Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/sched/sch_generic.c Generic packet scheduler routines.
4 *
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
7 * - Ingress support
8 */
9
10#include <linux/bitops.h>
11#include <linux/module.h>
12#include <linux/types.h>
13#include <linux/kernel.h>
14#include <linux/sched.h>
15#include <linux/string.h>
16#include <linux/errno.h>
17#include <linux/netdevice.h>
18#include <linux/skbuff.h>
19#include <linux/rtnetlink.h>
20#include <linux/init.h>
21#include <linux/rcupdate.h>
22#include <linux/list.h>
23#include <linux/slab.h>
24#include <linux/if_vlan.h>
25#include <linux/skb_array.h>
26#include <linux/if_macvlan.h>
27#include <net/sch_generic.h>
28#include <net/pkt_sched.h>
29#include <net/dst.h>
30#include <trace/events/qdisc.h>
31#include <trace/events/net.h>
32#include <net/xfrm.h>
33
34/* Qdisc to use by default */
35const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36EXPORT_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#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
50
51static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
52{
53 const struct netdev_queue *txq = q->dev_queue;
54 spinlock_t *lock = NULL;
55 struct sk_buff *skb;
56
57 if (q->flags & TCQ_F_NOLOCK) {
58 lock = qdisc_lock(q);
59 spin_lock(lock);
60 }
61
62 skb = skb_peek(&q->skb_bad_txq);
63 if (skb) {
64 /* check the reason of requeuing without tx lock first */
65 txq = skb_get_tx_queue(txq->dev, skb);
66 if (!netif_xmit_frozen_or_stopped(txq)) {
67 skb = __skb_dequeue(&q->skb_bad_txq);
68 if (qdisc_is_percpu_stats(q)) {
69 qdisc_qstats_cpu_backlog_dec(q, skb);
70 qdisc_qstats_cpu_qlen_dec(q);
71 } else {
72 qdisc_qstats_backlog_dec(q, skb);
73 q->q.qlen--;
74 }
75 } else {
76 skb = SKB_XOFF_MAGIC;
77 }
78 }
79
80 if (lock)
81 spin_unlock(lock);
82
83 return skb;
84}
85
86static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
87{
88 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
89
90 if (unlikely(skb))
91 skb = __skb_dequeue_bad_txq(q);
92
93 return skb;
94}
95
96static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
97 struct sk_buff *skb)
98{
99 spinlock_t *lock = NULL;
100
101 if (q->flags & TCQ_F_NOLOCK) {
102 lock = qdisc_lock(q);
103 spin_lock(lock);
104 }
105
106 __skb_queue_tail(&q->skb_bad_txq, skb);
107
108 if (qdisc_is_percpu_stats(q)) {
109 qdisc_qstats_cpu_backlog_inc(q, skb);
110 qdisc_qstats_cpu_qlen_inc(q);
111 } else {
112 qdisc_qstats_backlog_inc(q, skb);
113 q->q.qlen++;
114 }
115
116 if (lock)
117 spin_unlock(lock);
118}
119
120static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
121{
122 spinlock_t *lock = NULL;
123
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
126 spin_lock(lock);
127 }
128
129 while (skb) {
130 struct sk_buff *next = skb->next;
131
132 __skb_queue_tail(&q->gso_skb, skb);
133
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q)) {
136 qdisc_qstats_cpu_requeues_inc(q);
137 qdisc_qstats_cpu_backlog_inc(q, skb);
138 qdisc_qstats_cpu_qlen_inc(q);
139 } else {
140 q->qstats.requeues++;
141 qdisc_qstats_backlog_inc(q, skb);
142 q->q.qlen++;
143 }
144
145 skb = next;
146 }
147 if (lock)
148 spin_unlock(lock);
149 __netif_schedule(q);
150}
151
152static void try_bulk_dequeue_skb(struct Qdisc *q,
153 struct sk_buff *skb,
154 const struct netdev_queue *txq,
155 int *packets)
156{
157 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
158
159 while (bytelimit > 0) {
160 struct sk_buff *nskb = q->dequeue(q);
161
162 if (!nskb)
163 break;
164
165 bytelimit -= nskb->len; /* covers GSO len */
166 skb->next = nskb;
167 skb = nskb;
168 (*packets)++; /* GSO counts as one pkt */
169 }
170 skb_mark_not_on_list(skb);
171}
172
173/* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
175 */
176static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
177 struct sk_buff *skb,
178 int *packets)
179{
180 int mapping = skb_get_queue_mapping(skb);
181 struct sk_buff *nskb;
182 int cnt = 0;
183
184 do {
185 nskb = q->dequeue(q);
186 if (!nskb)
187 break;
188 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
189 qdisc_enqueue_skb_bad_txq(q, nskb);
190 break;
191 }
192 skb->next = nskb;
193 skb = nskb;
194 } while (++cnt < 8);
195 (*packets) += cnt;
196 skb_mark_not_on_list(skb);
197}
198
199/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
201 */
202static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
203 int *packets)
204{
205 const struct netdev_queue *txq = q->dev_queue;
206 struct sk_buff *skb = NULL;
207
208 *packets = 1;
209 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
210 spinlock_t *lock = NULL;
211
212 if (q->flags & TCQ_F_NOLOCK) {
213 lock = qdisc_lock(q);
214 spin_lock(lock);
215 }
216
217 skb = skb_peek(&q->gso_skb);
218
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
221 */
222 if (!skb) {
223 if (lock)
224 spin_unlock(lock);
225 goto validate;
226 }
227
228 /* skb in gso_skb were already validated */
229 *validate = false;
230 if (xfrm_offload(skb))
231 *validate = true;
232 /* check the reason of requeuing without tx lock first */
233 txq = skb_get_tx_queue(txq->dev, skb);
234 if (!netif_xmit_frozen_or_stopped(txq)) {
235 skb = __skb_dequeue(&q->gso_skb);
236 if (qdisc_is_percpu_stats(q)) {
237 qdisc_qstats_cpu_backlog_dec(q, skb);
238 qdisc_qstats_cpu_qlen_dec(q);
239 } else {
240 qdisc_qstats_backlog_dec(q, skb);
241 q->q.qlen--;
242 }
243 } else {
244 skb = NULL;
245 }
246 if (lock)
247 spin_unlock(lock);
248 goto trace;
249 }
250validate:
251 *validate = true;
252
253 if ((q->flags & TCQ_F_ONETXQUEUE) &&
254 netif_xmit_frozen_or_stopped(txq))
255 return skb;
256
257 skb = qdisc_dequeue_skb_bad_txq(q);
258 if (unlikely(skb)) {
259 if (skb == SKB_XOFF_MAGIC)
260 return NULL;
261 goto bulk;
262 }
263 skb = q->dequeue(q);
264 if (skb) {
265bulk:
266 if (qdisc_may_bulk(q))
267 try_bulk_dequeue_skb(q, skb, txq, packets);
268 else
269 try_bulk_dequeue_skb_slow(q, skb, packets);
270 }
271trace:
272 trace_qdisc_dequeue(q, txq, *packets, skb);
273 return skb;
274}
275
276/*
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
280 *
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
284 */
285bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
286 struct net_device *dev, struct netdev_queue *txq,
287 spinlock_t *root_lock, bool validate)
288{
289 int ret = NETDEV_TX_BUSY;
290 bool again = false;
291
292 /* And release qdisc */
293 if (root_lock)
294 spin_unlock(root_lock);
295
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
297 if (validate)
298 skb = validate_xmit_skb_list(skb, dev, &again);
299
300#ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again)) {
302 if (root_lock)
303 spin_lock(root_lock);
304
305 dev_requeue_skb(skb, q);
306 return false;
307 }
308#endif
309
310 if (likely(skb)) {
311 HARD_TX_LOCK(dev, txq, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq))
313 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
314
315 HARD_TX_UNLOCK(dev, txq);
316 } else {
317 if (root_lock)
318 spin_lock(root_lock);
319 return true;
320 }
321
322 if (root_lock)
323 spin_lock(root_lock);
324
325 if (!dev_xmit_complete(ret)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret != NETDEV_TX_BUSY))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev->name, ret, q->q.qlen);
330
331 dev_requeue_skb(skb, q);
332 return false;
333 }
334
335 return true;
336}
337
338/*
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
340 *
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
343 * this queue.
344 *
345 * netif_tx_lock serializes accesses to device driver.
346 *
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
349 *
350 * Note, that this procedure can be called by a watchdog timer
351 *
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
355 *
356 */
357static inline bool qdisc_restart(struct Qdisc *q, int *packets)
358{
359 spinlock_t *root_lock = NULL;
360 struct netdev_queue *txq;
361 struct net_device *dev;
362 struct sk_buff *skb;
363 bool validate;
364
365 /* Dequeue packet */
366 skb = dequeue_skb(q, &validate, packets);
367 if (unlikely(!skb))
368 return false;
369
370 if (!(q->flags & TCQ_F_NOLOCK))
371 root_lock = qdisc_lock(q);
372
373 dev = qdisc_dev(q);
374 txq = skb_get_tx_queue(dev, skb);
375
376 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
377}
378
379void __qdisc_run(struct Qdisc *q)
380{
381 int quota = dev_tx_weight;
382 int packets;
383
384 while (qdisc_restart(q, &packets)) {
385 /*
386 * Ordered by possible occurrence: Postpone processing if
387 * 1. we've exceeded packet quota
388 * 2. another process needs the CPU;
389 */
390 quota -= packets;
391 if (quota <= 0 || need_resched()) {
392 __netif_schedule(q);
393 break;
394 }
395 }
396}
397
398unsigned long dev_trans_start(struct net_device *dev)
399{
400 unsigned long val, res;
401 unsigned int i;
402
403 if (is_vlan_dev(dev))
404 dev = vlan_dev_real_dev(dev);
405 else if (netif_is_macvlan(dev))
406 dev = macvlan_dev_real_dev(dev);
407 res = netdev_get_tx_queue(dev, 0)->trans_start;
408 for (i = 1; i < dev->num_tx_queues; i++) {
409 val = netdev_get_tx_queue(dev, i)->trans_start;
410 if (val && time_after(val, res))
411 res = val;
412 }
413
414 return res;
415}
416EXPORT_SYMBOL(dev_trans_start);
417
418static void dev_watchdog(struct timer_list *t)
419{
420 struct net_device *dev = from_timer(dev, t, watchdog_timer);
421
422 netif_tx_lock(dev);
423 if (!qdisc_tx_is_noop(dev)) {
424 if (netif_device_present(dev) &&
425 netif_running(dev) &&
426 netif_carrier_ok(dev)) {
427 int some_queue_timedout = 0;
428 unsigned int i;
429 unsigned long trans_start;
430
431 for (i = 0; i < dev->num_tx_queues; i++) {
432 struct netdev_queue *txq;
433
434 txq = netdev_get_tx_queue(dev, i);
435 trans_start = txq->trans_start;
436 if (netif_xmit_stopped(txq) &&
437 time_after(jiffies, (trans_start +
438 dev->watchdog_timeo))) {
439 some_queue_timedout = 1;
440 txq->trans_timeout++;
441 break;
442 }
443 }
444
445 if (some_queue_timedout) {
446 trace_net_dev_xmit_timeout(dev, i);
447 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
448 dev->name, netdev_drivername(dev), i);
449 dev->netdev_ops->ndo_tx_timeout(dev);
450 }
451 if (!mod_timer(&dev->watchdog_timer,
452 round_jiffies(jiffies +
453 dev->watchdog_timeo)))
454 dev_hold(dev);
455 }
456 }
457 netif_tx_unlock(dev);
458
459 dev_put(dev);
460}
461
462void __netdev_watchdog_up(struct net_device *dev)
463{
464 if (dev->netdev_ops->ndo_tx_timeout) {
465 if (dev->watchdog_timeo <= 0)
466 dev->watchdog_timeo = 5*HZ;
467 if (!mod_timer(&dev->watchdog_timer,
468 round_jiffies(jiffies + dev->watchdog_timeo)))
469 dev_hold(dev);
470 }
471}
472
473static void dev_watchdog_up(struct net_device *dev)
474{
475 __netdev_watchdog_up(dev);
476}
477
478static void dev_watchdog_down(struct net_device *dev)
479{
480 netif_tx_lock_bh(dev);
481 if (del_timer(&dev->watchdog_timer))
482 dev_put(dev);
483 netif_tx_unlock_bh(dev);
484}
485
486/**
487 * netif_carrier_on - set carrier
488 * @dev: network device
489 *
490 * Device has detected acquisition of carrier.
491 */
492void netif_carrier_on(struct net_device *dev)
493{
494 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
495 if (dev->reg_state == NETREG_UNINITIALIZED)
496 return;
497 atomic_inc(&dev->carrier_up_count);
498 linkwatch_fire_event(dev);
499 if (netif_running(dev))
500 __netdev_watchdog_up(dev);
501 }
502}
503EXPORT_SYMBOL(netif_carrier_on);
504
505/**
506 * netif_carrier_off - clear carrier
507 * @dev: network device
508 *
509 * Device has detected loss of carrier.
510 */
511void netif_carrier_off(struct net_device *dev)
512{
513 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
514 if (dev->reg_state == NETREG_UNINITIALIZED)
515 return;
516 atomic_inc(&dev->carrier_down_count);
517 linkwatch_fire_event(dev);
518 }
519}
520EXPORT_SYMBOL(netif_carrier_off);
521
522/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
523 under all circumstances. It is difficult to invent anything faster or
524 cheaper.
525 */
526
527static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
528 struct sk_buff **to_free)
529{
530 __qdisc_drop(skb, to_free);
531 return NET_XMIT_CN;
532}
533
534static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
535{
536 return NULL;
537}
538
539struct Qdisc_ops noop_qdisc_ops __read_mostly = {
540 .id = "noop",
541 .priv_size = 0,
542 .enqueue = noop_enqueue,
543 .dequeue = noop_dequeue,
544 .peek = noop_dequeue,
545 .owner = THIS_MODULE,
546};
547
548static struct netdev_queue noop_netdev_queue = {
549 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
550 .qdisc_sleeping = &noop_qdisc,
551};
552
553struct Qdisc noop_qdisc = {
554 .enqueue = noop_enqueue,
555 .dequeue = noop_dequeue,
556 .flags = TCQ_F_BUILTIN,
557 .ops = &noop_qdisc_ops,
558 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
559 .dev_queue = &noop_netdev_queue,
560 .running = SEQCNT_ZERO(noop_qdisc.running),
561 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
562 .gso_skb = {
563 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
564 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
565 .qlen = 0,
566 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
567 },
568 .skb_bad_txq = {
569 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
570 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
571 .qlen = 0,
572 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
573 },
574};
575EXPORT_SYMBOL(noop_qdisc);
576
577static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
578 struct netlink_ext_ack *extack)
579{
580 /* register_qdisc() assigns a default of noop_enqueue if unset,
581 * but __dev_queue_xmit() treats noqueue only as such
582 * if this is NULL - so clear it here. */
583 qdisc->enqueue = NULL;
584 return 0;
585}
586
587struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
588 .id = "noqueue",
589 .priv_size = 0,
590 .init = noqueue_init,
591 .enqueue = noop_enqueue,
592 .dequeue = noop_dequeue,
593 .peek = noop_dequeue,
594 .owner = THIS_MODULE,
595};
596
597static const u8 prio2band[TC_PRIO_MAX + 1] = {
598 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
599};
600
601/* 3-band FIFO queue: old style, but should be a bit faster than
602 generic prio+fifo combination.
603 */
604
605#define PFIFO_FAST_BANDS 3
606
607/*
608 * Private data for a pfifo_fast scheduler containing:
609 * - rings for priority bands
610 */
611struct pfifo_fast_priv {
612 struct skb_array q[PFIFO_FAST_BANDS];
613};
614
615static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
616 int band)
617{
618 return &priv->q[band];
619}
620
621static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
622 struct sk_buff **to_free)
623{
624 int band = prio2band[skb->priority & TC_PRIO_MAX];
625 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
626 struct skb_array *q = band2list(priv, band);
627 unsigned int pkt_len = qdisc_pkt_len(skb);
628 int err;
629
630 err = skb_array_produce(q, skb);
631
632 if (unlikely(err)) {
633 if (qdisc_is_percpu_stats(qdisc))
634 return qdisc_drop_cpu(skb, qdisc, to_free);
635 else
636 return qdisc_drop(skb, qdisc, to_free);
637 }
638
639 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
640 return NET_XMIT_SUCCESS;
641}
642
643static 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(q);
656 }
657 if (likely(skb)) {
658 qdisc_update_stats_at_dequeue(qdisc, skb);
659 } else {
660 qdisc->empty = true;
661 }
662
663 return skb;
664}
665
666static 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
681static 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(q)) != NULL)
697 kfree_skb(skb);
698 }
699
700 if (qdisc_is_percpu_stats(qdisc)) {
701 for_each_possible_cpu(i) {
702 struct gnet_stats_queue *q;
703
704 q = per_cpu_ptr(qdisc->cpu_qstats, i);
705 q->backlog = 0;
706 q->qlen = 0;
707 }
708 }
709}
710
711static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
712{
713 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
714
715 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
716 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
717 goto nla_put_failure;
718 return skb->len;
719
720nla_put_failure:
721 return -1;
722}
723
724static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
725 struct netlink_ext_ack *extack)
726{
727 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
728 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
729 int prio;
730
731 /* guard against zero length rings */
732 if (!qlen)
733 return -EINVAL;
734
735 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
736 struct skb_array *q = band2list(priv, prio);
737 int err;
738
739 err = skb_array_init(q, qlen, GFP_KERNEL);
740 if (err)
741 return -ENOMEM;
742 }
743
744 /* Can by-pass the queue discipline */
745 qdisc->flags |= TCQ_F_CAN_BYPASS;
746 return 0;
747}
748
749static void pfifo_fast_destroy(struct Qdisc *sch)
750{
751 struct pfifo_fast_priv *priv = qdisc_priv(sch);
752 int prio;
753
754 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
755 struct skb_array *q = band2list(priv, prio);
756
757 /* NULL ring is possible if destroy path is due to a failed
758 * skb_array_init() in pfifo_fast_init() case.
759 */
760 if (!q->ring.queue)
761 continue;
762 /* Destroy ring but no need to kfree_skb because a call to
763 * pfifo_fast_reset() has already done that work.
764 */
765 ptr_ring_cleanup(&q->ring, NULL);
766 }
767}
768
769static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
770 unsigned int new_len)
771{
772 struct pfifo_fast_priv *priv = qdisc_priv(sch);
773 struct skb_array *bands[PFIFO_FAST_BANDS];
774 int prio;
775
776 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
777 struct skb_array *q = band2list(priv, prio);
778
779 bands[prio] = q;
780 }
781
782 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
783 GFP_KERNEL);
784}
785
786struct Qdisc_ops pfifo_fast_ops __read_mostly = {
787 .id = "pfifo_fast",
788 .priv_size = sizeof(struct pfifo_fast_priv),
789 .enqueue = pfifo_fast_enqueue,
790 .dequeue = pfifo_fast_dequeue,
791 .peek = pfifo_fast_peek,
792 .init = pfifo_fast_init,
793 .destroy = pfifo_fast_destroy,
794 .reset = pfifo_fast_reset,
795 .dump = pfifo_fast_dump,
796 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
797 .owner = THIS_MODULE,
798 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
799};
800EXPORT_SYMBOL(pfifo_fast_ops);
801
802struct 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 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
855 spin_lock_init(&sch->seqlock);
856 seqcount_init(&sch->running);
857
858 sch->ops = ops;
859 sch->flags = ops->static_flags;
860 sch->enqueue = ops->enqueue;
861 sch->dequeue = ops->dequeue;
862 sch->dev_queue = dev_queue;
863 sch->empty = true;
864 dev_hold(dev);
865 refcount_set(&sch->refcnt, 1);
866
867 if (sch != &noop_qdisc) {
868 lockdep_set_class(&sch->busylock, &dev->qdisc_tx_busylock_key);
869 lockdep_set_class(&sch->seqlock, &dev->qdisc_tx_busylock_key);
870 lockdep_set_class(&sch->running, &dev->qdisc_running_key);
871 }
872
873 return sch;
874errout1:
875 kfree(p);
876errout:
877 return ERR_PTR(err);
878}
879
880struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
881 const struct Qdisc_ops *ops,
882 unsigned int parentid,
883 struct netlink_ext_ack *extack)
884{
885 struct Qdisc *sch;
886
887 if (!try_module_get(ops->owner)) {
888 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
889 return NULL;
890 }
891
892 sch = qdisc_alloc(dev_queue, ops, extack);
893 if (IS_ERR(sch)) {
894 module_put(ops->owner);
895 return NULL;
896 }
897 sch->parent = parentid;
898
899 if (!ops->init || ops->init(sch, NULL, extack) == 0)
900 return sch;
901
902 qdisc_put(sch);
903 return NULL;
904}
905EXPORT_SYMBOL(qdisc_create_dflt);
906
907/* Under qdisc_lock(qdisc) and BH! */
908
909void qdisc_reset(struct Qdisc *qdisc)
910{
911 const struct Qdisc_ops *ops = qdisc->ops;
912 struct sk_buff *skb, *tmp;
913
914 if (ops->reset)
915 ops->reset(qdisc);
916
917 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
918 __skb_unlink(skb, &qdisc->gso_skb);
919 kfree_skb_list(skb);
920 }
921
922 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
923 __skb_unlink(skb, &qdisc->skb_bad_txq);
924 kfree_skb_list(skb);
925 }
926
927 qdisc->q.qlen = 0;
928 qdisc->qstats.backlog = 0;
929}
930EXPORT_SYMBOL(qdisc_reset);
931
932void qdisc_free(struct Qdisc *qdisc)
933{
934 if (qdisc_is_percpu_stats(qdisc)) {
935 free_percpu(qdisc->cpu_bstats);
936 free_percpu(qdisc->cpu_qstats);
937 }
938
939 kfree((char *) qdisc - qdisc->padded);
940}
941
942static void qdisc_free_cb(struct rcu_head *head)
943{
944 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
945
946 qdisc_free(q);
947}
948
949static void qdisc_destroy(struct Qdisc *qdisc)
950{
951 const struct Qdisc_ops *ops = qdisc->ops;
952 struct sk_buff *skb, *tmp;
953
954#ifdef CONFIG_NET_SCHED
955 qdisc_hash_del(qdisc);
956
957 qdisc_put_stab(rtnl_dereference(qdisc->stab));
958#endif
959 gen_kill_estimator(&qdisc->rate_est);
960 if (ops->reset)
961 ops->reset(qdisc);
962 if (ops->destroy)
963 ops->destroy(qdisc);
964
965 module_put(ops->owner);
966 dev_put(qdisc_dev(qdisc));
967
968 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
969 __skb_unlink(skb, &qdisc->gso_skb);
970 kfree_skb_list(skb);
971 }
972
973 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
974 __skb_unlink(skb, &qdisc->skb_bad_txq);
975 kfree_skb_list(skb);
976 }
977
978 call_rcu(&qdisc->rcu, qdisc_free_cb);
979}
980
981void qdisc_put(struct Qdisc *qdisc)
982{
983 if (!qdisc)
984 return;
985
986 if (qdisc->flags & TCQ_F_BUILTIN ||
987 !refcount_dec_and_test(&qdisc->refcnt))
988 return;
989
990 qdisc_destroy(qdisc);
991}
992EXPORT_SYMBOL(qdisc_put);
993
994/* Version of qdisc_put() that is called with rtnl mutex unlocked.
995 * Intended to be used as optimization, this function only takes rtnl lock if
996 * qdisc reference counter reached zero.
997 */
998
999void qdisc_put_unlocked(struct Qdisc *qdisc)
1000{
1001 if (qdisc->flags & TCQ_F_BUILTIN ||
1002 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1003 return;
1004
1005 qdisc_destroy(qdisc);
1006 rtnl_unlock();
1007}
1008EXPORT_SYMBOL(qdisc_put_unlocked);
1009
1010/* Attach toplevel qdisc to device queue. */
1011struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1012 struct Qdisc *qdisc)
1013{
1014 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1015 spinlock_t *root_lock;
1016
1017 root_lock = qdisc_lock(oqdisc);
1018 spin_lock_bh(root_lock);
1019
1020 /* ... and graft new one */
1021 if (qdisc == NULL)
1022 qdisc = &noop_qdisc;
1023 dev_queue->qdisc_sleeping = qdisc;
1024 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1025
1026 spin_unlock_bh(root_lock);
1027
1028 return oqdisc;
1029}
1030EXPORT_SYMBOL(dev_graft_qdisc);
1031
1032static void attach_one_default_qdisc(struct net_device *dev,
1033 struct netdev_queue *dev_queue,
1034 void *_unused)
1035{
1036 struct Qdisc *qdisc;
1037 const struct Qdisc_ops *ops = default_qdisc_ops;
1038
1039 if (dev->priv_flags & IFF_NO_QUEUE)
1040 ops = &noqueue_qdisc_ops;
1041 else if(dev->type == ARPHRD_CAN)
1042 ops = &pfifo_fast_ops;
1043
1044 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1045 if (!qdisc) {
1046 netdev_info(dev, "activation failed\n");
1047 return;
1048 }
1049 if (!netif_is_multiqueue(dev))
1050 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1051 dev_queue->qdisc_sleeping = qdisc;
1052}
1053
1054static void attach_default_qdiscs(struct net_device *dev)
1055{
1056 struct netdev_queue *txq;
1057 struct Qdisc *qdisc;
1058
1059 txq = netdev_get_tx_queue(dev, 0);
1060
1061 if (!netif_is_multiqueue(dev) ||
1062 dev->priv_flags & IFF_NO_QUEUE) {
1063 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1064 dev->qdisc = txq->qdisc_sleeping;
1065 qdisc_refcount_inc(dev->qdisc);
1066 } else {
1067 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1068 if (qdisc) {
1069 dev->qdisc = qdisc;
1070 qdisc->ops->attach(qdisc);
1071 }
1072 }
1073#ifdef CONFIG_NET_SCHED
1074 if (dev->qdisc != &noop_qdisc)
1075 qdisc_hash_add(dev->qdisc, false);
1076#endif
1077}
1078
1079static void transition_one_qdisc(struct net_device *dev,
1080 struct netdev_queue *dev_queue,
1081 void *_need_watchdog)
1082{
1083 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1084 int *need_watchdog_p = _need_watchdog;
1085
1086 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1087 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1088
1089 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1090 if (need_watchdog_p) {
1091 dev_queue->trans_start = 0;
1092 *need_watchdog_p = 1;
1093 }
1094}
1095
1096void dev_activate(struct net_device *dev)
1097{
1098 int need_watchdog;
1099
1100 /* No queueing discipline is attached to device;
1101 * create default one for devices, which need queueing
1102 * and noqueue_qdisc for virtual interfaces
1103 */
1104
1105 if (dev->qdisc == &noop_qdisc)
1106 attach_default_qdiscs(dev);
1107
1108 if (!netif_carrier_ok(dev))
1109 /* Delay activation until next carrier-on event */
1110 return;
1111
1112 need_watchdog = 0;
1113 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1114 if (dev_ingress_queue(dev))
1115 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1116
1117 if (need_watchdog) {
1118 netif_trans_update(dev);
1119 dev_watchdog_up(dev);
1120 }
1121}
1122EXPORT_SYMBOL(dev_activate);
1123
1124static void dev_deactivate_queue(struct net_device *dev,
1125 struct netdev_queue *dev_queue,
1126 void *_qdisc_default)
1127{
1128 struct Qdisc *qdisc_default = _qdisc_default;
1129 struct Qdisc *qdisc;
1130
1131 qdisc = rtnl_dereference(dev_queue->qdisc);
1132 if (qdisc) {
1133 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1134
1135 if (nolock)
1136 spin_lock_bh(&qdisc->seqlock);
1137 spin_lock_bh(qdisc_lock(qdisc));
1138
1139 if (!(qdisc->flags & TCQ_F_BUILTIN))
1140 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1141
1142 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1143 qdisc_reset(qdisc);
1144
1145 spin_unlock_bh(qdisc_lock(qdisc));
1146 if (nolock)
1147 spin_unlock_bh(&qdisc->seqlock);
1148 }
1149}
1150
1151static bool some_qdisc_is_busy(struct net_device *dev)
1152{
1153 unsigned int i;
1154
1155 for (i = 0; i < dev->num_tx_queues; i++) {
1156 struct netdev_queue *dev_queue;
1157 spinlock_t *root_lock;
1158 struct Qdisc *q;
1159 int val;
1160
1161 dev_queue = netdev_get_tx_queue(dev, i);
1162 q = dev_queue->qdisc_sleeping;
1163
1164 root_lock = qdisc_lock(q);
1165 spin_lock_bh(root_lock);
1166
1167 val = (qdisc_is_running(q) ||
1168 test_bit(__QDISC_STATE_SCHED, &q->state));
1169
1170 spin_unlock_bh(root_lock);
1171
1172 if (val)
1173 return true;
1174 }
1175 return false;
1176}
1177
1178static void dev_qdisc_reset(struct net_device *dev,
1179 struct netdev_queue *dev_queue,
1180 void *none)
1181{
1182 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1183
1184 if (qdisc)
1185 qdisc_reset(qdisc);
1186}
1187
1188/**
1189 * dev_deactivate_many - deactivate transmissions on several devices
1190 * @head: list of devices to deactivate
1191 *
1192 * This function returns only when all outstanding transmissions
1193 * have completed, unless all devices are in dismantle phase.
1194 */
1195void dev_deactivate_many(struct list_head *head)
1196{
1197 struct net_device *dev;
1198
1199 list_for_each_entry(dev, head, close_list) {
1200 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1201 &noop_qdisc);
1202 if (dev_ingress_queue(dev))
1203 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1204 &noop_qdisc);
1205
1206 dev_watchdog_down(dev);
1207 }
1208
1209 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1210 * This is avoided if all devices are in dismantle phase :
1211 * Caller will call synchronize_net() for us
1212 */
1213 synchronize_net();
1214
1215 /* Wait for outstanding qdisc_run calls. */
1216 list_for_each_entry(dev, head, close_list) {
1217 while (some_qdisc_is_busy(dev))
1218 yield();
1219 /* The new qdisc is assigned at this point so we can safely
1220 * unwind stale skb lists and qdisc statistics
1221 */
1222 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1223 if (dev_ingress_queue(dev))
1224 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1225 }
1226}
1227
1228void dev_deactivate(struct net_device *dev)
1229{
1230 LIST_HEAD(single);
1231
1232 list_add(&dev->close_list, &single);
1233 dev_deactivate_many(&single);
1234 list_del(&single);
1235}
1236EXPORT_SYMBOL(dev_deactivate);
1237
1238static int qdisc_change_tx_queue_len(struct net_device *dev,
1239 struct netdev_queue *dev_queue)
1240{
1241 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1242 const struct Qdisc_ops *ops = qdisc->ops;
1243
1244 if (ops->change_tx_queue_len)
1245 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1246 return 0;
1247}
1248
1249int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1250{
1251 bool up = dev->flags & IFF_UP;
1252 unsigned int i;
1253 int ret = 0;
1254
1255 if (up)
1256 dev_deactivate(dev);
1257
1258 for (i = 0; i < dev->num_tx_queues; i++) {
1259 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1260
1261 /* TODO: revert changes on a partial failure */
1262 if (ret)
1263 break;
1264 }
1265
1266 if (up)
1267 dev_activate(dev);
1268 return ret;
1269}
1270
1271static void dev_init_scheduler_queue(struct net_device *dev,
1272 struct netdev_queue *dev_queue,
1273 void *_qdisc)
1274{
1275 struct Qdisc *qdisc = _qdisc;
1276
1277 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1278 dev_queue->qdisc_sleeping = qdisc;
1279}
1280
1281void dev_init_scheduler(struct net_device *dev)
1282{
1283 dev->qdisc = &noop_qdisc;
1284 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1287
1288 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1289}
1290
1291static void shutdown_scheduler_queue(struct net_device *dev,
1292 struct netdev_queue *dev_queue,
1293 void *_qdisc_default)
1294{
1295 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1296 struct Qdisc *qdisc_default = _qdisc_default;
1297
1298 if (qdisc) {
1299 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1300 dev_queue->qdisc_sleeping = qdisc_default;
1301
1302 qdisc_put(qdisc);
1303 }
1304}
1305
1306void dev_shutdown(struct net_device *dev)
1307{
1308 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1309 if (dev_ingress_queue(dev))
1310 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311 qdisc_put(dev->qdisc);
1312 dev->qdisc = &noop_qdisc;
1313
1314 WARN_ON(timer_pending(&dev->watchdog_timer));
1315}
1316
1317void psched_ratecfg_precompute(struct psched_ratecfg *r,
1318 const struct tc_ratespec *conf,
1319 u64 rate64)
1320{
1321 memset(r, 0, sizeof(*r));
1322 r->overhead = conf->overhead;
1323 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1324 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1325 r->mult = 1;
1326 /*
1327 * The deal here is to replace a divide by a reciprocal one
1328 * in fast path (a reciprocal divide is a multiply and a shift)
1329 *
1330 * Normal formula would be :
1331 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1332 *
1333 * We compute mult/shift to use instead :
1334 * time_in_ns = (len * mult) >> shift;
1335 *
1336 * We try to get the highest possible mult value for accuracy,
1337 * but have to make sure no overflows will ever happen.
1338 */
1339 if (r->rate_bytes_ps > 0) {
1340 u64 factor = NSEC_PER_SEC;
1341
1342 for (;;) {
1343 r->mult = div64_u64(factor, r->rate_bytes_ps);
1344 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1345 break;
1346 factor <<= 1;
1347 r->shift++;
1348 }
1349 }
1350}
1351EXPORT_SYMBOL(psched_ratecfg_precompute);
1352
1353static void mini_qdisc_rcu_func(struct rcu_head *head)
1354{
1355}
1356
1357void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1358 struct tcf_proto *tp_head)
1359{
1360 /* Protected with chain0->filter_chain_lock.
1361 * Can't access chain directly because tp_head can be NULL.
1362 */
1363 struct mini_Qdisc *miniq_old =
1364 rcu_dereference_protected(*miniqp->p_miniq, 1);
1365 struct mini_Qdisc *miniq;
1366
1367 if (!tp_head) {
1368 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1369 /* Wait for flying RCU callback before it is freed. */
1370 rcu_barrier();
1371 return;
1372 }
1373
1374 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1375 &miniqp->miniq1 : &miniqp->miniq2;
1376
1377 /* We need to make sure that readers won't see the miniq
1378 * we are about to modify. So wait until previous call_rcu callback
1379 * is done.
1380 */
1381 rcu_barrier();
1382 miniq->filter_list = tp_head;
1383 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1384
1385 if (miniq_old)
1386 /* This is counterpart of the rcu barriers above. We need to
1387 * block potential new user of miniq_old until all readers
1388 * are not seeing it.
1389 */
1390 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1391}
1392EXPORT_SYMBOL(mini_qdisc_pair_swap);
1393
1394void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1395 struct mini_Qdisc __rcu **p_miniq)
1396{
1397 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1398 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1399 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1400 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1401 miniqp->p_miniq = p_miniq;
1402}
1403EXPORT_SYMBOL(mini_qdisc_pair_init);
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);