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 qdisc_qstats_backlog_inc(q, skb);
53 q->q.qlen++; /* it's still part of the queue */
54 __netif_schedule(q);
55
56 return 0;
57}
58
59static void try_bulk_dequeue_skb(struct Qdisc *q,
60 struct sk_buff *skb,
61 const struct netdev_queue *txq,
62 int *packets)
63{
64 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
65
66 while (bytelimit > 0) {
67 struct sk_buff *nskb = q->dequeue(q);
68
69 if (!nskb)
70 break;
71
72 bytelimit -= nskb->len; /* covers GSO len */
73 skb->next = nskb;
74 skb = nskb;
75 (*packets)++; /* GSO counts as one pkt */
76 }
77 skb->next = NULL;
78}
79
80/* This variant of try_bulk_dequeue_skb() makes sure
81 * all skbs in the chain are for the same txq
82 */
83static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
84 struct sk_buff *skb,
85 int *packets)
86{
87 int mapping = skb_get_queue_mapping(skb);
88 struct sk_buff *nskb;
89 int cnt = 0;
90
91 do {
92 nskb = q->dequeue(q);
93 if (!nskb)
94 break;
95 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
96 q->skb_bad_txq = nskb;
97 qdisc_qstats_backlog_inc(q, nskb);
98 q->q.qlen++;
99 break;
100 }
101 skb->next = nskb;
102 skb = nskb;
103 } while (++cnt < 8);
104 (*packets) += cnt;
105 skb->next = NULL;
106}
107
108/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
109 * A requeued skb (via q->gso_skb) can also be a SKB list.
110 */
111static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
112 int *packets)
113{
114 struct sk_buff *skb = q->gso_skb;
115 const struct netdev_queue *txq = q->dev_queue;
116
117 *packets = 1;
118 if (unlikely(skb)) {
119 /* skb in gso_skb were already validated */
120 *validate = false;
121 /* check the reason of requeuing without tx lock first */
122 txq = skb_get_tx_queue(txq->dev, skb);
123 if (!netif_xmit_frozen_or_stopped(txq)) {
124 q->gso_skb = NULL;
125 qdisc_qstats_backlog_dec(q, skb);
126 q->q.qlen--;
127 } else
128 skb = NULL;
129 return skb;
130 }
131 *validate = true;
132 skb = q->skb_bad_txq;
133 if (unlikely(skb)) {
134 /* check the reason of requeuing without tx lock first */
135 txq = skb_get_tx_queue(txq->dev, skb);
136 if (!netif_xmit_frozen_or_stopped(txq)) {
137 q->skb_bad_txq = NULL;
138 qdisc_qstats_backlog_dec(q, skb);
139 q->q.qlen--;
140 goto bulk;
141 }
142 return NULL;
143 }
144 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
145 !netif_xmit_frozen_or_stopped(txq))
146 skb = q->dequeue(q);
147 if (skb) {
148bulk:
149 if (qdisc_may_bulk(q))
150 try_bulk_dequeue_skb(q, skb, txq, packets);
151 else
152 try_bulk_dequeue_skb_slow(q, skb, packets);
153 }
154 return skb;
155}
156
157/*
158 * Transmit possibly several skbs, and handle the return status as
159 * required. Owning running seqcount bit guarantees that
160 * only one CPU can execute this function.
161 *
162 * Returns to the caller:
163 * 0 - queue is empty or throttled.
164 * >0 - queue is not empty.
165 */
166int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
167 struct net_device *dev, struct netdev_queue *txq,
168 spinlock_t *root_lock, bool validate)
169{
170 int ret = NETDEV_TX_BUSY;
171
172 /* And release qdisc */
173 spin_unlock(root_lock);
174
175 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
176 if (validate)
177 skb = validate_xmit_skb_list(skb, dev);
178
179 if (likely(skb)) {
180 HARD_TX_LOCK(dev, txq, smp_processor_id());
181 if (!netif_xmit_frozen_or_stopped(txq))
182 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
183
184 HARD_TX_UNLOCK(dev, txq);
185 } else {
186 spin_lock(root_lock);
187 return qdisc_qlen(q);
188 }
189 spin_lock(root_lock);
190
191 if (dev_xmit_complete(ret)) {
192 /* Driver sent out skb successfully or skb was consumed */
193 ret = qdisc_qlen(q);
194 } else {
195 /* Driver returned NETDEV_TX_BUSY - requeue skb */
196 if (unlikely(ret != NETDEV_TX_BUSY))
197 net_warn_ratelimited("BUG %s code %d qlen %d\n",
198 dev->name, ret, q->q.qlen);
199
200 ret = dev_requeue_skb(skb, q);
201 }
202
203 if (ret && netif_xmit_frozen_or_stopped(txq))
204 ret = 0;
205
206 return ret;
207}
208
209/*
210 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
211 *
212 * running seqcount guarantees only one CPU can process
213 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
214 * this queue.
215 *
216 * netif_tx_lock serializes accesses to device driver.
217 *
218 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
219 * if one is grabbed, another must be free.
220 *
221 * Note, that this procedure can be called by a watchdog timer
222 *
223 * Returns to the caller:
224 * 0 - queue is empty or throttled.
225 * >0 - queue is not empty.
226 *
227 */
228static inline int qdisc_restart(struct Qdisc *q, int *packets)
229{
230 struct netdev_queue *txq;
231 struct net_device *dev;
232 spinlock_t *root_lock;
233 struct sk_buff *skb;
234 bool validate;
235
236 /* Dequeue packet */
237 skb = dequeue_skb(q, &validate, packets);
238 if (unlikely(!skb))
239 return 0;
240
241 root_lock = qdisc_lock(q);
242 dev = qdisc_dev(q);
243 txq = skb_get_tx_queue(dev, skb);
244
245 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
246}
247
248void __qdisc_run(struct Qdisc *q)
249{
250 int quota = weight_p;
251 int packets;
252
253 while (qdisc_restart(q, &packets)) {
254 /*
255 * Ordered by possible occurrence: Postpone processing if
256 * 1. we've exceeded packet quota
257 * 2. another process needs the CPU;
258 */
259 quota -= packets;
260 if (quota <= 0 || need_resched()) {
261 __netif_schedule(q);
262 break;
263 }
264 }
265
266 qdisc_run_end(q);
267}
268
269unsigned long dev_trans_start(struct net_device *dev)
270{
271 unsigned long val, res;
272 unsigned int i;
273
274 if (is_vlan_dev(dev))
275 dev = vlan_dev_real_dev(dev);
276 res = netdev_get_tx_queue(dev, 0)->trans_start;
277 for (i = 1; i < dev->num_tx_queues; i++) {
278 val = netdev_get_tx_queue(dev, i)->trans_start;
279 if (val && time_after(val, res))
280 res = val;
281 }
282
283 return res;
284}
285EXPORT_SYMBOL(dev_trans_start);
286
287static void dev_watchdog(unsigned long arg)
288{
289 struct net_device *dev = (struct net_device *)arg;
290
291 netif_tx_lock(dev);
292 if (!qdisc_tx_is_noop(dev)) {
293 if (netif_device_present(dev) &&
294 netif_running(dev) &&
295 netif_carrier_ok(dev)) {
296 int some_queue_timedout = 0;
297 unsigned int i;
298 unsigned long trans_start;
299
300 for (i = 0; i < dev->num_tx_queues; i++) {
301 struct netdev_queue *txq;
302
303 txq = netdev_get_tx_queue(dev, i);
304 trans_start = txq->trans_start;
305 if (netif_xmit_stopped(txq) &&
306 time_after(jiffies, (trans_start +
307 dev->watchdog_timeo))) {
308 some_queue_timedout = 1;
309 txq->trans_timeout++;
310 break;
311 }
312 }
313
314 if (some_queue_timedout) {
315 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
316 dev->name, netdev_drivername(dev), i);
317 dev->netdev_ops->ndo_tx_timeout(dev);
318 }
319 if (!mod_timer(&dev->watchdog_timer,
320 round_jiffies(jiffies +
321 dev->watchdog_timeo)))
322 dev_hold(dev);
323 }
324 }
325 netif_tx_unlock(dev);
326
327 dev_put(dev);
328}
329
330void __netdev_watchdog_up(struct net_device *dev)
331{
332 if (dev->netdev_ops->ndo_tx_timeout) {
333 if (dev->watchdog_timeo <= 0)
334 dev->watchdog_timeo = 5*HZ;
335 if (!mod_timer(&dev->watchdog_timer,
336 round_jiffies(jiffies + dev->watchdog_timeo)))
337 dev_hold(dev);
338 }
339}
340
341static void dev_watchdog_up(struct net_device *dev)
342{
343 __netdev_watchdog_up(dev);
344}
345
346static void dev_watchdog_down(struct net_device *dev)
347{
348 netif_tx_lock_bh(dev);
349 if (del_timer(&dev->watchdog_timer))
350 dev_put(dev);
351 netif_tx_unlock_bh(dev);
352}
353
354/**
355 * netif_carrier_on - set carrier
356 * @dev: network device
357 *
358 * Device has detected that carrier.
359 */
360void netif_carrier_on(struct net_device *dev)
361{
362 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
363 if (dev->reg_state == NETREG_UNINITIALIZED)
364 return;
365 atomic_inc(&dev->carrier_changes);
366 linkwatch_fire_event(dev);
367 if (netif_running(dev))
368 __netdev_watchdog_up(dev);
369 }
370}
371EXPORT_SYMBOL(netif_carrier_on);
372
373/**
374 * netif_carrier_off - clear carrier
375 * @dev: network device
376 *
377 * Device has detected loss of carrier.
378 */
379void netif_carrier_off(struct net_device *dev)
380{
381 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
382 if (dev->reg_state == NETREG_UNINITIALIZED)
383 return;
384 atomic_inc(&dev->carrier_changes);
385 linkwatch_fire_event(dev);
386 }
387}
388EXPORT_SYMBOL(netif_carrier_off);
389
390/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
391 under all circumstances. It is difficult to invent anything faster or
392 cheaper.
393 */
394
395static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
396 struct sk_buff **to_free)
397{
398 __qdisc_drop(skb, to_free);
399 return NET_XMIT_CN;
400}
401
402static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
403{
404 return NULL;
405}
406
407struct Qdisc_ops noop_qdisc_ops __read_mostly = {
408 .id = "noop",
409 .priv_size = 0,
410 .enqueue = noop_enqueue,
411 .dequeue = noop_dequeue,
412 .peek = noop_dequeue,
413 .owner = THIS_MODULE,
414};
415
416static struct netdev_queue noop_netdev_queue = {
417 .qdisc = &noop_qdisc,
418 .qdisc_sleeping = &noop_qdisc,
419};
420
421struct Qdisc noop_qdisc = {
422 .enqueue = noop_enqueue,
423 .dequeue = noop_dequeue,
424 .flags = TCQ_F_BUILTIN,
425 .ops = &noop_qdisc_ops,
426 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
427 .dev_queue = &noop_netdev_queue,
428 .running = SEQCNT_ZERO(noop_qdisc.running),
429 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
430};
431EXPORT_SYMBOL(noop_qdisc);
432
433static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
434{
435 /* register_qdisc() assigns a default of noop_enqueue if unset,
436 * but __dev_queue_xmit() treats noqueue only as such
437 * if this is NULL - so clear it here. */
438 qdisc->enqueue = NULL;
439 return 0;
440}
441
442struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
443 .id = "noqueue",
444 .priv_size = 0,
445 .init = noqueue_init,
446 .enqueue = noop_enqueue,
447 .dequeue = noop_dequeue,
448 .peek = noop_dequeue,
449 .owner = THIS_MODULE,
450};
451
452static const u8 prio2band[TC_PRIO_MAX + 1] = {
453 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
454};
455
456/* 3-band FIFO queue: old style, but should be a bit faster than
457 generic prio+fifo combination.
458 */
459
460#define PFIFO_FAST_BANDS 3
461
462/*
463 * Private data for a pfifo_fast scheduler containing:
464 * - queues for the three band
465 * - bitmap indicating which of the bands contain skbs
466 */
467struct pfifo_fast_priv {
468 u32 bitmap;
469 struct qdisc_skb_head q[PFIFO_FAST_BANDS];
470};
471
472/*
473 * Convert a bitmap to the first band number where an skb is queued, where:
474 * bitmap=0 means there are no skbs on any band.
475 * bitmap=1 means there is an skb on band 0.
476 * bitmap=7 means there are skbs on all 3 bands, etc.
477 */
478static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
479
480static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
481 int band)
482{
483 return priv->q + band;
484}
485
486static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
487 struct sk_buff **to_free)
488{
489 if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
490 int band = prio2band[skb->priority & TC_PRIO_MAX];
491 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
492 struct qdisc_skb_head *list = band2list(priv, band);
493
494 priv->bitmap |= (1 << band);
495 qdisc->q.qlen++;
496 return __qdisc_enqueue_tail(skb, qdisc, list);
497 }
498
499 return qdisc_drop(skb, qdisc, to_free);
500}
501
502static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
503{
504 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
505 int band = bitmap2band[priv->bitmap];
506
507 if (likely(band >= 0)) {
508 struct qdisc_skb_head *qh = band2list(priv, band);
509 struct sk_buff *skb = __qdisc_dequeue_head(qh);
510
511 if (likely(skb != NULL)) {
512 qdisc_qstats_backlog_dec(qdisc, skb);
513 qdisc_bstats_update(qdisc, skb);
514 }
515
516 qdisc->q.qlen--;
517 if (qh->qlen == 0)
518 priv->bitmap &= ~(1 << band);
519
520 return skb;
521 }
522
523 return NULL;
524}
525
526static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
527{
528 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
529 int band = bitmap2band[priv->bitmap];
530
531 if (band >= 0) {
532 struct qdisc_skb_head *qh = band2list(priv, band);
533
534 return qh->head;
535 }
536
537 return NULL;
538}
539
540static void pfifo_fast_reset(struct Qdisc *qdisc)
541{
542 int prio;
543 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
544
545 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
546 __qdisc_reset_queue(band2list(priv, prio));
547
548 priv->bitmap = 0;
549 qdisc->qstats.backlog = 0;
550 qdisc->q.qlen = 0;
551}
552
553static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
554{
555 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
556
557 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
558 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
559 goto nla_put_failure;
560 return skb->len;
561
562nla_put_failure:
563 return -1;
564}
565
566static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
567{
568 int prio;
569 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
570
571 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
572 qdisc_skb_head_init(band2list(priv, prio));
573
574 /* Can by-pass the queue discipline */
575 qdisc->flags |= TCQ_F_CAN_BYPASS;
576 return 0;
577}
578
579struct Qdisc_ops pfifo_fast_ops __read_mostly = {
580 .id = "pfifo_fast",
581 .priv_size = sizeof(struct pfifo_fast_priv),
582 .enqueue = pfifo_fast_enqueue,
583 .dequeue = pfifo_fast_dequeue,
584 .peek = pfifo_fast_peek,
585 .init = pfifo_fast_init,
586 .reset = pfifo_fast_reset,
587 .dump = pfifo_fast_dump,
588 .owner = THIS_MODULE,
589};
590EXPORT_SYMBOL(pfifo_fast_ops);
591
592static struct lock_class_key qdisc_tx_busylock;
593static struct lock_class_key qdisc_running_key;
594
595struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
596 const struct Qdisc_ops *ops)
597{
598 void *p;
599 struct Qdisc *sch;
600 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
601 int err = -ENOBUFS;
602 struct net_device *dev = dev_queue->dev;
603
604 p = kzalloc_node(size, GFP_KERNEL,
605 netdev_queue_numa_node_read(dev_queue));
606
607 if (!p)
608 goto errout;
609 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
610 /* if we got non aligned memory, ask more and do alignment ourself */
611 if (sch != p) {
612 kfree(p);
613 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
614 netdev_queue_numa_node_read(dev_queue));
615 if (!p)
616 goto errout;
617 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
618 sch->padded = (char *) sch - (char *) p;
619 }
620 qdisc_skb_head_init(&sch->q);
621 spin_lock_init(&sch->q.lock);
622
623 spin_lock_init(&sch->busylock);
624 lockdep_set_class(&sch->busylock,
625 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
626
627 seqcount_init(&sch->running);
628 lockdep_set_class(&sch->running,
629 dev->qdisc_running_key ?: &qdisc_running_key);
630
631 sch->ops = ops;
632 sch->enqueue = ops->enqueue;
633 sch->dequeue = ops->dequeue;
634 sch->dev_queue = dev_queue;
635 dev_hold(dev);
636 atomic_set(&sch->refcnt, 1);
637
638 return sch;
639errout:
640 return ERR_PTR(err);
641}
642
643struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
644 const struct Qdisc_ops *ops,
645 unsigned int parentid)
646{
647 struct Qdisc *sch;
648
649 if (!try_module_get(ops->owner))
650 return NULL;
651
652 sch = qdisc_alloc(dev_queue, ops);
653 if (IS_ERR(sch)) {
654 module_put(ops->owner);
655 return NULL;
656 }
657 sch->parent = parentid;
658
659 if (!ops->init || ops->init(sch, NULL) == 0)
660 return sch;
661
662 qdisc_destroy(sch);
663 return NULL;
664}
665EXPORT_SYMBOL(qdisc_create_dflt);
666
667/* Under qdisc_lock(qdisc) and BH! */
668
669void qdisc_reset(struct Qdisc *qdisc)
670{
671 const struct Qdisc_ops *ops = qdisc->ops;
672
673 if (ops->reset)
674 ops->reset(qdisc);
675
676 kfree_skb(qdisc->skb_bad_txq);
677 qdisc->skb_bad_txq = NULL;
678
679 if (qdisc->gso_skb) {
680 kfree_skb_list(qdisc->gso_skb);
681 qdisc->gso_skb = NULL;
682 }
683 qdisc->q.qlen = 0;
684}
685EXPORT_SYMBOL(qdisc_reset);
686
687static void qdisc_rcu_free(struct rcu_head *head)
688{
689 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
690
691 if (qdisc_is_percpu_stats(qdisc)) {
692 free_percpu(qdisc->cpu_bstats);
693 free_percpu(qdisc->cpu_qstats);
694 }
695
696 kfree((char *) qdisc - qdisc->padded);
697}
698
699void qdisc_destroy(struct Qdisc *qdisc)
700{
701 const struct Qdisc_ops *ops = qdisc->ops;
702
703 if (qdisc->flags & TCQ_F_BUILTIN ||
704 !atomic_dec_and_test(&qdisc->refcnt))
705 return;
706
707#ifdef CONFIG_NET_SCHED
708 qdisc_hash_del(qdisc);
709
710 qdisc_put_stab(rtnl_dereference(qdisc->stab));
711#endif
712 gen_kill_estimator(&qdisc->rate_est);
713 if (ops->reset)
714 ops->reset(qdisc);
715 if (ops->destroy)
716 ops->destroy(qdisc);
717
718 module_put(ops->owner);
719 dev_put(qdisc_dev(qdisc));
720
721 kfree_skb_list(qdisc->gso_skb);
722 kfree_skb(qdisc->skb_bad_txq);
723 /*
724 * gen_estimator est_timer() might access qdisc->q.lock,
725 * wait a RCU grace period before freeing qdisc.
726 */
727 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
728}
729EXPORT_SYMBOL(qdisc_destroy);
730
731/* Attach toplevel qdisc to device queue. */
732struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
733 struct Qdisc *qdisc)
734{
735 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
736 spinlock_t *root_lock;
737
738 root_lock = qdisc_lock(oqdisc);
739 spin_lock_bh(root_lock);
740
741 /* Prune old scheduler */
742 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
743 qdisc_reset(oqdisc);
744
745 /* ... and graft new one */
746 if (qdisc == NULL)
747 qdisc = &noop_qdisc;
748 dev_queue->qdisc_sleeping = qdisc;
749 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
750
751 spin_unlock_bh(root_lock);
752
753 return oqdisc;
754}
755EXPORT_SYMBOL(dev_graft_qdisc);
756
757static void attach_one_default_qdisc(struct net_device *dev,
758 struct netdev_queue *dev_queue,
759 void *_unused)
760{
761 struct Qdisc *qdisc;
762 const struct Qdisc_ops *ops = default_qdisc_ops;
763
764 if (dev->priv_flags & IFF_NO_QUEUE)
765 ops = &noqueue_qdisc_ops;
766
767 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
768 if (!qdisc) {
769 netdev_info(dev, "activation failed\n");
770 return;
771 }
772 if (!netif_is_multiqueue(dev))
773 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
774 dev_queue->qdisc_sleeping = qdisc;
775}
776
777static void attach_default_qdiscs(struct net_device *dev)
778{
779 struct netdev_queue *txq;
780 struct Qdisc *qdisc;
781
782 txq = netdev_get_tx_queue(dev, 0);
783
784 if (!netif_is_multiqueue(dev) ||
785 dev->priv_flags & IFF_NO_QUEUE) {
786 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
787 dev->qdisc = txq->qdisc_sleeping;
788 atomic_inc(&dev->qdisc->refcnt);
789 } else {
790 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
791 if (qdisc) {
792 dev->qdisc = qdisc;
793 qdisc->ops->attach(qdisc);
794 }
795 }
796#ifdef CONFIG_NET_SCHED
797 if (dev->qdisc)
798 qdisc_hash_add(dev->qdisc);
799#endif
800}
801
802static void transition_one_qdisc(struct net_device *dev,
803 struct netdev_queue *dev_queue,
804 void *_need_watchdog)
805{
806 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
807 int *need_watchdog_p = _need_watchdog;
808
809 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
810 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
811
812 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
813 if (need_watchdog_p) {
814 dev_queue->trans_start = 0;
815 *need_watchdog_p = 1;
816 }
817}
818
819void dev_activate(struct net_device *dev)
820{
821 int need_watchdog;
822
823 /* No queueing discipline is attached to device;
824 * create default one for devices, which need queueing
825 * and noqueue_qdisc for virtual interfaces
826 */
827
828 if (dev->qdisc == &noop_qdisc)
829 attach_default_qdiscs(dev);
830
831 if (!netif_carrier_ok(dev))
832 /* Delay activation until next carrier-on event */
833 return;
834
835 need_watchdog = 0;
836 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
837 if (dev_ingress_queue(dev))
838 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
839
840 if (need_watchdog) {
841 netif_trans_update(dev);
842 dev_watchdog_up(dev);
843 }
844}
845EXPORT_SYMBOL(dev_activate);
846
847static void dev_deactivate_queue(struct net_device *dev,
848 struct netdev_queue *dev_queue,
849 void *_qdisc_default)
850{
851 struct Qdisc *qdisc_default = _qdisc_default;
852 struct Qdisc *qdisc;
853
854 qdisc = rtnl_dereference(dev_queue->qdisc);
855 if (qdisc) {
856 spin_lock_bh(qdisc_lock(qdisc));
857
858 if (!(qdisc->flags & TCQ_F_BUILTIN))
859 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
860
861 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
862 qdisc_reset(qdisc);
863
864 spin_unlock_bh(qdisc_lock(qdisc));
865 }
866}
867
868static bool some_qdisc_is_busy(struct net_device *dev)
869{
870 unsigned int i;
871
872 for (i = 0; i < dev->num_tx_queues; i++) {
873 struct netdev_queue *dev_queue;
874 spinlock_t *root_lock;
875 struct Qdisc *q;
876 int val;
877
878 dev_queue = netdev_get_tx_queue(dev, i);
879 q = dev_queue->qdisc_sleeping;
880 root_lock = qdisc_lock(q);
881
882 spin_lock_bh(root_lock);
883
884 val = (qdisc_is_running(q) ||
885 test_bit(__QDISC_STATE_SCHED, &q->state));
886
887 spin_unlock_bh(root_lock);
888
889 if (val)
890 return true;
891 }
892 return false;
893}
894
895/**
896 * dev_deactivate_many - deactivate transmissions on several devices
897 * @head: list of devices to deactivate
898 *
899 * This function returns only when all outstanding transmissions
900 * have completed, unless all devices are in dismantle phase.
901 */
902void dev_deactivate_many(struct list_head *head)
903{
904 struct net_device *dev;
905 bool sync_needed = false;
906
907 list_for_each_entry(dev, head, close_list) {
908 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
909 &noop_qdisc);
910 if (dev_ingress_queue(dev))
911 dev_deactivate_queue(dev, dev_ingress_queue(dev),
912 &noop_qdisc);
913
914 dev_watchdog_down(dev);
915 sync_needed |= !dev->dismantle;
916 }
917
918 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
919 * This is avoided if all devices are in dismantle phase :
920 * Caller will call synchronize_net() for us
921 */
922 if (sync_needed)
923 synchronize_net();
924
925 /* Wait for outstanding qdisc_run calls. */
926 list_for_each_entry(dev, head, close_list)
927 while (some_qdisc_is_busy(dev))
928 yield();
929}
930
931void dev_deactivate(struct net_device *dev)
932{
933 LIST_HEAD(single);
934
935 list_add(&dev->close_list, &single);
936 dev_deactivate_many(&single);
937 list_del(&single);
938}
939EXPORT_SYMBOL(dev_deactivate);
940
941static void dev_init_scheduler_queue(struct net_device *dev,
942 struct netdev_queue *dev_queue,
943 void *_qdisc)
944{
945 struct Qdisc *qdisc = _qdisc;
946
947 rcu_assign_pointer(dev_queue->qdisc, qdisc);
948 dev_queue->qdisc_sleeping = qdisc;
949}
950
951void dev_init_scheduler(struct net_device *dev)
952{
953 dev->qdisc = &noop_qdisc;
954 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
955 if (dev_ingress_queue(dev))
956 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
957
958 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
959}
960
961static void shutdown_scheduler_queue(struct net_device *dev,
962 struct netdev_queue *dev_queue,
963 void *_qdisc_default)
964{
965 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
966 struct Qdisc *qdisc_default = _qdisc_default;
967
968 if (qdisc) {
969 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
970 dev_queue->qdisc_sleeping = qdisc_default;
971
972 qdisc_destroy(qdisc);
973 }
974}
975
976void dev_shutdown(struct net_device *dev)
977{
978 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
979 if (dev_ingress_queue(dev))
980 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
981 qdisc_destroy(dev->qdisc);
982 dev->qdisc = &noop_qdisc;
983
984 WARN_ON(timer_pending(&dev->watchdog_timer));
985}
986
987void psched_ratecfg_precompute(struct psched_ratecfg *r,
988 const struct tc_ratespec *conf,
989 u64 rate64)
990{
991 memset(r, 0, sizeof(*r));
992 r->overhead = conf->overhead;
993 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
994 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
995 r->mult = 1;
996 /*
997 * The deal here is to replace a divide by a reciprocal one
998 * in fast path (a reciprocal divide is a multiply and a shift)
999 *
1000 * Normal formula would be :
1001 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1002 *
1003 * We compute mult/shift to use instead :
1004 * time_in_ns = (len * mult) >> shift;
1005 *
1006 * We try to get the highest possible mult value for accuracy,
1007 * but have to make sure no overflows will ever happen.
1008 */
1009 if (r->rate_bytes_ps > 0) {
1010 u64 factor = NSEC_PER_SEC;
1011
1012 for (;;) {
1013 r->mult = div64_u64(factor, r->rate_bytes_ps);
1014 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1015 break;
1016 factor <<= 1;
1017 r->shift++;
1018 }
1019 }
1020}
1021EXPORT_SYMBOL(psched_ratecfg_precompute);