1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/sched/sch_generic.c Generic packet scheduler routines.
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
10 #include <linux/bitops.h>
11 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/netdevice.h>
18 #include <linux/skbuff.h>
19 #include <linux/rtnetlink.h>
20 #include <linux/init.h>
21 #include <linux/rcupdate.h>
22 #include <linux/list.h>
23 #include <linux/slab.h>
24 #include <linux/if_vlan.h>
25 #include <linux/skb_array.h>
26 #include <linux/if_macvlan.h>
27 #include <net/sch_generic.h>
28 #include <net/pkt_sched.h>
30 #include <trace/events/qdisc.h>
31 #include <trace/events/net.h>
34 /* Qdisc to use by default */
35 const struct Qdisc_ops
*default_qdisc_ops
= &pfifo_fast_ops
;
36 EXPORT_SYMBOL(default_qdisc_ops
);
38 /* Main transmission queue. */
40 /* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
49 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
51 static inline struct sk_buff
*__skb_dequeue_bad_txq(struct Qdisc
*q
)
53 const struct netdev_queue
*txq
= q
->dev_queue
;
54 spinlock_t
*lock
= NULL
;
57 if (q
->flags
& TCQ_F_NOLOCK
) {
62 skb
= skb_peek(&q
->skb_bad_txq
);
64 /* check the reason of requeuing without tx lock first */
65 txq
= skb_get_tx_queue(txq
->dev
, skb
);
66 if (!netif_xmit_frozen_or_stopped(txq
)) {
67 skb
= __skb_dequeue(&q
->skb_bad_txq
);
68 if (qdisc_is_percpu_stats(q
)) {
69 qdisc_qstats_cpu_backlog_dec(q
, skb
);
70 qdisc_qstats_cpu_qlen_dec(q
);
72 qdisc_qstats_backlog_dec(q
, skb
);
86 static inline struct sk_buff
*qdisc_dequeue_skb_bad_txq(struct Qdisc
*q
)
88 struct sk_buff
*skb
= skb_peek(&q
->skb_bad_txq
);
91 skb
= __skb_dequeue_bad_txq(q
);
96 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc
*q
,
99 spinlock_t
*lock
= NULL
;
101 if (q
->flags
& TCQ_F_NOLOCK
) {
102 lock
= qdisc_lock(q
);
106 __skb_queue_tail(&q
->skb_bad_txq
, skb
);
108 if (qdisc_is_percpu_stats(q
)) {
109 qdisc_qstats_cpu_backlog_inc(q
, skb
);
110 qdisc_qstats_cpu_qlen_inc(q
);
112 qdisc_qstats_backlog_inc(q
, skb
);
120 static inline void dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
122 spinlock_t
*lock
= NULL
;
124 if (q
->flags
& TCQ_F_NOLOCK
) {
125 lock
= qdisc_lock(q
);
130 struct sk_buff
*next
= skb
->next
;
132 __skb_queue_tail(&q
->gso_skb
, skb
);
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q
)) {
136 qdisc_qstats_cpu_requeues_inc(q
);
137 qdisc_qstats_cpu_backlog_inc(q
, skb
);
138 qdisc_qstats_cpu_qlen_inc(q
);
140 q
->qstats
.requeues
++;
141 qdisc_qstats_backlog_inc(q
, skb
);
152 static void try_bulk_dequeue_skb(struct Qdisc
*q
,
154 const struct netdev_queue
*txq
,
157 int bytelimit
= qdisc_avail_bulklimit(txq
) - skb
->len
;
159 while (bytelimit
> 0) {
160 struct sk_buff
*nskb
= q
->dequeue(q
);
165 bytelimit
-= nskb
->len
; /* covers GSO len */
168 (*packets
)++; /* GSO counts as one pkt */
170 skb_mark_not_on_list(skb
);
173 /* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
176 static void try_bulk_dequeue_skb_slow(struct Qdisc
*q
,
180 int mapping
= skb_get_queue_mapping(skb
);
181 struct sk_buff
*nskb
;
185 nskb
= q
->dequeue(q
);
188 if (unlikely(skb_get_queue_mapping(nskb
) != mapping
)) {
189 qdisc_enqueue_skb_bad_txq(q
, nskb
);
196 skb_mark_not_on_list(skb
);
199 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
202 static struct sk_buff
*dequeue_skb(struct Qdisc
*q
, bool *validate
,
205 const struct netdev_queue
*txq
= q
->dev_queue
;
206 struct sk_buff
*skb
= NULL
;
209 if (unlikely(!skb_queue_empty(&q
->gso_skb
))) {
210 spinlock_t
*lock
= NULL
;
212 if (q
->flags
& TCQ_F_NOLOCK
) {
213 lock
= qdisc_lock(q
);
217 skb
= skb_peek(&q
->gso_skb
);
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
228 /* skb in gso_skb were already validated */
230 if (xfrm_offload(skb
))
232 /* check the reason of requeuing without tx lock first */
233 txq
= skb_get_tx_queue(txq
->dev
, skb
);
234 if (!netif_xmit_frozen_or_stopped(txq
)) {
235 skb
= __skb_dequeue(&q
->gso_skb
);
236 if (qdisc_is_percpu_stats(q
)) {
237 qdisc_qstats_cpu_backlog_dec(q
, skb
);
238 qdisc_qstats_cpu_qlen_dec(q
);
240 qdisc_qstats_backlog_dec(q
, skb
);
253 if ((q
->flags
& TCQ_F_ONETXQUEUE
) &&
254 netif_xmit_frozen_or_stopped(txq
))
257 skb
= qdisc_dequeue_skb_bad_txq(q
);
259 if (skb
== SKB_XOFF_MAGIC
)
266 if (qdisc_may_bulk(q
))
267 try_bulk_dequeue_skb(q
, skb
, txq
, packets
);
269 try_bulk_dequeue_skb_slow(q
, skb
, packets
);
272 trace_qdisc_dequeue(q
, txq
, *packets
, skb
);
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
285 bool sch_direct_xmit(struct sk_buff
*skb
, struct Qdisc
*q
,
286 struct net_device
*dev
, struct netdev_queue
*txq
,
287 spinlock_t
*root_lock
, bool validate
)
289 int ret
= NETDEV_TX_BUSY
;
292 /* And release qdisc */
294 spin_unlock(root_lock
);
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
298 skb
= validate_xmit_skb_list(skb
, dev
, &again
);
300 #ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again
)) {
303 spin_lock(root_lock
);
305 dev_requeue_skb(skb
, q
);
311 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq
))
313 skb
= dev_hard_start_xmit(skb
, dev
, txq
, &ret
);
315 HARD_TX_UNLOCK(dev
, txq
);
318 spin_lock(root_lock
);
323 spin_lock(root_lock
);
325 if (!dev_xmit_complete(ret
)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret
!= NETDEV_TX_BUSY
))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev
->name
, ret
, q
->q
.qlen
);
331 dev_requeue_skb(skb
, q
);
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
345 * netif_tx_lock serializes accesses to device driver.
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
350 * Note, that this procedure can be called by a watchdog timer
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
357 static inline bool qdisc_restart(struct Qdisc
*q
, int *packets
)
359 spinlock_t
*root_lock
= NULL
;
360 struct netdev_queue
*txq
;
361 struct net_device
*dev
;
366 skb
= dequeue_skb(q
, &validate
, packets
);
370 if (!(q
->flags
& TCQ_F_NOLOCK
))
371 root_lock
= qdisc_lock(q
);
374 txq
= skb_get_tx_queue(dev
, skb
);
376 return sch_direct_xmit(skb
, q
, dev
, txq
, root_lock
, validate
);
379 void __qdisc_run(struct Qdisc
*q
)
381 int quota
= dev_tx_weight
;
384 while (qdisc_restart(q
, &packets
)) {
393 unsigned long dev_trans_start(struct net_device
*dev
)
395 unsigned long val
, res
;
398 if (is_vlan_dev(dev
))
399 dev
= vlan_dev_real_dev(dev
);
400 else if (netif_is_macvlan(dev
))
401 dev
= macvlan_dev_real_dev(dev
);
402 res
= netdev_get_tx_queue(dev
, 0)->trans_start
;
403 for (i
= 1; i
< dev
->num_tx_queues
; i
++) {
404 val
= netdev_get_tx_queue(dev
, i
)->trans_start
;
405 if (val
&& time_after(val
, res
))
411 EXPORT_SYMBOL(dev_trans_start
);
413 static void dev_watchdog(struct timer_list
*t
)
415 struct net_device
*dev
= from_timer(dev
, t
, watchdog_timer
);
418 if (!qdisc_tx_is_noop(dev
)) {
419 if (netif_device_present(dev
) &&
420 netif_running(dev
) &&
421 netif_carrier_ok(dev
)) {
422 int some_queue_timedout
= 0;
424 unsigned long trans_start
;
426 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
427 struct netdev_queue
*txq
;
429 txq
= netdev_get_tx_queue(dev
, i
);
430 trans_start
= txq
->trans_start
;
431 if (netif_xmit_stopped(txq
) &&
432 time_after(jiffies
, (trans_start
+
433 dev
->watchdog_timeo
))) {
434 some_queue_timedout
= 1;
435 txq
->trans_timeout
++;
440 if (some_queue_timedout
) {
441 trace_net_dev_xmit_timeout(dev
, i
);
442 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
443 dev
->name
, netdev_drivername(dev
), i
);
444 dev
->netdev_ops
->ndo_tx_timeout(dev
, i
);
446 if (!mod_timer(&dev
->watchdog_timer
,
447 round_jiffies(jiffies
+
448 dev
->watchdog_timeo
)))
452 netif_tx_unlock(dev
);
457 void __netdev_watchdog_up(struct net_device
*dev
)
459 if (dev
->netdev_ops
->ndo_tx_timeout
) {
460 if (dev
->watchdog_timeo
<= 0)
461 dev
->watchdog_timeo
= 5*HZ
;
462 if (!mod_timer(&dev
->watchdog_timer
,
463 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
467 EXPORT_SYMBOL_GPL(__netdev_watchdog_up
);
469 static void dev_watchdog_up(struct net_device
*dev
)
471 __netdev_watchdog_up(dev
);
474 static void dev_watchdog_down(struct net_device
*dev
)
476 netif_tx_lock_bh(dev
);
477 if (del_timer(&dev
->watchdog_timer
))
479 netif_tx_unlock_bh(dev
);
483 * netif_carrier_on - set carrier
484 * @dev: network device
486 * Device has detected acquisition of carrier.
488 void netif_carrier_on(struct net_device
*dev
)
490 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
491 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
493 atomic_inc(&dev
->carrier_up_count
);
494 linkwatch_fire_event(dev
);
495 if (netif_running(dev
))
496 __netdev_watchdog_up(dev
);
499 EXPORT_SYMBOL(netif_carrier_on
);
502 * netif_carrier_off - clear carrier
503 * @dev: network device
505 * Device has detected loss of carrier.
507 void netif_carrier_off(struct net_device
*dev
)
509 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
510 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
512 atomic_inc(&dev
->carrier_down_count
);
513 linkwatch_fire_event(dev
);
516 EXPORT_SYMBOL(netif_carrier_off
);
518 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
519 under all circumstances. It is difficult to invent anything faster or
523 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
524 struct sk_buff
**to_free
)
526 __qdisc_drop(skb
, to_free
);
530 static struct sk_buff
*noop_dequeue(struct Qdisc
*qdisc
)
535 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
538 .enqueue
= noop_enqueue
,
539 .dequeue
= noop_dequeue
,
540 .peek
= noop_dequeue
,
541 .owner
= THIS_MODULE
,
544 static struct netdev_queue noop_netdev_queue
= {
545 RCU_POINTER_INITIALIZER(qdisc
, &noop_qdisc
),
546 .qdisc_sleeping
= &noop_qdisc
,
549 struct Qdisc noop_qdisc
= {
550 .enqueue
= noop_enqueue
,
551 .dequeue
= noop_dequeue
,
552 .flags
= TCQ_F_BUILTIN
,
553 .ops
= &noop_qdisc_ops
,
554 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
555 .dev_queue
= &noop_netdev_queue
,
556 .running
= SEQCNT_ZERO(noop_qdisc
.running
),
557 .busylock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.busylock
),
559 .next
= (struct sk_buff
*)&noop_qdisc
.gso_skb
,
560 .prev
= (struct sk_buff
*)&noop_qdisc
.gso_skb
,
562 .lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.gso_skb
.lock
),
565 .next
= (struct sk_buff
*)&noop_qdisc
.skb_bad_txq
,
566 .prev
= (struct sk_buff
*)&noop_qdisc
.skb_bad_txq
,
568 .lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.skb_bad_txq
.lock
),
571 EXPORT_SYMBOL(noop_qdisc
);
573 static int noqueue_init(struct Qdisc
*qdisc
, struct nlattr
*opt
,
574 struct netlink_ext_ack
*extack
)
576 /* register_qdisc() assigns a default of noop_enqueue if unset,
577 * but __dev_queue_xmit() treats noqueue only as such
578 * if this is NULL - so clear it here. */
579 qdisc
->enqueue
= NULL
;
583 struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
586 .init
= noqueue_init
,
587 .enqueue
= noop_enqueue
,
588 .dequeue
= noop_dequeue
,
589 .peek
= noop_dequeue
,
590 .owner
= THIS_MODULE
,
593 static const u8 prio2band
[TC_PRIO_MAX
+ 1] = {
594 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
597 /* 3-band FIFO queue: old style, but should be a bit faster than
598 generic prio+fifo combination.
601 #define PFIFO_FAST_BANDS 3
604 * Private data for a pfifo_fast scheduler containing:
605 * - rings for priority bands
607 struct pfifo_fast_priv
{
608 struct skb_array q
[PFIFO_FAST_BANDS
];
611 static inline struct skb_array
*band2list(struct pfifo_fast_priv
*priv
,
614 return &priv
->q
[band
];
617 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
*qdisc
,
618 struct sk_buff
**to_free
)
620 int band
= prio2band
[skb
->priority
& TC_PRIO_MAX
];
621 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
622 struct skb_array
*q
= band2list(priv
, band
);
623 unsigned int pkt_len
= qdisc_pkt_len(skb
);
626 err
= skb_array_produce(q
, skb
);
629 if (qdisc_is_percpu_stats(qdisc
))
630 return qdisc_drop_cpu(skb
, qdisc
, to_free
);
632 return qdisc_drop(skb
, qdisc
, to_free
);
635 qdisc_update_stats_at_enqueue(qdisc
, pkt_len
);
636 return NET_XMIT_SUCCESS
;
639 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
*qdisc
)
641 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
642 struct sk_buff
*skb
= NULL
;
645 for (band
= 0; band
< PFIFO_FAST_BANDS
&& !skb
; band
++) {
646 struct skb_array
*q
= band2list(priv
, band
);
648 if (__skb_array_empty(q
))
651 skb
= __skb_array_consume(q
);
654 qdisc_update_stats_at_dequeue(qdisc
, skb
);
656 WRITE_ONCE(qdisc
->empty
, true);
662 static struct sk_buff
*pfifo_fast_peek(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 skb
= __skb_array_peek(q
);
677 static void pfifo_fast_reset(struct Qdisc
*qdisc
)
680 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
682 for (band
= 0; band
< PFIFO_FAST_BANDS
; band
++) {
683 struct skb_array
*q
= band2list(priv
, band
);
686 /* NULL ring is possible if destroy path is due to a failed
687 * skb_array_init() in pfifo_fast_init() case.
692 while ((skb
= __skb_array_consume(q
)) != NULL
)
696 if (qdisc_is_percpu_stats(qdisc
)) {
697 for_each_possible_cpu(i
) {
698 struct gnet_stats_queue
*q
;
700 q
= per_cpu_ptr(qdisc
->cpu_qstats
, i
);
707 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
709 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
711 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+ 1);
712 if (nla_put(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
))
713 goto nla_put_failure
;
720 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
,
721 struct netlink_ext_ack
*extack
)
723 unsigned int qlen
= qdisc_dev(qdisc
)->tx_queue_len
;
724 struct pfifo_fast_priv
*priv
= qdisc_priv(qdisc
);
727 /* guard against zero length rings */
731 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
732 struct skb_array
*q
= band2list(priv
, prio
);
735 err
= skb_array_init(q
, qlen
, GFP_KERNEL
);
740 /* Can by-pass the queue discipline */
741 qdisc
->flags
|= TCQ_F_CAN_BYPASS
;
745 static void pfifo_fast_destroy(struct Qdisc
*sch
)
747 struct pfifo_fast_priv
*priv
= qdisc_priv(sch
);
750 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
751 struct skb_array
*q
= band2list(priv
, prio
);
753 /* NULL ring is possible if destroy path is due to a failed
754 * skb_array_init() in pfifo_fast_init() case.
758 /* Destroy ring but no need to kfree_skb because a call to
759 * pfifo_fast_reset() has already done that work.
761 ptr_ring_cleanup(&q
->ring
, NULL
);
765 static int pfifo_fast_change_tx_queue_len(struct Qdisc
*sch
,
766 unsigned int new_len
)
768 struct pfifo_fast_priv
*priv
= qdisc_priv(sch
);
769 struct skb_array
*bands
[PFIFO_FAST_BANDS
];
772 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
773 struct skb_array
*q
= band2list(priv
, prio
);
778 return skb_array_resize_multiple(bands
, PFIFO_FAST_BANDS
, new_len
,
782 struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
784 .priv_size
= sizeof(struct pfifo_fast_priv
),
785 .enqueue
= pfifo_fast_enqueue
,
786 .dequeue
= pfifo_fast_dequeue
,
787 .peek
= pfifo_fast_peek
,
788 .init
= pfifo_fast_init
,
789 .destroy
= pfifo_fast_destroy
,
790 .reset
= pfifo_fast_reset
,
791 .dump
= pfifo_fast_dump
,
792 .change_tx_queue_len
= pfifo_fast_change_tx_queue_len
,
793 .owner
= THIS_MODULE
,
794 .static_flags
= TCQ_F_NOLOCK
| TCQ_F_CPUSTATS
,
796 EXPORT_SYMBOL(pfifo_fast_ops
);
798 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
799 const struct Qdisc_ops
*ops
,
800 struct netlink_ext_ack
*extack
)
804 unsigned int size
= QDISC_ALIGN(sizeof(*sch
)) + ops
->priv_size
;
806 struct net_device
*dev
;
809 NL_SET_ERR_MSG(extack
, "No device queue given");
814 dev
= dev_queue
->dev
;
815 p
= kzalloc_node(size
, GFP_KERNEL
,
816 netdev_queue_numa_node_read(dev_queue
));
820 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
821 /* if we got non aligned memory, ask more and do alignment ourself */
824 p
= kzalloc_node(size
+ QDISC_ALIGNTO
- 1, GFP_KERNEL
,
825 netdev_queue_numa_node_read(dev_queue
));
828 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
829 sch
->padded
= (char *) sch
- (char *) p
;
831 __skb_queue_head_init(&sch
->gso_skb
);
832 __skb_queue_head_init(&sch
->skb_bad_txq
);
833 qdisc_skb_head_init(&sch
->q
);
834 spin_lock_init(&sch
->q
.lock
);
836 if (ops
->static_flags
& TCQ_F_CPUSTATS
) {
838 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu
);
839 if (!sch
->cpu_bstats
)
842 sch
->cpu_qstats
= alloc_percpu(struct gnet_stats_queue
);
843 if (!sch
->cpu_qstats
) {
844 free_percpu(sch
->cpu_bstats
);
849 spin_lock_init(&sch
->busylock
);
850 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
851 spin_lock_init(&sch
->seqlock
);
852 seqcount_init(&sch
->running
);
855 sch
->flags
= ops
->static_flags
;
856 sch
->enqueue
= ops
->enqueue
;
857 sch
->dequeue
= ops
->dequeue
;
858 sch
->dev_queue
= dev_queue
;
861 refcount_set(&sch
->refcnt
, 1);
863 if (sch
!= &noop_qdisc
) {
864 lockdep_set_class(&sch
->busylock
, &dev
->qdisc_tx_busylock_key
);
865 lockdep_set_class(&sch
->seqlock
, &dev
->qdisc_tx_busylock_key
);
866 lockdep_set_class(&sch
->running
, &dev
->qdisc_running_key
);
876 struct Qdisc
*qdisc_create_dflt(struct netdev_queue
*dev_queue
,
877 const struct Qdisc_ops
*ops
,
878 unsigned int parentid
,
879 struct netlink_ext_ack
*extack
)
883 if (!try_module_get(ops
->owner
)) {
884 NL_SET_ERR_MSG(extack
, "Failed to increase module reference counter");
888 sch
= qdisc_alloc(dev_queue
, ops
, extack
);
890 module_put(ops
->owner
);
893 sch
->parent
= parentid
;
895 if (!ops
->init
|| ops
->init(sch
, NULL
, extack
) == 0)
901 EXPORT_SYMBOL(qdisc_create_dflt
);
903 /* Under qdisc_lock(qdisc) and BH! */
905 void qdisc_reset(struct Qdisc
*qdisc
)
907 const struct Qdisc_ops
*ops
= qdisc
->ops
;
908 struct sk_buff
*skb
, *tmp
;
913 skb_queue_walk_safe(&qdisc
->gso_skb
, skb
, tmp
) {
914 __skb_unlink(skb
, &qdisc
->gso_skb
);
918 skb_queue_walk_safe(&qdisc
->skb_bad_txq
, skb
, tmp
) {
919 __skb_unlink(skb
, &qdisc
->skb_bad_txq
);
924 qdisc
->qstats
.backlog
= 0;
926 EXPORT_SYMBOL(qdisc_reset
);
928 void qdisc_free(struct Qdisc
*qdisc
)
930 if (qdisc_is_percpu_stats(qdisc
)) {
931 free_percpu(qdisc
->cpu_bstats
);
932 free_percpu(qdisc
->cpu_qstats
);
935 kfree((char *) qdisc
- qdisc
->padded
);
938 static void qdisc_free_cb(struct rcu_head
*head
)
940 struct Qdisc
*q
= container_of(head
, struct Qdisc
, rcu
);
945 static void qdisc_destroy(struct Qdisc
*qdisc
)
947 const struct Qdisc_ops
*ops
= qdisc
->ops
;
948 struct sk_buff
*skb
, *tmp
;
950 #ifdef CONFIG_NET_SCHED
951 qdisc_hash_del(qdisc
);
953 qdisc_put_stab(rtnl_dereference(qdisc
->stab
));
955 gen_kill_estimator(&qdisc
->rate_est
);
961 module_put(ops
->owner
);
962 dev_put(qdisc_dev(qdisc
));
964 skb_queue_walk_safe(&qdisc
->gso_skb
, skb
, tmp
) {
965 __skb_unlink(skb
, &qdisc
->gso_skb
);
969 skb_queue_walk_safe(&qdisc
->skb_bad_txq
, skb
, tmp
) {
970 __skb_unlink(skb
, &qdisc
->skb_bad_txq
);
974 call_rcu(&qdisc
->rcu
, qdisc_free_cb
);
977 void qdisc_put(struct Qdisc
*qdisc
)
982 if (qdisc
->flags
& TCQ_F_BUILTIN
||
983 !refcount_dec_and_test(&qdisc
->refcnt
))
986 qdisc_destroy(qdisc
);
988 EXPORT_SYMBOL(qdisc_put
);
990 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
991 * Intended to be used as optimization, this function only takes rtnl lock if
992 * qdisc reference counter reached zero.
995 void qdisc_put_unlocked(struct Qdisc
*qdisc
)
997 if (qdisc
->flags
& TCQ_F_BUILTIN
||
998 !refcount_dec_and_rtnl_lock(&qdisc
->refcnt
))
1001 qdisc_destroy(qdisc
);
1004 EXPORT_SYMBOL(qdisc_put_unlocked
);
1006 /* Attach toplevel qdisc to device queue. */
1007 struct Qdisc
*dev_graft_qdisc(struct netdev_queue
*dev_queue
,
1008 struct Qdisc
*qdisc
)
1010 struct Qdisc
*oqdisc
= dev_queue
->qdisc_sleeping
;
1011 spinlock_t
*root_lock
;
1013 root_lock
= qdisc_lock(oqdisc
);
1014 spin_lock_bh(root_lock
);
1016 /* ... and graft new one */
1018 qdisc
= &noop_qdisc
;
1019 dev_queue
->qdisc_sleeping
= qdisc
;
1020 rcu_assign_pointer(dev_queue
->qdisc
, &noop_qdisc
);
1022 spin_unlock_bh(root_lock
);
1026 EXPORT_SYMBOL(dev_graft_qdisc
);
1028 static void attach_one_default_qdisc(struct net_device
*dev
,
1029 struct netdev_queue
*dev_queue
,
1032 struct Qdisc
*qdisc
;
1033 const struct Qdisc_ops
*ops
= default_qdisc_ops
;
1035 if (dev
->priv_flags
& IFF_NO_QUEUE
)
1036 ops
= &noqueue_qdisc_ops
;
1037 else if(dev
->type
== ARPHRD_CAN
)
1038 ops
= &pfifo_fast_ops
;
1040 qdisc
= qdisc_create_dflt(dev_queue
, ops
, TC_H_ROOT
, NULL
);
1042 netdev_info(dev
, "activation failed\n");
1045 if (!netif_is_multiqueue(dev
))
1046 qdisc
->flags
|= TCQ_F_ONETXQUEUE
| TCQ_F_NOPARENT
;
1047 dev_queue
->qdisc_sleeping
= qdisc
;
1050 static void attach_default_qdiscs(struct net_device
*dev
)
1052 struct netdev_queue
*txq
;
1053 struct Qdisc
*qdisc
;
1055 txq
= netdev_get_tx_queue(dev
, 0);
1057 if (!netif_is_multiqueue(dev
) ||
1058 dev
->priv_flags
& IFF_NO_QUEUE
) {
1059 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
1060 dev
->qdisc
= txq
->qdisc_sleeping
;
1061 qdisc_refcount_inc(dev
->qdisc
);
1063 qdisc
= qdisc_create_dflt(txq
, &mq_qdisc_ops
, TC_H_ROOT
, NULL
);
1066 qdisc
->ops
->attach(qdisc
);
1069 #ifdef CONFIG_NET_SCHED
1070 if (dev
->qdisc
!= &noop_qdisc
)
1071 qdisc_hash_add(dev
->qdisc
, false);
1075 static void transition_one_qdisc(struct net_device
*dev
,
1076 struct netdev_queue
*dev_queue
,
1077 void *_need_watchdog
)
1079 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
1080 int *need_watchdog_p
= _need_watchdog
;
1082 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
1083 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
1085 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
1086 if (need_watchdog_p
) {
1087 dev_queue
->trans_start
= 0;
1088 *need_watchdog_p
= 1;
1092 void dev_activate(struct net_device
*dev
)
1096 /* No queueing discipline is attached to device;
1097 * create default one for devices, which need queueing
1098 * and noqueue_qdisc for virtual interfaces
1101 if (dev
->qdisc
== &noop_qdisc
)
1102 attach_default_qdiscs(dev
);
1104 if (!netif_carrier_ok(dev
))
1105 /* Delay activation until next carrier-on event */
1109 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
1110 if (dev_ingress_queue(dev
))
1111 transition_one_qdisc(dev
, dev_ingress_queue(dev
), NULL
);
1113 if (need_watchdog
) {
1114 netif_trans_update(dev
);
1115 dev_watchdog_up(dev
);
1118 EXPORT_SYMBOL(dev_activate
);
1120 static void dev_deactivate_queue(struct net_device
*dev
,
1121 struct netdev_queue
*dev_queue
,
1122 void *_qdisc_default
)
1124 struct Qdisc
*qdisc_default
= _qdisc_default
;
1125 struct Qdisc
*qdisc
;
1127 qdisc
= rtnl_dereference(dev_queue
->qdisc
);
1129 bool nolock
= qdisc
->flags
& TCQ_F_NOLOCK
;
1132 spin_lock_bh(&qdisc
->seqlock
);
1133 spin_lock_bh(qdisc_lock(qdisc
));
1135 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
1136 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
1138 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
1141 spin_unlock_bh(qdisc_lock(qdisc
));
1143 spin_unlock_bh(&qdisc
->seqlock
);
1147 static bool some_qdisc_is_busy(struct net_device
*dev
)
1151 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
1152 struct netdev_queue
*dev_queue
;
1153 spinlock_t
*root_lock
;
1157 dev_queue
= netdev_get_tx_queue(dev
, i
);
1158 q
= dev_queue
->qdisc_sleeping
;
1160 root_lock
= qdisc_lock(q
);
1161 spin_lock_bh(root_lock
);
1163 val
= (qdisc_is_running(q
) ||
1164 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
1166 spin_unlock_bh(root_lock
);
1174 static void dev_qdisc_reset(struct net_device
*dev
,
1175 struct netdev_queue
*dev_queue
,
1178 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1185 * dev_deactivate_many - deactivate transmissions on several devices
1186 * @head: list of devices to deactivate
1188 * This function returns only when all outstanding transmissions
1189 * have completed, unless all devices are in dismantle phase.
1191 void dev_deactivate_many(struct list_head
*head
)
1193 struct net_device
*dev
;
1195 list_for_each_entry(dev
, head
, close_list
) {
1196 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
,
1198 if (dev_ingress_queue(dev
))
1199 dev_deactivate_queue(dev
, dev_ingress_queue(dev
),
1202 dev_watchdog_down(dev
);
1205 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1206 * This is avoided if all devices are in dismantle phase :
1207 * Caller will call synchronize_net() for us
1211 /* Wait for outstanding qdisc_run calls. */
1212 list_for_each_entry(dev
, head
, close_list
) {
1213 while (some_qdisc_is_busy(dev
)) {
1214 /* wait_event() would avoid this sleep-loop but would
1215 * require expensive checks in the fast paths of packet
1216 * processing which isn't worth it.
1218 schedule_timeout_uninterruptible(1);
1220 /* The new qdisc is assigned at this point so we can safely
1221 * unwind stale skb lists and qdisc statistics
1223 netdev_for_each_tx_queue(dev
, dev_qdisc_reset
, NULL
);
1224 if (dev_ingress_queue(dev
))
1225 dev_qdisc_reset(dev
, dev_ingress_queue(dev
), NULL
);
1229 void dev_deactivate(struct net_device
*dev
)
1233 list_add(&dev
->close_list
, &single
);
1234 dev_deactivate_many(&single
);
1237 EXPORT_SYMBOL(dev_deactivate
);
1239 static int qdisc_change_tx_queue_len(struct net_device
*dev
,
1240 struct netdev_queue
*dev_queue
)
1242 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1243 const struct Qdisc_ops
*ops
= qdisc
->ops
;
1245 if (ops
->change_tx_queue_len
)
1246 return ops
->change_tx_queue_len(qdisc
, dev
->tx_queue_len
);
1250 int dev_qdisc_change_tx_queue_len(struct net_device
*dev
)
1252 bool up
= dev
->flags
& IFF_UP
;
1257 dev_deactivate(dev
);
1259 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
1260 ret
= qdisc_change_tx_queue_len(dev
, &dev
->_tx
[i
]);
1262 /* TODO: revert changes on a partial failure */
1272 static void dev_init_scheduler_queue(struct net_device
*dev
,
1273 struct netdev_queue
*dev_queue
,
1276 struct Qdisc
*qdisc
= _qdisc
;
1278 rcu_assign_pointer(dev_queue
->qdisc
, qdisc
);
1279 dev_queue
->qdisc_sleeping
= qdisc
;
1282 void dev_init_scheduler(struct net_device
*dev
)
1284 dev
->qdisc
= &noop_qdisc
;
1285 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
1286 if (dev_ingress_queue(dev
))
1287 dev_init_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
1289 timer_setup(&dev
->watchdog_timer
, dev_watchdog
, 0);
1292 static void shutdown_scheduler_queue(struct net_device
*dev
,
1293 struct netdev_queue
*dev_queue
,
1294 void *_qdisc_default
)
1296 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
1297 struct Qdisc
*qdisc_default
= _qdisc_default
;
1300 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
1301 dev_queue
->qdisc_sleeping
= qdisc_default
;
1307 void dev_shutdown(struct net_device
*dev
)
1309 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
1310 if (dev_ingress_queue(dev
))
1311 shutdown_scheduler_queue(dev
, dev_ingress_queue(dev
), &noop_qdisc
);
1312 qdisc_put(dev
->qdisc
);
1313 dev
->qdisc
= &noop_qdisc
;
1315 WARN_ON(timer_pending(&dev
->watchdog_timer
));
1318 void psched_ratecfg_precompute(struct psched_ratecfg
*r
,
1319 const struct tc_ratespec
*conf
,
1322 memset(r
, 0, sizeof(*r
));
1323 r
->overhead
= conf
->overhead
;
1324 r
->rate_bytes_ps
= max_t(u64
, conf
->rate
, rate64
);
1325 r
->linklayer
= (conf
->linklayer
& TC_LINKLAYER_MASK
);
1328 * The deal here is to replace a divide by a reciprocal one
1329 * in fast path (a reciprocal divide is a multiply and a shift)
1331 * Normal formula would be :
1332 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1334 * We compute mult/shift to use instead :
1335 * time_in_ns = (len * mult) >> shift;
1337 * We try to get the highest possible mult value for accuracy,
1338 * but have to make sure no overflows will ever happen.
1340 if (r
->rate_bytes_ps
> 0) {
1341 u64 factor
= NSEC_PER_SEC
;
1344 r
->mult
= div64_u64(factor
, r
->rate_bytes_ps
);
1345 if (r
->mult
& (1U << 31) || factor
& (1ULL << 63))
1352 EXPORT_SYMBOL(psched_ratecfg_precompute
);
1354 static void mini_qdisc_rcu_func(struct rcu_head
*head
)
1358 void mini_qdisc_pair_swap(struct mini_Qdisc_pair
*miniqp
,
1359 struct tcf_proto
*tp_head
)
1361 /* Protected with chain0->filter_chain_lock.
1362 * Can't access chain directly because tp_head can be NULL.
1364 struct mini_Qdisc
*miniq_old
=
1365 rcu_dereference_protected(*miniqp
->p_miniq
, 1);
1366 struct mini_Qdisc
*miniq
;
1369 RCU_INIT_POINTER(*miniqp
->p_miniq
, NULL
);
1370 /* Wait for flying RCU callback before it is freed. */
1375 miniq
= !miniq_old
|| miniq_old
== &miniqp
->miniq2
?
1376 &miniqp
->miniq1
: &miniqp
->miniq2
;
1378 /* We need to make sure that readers won't see the miniq
1379 * we are about to modify. So wait until previous call_rcu callback
1383 miniq
->filter_list
= tp_head
;
1384 rcu_assign_pointer(*miniqp
->p_miniq
, miniq
);
1387 /* This is counterpart of the rcu barriers above. We need to
1388 * block potential new user of miniq_old until all readers
1389 * are not seeing it.
1391 call_rcu(&miniq_old
->rcu
, mini_qdisc_rcu_func
);
1393 EXPORT_SYMBOL(mini_qdisc_pair_swap
);
1395 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair
*miniqp
,
1396 struct tcf_block
*block
)
1398 miniqp
->miniq1
.block
= block
;
1399 miniqp
->miniq2
.block
= block
;
1401 EXPORT_SYMBOL(mini_qdisc_pair_block_init
);
1403 void mini_qdisc_pair_init(struct mini_Qdisc_pair
*miniqp
, struct Qdisc
*qdisc
,
1404 struct mini_Qdisc __rcu
**p_miniq
)
1406 miniqp
->miniq1
.cpu_bstats
= qdisc
->cpu_bstats
;
1407 miniqp
->miniq1
.cpu_qstats
= qdisc
->cpu_qstats
;
1408 miniqp
->miniq2
.cpu_bstats
= qdisc
->cpu_bstats
;
1409 miniqp
->miniq2
.cpu_qstats
= qdisc
->cpu_qstats
;
1410 miniqp
->p_miniq
= p_miniq
;
1412 EXPORT_SYMBOL(mini_qdisc_pair_init
);