?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.16-rc1
[cris-mirror.git] / net / sched / sch_generic.c
blob190570f21b208d5a17943360a3a6f85e1c2a2187
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 <linux/skb_array.h>
30 #include <linux/if_macvlan.h>
31 #include <net/sch_generic.h>
32 #include <net/pkt_sched.h>
33 #include <net/dst.h>
34 #include <trace/events/qdisc.h>
35 #include <net/xfrm.h>
36
37 /* Qdisc to use by default */
38 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
39 EXPORT_SYMBOL(default_qdisc_ops);
40
41 /* Main transmission queue. */
42
43 /* Modifications to data participating in scheduling must be protected with
44 * qdisc_lock(qdisc) spinlock.
45 *
46 * The idea is the following:
47 * - enqueue, dequeue are serialized via qdisc root lock
48 * - ingress filtering is also serialized via qdisc root lock
49 * - updates to tree and tree walking are only done under the rtnl mutex.
50 */
51
52 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53 {
54 const struct netdev_queue *txq = q->dev_queue;
55 spinlock_t *lock = NULL;
56 struct sk_buff *skb;
57
58 if (q->flags & TCQ_F_NOLOCK) {
59 lock = qdisc_lock(q);
60 spin_lock(lock);
61 }
62
63 skb = skb_peek(&q->skb_bad_txq);
64 if (skb) {
65 /* check the reason of requeuing without tx lock first */
66 txq = skb_get_tx_queue(txq->dev, skb);
67 if (!netif_xmit_frozen_or_stopped(txq)) {
68 skb = __skb_dequeue(&q->skb_bad_txq);
69 if (qdisc_is_percpu_stats(q)) {
70 qdisc_qstats_cpu_backlog_dec(q, skb);
71 qdisc_qstats_cpu_qlen_dec(q);
72 } else {
73 qdisc_qstats_backlog_dec(q, skb);
74 q->q.qlen--;
75 }
76 } else {
77 skb = NULL;
78 }
79 }
80
81 if (lock)
82 spin_unlock(lock);
83
84 return skb;
85 }
86
87 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88 {
89 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
90
91 if (unlikely(skb))
92 skb = __skb_dequeue_bad_txq(q);
93
94 return skb;
95 }
96
97 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
98 struct sk_buff *skb)
99 {
100 spinlock_t *lock = NULL;
101
102 if (q->flags & TCQ_F_NOLOCK) {
103 lock = qdisc_lock(q);
104 spin_lock(lock);
105 }
106
107 __skb_queue_tail(&q->skb_bad_txq, skb);
108
109 if (lock)
110 spin_unlock(lock);
111 }
112
113 static inline int __dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
114 {
115 while (skb) {
116 struct sk_buff *next = skb->next;
117
118 __skb_queue_tail(&q->gso_skb, skb);
119 q->qstats.requeues++;
120 qdisc_qstats_backlog_inc(q, skb);
121 q->q.qlen++; /* it's still part of the queue */
122
123 skb = next;
124 }
125 __netif_schedule(q);
126
127 return 0;
128 }
129
130 static inline int dev_requeue_skb_locked(struct sk_buff *skb, struct Qdisc *q)
131 {
132 spinlock_t *lock = qdisc_lock(q);
133
134 spin_lock(lock);
135 while (skb) {
136 struct sk_buff *next = skb->next;
137
138 __skb_queue_tail(&q->gso_skb, skb);
139
140 qdisc_qstats_cpu_requeues_inc(q);
141 qdisc_qstats_cpu_backlog_inc(q, skb);
142 qdisc_qstats_cpu_qlen_inc(q);
143
144 skb = next;
145 }
146 spin_unlock(lock);
147
148 __netif_schedule(q);
149
150 return 0;
151 }
152
153 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
154 {
155 if (q->flags & TCQ_F_NOLOCK)
156 return dev_requeue_skb_locked(skb, q);
157 else
158 return __dev_requeue_skb(skb, q);
159 }
160
161 static void try_bulk_dequeue_skb(struct Qdisc *q,
162 struct sk_buff *skb,
163 const struct netdev_queue *txq,
164 int *packets)
165 {
166 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
167
168 while (bytelimit > 0) {
169 struct sk_buff *nskb = q->dequeue(q);
170
171 if (!nskb)
172 break;
173
174 bytelimit -= nskb->len; /* covers GSO len */
175 skb->next = nskb;
176 skb = nskb;
177 (*packets)++; /* GSO counts as one pkt */
178 }
179 skb->next = NULL;
180 }
181
182 /* This variant of try_bulk_dequeue_skb() makes sure
183 * all skbs in the chain are for the same txq
184 */
185 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
186 struct sk_buff *skb,
187 int *packets)
188 {
189 int mapping = skb_get_queue_mapping(skb);
190 struct sk_buff *nskb;
191 int cnt = 0;
192
193 do {
194 nskb = q->dequeue(q);
195 if (!nskb)
196 break;
197 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
198 qdisc_enqueue_skb_bad_txq(q, nskb);
199
200 if (qdisc_is_percpu_stats(q)) {
201 qdisc_qstats_cpu_backlog_inc(q, nskb);
202 qdisc_qstats_cpu_qlen_inc(q);
203 } else {
204 qdisc_qstats_backlog_inc(q, nskb);
205 q->q.qlen++;
206 }
207 break;
208 }
209 skb->next = nskb;
210 skb = nskb;
211 } while (++cnt < 8);
212 (*packets) += cnt;
213 skb->next = NULL;
214 }
215
216 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
217 * A requeued skb (via q->gso_skb) can also be a SKB list.
218 */
219 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
220 int *packets)
221 {
222 const struct netdev_queue *txq = q->dev_queue;
223 struct sk_buff *skb = NULL;
224
225 *packets = 1;
226 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
227 spinlock_t *lock = NULL;
228
229 if (q->flags & TCQ_F_NOLOCK) {
230 lock = qdisc_lock(q);
231 spin_lock(lock);
232 }
233
234 skb = skb_peek(&q->gso_skb);
235
236 /* skb may be null if another cpu pulls gso_skb off in between
237 * empty check and lock.
238 */
239 if (!skb) {
240 if (lock)
241 spin_unlock(lock);
242 goto validate;
243 }
244
245 /* skb in gso_skb were already validated */
246 *validate = false;
247 if (xfrm_offload(skb))
248 *validate = true;
249 /* check the reason of requeuing without tx lock first */
250 txq = skb_get_tx_queue(txq->dev, skb);
251 if (!netif_xmit_frozen_or_stopped(txq)) {
252 skb = __skb_dequeue(&q->gso_skb);
253 if (qdisc_is_percpu_stats(q)) {
254 qdisc_qstats_cpu_backlog_dec(q, skb);
255 qdisc_qstats_cpu_qlen_dec(q);
256 } else {
257 qdisc_qstats_backlog_dec(q, skb);
258 q->q.qlen--;
259 }
260 } else {
261 skb = NULL;
262 }
263 if (lock)
264 spin_unlock(lock);
265 goto trace;
266 }
267 validate:
268 *validate = true;
269
270 if ((q->flags & TCQ_F_ONETXQUEUE) &&
271 netif_xmit_frozen_or_stopped(txq))
272 return skb;
273
274 skb = qdisc_dequeue_skb_bad_txq(q);
275 if (unlikely(skb))
276 goto bulk;
277 skb = q->dequeue(q);
278 if (skb) {
279 bulk:
280 if (qdisc_may_bulk(q))
281 try_bulk_dequeue_skb(q, skb, txq, packets);
282 else
283 try_bulk_dequeue_skb_slow(q, skb, packets);
284 }
285 trace:
286 trace_qdisc_dequeue(q, txq, *packets, skb);
287 return skb;
288 }
289
290 /*
291 * Transmit possibly several skbs, and handle the return status as
292 * required. Owning running seqcount bit guarantees that
293 * only one CPU can execute this function.
294 *
295 * Returns to the caller:
296 * false - hardware queue frozen backoff
297 * true - feel free to send more pkts
298 */
299 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
300 struct net_device *dev, struct netdev_queue *txq,
301 spinlock_t *root_lock, bool validate)
302 {
303 int ret = NETDEV_TX_BUSY;
304 bool again = false;
305
306 /* And release qdisc */
307 if (root_lock)
308 spin_unlock(root_lock);
309
310 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
311 if (validate)
312 skb = validate_xmit_skb_list(skb, dev, &again);
313
314 #ifdef CONFIG_XFRM_OFFLOAD
315 if (unlikely(again)) {
316 if (root_lock)
317 spin_lock(root_lock);
318
319 dev_requeue_skb(skb, q);
320 return false;
321 }
322 #endif
323
324 if (likely(skb)) {
325 HARD_TX_LOCK(dev, txq, smp_processor_id());
326 if (!netif_xmit_frozen_or_stopped(txq))
327 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
328
329 HARD_TX_UNLOCK(dev, txq);
330 } else {
331 if (root_lock)
332 spin_lock(root_lock);
333 return true;
334 }
335
336 if (root_lock)
337 spin_lock(root_lock);
338
339 if (!dev_xmit_complete(ret)) {
340 /* Driver returned NETDEV_TX_BUSY - requeue skb */
341 if (unlikely(ret != NETDEV_TX_BUSY))
342 net_warn_ratelimited("BUG %s code %d qlen %d\n",
343 dev->name, ret, q->q.qlen);
344
345 dev_requeue_skb(skb, q);
346 return false;
347 }
348
349 if (ret && netif_xmit_frozen_or_stopped(txq))
350 return false;
351
352 return true;
353 }
354
355 /*
356 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
357 *
358 * running seqcount guarantees only one CPU can process
359 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
360 * this queue.
361 *
362 * netif_tx_lock serializes accesses to device driver.
363 *
364 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
365 * if one is grabbed, another must be free.
366 *
367 * Note, that this procedure can be called by a watchdog timer
368 *
369 * Returns to the caller:
370 * 0 - queue is empty or throttled.
371 * >0 - queue is not empty.
372 *
373 */
374 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
375 {
376 spinlock_t *root_lock = NULL;
377 struct netdev_queue *txq;
378 struct net_device *dev;
379 struct sk_buff *skb;
380 bool validate;
381
382 /* Dequeue packet */
383 skb = dequeue_skb(q, &validate, packets);
384 if (unlikely(!skb))
385 return false;
386
387 if (!(q->flags & TCQ_F_NOLOCK))
388 root_lock = qdisc_lock(q);
389
390 dev = qdisc_dev(q);
391 txq = skb_get_tx_queue(dev, skb);
392
393 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
394 }
395
396 void __qdisc_run(struct Qdisc *q)
397 {
398 int quota = dev_tx_weight;
399 int packets;
400
401 while (qdisc_restart(q, &packets)) {
402 /*
403 * Ordered by possible occurrence: Postpone processing if
404 * 1. we've exceeded packet quota
405 * 2. another process needs the CPU;
406 */
407 quota -= packets;
408 if (quota <= 0 || need_resched()) {
409 __netif_schedule(q);
410 break;
411 }
412 }
413 }
414
415 unsigned long dev_trans_start(struct net_device *dev)
416 {
417 unsigned long val, res;
418 unsigned int i;
419
420 if (is_vlan_dev(dev))
421 dev = vlan_dev_real_dev(dev);
422 else if (netif_is_macvlan(dev))
423 dev = macvlan_dev_real_dev(dev);
424 res = netdev_get_tx_queue(dev, 0)->trans_start;
425 for (i = 1; i < dev->num_tx_queues; i++) {
426 val = netdev_get_tx_queue(dev, i)->trans_start;
427 if (val && time_after(val, res))
428 res = val;
429 }
430
431 return res;
432 }
433 EXPORT_SYMBOL(dev_trans_start);
434
435 static void dev_watchdog(struct timer_list *t)
436 {
437 struct net_device *dev = from_timer(dev, t, watchdog_timer);
438
439 netif_tx_lock(dev);
440 if (!qdisc_tx_is_noop(dev)) {
441 if (netif_device_present(dev) &&
442 netif_running(dev) &&
443 netif_carrier_ok(dev)) {
444 int some_queue_timedout = 0;
445 unsigned int i;
446 unsigned long trans_start;
447
448 for (i = 0; i < dev->num_tx_queues; i++) {
449 struct netdev_queue *txq;
450
451 txq = netdev_get_tx_queue(dev, i);
452 trans_start = txq->trans_start;
453 if (netif_xmit_stopped(txq) &&
454 time_after(jiffies, (trans_start +
455 dev->watchdog_timeo))) {
456 some_queue_timedout = 1;
457 txq->trans_timeout++;
458 break;
459 }
460 }
461
462 if (some_queue_timedout) {
463 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
464 dev->name, netdev_drivername(dev), i);
465 dev->netdev_ops->ndo_tx_timeout(dev);
466 }
467 if (!mod_timer(&dev->watchdog_timer,
468 round_jiffies(jiffies +
469 dev->watchdog_timeo)))
470 dev_hold(dev);
471 }
472 }
473 netif_tx_unlock(dev);
474
475 dev_put(dev);
476 }
477
478 void __netdev_watchdog_up(struct net_device *dev)
479 {
480 if (dev->netdev_ops->ndo_tx_timeout) {
481 if (dev->watchdog_timeo <= 0)
482 dev->watchdog_timeo = 5*HZ;
483 if (!mod_timer(&dev->watchdog_timer,
484 round_jiffies(jiffies + dev->watchdog_timeo)))
485 dev_hold(dev);
486 }
487 }
488
489 static void dev_watchdog_up(struct net_device *dev)
490 {
491 __netdev_watchdog_up(dev);
492 }
493
494 static void dev_watchdog_down(struct net_device *dev)
495 {
496 netif_tx_lock_bh(dev);
497 if (del_timer(&dev->watchdog_timer))
498 dev_put(dev);
499 netif_tx_unlock_bh(dev);
500 }
501
502 /**
503 * netif_carrier_on - set carrier
504 * @dev: network device
505 *
506 * Device has detected that carrier.
507 */
508 void netif_carrier_on(struct net_device *dev)
509 {
510 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
511 if (dev->reg_state == NETREG_UNINITIALIZED)
512 return;
513 atomic_inc(&dev->carrier_up_count);
514 linkwatch_fire_event(dev);
515 if (netif_running(dev))
516 __netdev_watchdog_up(dev);
517 }
518 }
519 EXPORT_SYMBOL(netif_carrier_on);
520
521 /**
522 * netif_carrier_off - clear carrier
523 * @dev: network device
524 *
525 * Device has detected loss of carrier.
526 */
527 void netif_carrier_off(struct net_device *dev)
528 {
529 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
530 if (dev->reg_state == NETREG_UNINITIALIZED)
531 return;
532 atomic_inc(&dev->carrier_down_count);
533 linkwatch_fire_event(dev);
534 }
535 }
536 EXPORT_SYMBOL(netif_carrier_off);
537
538 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
539 under all circumstances. It is difficult to invent anything faster or
540 cheaper.
541 */
542
543 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
544 struct sk_buff **to_free)
545 {
546 __qdisc_drop(skb, to_free);
547 return NET_XMIT_CN;
548 }
549
550 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
551 {
552 return NULL;
553 }
554
555 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
556 .id = "noop",
557 .priv_size = 0,
558 .enqueue = noop_enqueue,
559 .dequeue = noop_dequeue,
560 .peek = noop_dequeue,
561 .owner = THIS_MODULE,
562 };
563
564 static struct netdev_queue noop_netdev_queue = {
565 .qdisc = &noop_qdisc,
566 .qdisc_sleeping = &noop_qdisc,
567 };
568
569 struct Qdisc noop_qdisc = {
570 .enqueue = noop_enqueue,
571 .dequeue = noop_dequeue,
572 .flags = TCQ_F_BUILTIN,
573 .ops = &noop_qdisc_ops,
574 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
575 .dev_queue = &noop_netdev_queue,
576 .running = SEQCNT_ZERO(noop_qdisc.running),
577 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
578 };
579 EXPORT_SYMBOL(noop_qdisc);
580
581 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
582 struct netlink_ext_ack *extack)
583 {
584 /* register_qdisc() assigns a default of noop_enqueue if unset,
585 * but __dev_queue_xmit() treats noqueue only as such
586 * if this is NULL - so clear it here. */
587 qdisc->enqueue = NULL;
588 return 0;
589 }
590
591 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
592 .id = "noqueue",
593 .priv_size = 0,
594 .init = noqueue_init,
595 .enqueue = noop_enqueue,
596 .dequeue = noop_dequeue,
597 .peek = noop_dequeue,
598 .owner = THIS_MODULE,
599 };
600
601 static const u8 prio2band[TC_PRIO_MAX + 1] = {
602 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
603 };
604
605 /* 3-band FIFO queue: old style, but should be a bit faster than
606 generic prio+fifo combination.
607 */
608
609 #define PFIFO_FAST_BANDS 3
610
611 /*
612 * Private data for a pfifo_fast scheduler containing:
613 * - rings for priority bands
614 */
615 struct pfifo_fast_priv {
616 struct skb_array q[PFIFO_FAST_BANDS];
617 };
618
619 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
620 int band)
621 {
622 return &priv->q[band];
623 }
624
625 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
626 struct sk_buff **to_free)
627 {
628 int band = prio2band[skb->priority & TC_PRIO_MAX];
629 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
630 struct skb_array *q = band2list(priv, band);
631 int err;
632
633 err = skb_array_produce(q, skb);
634
635 if (unlikely(err))
636 return qdisc_drop_cpu(skb, qdisc, to_free);
637
638 qdisc_qstats_cpu_qlen_inc(qdisc);
639 qdisc_qstats_cpu_backlog_inc(qdisc, skb);
640 return NET_XMIT_SUCCESS;
641 }
642
643 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
644 {
645 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
646 struct sk_buff *skb = NULL;
647 int band;
648
649 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
650 struct skb_array *q = band2list(priv, band);
651
652 if (__skb_array_empty(q))
653 continue;
654
655 skb = skb_array_consume_bh(q);
656 }
657 if (likely(skb)) {
658 qdisc_qstats_cpu_backlog_dec(qdisc, skb);
659 qdisc_bstats_cpu_update(qdisc, skb);
660 qdisc_qstats_cpu_qlen_dec(qdisc);
661 }
662
663 return skb;
664 }
665
666 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
667 {
668 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
669 struct sk_buff *skb = NULL;
670 int band;
671
672 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
673 struct skb_array *q = band2list(priv, band);
674
675 skb = __skb_array_peek(q);
676 }
677
678 return skb;
679 }
680
681 static void pfifo_fast_reset(struct Qdisc *qdisc)
682 {
683 int i, band;
684 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
685
686 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
687 struct skb_array *q = band2list(priv, band);
688 struct sk_buff *skb;
689
690 /* NULL ring is possible if destroy path is due to a failed
691 * skb_array_init() in pfifo_fast_init() case.
692 */
693 if (!q->ring.queue)
694 continue;
695
696 while ((skb = skb_array_consume_bh(q)) != NULL)
697 kfree_skb(skb);
698 }
699
700 for_each_possible_cpu(i) {
701 struct gnet_stats_queue *q = per_cpu_ptr(qdisc->cpu_qstats, i);
702
703 q->backlog = 0;
704 q->qlen = 0;
705 }
706 }
707
708 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
709 {
710 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
711
712 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
713 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
714 goto nla_put_failure;
715 return skb->len;
716
717 nla_put_failure:
718 return -1;
719 }
720
721 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
722 struct netlink_ext_ack *extack)
723 {
724 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
725 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
726 int prio;
727
728 /* guard against zero length rings */
729 if (!qlen)
730 return -EINVAL;
731
732 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
733 struct skb_array *q = band2list(priv, prio);
734 int err;
735
736 err = skb_array_init(q, qlen, GFP_KERNEL);
737 if (err)
738 return -ENOMEM;
739 }
740
741 /* Can by-pass the queue discipline */
742 qdisc->flags |= TCQ_F_CAN_BYPASS;
743 return 0;
744 }
745
746 static void pfifo_fast_destroy(struct Qdisc *sch)
747 {
748 struct pfifo_fast_priv *priv = qdisc_priv(sch);
749 int prio;
750
751 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
752 struct skb_array *q = band2list(priv, prio);
753
754 /* NULL ring is possible if destroy path is due to a failed
755 * skb_array_init() in pfifo_fast_init() case.
756 */
757 if (!q->ring.queue)
758 continue;
759 /* Destroy ring but no need to kfree_skb because a call to
760 * pfifo_fast_reset() has already done that work.
761 */
762 ptr_ring_cleanup(&q->ring, NULL);
763 }
764 }
765
766 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
767 unsigned int new_len)
768 {
769 struct pfifo_fast_priv *priv = qdisc_priv(sch);
770 struct skb_array *bands[PFIFO_FAST_BANDS];
771 int prio;
772
773 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
774 struct skb_array *q = band2list(priv, prio);
775
776 bands[prio] = q;
777 }
778
779 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
780 GFP_KERNEL);
781 }
782
783 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
784 .id = "pfifo_fast",
785 .priv_size = sizeof(struct pfifo_fast_priv),
786 .enqueue = pfifo_fast_enqueue,
787 .dequeue = pfifo_fast_dequeue,
788 .peek = pfifo_fast_peek,
789 .init = pfifo_fast_init,
790 .destroy = pfifo_fast_destroy,
791 .reset = pfifo_fast_reset,
792 .dump = pfifo_fast_dump,
793 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
794 .owner = THIS_MODULE,
795 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
796 };
797 EXPORT_SYMBOL(pfifo_fast_ops);
798
799 static struct lock_class_key qdisc_tx_busylock;
800 static struct lock_class_key qdisc_running_key;
801
802 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
803 const struct Qdisc_ops *ops,
804 struct netlink_ext_ack *extack)
805 {
806 void *p;
807 struct Qdisc *sch;
808 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
809 int err = -ENOBUFS;
810 struct net_device *dev;
811
812 if (!dev_queue) {
813 NL_SET_ERR_MSG(extack, "No device queue given");
814 err = -EINVAL;
815 goto errout;
816 }
817
818 dev = dev_queue->dev;
819 p = kzalloc_node(size, GFP_KERNEL,
820 netdev_queue_numa_node_read(dev_queue));
821
822 if (!p)
823 goto errout;
824 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
825 /* if we got non aligned memory, ask more and do alignment ourself */
826 if (sch != p) {
827 kfree(p);
828 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
829 netdev_queue_numa_node_read(dev_queue));
830 if (!p)
831 goto errout;
832 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
833 sch->padded = (char *) sch - (char *) p;
834 }
835 __skb_queue_head_init(&sch->gso_skb);
836 __skb_queue_head_init(&sch->skb_bad_txq);
837 qdisc_skb_head_init(&sch->q);
838 spin_lock_init(&sch->q.lock);
839
840 if (ops->static_flags & TCQ_F_CPUSTATS) {
841 sch->cpu_bstats =
842 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
843 if (!sch->cpu_bstats)
844 goto errout1;
845
846 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
847 if (!sch->cpu_qstats) {
848 free_percpu(sch->cpu_bstats);
849 goto errout1;
850 }
851 }
852
853 spin_lock_init(&sch->busylock);
854 lockdep_set_class(&sch->busylock,
855 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
856
857 seqcount_init(&sch->running);
858 lockdep_set_class(&sch->running,
859 dev->qdisc_running_key ?: &qdisc_running_key);
860
861 sch->ops = ops;
862 sch->flags = ops->static_flags;
863 sch->enqueue = ops->enqueue;
864 sch->dequeue = ops->dequeue;
865 sch->dev_queue = dev_queue;
866 dev_hold(dev);
867 refcount_set(&sch->refcnt, 1);
868
869 return sch;
870 errout1:
871 kfree(p);
872 errout:
873 return ERR_PTR(err);
874 }
875
876 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
877 const struct Qdisc_ops *ops,
878 unsigned int parentid,
879 struct netlink_ext_ack *extack)
880 {
881 struct Qdisc *sch;
882
883 if (!try_module_get(ops->owner)) {
884 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
885 return NULL;
886 }
887
888 sch = qdisc_alloc(dev_queue, ops, extack);
889 if (IS_ERR(sch)) {
890 module_put(ops->owner);
891 return NULL;
892 }
893 sch->parent = parentid;
894
895 if (!ops->init || ops->init(sch, NULL, extack) == 0)
896 return sch;
897
898 qdisc_destroy(sch);
899 return NULL;
900 }
901 EXPORT_SYMBOL(qdisc_create_dflt);
902
903 /* Under qdisc_lock(qdisc) and BH! */
904
905 void qdisc_reset(struct Qdisc *qdisc)
906 {
907 const struct Qdisc_ops *ops = qdisc->ops;
908 struct sk_buff *skb, *tmp;
909
910 if (ops->reset)
911 ops->reset(qdisc);
912
913 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
914 __skb_unlink(skb, &qdisc->gso_skb);
915 kfree_skb_list(skb);
916 }
917
918 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
919 __skb_unlink(skb, &qdisc->skb_bad_txq);
920 kfree_skb_list(skb);
921 }
922
923 qdisc->q.qlen = 0;
924 qdisc->qstats.backlog = 0;
925 }
926 EXPORT_SYMBOL(qdisc_reset);
927
928 void qdisc_free(struct Qdisc *qdisc)
929 {
930 if (qdisc_is_percpu_stats(qdisc)) {
931 free_percpu(qdisc->cpu_bstats);
932 free_percpu(qdisc->cpu_qstats);
933 }
934
935 kfree((char *) qdisc - qdisc->padded);
936 }
937
938 void qdisc_destroy(struct Qdisc *qdisc)
939 {
940 const struct Qdisc_ops *ops = qdisc->ops;
941 struct sk_buff *skb, *tmp;
942
943 if (qdisc->flags & TCQ_F_BUILTIN ||
944 !refcount_dec_and_test(&qdisc->refcnt))
945 return;
946
947 #ifdef CONFIG_NET_SCHED
948 qdisc_hash_del(qdisc);
949
950 qdisc_put_stab(rtnl_dereference(qdisc->stab));
951 #endif
952 gen_kill_estimator(&qdisc->rate_est);
953 if (ops->reset)
954 ops->reset(qdisc);
955 if (ops->destroy)
956 ops->destroy(qdisc);
957
958 module_put(ops->owner);
959 dev_put(qdisc_dev(qdisc));
960
961 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
962 __skb_unlink(skb, &qdisc->gso_skb);
963 kfree_skb_list(skb);
964 }
965
966 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
967 __skb_unlink(skb, &qdisc->skb_bad_txq);
968 kfree_skb_list(skb);
969 }
970
971 qdisc_free(qdisc);
972 }
973 EXPORT_SYMBOL(qdisc_destroy);
974
975 /* Attach toplevel qdisc to device queue. */
976 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
977 struct Qdisc *qdisc)
978 {
979 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
980 spinlock_t *root_lock;
981
982 root_lock = qdisc_lock(oqdisc);
983 spin_lock_bh(root_lock);
984
985 /* ... and graft new one */
986 if (qdisc == NULL)
987 qdisc = &noop_qdisc;
988 dev_queue->qdisc_sleeping = qdisc;
989 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
990
991 spin_unlock_bh(root_lock);
992
993 return oqdisc;
994 }
995 EXPORT_SYMBOL(dev_graft_qdisc);
996
997 static void attach_one_default_qdisc(struct net_device *dev,
998 struct netdev_queue *dev_queue,
999 void *_unused)
1000 {
1001 struct Qdisc *qdisc;
1002 const struct Qdisc_ops *ops = default_qdisc_ops;
1003
1004 if (dev->priv_flags & IFF_NO_QUEUE)
1005 ops = &noqueue_qdisc_ops;
1006
1007 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1008 if (!qdisc) {
1009 netdev_info(dev, "activation failed\n");
1010 return;
1011 }
1012 if (!netif_is_multiqueue(dev))
1013 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1014 dev_queue->qdisc_sleeping = qdisc;
1015 }
1016
1017 static void attach_default_qdiscs(struct net_device *dev)
1018 {
1019 struct netdev_queue *txq;
1020 struct Qdisc *qdisc;
1021
1022 txq = netdev_get_tx_queue(dev, 0);
1023
1024 if (!netif_is_multiqueue(dev) ||
1025 dev->priv_flags & IFF_NO_QUEUE) {
1026 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1027 dev->qdisc = txq->qdisc_sleeping;
1028 qdisc_refcount_inc(dev->qdisc);
1029 } else {
1030 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1031 if (qdisc) {
1032 dev->qdisc = qdisc;
1033 qdisc->ops->attach(qdisc);
1034 }
1035 }
1036 #ifdef CONFIG_NET_SCHED
1037 if (dev->qdisc != &noop_qdisc)
1038 qdisc_hash_add(dev->qdisc, false);
1039 #endif
1040 }
1041
1042 static void transition_one_qdisc(struct net_device *dev,
1043 struct netdev_queue *dev_queue,
1044 void *_need_watchdog)
1045 {
1046 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1047 int *need_watchdog_p = _need_watchdog;
1048
1049 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1050 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1051
1052 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1053 if (need_watchdog_p) {
1054 dev_queue->trans_start = 0;
1055 *need_watchdog_p = 1;
1056 }
1057 }
1058
1059 void dev_activate(struct net_device *dev)
1060 {
1061 int need_watchdog;
1062
1063 /* No queueing discipline is attached to device;
1064 * create default one for devices, which need queueing
1065 * and noqueue_qdisc for virtual interfaces
1066 */
1067
1068 if (dev->qdisc == &noop_qdisc)
1069 attach_default_qdiscs(dev);
1070
1071 if (!netif_carrier_ok(dev))
1072 /* Delay activation until next carrier-on event */
1073 return;
1074
1075 need_watchdog = 0;
1076 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1077 if (dev_ingress_queue(dev))
1078 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1079
1080 if (need_watchdog) {
1081 netif_trans_update(dev);
1082 dev_watchdog_up(dev);
1083 }
1084 }
1085 EXPORT_SYMBOL(dev_activate);
1086
1087 static void dev_deactivate_queue(struct net_device *dev,
1088 struct netdev_queue *dev_queue,
1089 void *_qdisc_default)
1090 {
1091 struct Qdisc *qdisc_default = _qdisc_default;
1092 struct Qdisc *qdisc;
1093
1094 qdisc = rtnl_dereference(dev_queue->qdisc);
1095 if (qdisc) {
1096 spin_lock_bh(qdisc_lock(qdisc));
1097
1098 if (!(qdisc->flags & TCQ_F_BUILTIN))
1099 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1100
1101 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1102 qdisc_reset(qdisc);
1103
1104 spin_unlock_bh(qdisc_lock(qdisc));
1105 }
1106 }
1107
1108 static bool some_qdisc_is_busy(struct net_device *dev)
1109 {
1110 unsigned int i;
1111
1112 for (i = 0; i < dev->num_tx_queues; i++) {
1113 struct netdev_queue *dev_queue;
1114 spinlock_t *root_lock;
1115 struct Qdisc *q;
1116 int val;
1117
1118 dev_queue = netdev_get_tx_queue(dev, i);
1119 q = dev_queue->qdisc_sleeping;
1120
1121 if (q->flags & TCQ_F_NOLOCK) {
1122 val = test_bit(__QDISC_STATE_SCHED, &q->state);
1123 } else {
1124 root_lock = qdisc_lock(q);
1125 spin_lock_bh(root_lock);
1126
1127 val = (qdisc_is_running(q) ||
1128 test_bit(__QDISC_STATE_SCHED, &q->state));
1129
1130 spin_unlock_bh(root_lock);
1131 }
1132
1133 if (val)
1134 return true;
1135 }
1136 return false;
1137 }
1138
1139 static void dev_qdisc_reset(struct net_device *dev,
1140 struct netdev_queue *dev_queue,
1141 void *none)
1142 {
1143 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1144
1145 if (qdisc)
1146 qdisc_reset(qdisc);
1147 }
1148
1149 /**
1150 * dev_deactivate_many - deactivate transmissions on several devices
1151 * @head: list of devices to deactivate
1152 *
1153 * This function returns only when all outstanding transmissions
1154 * have completed, unless all devices are in dismantle phase.
1155 */
1156 void dev_deactivate_many(struct list_head *head)
1157 {
1158 struct net_device *dev;
1159
1160 list_for_each_entry(dev, head, close_list) {
1161 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1162 &noop_qdisc);
1163 if (dev_ingress_queue(dev))
1164 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1165 &noop_qdisc);
1166
1167 dev_watchdog_down(dev);
1168 }
1169
1170 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1171 * This is avoided if all devices are in dismantle phase :
1172 * Caller will call synchronize_net() for us
1173 */
1174 synchronize_net();
1175
1176 /* Wait for outstanding qdisc_run calls. */
1177 list_for_each_entry(dev, head, close_list) {
1178 while (some_qdisc_is_busy(dev))
1179 yield();
1180 /* The new qdisc is assigned at this point so we can safely
1181 * unwind stale skb lists and qdisc statistics
1182 */
1183 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1184 if (dev_ingress_queue(dev))
1185 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1186 }
1187 }
1188
1189 void dev_deactivate(struct net_device *dev)
1190 {
1191 LIST_HEAD(single);
1192
1193 list_add(&dev->close_list, &single);
1194 dev_deactivate_many(&single);
1195 list_del(&single);
1196 }
1197 EXPORT_SYMBOL(dev_deactivate);
1198
1199 static int qdisc_change_tx_queue_len(struct net_device *dev,
1200 struct netdev_queue *dev_queue)
1201 {
1202 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1203 const struct Qdisc_ops *ops = qdisc->ops;
1204
1205 if (ops->change_tx_queue_len)
1206 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1207 return 0;
1208 }
1209
1210 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1211 {
1212 bool up = dev->flags & IFF_UP;
1213 unsigned int i;
1214 int ret = 0;
1215
1216 if (up)
1217 dev_deactivate(dev);
1218
1219 for (i = 0; i < dev->num_tx_queues; i++) {
1220 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1221
1222 /* TODO: revert changes on a partial failure */
1223 if (ret)
1224 break;
1225 }
1226
1227 if (up)
1228 dev_activate(dev);
1229 return ret;
1230 }
1231
1232 static void dev_init_scheduler_queue(struct net_device *dev,
1233 struct netdev_queue *dev_queue,
1234 void *_qdisc)
1235 {
1236 struct Qdisc *qdisc = _qdisc;
1237
1238 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1239 dev_queue->qdisc_sleeping = qdisc;
1240 __skb_queue_head_init(&qdisc->gso_skb);
1241 __skb_queue_head_init(&qdisc->skb_bad_txq);
1242 }
1243
1244 void dev_init_scheduler(struct net_device *dev)
1245 {
1246 dev->qdisc = &noop_qdisc;
1247 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1248 if (dev_ingress_queue(dev))
1249 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1250
1251 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1252 }
1253
1254 static void shutdown_scheduler_queue(struct net_device *dev,
1255 struct netdev_queue *dev_queue,
1256 void *_qdisc_default)
1257 {
1258 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1259 struct Qdisc *qdisc_default = _qdisc_default;
1260
1261 if (qdisc) {
1262 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1263 dev_queue->qdisc_sleeping = qdisc_default;
1264
1265 qdisc_destroy(qdisc);
1266 }
1267 }
1268
1269 void dev_shutdown(struct net_device *dev)
1270 {
1271 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1272 if (dev_ingress_queue(dev))
1273 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1274 qdisc_destroy(dev->qdisc);
1275 dev->qdisc = &noop_qdisc;
1276
1277 WARN_ON(timer_pending(&dev->watchdog_timer));
1278 }
1279
1280 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1281 const struct tc_ratespec *conf,
1282 u64 rate64)
1283 {
1284 memset(r, 0, sizeof(*r));
1285 r->overhead = conf->overhead;
1286 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1287 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1288 r->mult = 1;
1289 /*
1290 * The deal here is to replace a divide by a reciprocal one
1291 * in fast path (a reciprocal divide is a multiply and a shift)
1292 *
1293 * Normal formula would be :
1294 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1295 *
1296 * We compute mult/shift to use instead :
1297 * time_in_ns = (len * mult) >> shift;
1298 *
1299 * We try to get the highest possible mult value for accuracy,
1300 * but have to make sure no overflows will ever happen.
1301 */
1302 if (r->rate_bytes_ps > 0) {
1303 u64 factor = NSEC_PER_SEC;
1304
1305 for (;;) {
1306 r->mult = div64_u64(factor, r->rate_bytes_ps);
1307 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1308 break;
1309 factor <<= 1;
1310 r->shift++;
1311 }
1312 }
1313 }
1314 EXPORT_SYMBOL(psched_ratecfg_precompute);
1315
1316 static void mini_qdisc_rcu_func(struct rcu_head *head)
1317 {
1318 }
1319
1320 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1321 struct tcf_proto *tp_head)
1322 {
1323 struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1324 struct mini_Qdisc *miniq;
1325
1326 if (!tp_head) {
1327 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1328 /* Wait for flying RCU callback before it is freed. */
1329 rcu_barrier_bh();
1330 return;
1331 }
1332
1333 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1334 &miniqp->miniq1 : &miniqp->miniq2;
1335
1336 /* We need to make sure that readers won't see the miniq
1337 * we are about to modify. So wait until previous call_rcu_bh callback
1338 * is done.
1339 */
1340 rcu_barrier_bh();
1341 miniq->filter_list = tp_head;
1342 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1343
1344 if (miniq_old)
1345 /* This is counterpart of the rcu barriers above. We need to
1346 * block potential new user of miniq_old until all readers
1347 * are not seeing it.
1348 */
1349 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1350 }
1351 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1352
1353 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1354 struct mini_Qdisc __rcu **p_miniq)
1355 {
1356 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1357 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1358 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1359 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1360 miniqp->p_miniq = p_miniq;
1361 }
1362 EXPORT_SYMBOL(mini_qdisc_pair_init);
CRIS linux kernel tree