Loading...
1/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
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.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
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 <net/pkt_sched.h>
29#include <net/dst.h>
30
31/* Main transmission queue. */
32
33/* Modifications to data participating in scheduling must be protected with
34 * qdisc_lock(qdisc) spinlock.
35 *
36 * The idea is the following:
37 * - enqueue, dequeue are serialized via qdisc root lock
38 * - ingress filtering is also serialized via qdisc root lock
39 * - updates to tree and tree walking are only done under the rtnl mutex.
40 */
41
42static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
43{
44 skb_dst_force(skb);
45 q->gso_skb = skb;
46 q->qstats.requeues++;
47 q->q.qlen++; /* it's still part of the queue */
48 __netif_schedule(q);
49
50 return 0;
51}
52
53static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
54{
55 struct sk_buff *skb = q->gso_skb;
56
57 if (unlikely(skb)) {
58 struct net_device *dev = qdisc_dev(q);
59 struct netdev_queue *txq;
60
61 /* check the reason of requeuing without tx lock first */
62 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
63 if (!netif_tx_queue_frozen_or_stopped(txq)) {
64 q->gso_skb = NULL;
65 q->q.qlen--;
66 } else
67 skb = NULL;
68 } else {
69 skb = q->dequeue(q);
70 }
71
72 return skb;
73}
74
75static inline int handle_dev_cpu_collision(struct sk_buff *skb,
76 struct netdev_queue *dev_queue,
77 struct Qdisc *q)
78{
79 int ret;
80
81 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
82 /*
83 * Same CPU holding the lock. It may be a transient
84 * configuration error, when hard_start_xmit() recurses. We
85 * detect it by checking xmit owner and drop the packet when
86 * deadloop is detected. Return OK to try the next skb.
87 */
88 kfree_skb(skb);
89 if (net_ratelimit())
90 pr_warning("Dead loop on netdevice %s, fix it urgently!\n",
91 dev_queue->dev->name);
92 ret = qdisc_qlen(q);
93 } else {
94 /*
95 * Another cpu is holding lock, requeue & delay xmits for
96 * some time.
97 */
98 __this_cpu_inc(softnet_data.cpu_collision);
99 ret = dev_requeue_skb(skb, q);
100 }
101
102 return ret;
103}
104
105/*
106 * Transmit one skb, and handle the return status as required. Holding the
107 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
108 * function.
109 *
110 * Returns to the caller:
111 * 0 - queue is empty or throttled.
112 * >0 - queue is not empty.
113 */
114int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
115 struct net_device *dev, struct netdev_queue *txq,
116 spinlock_t *root_lock)
117{
118 int ret = NETDEV_TX_BUSY;
119
120 /* And release qdisc */
121 spin_unlock(root_lock);
122
123 HARD_TX_LOCK(dev, txq, smp_processor_id());
124 if (!netif_tx_queue_frozen_or_stopped(txq))
125 ret = dev_hard_start_xmit(skb, dev, txq);
126
127 HARD_TX_UNLOCK(dev, txq);
128
129 spin_lock(root_lock);
130
131 if (dev_xmit_complete(ret)) {
132 /* Driver sent out skb successfully or skb was consumed */
133 ret = qdisc_qlen(q);
134 } else if (ret == NETDEV_TX_LOCKED) {
135 /* Driver try lock failed */
136 ret = handle_dev_cpu_collision(skb, txq, q);
137 } else {
138 /* Driver returned NETDEV_TX_BUSY - requeue skb */
139 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
140 pr_warning("BUG %s code %d qlen %d\n",
141 dev->name, ret, q->q.qlen);
142
143 ret = dev_requeue_skb(skb, q);
144 }
145
146 if (ret && netif_tx_queue_frozen_or_stopped(txq))
147 ret = 0;
148
149 return ret;
150}
151
152/*
153 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
154 *
155 * __QDISC_STATE_RUNNING guarantees only one CPU can process
156 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
157 * this queue.
158 *
159 * netif_tx_lock serializes accesses to device driver.
160 *
161 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
162 * if one is grabbed, another must be free.
163 *
164 * Note, that this procedure can be called by a watchdog timer
165 *
166 * Returns to the caller:
167 * 0 - queue is empty or throttled.
168 * >0 - queue is not empty.
169 *
170 */
171static inline int qdisc_restart(struct Qdisc *q)
172{
173 struct netdev_queue *txq;
174 struct net_device *dev;
175 spinlock_t *root_lock;
176 struct sk_buff *skb;
177
178 /* Dequeue packet */
179 skb = dequeue_skb(q);
180 if (unlikely(!skb))
181 return 0;
182 WARN_ON_ONCE(skb_dst_is_noref(skb));
183 root_lock = qdisc_lock(q);
184 dev = qdisc_dev(q);
185 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
186
187 return sch_direct_xmit(skb, q, dev, txq, root_lock);
188}
189
190void __qdisc_run(struct Qdisc *q)
191{
192 int quota = weight_p;
193
194 while (qdisc_restart(q)) {
195 /*
196 * Ordered by possible occurrence: Postpone processing if
197 * 1. we've exceeded packet quota
198 * 2. another process needs the CPU;
199 */
200 if (--quota <= 0 || need_resched()) {
201 __netif_schedule(q);
202 break;
203 }
204 }
205
206 qdisc_run_end(q);
207}
208
209unsigned long dev_trans_start(struct net_device *dev)
210{
211 unsigned long val, res = dev->trans_start;
212 unsigned int i;
213
214 for (i = 0; i < dev->num_tx_queues; i++) {
215 val = netdev_get_tx_queue(dev, i)->trans_start;
216 if (val && time_after(val, res))
217 res = val;
218 }
219 dev->trans_start = res;
220 return res;
221}
222EXPORT_SYMBOL(dev_trans_start);
223
224static void dev_watchdog(unsigned long arg)
225{
226 struct net_device *dev = (struct net_device *)arg;
227
228 netif_tx_lock(dev);
229 if (!qdisc_tx_is_noop(dev)) {
230 if (netif_device_present(dev) &&
231 netif_running(dev) &&
232 netif_carrier_ok(dev)) {
233 int some_queue_timedout = 0;
234 unsigned int i;
235 unsigned long trans_start;
236
237 for (i = 0; i < dev->num_tx_queues; i++) {
238 struct netdev_queue *txq;
239
240 txq = netdev_get_tx_queue(dev, i);
241 /*
242 * old device drivers set dev->trans_start
243 */
244 trans_start = txq->trans_start ? : dev->trans_start;
245 if (netif_tx_queue_stopped(txq) &&
246 time_after(jiffies, (trans_start +
247 dev->watchdog_timeo))) {
248 some_queue_timedout = 1;
249 break;
250 }
251 }
252
253 if (some_queue_timedout) {
254 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
255 dev->name, netdev_drivername(dev), i);
256 dev->netdev_ops->ndo_tx_timeout(dev);
257 }
258 if (!mod_timer(&dev->watchdog_timer,
259 round_jiffies(jiffies +
260 dev->watchdog_timeo)))
261 dev_hold(dev);
262 }
263 }
264 netif_tx_unlock(dev);
265
266 dev_put(dev);
267}
268
269void __netdev_watchdog_up(struct net_device *dev)
270{
271 if (dev->netdev_ops->ndo_tx_timeout) {
272 if (dev->watchdog_timeo <= 0)
273 dev->watchdog_timeo = 5*HZ;
274 if (!mod_timer(&dev->watchdog_timer,
275 round_jiffies(jiffies + dev->watchdog_timeo)))
276 dev_hold(dev);
277 }
278}
279
280static void dev_watchdog_up(struct net_device *dev)
281{
282 __netdev_watchdog_up(dev);
283}
284
285static void dev_watchdog_down(struct net_device *dev)
286{
287 netif_tx_lock_bh(dev);
288 if (del_timer(&dev->watchdog_timer))
289 dev_put(dev);
290 netif_tx_unlock_bh(dev);
291}
292
293/**
294 * netif_carrier_on - set carrier
295 * @dev: network device
296 *
297 * Device has detected that carrier.
298 */
299void netif_carrier_on(struct net_device *dev)
300{
301 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
302 if (dev->reg_state == NETREG_UNINITIALIZED)
303 return;
304 linkwatch_fire_event(dev);
305 if (netif_running(dev))
306 __netdev_watchdog_up(dev);
307 }
308}
309EXPORT_SYMBOL(netif_carrier_on);
310
311/**
312 * netif_carrier_off - clear carrier
313 * @dev: network device
314 *
315 * Device has detected loss of carrier.
316 */
317void netif_carrier_off(struct net_device *dev)
318{
319 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
320 if (dev->reg_state == NETREG_UNINITIALIZED)
321 return;
322 linkwatch_fire_event(dev);
323 }
324}
325EXPORT_SYMBOL(netif_carrier_off);
326
327/**
328 * netif_notify_peers - notify network peers about existence of @dev
329 * @dev: network device
330 *
331 * Generate traffic such that interested network peers are aware of
332 * @dev, such as by generating a gratuitous ARP. This may be used when
333 * a device wants to inform the rest of the network about some sort of
334 * reconfiguration such as a failover event or virtual machine
335 * migration.
336 */
337void netif_notify_peers(struct net_device *dev)
338{
339 rtnl_lock();
340 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
341 rtnl_unlock();
342}
343EXPORT_SYMBOL(netif_notify_peers);
344
345/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
346 under all circumstances. It is difficult to invent anything faster or
347 cheaper.
348 */
349
350static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
351{
352 kfree_skb(skb);
353 return NET_XMIT_CN;
354}
355
356static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
357{
358 return NULL;
359}
360
361struct Qdisc_ops noop_qdisc_ops __read_mostly = {
362 .id = "noop",
363 .priv_size = 0,
364 .enqueue = noop_enqueue,
365 .dequeue = noop_dequeue,
366 .peek = noop_dequeue,
367 .owner = THIS_MODULE,
368};
369
370static struct netdev_queue noop_netdev_queue = {
371 .qdisc = &noop_qdisc,
372 .qdisc_sleeping = &noop_qdisc,
373};
374
375struct Qdisc noop_qdisc = {
376 .enqueue = noop_enqueue,
377 .dequeue = noop_dequeue,
378 .flags = TCQ_F_BUILTIN,
379 .ops = &noop_qdisc_ops,
380 .list = LIST_HEAD_INIT(noop_qdisc.list),
381 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
382 .dev_queue = &noop_netdev_queue,
383 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
384};
385EXPORT_SYMBOL(noop_qdisc);
386
387static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
388 .id = "noqueue",
389 .priv_size = 0,
390 .enqueue = noop_enqueue,
391 .dequeue = noop_dequeue,
392 .peek = noop_dequeue,
393 .owner = THIS_MODULE,
394};
395
396static struct Qdisc noqueue_qdisc;
397static struct netdev_queue noqueue_netdev_queue = {
398 .qdisc = &noqueue_qdisc,
399 .qdisc_sleeping = &noqueue_qdisc,
400};
401
402static struct Qdisc noqueue_qdisc = {
403 .enqueue = NULL,
404 .dequeue = noop_dequeue,
405 .flags = TCQ_F_BUILTIN,
406 .ops = &noqueue_qdisc_ops,
407 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
408 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
409 .dev_queue = &noqueue_netdev_queue,
410 .busylock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
411};
412
413
414static const u8 prio2band[TC_PRIO_MAX + 1] = {
415 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
416};
417
418/* 3-band FIFO queue: old style, but should be a bit faster than
419 generic prio+fifo combination.
420 */
421
422#define PFIFO_FAST_BANDS 3
423
424/*
425 * Private data for a pfifo_fast scheduler containing:
426 * - queues for the three band
427 * - bitmap indicating which of the bands contain skbs
428 */
429struct pfifo_fast_priv {
430 u32 bitmap;
431 struct sk_buff_head q[PFIFO_FAST_BANDS];
432};
433
434/*
435 * Convert a bitmap to the first band number where an skb is queued, where:
436 * bitmap=0 means there are no skbs on any band.
437 * bitmap=1 means there is an skb on band 0.
438 * bitmap=7 means there are skbs on all 3 bands, etc.
439 */
440static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
441
442static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
443 int band)
444{
445 return priv->q + band;
446}
447
448static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
449{
450 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
451 int band = prio2band[skb->priority & TC_PRIO_MAX];
452 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
453 struct sk_buff_head *list = band2list(priv, band);
454
455 priv->bitmap |= (1 << band);
456 qdisc->q.qlen++;
457 return __qdisc_enqueue_tail(skb, qdisc, list);
458 }
459
460 return qdisc_drop(skb, qdisc);
461}
462
463static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
464{
465 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
466 int band = bitmap2band[priv->bitmap];
467
468 if (likely(band >= 0)) {
469 struct sk_buff_head *list = band2list(priv, band);
470 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
471
472 qdisc->q.qlen--;
473 if (skb_queue_empty(list))
474 priv->bitmap &= ~(1 << band);
475
476 return skb;
477 }
478
479 return NULL;
480}
481
482static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
483{
484 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
485 int band = bitmap2band[priv->bitmap];
486
487 if (band >= 0) {
488 struct sk_buff_head *list = band2list(priv, band);
489
490 return skb_peek(list);
491 }
492
493 return NULL;
494}
495
496static void pfifo_fast_reset(struct Qdisc *qdisc)
497{
498 int prio;
499 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
500
501 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
502 __qdisc_reset_queue(qdisc, band2list(priv, prio));
503
504 priv->bitmap = 0;
505 qdisc->qstats.backlog = 0;
506 qdisc->q.qlen = 0;
507}
508
509static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
510{
511 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
512
513 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
514 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
515 return skb->len;
516
517nla_put_failure:
518 return -1;
519}
520
521static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
522{
523 int prio;
524 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
525
526 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
527 skb_queue_head_init(band2list(priv, prio));
528
529 /* Can by-pass the queue discipline */
530 qdisc->flags |= TCQ_F_CAN_BYPASS;
531 return 0;
532}
533
534struct Qdisc_ops pfifo_fast_ops __read_mostly = {
535 .id = "pfifo_fast",
536 .priv_size = sizeof(struct pfifo_fast_priv),
537 .enqueue = pfifo_fast_enqueue,
538 .dequeue = pfifo_fast_dequeue,
539 .peek = pfifo_fast_peek,
540 .init = pfifo_fast_init,
541 .reset = pfifo_fast_reset,
542 .dump = pfifo_fast_dump,
543 .owner = THIS_MODULE,
544};
545EXPORT_SYMBOL(pfifo_fast_ops);
546
547struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
548 struct Qdisc_ops *ops)
549{
550 void *p;
551 struct Qdisc *sch;
552 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
553 int err = -ENOBUFS;
554
555 p = kzalloc_node(size, GFP_KERNEL,
556 netdev_queue_numa_node_read(dev_queue));
557
558 if (!p)
559 goto errout;
560 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
561 /* if we got non aligned memory, ask more and do alignment ourself */
562 if (sch != p) {
563 kfree(p);
564 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
565 netdev_queue_numa_node_read(dev_queue));
566 if (!p)
567 goto errout;
568 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
569 sch->padded = (char *) sch - (char *) p;
570 }
571 INIT_LIST_HEAD(&sch->list);
572 skb_queue_head_init(&sch->q);
573 spin_lock_init(&sch->busylock);
574 sch->ops = ops;
575 sch->enqueue = ops->enqueue;
576 sch->dequeue = ops->dequeue;
577 sch->dev_queue = dev_queue;
578 dev_hold(qdisc_dev(sch));
579 atomic_set(&sch->refcnt, 1);
580
581 return sch;
582errout:
583 return ERR_PTR(err);
584}
585
586struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
587 struct Qdisc_ops *ops, unsigned int parentid)
588{
589 struct Qdisc *sch;
590
591 sch = qdisc_alloc(dev_queue, ops);
592 if (IS_ERR(sch))
593 goto errout;
594 sch->parent = parentid;
595
596 if (!ops->init || ops->init(sch, NULL) == 0)
597 return sch;
598
599 qdisc_destroy(sch);
600errout:
601 return NULL;
602}
603EXPORT_SYMBOL(qdisc_create_dflt);
604
605/* Under qdisc_lock(qdisc) and BH! */
606
607void qdisc_reset(struct Qdisc *qdisc)
608{
609 const struct Qdisc_ops *ops = qdisc->ops;
610
611 if (ops->reset)
612 ops->reset(qdisc);
613
614 if (qdisc->gso_skb) {
615 kfree_skb(qdisc->gso_skb);
616 qdisc->gso_skb = NULL;
617 qdisc->q.qlen = 0;
618 }
619}
620EXPORT_SYMBOL(qdisc_reset);
621
622static void qdisc_rcu_free(struct rcu_head *head)
623{
624 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
625
626 kfree((char *) qdisc - qdisc->padded);
627}
628
629void qdisc_destroy(struct Qdisc *qdisc)
630{
631 const struct Qdisc_ops *ops = qdisc->ops;
632
633 if (qdisc->flags & TCQ_F_BUILTIN ||
634 !atomic_dec_and_test(&qdisc->refcnt))
635 return;
636
637#ifdef CONFIG_NET_SCHED
638 qdisc_list_del(qdisc);
639
640 qdisc_put_stab(rtnl_dereference(qdisc->stab));
641#endif
642 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
643 if (ops->reset)
644 ops->reset(qdisc);
645 if (ops->destroy)
646 ops->destroy(qdisc);
647
648 module_put(ops->owner);
649 dev_put(qdisc_dev(qdisc));
650
651 kfree_skb(qdisc->gso_skb);
652 /*
653 * gen_estimator est_timer() might access qdisc->q.lock,
654 * wait a RCU grace period before freeing qdisc.
655 */
656 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
657}
658EXPORT_SYMBOL(qdisc_destroy);
659
660/* Attach toplevel qdisc to device queue. */
661struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
662 struct Qdisc *qdisc)
663{
664 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
665 spinlock_t *root_lock;
666
667 root_lock = qdisc_lock(oqdisc);
668 spin_lock_bh(root_lock);
669
670 /* Prune old scheduler */
671 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
672 qdisc_reset(oqdisc);
673
674 /* ... and graft new one */
675 if (qdisc == NULL)
676 qdisc = &noop_qdisc;
677 dev_queue->qdisc_sleeping = qdisc;
678 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
679
680 spin_unlock_bh(root_lock);
681
682 return oqdisc;
683}
684EXPORT_SYMBOL(dev_graft_qdisc);
685
686static void attach_one_default_qdisc(struct net_device *dev,
687 struct netdev_queue *dev_queue,
688 void *_unused)
689{
690 struct Qdisc *qdisc = &noqueue_qdisc;
691
692 if (dev->tx_queue_len) {
693 qdisc = qdisc_create_dflt(dev_queue,
694 &pfifo_fast_ops, TC_H_ROOT);
695 if (!qdisc) {
696 netdev_info(dev, "activation failed\n");
697 return;
698 }
699 }
700 dev_queue->qdisc_sleeping = qdisc;
701}
702
703static void attach_default_qdiscs(struct net_device *dev)
704{
705 struct netdev_queue *txq;
706 struct Qdisc *qdisc;
707
708 txq = netdev_get_tx_queue(dev, 0);
709
710 if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
711 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
712 dev->qdisc = txq->qdisc_sleeping;
713 atomic_inc(&dev->qdisc->refcnt);
714 } else {
715 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
716 if (qdisc) {
717 qdisc->ops->attach(qdisc);
718 dev->qdisc = qdisc;
719 }
720 }
721}
722
723static void transition_one_qdisc(struct net_device *dev,
724 struct netdev_queue *dev_queue,
725 void *_need_watchdog)
726{
727 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
728 int *need_watchdog_p = _need_watchdog;
729
730 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
731 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
732
733 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
734 if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
735 dev_queue->trans_start = 0;
736 *need_watchdog_p = 1;
737 }
738}
739
740void dev_activate(struct net_device *dev)
741{
742 int need_watchdog;
743
744 /* No queueing discipline is attached to device;
745 create default one i.e. pfifo_fast for devices,
746 which need queueing and noqueue_qdisc for
747 virtual interfaces
748 */
749
750 if (dev->qdisc == &noop_qdisc)
751 attach_default_qdiscs(dev);
752
753 if (!netif_carrier_ok(dev))
754 /* Delay activation until next carrier-on event */
755 return;
756
757 need_watchdog = 0;
758 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
759 if (dev_ingress_queue(dev))
760 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
761
762 if (need_watchdog) {
763 dev->trans_start = jiffies;
764 dev_watchdog_up(dev);
765 }
766}
767EXPORT_SYMBOL(dev_activate);
768
769static void dev_deactivate_queue(struct net_device *dev,
770 struct netdev_queue *dev_queue,
771 void *_qdisc_default)
772{
773 struct Qdisc *qdisc_default = _qdisc_default;
774 struct Qdisc *qdisc;
775
776 qdisc = dev_queue->qdisc;
777 if (qdisc) {
778 spin_lock_bh(qdisc_lock(qdisc));
779
780 if (!(qdisc->flags & TCQ_F_BUILTIN))
781 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
782
783 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
784 qdisc_reset(qdisc);
785
786 spin_unlock_bh(qdisc_lock(qdisc));
787 }
788}
789
790static bool some_qdisc_is_busy(struct net_device *dev)
791{
792 unsigned int i;
793
794 for (i = 0; i < dev->num_tx_queues; i++) {
795 struct netdev_queue *dev_queue;
796 spinlock_t *root_lock;
797 struct Qdisc *q;
798 int val;
799
800 dev_queue = netdev_get_tx_queue(dev, i);
801 q = dev_queue->qdisc_sleeping;
802 root_lock = qdisc_lock(q);
803
804 spin_lock_bh(root_lock);
805
806 val = (qdisc_is_running(q) ||
807 test_bit(__QDISC_STATE_SCHED, &q->state));
808
809 spin_unlock_bh(root_lock);
810
811 if (val)
812 return true;
813 }
814 return false;
815}
816
817/**
818 * dev_deactivate_many - deactivate transmissions on several devices
819 * @head: list of devices to deactivate
820 *
821 * This function returns only when all outstanding transmissions
822 * have completed, unless all devices are in dismantle phase.
823 */
824void dev_deactivate_many(struct list_head *head)
825{
826 struct net_device *dev;
827 bool sync_needed = false;
828
829 list_for_each_entry(dev, head, unreg_list) {
830 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
831 &noop_qdisc);
832 if (dev_ingress_queue(dev))
833 dev_deactivate_queue(dev, dev_ingress_queue(dev),
834 &noop_qdisc);
835
836 dev_watchdog_down(dev);
837 sync_needed |= !dev->dismantle;
838 }
839
840 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
841 * This is avoided if all devices are in dismantle phase :
842 * Caller will call synchronize_net() for us
843 */
844 if (sync_needed)
845 synchronize_net();
846
847 /* Wait for outstanding qdisc_run calls. */
848 list_for_each_entry(dev, head, unreg_list)
849 while (some_qdisc_is_busy(dev))
850 yield();
851}
852
853void dev_deactivate(struct net_device *dev)
854{
855 LIST_HEAD(single);
856
857 list_add(&dev->unreg_list, &single);
858 dev_deactivate_many(&single);
859 list_del(&single);
860}
861EXPORT_SYMBOL(dev_deactivate);
862
863static void dev_init_scheduler_queue(struct net_device *dev,
864 struct netdev_queue *dev_queue,
865 void *_qdisc)
866{
867 struct Qdisc *qdisc = _qdisc;
868
869 dev_queue->qdisc = qdisc;
870 dev_queue->qdisc_sleeping = qdisc;
871}
872
873void dev_init_scheduler(struct net_device *dev)
874{
875 dev->qdisc = &noop_qdisc;
876 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
877 if (dev_ingress_queue(dev))
878 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
879
880 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
881}
882
883static void shutdown_scheduler_queue(struct net_device *dev,
884 struct netdev_queue *dev_queue,
885 void *_qdisc_default)
886{
887 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
888 struct Qdisc *qdisc_default = _qdisc_default;
889
890 if (qdisc) {
891 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
892 dev_queue->qdisc_sleeping = qdisc_default;
893
894 qdisc_destroy(qdisc);
895 }
896}
897
898void dev_shutdown(struct net_device *dev)
899{
900 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
901 if (dev_ingress_queue(dev))
902 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
903 qdisc_destroy(dev->qdisc);
904 dev->qdisc = &noop_qdisc;
905
906 WARN_ON(timer_pending(&dev->watchdog_timer));
907}
1/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
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.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
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 <net/sch_generic.h>
30#include <net/pkt_sched.h>
31#include <net/dst.h>
32
33/* Qdisc to use by default */
34const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35EXPORT_SYMBOL(default_qdisc_ops);
36
37/* Main transmission queue. */
38
39/* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
41 *
42 * The idea is the following:
43 * - enqueue, dequeue are serialized via qdisc root lock
44 * - ingress filtering is also serialized via qdisc root lock
45 * - updates to tree and tree walking are only done under the rtnl mutex.
46 */
47
48static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
49{
50 skb_dst_force(skb);
51 q->gso_skb = skb;
52 q->qstats.requeues++;
53 q->q.qlen++; /* it's still part of the queue */
54 __netif_schedule(q);
55
56 return 0;
57}
58
59static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
60{
61 struct sk_buff *skb = q->gso_skb;
62 const struct netdev_queue *txq = q->dev_queue;
63
64 if (unlikely(skb)) {
65 /* check the reason of requeuing without tx lock first */
66 txq = netdev_get_tx_queue(txq->dev, skb_get_queue_mapping(skb));
67 if (!netif_xmit_frozen_or_stopped(txq)) {
68 q->gso_skb = NULL;
69 q->q.qlen--;
70 } else
71 skb = NULL;
72 } else {
73 if (!(q->flags & TCQ_F_ONETXQUEUE) || !netif_xmit_frozen_or_stopped(txq))
74 skb = q->dequeue(q);
75 }
76
77 return skb;
78}
79
80static inline int handle_dev_cpu_collision(struct sk_buff *skb,
81 struct netdev_queue *dev_queue,
82 struct Qdisc *q)
83{
84 int ret;
85
86 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
87 /*
88 * Same CPU holding the lock. It may be a transient
89 * configuration error, when hard_start_xmit() recurses. We
90 * detect it by checking xmit owner and drop the packet when
91 * deadloop is detected. Return OK to try the next skb.
92 */
93 kfree_skb(skb);
94 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
95 dev_queue->dev->name);
96 ret = qdisc_qlen(q);
97 } else {
98 /*
99 * Another cpu is holding lock, requeue & delay xmits for
100 * some time.
101 */
102 __this_cpu_inc(softnet_data.cpu_collision);
103 ret = dev_requeue_skb(skb, q);
104 }
105
106 return ret;
107}
108
109/*
110 * Transmit one skb, and handle the return status as required. Holding the
111 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
112 * function.
113 *
114 * Returns to the caller:
115 * 0 - queue is empty or throttled.
116 * >0 - queue is not empty.
117 */
118int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
119 struct net_device *dev, struct netdev_queue *txq,
120 spinlock_t *root_lock)
121{
122 int ret = NETDEV_TX_BUSY;
123
124 /* And release qdisc */
125 spin_unlock(root_lock);
126
127 HARD_TX_LOCK(dev, txq, smp_processor_id());
128 if (!netif_xmit_frozen_or_stopped(txq))
129 ret = dev_hard_start_xmit(skb, dev, txq);
130
131 HARD_TX_UNLOCK(dev, txq);
132
133 spin_lock(root_lock);
134
135 if (dev_xmit_complete(ret)) {
136 /* Driver sent out skb successfully or skb was consumed */
137 ret = qdisc_qlen(q);
138 } else if (ret == NETDEV_TX_LOCKED) {
139 /* Driver try lock failed */
140 ret = handle_dev_cpu_collision(skb, txq, q);
141 } else {
142 /* Driver returned NETDEV_TX_BUSY - requeue skb */
143 if (unlikely(ret != NETDEV_TX_BUSY))
144 net_warn_ratelimited("BUG %s code %d qlen %d\n",
145 dev->name, ret, q->q.qlen);
146
147 ret = dev_requeue_skb(skb, q);
148 }
149
150 if (ret && netif_xmit_frozen_or_stopped(txq))
151 ret = 0;
152
153 return ret;
154}
155
156/*
157 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
158 *
159 * __QDISC_STATE_RUNNING guarantees only one CPU can process
160 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
161 * this queue.
162 *
163 * netif_tx_lock serializes accesses to device driver.
164 *
165 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
166 * if one is grabbed, another must be free.
167 *
168 * Note, that this procedure can be called by a watchdog timer
169 *
170 * Returns to the caller:
171 * 0 - queue is empty or throttled.
172 * >0 - queue is not empty.
173 *
174 */
175static inline int qdisc_restart(struct Qdisc *q)
176{
177 struct netdev_queue *txq;
178 struct net_device *dev;
179 spinlock_t *root_lock;
180 struct sk_buff *skb;
181
182 /* Dequeue packet */
183 skb = dequeue_skb(q);
184 if (unlikely(!skb))
185 return 0;
186 WARN_ON_ONCE(skb_dst_is_noref(skb));
187 root_lock = qdisc_lock(q);
188 dev = qdisc_dev(q);
189 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
190
191 return sch_direct_xmit(skb, q, dev, txq, root_lock);
192}
193
194void __qdisc_run(struct Qdisc *q)
195{
196 int quota = weight_p;
197
198 while (qdisc_restart(q)) {
199 /*
200 * Ordered by possible occurrence: Postpone processing if
201 * 1. we've exceeded packet quota
202 * 2. another process needs the CPU;
203 */
204 if (--quota <= 0 || need_resched()) {
205 __netif_schedule(q);
206 break;
207 }
208 }
209
210 qdisc_run_end(q);
211}
212
213unsigned long dev_trans_start(struct net_device *dev)
214{
215 unsigned long val, res;
216 unsigned int i;
217
218 if (is_vlan_dev(dev))
219 dev = vlan_dev_real_dev(dev);
220 res = dev->trans_start;
221 for (i = 0; i < dev->num_tx_queues; i++) {
222 val = netdev_get_tx_queue(dev, i)->trans_start;
223 if (val && time_after(val, res))
224 res = val;
225 }
226 dev->trans_start = res;
227
228 return res;
229}
230EXPORT_SYMBOL(dev_trans_start);
231
232static void dev_watchdog(unsigned long arg)
233{
234 struct net_device *dev = (struct net_device *)arg;
235
236 netif_tx_lock(dev);
237 if (!qdisc_tx_is_noop(dev)) {
238 if (netif_device_present(dev) &&
239 netif_running(dev) &&
240 netif_carrier_ok(dev)) {
241 int some_queue_timedout = 0;
242 unsigned int i;
243 unsigned long trans_start;
244
245 for (i = 0; i < dev->num_tx_queues; i++) {
246 struct netdev_queue *txq;
247
248 txq = netdev_get_tx_queue(dev, i);
249 /*
250 * old device drivers set dev->trans_start
251 */
252 trans_start = txq->trans_start ? : dev->trans_start;
253 if (netif_xmit_stopped(txq) &&
254 time_after(jiffies, (trans_start +
255 dev->watchdog_timeo))) {
256 some_queue_timedout = 1;
257 txq->trans_timeout++;
258 break;
259 }
260 }
261
262 if (some_queue_timedout) {
263 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
264 dev->name, netdev_drivername(dev), i);
265 dev->netdev_ops->ndo_tx_timeout(dev);
266 }
267 if (!mod_timer(&dev->watchdog_timer,
268 round_jiffies(jiffies +
269 dev->watchdog_timeo)))
270 dev_hold(dev);
271 }
272 }
273 netif_tx_unlock(dev);
274
275 dev_put(dev);
276}
277
278void __netdev_watchdog_up(struct net_device *dev)
279{
280 if (dev->netdev_ops->ndo_tx_timeout) {
281 if (dev->watchdog_timeo <= 0)
282 dev->watchdog_timeo = 5*HZ;
283 if (!mod_timer(&dev->watchdog_timer,
284 round_jiffies(jiffies + dev->watchdog_timeo)))
285 dev_hold(dev);
286 }
287}
288
289static void dev_watchdog_up(struct net_device *dev)
290{
291 __netdev_watchdog_up(dev);
292}
293
294static void dev_watchdog_down(struct net_device *dev)
295{
296 netif_tx_lock_bh(dev);
297 if (del_timer(&dev->watchdog_timer))
298 dev_put(dev);
299 netif_tx_unlock_bh(dev);
300}
301
302/**
303 * netif_carrier_on - set carrier
304 * @dev: network device
305 *
306 * Device has detected that carrier.
307 */
308void netif_carrier_on(struct net_device *dev)
309{
310 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
311 if (dev->reg_state == NETREG_UNINITIALIZED)
312 return;
313 atomic_inc(&dev->carrier_changes);
314 linkwatch_fire_event(dev);
315 if (netif_running(dev))
316 __netdev_watchdog_up(dev);
317 }
318}
319EXPORT_SYMBOL(netif_carrier_on);
320
321/**
322 * netif_carrier_off - clear carrier
323 * @dev: network device
324 *
325 * Device has detected loss of carrier.
326 */
327void netif_carrier_off(struct net_device *dev)
328{
329 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
330 if (dev->reg_state == NETREG_UNINITIALIZED)
331 return;
332 atomic_inc(&dev->carrier_changes);
333 linkwatch_fire_event(dev);
334 }
335}
336EXPORT_SYMBOL(netif_carrier_off);
337
338/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
339 under all circumstances. It is difficult to invent anything faster or
340 cheaper.
341 */
342
343static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
344{
345 kfree_skb(skb);
346 return NET_XMIT_CN;
347}
348
349static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
350{
351 return NULL;
352}
353
354struct Qdisc_ops noop_qdisc_ops __read_mostly = {
355 .id = "noop",
356 .priv_size = 0,
357 .enqueue = noop_enqueue,
358 .dequeue = noop_dequeue,
359 .peek = noop_dequeue,
360 .owner = THIS_MODULE,
361};
362
363static struct netdev_queue noop_netdev_queue = {
364 .qdisc = &noop_qdisc,
365 .qdisc_sleeping = &noop_qdisc,
366};
367
368struct Qdisc noop_qdisc = {
369 .enqueue = noop_enqueue,
370 .dequeue = noop_dequeue,
371 .flags = TCQ_F_BUILTIN,
372 .ops = &noop_qdisc_ops,
373 .list = LIST_HEAD_INIT(noop_qdisc.list),
374 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
375 .dev_queue = &noop_netdev_queue,
376 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
377};
378EXPORT_SYMBOL(noop_qdisc);
379
380static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
381 .id = "noqueue",
382 .priv_size = 0,
383 .enqueue = noop_enqueue,
384 .dequeue = noop_dequeue,
385 .peek = noop_dequeue,
386 .owner = THIS_MODULE,
387};
388
389static struct Qdisc noqueue_qdisc;
390static struct netdev_queue noqueue_netdev_queue = {
391 .qdisc = &noqueue_qdisc,
392 .qdisc_sleeping = &noqueue_qdisc,
393};
394
395static struct Qdisc noqueue_qdisc = {
396 .enqueue = NULL,
397 .dequeue = noop_dequeue,
398 .flags = TCQ_F_BUILTIN,
399 .ops = &noqueue_qdisc_ops,
400 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
401 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
402 .dev_queue = &noqueue_netdev_queue,
403 .busylock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
404};
405
406
407static const u8 prio2band[TC_PRIO_MAX + 1] = {
408 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
409};
410
411/* 3-band FIFO queue: old style, but should be a bit faster than
412 generic prio+fifo combination.
413 */
414
415#define PFIFO_FAST_BANDS 3
416
417/*
418 * Private data for a pfifo_fast scheduler containing:
419 * - queues for the three band
420 * - bitmap indicating which of the bands contain skbs
421 */
422struct pfifo_fast_priv {
423 u32 bitmap;
424 struct sk_buff_head q[PFIFO_FAST_BANDS];
425};
426
427/*
428 * Convert a bitmap to the first band number where an skb is queued, where:
429 * bitmap=0 means there are no skbs on any band.
430 * bitmap=1 means there is an skb on band 0.
431 * bitmap=7 means there are skbs on all 3 bands, etc.
432 */
433static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
434
435static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
436 int band)
437{
438 return priv->q + band;
439}
440
441static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
442{
443 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
444 int band = prio2band[skb->priority & TC_PRIO_MAX];
445 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
446 struct sk_buff_head *list = band2list(priv, band);
447
448 priv->bitmap |= (1 << band);
449 qdisc->q.qlen++;
450 return __qdisc_enqueue_tail(skb, qdisc, list);
451 }
452
453 return qdisc_drop(skb, qdisc);
454}
455
456static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
457{
458 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
459 int band = bitmap2band[priv->bitmap];
460
461 if (likely(band >= 0)) {
462 struct sk_buff_head *list = band2list(priv, band);
463 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
464
465 qdisc->q.qlen--;
466 if (skb_queue_empty(list))
467 priv->bitmap &= ~(1 << band);
468
469 return skb;
470 }
471
472 return NULL;
473}
474
475static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
476{
477 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
478 int band = bitmap2band[priv->bitmap];
479
480 if (band >= 0) {
481 struct sk_buff_head *list = band2list(priv, band);
482
483 return skb_peek(list);
484 }
485
486 return NULL;
487}
488
489static void pfifo_fast_reset(struct Qdisc *qdisc)
490{
491 int prio;
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493
494 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
495 __qdisc_reset_queue(qdisc, band2list(priv, prio));
496
497 priv->bitmap = 0;
498 qdisc->qstats.backlog = 0;
499 qdisc->q.qlen = 0;
500}
501
502static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
503{
504 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
505
506 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
507 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
508 goto nla_put_failure;
509 return skb->len;
510
511nla_put_failure:
512 return -1;
513}
514
515static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
516{
517 int prio;
518 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
519
520 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
521 skb_queue_head_init(band2list(priv, prio));
522
523 /* Can by-pass the queue discipline */
524 qdisc->flags |= TCQ_F_CAN_BYPASS;
525 return 0;
526}
527
528struct Qdisc_ops pfifo_fast_ops __read_mostly = {
529 .id = "pfifo_fast",
530 .priv_size = sizeof(struct pfifo_fast_priv),
531 .enqueue = pfifo_fast_enqueue,
532 .dequeue = pfifo_fast_dequeue,
533 .peek = pfifo_fast_peek,
534 .init = pfifo_fast_init,
535 .reset = pfifo_fast_reset,
536 .dump = pfifo_fast_dump,
537 .owner = THIS_MODULE,
538};
539
540static struct lock_class_key qdisc_tx_busylock;
541
542struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
543 const struct Qdisc_ops *ops)
544{
545 void *p;
546 struct Qdisc *sch;
547 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
548 int err = -ENOBUFS;
549 struct net_device *dev = dev_queue->dev;
550
551 p = kzalloc_node(size, GFP_KERNEL,
552 netdev_queue_numa_node_read(dev_queue));
553
554 if (!p)
555 goto errout;
556 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
557 /* if we got non aligned memory, ask more and do alignment ourself */
558 if (sch != p) {
559 kfree(p);
560 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
561 netdev_queue_numa_node_read(dev_queue));
562 if (!p)
563 goto errout;
564 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
565 sch->padded = (char *) sch - (char *) p;
566 }
567 INIT_LIST_HEAD(&sch->list);
568 skb_queue_head_init(&sch->q);
569
570 spin_lock_init(&sch->busylock);
571 lockdep_set_class(&sch->busylock,
572 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
573
574 sch->ops = ops;
575 sch->enqueue = ops->enqueue;
576 sch->dequeue = ops->dequeue;
577 sch->dev_queue = dev_queue;
578 dev_hold(dev);
579 atomic_set(&sch->refcnt, 1);
580
581 return sch;
582errout:
583 return ERR_PTR(err);
584}
585
586struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
587 const struct Qdisc_ops *ops,
588 unsigned int parentid)
589{
590 struct Qdisc *sch;
591
592 if (!try_module_get(ops->owner))
593 goto errout;
594
595 sch = qdisc_alloc(dev_queue, ops);
596 if (IS_ERR(sch))
597 goto errout;
598 sch->parent = parentid;
599
600 if (!ops->init || ops->init(sch, NULL) == 0)
601 return sch;
602
603 qdisc_destroy(sch);
604errout:
605 return NULL;
606}
607EXPORT_SYMBOL(qdisc_create_dflt);
608
609/* Under qdisc_lock(qdisc) and BH! */
610
611void qdisc_reset(struct Qdisc *qdisc)
612{
613 const struct Qdisc_ops *ops = qdisc->ops;
614
615 if (ops->reset)
616 ops->reset(qdisc);
617
618 if (qdisc->gso_skb) {
619 kfree_skb(qdisc->gso_skb);
620 qdisc->gso_skb = NULL;
621 qdisc->q.qlen = 0;
622 }
623}
624EXPORT_SYMBOL(qdisc_reset);
625
626static void qdisc_rcu_free(struct rcu_head *head)
627{
628 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
629
630 kfree((char *) qdisc - qdisc->padded);
631}
632
633void qdisc_destroy(struct Qdisc *qdisc)
634{
635 const struct Qdisc_ops *ops = qdisc->ops;
636
637 if (qdisc->flags & TCQ_F_BUILTIN ||
638 !atomic_dec_and_test(&qdisc->refcnt))
639 return;
640
641#ifdef CONFIG_NET_SCHED
642 qdisc_list_del(qdisc);
643
644 qdisc_put_stab(rtnl_dereference(qdisc->stab));
645#endif
646 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
647 if (ops->reset)
648 ops->reset(qdisc);
649 if (ops->destroy)
650 ops->destroy(qdisc);
651
652 module_put(ops->owner);
653 dev_put(qdisc_dev(qdisc));
654
655 kfree_skb(qdisc->gso_skb);
656 /*
657 * gen_estimator est_timer() might access qdisc->q.lock,
658 * wait a RCU grace period before freeing qdisc.
659 */
660 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
661}
662EXPORT_SYMBOL(qdisc_destroy);
663
664/* Attach toplevel qdisc to device queue. */
665struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
666 struct Qdisc *qdisc)
667{
668 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
669 spinlock_t *root_lock;
670
671 root_lock = qdisc_lock(oqdisc);
672 spin_lock_bh(root_lock);
673
674 /* Prune old scheduler */
675 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
676 qdisc_reset(oqdisc);
677
678 /* ... and graft new one */
679 if (qdisc == NULL)
680 qdisc = &noop_qdisc;
681 dev_queue->qdisc_sleeping = qdisc;
682 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
683
684 spin_unlock_bh(root_lock);
685
686 return oqdisc;
687}
688EXPORT_SYMBOL(dev_graft_qdisc);
689
690static void attach_one_default_qdisc(struct net_device *dev,
691 struct netdev_queue *dev_queue,
692 void *_unused)
693{
694 struct Qdisc *qdisc = &noqueue_qdisc;
695
696 if (dev->tx_queue_len) {
697 qdisc = qdisc_create_dflt(dev_queue,
698 default_qdisc_ops, TC_H_ROOT);
699 if (!qdisc) {
700 netdev_info(dev, "activation failed\n");
701 return;
702 }
703 if (!netif_is_multiqueue(dev))
704 qdisc->flags |= TCQ_F_ONETXQUEUE;
705 }
706 dev_queue->qdisc_sleeping = qdisc;
707}
708
709static void attach_default_qdiscs(struct net_device *dev)
710{
711 struct netdev_queue *txq;
712 struct Qdisc *qdisc;
713
714 txq = netdev_get_tx_queue(dev, 0);
715
716 if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
717 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
718 dev->qdisc = txq->qdisc_sleeping;
719 atomic_inc(&dev->qdisc->refcnt);
720 } else {
721 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
722 if (qdisc) {
723 dev->qdisc = qdisc;
724 qdisc->ops->attach(qdisc);
725 }
726 }
727}
728
729static void transition_one_qdisc(struct net_device *dev,
730 struct netdev_queue *dev_queue,
731 void *_need_watchdog)
732{
733 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
734 int *need_watchdog_p = _need_watchdog;
735
736 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
737 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
738
739 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
740 if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
741 dev_queue->trans_start = 0;
742 *need_watchdog_p = 1;
743 }
744}
745
746void dev_activate(struct net_device *dev)
747{
748 int need_watchdog;
749
750 /* No queueing discipline is attached to device;
751 * create default one for devices, which need queueing
752 * and noqueue_qdisc for virtual interfaces
753 */
754
755 if (dev->qdisc == &noop_qdisc)
756 attach_default_qdiscs(dev);
757
758 if (!netif_carrier_ok(dev))
759 /* Delay activation until next carrier-on event */
760 return;
761
762 need_watchdog = 0;
763 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
764 if (dev_ingress_queue(dev))
765 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
766
767 if (need_watchdog) {
768 dev->trans_start = jiffies;
769 dev_watchdog_up(dev);
770 }
771}
772EXPORT_SYMBOL(dev_activate);
773
774static void dev_deactivate_queue(struct net_device *dev,
775 struct netdev_queue *dev_queue,
776 void *_qdisc_default)
777{
778 struct Qdisc *qdisc_default = _qdisc_default;
779 struct Qdisc *qdisc;
780
781 qdisc = dev_queue->qdisc;
782 if (qdisc) {
783 spin_lock_bh(qdisc_lock(qdisc));
784
785 if (!(qdisc->flags & TCQ_F_BUILTIN))
786 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
787
788 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
789 qdisc_reset(qdisc);
790
791 spin_unlock_bh(qdisc_lock(qdisc));
792 }
793}
794
795static bool some_qdisc_is_busy(struct net_device *dev)
796{
797 unsigned int i;
798
799 for (i = 0; i < dev->num_tx_queues; i++) {
800 struct netdev_queue *dev_queue;
801 spinlock_t *root_lock;
802 struct Qdisc *q;
803 int val;
804
805 dev_queue = netdev_get_tx_queue(dev, i);
806 q = dev_queue->qdisc_sleeping;
807 root_lock = qdisc_lock(q);
808
809 spin_lock_bh(root_lock);
810
811 val = (qdisc_is_running(q) ||
812 test_bit(__QDISC_STATE_SCHED, &q->state));
813
814 spin_unlock_bh(root_lock);
815
816 if (val)
817 return true;
818 }
819 return false;
820}
821
822/**
823 * dev_deactivate_many - deactivate transmissions on several devices
824 * @head: list of devices to deactivate
825 *
826 * This function returns only when all outstanding transmissions
827 * have completed, unless all devices are in dismantle phase.
828 */
829void dev_deactivate_many(struct list_head *head)
830{
831 struct net_device *dev;
832 bool sync_needed = false;
833
834 list_for_each_entry(dev, head, close_list) {
835 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
836 &noop_qdisc);
837 if (dev_ingress_queue(dev))
838 dev_deactivate_queue(dev, dev_ingress_queue(dev),
839 &noop_qdisc);
840
841 dev_watchdog_down(dev);
842 sync_needed |= !dev->dismantle;
843 }
844
845 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
846 * This is avoided if all devices are in dismantle phase :
847 * Caller will call synchronize_net() for us
848 */
849 if (sync_needed)
850 synchronize_net();
851
852 /* Wait for outstanding qdisc_run calls. */
853 list_for_each_entry(dev, head, close_list)
854 while (some_qdisc_is_busy(dev))
855 yield();
856}
857
858void dev_deactivate(struct net_device *dev)
859{
860 LIST_HEAD(single);
861
862 list_add(&dev->close_list, &single);
863 dev_deactivate_many(&single);
864 list_del(&single);
865}
866EXPORT_SYMBOL(dev_deactivate);
867
868static void dev_init_scheduler_queue(struct net_device *dev,
869 struct netdev_queue *dev_queue,
870 void *_qdisc)
871{
872 struct Qdisc *qdisc = _qdisc;
873
874 dev_queue->qdisc = qdisc;
875 dev_queue->qdisc_sleeping = qdisc;
876}
877
878void dev_init_scheduler(struct net_device *dev)
879{
880 dev->qdisc = &noop_qdisc;
881 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
882 if (dev_ingress_queue(dev))
883 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
884
885 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
886}
887
888static void shutdown_scheduler_queue(struct net_device *dev,
889 struct netdev_queue *dev_queue,
890 void *_qdisc_default)
891{
892 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
893 struct Qdisc *qdisc_default = _qdisc_default;
894
895 if (qdisc) {
896 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
897 dev_queue->qdisc_sleeping = qdisc_default;
898
899 qdisc_destroy(qdisc);
900 }
901}
902
903void dev_shutdown(struct net_device *dev)
904{
905 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
906 if (dev_ingress_queue(dev))
907 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
908 qdisc_destroy(dev->qdisc);
909 dev->qdisc = &noop_qdisc;
910
911 WARN_ON(timer_pending(&dev->watchdog_timer));
912}
913
914void psched_ratecfg_precompute(struct psched_ratecfg *r,
915 const struct tc_ratespec *conf,
916 u64 rate64)
917{
918 memset(r, 0, sizeof(*r));
919 r->overhead = conf->overhead;
920 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
921 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
922 r->mult = 1;
923 /*
924 * The deal here is to replace a divide by a reciprocal one
925 * in fast path (a reciprocal divide is a multiply and a shift)
926 *
927 * Normal formula would be :
928 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
929 *
930 * We compute mult/shift to use instead :
931 * time_in_ns = (len * mult) >> shift;
932 *
933 * We try to get the highest possible mult value for accuracy,
934 * but have to make sure no overflows will ever happen.
935 */
936 if (r->rate_bytes_ps > 0) {
937 u64 factor = NSEC_PER_SEC;
938
939 for (;;) {
940 r->mult = div64_u64(factor, r->rate_bytes_ps);
941 if (r->mult & (1U << 31) || factor & (1ULL << 63))
942 break;
943 factor <<= 1;
944 r->shift++;
945 }
946 }
947}
948EXPORT_SYMBOL(psched_ratecfg_precompute);