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 <linux/if_macvlan.h>
30 #include <net/sch_generic.h>
31 #include <net/pkt_sched.h>
33 #include <trace/events/qdisc.h>
35 /* Qdisc to use by default */
36 const struct Qdisc_ops
*default_qdisc_ops
= &pfifo_fast_ops
;
37 EXPORT_SYMBOL(default_qdisc_ops
);
39 /* Main transmission queue. */
41 /* Modifications to data participating in scheduling must be protected with
42 * qdisc_lock(qdisc) spinlock.
44 * The idea is the following:
45 * - enqueue, dequeue are serialized via qdisc root lock
46 * - ingress filtering is also serialized via qdisc root lock
47 * - updates to tree and tree walking are only done under the rtnl mutex.
50 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
54 qdisc_qstats_backlog_inc(q
, skb
);
55 q
->q
.qlen
++; /* it's still part of the queue */
61 static void try_bulk_dequeue_skb(struct Qdisc
*q
,
63 const struct netdev_queue
*txq
,
66 int bytelimit
= qdisc_avail_bulklimit(txq
) - skb
->len
;
68 while (bytelimit
> 0) {
69 struct sk_buff
*nskb
= q
->dequeue(q
);
74 bytelimit
-= nskb
->len
; /* covers GSO len */
77 (*packets
)++; /* GSO counts as one pkt */
82 /* This variant of try_bulk_dequeue_skb() makes sure
83 * all skbs in the chain are for the same txq
85 static void try_bulk_dequeue_skb_slow(struct Qdisc
*q
,
89 int mapping
= skb_get_queue_mapping(skb
);
97 if (unlikely(skb_get_queue_mapping(nskb
) != mapping
)) {
98 q
->skb_bad_txq
= nskb
;
99 qdisc_qstats_backlog_inc(q
, nskb
);
110 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
111 * A requeued skb (via q->gso_skb) can also be a SKB list.
113 static struct sk_buff
*dequeue_skb(struct Qdisc
*q
, bool *validate
,
116 struct sk_buff
*skb
= q
->gso_skb
;
117 const struct netdev_queue
*txq
= q
->dev_queue
;
121 /* skb in gso_skb were already validated */
123 /* check the reason of requeuing without tx lock first */
124 txq
= skb_get_tx_queue(txq
->dev
, skb
);
125 if (!netif_xmit_frozen_or_stopped(txq
)) {
127 qdisc_qstats_backlog_dec(q
, skb
);
134 skb
= q
->skb_bad_txq
;
136 /* check the reason of requeuing without tx lock first */
137 txq
= skb_get_tx_queue(txq
->dev
, skb
);
138 if (!netif_xmit_frozen_or_stopped(txq
)) {
139 q
->skb_bad_txq
= NULL
;
140 qdisc_qstats_backlog_dec(q
, skb
);
147 if (!(q
->flags
& TCQ_F_ONETXQUEUE
) ||
148 !netif_xmit_frozen_or_stopped(txq
))
152 if (qdisc_may_bulk(q
))
153 try_bulk_dequeue_skb(q
, skb
, txq
, packets
);
155 try_bulk_dequeue_skb_slow(q
, skb
, packets
);
158 trace_qdisc_dequeue(q
, txq
, *packets
, skb
);
163 * Transmit possibly several skbs, and handle the return status as
164 * required. Owning running seqcount bit guarantees that
165 * only one CPU can execute this function.
167 * Returns to the caller:
168 * 0 - queue is empty or throttled.
169 * >0 - queue is not empty.
171 int sch_direct_xmit(struct sk_buff
*skb
, struct Qdisc
*q
,
172 struct net_device
*dev
, struct netdev_queue
*txq
,
173 spinlock_t
*root_lock
, bool validate
)
175 int ret
= NETDEV_TX_BUSY
;
177 /* And release qdisc */
178 spin_unlock(root_lock
);
180 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
182 skb
= validate_xmit_skb_list(skb
, dev
);
185 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
186 if (!netif_xmit_frozen_or_stopped(txq
))
187 skb
= dev_hard_start_xmit(skb
, dev
, txq
, &ret
);
189 HARD_TX_UNLOCK(dev
, txq
);
191 spin_lock(root_lock
);
192 return qdisc_qlen(q
);
194 spin_lock(root_lock
);
196 if (dev_xmit_complete(ret
)) {
197 /* Driver sent out skb successfully or skb was consumed */
200 /* Driver returned NETDEV_TX_BUSY - requeue skb */
201 if (unlikely(ret
!= NETDEV_TX_BUSY
))
202 net_warn_ratelimited("BUG %s code %d qlen %d\n",
203 dev
->name
, ret
, q
->q
.qlen
);
205 ret
= dev_requeue_skb(skb
, q
);
208 if (ret
&& netif_xmit_frozen_or_stopped(txq
))
215 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
217 * running seqcount guarantees only one CPU can process
218 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
221 * netif_tx_lock serializes accesses to device driver.
223 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
224 * if one is grabbed, another must be free.
226 * Note, that this procedure can be called by a watchdog timer
228 * Returns to the caller:
229 * 0 - queue is empty or throttled.
230 * >0 - queue is not empty.
233 static inline int qdisc_restart(struct Qdisc
*q
, int *packets
)
235 struct netdev_queue
*txq
;
236 struct net_device
*dev
;
237 spinlock_t
*root_lock
;
242 skb
= dequeue_skb(q
, &validate
, packets
);
246 root_lock
= qdisc_lock(q
);
248 txq
= skb_get_tx_queue(dev
, skb
);
250 return sch_direct_xmit(skb
, q
, dev
, txq
, root_lock
, validate
);
253 void __qdisc_run(struct Qdisc
*q
)
255 int quota
= dev_tx_weight
;
258 while (qdisc_restart(q
, &packets
)) {
260 * Ordered by possible occurrence: Postpone processing if
261 * 1. we've exceeded packet quota
262 * 2. another process needs the CPU;
265 if (quota
<= 0 || need_resched()) {
274 unsigned long dev_trans_start(struct net_device
*dev
)
276 unsigned long val
, res
;
279 if (is_vlan_dev(dev
))
280 dev
= vlan_dev_real_dev(dev
);
281 else if (netif_is_macvlan(dev
))
282 dev
= macvlan_dev_real_dev(dev
);
283 res
= netdev_get_tx_queue(dev
, 0)->trans_start
;
284 for (i
= 1; i
< dev
->num_tx_queues
; i
++) {
285 val
= netdev_get_tx_queue(dev
, i
)->trans_start
;
286 if (val
&& time_after(val
, res
))
292 EXPORT_SYMBOL(dev_trans_start
);
294 static void dev_watchdog(struct timer_list
*t
)
296 struct net_device
*dev
= from_timer(dev
, t
, watchdog_timer
);
299 if (!qdisc_tx_is_noop(dev
)) {
300 if (netif_device_present(dev
) &&
301 netif_running(dev
) &&
302 netif_carrier_ok(dev
)) {
303 int some_queue_timedout
= 0;
305 unsigned long trans_start
;
307 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
308 struct netdev_queue
*txq
;
310 txq
= netdev_get_tx_queue(dev
, i
);
311 trans_start
= txq
->trans_start
;
312 if (netif_xmit_stopped(txq
) &&
313 time_after(jiffies
, (trans_start
+
314 dev
->watchdog_timeo
))) {
315 some_queue_timedout
= 1;
316 txq
->trans_timeout
++;
321 if (some_queue_timedout
) {
322 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
323 dev
->name
, netdev_drivername(dev
), i
);
324 dev
->netdev_ops
->ndo_tx_timeout(dev
);
326 if (!mod_timer(&dev
->watchdog_timer
,
327 round_jiffies(jiffies
+
328 dev
->watchdog_timeo
)))
332 netif_tx_unlock(dev
);
337 void __netdev_watchdog_up(struct net_device
*dev
)
339 if (dev
->netdev_ops
->ndo_tx_timeout
) {
340 if (dev
->watchdog_timeo
<= 0)
341 dev
->watchdog_timeo
= 5*HZ
;
342 if (!mod_timer(&dev
->watchdog_timer
,
343 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
348 static void dev_watchdog_up(struct net_device
*dev
)
350 __netdev_watchdog_up(dev
);
353 static void dev_watchdog_down(struct net_device
*dev
)
355 netif_tx_lock_bh(dev
);
356 if (del_timer(&dev
->watchdog_timer
))
358 netif_tx_unlock_bh(dev
);
362 * netif_carrier_on - set carrier
363 * @dev: network device
365 * Device has detected that carrier.
367 void netif_carrier_on(struct net_device
*dev
)
369 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
370 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
372 atomic_inc(&dev
->carrier_changes
);
373 linkwatch_fire_event(dev
);
374 if (netif_running(dev
))
375 __netdev_watchdog_up(dev
);
378 EXPORT_SYMBOL(netif_carrier_on
);
381 * netif_carrier_off - clear carrier
382 * @dev: network device
384 * Device has detected loss of carrier.
386 void netif_carrier_off(struct net_device
*dev
)
388 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
389 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
391 atomic_inc(&dev
->carrier_changes
);
392 linkwatch_fire_event(dev
);
395 EXPORT_SYMBOL(netif_carrier_off
);
397 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
398 under all circumstances. It is difficult to invent anything faster or
402 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
403 struct sk_buff
**to_free
)
405 __qdisc_drop(skb
, to_free
);
409 static struct sk_buff
*noop_dequeue(struct Qdisc
*qdisc
)
414 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
417 .enqueue
= noop_enqueue
,
418 .dequeue
= noop_dequeue
,
419 .peek
= noop_dequeue
,
420 .owner
= THIS_MODULE
,
423 static struct netdev_queue noop_netdev_queue
= {
424 .qdisc
= &noop_qdisc
,
425 .qdisc_sleeping
= &noop_qdisc
,
428 struct Qdisc noop_qdisc
= {
429 .enqueue
= noop_enqueue
,
430 .dequeue
= noop_dequeue
,
431 .flags
= TCQ_F_BUILTIN
,
432 .ops
= &noop_qdisc_ops
,
433 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
434 .dev_queue
= &noop_netdev_queue
,
435 .running
= SEQCNT_ZERO(noop_qdisc
.running
),
436 .busylock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.busylock
),
438 EXPORT_SYMBOL(noop_qdisc
);
440 static int noqueue_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
442 /* register_qdisc() assigns a default of noop_enqueue if unset,
443 * but __dev_queue_xmit() treats noqueue only as such
444 * if this is NULL - so clear it here. */
445 qdisc
->enqueue
= NULL
;
449 struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
452 .init
= noqueue_init
,
453 .enqueue
= noop_enqueue
,
454 .dequeue
= noop_dequeue
,
455 .peek
= noop_dequeue
,
456 .owner
= THIS_MODULE
,
459 static const u8 prio2band
[TC_PRIO_MAX
+ 1] = {
460 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
463 /* 3-band FIFO queue: old style, but should be a bit faster than
464 generic prio+fifo combination.
467 #define PFIFO_FAST_BANDS 3
470 * Private data for a pfifo_fast scheduler containing:
471 * - queues for the three band
472 * - bitmap indicating which of the bands contain skbs
474 struct pfifo_fast_priv
{
476 struct qdisc_skb_head q
[PFIFO_FAST_BANDS
];
480 * Convert a bitmap to the first band number where an skb is queued, where:
481 * bitmap=0 means there are no skbs on any band.
482 * bitmap=1 means there is an skb on band 0.
483 * bitmap=7 means there are skbs on all 3 bands, etc.
485 static const int bitmap2band
[] = {-1, 0, 1, 0, 2, 0, 1, 0};
487 static inline struct qdisc_skb_head
*band2list(struct pfifo_fast_priv
*priv
,
490 return priv
->q
+ band
;
493 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
494 struct sk_buff
**to_free
)
496 if (qdisc
->q
.qlen
< qdisc_dev(qdisc
)->tx_queue_len
) {
497 int band
= prio2band
[skb
->priority
& TC_PRIO_MAX
];
498 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
499 struct qdisc_skb_head
*list
= band2list(priv
, band
);
501 priv
->bitmap
|= (1 << band
);
503 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
506 return qdisc_drop(skb
, qdisc
, to_free
);
509 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
*qdisc
)
511 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
512 int band
= bitmap2band
[priv
->bitmap
];
514 if (likely(band
>= 0)) {
515 struct qdisc_skb_head
*qh
= band2list(priv
, band
);
516 struct sk_buff
*skb
= __qdisc_dequeue_head(qh
);
518 if (likely(skb
!= NULL
)) {
519 qdisc_qstats_backlog_dec(qdisc
, skb
);
520 qdisc_bstats_update(qdisc
, skb
);
525 priv
->bitmap
&= ~(1 << band
);
533 static struct sk_buff
*pfifo_fast_peek(struct Qdisc
*qdisc
)
535 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
536 int band
= bitmap2band
[priv
->bitmap
];
539 struct qdisc_skb_head
*qh
= band2list(priv
, band
);
547 static void pfifo_fast_reset(struct Qdisc
*qdisc
)
550 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
552 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
553 __qdisc_reset_queue(band2list(priv
, prio
));
556 qdisc
->qstats
.backlog
= 0;
560 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
562 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
564 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+ 1);
565 if (nla_put(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
))
566 goto nla_put_failure
;
573 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
576 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
578 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
579 qdisc_skb_head_init(band2list(priv
, prio
));
581 /* Can by-pass the queue discipline */
582 qdisc
->flags
|= TCQ_F_CAN_BYPASS
;
586 struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
588 .priv_size
= sizeof(struct pfifo_fast_priv
),
589 .enqueue
= pfifo_fast_enqueue
,
590 .dequeue
= pfifo_fast_dequeue
,
591 .peek
= pfifo_fast_peek
,
592 .init
= pfifo_fast_init
,
593 .reset
= pfifo_fast_reset
,
594 .dump
= pfifo_fast_dump
,
595 .owner
= THIS_MODULE
,
597 EXPORT_SYMBOL(pfifo_fast_ops
);
599 static struct lock_class_key qdisc_tx_busylock
;
600 static struct lock_class_key qdisc_running_key
;
602 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
603 const struct Qdisc_ops
*ops
)
607 unsigned int size
= QDISC_ALIGN(sizeof(*sch
)) + ops
->priv_size
;
609 struct net_device
*dev
;
616 dev
= dev_queue
->dev
;
617 p
= kzalloc_node(size
, GFP_KERNEL
,
618 netdev_queue_numa_node_read(dev_queue
));
622 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
623 /* if we got non aligned memory, ask more and do alignment ourself */
626 p
= kzalloc_node(size
+ QDISC_ALIGNTO
- 1, GFP_KERNEL
,
627 netdev_queue_numa_node_read(dev_queue
));
630 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
631 sch
->padded
= (char *) sch
- (char *) p
;
633 qdisc_skb_head_init(&sch
->q
);
634 spin_lock_init(&sch
->q
.lock
);
636 spin_lock_init(&sch
->busylock
);
637 lockdep_set_class(&sch
->busylock
,
638 dev
->qdisc_tx_busylock
?: &qdisc_tx_busylock
);
640 seqcount_init(&sch
->running
);
641 lockdep_set_class(&sch
->running
,
642 dev
->qdisc_running_key
?: &qdisc_running_key
);
645 sch
->enqueue
= ops
->enqueue
;
646 sch
->dequeue
= ops
->dequeue
;
647 sch
->dev_queue
= dev_queue
;
649 refcount_set(&sch
->refcnt
, 1);
656 struct Qdisc
*qdisc_create_dflt(struct netdev_queue
*dev_queue
,
657 const struct Qdisc_ops
*ops
,
658 unsigned int parentid
)
662 if (!try_module_get(ops
->owner
))
665 sch
= qdisc_alloc(dev_queue
, ops
);
667 module_put(ops
->owner
);
670 sch
->parent
= parentid
;
672 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
678 EXPORT_SYMBOL(qdisc_create_dflt
);
680 /* Under qdisc_lock(qdisc) and BH! */
682 void qdisc_reset(struct Qdisc
*qdisc
)
684 const struct Qdisc_ops
*ops
= qdisc
->ops
;
689 kfree_skb(qdisc
->skb_bad_txq
);
690 qdisc
->skb_bad_txq
= NULL
;
692 if (qdisc
->gso_skb
) {
693 kfree_skb_list(qdisc
->gso_skb
);
694 qdisc
->gso_skb
= NULL
;
697 qdisc
->qstats
.backlog
= 0;
699 EXPORT_SYMBOL(qdisc_reset
);
701 static void qdisc_free(struct Qdisc
*qdisc
)
703 if (qdisc_is_percpu_stats(qdisc
)) {
704 free_percpu(qdisc
->cpu_bstats
);
705 free_percpu(qdisc
->cpu_qstats
);
708 kfree((char *) qdisc
- qdisc
->padded
);
711 void qdisc_destroy(struct Qdisc
*qdisc
)
713 const struct Qdisc_ops
*ops
= qdisc
->ops
;
715 if (qdisc
->flags
& TCQ_F_BUILTIN
||
716 !refcount_dec_and_test(&qdisc
->refcnt
))
719 #ifdef CONFIG_NET_SCHED
720 qdisc_hash_del(qdisc
);
722 qdisc_put_stab(rtnl_dereference(qdisc
->stab
));
724 gen_kill_estimator(&qdisc
->rate_est
);
730 module_put(ops
->owner
);
731 dev_put(qdisc_dev(qdisc
));
733 kfree_skb_list(qdisc
->gso_skb
);
734 kfree_skb(qdisc
->skb_bad_txq
);
737 EXPORT_SYMBOL(qdisc_destroy
);
739 /* Attach toplevel qdisc to device queue. */
740 struct Qdisc
*dev_graft_qdisc(struct netdev_queue
*dev_queue
,
743 struct Qdisc
*oqdisc
= dev_queue
->qdisc_sleeping
;
744 spinlock_t
*root_lock
;
746 root_lock
= qdisc_lock(oqdisc
);
747 spin_lock_bh(root_lock
);
749 /* Prune old scheduler */
750 if (oqdisc
&& refcount_read(&oqdisc
->refcnt
) <= 1)
753 /* ... and graft new one */
756 dev_queue
->qdisc_sleeping
= qdisc
;
757 rcu_assign_pointer(dev_queue
->qdisc
, &noop_qdisc
);
759 spin_unlock_bh(root_lock
);
763 EXPORT_SYMBOL(dev_graft_qdisc
);
765 static void attach_one_default_qdisc(struct net_device
*dev
,
766 struct netdev_queue
*dev_queue
,
770 const struct Qdisc_ops
*ops
= default_qdisc_ops
;
772 if (dev
->priv_flags
& IFF_NO_QUEUE
)
773 ops
= &noqueue_qdisc_ops
;
775 qdisc
= qdisc_create_dflt(dev_queue
, ops
, TC_H_ROOT
);
777 netdev_info(dev
, "activation failed\n");
780 if (!netif_is_multiqueue(dev
))
781 qdisc
->flags
|= TCQ_F_ONETXQUEUE
| TCQ_F_NOPARENT
;
782 dev_queue
->qdisc_sleeping
= qdisc
;
785 static void attach_default_qdiscs(struct net_device
*dev
)
787 struct netdev_queue
*txq
;
790 txq
= netdev_get_tx_queue(dev
, 0);
792 if (!netif_is_multiqueue(dev
) ||
793 dev
->priv_flags
& IFF_NO_QUEUE
) {
794 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
795 dev
->qdisc
= txq
->qdisc_sleeping
;
796 qdisc_refcount_inc(dev
->qdisc
);
798 qdisc
= qdisc_create_dflt(txq
, &mq_qdisc_ops
, TC_H_ROOT
);
801 qdisc
->ops
->attach(qdisc
);
804 #ifdef CONFIG_NET_SCHED
805 if (dev
->qdisc
!= &noop_qdisc
)
806 qdisc_hash_add(dev
->qdisc
, false);
810 static void transition_one_qdisc(struct net_device
*dev
,
811 struct netdev_queue
*dev_queue
,
812 void *_need_watchdog
)
814 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
815 int *need_watchdog_p
= _need_watchdog
;
817 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
818 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
820 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
821 if (need_watchdog_p
) {
822 dev_queue
->trans_start
= 0;
823 *need_watchdog_p
= 1;
827 void dev_activate(struct net_device
*dev
)
831 /* No queueing discipline is attached to device;
832 * create default one for devices, which need queueing
833 * and noqueue_qdisc for virtual interfaces
836 if (dev
->qdisc
== &noop_qdisc
)
837 attach_default_qdiscs(dev
);
839 if (!netif_carrier_ok(dev
))
840 /* Delay activation until next carrier-on event */
844 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
845 if (dev_ingress_queue(dev
))
846 transition_one_qdisc(dev
, dev_ingress_queue(dev
), NULL
);
849 netif_trans_update(dev
);
850 dev_watchdog_up(dev
);
853 EXPORT_SYMBOL(dev_activate
);
855 static void dev_deactivate_queue(struct net_device
*dev
,
856 struct netdev_queue
*dev_queue
,
857 void *_qdisc_default
)
859 struct Qdisc
*qdisc_default
= _qdisc_default
;
862 qdisc
= rtnl_dereference(dev_queue
->qdisc
);
864 spin_lock_bh(qdisc_lock(qdisc
));
866 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
867 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
869 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
872 spin_unlock_bh(qdisc_lock(qdisc
));
876 static bool some_qdisc_is_busy(struct net_device
*dev
)
880 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
881 struct netdev_queue
*dev_queue
;
882 spinlock_t
*root_lock
;
886 dev_queue
= netdev_get_tx_queue(dev
, i
);
887 q
= dev_queue
->qdisc_sleeping
;
888 root_lock
= qdisc_lock(q
);
890 spin_lock_bh(root_lock
);
892 val
= (qdisc_is_running(q
) ||
893 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
895 spin_unlock_bh(root_lock
);
904 * dev_deactivate_many - deactivate transmissions on several devices
905 * @head: list of devices to deactivate
907 * This function returns only when all outstanding transmissions
908 * have completed, unless all devices are in dismantle phase.
910 void dev_deactivate_many(struct list_head
*head
)
912 struct net_device
*dev
;
913 bool sync_needed
= false;
915 list_for_each_entry(dev
, head
, close_list
) {
916 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
,
918 if (dev_ingress_queue(dev
))
919 dev_deactivate_queue(dev
, dev_ingress_queue(dev
),
922 dev_watchdog_down(dev
);
923 sync_needed
|= !dev
->dismantle
;
926 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
927 * This is avoided if all devices are in dismantle phase :
928 * Caller will call synchronize_net() for us
933 /* Wait for outstanding qdisc_run calls. */
934 list_for_each_entry(dev
, head
, close_list
)
935 while (some_qdisc_is_busy(dev
))
939 void dev_deactivate(struct net_device
*dev
)
943 list_add(&dev
->close_list
, &single
);
944 dev_deactivate_many(&single
);
947 EXPORT_SYMBOL(dev_deactivate
);
949 static void dev_init_scheduler_queue(struct net_device
*dev
,
950 struct netdev_queue
*dev_queue
,
953 struct Qdisc
*qdisc
= _qdisc
;
955 rcu_assign_pointer(dev_queue
->qdisc
, qdisc
);
956 dev_queue
->qdisc_sleeping
= qdisc
;
959 void dev_init_scheduler(struct net_device
*dev
)
961 dev
->qdisc
= &noop_qdisc
;
962 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
963 if (dev_ingress_queue(dev
))
964 dev_init_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
966 timer_setup(&dev
->watchdog_timer
, dev_watchdog
, 0);
969 static void shutdown_scheduler_queue(struct net_device
*dev
,
970 struct netdev_queue
*dev_queue
,
971 void *_qdisc_default
)
973 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
974 struct Qdisc
*qdisc_default
= _qdisc_default
;
977 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
978 dev_queue
->qdisc_sleeping
= qdisc_default
;
980 qdisc_destroy(qdisc
);
984 void dev_shutdown(struct net_device
*dev
)
986 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
987 if (dev_ingress_queue(dev
))
988 shutdown_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
989 qdisc_destroy(dev
->qdisc
);
990 dev
->qdisc
= &noop_qdisc
;
992 WARN_ON(timer_pending(&dev
->watchdog_timer
));
995 void psched_ratecfg_precompute(struct psched_ratecfg
*r
,
996 const struct tc_ratespec
*conf
,
999 memset(r
, 0, sizeof(*r
));
1000 r
->overhead
= conf
->overhead
;
1001 r
->rate_bytes_ps
= max_t(u64
, conf
->rate
, rate64
);
1002 r
->linklayer
= (conf
->linklayer
& TC_LINKLAYER_MASK
);
1005 * The deal here is to replace a divide by a reciprocal one
1006 * in fast path (a reciprocal divide is a multiply and a shift)
1008 * Normal formula would be :
1009 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1011 * We compute mult/shift to use instead :
1012 * time_in_ns = (len * mult) >> shift;
1014 * We try to get the highest possible mult value for accuracy,
1015 * but have to make sure no overflows will ever happen.
1017 if (r
->rate_bytes_ps
> 0) {
1018 u64 factor
= NSEC_PER_SEC
;
1021 r
->mult
= div64_u64(factor
, r
->rate_bytes_ps
);
1022 if (r
->mult
& (1U << 31) || factor
& (1ULL << 63))
1029 EXPORT_SYMBOL(psched_ratecfg_precompute
);
1031 static void mini_qdisc_rcu_func(struct rcu_head
*head
)
1035 void mini_qdisc_pair_swap(struct mini_Qdisc_pair
*miniqp
,
1036 struct tcf_proto
*tp_head
)
1038 struct mini_Qdisc
*miniq_old
= rtnl_dereference(*miniqp
->p_miniq
);
1039 struct mini_Qdisc
*miniq
;
1042 RCU_INIT_POINTER(*miniqp
->p_miniq
, NULL
);
1043 /* Wait for flying RCU callback before it is freed. */
1048 miniq
= !miniq_old
|| miniq_old
== &miniqp
->miniq2
?
1049 &miniqp
->miniq1
: &miniqp
->miniq2
;
1051 /* We need to make sure that readers won't see the miniq
1052 * we are about to modify. So wait until previous call_rcu_bh callback
1056 miniq
->filter_list
= tp_head
;
1057 rcu_assign_pointer(*miniqp
->p_miniq
, miniq
);
1060 /* This is counterpart of the rcu barriers above. We need to
1061 * block potential new user of miniq_old until all readers
1062 * are not seeing it.
1064 call_rcu_bh(&miniq_old
->rcu
, mini_qdisc_rcu_func
);
1066 EXPORT_SYMBOL(mini_qdisc_pair_swap
);
1068 void mini_qdisc_pair_init(struct mini_Qdisc_pair
*miniqp
, struct Qdisc
*qdisc
,
1069 struct mini_Qdisc __rcu
**p_miniq
)
1071 miniqp
->miniq1
.cpu_bstats
= qdisc
->cpu_bstats
;
1072 miniqp
->miniq1
.cpu_qstats
= qdisc
->cpu_qstats
;
1073 miniqp
->miniq2
.cpu_bstats
= qdisc
->cpu_bstats
;
1074 miniqp
->miniq2
.cpu_qstats
= qdisc
->cpu_qstats
;
1075 miniqp
->p_miniq
= p_miniq
;
1077 EXPORT_SYMBOL(mini_qdisc_pair_init
);