2 * net/sched/sch_generic.c Generic packet scheduler routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
14 #include <linux/bitops.h>
15 #include <linux/module.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/sched.h>
19 #include <linux/string.h>
20 #include <linux/errno.h>
21 #include <linux/netdevice.h>
22 #include <linux/skbuff.h>
23 #include <linux/rtnetlink.h>
24 #include <linux/init.h>
25 #include <linux/rcupdate.h>
26 #include <linux/list.h>
27 #include <net/pkt_sched.h>
29 /* Main transmission queue. */
31 /* Modifications to data participating in scheduling must be protected with
32 * qdisc_lock(qdisc) spinlock.
34 * The idea is the following:
35 * - enqueue, dequeue are serialized via qdisc root lock
36 * - ingress filtering is also serialized via qdisc root lock
37 * - updates to tree and tree walking are only done under the rtnl mutex.
40 static inline int dev_requeue_skb(struct sk_buff
*skb
, struct Qdisc
*q
)
44 q
->q
.qlen
++; /* it's still part of the queue */
50 static inline struct sk_buff
*dequeue_skb(struct Qdisc
*q
)
52 struct sk_buff
*skb
= q
->gso_skb
;
55 struct net_device
*dev
= qdisc_dev(q
);
56 struct netdev_queue
*txq
;
58 /* check the reason of requeuing without tx lock first */
59 txq
= netdev_get_tx_queue(dev
, skb_get_queue_mapping(skb
));
60 if (!netif_tx_queue_stopped(txq
) &&
61 !netif_tx_queue_frozen(txq
)) {
73 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
74 struct netdev_queue
*dev_queue
,
79 if (unlikely(dev_queue
->xmit_lock_owner
== smp_processor_id())) {
81 * Same CPU holding the lock. It may be a transient
82 * configuration error, when hard_start_xmit() recurses. We
83 * detect it by checking xmit owner and drop the packet when
84 * deadloop is detected. Return OK to try the next skb.
88 printk(KERN_WARNING
"Dead loop on netdevice %s, "
89 "fix it urgently!\n", dev_queue
->dev
->name
);
93 * Another cpu is holding lock, requeue & delay xmits for
96 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
97 ret
= dev_requeue_skb(skb
, q
);
104 * Transmit one skb, and handle the return status as required. Holding the
105 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
108 * Returns to the caller:
109 * 0 - queue is empty or throttled.
110 * >0 - queue is not empty.
112 int sch_direct_xmit(struct sk_buff
*skb
, struct Qdisc
*q
,
113 struct net_device
*dev
, struct netdev_queue
*txq
,
114 spinlock_t
*root_lock
)
116 int ret
= NETDEV_TX_BUSY
;
118 /* And release qdisc */
119 spin_unlock(root_lock
);
121 HARD_TX_LOCK(dev
, txq
, smp_processor_id());
122 if (!netif_tx_queue_stopped(txq
) &&
123 !netif_tx_queue_frozen(txq
))
124 ret
= dev_hard_start_xmit(skb
, dev
, txq
);
125 HARD_TX_UNLOCK(dev
, txq
);
127 spin_lock(root_lock
);
131 /* Driver sent out skb successfully */
135 case NETDEV_TX_LOCKED
:
136 /* Driver try lock failed */
137 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
141 /* Driver returned NETDEV_TX_BUSY - requeue skb */
142 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
143 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
144 dev
->name
, ret
, q
->q
.qlen
);
146 ret
= dev_requeue_skb(skb
, q
);
150 if (ret
&& (netif_tx_queue_stopped(txq
) ||
151 netif_tx_queue_frozen(txq
)))
158 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
160 * __QDISC_STATE_RUNNING guarantees only one CPU can process
161 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
164 * netif_tx_lock serializes accesses to device driver.
166 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
167 * if one is grabbed, another must be free.
169 * Note, that this procedure can be called by a watchdog timer
171 * Returns to the caller:
172 * 0 - queue is empty or throttled.
173 * >0 - queue is not empty.
176 static inline int qdisc_restart(struct Qdisc
*q
)
178 struct netdev_queue
*txq
;
179 struct net_device
*dev
;
180 spinlock_t
*root_lock
;
184 skb
= dequeue_skb(q
);
188 root_lock
= qdisc_lock(q
);
190 txq
= netdev_get_tx_queue(dev
, skb_get_queue_mapping(skb
));
192 return sch_direct_xmit(skb
, q
, dev
, txq
, root_lock
);
195 void __qdisc_run(struct Qdisc
*q
)
197 unsigned long start_time
= jiffies
;
199 while (qdisc_restart(q
)) {
201 * Postpone processing if
202 * 1. another process needs the CPU;
203 * 2. we've been doing it for too long.
205 if (need_resched() || jiffies
!= start_time
) {
211 clear_bit(__QDISC_STATE_RUNNING
, &q
->state
);
214 unsigned long dev_trans_start(struct net_device
*dev
)
216 unsigned long val
, res
= dev
->trans_start
;
219 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
220 val
= netdev_get_tx_queue(dev
, i
)->trans_start
;
221 if (val
&& time_after(val
, res
))
224 dev
->trans_start
= res
;
227 EXPORT_SYMBOL(dev_trans_start
);
229 static void dev_watchdog(unsigned long arg
)
231 struct net_device
*dev
= (struct net_device
*)arg
;
234 if (!qdisc_tx_is_noop(dev
)) {
235 if (netif_device_present(dev
) &&
236 netif_running(dev
) &&
237 netif_carrier_ok(dev
)) {
238 int some_queue_timedout
= 0;
240 unsigned long trans_start
;
242 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
243 struct netdev_queue
*txq
;
245 txq
= netdev_get_tx_queue(dev
, i
);
247 * old device drivers set dev->trans_start
249 trans_start
= txq
->trans_start
? : dev
->trans_start
;
250 if (netif_tx_queue_stopped(txq
) &&
251 time_after(jiffies
, (trans_start
+
252 dev
->watchdog_timeo
))) {
253 some_queue_timedout
= 1;
258 if (some_queue_timedout
) {
260 WARN_ONCE(1, KERN_INFO
"NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
261 dev
->name
, netdev_drivername(dev
, drivername
, 64), i
);
262 dev
->netdev_ops
->ndo_tx_timeout(dev
);
264 if (!mod_timer(&dev
->watchdog_timer
,
265 round_jiffies(jiffies
+
266 dev
->watchdog_timeo
)))
270 netif_tx_unlock(dev
);
275 void __netdev_watchdog_up(struct net_device
*dev
)
277 if (dev
->netdev_ops
->ndo_tx_timeout
) {
278 if (dev
->watchdog_timeo
<= 0)
279 dev
->watchdog_timeo
= 5*HZ
;
280 if (!mod_timer(&dev
->watchdog_timer
,
281 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
286 static void dev_watchdog_up(struct net_device
*dev
)
288 __netdev_watchdog_up(dev
);
291 static void dev_watchdog_down(struct net_device
*dev
)
293 netif_tx_lock_bh(dev
);
294 if (del_timer(&dev
->watchdog_timer
))
296 netif_tx_unlock_bh(dev
);
300 * netif_carrier_on - set carrier
301 * @dev: network device
303 * Device has detected that carrier.
305 void netif_carrier_on(struct net_device
*dev
)
307 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
308 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
310 linkwatch_fire_event(dev
);
311 if (netif_running(dev
))
312 __netdev_watchdog_up(dev
);
315 EXPORT_SYMBOL(netif_carrier_on
);
318 * netif_carrier_off - clear carrier
319 * @dev: network device
321 * Device has detected loss of carrier.
323 void netif_carrier_off(struct net_device
*dev
)
325 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
326 if (dev
->reg_state
== NETREG_UNINITIALIZED
)
328 linkwatch_fire_event(dev
);
331 EXPORT_SYMBOL(netif_carrier_off
);
333 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
334 under all circumstances. It is difficult to invent anything faster or
338 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
344 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
349 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
352 .enqueue
= noop_enqueue
,
353 .dequeue
= noop_dequeue
,
354 .peek
= noop_dequeue
,
355 .owner
= THIS_MODULE
,
358 static struct netdev_queue noop_netdev_queue
= {
359 .qdisc
= &noop_qdisc
,
360 .qdisc_sleeping
= &noop_qdisc
,
363 struct Qdisc noop_qdisc
= {
364 .enqueue
= noop_enqueue
,
365 .dequeue
= noop_dequeue
,
366 .flags
= TCQ_F_BUILTIN
,
367 .ops
= &noop_qdisc_ops
,
368 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
369 .q
.lock
= __SPIN_LOCK_UNLOCKED(noop_qdisc
.q
.lock
),
370 .dev_queue
= &noop_netdev_queue
,
372 EXPORT_SYMBOL(noop_qdisc
);
374 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
377 .enqueue
= noop_enqueue
,
378 .dequeue
= noop_dequeue
,
379 .peek
= noop_dequeue
,
380 .owner
= THIS_MODULE
,
383 static struct Qdisc noqueue_qdisc
;
384 static struct netdev_queue noqueue_netdev_queue
= {
385 .qdisc
= &noqueue_qdisc
,
386 .qdisc_sleeping
= &noqueue_qdisc
,
389 static struct Qdisc noqueue_qdisc
= {
391 .dequeue
= noop_dequeue
,
392 .flags
= TCQ_F_BUILTIN
,
393 .ops
= &noqueue_qdisc_ops
,
394 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
395 .q
.lock
= __SPIN_LOCK_UNLOCKED(noqueue_qdisc
.q
.lock
),
396 .dev_queue
= &noqueue_netdev_queue
,
400 static const u8 prio2band
[TC_PRIO_MAX
+1] =
401 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
403 /* 3-band FIFO queue: old style, but should be a bit faster than
404 generic prio+fifo combination.
407 #define PFIFO_FAST_BANDS 3
409 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
412 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
413 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
416 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
418 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
420 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
422 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
425 return qdisc_drop(skb
, qdisc
);
428 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
431 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
433 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
434 if (!skb_queue_empty(list
+ prio
)) {
436 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
443 static struct sk_buff
*pfifo_fast_peek(struct Qdisc
* qdisc
)
446 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
448 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
449 if (!skb_queue_empty(list
+ prio
))
450 return skb_peek(list
+ prio
);
456 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
459 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
461 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
462 __qdisc_reset_queue(qdisc
, list
+ prio
);
464 qdisc
->qstats
.backlog
= 0;
468 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
470 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
472 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
473 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
480 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
483 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
485 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
486 skb_queue_head_init(list
+ prio
);
491 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
493 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
494 .enqueue
= pfifo_fast_enqueue
,
495 .dequeue
= pfifo_fast_dequeue
,
496 .peek
= pfifo_fast_peek
,
497 .init
= pfifo_fast_init
,
498 .reset
= pfifo_fast_reset
,
499 .dump
= pfifo_fast_dump
,
500 .owner
= THIS_MODULE
,
503 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
504 struct Qdisc_ops
*ops
)
511 /* ensure that the Qdisc and the private data are 32-byte aligned */
512 size
= QDISC_ALIGN(sizeof(*sch
));
513 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
515 p
= kzalloc(size
, GFP_KERNEL
);
518 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
519 sch
->padded
= (char *) sch
- (char *) p
;
521 INIT_LIST_HEAD(&sch
->list
);
522 skb_queue_head_init(&sch
->q
);
524 sch
->enqueue
= ops
->enqueue
;
525 sch
->dequeue
= ops
->dequeue
;
526 sch
->dev_queue
= dev_queue
;
527 dev_hold(qdisc_dev(sch
));
528 atomic_set(&sch
->refcnt
, 1);
535 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
536 struct netdev_queue
*dev_queue
,
537 struct Qdisc_ops
*ops
,
538 unsigned int parentid
)
542 sch
= qdisc_alloc(dev_queue
, ops
);
545 sch
->parent
= parentid
;
547 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
554 EXPORT_SYMBOL(qdisc_create_dflt
);
556 /* Under qdisc_lock(qdisc) and BH! */
558 void qdisc_reset(struct Qdisc
*qdisc
)
560 const struct Qdisc_ops
*ops
= qdisc
->ops
;
565 if (qdisc
->gso_skb
) {
566 kfree_skb(qdisc
->gso_skb
);
567 qdisc
->gso_skb
= NULL
;
571 EXPORT_SYMBOL(qdisc_reset
);
573 void qdisc_destroy(struct Qdisc
*qdisc
)
575 const struct Qdisc_ops
*ops
= qdisc
->ops
;
577 if (qdisc
->flags
& TCQ_F_BUILTIN
||
578 !atomic_dec_and_test(&qdisc
->refcnt
))
581 #ifdef CONFIG_NET_SCHED
582 qdisc_list_del(qdisc
);
584 qdisc_put_stab(qdisc
->stab
);
586 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
592 module_put(ops
->owner
);
593 dev_put(qdisc_dev(qdisc
));
595 kfree_skb(qdisc
->gso_skb
);
596 kfree((char *) qdisc
- qdisc
->padded
);
598 EXPORT_SYMBOL(qdisc_destroy
);
600 static bool dev_all_qdisc_sleeping_noop(struct net_device
*dev
)
604 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
605 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, i
);
607 if (txq
->qdisc_sleeping
!= &noop_qdisc
)
613 static void attach_one_default_qdisc(struct net_device
*dev
,
614 struct netdev_queue
*dev_queue
,
619 if (dev
->tx_queue_len
) {
620 qdisc
= qdisc_create_dflt(dev
, dev_queue
,
621 &pfifo_fast_ops
, TC_H_ROOT
);
623 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
627 /* Can by-pass the queue discipline for default qdisc */
628 qdisc
->flags
|= TCQ_F_CAN_BYPASS
;
630 qdisc
= &noqueue_qdisc
;
632 dev_queue
->qdisc_sleeping
= qdisc
;
635 static void transition_one_qdisc(struct net_device
*dev
,
636 struct netdev_queue
*dev_queue
,
637 void *_need_watchdog
)
639 struct Qdisc
*new_qdisc
= dev_queue
->qdisc_sleeping
;
640 int *need_watchdog_p
= _need_watchdog
;
642 if (!(new_qdisc
->flags
& TCQ_F_BUILTIN
))
643 clear_bit(__QDISC_STATE_DEACTIVATED
, &new_qdisc
->state
);
645 rcu_assign_pointer(dev_queue
->qdisc
, new_qdisc
);
646 if (need_watchdog_p
&& new_qdisc
!= &noqueue_qdisc
) {
647 dev_queue
->trans_start
= 0;
648 *need_watchdog_p
= 1;
652 void dev_activate(struct net_device
*dev
)
656 /* No queueing discipline is attached to device;
657 create default one i.e. pfifo_fast for devices,
658 which need queueing and noqueue_qdisc for
662 if (dev_all_qdisc_sleeping_noop(dev
))
663 netdev_for_each_tx_queue(dev
, attach_one_default_qdisc
, NULL
);
665 if (!netif_carrier_ok(dev
))
666 /* Delay activation until next carrier-on event */
670 netdev_for_each_tx_queue(dev
, transition_one_qdisc
, &need_watchdog
);
671 transition_one_qdisc(dev
, &dev
->rx_queue
, NULL
);
674 dev
->trans_start
= jiffies
;
675 dev_watchdog_up(dev
);
679 static void dev_deactivate_queue(struct net_device
*dev
,
680 struct netdev_queue
*dev_queue
,
681 void *_qdisc_default
)
683 struct Qdisc
*qdisc_default
= _qdisc_default
;
686 qdisc
= dev_queue
->qdisc
;
688 spin_lock_bh(qdisc_lock(qdisc
));
690 if (!(qdisc
->flags
& TCQ_F_BUILTIN
))
691 set_bit(__QDISC_STATE_DEACTIVATED
, &qdisc
->state
);
693 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
696 spin_unlock_bh(qdisc_lock(qdisc
));
700 static bool some_qdisc_is_busy(struct net_device
*dev
)
704 for (i
= 0; i
< dev
->num_tx_queues
; i
++) {
705 struct netdev_queue
*dev_queue
;
706 spinlock_t
*root_lock
;
710 dev_queue
= netdev_get_tx_queue(dev
, i
);
711 q
= dev_queue
->qdisc_sleeping
;
712 root_lock
= qdisc_lock(q
);
714 spin_lock_bh(root_lock
);
716 val
= (test_bit(__QDISC_STATE_RUNNING
, &q
->state
) ||
717 test_bit(__QDISC_STATE_SCHED
, &q
->state
));
719 spin_unlock_bh(root_lock
);
727 void dev_deactivate(struct net_device
*dev
)
729 netdev_for_each_tx_queue(dev
, dev_deactivate_queue
, &noop_qdisc
);
730 dev_deactivate_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
732 dev_watchdog_down(dev
);
734 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
737 /* Wait for outstanding qdisc_run calls. */
738 while (some_qdisc_is_busy(dev
))
742 static void dev_init_scheduler_queue(struct net_device
*dev
,
743 struct netdev_queue
*dev_queue
,
746 struct Qdisc
*qdisc
= _qdisc
;
748 dev_queue
->qdisc
= qdisc
;
749 dev_queue
->qdisc_sleeping
= qdisc
;
752 void dev_init_scheduler(struct net_device
*dev
)
754 netdev_for_each_tx_queue(dev
, dev_init_scheduler_queue
, &noop_qdisc
);
755 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
757 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
760 static void shutdown_scheduler_queue(struct net_device
*dev
,
761 struct netdev_queue
*dev_queue
,
762 void *_qdisc_default
)
764 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
765 struct Qdisc
*qdisc_default
= _qdisc_default
;
768 rcu_assign_pointer(dev_queue
->qdisc
, qdisc_default
);
769 dev_queue
->qdisc_sleeping
= qdisc_default
;
771 qdisc_destroy(qdisc
);
775 void dev_shutdown(struct net_device
*dev
)
777 netdev_for_each_tx_queue(dev
, shutdown_scheduler_queue
, &noop_qdisc
);
778 shutdown_scheduler_queue(dev
, &dev
->rx_queue
, &noop_qdisc
);
779 WARN_ON(timer_pending(&dev
->watchdog_timer
));