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 * queue->lock spinlock.
34 * The idea is the following:
35 * - enqueue, dequeue are serialized via top level device
36 * spinlock queue->lock.
37 * - ingress filtering is serialized via top level device
38 * spinlock dev->rx_queue.lock.
39 * - updates to tree and tree walking are only done under the rtnl mutex.
42 void qdisc_lock_tree(struct net_device
*dev
)
43 __acquires(dev
->tx_queue
.lock
)
44 __acquires(dev
->rx_queue
.lock
)
46 spin_lock_bh(&dev
->tx_queue
.lock
);
47 spin_lock(&dev
->rx_queue
.lock
);
49 EXPORT_SYMBOL(qdisc_lock_tree
);
51 void qdisc_unlock_tree(struct net_device
*dev
)
52 __releases(dev
->rx_queue
.lock
)
53 __releases(dev
->tx_queue
.lock
)
55 spin_unlock(&dev
->rx_queue
.lock
);
56 spin_unlock_bh(&dev
->tx_queue
.lock
);
58 EXPORT_SYMBOL(qdisc_unlock_tree
);
60 static inline int qdisc_qlen(struct Qdisc
*q
)
65 static inline int dev_requeue_skb(struct sk_buff
*skb
,
66 struct netdev_queue
*dev_queue
,
69 if (unlikely(skb
->next
))
70 dev_queue
->gso_skb
= skb
;
72 q
->ops
->requeue(skb
, q
);
74 netif_schedule_queue(dev_queue
);
78 static inline struct sk_buff
*dequeue_skb(struct netdev_queue
*dev_queue
,
83 if ((skb
= dev_queue
->gso_skb
))
84 dev_queue
->gso_skb
= NULL
;
91 static inline int handle_dev_cpu_collision(struct sk_buff
*skb
,
92 struct netdev_queue
*dev_queue
,
95 struct net_device
*dev
= dev_queue
->dev
;
98 if (unlikely(dev
->xmit_lock_owner
== smp_processor_id())) {
100 * Same CPU holding the lock. It may be a transient
101 * configuration error, when hard_start_xmit() recurses. We
102 * detect it by checking xmit owner and drop the packet when
103 * deadloop is detected. Return OK to try the next skb.
107 printk(KERN_WARNING
"Dead loop on netdevice %s, "
108 "fix it urgently!\n", dev
->name
);
112 * Another cpu is holding lock, requeue & delay xmits for
115 __get_cpu_var(netdev_rx_stat
).cpu_collision
++;
116 ret
= dev_requeue_skb(skb
, dev_queue
, q
);
123 * NOTE: Called under queue->lock with locally disabled BH.
125 * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
126 * device at a time. queue->lock serializes queue accesses for
127 * this device AND txq->qdisc pointer itself.
129 * netif_tx_lock serializes accesses to device driver.
131 * queue->lock and netif_tx_lock are mutually exclusive,
132 * if one is grabbed, another must be free.
134 * Note, that this procedure can be called by a watchdog timer
136 * Returns to the caller:
137 * 0 - queue is empty or throttled.
138 * >0 - queue is not empty.
141 static inline int qdisc_restart(struct netdev_queue
*txq
)
143 struct Qdisc
*q
= txq
->qdisc
;
144 int ret
= NETDEV_TX_BUSY
;
145 struct net_device
*dev
;
149 if (unlikely((skb
= dequeue_skb(txq
, q
)) == NULL
))
153 /* And release queue */
154 spin_unlock(&txq
->lock
);
158 HARD_TX_LOCK(dev
, smp_processor_id());
159 if (!netif_subqueue_stopped(dev
, skb
))
160 ret
= dev_hard_start_xmit(skb
, dev
);
163 spin_lock(&txq
->lock
);
168 /* Driver sent out skb successfully */
172 case NETDEV_TX_LOCKED
:
173 /* Driver try lock failed */
174 ret
= handle_dev_cpu_collision(skb
, txq
, q
);
178 /* Driver returned NETDEV_TX_BUSY - requeue skb */
179 if (unlikely (ret
!= NETDEV_TX_BUSY
&& net_ratelimit()))
180 printk(KERN_WARNING
"BUG %s code %d qlen %d\n",
181 dev
->name
, ret
, q
->q
.qlen
);
183 ret
= dev_requeue_skb(skb
, txq
, q
);
190 void __qdisc_run(struct netdev_queue
*txq
)
192 struct net_device
*dev
= txq
->dev
;
193 unsigned long start_time
= jiffies
;
195 while (qdisc_restart(txq
)) {
196 if (netif_queue_stopped(dev
))
200 * Postpone processing if
201 * 1. another process needs the CPU;
202 * 2. we've been doing it for too long.
204 if (need_resched() || jiffies
!= start_time
) {
205 netif_schedule_queue(txq
);
210 clear_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
213 static void dev_watchdog(unsigned long arg
)
215 struct net_device
*dev
= (struct net_device
*)arg
;
216 struct netdev_queue
*txq
= &dev
->tx_queue
;
219 if (txq
->qdisc
!= &noop_qdisc
) {
220 if (netif_device_present(dev
) &&
221 netif_running(dev
) &&
222 netif_carrier_ok(dev
)) {
223 if (netif_queue_stopped(dev
) &&
224 time_after(jiffies
, dev
->trans_start
+ dev
->watchdog_timeo
)) {
226 printk(KERN_INFO
"NETDEV WATCHDOG: %s: transmit timed out\n",
228 dev
->tx_timeout(dev
);
231 if (!mod_timer(&dev
->watchdog_timer
, round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
235 netif_tx_unlock(dev
);
240 void __netdev_watchdog_up(struct net_device
*dev
)
242 if (dev
->tx_timeout
) {
243 if (dev
->watchdog_timeo
<= 0)
244 dev
->watchdog_timeo
= 5*HZ
;
245 if (!mod_timer(&dev
->watchdog_timer
,
246 round_jiffies(jiffies
+ dev
->watchdog_timeo
)))
251 static void dev_watchdog_up(struct net_device
*dev
)
253 __netdev_watchdog_up(dev
);
256 static void dev_watchdog_down(struct net_device
*dev
)
258 netif_tx_lock_bh(dev
);
259 if (del_timer(&dev
->watchdog_timer
))
261 netif_tx_unlock_bh(dev
);
265 * netif_carrier_on - set carrier
266 * @dev: network device
268 * Device has detected that carrier.
270 void netif_carrier_on(struct net_device
*dev
)
272 if (test_and_clear_bit(__LINK_STATE_NOCARRIER
, &dev
->state
)) {
273 linkwatch_fire_event(dev
);
274 if (netif_running(dev
))
275 __netdev_watchdog_up(dev
);
278 EXPORT_SYMBOL(netif_carrier_on
);
281 * netif_carrier_off - clear carrier
282 * @dev: network device
284 * Device has detected loss of carrier.
286 void netif_carrier_off(struct net_device
*dev
)
288 if (!test_and_set_bit(__LINK_STATE_NOCARRIER
, &dev
->state
))
289 linkwatch_fire_event(dev
);
291 EXPORT_SYMBOL(netif_carrier_off
);
293 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
294 under all circumstances. It is difficult to invent anything faster or
298 static int noop_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
304 static struct sk_buff
*noop_dequeue(struct Qdisc
* qdisc
)
309 static int noop_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
312 printk(KERN_DEBUG
"%s deferred output. It is buggy.\n",
318 struct Qdisc_ops noop_qdisc_ops __read_mostly
= {
321 .enqueue
= noop_enqueue
,
322 .dequeue
= noop_dequeue
,
323 .requeue
= noop_requeue
,
324 .owner
= THIS_MODULE
,
327 struct Qdisc noop_qdisc
= {
328 .enqueue
= noop_enqueue
,
329 .dequeue
= noop_dequeue
,
330 .flags
= TCQ_F_BUILTIN
,
331 .ops
= &noop_qdisc_ops
,
332 .list
= LIST_HEAD_INIT(noop_qdisc
.list
),
334 EXPORT_SYMBOL(noop_qdisc
);
336 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly
= {
339 .enqueue
= noop_enqueue
,
340 .dequeue
= noop_dequeue
,
341 .requeue
= noop_requeue
,
342 .owner
= THIS_MODULE
,
345 static struct Qdisc noqueue_qdisc
= {
347 .dequeue
= noop_dequeue
,
348 .flags
= TCQ_F_BUILTIN
,
349 .ops
= &noqueue_qdisc_ops
,
350 .list
= LIST_HEAD_INIT(noqueue_qdisc
.list
),
354 static const u8 prio2band
[TC_PRIO_MAX
+1] =
355 { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
357 /* 3-band FIFO queue: old style, but should be a bit faster than
358 generic prio+fifo combination.
361 #define PFIFO_FAST_BANDS 3
363 static inline struct sk_buff_head
*prio2list(struct sk_buff
*skb
,
366 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
367 return list
+ prio2band
[skb
->priority
& TC_PRIO_MAX
];
370 static int pfifo_fast_enqueue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
372 struct sk_buff_head
*list
= prio2list(skb
, qdisc
);
374 if (skb_queue_len(list
) < qdisc_dev(qdisc
)->tx_queue_len
) {
376 return __qdisc_enqueue_tail(skb
, qdisc
, list
);
379 return qdisc_drop(skb
, qdisc
);
382 static struct sk_buff
*pfifo_fast_dequeue(struct Qdisc
* qdisc
)
385 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
387 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++) {
388 if (!skb_queue_empty(list
+ prio
)) {
390 return __qdisc_dequeue_head(qdisc
, list
+ prio
);
397 static int pfifo_fast_requeue(struct sk_buff
*skb
, struct Qdisc
* qdisc
)
400 return __qdisc_requeue(skb
, qdisc
, prio2list(skb
, qdisc
));
403 static void pfifo_fast_reset(struct Qdisc
* qdisc
)
406 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
408 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
409 __qdisc_reset_queue(qdisc
, list
+ prio
);
411 qdisc
->qstats
.backlog
= 0;
415 static int pfifo_fast_dump(struct Qdisc
*qdisc
, struct sk_buff
*skb
)
417 struct tc_prio_qopt opt
= { .bands
= PFIFO_FAST_BANDS
};
419 memcpy(&opt
.priomap
, prio2band
, TC_PRIO_MAX
+1);
420 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
427 static int pfifo_fast_init(struct Qdisc
*qdisc
, struct nlattr
*opt
)
430 struct sk_buff_head
*list
= qdisc_priv(qdisc
);
432 for (prio
= 0; prio
< PFIFO_FAST_BANDS
; prio
++)
433 skb_queue_head_init(list
+ prio
);
438 static struct Qdisc_ops pfifo_fast_ops __read_mostly
= {
440 .priv_size
= PFIFO_FAST_BANDS
* sizeof(struct sk_buff_head
),
441 .enqueue
= pfifo_fast_enqueue
,
442 .dequeue
= pfifo_fast_dequeue
,
443 .requeue
= pfifo_fast_requeue
,
444 .init
= pfifo_fast_init
,
445 .reset
= pfifo_fast_reset
,
446 .dump
= pfifo_fast_dump
,
447 .owner
= THIS_MODULE
,
450 struct Qdisc
*qdisc_alloc(struct netdev_queue
*dev_queue
,
451 struct Qdisc_ops
*ops
)
458 /* ensure that the Qdisc and the private data are 32-byte aligned */
459 size
= QDISC_ALIGN(sizeof(*sch
));
460 size
+= ops
->priv_size
+ (QDISC_ALIGNTO
- 1);
462 p
= kzalloc(size
, GFP_KERNEL
);
465 sch
= (struct Qdisc
*) QDISC_ALIGN((unsigned long) p
);
466 sch
->padded
= (char *) sch
- (char *) p
;
468 INIT_LIST_HEAD(&sch
->list
);
469 skb_queue_head_init(&sch
->q
);
471 sch
->enqueue
= ops
->enqueue
;
472 sch
->dequeue
= ops
->dequeue
;
473 sch
->dev_queue
= dev_queue
;
474 dev_hold(qdisc_dev(sch
));
475 atomic_set(&sch
->refcnt
, 1);
482 struct Qdisc
* qdisc_create_dflt(struct net_device
*dev
,
483 struct netdev_queue
*dev_queue
,
484 struct Qdisc_ops
*ops
,
485 unsigned int parentid
)
489 sch
= qdisc_alloc(dev_queue
, ops
);
492 sch
->parent
= parentid
;
494 if (!ops
->init
|| ops
->init(sch
, NULL
) == 0)
501 EXPORT_SYMBOL(qdisc_create_dflt
);
503 /* Under queue->lock and BH! */
505 void qdisc_reset(struct Qdisc
*qdisc
)
507 const struct Qdisc_ops
*ops
= qdisc
->ops
;
512 EXPORT_SYMBOL(qdisc_reset
);
514 /* this is the rcu callback function to clean up a qdisc when there
515 * are no further references to it */
517 static void __qdisc_destroy(struct rcu_head
*head
)
519 struct Qdisc
*qdisc
= container_of(head
, struct Qdisc
, q_rcu
);
520 kfree((char *) qdisc
- qdisc
->padded
);
523 /* Under queue->lock and BH! */
525 void qdisc_destroy(struct Qdisc
*qdisc
)
527 const struct Qdisc_ops
*ops
= qdisc
->ops
;
529 if (qdisc
->flags
& TCQ_F_BUILTIN
||
530 !atomic_dec_and_test(&qdisc
->refcnt
))
533 list_del(&qdisc
->list
);
534 gen_kill_estimator(&qdisc
->bstats
, &qdisc
->rate_est
);
540 module_put(ops
->owner
);
541 dev_put(qdisc_dev(qdisc
));
542 call_rcu(&qdisc
->q_rcu
, __qdisc_destroy
);
544 EXPORT_SYMBOL(qdisc_destroy
);
546 void dev_activate(struct net_device
*dev
)
548 struct netdev_queue
*txq
= &dev
->tx_queue
;
550 /* No queueing discipline is attached to device;
551 create default one i.e. pfifo_fast for devices,
552 which need queueing and noqueue_qdisc for
556 if (txq
->qdisc_sleeping
== &noop_qdisc
) {
558 if (dev
->tx_queue_len
) {
559 qdisc
= qdisc_create_dflt(dev
, txq
,
563 printk(KERN_INFO
"%s: activation failed\n", dev
->name
);
566 list_add_tail(&qdisc
->list
, &txq
->qdisc_list
);
568 qdisc
= &noqueue_qdisc
;
570 txq
->qdisc_sleeping
= qdisc
;
573 if (!netif_carrier_ok(dev
))
574 /* Delay activation until next carrier-on event */
577 spin_lock_bh(&txq
->lock
);
578 rcu_assign_pointer(txq
->qdisc
, txq
->qdisc_sleeping
);
579 if (txq
->qdisc
!= &noqueue_qdisc
) {
580 dev
->trans_start
= jiffies
;
581 dev_watchdog_up(dev
);
583 spin_unlock_bh(&txq
->lock
);
586 static void dev_deactivate_queue(struct netdev_queue
*dev_queue
,
587 struct Qdisc
*qdisc_default
)
592 spin_lock_bh(&dev_queue
->lock
);
594 qdisc
= dev_queue
->qdisc
;
596 dev_queue
->qdisc
= qdisc_default
;
599 skb
= dev_queue
->gso_skb
;
600 dev_queue
->gso_skb
= NULL
;
602 spin_unlock_bh(&dev_queue
->lock
);
607 void dev_deactivate(struct net_device
*dev
)
611 dev_deactivate_queue(&dev
->tx_queue
, &noop_qdisc
);
613 dev_watchdog_down(dev
);
615 /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
618 /* Wait for outstanding qdisc_run calls. */
620 while (test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
))
624 * Double-check inside queue lock to ensure that all effects
625 * of the queue run are visible when we return.
627 spin_lock_bh(&dev
->tx_queue
.lock
);
628 running
= test_bit(__LINK_STATE_QDISC_RUNNING
, &dev
->state
);
629 spin_unlock_bh(&dev
->tx_queue
.lock
);
632 * The running flag should never be set at this point because
633 * we've already set dev->qdisc to noop_qdisc *inside* the same
634 * pair of spin locks. That is, if any qdisc_run starts after
635 * our initial test it should see the noop_qdisc and then
636 * clear the RUNNING bit before dropping the queue lock. So
637 * if it is set here then we've found a bug.
639 } while (WARN_ON_ONCE(running
));
642 static void dev_init_scheduler_queue(struct net_device
*dev
,
643 struct netdev_queue
*dev_queue
,
646 dev_queue
->qdisc
= qdisc
;
647 dev_queue
->qdisc_sleeping
= qdisc
;
648 INIT_LIST_HEAD(&dev_queue
->qdisc_list
);
651 void dev_init_scheduler(struct net_device
*dev
)
653 qdisc_lock_tree(dev
);
654 dev_init_scheduler_queue(dev
, &dev
->tx_queue
, &noop_qdisc
);
655 dev_init_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
656 qdisc_unlock_tree(dev
);
658 setup_timer(&dev
->watchdog_timer
, dev_watchdog
, (unsigned long)dev
);
661 static void dev_shutdown_scheduler_queue(struct net_device
*dev
,
662 struct netdev_queue
*dev_queue
,
663 struct Qdisc
*qdisc_default
)
665 struct Qdisc
*qdisc
= dev_queue
->qdisc_sleeping
;
668 dev_queue
->qdisc
= qdisc_default
;
669 dev_queue
->qdisc_sleeping
= qdisc_default
;
671 qdisc_destroy(qdisc
);
675 void dev_shutdown(struct net_device
*dev
)
677 qdisc_lock_tree(dev
);
678 dev_shutdown_scheduler_queue(dev
, &dev
->tx_queue
, &noop_qdisc
);
679 dev_shutdown_scheduler_queue(dev
, &dev
->rx_queue
, NULL
);
680 BUG_TRAP(!timer_pending(&dev
->watchdog_timer
));
681 qdisc_unlock_tree(dev
);