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 q->gso_skb = skb;
51 q->qstats.requeues++;
52 q->q.qlen++; /* it's still part of the queue */
53 __netif_schedule(q);
54
55 return 0;
56}
57
58static void try_bulk_dequeue_skb(struct Qdisc *q,
59 struct sk_buff *skb,
60 const struct netdev_queue *txq,
61 int *packets)
62{
63 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
64
65 while (bytelimit > 0) {
66 struct sk_buff *nskb = q->dequeue(q);
67
68 if (!nskb)
69 break;
70
71 bytelimit -= nskb->len; /* covers GSO len */
72 skb->next = nskb;
73 skb = nskb;
74 (*packets)++; /* GSO counts as one pkt */
75 }
76 skb->next = NULL;
77}
78
79/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
80 * A requeued skb (via q->gso_skb) can also be a SKB list.
81 */
82static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
83 int *packets)
84{
85 struct sk_buff *skb = q->gso_skb;
86 const struct netdev_queue *txq = q->dev_queue;
87
88 *packets = 1;
89 *validate = true;
90 if (unlikely(skb)) {
91 /* check the reason of requeuing without tx lock first */
92 txq = skb_get_tx_queue(txq->dev, skb);
93 if (!netif_xmit_frozen_or_stopped(txq)) {
94 q->gso_skb = NULL;
95 q->q.qlen--;
96 } else
97 skb = NULL;
98 /* skb in gso_skb were already validated */
99 *validate = false;
100 } else {
101 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
102 !netif_xmit_frozen_or_stopped(txq)) {
103 skb = q->dequeue(q);
104 if (skb && qdisc_may_bulk(q))
105 try_bulk_dequeue_skb(q, skb, txq, packets);
106 }
107 }
108 return skb;
109}
110
111static inline int handle_dev_cpu_collision(struct sk_buff *skb,
112 struct netdev_queue *dev_queue,
113 struct Qdisc *q)
114{
115 int ret;
116
117 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
118 /*
119 * Same CPU holding the lock. It may be a transient
120 * configuration error, when hard_start_xmit() recurses. We
121 * detect it by checking xmit owner and drop the packet when
122 * deadloop is detected. Return OK to try the next skb.
123 */
124 kfree_skb_list(skb);
125 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
126 dev_queue->dev->name);
127 ret = qdisc_qlen(q);
128 } else {
129 /*
130 * Another cpu is holding lock, requeue & delay xmits for
131 * some time.
132 */
133 __this_cpu_inc(softnet_data.cpu_collision);
134 ret = dev_requeue_skb(skb, q);
135 }
136
137 return ret;
138}
139
140/*
141 * Transmit possibly several skbs, and handle the return status as
142 * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
143 * only one CPU can execute this function.
144 *
145 * Returns to the caller:
146 * 0 - queue is empty or throttled.
147 * >0 - queue is not empty.
148 */
149int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
150 struct net_device *dev, struct netdev_queue *txq,
151 spinlock_t *root_lock, bool validate)
152{
153 int ret = NETDEV_TX_BUSY;
154
155 /* And release qdisc */
156 spin_unlock(root_lock);
157
158 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
159 if (validate)
160 skb = validate_xmit_skb_list(skb, dev);
161
162 if (likely(skb)) {
163 HARD_TX_LOCK(dev, txq, smp_processor_id());
164 if (!netif_xmit_frozen_or_stopped(txq))
165 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
166
167 HARD_TX_UNLOCK(dev, txq);
168 } else {
169 spin_lock(root_lock);
170 return qdisc_qlen(q);
171 }
172 spin_lock(root_lock);
173
174 if (dev_xmit_complete(ret)) {
175 /* Driver sent out skb successfully or skb was consumed */
176 ret = qdisc_qlen(q);
177 } else if (ret == NETDEV_TX_LOCKED) {
178 /* Driver try lock failed */
179 ret = handle_dev_cpu_collision(skb, txq, q);
180 } else {
181 /* Driver returned NETDEV_TX_BUSY - requeue skb */
182 if (unlikely(ret != NETDEV_TX_BUSY))
183 net_warn_ratelimited("BUG %s code %d qlen %d\n",
184 dev->name, ret, q->q.qlen);
185
186 ret = dev_requeue_skb(skb, q);
187 }
188
189 if (ret && netif_xmit_frozen_or_stopped(txq))
190 ret = 0;
191
192 return ret;
193}
194
195/*
196 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
197 *
198 * __QDISC___STATE_RUNNING guarantees only one CPU can process
199 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
200 * this queue.
201 *
202 * netif_tx_lock serializes accesses to device driver.
203 *
204 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
205 * if one is grabbed, another must be free.
206 *
207 * Note, that this procedure can be called by a watchdog timer
208 *
209 * Returns to the caller:
210 * 0 - queue is empty or throttled.
211 * >0 - queue is not empty.
212 *
213 */
214static inline int qdisc_restart(struct Qdisc *q, int *packets)
215{
216 struct netdev_queue *txq;
217 struct net_device *dev;
218 spinlock_t *root_lock;
219 struct sk_buff *skb;
220 bool validate;
221
222 /* Dequeue packet */
223 skb = dequeue_skb(q, &validate, packets);
224 if (unlikely(!skb))
225 return 0;
226
227 root_lock = qdisc_lock(q);
228 dev = qdisc_dev(q);
229 txq = skb_get_tx_queue(dev, skb);
230
231 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
232}
233
234void __qdisc_run(struct Qdisc *q)
235{
236 int quota = weight_p;
237 int packets;
238
239 while (qdisc_restart(q, &packets)) {
240 /*
241 * Ordered by possible occurrence: Postpone processing if
242 * 1. we've exceeded packet quota
243 * 2. another process needs the CPU;
244 */
245 quota -= packets;
246 if (quota <= 0 || need_resched()) {
247 __netif_schedule(q);
248 break;
249 }
250 }
251
252 qdisc_run_end(q);
253}
254
255unsigned long dev_trans_start(struct net_device *dev)
256{
257 unsigned long val, res;
258 unsigned int i;
259
260 if (is_vlan_dev(dev))
261 dev = vlan_dev_real_dev(dev);
262 res = dev->trans_start;
263 for (i = 0; i < dev->num_tx_queues; i++) {
264 val = netdev_get_tx_queue(dev, i)->trans_start;
265 if (val && time_after(val, res))
266 res = val;
267 }
268 dev->trans_start = res;
269
270 return res;
271}
272EXPORT_SYMBOL(dev_trans_start);
273
274static void dev_watchdog(unsigned long arg)
275{
276 struct net_device *dev = (struct net_device *)arg;
277
278 netif_tx_lock(dev);
279 if (!qdisc_tx_is_noop(dev)) {
280 if (netif_device_present(dev) &&
281 netif_running(dev) &&
282 netif_carrier_ok(dev)) {
283 int some_queue_timedout = 0;
284 unsigned int i;
285 unsigned long trans_start;
286
287 for (i = 0; i < dev->num_tx_queues; i++) {
288 struct netdev_queue *txq;
289
290 txq = netdev_get_tx_queue(dev, i);
291 /*
292 * old device drivers set dev->trans_start
293 */
294 trans_start = txq->trans_start ? : dev->trans_start;
295 if (netif_xmit_stopped(txq) &&
296 time_after(jiffies, (trans_start +
297 dev->watchdog_timeo))) {
298 some_queue_timedout = 1;
299 txq->trans_timeout++;
300 break;
301 }
302 }
303
304 if (some_queue_timedout) {
305 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
306 dev->name, netdev_drivername(dev), i);
307 dev->netdev_ops->ndo_tx_timeout(dev);
308 }
309 if (!mod_timer(&dev->watchdog_timer,
310 round_jiffies(jiffies +
311 dev->watchdog_timeo)))
312 dev_hold(dev);
313 }
314 }
315 netif_tx_unlock(dev);
316
317 dev_put(dev);
318}
319
320void __netdev_watchdog_up(struct net_device *dev)
321{
322 if (dev->netdev_ops->ndo_tx_timeout) {
323 if (dev->watchdog_timeo <= 0)
324 dev->watchdog_timeo = 5*HZ;
325 if (!mod_timer(&dev->watchdog_timer,
326 round_jiffies(jiffies + dev->watchdog_timeo)))
327 dev_hold(dev);
328 }
329}
330
331static void dev_watchdog_up(struct net_device *dev)
332{
333 __netdev_watchdog_up(dev);
334}
335
336static void dev_watchdog_down(struct net_device *dev)
337{
338 netif_tx_lock_bh(dev);
339 if (del_timer(&dev->watchdog_timer))
340 dev_put(dev);
341 netif_tx_unlock_bh(dev);
342}
343
344/**
345 * netif_carrier_on - set carrier
346 * @dev: network device
347 *
348 * Device has detected that carrier.
349 */
350void netif_carrier_on(struct net_device *dev)
351{
352 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
353 if (dev->reg_state == NETREG_UNINITIALIZED)
354 return;
355 atomic_inc(&dev->carrier_changes);
356 linkwatch_fire_event(dev);
357 if (netif_running(dev))
358 __netdev_watchdog_up(dev);
359 }
360}
361EXPORT_SYMBOL(netif_carrier_on);
362
363/**
364 * netif_carrier_off - clear carrier
365 * @dev: network device
366 *
367 * Device has detected loss of carrier.
368 */
369void netif_carrier_off(struct net_device *dev)
370{
371 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
372 if (dev->reg_state == NETREG_UNINITIALIZED)
373 return;
374 atomic_inc(&dev->carrier_changes);
375 linkwatch_fire_event(dev);
376 }
377}
378EXPORT_SYMBOL(netif_carrier_off);
379
380/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
381 under all circumstances. It is difficult to invent anything faster or
382 cheaper.
383 */
384
385static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
386{
387 kfree_skb(skb);
388 return NET_XMIT_CN;
389}
390
391static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
392{
393 return NULL;
394}
395
396struct Qdisc_ops noop_qdisc_ops __read_mostly = {
397 .id = "noop",
398 .priv_size = 0,
399 .enqueue = noop_enqueue,
400 .dequeue = noop_dequeue,
401 .peek = noop_dequeue,
402 .owner = THIS_MODULE,
403};
404
405static struct netdev_queue noop_netdev_queue = {
406 .qdisc = &noop_qdisc,
407 .qdisc_sleeping = &noop_qdisc,
408};
409
410struct Qdisc noop_qdisc = {
411 .enqueue = noop_enqueue,
412 .dequeue = noop_dequeue,
413 .flags = TCQ_F_BUILTIN,
414 .ops = &noop_qdisc_ops,
415 .list = LIST_HEAD_INIT(noop_qdisc.list),
416 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
417 .dev_queue = &noop_netdev_queue,
418 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
419};
420EXPORT_SYMBOL(noop_qdisc);
421
422static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
423{
424 /* register_qdisc() assigns a default of noop_enqueue if unset,
425 * but __dev_queue_xmit() treats noqueue only as such
426 * if this is NULL - so clear it here. */
427 qdisc->enqueue = NULL;
428 return 0;
429}
430
431struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
432 .id = "noqueue",
433 .priv_size = 0,
434 .init = noqueue_init,
435 .enqueue = noop_enqueue,
436 .dequeue = noop_dequeue,
437 .peek = noop_dequeue,
438 .owner = THIS_MODULE,
439};
440
441static const u8 prio2band[TC_PRIO_MAX + 1] = {
442 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
443};
444
445/* 3-band FIFO queue: old style, but should be a bit faster than
446 generic prio+fifo combination.
447 */
448
449#define PFIFO_FAST_BANDS 3
450
451/*
452 * Private data for a pfifo_fast scheduler containing:
453 * - queues for the three band
454 * - bitmap indicating which of the bands contain skbs
455 */
456struct pfifo_fast_priv {
457 u32 bitmap;
458 struct sk_buff_head q[PFIFO_FAST_BANDS];
459};
460
461/*
462 * Convert a bitmap to the first band number where an skb is queued, where:
463 * bitmap=0 means there are no skbs on any band.
464 * bitmap=1 means there is an skb on band 0.
465 * bitmap=7 means there are skbs on all 3 bands, etc.
466 */
467static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
468
469static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
470 int band)
471{
472 return priv->q + band;
473}
474
475static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
476{
477 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
478 int band = prio2band[skb->priority & TC_PRIO_MAX];
479 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
480 struct sk_buff_head *list = band2list(priv, band);
481
482 priv->bitmap |= (1 << band);
483 qdisc->q.qlen++;
484 return __qdisc_enqueue_tail(skb, qdisc, list);
485 }
486
487 return qdisc_drop(skb, qdisc);
488}
489
490static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
491{
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493 int band = bitmap2band[priv->bitmap];
494
495 if (likely(band >= 0)) {
496 struct sk_buff_head *list = band2list(priv, band);
497 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
498
499 qdisc->q.qlen--;
500 if (skb_queue_empty(list))
501 priv->bitmap &= ~(1 << band);
502
503 return skb;
504 }
505
506 return NULL;
507}
508
509static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
510{
511 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
512 int band = bitmap2band[priv->bitmap];
513
514 if (band >= 0) {
515 struct sk_buff_head *list = band2list(priv, band);
516
517 return skb_peek(list);
518 }
519
520 return NULL;
521}
522
523static void pfifo_fast_reset(struct Qdisc *qdisc)
524{
525 int prio;
526 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
527
528 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
529 __qdisc_reset_queue(qdisc, band2list(priv, prio));
530
531 priv->bitmap = 0;
532 qdisc->qstats.backlog = 0;
533 qdisc->q.qlen = 0;
534}
535
536static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
537{
538 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
539
540 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
541 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
542 goto nla_put_failure;
543 return skb->len;
544
545nla_put_failure:
546 return -1;
547}
548
549static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
550{
551 int prio;
552 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
553
554 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
555 __skb_queue_head_init(band2list(priv, prio));
556
557 /* Can by-pass the queue discipline */
558 qdisc->flags |= TCQ_F_CAN_BYPASS;
559 return 0;
560}
561
562struct Qdisc_ops pfifo_fast_ops __read_mostly = {
563 .id = "pfifo_fast",
564 .priv_size = sizeof(struct pfifo_fast_priv),
565 .enqueue = pfifo_fast_enqueue,
566 .dequeue = pfifo_fast_dequeue,
567 .peek = pfifo_fast_peek,
568 .init = pfifo_fast_init,
569 .reset = pfifo_fast_reset,
570 .dump = pfifo_fast_dump,
571 .owner = THIS_MODULE,
572};
573EXPORT_SYMBOL(pfifo_fast_ops);
574
575static struct lock_class_key qdisc_tx_busylock;
576
577struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
578 const struct Qdisc_ops *ops)
579{
580 void *p;
581 struct Qdisc *sch;
582 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
583 int err = -ENOBUFS;
584 struct net_device *dev = dev_queue->dev;
585
586 p = kzalloc_node(size, GFP_KERNEL,
587 netdev_queue_numa_node_read(dev_queue));
588
589 if (!p)
590 goto errout;
591 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
592 /* if we got non aligned memory, ask more and do alignment ourself */
593 if (sch != p) {
594 kfree(p);
595 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
596 netdev_queue_numa_node_read(dev_queue));
597 if (!p)
598 goto errout;
599 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
600 sch->padded = (char *) sch - (char *) p;
601 }
602 INIT_LIST_HEAD(&sch->list);
603 skb_queue_head_init(&sch->q);
604
605 spin_lock_init(&sch->busylock);
606 lockdep_set_class(&sch->busylock,
607 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
608
609 sch->ops = ops;
610 sch->enqueue = ops->enqueue;
611 sch->dequeue = ops->dequeue;
612 sch->dev_queue = dev_queue;
613 dev_hold(dev);
614 atomic_set(&sch->refcnt, 1);
615
616 return sch;
617errout:
618 return ERR_PTR(err);
619}
620
621struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
622 const struct Qdisc_ops *ops,
623 unsigned int parentid)
624{
625 struct Qdisc *sch;
626
627 if (!try_module_get(ops->owner))
628 goto errout;
629
630 sch = qdisc_alloc(dev_queue, ops);
631 if (IS_ERR(sch))
632 goto errout;
633 sch->parent = parentid;
634
635 if (!ops->init || ops->init(sch, NULL) == 0)
636 return sch;
637
638 qdisc_destroy(sch);
639errout:
640 return NULL;
641}
642EXPORT_SYMBOL(qdisc_create_dflt);
643
644/* Under qdisc_lock(qdisc) and BH! */
645
646void qdisc_reset(struct Qdisc *qdisc)
647{
648 const struct Qdisc_ops *ops = qdisc->ops;
649
650 if (ops->reset)
651 ops->reset(qdisc);
652
653 if (qdisc->gso_skb) {
654 kfree_skb_list(qdisc->gso_skb);
655 qdisc->gso_skb = NULL;
656 qdisc->q.qlen = 0;
657 }
658}
659EXPORT_SYMBOL(qdisc_reset);
660
661static void qdisc_rcu_free(struct rcu_head *head)
662{
663 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
664
665 if (qdisc_is_percpu_stats(qdisc)) {
666 free_percpu(qdisc->cpu_bstats);
667 free_percpu(qdisc->cpu_qstats);
668 }
669
670 kfree((char *) qdisc - qdisc->padded);
671}
672
673void qdisc_destroy(struct Qdisc *qdisc)
674{
675 const struct Qdisc_ops *ops = qdisc->ops;
676
677 if (qdisc->flags & TCQ_F_BUILTIN ||
678 !atomic_dec_and_test(&qdisc->refcnt))
679 return;
680
681#ifdef CONFIG_NET_SCHED
682 qdisc_list_del(qdisc);
683
684 qdisc_put_stab(rtnl_dereference(qdisc->stab));
685#endif
686 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
687 if (ops->reset)
688 ops->reset(qdisc);
689 if (ops->destroy)
690 ops->destroy(qdisc);
691
692 module_put(ops->owner);
693 dev_put(qdisc_dev(qdisc));
694
695 kfree_skb_list(qdisc->gso_skb);
696 /*
697 * gen_estimator est_timer() might access qdisc->q.lock,
698 * wait a RCU grace period before freeing qdisc.
699 */
700 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
701}
702EXPORT_SYMBOL(qdisc_destroy);
703
704/* Attach toplevel qdisc to device queue. */
705struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
706 struct Qdisc *qdisc)
707{
708 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
709 spinlock_t *root_lock;
710
711 root_lock = qdisc_lock(oqdisc);
712 spin_lock_bh(root_lock);
713
714 /* Prune old scheduler */
715 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
716 qdisc_reset(oqdisc);
717
718 /* ... and graft new one */
719 if (qdisc == NULL)
720 qdisc = &noop_qdisc;
721 dev_queue->qdisc_sleeping = qdisc;
722 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
723
724 spin_unlock_bh(root_lock);
725
726 return oqdisc;
727}
728EXPORT_SYMBOL(dev_graft_qdisc);
729
730static void attach_one_default_qdisc(struct net_device *dev,
731 struct netdev_queue *dev_queue,
732 void *_unused)
733{
734 struct Qdisc *qdisc;
735 const struct Qdisc_ops *ops = default_qdisc_ops;
736
737 if (dev->priv_flags & IFF_NO_QUEUE)
738 ops = &noqueue_qdisc_ops;
739
740 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
741 if (!qdisc) {
742 netdev_info(dev, "activation failed\n");
743 return;
744 }
745 if (!netif_is_multiqueue(dev))
746 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
747 dev_queue->qdisc_sleeping = qdisc;
748}
749
750static void attach_default_qdiscs(struct net_device *dev)
751{
752 struct netdev_queue *txq;
753 struct Qdisc *qdisc;
754
755 txq = netdev_get_tx_queue(dev, 0);
756
757 if (!netif_is_multiqueue(dev) ||
758 dev->priv_flags & IFF_NO_QUEUE) {
759 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
760 dev->qdisc = txq->qdisc_sleeping;
761 atomic_inc(&dev->qdisc->refcnt);
762 } else {
763 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
764 if (qdisc) {
765 dev->qdisc = qdisc;
766 qdisc->ops->attach(qdisc);
767 }
768 }
769}
770
771static void transition_one_qdisc(struct net_device *dev,
772 struct netdev_queue *dev_queue,
773 void *_need_watchdog)
774{
775 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
776 int *need_watchdog_p = _need_watchdog;
777
778 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
779 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
780
781 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
782 if (need_watchdog_p) {
783 dev_queue->trans_start = 0;
784 *need_watchdog_p = 1;
785 }
786}
787
788void dev_activate(struct net_device *dev)
789{
790 int need_watchdog;
791
792 /* No queueing discipline is attached to device;
793 * create default one for devices, which need queueing
794 * and noqueue_qdisc for virtual interfaces
795 */
796
797 if (dev->qdisc == &noop_qdisc)
798 attach_default_qdiscs(dev);
799
800 if (!netif_carrier_ok(dev))
801 /* Delay activation until next carrier-on event */
802 return;
803
804 need_watchdog = 0;
805 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
806 if (dev_ingress_queue(dev))
807 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
808
809 if (need_watchdog) {
810 dev->trans_start = jiffies;
811 dev_watchdog_up(dev);
812 }
813}
814EXPORT_SYMBOL(dev_activate);
815
816static void dev_deactivate_queue(struct net_device *dev,
817 struct netdev_queue *dev_queue,
818 void *_qdisc_default)
819{
820 struct Qdisc *qdisc_default = _qdisc_default;
821 struct Qdisc *qdisc;
822
823 qdisc = rtnl_dereference(dev_queue->qdisc);
824 if (qdisc) {
825 spin_lock_bh(qdisc_lock(qdisc));
826
827 if (!(qdisc->flags & TCQ_F_BUILTIN))
828 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
829
830 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
831 qdisc_reset(qdisc);
832
833 spin_unlock_bh(qdisc_lock(qdisc));
834 }
835}
836
837static bool some_qdisc_is_busy(struct net_device *dev)
838{
839 unsigned int i;
840
841 for (i = 0; i < dev->num_tx_queues; i++) {
842 struct netdev_queue *dev_queue;
843 spinlock_t *root_lock;
844 struct Qdisc *q;
845 int val;
846
847 dev_queue = netdev_get_tx_queue(dev, i);
848 q = dev_queue->qdisc_sleeping;
849 root_lock = qdisc_lock(q);
850
851 spin_lock_bh(root_lock);
852
853 val = (qdisc_is_running(q) ||
854 test_bit(__QDISC_STATE_SCHED, &q->state));
855
856 spin_unlock_bh(root_lock);
857
858 if (val)
859 return true;
860 }
861 return false;
862}
863
864/**
865 * dev_deactivate_many - deactivate transmissions on several devices
866 * @head: list of devices to deactivate
867 *
868 * This function returns only when all outstanding transmissions
869 * have completed, unless all devices are in dismantle phase.
870 */
871void dev_deactivate_many(struct list_head *head)
872{
873 struct net_device *dev;
874 bool sync_needed = false;
875
876 list_for_each_entry(dev, head, close_list) {
877 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
878 &noop_qdisc);
879 if (dev_ingress_queue(dev))
880 dev_deactivate_queue(dev, dev_ingress_queue(dev),
881 &noop_qdisc);
882
883 dev_watchdog_down(dev);
884 sync_needed |= !dev->dismantle;
885 }
886
887 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
888 * This is avoided if all devices are in dismantle phase :
889 * Caller will call synchronize_net() for us
890 */
891 if (sync_needed)
892 synchronize_net();
893
894 /* Wait for outstanding qdisc_run calls. */
895 list_for_each_entry(dev, head, close_list)
896 while (some_qdisc_is_busy(dev))
897 yield();
898}
899
900void dev_deactivate(struct net_device *dev)
901{
902 LIST_HEAD(single);
903
904 list_add(&dev->close_list, &single);
905 dev_deactivate_many(&single);
906 list_del(&single);
907}
908EXPORT_SYMBOL(dev_deactivate);
909
910static void dev_init_scheduler_queue(struct net_device *dev,
911 struct netdev_queue *dev_queue,
912 void *_qdisc)
913{
914 struct Qdisc *qdisc = _qdisc;
915
916 rcu_assign_pointer(dev_queue->qdisc, qdisc);
917 dev_queue->qdisc_sleeping = qdisc;
918}
919
920void dev_init_scheduler(struct net_device *dev)
921{
922 dev->qdisc = &noop_qdisc;
923 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
924 if (dev_ingress_queue(dev))
925 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
926
927 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
928}
929
930static void shutdown_scheduler_queue(struct net_device *dev,
931 struct netdev_queue *dev_queue,
932 void *_qdisc_default)
933{
934 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
935 struct Qdisc *qdisc_default = _qdisc_default;
936
937 if (qdisc) {
938 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
939 dev_queue->qdisc_sleeping = qdisc_default;
940
941 qdisc_destroy(qdisc);
942 }
943}
944
945void dev_shutdown(struct net_device *dev)
946{
947 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
948 if (dev_ingress_queue(dev))
949 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
950 qdisc_destroy(dev->qdisc);
951 dev->qdisc = &noop_qdisc;
952
953 WARN_ON(timer_pending(&dev->watchdog_timer));
954}
955
956void psched_ratecfg_precompute(struct psched_ratecfg *r,
957 const struct tc_ratespec *conf,
958 u64 rate64)
959{
960 memset(r, 0, sizeof(*r));
961 r->overhead = conf->overhead;
962 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
963 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
964 r->mult = 1;
965 /*
966 * The deal here is to replace a divide by a reciprocal one
967 * in fast path (a reciprocal divide is a multiply and a shift)
968 *
969 * Normal formula would be :
970 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
971 *
972 * We compute mult/shift to use instead :
973 * time_in_ns = (len * mult) >> shift;
974 *
975 * We try to get the highest possible mult value for accuracy,
976 * but have to make sure no overflows will ever happen.
977 */
978 if (r->rate_bytes_ps > 0) {
979 u64 factor = NSEC_PER_SEC;
980
981 for (;;) {
982 r->mult = div64_u64(factor, r->rate_bytes_ps);
983 if (r->mult & (1U << 31) || factor & (1ULL << 63))
984 break;
985 factor <<= 1;
986 r->shift++;
987 }
988 }
989}
990EXPORT_SYMBOL(psched_ratecfg_precompute);