?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ARC: [*defconfig] Reenable soft lock-up detector
[linux/fpc-iii.git] / net / sched / sch_generic.c
blobbf8c81e07c70457f37eeb167764a2aa33be34ace
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 #include <trace/events/qdisc.h>
33
34 /* Qdisc to use by default */
35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36 EXPORT_SYMBOL(default_qdisc_ops);
37
38 /* Main transmission queue. */
39
40 /* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
42 *
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
47 */
48
49 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
50 {
51 q->gso_skb = skb;
52 q->qstats.requeues++;
53 qdisc_qstats_backlog_inc(q, skb);
54 q->q.qlen++; /* it's still part of the queue */
55 __netif_schedule(q);
56
57 return 0;
58 }
59
60 static void try_bulk_dequeue_skb(struct Qdisc *q,
61 struct sk_buff *skb,
62 const struct netdev_queue *txq,
63 int *packets)
64 {
65 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
66
67 while (bytelimit > 0) {
68 struct sk_buff *nskb = q->dequeue(q);
69
70 if (!nskb)
71 break;
72
73 bytelimit -= nskb->len; /* covers GSO len */
74 skb->next = nskb;
75 skb = nskb;
76 (*packets)++; /* GSO counts as one pkt */
77 }
78 skb->next = NULL;
79 }
80
81 /* This variant of try_bulk_dequeue_skb() makes sure
82 * all skbs in the chain are for the same txq
83 */
84 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
85 struct sk_buff *skb,
86 int *packets)
87 {
88 int mapping = skb_get_queue_mapping(skb);
89 struct sk_buff *nskb;
90 int cnt = 0;
91
92 do {
93 nskb = q->dequeue(q);
94 if (!nskb)
95 break;
96 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
97 q->skb_bad_txq = nskb;
98 qdisc_qstats_backlog_inc(q, nskb);
99 q->q.qlen++;
100 break;
101 }
102 skb->next = nskb;
103 skb = nskb;
104 } while (++cnt < 8);
105 (*packets) += cnt;
106 skb->next = NULL;
107 }
108
109 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
110 * A requeued skb (via q->gso_skb) can also be a SKB list.
111 */
112 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
113 int *packets)
114 {
115 struct sk_buff *skb = q->gso_skb;
116 const struct netdev_queue *txq = q->dev_queue;
117
118 *packets = 1;
119 if (unlikely(skb)) {
120 /* skb in gso_skb were already validated */
121 *validate = false;
122 /* check the reason of requeuing without tx lock first */
123 txq = skb_get_tx_queue(txq->dev, skb);
124 if (!netif_xmit_frozen_or_stopped(txq)) {
125 q->gso_skb = NULL;
126 qdisc_qstats_backlog_dec(q, skb);
127 q->q.qlen--;
128 } else
129 skb = NULL;
130 goto trace;
131 }
132 *validate = true;
133 skb = q->skb_bad_txq;
134 if (unlikely(skb)) {
135 /* check the reason of requeuing without tx lock first */
136 txq = skb_get_tx_queue(txq->dev, skb);
137 if (!netif_xmit_frozen_or_stopped(txq)) {
138 q->skb_bad_txq = NULL;
139 qdisc_qstats_backlog_dec(q, skb);
140 q->q.qlen--;
141 goto bulk;
142 }
143 skb = NULL;
144 goto trace;
145 }
146 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
147 !netif_xmit_frozen_or_stopped(txq))
148 skb = q->dequeue(q);
149 if (skb) {
150 bulk:
151 if (qdisc_may_bulk(q))
152 try_bulk_dequeue_skb(q, skb, txq, packets);
153 else
154 try_bulk_dequeue_skb_slow(q, skb, packets);
155 }
156 trace:
157 trace_qdisc_dequeue(q, txq, *packets, skb);
158 return skb;
159 }
160
161 /*
162 * Transmit possibly several skbs, and handle the return status as
163 * required. Owning running seqcount bit guarantees that
164 * only one CPU can execute this function.
165 *
166 * Returns to the caller:
167 * 0 - queue is empty or throttled.
168 * >0 - queue is not empty.
169 */
170 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
171 struct net_device *dev, struct netdev_queue *txq,
172 spinlock_t *root_lock, bool validate)
173 {
174 int ret = NETDEV_TX_BUSY;
175
176 /* And release qdisc */
177 spin_unlock(root_lock);
178
179 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
180 if (validate)
181 skb = validate_xmit_skb_list(skb, dev);
182
183 if (likely(skb)) {
184 HARD_TX_LOCK(dev, txq, smp_processor_id());
185 if (!netif_xmit_frozen_or_stopped(txq))
186 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
187
188 HARD_TX_UNLOCK(dev, txq);
189 } else {
190 spin_lock(root_lock);
191 return qdisc_qlen(q);
192 }
193 spin_lock(root_lock);
194
195 if (dev_xmit_complete(ret)) {
196 /* Driver sent out skb successfully or skb was consumed */
197 ret = qdisc_qlen(q);
198 } else {
199 /* Driver returned NETDEV_TX_BUSY - requeue skb */
200 if (unlikely(ret != NETDEV_TX_BUSY))
201 net_warn_ratelimited("BUG %s code %d qlen %d\n",
202 dev->name, ret, q->q.qlen);
203
204 ret = dev_requeue_skb(skb, q);
205 }
206
207 if (ret && netif_xmit_frozen_or_stopped(txq))
208 ret = 0;
209
210 return ret;
211 }
212
213 /*
214 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
215 *
216 * running seqcount guarantees only one CPU can process
217 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
218 * this queue.
219 *
220 * netif_tx_lock serializes accesses to device driver.
221 *
222 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
223 * if one is grabbed, another must be free.
224 *
225 * Note, that this procedure can be called by a watchdog timer
226 *
227 * Returns to the caller:
228 * 0 - queue is empty or throttled.
229 * >0 - queue is not empty.
230 *
231 */
232 static inline int qdisc_restart(struct Qdisc *q, int *packets)
233 {
234 struct netdev_queue *txq;
235 struct net_device *dev;
236 spinlock_t *root_lock;
237 struct sk_buff *skb;
238 bool validate;
239
240 /* Dequeue packet */
241 skb = dequeue_skb(q, &validate, packets);
242 if (unlikely(!skb))
243 return 0;
244
245 root_lock = qdisc_lock(q);
246 dev = qdisc_dev(q);
247 txq = skb_get_tx_queue(dev, skb);
248
249 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
250 }
251
252 void __qdisc_run(struct Qdisc *q)
253 {
254 int quota = dev_tx_weight;
255 int packets;
256
257 while (qdisc_restart(q, &packets)) {
258 /*
259 * Ordered by possible occurrence: Postpone processing if
260 * 1. we've exceeded packet quota
261 * 2. another process needs the CPU;
262 */
263 quota -= packets;
264 if (quota <= 0 || need_resched()) {
265 __netif_schedule(q);
266 break;
267 }
268 }
269
270 qdisc_run_end(q);
271 }
272
273 unsigned long dev_trans_start(struct net_device *dev)
274 {
275 unsigned long val, res;
276 unsigned int i;
277
278 if (is_vlan_dev(dev))
279 dev = vlan_dev_real_dev(dev);
280 res = netdev_get_tx_queue(dev, 0)->trans_start;
281 for (i = 1; i < dev->num_tx_queues; i++) {
282 val = netdev_get_tx_queue(dev, i)->trans_start;
283 if (val && time_after(val, res))
284 res = val;
285 }
286
287 return res;
288 }
289 EXPORT_SYMBOL(dev_trans_start);
290
291 static void dev_watchdog(unsigned long arg)
292 {
293 struct net_device *dev = (struct net_device *)arg;
294
295 netif_tx_lock(dev);
296 if (!qdisc_tx_is_noop(dev)) {
297 if (netif_device_present(dev) &&
298 netif_running(dev) &&
299 netif_carrier_ok(dev)) {
300 int some_queue_timedout = 0;
301 unsigned int i;
302 unsigned long trans_start;
303
304 for (i = 0; i < dev->num_tx_queues; i++) {
305 struct netdev_queue *txq;
306
307 txq = netdev_get_tx_queue(dev, i);
308 trans_start = txq->trans_start;
309 if (netif_xmit_stopped(txq) &&
310 time_after(jiffies, (trans_start +
311 dev->watchdog_timeo))) {
312 some_queue_timedout = 1;
313 txq->trans_timeout++;
314 break;
315 }
316 }
317
318 if (some_queue_timedout) {
319 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
320 dev->name, netdev_drivername(dev), i);
321 dev->netdev_ops->ndo_tx_timeout(dev);
322 }
323 if (!mod_timer(&dev->watchdog_timer,
324 round_jiffies(jiffies +
325 dev->watchdog_timeo)))
326 dev_hold(dev);
327 }
328 }
329 netif_tx_unlock(dev);
330
331 dev_put(dev);
332 }
333
334 void __netdev_watchdog_up(struct net_device *dev)
335 {
336 if (dev->netdev_ops->ndo_tx_timeout) {
337 if (dev->watchdog_timeo <= 0)
338 dev->watchdog_timeo = 5*HZ;
339 if (!mod_timer(&dev->watchdog_timer,
340 round_jiffies(jiffies + dev->watchdog_timeo)))
341 dev_hold(dev);
342 }
343 }
344
345 static void dev_watchdog_up(struct net_device *dev)
346 {
347 __netdev_watchdog_up(dev);
348 }
349
350 static void dev_watchdog_down(struct net_device *dev)
351 {
352 netif_tx_lock_bh(dev);
353 if (del_timer(&dev->watchdog_timer))
354 dev_put(dev);
355 netif_tx_unlock_bh(dev);
356 }
357
358 /**
359 * netif_carrier_on - set carrier
360 * @dev: network device
361 *
362 * Device has detected that carrier.
363 */
364 void netif_carrier_on(struct net_device *dev)
365 {
366 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
367 if (dev->reg_state == NETREG_UNINITIALIZED)
368 return;
369 atomic_inc(&dev->carrier_changes);
370 linkwatch_fire_event(dev);
371 if (netif_running(dev))
372 __netdev_watchdog_up(dev);
373 }
374 }
375 EXPORT_SYMBOL(netif_carrier_on);
376
377 /**
378 * netif_carrier_off - clear carrier
379 * @dev: network device
380 *
381 * Device has detected loss of carrier.
382 */
383 void netif_carrier_off(struct net_device *dev)
384 {
385 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
386 if (dev->reg_state == NETREG_UNINITIALIZED)
387 return;
388 atomic_inc(&dev->carrier_changes);
389 linkwatch_fire_event(dev);
390 }
391 }
392 EXPORT_SYMBOL(netif_carrier_off);
393
394 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
395 under all circumstances. It is difficult to invent anything faster or
396 cheaper.
397 */
398
399 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
400 struct sk_buff **to_free)
401 {
402 __qdisc_drop(skb, to_free);
403 return NET_XMIT_CN;
404 }
405
406 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
407 {
408 return NULL;
409 }
410
411 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
412 .id = "noop",
413 .priv_size = 0,
414 .enqueue = noop_enqueue,
415 .dequeue = noop_dequeue,
416 .peek = noop_dequeue,
417 .owner = THIS_MODULE,
418 };
419
420 static struct netdev_queue noop_netdev_queue = {
421 .qdisc = &noop_qdisc,
422 .qdisc_sleeping = &noop_qdisc,
423 };
424
425 struct Qdisc noop_qdisc = {
426 .enqueue = noop_enqueue,
427 .dequeue = noop_dequeue,
428 .flags = TCQ_F_BUILTIN,
429 .ops = &noop_qdisc_ops,
430 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
431 .dev_queue = &noop_netdev_queue,
432 .running = SEQCNT_ZERO(noop_qdisc.running),
433 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
434 };
435 EXPORT_SYMBOL(noop_qdisc);
436
437 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
438 {
439 /* register_qdisc() assigns a default of noop_enqueue if unset,
440 * but __dev_queue_xmit() treats noqueue only as such
441 * if this is NULL - so clear it here. */
442 qdisc->enqueue = NULL;
443 return 0;
444 }
445
446 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
447 .id = "noqueue",
448 .priv_size = 0,
449 .init = noqueue_init,
450 .enqueue = noop_enqueue,
451 .dequeue = noop_dequeue,
452 .peek = noop_dequeue,
453 .owner = THIS_MODULE,
454 };
455
456 static const u8 prio2band[TC_PRIO_MAX + 1] = {
457 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
458 };
459
460 /* 3-band FIFO queue: old style, but should be a bit faster than
461 generic prio+fifo combination.
462 */
463
464 #define PFIFO_FAST_BANDS 3
465
466 /*
467 * Private data for a pfifo_fast scheduler containing:
468 * - queues for the three band
469 * - bitmap indicating which of the bands contain skbs
470 */
471 struct pfifo_fast_priv {
472 u32 bitmap;
473 struct qdisc_skb_head q[PFIFO_FAST_BANDS];
474 };
475
476 /*
477 * Convert a bitmap to the first band number where an skb is queued, where:
478 * bitmap=0 means there are no skbs on any band.
479 * bitmap=1 means there is an skb on band 0.
480 * bitmap=7 means there are skbs on all 3 bands, etc.
481 */
482 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
483
484 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
485 int band)
486 {
487 return priv->q + band;
488 }
489
490 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
491 struct sk_buff **to_free)
492 {
493 if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
494 int band = prio2band[skb->priority & TC_PRIO_MAX];
495 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
496 struct qdisc_skb_head *list = band2list(priv, band);
497
498 priv->bitmap |= (1 << band);
499 qdisc->q.qlen++;
500 return __qdisc_enqueue_tail(skb, qdisc, list);
501 }
502
503 return qdisc_drop(skb, qdisc, to_free);
504 }
505
506 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
507 {
508 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
509 int band = bitmap2band[priv->bitmap];
510
511 if (likely(band >= 0)) {
512 struct qdisc_skb_head *qh = band2list(priv, band);
513 struct sk_buff *skb = __qdisc_dequeue_head(qh);
514
515 if (likely(skb != NULL)) {
516 qdisc_qstats_backlog_dec(qdisc, skb);
517 qdisc_bstats_update(qdisc, skb);
518 }
519
520 qdisc->q.qlen--;
521 if (qh->qlen == 0)
522 priv->bitmap &= ~(1 << band);
523
524 return skb;
525 }
526
527 return NULL;
528 }
529
530 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
531 {
532 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
533 int band = bitmap2band[priv->bitmap];
534
535 if (band >= 0) {
536 struct qdisc_skb_head *qh = band2list(priv, band);
537
538 return qh->head;
539 }
540
541 return NULL;
542 }
543
544 static void pfifo_fast_reset(struct Qdisc *qdisc)
545 {
546 int prio;
547 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
548
549 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
550 __qdisc_reset_queue(band2list(priv, prio));
551
552 priv->bitmap = 0;
553 qdisc->qstats.backlog = 0;
554 qdisc->q.qlen = 0;
555 }
556
557 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
558 {
559 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
560
561 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
562 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
563 goto nla_put_failure;
564 return skb->len;
565
566 nla_put_failure:
567 return -1;
568 }
569
570 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
571 {
572 int prio;
573 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
574
575 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
576 qdisc_skb_head_init(band2list(priv, prio));
577
578 /* Can by-pass the queue discipline */
579 qdisc->flags |= TCQ_F_CAN_BYPASS;
580 return 0;
581 }
582
583 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
584 .id = "pfifo_fast",
585 .priv_size = sizeof(struct pfifo_fast_priv),
586 .enqueue = pfifo_fast_enqueue,
587 .dequeue = pfifo_fast_dequeue,
588 .peek = pfifo_fast_peek,
589 .init = pfifo_fast_init,
590 .reset = pfifo_fast_reset,
591 .dump = pfifo_fast_dump,
592 .owner = THIS_MODULE,
593 };
594 EXPORT_SYMBOL(pfifo_fast_ops);
595
596 static struct lock_class_key qdisc_tx_busylock;
597 static struct lock_class_key qdisc_running_key;
598
599 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
600 const struct Qdisc_ops *ops)
601 {
602 void *p;
603 struct Qdisc *sch;
604 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
605 int err = -ENOBUFS;
606 struct net_device *dev = dev_queue->dev;
607
608 p = kzalloc_node(size, GFP_KERNEL,
609 netdev_queue_numa_node_read(dev_queue));
610
611 if (!p)
612 goto errout;
613 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
614 /* if we got non aligned memory, ask more and do alignment ourself */
615 if (sch != p) {
616 kfree(p);
617 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
618 netdev_queue_numa_node_read(dev_queue));
619 if (!p)
620 goto errout;
621 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
622 sch->padded = (char *) sch - (char *) p;
623 }
624 qdisc_skb_head_init(&sch->q);
625 spin_lock_init(&sch->q.lock);
626
627 spin_lock_init(&sch->busylock);
628 lockdep_set_class(&sch->busylock,
629 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
630
631 seqcount_init(&sch->running);
632 lockdep_set_class(&sch->running,
633 dev->qdisc_running_key ?: &qdisc_running_key);
634
635 sch->ops = ops;
636 sch->enqueue = ops->enqueue;
637 sch->dequeue = ops->dequeue;
638 sch->dev_queue = dev_queue;
639 dev_hold(dev);
640 refcount_set(&sch->refcnt, 1);
641
642 return sch;
643 errout:
644 return ERR_PTR(err);
645 }
646
647 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
648 const struct Qdisc_ops *ops,
649 unsigned int parentid)
650 {
651 struct Qdisc *sch;
652
653 if (!try_module_get(ops->owner))
654 return NULL;
655
656 sch = qdisc_alloc(dev_queue, ops);
657 if (IS_ERR(sch)) {
658 module_put(ops->owner);
659 return NULL;
660 }
661 sch->parent = parentid;
662
663 if (!ops->init || ops->init(sch, NULL) == 0)
664 return sch;
665
666 qdisc_destroy(sch);
667 return NULL;
668 }
669 EXPORT_SYMBOL(qdisc_create_dflt);
670
671 /* Under qdisc_lock(qdisc) and BH! */
672
673 void qdisc_reset(struct Qdisc *qdisc)
674 {
675 const struct Qdisc_ops *ops = qdisc->ops;
676
677 if (ops->reset)
678 ops->reset(qdisc);
679
680 kfree_skb(qdisc->skb_bad_txq);
681 qdisc->skb_bad_txq = NULL;
682
683 if (qdisc->gso_skb) {
684 kfree_skb_list(qdisc->gso_skb);
685 qdisc->gso_skb = NULL;
686 }
687 qdisc->q.qlen = 0;
688 qdisc->qstats.backlog = 0;
689 }
690 EXPORT_SYMBOL(qdisc_reset);
691
692 static void qdisc_rcu_free(struct rcu_head *head)
693 {
694 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
695
696 if (qdisc_is_percpu_stats(qdisc)) {
697 free_percpu(qdisc->cpu_bstats);
698 free_percpu(qdisc->cpu_qstats);
699 }
700
701 kfree((char *) qdisc - qdisc->padded);
702 }
703
704 void qdisc_destroy(struct Qdisc *qdisc)
705 {
706 const struct Qdisc_ops *ops = qdisc->ops;
707
708 if (qdisc->flags & TCQ_F_BUILTIN ||
709 !refcount_dec_and_test(&qdisc->refcnt))
710 return;
711
712 #ifdef CONFIG_NET_SCHED
713 qdisc_hash_del(qdisc);
714
715 qdisc_put_stab(rtnl_dereference(qdisc->stab));
716 #endif
717 gen_kill_estimator(&qdisc->rate_est);
718 if (ops->reset)
719 ops->reset(qdisc);
720 if (ops->destroy)
721 ops->destroy(qdisc);
722
723 module_put(ops->owner);
724 dev_put(qdisc_dev(qdisc));
725
726 kfree_skb_list(qdisc->gso_skb);
727 kfree_skb(qdisc->skb_bad_txq);
728 /*
729 * gen_estimator est_timer() might access qdisc->q.lock,
730 * wait a RCU grace period before freeing qdisc.
731 */
732 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
733 }
734 EXPORT_SYMBOL(qdisc_destroy);
735
736 /* Attach toplevel qdisc to device queue. */
737 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
738 struct Qdisc *qdisc)
739 {
740 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
741 spinlock_t *root_lock;
742
743 root_lock = qdisc_lock(oqdisc);
744 spin_lock_bh(root_lock);
745
746 /* Prune old scheduler */
747 if (oqdisc && refcount_read(&oqdisc->refcnt) <= 1)
748 qdisc_reset(oqdisc);
749
750 /* ... and graft new one */
751 if (qdisc == NULL)
752 qdisc = &noop_qdisc;
753 dev_queue->qdisc_sleeping = qdisc;
754 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
755
756 spin_unlock_bh(root_lock);
757
758 return oqdisc;
759 }
760 EXPORT_SYMBOL(dev_graft_qdisc);
761
762 static void attach_one_default_qdisc(struct net_device *dev,
763 struct netdev_queue *dev_queue,
764 void *_unused)
765 {
766 struct Qdisc *qdisc;
767 const struct Qdisc_ops *ops = default_qdisc_ops;
768
769 if (dev->priv_flags & IFF_NO_QUEUE)
770 ops = &noqueue_qdisc_ops;
771
772 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
773 if (!qdisc) {
774 netdev_info(dev, "activation failed\n");
775 return;
776 }
777 if (!netif_is_multiqueue(dev))
778 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
779 dev_queue->qdisc_sleeping = qdisc;
780 }
781
782 static void attach_default_qdiscs(struct net_device *dev)
783 {
784 struct netdev_queue *txq;
785 struct Qdisc *qdisc;
786
787 txq = netdev_get_tx_queue(dev, 0);
788
789 if (!netif_is_multiqueue(dev) ||
790 dev->priv_flags & IFF_NO_QUEUE) {
791 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
792 dev->qdisc = txq->qdisc_sleeping;
793 qdisc_refcount_inc(dev->qdisc);
794 } else {
795 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
796 if (qdisc) {
797 dev->qdisc = qdisc;
798 qdisc->ops->attach(qdisc);
799 }
800 }
801 #ifdef CONFIG_NET_SCHED
802 if (dev->qdisc != &noop_qdisc)
803 qdisc_hash_add(dev->qdisc, false);
804 #endif
805 }
806
807 static void transition_one_qdisc(struct net_device *dev,
808 struct netdev_queue *dev_queue,
809 void *_need_watchdog)
810 {
811 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
812 int *need_watchdog_p = _need_watchdog;
813
814 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
815 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
816
817 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
818 if (need_watchdog_p) {
819 dev_queue->trans_start = 0;
820 *need_watchdog_p = 1;
821 }
822 }
823
824 void dev_activate(struct net_device *dev)
825 {
826 int need_watchdog;
827
828 /* No queueing discipline is attached to device;
829 * create default one for devices, which need queueing
830 * and noqueue_qdisc for virtual interfaces
831 */
832
833 if (dev->qdisc == &noop_qdisc)
834 attach_default_qdiscs(dev);
835
836 if (!netif_carrier_ok(dev))
837 /* Delay activation until next carrier-on event */
838 return;
839
840 need_watchdog = 0;
841 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
842 if (dev_ingress_queue(dev))
843 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
844
845 if (need_watchdog) {
846 netif_trans_update(dev);
847 dev_watchdog_up(dev);
848 }
849 }
850 EXPORT_SYMBOL(dev_activate);
851
852 static void dev_deactivate_queue(struct net_device *dev,
853 struct netdev_queue *dev_queue,
854 void *_qdisc_default)
855 {
856 struct Qdisc *qdisc_default = _qdisc_default;
857 struct Qdisc *qdisc;
858
859 qdisc = rtnl_dereference(dev_queue->qdisc);
860 if (qdisc) {
861 spin_lock_bh(qdisc_lock(qdisc));
862
863 if (!(qdisc->flags & TCQ_F_BUILTIN))
864 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
865
866 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
867 qdisc_reset(qdisc);
868
869 spin_unlock_bh(qdisc_lock(qdisc));
870 }
871 }
872
873 static bool some_qdisc_is_busy(struct net_device *dev)
874 {
875 unsigned int i;
876
877 for (i = 0; i < dev->num_tx_queues; i++) {
878 struct netdev_queue *dev_queue;
879 spinlock_t *root_lock;
880 struct Qdisc *q;
881 int val;
882
883 dev_queue = netdev_get_tx_queue(dev, i);
884 q = dev_queue->qdisc_sleeping;
885 root_lock = qdisc_lock(q);
886
887 spin_lock_bh(root_lock);
888
889 val = (qdisc_is_running(q) ||
890 test_bit(__QDISC_STATE_SCHED, &q->state));
891
892 spin_unlock_bh(root_lock);
893
894 if (val)
895 return true;
896 }
897 return false;
898 }
899
900 /**
901 * dev_deactivate_many - deactivate transmissions on several devices
902 * @head: list of devices to deactivate
903 *
904 * This function returns only when all outstanding transmissions
905 * have completed, unless all devices are in dismantle phase.
906 */
907 void dev_deactivate_many(struct list_head *head)
908 {
909 struct net_device *dev;
910 bool sync_needed = false;
911
912 list_for_each_entry(dev, head, close_list) {
913 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
914 &noop_qdisc);
915 if (dev_ingress_queue(dev))
916 dev_deactivate_queue(dev, dev_ingress_queue(dev),
917 &noop_qdisc);
918
919 dev_watchdog_down(dev);
920 sync_needed |= !dev->dismantle;
921 }
922
923 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
924 * This is avoided if all devices are in dismantle phase :
925 * Caller will call synchronize_net() for us
926 */
927 if (sync_needed)
928 synchronize_net();
929
930 /* Wait for outstanding qdisc_run calls. */
931 list_for_each_entry(dev, head, close_list)
932 while (some_qdisc_is_busy(dev))
933 yield();
934 }
935
936 void dev_deactivate(struct net_device *dev)
937 {
938 LIST_HEAD(single);
939
940 list_add(&dev->close_list, &single);
941 dev_deactivate_many(&single);
942 list_del(&single);
943 }
944 EXPORT_SYMBOL(dev_deactivate);
945
946 static void dev_init_scheduler_queue(struct net_device *dev,
947 struct netdev_queue *dev_queue,
948 void *_qdisc)
949 {
950 struct Qdisc *qdisc = _qdisc;
951
952 rcu_assign_pointer(dev_queue->qdisc, qdisc);
953 dev_queue->qdisc_sleeping = qdisc;
954 }
955
956 void dev_init_scheduler(struct net_device *dev)
957 {
958 dev->qdisc = &noop_qdisc;
959 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
960 if (dev_ingress_queue(dev))
961 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
962
963 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
964 }
965
966 static void shutdown_scheduler_queue(struct net_device *dev,
967 struct netdev_queue *dev_queue,
968 void *_qdisc_default)
969 {
970 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
971 struct Qdisc *qdisc_default = _qdisc_default;
972
973 if (qdisc) {
974 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
975 dev_queue->qdisc_sleeping = qdisc_default;
976
977 qdisc_destroy(qdisc);
978 }
979 }
980
981 void dev_shutdown(struct net_device *dev)
982 {
983 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
984 if (dev_ingress_queue(dev))
985 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
986 qdisc_destroy(dev->qdisc);
987 dev->qdisc = &noop_qdisc;
988
989 WARN_ON(timer_pending(&dev->watchdog_timer));
990 }
991
992 void psched_ratecfg_precompute(struct psched_ratecfg *r,
993 const struct tc_ratespec *conf,
994 u64 rate64)
995 {
996 memset(r, 0, sizeof(*r));
997 r->overhead = conf->overhead;
998 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
999 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1000 r->mult = 1;
1001 /*
1002 * The deal here is to replace a divide by a reciprocal one
1003 * in fast path (a reciprocal divide is a multiply and a shift)
1004 *
1005 * Normal formula would be :
1006 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1007 *
1008 * We compute mult/shift to use instead :
1009 * time_in_ns = (len * mult) >> shift;
1010 *
1011 * We try to get the highest possible mult value for accuracy,
1012 * but have to make sure no overflows will ever happen.
1013 */
1014 if (r->rate_bytes_ps > 0) {
1015 u64 factor = NSEC_PER_SEC;
1016
1017 for (;;) {
1018 r->mult = div64_u64(factor, r->rate_bytes_ps);
1019 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1020 break;
1021 factor <<= 1;
1022 r->shift++;
1023 }
1024 }
1025 }
1026 EXPORT_SYMBOL(psched_ratecfg_precompute);
Linux with added FPC-III support