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/skb_array.h>
30 #include <linux/if_macvlan.h>
31 #include <net/sch_generic.h>
32 #include <net/pkt_sched.h>
34 #include <trace/events/qdisc.h>
37 /* Qdisc to use by default */
38 const struct Qdisc_ops
*default_qdisc_ops
= &pfifo_fast_ops
;
39 EXPORT_SYMBOL(default_qdisc_ops
);
41 /* Main transmission queue. */
43 /* Modifications to data participating in scheduling must be protected with
44 * qdisc_lock(qdisc) spinlock.
46 * The idea is the following:
47 * - enqueue, dequeue are serialized via qdisc root lock
48 * - ingress filtering is also serialized via qdisc root lock
49 * - updates to tree and tree walking are only done under the rtnl mutex.
52 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
54 static inline struct sk_buff
*__skb_dequeue_bad_txq(struct Qdisc
*q
)
56 const struct netdev_queue
*txq
= q
->dev_queue
;
57 spinlock_t
*lock
= NULL
;
60 if (q
->flags
& TCQ_F_NOLOCK
) {
65 skb
= skb_peek(&q
->skb_bad_txq
);
67 /* check the reason of requeuing without tx lock first */
68 txq
= skb_get_tx_queue(txq
->dev
, skb
);
69 if (!netif_xmit_frozen_or_stopped(txq
)) {
70 skb
= __skb_dequeue(&q
->skb_bad_txq
);
71 if (qdisc_is_percpu_stats(q
)) {
72 qdisc_qstats_cpu_backlog_dec(q
, skb
);
73 qdisc_qstats_atomic_qlen_dec(q
);
75 qdisc_qstats_backlog_dec(q
, skb
);
89 static inline struct sk_buff
*qdisc_dequeue_skb_bad_txq(struct Qdisc
*q
)
91 struct sk_buff
*skb
= skb_peek(&q
->skb_bad_txq
);
94 skb
= __skb_dequeue_bad_txq(q
);
99 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc
*q
,
102 spinlock_t
*lock
= NULL
;
104 if (q
->flags
& TCQ_F_NOLOCK
) {
105 lock
= qdisc_lock(q
);
109 __skb_queue_tail(&q
->skb_bad_txq
, skb
);
111 if (qdisc_is_percpu_stats(q
)) {
112 qdisc_qstats_cpu_backlog_inc(q
, skb
);
113 qdisc_qstats_atomic_qlen_inc(q
);
115 qdisc_qstats_backlog_inc(q
, skb
);
123 static inline int __dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
126 struct sk_buff
*next
= skb
->next
;
128 __skb_queue_tail(&q
->gso_skb
, skb
);
129 q
->qstats
.requeues
++;
130 qdisc_qstats_backlog_inc(q
, skb
);
131 q
->q
.qlen
++; /* it's still part of the queue */
140 static inline int dev_requeue_skb_locked(struct sk_buff
*skb
, struct Qdisc
*q
)
142 spinlock_t
*lock
= qdisc_lock(q
);
146 struct sk_buff
*next
= skb
->next
;
148 __skb_queue_tail(&q
->gso_skb
, skb
);
150 qdisc_qstats_cpu_requeues_inc(q
);
151 qdisc_qstats_cpu_backlog_inc(q
, skb
);
152 qdisc_qstats_atomic_qlen_inc(q
);
163 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
165 if (q
->flags
& TCQ_F_NOLOCK
)
166 return dev_requeue_skb_locked(skb
, q
);
168 return __dev_requeue_skb(skb
, q
);
171 static void try_bulk_dequeue_skb(struct Qdisc
*q
,
173 const struct netdev_queue
*txq
,
176 int bytelimit
= qdisc_avail_bulklimit(txq
) - skb
->len
;
178 while (bytelimit
> 0) {
179 struct sk_buff
*nskb
= q
->dequeue(q
);
184 bytelimit
-= nskb
->len
; /* covers GSO len */
187 (*packets
)++; /* GSO counts as one pkt */
192 /* This variant of try_bulk_dequeue_skb() makes sure
193 * all skbs in the chain are for the same txq
195 static void try_bulk_dequeue_skb_slow(struct Qdisc
*q
,
199 int mapping
= skb_get_queue_mapping(skb
);
200 struct sk_buff
*nskb
;
204 nskb
= q
->dequeue(q
);
207 if (unlikely(skb_get_queue_mapping(nskb
) != mapping
)) {
208 qdisc_enqueue_skb_bad_txq(q
, nskb
);
218 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
219 * A requeued skb (via q->gso_skb) can also be a SKB list.
221 static struct sk_buff
*dequeue_skb(struct Qdisc
*q
, bool *validate
,
224 const struct netdev_queue
*txq
= q
->dev_queue
;
225 struct sk_buff
*skb
= NULL
;
228 if (unlikely(!skb_queue_empty(&q
->gso_skb
))) {
229 spinlock_t
*lock
= NULL
;
231 if (q
->flags
& TCQ_F_NOLOCK
) {
232 lock
= qdisc_lock(q
);
236 skb
= skb_peek(&q
->gso_skb
);
238 /* skb may be null if another cpu pulls gso_skb off in between
239 * empty check and lock.
247 /* skb in gso_skb were already validated */
249 if (xfrm_offload(skb
))
251 /* check the reason of requeuing without tx lock first */
252 txq
= skb_get_tx_queue(txq
->dev
, skb
);
253 if (!netif_xmit_frozen_or_stopped(txq
)) {
254 skb
= __skb_dequeue(&q
->gso_skb
);
255 if (qdisc_is_percpu_stats(q
)) {
256 qdisc_qstats_cpu_backlog_dec(q
, skb
);
257 qdisc_qstats_atomic_qlen_dec(q
);
259 qdisc_qstats_backlog_dec(q
, skb
);
272 if ((q
->flags
& TCQ_F_ONETXQUEUE
) &&
273 netif_xmit_frozen_or_stopped(txq
))
276 skb
= qdisc_dequeue_skb_bad_txq(q
);
278 if (skb
== SKB_XOFF_MAGIC
)
285 if (qdisc_may_bulk(q
))
286 try_bulk_dequeue_skb(q
, skb
, txq
, packets
);
288 try_bulk_dequeue_skb_slow(q
, skb
, packets
);
291 trace_qdisc_dequeue(q
, txq
, *packets
, skb
);
296 * Transmit possibly several skbs, and handle the return status as
297 * required. Owning running seqcount bit guarantees that
298 * only one CPU can execute this function.
300 * Returns to the caller:
301 * false - hardware queue frozen backoff
302 * true - feel free to send more pkts
304 bool sch_direct_xmit(struct sk_buff
*skb
, struct Qdisc
*q
,
305 struct net_device
*dev
, struct netdev_queue
*txq
,
306 spinlock_t
*root_lock
, bool validate
)
308 int ret
= NETDEV_TX_BUSY
;
311 /* And release qdisc */
313 spin_unlock(root_lock
);
315 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
317 skb
= validate_xmit_skb_list(skb
, dev
, &again
);
319 #ifdef CONFIG_XFRM_OFFLOAD
320 if (unlikely(again
)) {
322 spin_lock(root_lock
);
324 dev_requeue_skb(skb
, q
);
330 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
331 if (!netif_xmit_frozen_or_stopped(txq
))
332 skb
= dev_hard_start_xmit(skb
, dev
, txq
, &ret
);
334 HARD_TX_UNLOCK(dev
, txq
);
337 spin_lock(root_lock
);
342 spin_lock(root_lock
);
344 if (!dev_xmit_complete(ret
)) {
345 /* Driver returned NETDEV_TX_BUSY - requeue skb */
346 if (unlikely(ret
!= NETDEV_TX_BUSY
))
347 net_warn_ratelimited("BUG %s code %d qlen %d\n",
348 dev
->name
, ret
, q
->q
.qlen
);
350 dev_requeue_skb(skb
, q
);
358 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
360 * running seqcount guarantees only one CPU can process
361 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
364 * netif_tx_lock serializes accesses to device driver.
366 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
367 * if one is grabbed, another must be free.
369 * Note, that this procedure can be called by a watchdog timer
371 * Returns to the caller:
372 * 0 - queue is empty or throttled.
373 * >0 - queue is not empty.
376 static inline bool qdisc_restart(struct Qdisc
*q
, int *packets
)
378 spinlock_t
*root_lock
= NULL
;
379 struct netdev_queue
*txq
;
380 struct net_device
*dev
;
385 skb
= dequeue_skb(q
, &validate
, packets
);
389 if (!(q
->flags
& TCQ_F_NOLOCK
))
390 root_lock
= qdisc_lock(q
);
393 txq
= skb_get_tx_queue(dev
, skb
);
395 return sch_direct_xmit(skb
, q
, dev
, txq
, root_lock
, validate
);
398 void __qdisc_run(struct Qdisc
*q
)
400 int quota
= dev_tx_weight
;
403 while (qdisc_restart(q
, &packets
)) {
405 * Ordered by possible occurrence: Postpone processing if
406 * 1. we've exceeded packet quota
407 * 2. another process needs the CPU;
410 if (quota
<= 0 || need_resched()) {
417 unsigned long dev_trans_start(struct net_device
*dev
)
419 unsigned long val
, res
;
422 if (is_vlan_dev(dev
))
423 dev
= vlan_dev_real_dev(dev
);
424 else if (netif_is_macvlan(dev
))
425 dev
= macvlan_dev_real_dev(dev
);
426 res
= netdev_get_tx_queue(dev
, 0)->trans_start
;
427 for (i
= 1; i
< dev
->num_tx_queues
; i
++) {
428 val
= netdev_get_tx_queue(dev
, i
)->trans_start
;
429 if (val
&& time_after(val
, res
))
435 EXPORT_SYMBOL(dev_trans_start
);
437 static void dev_watchdog(struct timer_list
*t
)
439 struct net_device
*dev
= from_timer(dev
, t
, watchdog_timer
);
442 if (!qdisc_tx_is_noop(dev
)) {
443 if (netif_device_present(dev
) &&
444 netif_running(dev
) &&
445 netif_carrier_ok(dev
)) {
446 int some_queue_timedout
= 0;
448 unsigned long trans_start
;
450 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
451 struct netdev_queue
*txq
;
453 txq
= netdev_get_tx_queue(dev
, i
);
454 trans_start
= txq
->trans_start
;
455 if (netif_xmit_stopped(txq
) &&
456 time_after(jiffies
, (trans_start
+
457 dev
->watchdog_timeo
))) {
458 some_queue_timedout
= 1;
459 txq
->trans_timeout
++;
464 if (some_queue_timedout
) {
465 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
466 dev
->name
, netdev_drivername(dev
), i
);
467 dev
->netdev_ops
->ndo_tx_timeout(dev
);
469 if (!mod_timer(&dev
->watchdog_timer
,
470 round_jiffies(jiffies
+
471 dev
->watchdog_timeo
)))
475 netif_tx_unlock(dev
);
480 void __netdev_watchdog_up(struct net_device
*dev
)
482 if (dev
->netdev_ops
->ndo_tx_timeout
) {
483 if (dev
->watchdog_timeo
<= 0)
484 dev
->watchdog_timeo
= 5*HZ
;
485 if (!mod_timer(&dev
->watchdog_timer
,
486 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
490 EXPORT_SYMBOL_GPL(__netdev_watchdog_up
);
492 static void dev_watchdog_up(struct net_device
*dev
)
494 __netdev_watchdog_up(dev
);
497 static void dev_watchdog_down(struct net_device
*dev
)
499 netif_tx_lock_bh(dev
);
500 if (del_timer(&dev
->watchdog_timer
))
502 netif_tx_unlock_bh(dev
);
506 * netif_carrier_on - set carrier
507 * @dev: network device
509 * Device has detected that carrier.
511 void netif_carrier_on(struct net_device
*dev
)
513 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
514 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
516 atomic_inc(&dev
->carrier_up_count
);
517 linkwatch_fire_event(dev
);
518 if (netif_running(dev
))
519 __netdev_watchdog_up(dev
);
522 EXPORT_SYMBOL(netif_carrier_on
);
525 * netif_carrier_off - clear carrier
526 * @dev: network device
528 * Device has detected loss of carrier.
530 void netif_carrier_off(struct net_device
*dev
)
532 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
533 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
535 atomic_inc(&dev
->carrier_down_count
);
536 linkwatch_fire_event(dev
);
539 EXPORT_SYMBOL(netif_carrier_off
);
541 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
542 under all circumstances. It is difficult to invent anything faster or
546 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
547 struct sk_buff
**to_free
)
549 __qdisc_drop(skb
, to_free
);
553 static struct sk_buff
*noop_dequeue(struct Qdisc
*qdisc
)
558 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
561 .enqueue
= noop_enqueue
,
562 .dequeue
= noop_dequeue
,
563 .peek
= noop_dequeue
,
564 .owner
= THIS_MODULE
,
567 static struct netdev_queue noop_netdev_queue
= {
568 .qdisc
= &noop_qdisc
,
569 .qdisc_sleeping
= &noop_qdisc
,
572 struct Qdisc noop_qdisc
= {
573 .enqueue
= noop_enqueue
,
574 .dequeue
= noop_dequeue
,
575 .flags
= TCQ_F_BUILTIN
,
576 .ops
= &noop_qdisc_ops
,
577 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
578 .dev_queue
= &noop_netdev_queue
,
579 .running
= SEQCNT_ZERO(noop_qdisc
.running
),
580 .busylock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.busylock
),
582 .next
= (struct sk_buff
*)&noop_qdisc
.gso_skb
,
583 .prev
= (struct sk_buff
*)&noop_qdisc
.gso_skb
,
585 .lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.gso_skb
.lock
),
588 .next
= (struct sk_buff
*)&noop_qdisc
.skb_bad_txq
,
589 .prev
= (struct sk_buff
*)&noop_qdisc
.skb_bad_txq
,
591 .lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.skb_bad_txq
.lock
),
594 EXPORT_SYMBOL(noop_qdisc
);
596 static int noqueue_init(struct Qdisc
*qdisc
, struct nlattr
*opt
,
597 struct netlink_ext_ack
*extack
)
599 /* register_qdisc() assigns a default of noop_enqueue if unset,
600 * but __dev_queue_xmit() treats noqueue only as such
601 * if this is NULL - so clear it here. */
602 qdisc
->enqueue
= NULL
;
606 struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
609 .init
= noqueue_init
,
610 .enqueue
= noop_enqueue
,
611 .dequeue
= noop_dequeue
,
612 .peek
= noop_dequeue
,
613 .owner
= THIS_MODULE
,
616 static const u8 prio2band
[TC_PRIO_MAX
+ 1] = {
617 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
620 /* 3-band FIFO queue: old style, but should be a bit faster than
621 generic prio+fifo combination.
624 #define PFIFO_FAST_BANDS 3
627 * Private data for a pfifo_fast scheduler containing:
628 * - rings for priority bands
630 struct pfifo_fast_priv
{
631 struct skb_array q
[PFIFO_FAST_BANDS
];
634 static inline struct skb_array
*band2list(struct pfifo_fast_priv
*priv
,
637 return &priv
->q
[band
];
640 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
641 struct sk_buff
**to_free
)
643 int band
= prio2band
[skb
->priority
& TC_PRIO_MAX
];
644 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
645 struct skb_array
*q
= band2list(priv
, band
);
646 unsigned int pkt_len
= qdisc_pkt_len(skb
);
649 err
= skb_array_produce(q
, skb
);
652 return qdisc_drop_cpu(skb
, qdisc
, to_free
);
654 qdisc_qstats_atomic_qlen_inc(qdisc
);
655 /* Note: skb can not be used after skb_array_produce(),
656 * so we better not use qdisc_qstats_cpu_backlog_inc()
658 this_cpu_add(qdisc
->cpu_qstats
->backlog
, pkt_len
);
659 return NET_XMIT_SUCCESS
;
662 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
*qdisc
)
664 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
665 struct sk_buff
*skb
= NULL
;
668 for (band
= 0; band
< PFIFO_FAST_BANDS
&& !skb
; band
++) {
669 struct skb_array
*q
= band2list(priv
, band
);
671 if (__skb_array_empty(q
))
674 skb
= __skb_array_consume(q
);
677 qdisc_qstats_cpu_backlog_dec(qdisc
, skb
);
678 qdisc_bstats_cpu_update(qdisc
, skb
);
679 qdisc_qstats_atomic_qlen_dec(qdisc
);
685 static struct sk_buff
*pfifo_fast_peek(struct Qdisc
*qdisc
)
687 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
688 struct sk_buff
*skb
= NULL
;
691 for (band
= 0; band
< PFIFO_FAST_BANDS
&& !skb
; band
++) {
692 struct skb_array
*q
= band2list(priv
, band
);
694 skb
= __skb_array_peek(q
);
700 static void pfifo_fast_reset(struct Qdisc
*qdisc
)
703 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
705 for (band
= 0; band
< PFIFO_FAST_BANDS
; band
++) {
706 struct skb_array
*q
= band2list(priv
, band
);
709 /* NULL ring is possible if destroy path is due to a failed
710 * skb_array_init() in pfifo_fast_init() case.
715 while ((skb
= __skb_array_consume(q
)) != NULL
)
719 for_each_possible_cpu(i
) {
720 struct gnet_stats_queue
*q
= per_cpu_ptr(qdisc
->cpu_qstats
, i
);
726 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
728 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
730 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+ 1);
731 if (nla_put(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
))
732 goto nla_put_failure
;
739 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
,
740 struct netlink_ext_ack
*extack
)
742 unsigned int qlen
= qdisc_dev(qdisc
)->tx_queue_len
;
743 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
746 /* guard against zero length rings */
750 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
751 struct skb_array
*q
= band2list(priv
, prio
);
754 err
= skb_array_init(q
, qlen
, GFP_KERNEL
);
759 /* Can by-pass the queue discipline */
760 qdisc
->flags
|= TCQ_F_CAN_BYPASS
;
764 static void pfifo_fast_destroy(struct Qdisc
*sch
)
766 struct pfifo_fast_priv
*priv
= qdisc_priv(sch
);
769 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
770 struct skb_array
*q
= band2list(priv
, prio
);
772 /* NULL ring is possible if destroy path is due to a failed
773 * skb_array_init() in pfifo_fast_init() case.
777 /* Destroy ring but no need to kfree_skb because a call to
778 * pfifo_fast_reset() has already done that work.
780 ptr_ring_cleanup(&q
->ring
, NULL
);
784 static int pfifo_fast_change_tx_queue_len(struct Qdisc
*sch
,
785 unsigned int new_len
)
787 struct pfifo_fast_priv
*priv
= qdisc_priv(sch
);
788 struct skb_array
*bands
[PFIFO_FAST_BANDS
];
791 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
792 struct skb_array
*q
= band2list(priv
, prio
);
797 return skb_array_resize_multiple(bands
, PFIFO_FAST_BANDS
, new_len
,
801 struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
803 .priv_size
= sizeof(struct pfifo_fast_priv
),
804 .enqueue
= pfifo_fast_enqueue
,
805 .dequeue
= pfifo_fast_dequeue
,
806 .peek
= pfifo_fast_peek
,
807 .init
= pfifo_fast_init
,
808 .destroy
= pfifo_fast_destroy
,
809 .reset
= pfifo_fast_reset
,
810 .dump
= pfifo_fast_dump
,
811 .change_tx_queue_len
= pfifo_fast_change_tx_queue_len
,
812 .owner
= THIS_MODULE
,
813 .static_flags
= TCQ_F_NOLOCK
| TCQ_F_CPUSTATS
,
815 EXPORT_SYMBOL(pfifo_fast_ops
);
817 static struct lock_class_key qdisc_tx_busylock
;
818 static struct lock_class_key qdisc_running_key
;
820 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
821 const struct Qdisc_ops
*ops
,
822 struct netlink_ext_ack
*extack
)
826 unsigned int size
= QDISC_ALIGN(sizeof(*sch
)) + ops
->priv_size
;
828 struct net_device
*dev
;
831 NL_SET_ERR_MSG(extack
, "No device queue given");
836 dev
= dev_queue
->dev
;
837 p
= kzalloc_node(size
, GFP_KERNEL
,
838 netdev_queue_numa_node_read(dev_queue
));
842 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
843 /* if we got non aligned memory, ask more and do alignment ourself */
846 p
= kzalloc_node(size
+ QDISC_ALIGNTO
- 1, GFP_KERNEL
,
847 netdev_queue_numa_node_read(dev_queue
));
850 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
851 sch
->padded
= (char *) sch
- (char *) p
;
853 __skb_queue_head_init(&sch
->gso_skb
);
854 __skb_queue_head_init(&sch
->skb_bad_txq
);
855 qdisc_skb_head_init(&sch
->q
);
856 spin_lock_init(&sch
->q
.lock
);
858 if (ops
->static_flags
& TCQ_F_CPUSTATS
) {
860 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu
);
861 if (!sch
->cpu_bstats
)
864 sch
->cpu_qstats
= alloc_percpu(struct gnet_stats_queue
);
865 if (!sch
->cpu_qstats
) {
866 free_percpu(sch
->cpu_bstats
);
871 spin_lock_init(&sch
->busylock
);
872 lockdep_set_class(&sch
->busylock
,
873 dev
->qdisc_tx_busylock
?: &qdisc_tx_busylock
);
875 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
876 spin_lock_init(&sch
->seqlock
);
877 lockdep_set_class(&sch
->busylock
,
878 dev
->qdisc_tx_busylock
?: &qdisc_tx_busylock
);
880 seqcount_init(&sch
->running
);
881 lockdep_set_class(&sch
->running
,
882 dev
->qdisc_running_key
?: &qdisc_running_key
);
885 sch
->flags
= ops
->static_flags
;
886 sch
->enqueue
= ops
->enqueue
;
887 sch
->dequeue
= ops
->dequeue
;
888 sch
->dev_queue
= dev_queue
;
890 refcount_set(&sch
->refcnt
, 1);
899 struct Qdisc
*qdisc_create_dflt(struct netdev_queue
*dev_queue
,
900 const struct Qdisc_ops
*ops
,
901 unsigned int parentid
,
902 struct netlink_ext_ack
*extack
)
906 if (!try_module_get(ops
->owner
)) {
907 NL_SET_ERR_MSG(extack
, "Failed to increase module reference counter");
911 sch
= qdisc_alloc(dev_queue
, ops
, extack
);
913 module_put(ops
->owner
);
916 sch
->parent
= parentid
;
918 if (!ops
->init
|| ops
->init(sch
, NULL
, extack
) == 0)
924 EXPORT_SYMBOL(qdisc_create_dflt
);
926 /* Under qdisc_lock(qdisc) and BH! */
928 void qdisc_reset(struct Qdisc
*qdisc
)
930 const struct Qdisc_ops
*ops
= qdisc
->ops
;
931 struct sk_buff
*skb
, *tmp
;
936 skb_queue_walk_safe(&qdisc
->gso_skb
, skb
, tmp
) {
937 __skb_unlink(skb
, &qdisc
->gso_skb
);
941 skb_queue_walk_safe(&qdisc
->skb_bad_txq
, skb
, tmp
) {
942 __skb_unlink(skb
, &qdisc
->skb_bad_txq
);
947 qdisc
->qstats
.backlog
= 0;
949 EXPORT_SYMBOL(qdisc_reset
);
951 void qdisc_free(struct Qdisc
*qdisc
)
953 if (qdisc_is_percpu_stats(qdisc
)) {
954 free_percpu(qdisc
->cpu_bstats
);
955 free_percpu(qdisc
->cpu_qstats
);
958 kfree((char *) qdisc
- qdisc
->padded
);
961 void qdisc_destroy(struct Qdisc
*qdisc
)
963 const struct Qdisc_ops
*ops
;
964 struct sk_buff
*skb
, *tmp
;
970 if (qdisc
->flags
& TCQ_F_BUILTIN
||
971 !refcount_dec_and_test(&qdisc
->refcnt
))
974 #ifdef CONFIG_NET_SCHED
975 qdisc_hash_del(qdisc
);
977 qdisc_put_stab(rtnl_dereference(qdisc
->stab
));
979 gen_kill_estimator(&qdisc
->rate_est
);
985 module_put(ops
->owner
);
986 dev_put(qdisc_dev(qdisc
));
988 skb_queue_walk_safe(&qdisc
->gso_skb
, skb
, tmp
) {
989 __skb_unlink(skb
, &qdisc
->gso_skb
);
993 skb_queue_walk_safe(&qdisc
->skb_bad_txq
, skb
, tmp
) {
994 __skb_unlink(skb
, &qdisc
->skb_bad_txq
);
1000 EXPORT_SYMBOL(qdisc_destroy
);
1002 /* Attach toplevel qdisc to device queue. */
1003 struct Qdisc
*dev_graft_qdisc(struct netdev_queue
*dev_queue
,
1004 struct Qdisc
*qdisc
)
1006 struct Qdisc
*oqdisc
= dev_queue
->qdisc_sleeping
;
1007 spinlock_t
*root_lock
;
1009 root_lock
= qdisc_lock(oqdisc
);
1010 spin_lock_bh(root_lock
);
1012 /* ... and graft new one */
1014 qdisc
= &noop_qdisc
;
1015 dev_queue
->qdisc_sleeping
= qdisc
;
1016 rcu_assign_pointer(dev_queue
->qdisc
, &noop_qdisc
);
1018 spin_unlock_bh(root_lock
);
1022 EXPORT_SYMBOL(dev_graft_qdisc
);
1024 static void attach_one_default_qdisc(struct net_device
*dev
,
1025 struct netdev_queue
*dev_queue
,
1028 struct Qdisc
*qdisc
;
1029 const struct Qdisc_ops
*ops
= default_qdisc_ops
;
1031 if (dev
->priv_flags
& IFF_NO_QUEUE
)
1032 ops
= &noqueue_qdisc_ops
;
1034 qdisc
= qdisc_create_dflt(dev_queue
, ops
, TC_H_ROOT
, NULL
);
1036 netdev_info(dev
, "activation failed\n");
1039 if (!netif_is_multiqueue(dev
))
1040 qdisc
->flags
|= TCQ_F_ONETXQUEUE
| TCQ_F_NOPARENT
;
1041 dev_queue
->qdisc_sleeping
= qdisc
;
1044 static void attach_default_qdiscs(struct net_device
*dev
)
1046 struct netdev_queue
*txq
;
1047 struct Qdisc
*qdisc
;
1049 txq
= netdev_get_tx_queue(dev
, 0);
1051 if (!netif_is_multiqueue(dev
) ||
1052 dev
->priv_flags
& IFF_NO_QUEUE
) {
1053 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
1054 dev
->qdisc
= txq
->qdisc_sleeping
;
1055 qdisc_refcount_inc(dev
->qdisc
);
1057 qdisc
= qdisc_create_dflt(txq
, &mq_qdisc_ops
, TC_H_ROOT
, NULL
);
1060 qdisc
->ops
->attach(qdisc
);
1063 #ifdef CONFIG_NET_SCHED
1064 if (dev
->qdisc
!= &noop_qdisc
)
1065 qdisc_hash_add(dev
->qdisc
, false);
1069 static void transition_one_qdisc(struct net_device
*dev
,
1070 struct netdev_queue
*dev_queue
,
1071 void *_need_watchdog
)
1073 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
1074 int *need_watchdog_p
= _need_watchdog
;
1076 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
1077 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
1079 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
1080 if (need_watchdog_p
) {
1081 dev_queue
->trans_start
= 0;
1082 *need_watchdog_p
= 1;
1086 void dev_activate(struct net_device
*dev
)
1090 /* No queueing discipline is attached to device;
1091 * create default one for devices, which need queueing
1092 * and noqueue_qdisc for virtual interfaces
1095 if (dev
->qdisc
== &noop_qdisc
)
1096 attach_default_qdiscs(dev
);
1098 if (!netif_carrier_ok(dev
))
1099 /* Delay activation until next carrier-on event */
1103 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
1104 if (dev_ingress_queue(dev
))
1105 transition_one_qdisc(dev
, dev_ingress_queue(dev
), NULL
);
1107 if (need_watchdog
) {
1108 netif_trans_update(dev
);
1109 dev_watchdog_up(dev
);
1112 EXPORT_SYMBOL(dev_activate
);
1114 static void dev_deactivate_queue(struct net_device
*dev
,
1115 struct netdev_queue
*dev_queue
,
1116 void *_qdisc_default
)
1118 struct Qdisc
*qdisc
= rtnl_dereference(dev_queue
->qdisc
);
1119 struct Qdisc
*qdisc_default
= _qdisc_default
;
1122 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
1123 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
1125 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
1129 static void dev_reset_queue(struct net_device
*dev
,
1130 struct netdev_queue
*dev_queue
,
1133 struct Qdisc
*qdisc
;
1136 qdisc
= dev_queue
->qdisc_sleeping
;
1140 nolock
= qdisc
->flags
& TCQ_F_NOLOCK
;
1143 spin_lock_bh(&qdisc
->seqlock
);
1144 spin_lock_bh(qdisc_lock(qdisc
));
1148 spin_unlock_bh(qdisc_lock(qdisc
));
1150 spin_unlock_bh(&qdisc
->seqlock
);
1153 static bool some_qdisc_is_busy(struct net_device
*dev
)
1157 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
1158 struct netdev_queue
*dev_queue
;
1159 spinlock_t
*root_lock
;
1163 dev_queue
= netdev_get_tx_queue(dev
, i
);
1164 q
= dev_queue
->qdisc_sleeping
;
1166 root_lock
= qdisc_lock(q
);
1167 spin_lock_bh(root_lock
);
1169 val
= (qdisc_is_running(q
) ||
1170 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
1172 spin_unlock_bh(root_lock
);
1180 static void dev_qdisc_reset(struct net_device
*dev
,
1181 struct netdev_queue
*dev_queue
,
1184 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1191 * dev_deactivate_many - deactivate transmissions on several devices
1192 * @head: list of devices to deactivate
1194 * This function returns only when all outstanding transmissions
1195 * have completed, unless all devices are in dismantle phase.
1197 void dev_deactivate_many(struct list_head
*head
)
1199 struct net_device
*dev
;
1201 list_for_each_entry(dev
, head
, close_list
) {
1202 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
,
1204 if (dev_ingress_queue(dev
))
1205 dev_deactivate_queue(dev
, dev_ingress_queue(dev
),
1208 dev_watchdog_down(dev
);
1211 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1212 * outstanding qdisc enqueuing calls.
1213 * This is avoided if all devices are in dismantle phase :
1214 * Caller will call synchronize_net() for us
1218 list_for_each_entry(dev
, head
, close_list
) {
1219 netdev_for_each_tx_queue(dev
, dev_reset_queue
, NULL
);
1221 if (dev_ingress_queue(dev
))
1222 dev_reset_queue(dev
, dev_ingress_queue(dev
), NULL
);
1225 /* Wait for outstanding qdisc_run calls. */
1226 list_for_each_entry(dev
, head
, close_list
) {
1227 while (some_qdisc_is_busy(dev
))
1229 /* The new qdisc is assigned at this point so we can safely
1230 * unwind stale skb lists and qdisc statistics
1232 netdev_for_each_tx_queue(dev
, dev_qdisc_reset
, NULL
);
1233 if (dev_ingress_queue(dev
))
1234 dev_qdisc_reset(dev
, dev_ingress_queue(dev
), NULL
);
1238 void dev_deactivate(struct net_device
*dev
)
1242 list_add(&dev
->close_list
, &single
);
1243 dev_deactivate_many(&single
);
1246 EXPORT_SYMBOL(dev_deactivate
);
1248 static int qdisc_change_tx_queue_len(struct net_device
*dev
,
1249 struct netdev_queue
*dev_queue
)
1251 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1252 const struct Qdisc_ops
*ops
= qdisc
->ops
;
1254 if (ops
->change_tx_queue_len
)
1255 return ops
->change_tx_queue_len(qdisc
, dev
->tx_queue_len
);
1259 int dev_qdisc_change_tx_queue_len(struct net_device
*dev
)
1261 bool up
= dev
->flags
& IFF_UP
;
1266 dev_deactivate(dev
);
1268 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
1269 ret
= qdisc_change_tx_queue_len(dev
, &dev
->_tx
[i
]);
1271 /* TODO: revert changes on a partial failure */
1281 static void dev_init_scheduler_queue(struct net_device
*dev
,
1282 struct netdev_queue
*dev_queue
,
1285 struct Qdisc
*qdisc
= _qdisc
;
1287 rcu_assign_pointer(dev_queue
->qdisc
, qdisc
);
1288 dev_queue
->qdisc_sleeping
= qdisc
;
1291 void dev_init_scheduler(struct net_device
*dev
)
1293 dev
->qdisc
= &noop_qdisc
;
1294 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
1295 if (dev_ingress_queue(dev
))
1296 dev_init_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
1298 timer_setup(&dev
->watchdog_timer
, dev_watchdog
, 0);
1301 static void shutdown_scheduler_queue(struct net_device
*dev
,
1302 struct netdev_queue
*dev_queue
,
1303 void *_qdisc_default
)
1305 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1306 struct Qdisc
*qdisc_default
= _qdisc_default
;
1309 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
1310 dev_queue
->qdisc_sleeping
= qdisc_default
;
1312 qdisc_destroy(qdisc
);
1316 void dev_shutdown(struct net_device
*dev
)
1318 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
1319 if (dev_ingress_queue(dev
))
1320 shutdown_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
1321 qdisc_destroy(dev
->qdisc
);
1322 dev
->qdisc
= &noop_qdisc
;
1324 WARN_ON(timer_pending(&dev
->watchdog_timer
));
1327 void psched_ratecfg_precompute(struct psched_ratecfg
*r
,
1328 const struct tc_ratespec
*conf
,
1331 memset(r
, 0, sizeof(*r
));
1332 r
->overhead
= conf
->overhead
;
1333 r
->rate_bytes_ps
= max_t(u64
, conf
->rate
, rate64
);
1334 r
->linklayer
= (conf
->linklayer
& TC_LINKLAYER_MASK
);
1337 * The deal here is to replace a divide by a reciprocal one
1338 * in fast path (a reciprocal divide is a multiply and a shift)
1340 * Normal formula would be :
1341 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1343 * We compute mult/shift to use instead :
1344 * time_in_ns = (len * mult) >> shift;
1346 * We try to get the highest possible mult value for accuracy,
1347 * but have to make sure no overflows will ever happen.
1349 if (r
->rate_bytes_ps
> 0) {
1350 u64 factor
= NSEC_PER_SEC
;
1353 r
->mult
= div64_u64(factor
, r
->rate_bytes_ps
);
1354 if (r
->mult
& (1U << 31) || factor
& (1ULL << 63))
1361 EXPORT_SYMBOL(psched_ratecfg_precompute
);
1363 static void mini_qdisc_rcu_func(struct rcu_head
*head
)
1367 void mini_qdisc_pair_swap(struct mini_Qdisc_pair
*miniqp
,
1368 struct tcf_proto
*tp_head
)
1370 struct mini_Qdisc
*miniq_old
= rtnl_dereference(*miniqp
->p_miniq
);
1371 struct mini_Qdisc
*miniq
;
1374 RCU_INIT_POINTER(*miniqp
->p_miniq
, NULL
);
1375 /* Wait for flying RCU callback before it is freed. */
1380 miniq
= !miniq_old
|| miniq_old
== &miniqp
->miniq2
?
1381 &miniqp
->miniq1
: &miniqp
->miniq2
;
1383 /* We need to make sure that readers won't see the miniq
1384 * we are about to modify. So wait until previous call_rcu_bh callback
1388 miniq
->filter_list
= tp_head
;
1389 rcu_assign_pointer(*miniqp
->p_miniq
, miniq
);
1392 /* This is counterpart of the rcu barriers above. We need to
1393 * block potential new user of miniq_old until all readers
1394 * are not seeing it.
1396 call_rcu_bh(&miniq_old
->rcu
, mini_qdisc_rcu_func
);
1398 EXPORT_SYMBOL(mini_qdisc_pair_swap
);
1400 void mini_qdisc_pair_init(struct mini_Qdisc_pair
*miniqp
, struct Qdisc
*qdisc
,
1401 struct mini_Qdisc __rcu
**p_miniq
)
1403 miniqp
->miniq1
.cpu_bstats
= qdisc
->cpu_bstats
;
1404 miniqp
->miniq1
.cpu_qstats
= qdisc
->cpu_qstats
;
1405 miniqp
->miniq2
.cpu_bstats
= qdisc
->cpu_bstats
;
1406 miniqp
->miniq2
.cpu_qstats
= qdisc
->cpu_qstats
;
1407 miniqp
->p_miniq
= p_miniq
;
1409 EXPORT_SYMBOL(mini_qdisc_pair_init
);