN(sizeof(*sch
)) + ops
->priv_size
;
581 struct net_device
*dev
= dev_queue
->dev
;
583 p
= kzalloc_node(size
, GFP_KERNEL
,
584 netdev_queue_numa_node_read(dev_queue
));
588 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
589 /* if we got non aligned memory, ask more and do alignment ourself */
592 p
= kzalloc_node(size
+ QDISC_ALIGNTO
- 1, GFP_KERNEL
,
593 netdev_queue_numa_node_read(dev_queue
));
596 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
597 sch
->padded
= (char *) sch
- (char *) p
;
599 INIT_LIST_HEAD(&sch
->list
);
600 skb_queue_head_init(&sch
->q
);
602 spin_lock_init(&sch
->busylock
);
603 lockdep_set_class(&sch
->busylock
,
604 dev
->qdisc_tx_busylock
?: &qdisc_tx_busylock
);
607 sch
->enqueue
= ops
->enqueue
;
608 sch
->dequeue
= ops
->dequeue
;
609 sch
->dev_queue
= dev_queue
;
611 atomic_set(&sch
->refcnt
, 1);
618 struct Qdisc
*qdisc_create_dflt(struct netdev_queue
*dev_queue
,
619 const struct Qdisc_ops
*ops
,
620 unsigned int parentid
)
624 if (!try_module_get(ops
->owner
))
627 sch
= qdisc_alloc(dev_queue
, ops
);
630 sch
->parent
= parentid
;
632 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
639 EXPORT_SYMBOL(qdisc_create_dflt
);
641 /* Under qdisc_lock(qdisc) and BH! */
643 void qdisc_reset(struct Qdisc
*qdisc
)
645 const struct Qdisc_ops
*ops
= qdisc
->ops
;
650 if (qdisc
->gso_skb
) {
651 kfree_skb_list(qdisc
->gso_skb
);
652 qdisc
->gso_skb
= NULL
;
656 EXPORT_SYMBOL(qdisc_reset
);
658 static void qdisc_rcu_free(struct rcu_head
*head
)
660 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, rcu_head
);
662 if (qdisc_is_percpu_stats(qdisc
)) {
663 free_percpu(qdisc
->cpu_bstats
);
664 free_percpu(qdisc
->cpu_qstats
);
667 kfree((char *) qdisc
- qdisc
->padded
);
670 void qdisc_destroy(struct Qdisc
*qdisc
)
672 const struct Qdisc_ops
*ops
= qdisc
->ops
;
674 if (qdisc
->flags
& TCQ_F_BUILTIN
||
675 !atomic_dec_and_test(&qdisc
->refcnt
))
678 #ifdef CONFIG_NET_SCHED
679 qdisc_list_del(qdisc
);
681 qdisc_put_stab(rtnl_dereference(qdisc
->stab
));
683 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
689 module_put(ops
->owner
);
690 dev_put(qdisc_dev(qdisc
));
692 kfree_skb_list(qdisc
->gso_skb
);
694 * gen_estimator est_timer() might access qdisc->q.lock,
695 * wait a RCU grace period before freeing qdisc.
697 call_rcu(&qdisc
->rcu_head
, qdisc_rcu_free
);
699 EXPORT_SYMBOL(qdisc_destroy
);
701 /* Attach toplevel qdisc to device queue. */
702 struct Qdisc
*dev_graft_qdisc(struct netdev_queue
*dev_queue
,
705 struct Qdisc
*oqdisc
= dev_queue
->qdisc_sleeping
;
706 spinlock_t
*root_lock
;
708 root_lock
= qdisc_lock(oqdisc
);
709 spin_lock_bh(root_lock
);
711 /* Prune old scheduler */
712 if (oqdisc
&& atomic_read(&oqdisc
->refcnt
) <= 1)
715 /* ... and graft new one */
718 dev_queue
->qdisc_sleeping
= qdisc
;
719 rcu_assign_pointer(dev_queue
->qdisc
, &noop_qdisc
);
721 spin_unlock_bh(root_lock
);
725 EXPORT_SYMBOL(dev_graft_qdisc
);
727 static void attach_one_default_qdisc(struct net_device
*dev
,
728 struct netdev_queue
*dev_queue
,
732 const struct Qdisc_ops
*ops
= default_qdisc_ops
;
734 if (dev
->priv_flags
& IFF_NO_QUEUE
)
735 ops
= &noqueue_qdisc_ops
;
737 qdisc
= qdisc_create_dflt(dev_queue
, ops
, TC_H_ROOT
);
739 netdev_info(dev
, "activation failed\n");
742 if (!netif_is_multiqueue(dev
))
743 qdisc
->flags
|= TCQ_F_ONETXQUEUE
| TCQ_F_NOPARENT
;
744 dev_queue
->qdisc_sleeping
= qdisc
;
747 static void attach_default_qdiscs(struct net_device
*dev
)
749 struct netdev_queue
*txq
;
752 txq
= netdev_get_tx_queue(dev
, 0);
754 if (!netif_is_multiqueue(dev
) ||
755 dev
->priv_flags
& IFF_NO_QUEUE
) {
756 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
757 dev
->qdisc
= txq
->qdisc_sleeping
;
758 atomic_inc(&dev
->qdisc
->refcnt
);
760 qdisc
= qdisc_create_dflt(txq
, &mq_qdisc_ops
, TC_H_ROOT
);
763 qdisc
->ops
->attach(qdisc
);
768 static void transition_one_qdisc(struct net_device
*dev
,
769 struct netdev_queue
*dev_queue
,
770 void *_need_watchdog
)
772 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
773 int *need_watchdog_p
= _need_watchdog
;
775 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
776 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
778 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
779 if (need_watchdog_p
) {
780 dev_queue
->trans_start
= 0;
781 *need_watchdog_p
= 1;
785 void dev_activate(struct net_device
*dev
)
789 /* No queueing discipline is attached to device;
790 * create default one for devices, which need queueing
791 * and noqueue_qdisc for virtual interfaces
794 if (dev
->qdisc
== &noop_qdisc
)
795 attach_default_qdiscs(dev
);
797 if (!netif_carrier_ok(dev
))
798 /* Delay activation until next carrier-on event */
802 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
803 if (dev_ingress_queue(dev
))
804 transition_one_qdisc(dev
, dev_ingress_queue(dev
), NULL
);
807 dev
->trans_start
= jiffies
;
808 dev_watchdog_up(dev
);
811 EXPORT_SYMBOL(dev_activate
);
813 static void dev_deactivate_queue(struct net_device
*dev
,
814 struct netdev_queue
*dev_queue
,
815 void *_qdisc_default
)
817 struct Qdisc
*qdisc_default
= _qdisc_default
;
820 qdisc
= rtnl_dereference(dev_queue
->qdisc
);
822 spin_lock_bh(qdisc_lock(qdisc
));
824 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
825 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
827 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
830 spin_unlock_bh(qdisc_lock(qdisc
));
834 static bool some_qdisc_is_busy(struct net_device
*dev
)
838 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
839 struct netdev_queue
*dev_queue
;
840 spinlock_t
*root_lock
;
844 dev_queue
= netdev_get_tx_queue(dev
, i
);
845 q
= dev_queue
->qdisc_sleeping
;
846 root_lock
= qdisc_lock(q
);
848 spin_lock_bh(root_lock
);
850 val
= (qdisc_is_running(q
) ||
851 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
853 spin_unlock_bh(root_lock
);
862 * dev_deactivate_many - deactivate transmissions on several devices
863 * @head: list of devices to deactivate
865 * This function returns only when all outstanding transmissions
866 * have completed, unless all devices are in dismantle phase.
868 void dev_deactivate_many(struct list_head
*head
)
870 struct net_device
*dev
;
871 bool sync_needed
= false;
873 list_for_each_entry(dev
, head
, close_list
) {
874 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
,
876 if (dev_ingress_queue(dev
))
877 dev_deactivate_queue(dev
, dev_ingress_queue(dev
),
880 dev_watchdog_down(dev
);
881 sync_needed
|= !dev
->dismantle
;
884 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
885 * This is avoided if all devices are in dismantle phase :
886 * Caller will call synchronize_net() for us
891 /* Wait for outstanding qdisc_run calls. */
892 list_for_each_entry(dev
, head
, close_list
)
893 while (some_qdisc_is_busy(dev
))
897 void dev_deactivate(struct net_device
*dev
)
901 list_add(&dev
->close_list
, &single
);
902 dev_deactivate_many(&single
);
905 EXPORT_SYMBOL(dev_deactivate
);
907 static void dev_init_scheduler_queue(struct net_device
*dev
,
908 struct netdev_queue
*dev_queue
,
911 struct Qdisc
*qdisc
= _qdisc
;
913 rcu_assign_pointer(dev_queue
->qdisc
, qdisc
);
914 dev_queue
->qdisc_sleeping
= qdisc
;
917 void dev_init_scheduler(struct net_device
*dev
)
919 dev
->qdisc
= &noop_qdisc
;
920 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
921 if (dev_ingress_queue(dev
))
922 dev_init_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
924 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
927 static void shutdown_scheduler_queue(struct net_device
*dev
,
928 struct netdev_queue
*dev_queue
,
929 void *_qdisc_default
)
931 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
932 struct Qdisc
*qdisc_default
= _qdisc_default
;
935 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
936 dev_queue
->qdisc_sleeping
= qdisc_default
;
938 qdisc_destroy(qdisc
);
942 void dev_shutdown(struct net_device
*dev
)
944 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
945 if (dev_ingress_queue(dev
))
946 shutdown_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
947 qdisc_destroy(dev
->qdisc
);
948 dev
->qdisc
= &noop_qdisc
;
950 WARN_ON(timer_pending(&dev
->watchdog_timer
));
953 void psched_ratecfg_precompute(struct psched_ratecfg
*r
,
954 const struct tc_ratespec
*conf
,
957 memset(r
, 0, sizeof(*r
));
958 r
->overhead
= conf
->overhead
;
959 r
->rate_bytes_ps
= max_t(u64
, conf
->rate
, rate64
);
960 r
->linklayer
= (conf
->linklayer
& TC_LINKLAYER_MASK
);
963 * The deal here is to replace a divide by a reciprocal one
964 * in fast path (a reciprocal divide is a multiply and a shift)
966 * Normal formula would be :
967 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
969 * We compute mult/shift to use instead :
970 * time_in_ns = (len * mult) >> shift;
972 * We try to get the highest possible mult value for accuracy,
973 * but have to make sure no overflows will ever happen.
975 if (r
->rate_bytes_ps
> 0) {
976 u64 factor
= NSEC_PER_SEC
;
979 r
->mult
= div64_u64(factor
, r
->rate_bytes_ps
);
980 if (r
->mult
& (1U << 31) || factor
& (1ULL << 63))
987 EXPORT_SYMBOL(psched_ratecfg_precompute
);