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
38static void qdisc_maybe_clear_missed(struct Qdisc *q,
39 const struct netdev_queue *txq)
40{
41 clear_bit(__QDISC_STATE_MISSED, &q->state);
42
43 /* Make sure the below netif_xmit_frozen_or_stopped()
44 * checking happens after clearing STATE_MISSED.
45 */
46 smp_mb__after_atomic();
47
48 /* Checking netif_xmit_frozen_or_stopped() again to
49 * make sure STATE_MISSED is set if the STATE_MISSED
50 * set by netif_tx_wake_queue()'s rescheduling of
51 * net_tx_action() is cleared by the above clear_bit().
52 */
53 if (!netif_xmit_frozen_or_stopped(txq))
54 set_bit(__QDISC_STATE_MISSED, &q->state);
55 else
56 set_bit(__QDISC_STATE_DRAINING, &q->state);
57}
58
59/* Main transmission queue. */
60
61/* Modifications to data participating in scheduling must be protected with
62 * qdisc_lock(qdisc) spinlock.
63 *
64 * The idea is the following:
65 * - enqueue, dequeue are serialized via qdisc root lock
66 * - ingress filtering is also serialized via qdisc root lock
67 * - updates to tree and tree walking are only done under the rtnl mutex.
68 */
69
70#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
71
72static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
73{
74 const struct netdev_queue *txq = q->dev_queue;
75 spinlock_t *lock = NULL;
76 struct sk_buff *skb;
77
78 if (q->flags & TCQ_F_NOLOCK) {
79 lock = qdisc_lock(q);
80 spin_lock(lock);
81 }
82
83 skb = skb_peek(&q->skb_bad_txq);
84 if (skb) {
85 /* check the reason of requeuing without tx lock first */
86 txq = skb_get_tx_queue(txq->dev, skb);
87 if (!netif_xmit_frozen_or_stopped(txq)) {
88 skb = __skb_dequeue(&q->skb_bad_txq);
89 if (qdisc_is_percpu_stats(q)) {
90 qdisc_qstats_cpu_backlog_dec(q, skb);
91 qdisc_qstats_cpu_qlen_dec(q);
92 } else {
93 qdisc_qstats_backlog_dec(q, skb);
94 q->q.qlen--;
95 }
96 } else {
97 skb = SKB_XOFF_MAGIC;
98 qdisc_maybe_clear_missed(q, txq);
99 }
100 }
101
102 if (lock)
103 spin_unlock(lock);
104
105 return skb;
106}
107
108static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
109{
110 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
111
112 if (unlikely(skb))
113 skb = __skb_dequeue_bad_txq(q);
114
115 return skb;
116}
117
118static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
119 struct sk_buff *skb)
120{
121 spinlock_t *lock = NULL;
122
123 if (q->flags & TCQ_F_NOLOCK) {
124 lock = qdisc_lock(q);
125 spin_lock(lock);
126 }
127
128 __skb_queue_tail(&q->skb_bad_txq, skb);
129
130 if (qdisc_is_percpu_stats(q)) {
131 qdisc_qstats_cpu_backlog_inc(q, skb);
132 qdisc_qstats_cpu_qlen_inc(q);
133 } else {
134 qdisc_qstats_backlog_inc(q, skb);
135 q->q.qlen++;
136 }
137
138 if (lock)
139 spin_unlock(lock);
140}
141
142static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
143{
144 spinlock_t *lock = NULL;
145
146 if (q->flags & TCQ_F_NOLOCK) {
147 lock = qdisc_lock(q);
148 spin_lock(lock);
149 }
150
151 while (skb) {
152 struct sk_buff *next = skb->next;
153
154 __skb_queue_tail(&q->gso_skb, skb);
155
156 /* it's still part of the queue */
157 if (qdisc_is_percpu_stats(q)) {
158 qdisc_qstats_cpu_requeues_inc(q);
159 qdisc_qstats_cpu_backlog_inc(q, skb);
160 qdisc_qstats_cpu_qlen_inc(q);
161 } else {
162 q->qstats.requeues++;
163 qdisc_qstats_backlog_inc(q, skb);
164 q->q.qlen++;
165 }
166
167 skb = next;
168 }
169
170 if (lock) {
171 spin_unlock(lock);
172 set_bit(__QDISC_STATE_MISSED, &q->state);
173 } else {
174 __netif_schedule(q);
175 }
176}
177
178static void try_bulk_dequeue_skb(struct Qdisc *q,
179 struct sk_buff *skb,
180 const struct netdev_queue *txq,
181 int *packets)
182{
183 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
184
185 while (bytelimit > 0) {
186 struct sk_buff *nskb = q->dequeue(q);
187
188 if (!nskb)
189 break;
190
191 bytelimit -= nskb->len; /* covers GSO len */
192 skb->next = nskb;
193 skb = nskb;
194 (*packets)++; /* GSO counts as one pkt */
195 }
196 skb_mark_not_on_list(skb);
197}
198
199/* This variant of try_bulk_dequeue_skb() makes sure
200 * all skbs in the chain are for the same txq
201 */
202static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
203 struct sk_buff *skb,
204 int *packets)
205{
206 int mapping = skb_get_queue_mapping(skb);
207 struct sk_buff *nskb;
208 int cnt = 0;
209
210 do {
211 nskb = q->dequeue(q);
212 if (!nskb)
213 break;
214 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
215 qdisc_enqueue_skb_bad_txq(q, nskb);
216 break;
217 }
218 skb->next = nskb;
219 skb = nskb;
220 } while (++cnt < 8);
221 (*packets) += cnt;
222 skb_mark_not_on_list(skb);
223}
224
225/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
226 * A requeued skb (via q->gso_skb) can also be a SKB list.
227 */
228static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
229 int *packets)
230{
231 const struct netdev_queue *txq = q->dev_queue;
232 struct sk_buff *skb = NULL;
233
234 *packets = 1;
235 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
236 spinlock_t *lock = NULL;
237
238 if (q->flags & TCQ_F_NOLOCK) {
239 lock = qdisc_lock(q);
240 spin_lock(lock);
241 }
242
243 skb = skb_peek(&q->gso_skb);
244
245 /* skb may be null if another cpu pulls gso_skb off in between
246 * empty check and lock.
247 */
248 if (!skb) {
249 if (lock)
250 spin_unlock(lock);
251 goto validate;
252 }
253
254 /* skb in gso_skb were already validated */
255 *validate = false;
256 if (xfrm_offload(skb))
257 *validate = true;
258 /* check the reason of requeuing without tx lock first */
259 txq = skb_get_tx_queue(txq->dev, skb);
260 if (!netif_xmit_frozen_or_stopped(txq)) {
261 skb = __skb_dequeue(&q->gso_skb);
262 if (qdisc_is_percpu_stats(q)) {
263 qdisc_qstats_cpu_backlog_dec(q, skb);
264 qdisc_qstats_cpu_qlen_dec(q);
265 } else {
266 qdisc_qstats_backlog_dec(q, skb);
267 q->q.qlen--;
268 }
269 } else {
270 skb = NULL;
271 qdisc_maybe_clear_missed(q, txq);
272 }
273 if (lock)
274 spin_unlock(lock);
275 goto trace;
276 }
277validate:
278 *validate = true;
279
280 if ((q->flags & TCQ_F_ONETXQUEUE) &&
281 netif_xmit_frozen_or_stopped(txq)) {
282 qdisc_maybe_clear_missed(q, txq);
283 return skb;
284 }
285
286 skb = qdisc_dequeue_skb_bad_txq(q);
287 if (unlikely(skb)) {
288 if (skb == SKB_XOFF_MAGIC)
289 return NULL;
290 goto bulk;
291 }
292 skb = q->dequeue(q);
293 if (skb) {
294bulk:
295 if (qdisc_may_bulk(q))
296 try_bulk_dequeue_skb(q, skb, txq, packets);
297 else
298 try_bulk_dequeue_skb_slow(q, skb, packets);
299 }
300trace:
301 trace_qdisc_dequeue(q, txq, *packets, skb);
302 return skb;
303}
304
305/*
306 * Transmit possibly several skbs, and handle the return status as
307 * required. Owning running seqcount bit guarantees that
308 * only one CPU can execute this function.
309 *
310 * Returns to the caller:
311 * false - hardware queue frozen backoff
312 * true - feel free to send more pkts
313 */
314bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
315 struct net_device *dev, struct netdev_queue *txq,
316 spinlock_t *root_lock, bool validate)
317{
318 int ret = NETDEV_TX_BUSY;
319 bool again = false;
320
321 /* And release qdisc */
322 if (root_lock)
323 spin_unlock(root_lock);
324
325 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
326 if (validate)
327 skb = validate_xmit_skb_list(skb, dev, &again);
328
329#ifdef CONFIG_XFRM_OFFLOAD
330 if (unlikely(again)) {
331 if (root_lock)
332 spin_lock(root_lock);
333
334 dev_requeue_skb(skb, q);
335 return false;
336 }
337#endif
338
339 if (likely(skb)) {
340 HARD_TX_LOCK(dev, txq, smp_processor_id());
341 if (!netif_xmit_frozen_or_stopped(txq))
342 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
343 else
344 qdisc_maybe_clear_missed(q, txq);
345
346 HARD_TX_UNLOCK(dev, txq);
347 } else {
348 if (root_lock)
349 spin_lock(root_lock);
350 return true;
351 }
352
353 if (root_lock)
354 spin_lock(root_lock);
355
356 if (!dev_xmit_complete(ret)) {
357 /* Driver returned NETDEV_TX_BUSY - requeue skb */
358 if (unlikely(ret != NETDEV_TX_BUSY))
359 net_warn_ratelimited("BUG %s code %d qlen %d\n",
360 dev->name, ret, q->q.qlen);
361
362 dev_requeue_skb(skb, q);
363 return false;
364 }
365
366 return true;
367}
368
369/*
370 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
371 *
372 * running seqcount guarantees only one CPU can process
373 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
374 * this queue.
375 *
376 * netif_tx_lock serializes accesses to device driver.
377 *
378 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
379 * if one is grabbed, another must be free.
380 *
381 * Note, that this procedure can be called by a watchdog timer
382 *
383 * Returns to the caller:
384 * 0 - queue is empty or throttled.
385 * >0 - queue is not empty.
386 *
387 */
388static inline bool qdisc_restart(struct Qdisc *q, int *packets)
389{
390 spinlock_t *root_lock = NULL;
391 struct netdev_queue *txq;
392 struct net_device *dev;
393 struct sk_buff *skb;
394 bool validate;
395
396 /* Dequeue packet */
397 skb = dequeue_skb(q, &validate, packets);
398 if (unlikely(!skb))
399 return false;
400
401 if (!(q->flags & TCQ_F_NOLOCK))
402 root_lock = qdisc_lock(q);
403
404 dev = qdisc_dev(q);
405 txq = skb_get_tx_queue(dev, skb);
406
407 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
408}
409
410void __qdisc_run(struct Qdisc *q)
411{
412 int quota = dev_tx_weight;
413 int packets;
414
415 while (qdisc_restart(q, &packets)) {
416 quota -= packets;
417 if (quota <= 0) {
418 if (q->flags & TCQ_F_NOLOCK)
419 set_bit(__QDISC_STATE_MISSED, &q->state);
420 else
421 __netif_schedule(q);
422
423 break;
424 }
425 }
426}
427
428unsigned long dev_trans_start(struct net_device *dev)
429{
430 unsigned long val, res;
431 unsigned int i;
432
433 if (is_vlan_dev(dev))
434 dev = vlan_dev_real_dev(dev);
435 else if (netif_is_macvlan(dev))
436 dev = macvlan_dev_real_dev(dev);
437 res = netdev_get_tx_queue(dev, 0)->trans_start;
438 for (i = 1; i < dev->num_tx_queues; i++) {
439 val = netdev_get_tx_queue(dev, i)->trans_start;
440 if (val && time_after(val, res))
441 res = val;
442 }
443
444 return res;
445}
446EXPORT_SYMBOL(dev_trans_start);
447
448static void dev_watchdog(struct timer_list *t)
449{
450 struct net_device *dev = from_timer(dev, t, watchdog_timer);
451
452 netif_tx_lock(dev);
453 if (!qdisc_tx_is_noop(dev)) {
454 if (netif_device_present(dev) &&
455 netif_running(dev) &&
456 netif_carrier_ok(dev)) {
457 int some_queue_timedout = 0;
458 unsigned int i;
459 unsigned long trans_start;
460
461 for (i = 0; i < dev->num_tx_queues; i++) {
462 struct netdev_queue *txq;
463
464 txq = netdev_get_tx_queue(dev, i);
465 trans_start = txq->trans_start;
466 if (netif_xmit_stopped(txq) &&
467 time_after(jiffies, (trans_start +
468 dev->watchdog_timeo))) {
469 some_queue_timedout = 1;
470 txq->trans_timeout++;
471 break;
472 }
473 }
474
475 if (some_queue_timedout) {
476 trace_net_dev_xmit_timeout(dev, i);
477 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
478 dev->name, netdev_drivername(dev), i);
479 dev->netdev_ops->ndo_tx_timeout(dev, i);
480 }
481 if (!mod_timer(&dev->watchdog_timer,
482 round_jiffies(jiffies +
483 dev->watchdog_timeo)))
484 dev_hold(dev);
485 }
486 }
487 netif_tx_unlock(dev);
488
489 dev_put(dev);
490}
491
492void __netdev_watchdog_up(struct net_device *dev)
493{
494 if (dev->netdev_ops->ndo_tx_timeout) {
495 if (dev->watchdog_timeo <= 0)
496 dev->watchdog_timeo = 5*HZ;
497 if (!mod_timer(&dev->watchdog_timer,
498 round_jiffies(jiffies + dev->watchdog_timeo)))
499 dev_hold(dev);
500 }
501}
502EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
503
504static void dev_watchdog_up(struct net_device *dev)
505{
506 __netdev_watchdog_up(dev);
507}
508
509static void dev_watchdog_down(struct net_device *dev)
510{
511 netif_tx_lock_bh(dev);
512 if (del_timer(&dev->watchdog_timer))
513 dev_put(dev);
514 netif_tx_unlock_bh(dev);
515}
516
517/**
518 * netif_carrier_on - set carrier
519 * @dev: network device
520 *
521 * Device has detected acquisition of carrier.
522 */
523void netif_carrier_on(struct net_device *dev)
524{
525 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
526 if (dev->reg_state == NETREG_UNINITIALIZED)
527 return;
528 atomic_inc(&dev->carrier_up_count);
529 linkwatch_fire_event(dev);
530 if (netif_running(dev))
531 __netdev_watchdog_up(dev);
532 }
533}
534EXPORT_SYMBOL(netif_carrier_on);
535
536/**
537 * netif_carrier_off - clear carrier
538 * @dev: network device
539 *
540 * Device has detected loss of carrier.
541 */
542void netif_carrier_off(struct net_device *dev)
543{
544 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
545 if (dev->reg_state == NETREG_UNINITIALIZED)
546 return;
547 atomic_inc(&dev->carrier_down_count);
548 linkwatch_fire_event(dev);
549 }
550}
551EXPORT_SYMBOL(netif_carrier_off);
552
553/**
554 * netif_carrier_event - report carrier state event
555 * @dev: network device
556 *
557 * Device has detected a carrier event but the carrier state wasn't changed.
558 * Use in drivers when querying carrier state asynchronously, to avoid missing
559 * events (link flaps) if link recovers before it's queried.
560 */
561void netif_carrier_event(struct net_device *dev)
562{
563 if (dev->reg_state == NETREG_UNINITIALIZED)
564 return;
565 atomic_inc(&dev->carrier_up_count);
566 atomic_inc(&dev->carrier_down_count);
567 linkwatch_fire_event(dev);
568}
569EXPORT_SYMBOL_GPL(netif_carrier_event);
570
571/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
572 under all circumstances. It is difficult to invent anything faster or
573 cheaper.
574 */
575
576static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
577 struct sk_buff **to_free)
578{
579 __qdisc_drop(skb, to_free);
580 return NET_XMIT_CN;
581}
582
583static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
584{
585 return NULL;
586}
587
588struct Qdisc_ops noop_qdisc_ops __read_mostly = {
589 .id = "noop",
590 .priv_size = 0,
591 .enqueue = noop_enqueue,
592 .dequeue = noop_dequeue,
593 .peek = noop_dequeue,
594 .owner = THIS_MODULE,
595};
596
597static struct netdev_queue noop_netdev_queue = {
598 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
599 .qdisc_sleeping = &noop_qdisc,
600};
601
602struct Qdisc noop_qdisc = {
603 .enqueue = noop_enqueue,
604 .dequeue = noop_dequeue,
605 .flags = TCQ_F_BUILTIN,
606 .ops = &noop_qdisc_ops,
607 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
608 .dev_queue = &noop_netdev_queue,
609 .running = SEQCNT_ZERO(noop_qdisc.running),
610 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
611 .gso_skb = {
612 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
613 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
614 .qlen = 0,
615 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
616 },
617 .skb_bad_txq = {
618 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
619 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
620 .qlen = 0,
621 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
622 },
623};
624EXPORT_SYMBOL(noop_qdisc);
625
626static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
627 struct netlink_ext_ack *extack)
628{
629 /* register_qdisc() assigns a default of noop_enqueue if unset,
630 * but __dev_queue_xmit() treats noqueue only as such
631 * if this is NULL - so clear it here. */
632 qdisc->enqueue = NULL;
633 return 0;
634}
635
636struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
637 .id = "noqueue",
638 .priv_size = 0,
639 .init = noqueue_init,
640 .enqueue = noop_enqueue,
641 .dequeue = noop_dequeue,
642 .peek = noop_dequeue,
643 .owner = THIS_MODULE,
644};
645
646static const u8 prio2band[TC_PRIO_MAX + 1] = {
647 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
648};
649
650/* 3-band FIFO queue: old style, but should be a bit faster than
651 generic prio+fifo combination.
652 */
653
654#define PFIFO_FAST_BANDS 3
655
656/*
657 * Private data for a pfifo_fast scheduler containing:
658 * - rings for priority bands
659 */
660struct pfifo_fast_priv {
661 struct skb_array q[PFIFO_FAST_BANDS];
662};
663
664static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
665 int band)
666{
667 return &priv->q[band];
668}
669
670static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
671 struct sk_buff **to_free)
672{
673 int band = prio2band[skb->priority & TC_PRIO_MAX];
674 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
675 struct skb_array *q = band2list(priv, band);
676 unsigned int pkt_len = qdisc_pkt_len(skb);
677 int err;
678
679 err = skb_array_produce(q, skb);
680
681 if (unlikely(err)) {
682 if (qdisc_is_percpu_stats(qdisc))
683 return qdisc_drop_cpu(skb, qdisc, to_free);
684 else
685 return qdisc_drop(skb, qdisc, to_free);
686 }
687
688 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
689 return NET_XMIT_SUCCESS;
690}
691
692static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
693{
694 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
695 struct sk_buff *skb = NULL;
696 bool need_retry = true;
697 int band;
698
699retry:
700 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
701 struct skb_array *q = band2list(priv, band);
702
703 if (__skb_array_empty(q))
704 continue;
705
706 skb = __skb_array_consume(q);
707 }
708 if (likely(skb)) {
709 qdisc_update_stats_at_dequeue(qdisc, skb);
710 } else if (need_retry &&
711 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
712 /* Delay clearing the STATE_MISSED here to reduce
713 * the overhead of the second spin_trylock() in
714 * qdisc_run_begin() and __netif_schedule() calling
715 * in qdisc_run_end().
716 */
717 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
718 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
719
720 /* Make sure dequeuing happens after clearing
721 * STATE_MISSED.
722 */
723 smp_mb__after_atomic();
724
725 need_retry = false;
726
727 goto retry;
728 }
729
730 return skb;
731}
732
733static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
734{
735 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
736 struct sk_buff *skb = NULL;
737 int band;
738
739 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
740 struct skb_array *q = band2list(priv, band);
741
742 skb = __skb_array_peek(q);
743 }
744
745 return skb;
746}
747
748static void pfifo_fast_reset(struct Qdisc *qdisc)
749{
750 int i, band;
751 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
752
753 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
754 struct skb_array *q = band2list(priv, band);
755 struct sk_buff *skb;
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
763 while ((skb = __skb_array_consume(q)) != NULL)
764 kfree_skb(skb);
765 }
766
767 if (qdisc_is_percpu_stats(qdisc)) {
768 for_each_possible_cpu(i) {
769 struct gnet_stats_queue *q;
770
771 q = per_cpu_ptr(qdisc->cpu_qstats, i);
772 q->backlog = 0;
773 q->qlen = 0;
774 }
775 }
776}
777
778static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
779{
780 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
781
782 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
783 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
784 goto nla_put_failure;
785 return skb->len;
786
787nla_put_failure:
788 return -1;
789}
790
791static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
792 struct netlink_ext_ack *extack)
793{
794 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
795 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
796 int prio;
797
798 /* guard against zero length rings */
799 if (!qlen)
800 return -EINVAL;
801
802 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
803 struct skb_array *q = band2list(priv, prio);
804 int err;
805
806 err = skb_array_init(q, qlen, GFP_KERNEL);
807 if (err)
808 return -ENOMEM;
809 }
810
811 /* Can by-pass the queue discipline */
812 qdisc->flags |= TCQ_F_CAN_BYPASS;
813 return 0;
814}
815
816static void pfifo_fast_destroy(struct Qdisc *sch)
817{
818 struct pfifo_fast_priv *priv = qdisc_priv(sch);
819 int prio;
820
821 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
822 struct skb_array *q = band2list(priv, prio);
823
824 /* NULL ring is possible if destroy path is due to a failed
825 * skb_array_init() in pfifo_fast_init() case.
826 */
827 if (!q->ring.queue)
828 continue;
829 /* Destroy ring but no need to kfree_skb because a call to
830 * pfifo_fast_reset() has already done that work.
831 */
832 ptr_ring_cleanup(&q->ring, NULL);
833 }
834}
835
836static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
837 unsigned int new_len)
838{
839 struct pfifo_fast_priv *priv = qdisc_priv(sch);
840 struct skb_array *bands[PFIFO_FAST_BANDS];
841 int prio;
842
843 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
844 struct skb_array *q = band2list(priv, prio);
845
846 bands[prio] = q;
847 }
848
849 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
850 GFP_KERNEL);
851}
852
853struct Qdisc_ops pfifo_fast_ops __read_mostly = {
854 .id = "pfifo_fast",
855 .priv_size = sizeof(struct pfifo_fast_priv),
856 .enqueue = pfifo_fast_enqueue,
857 .dequeue = pfifo_fast_dequeue,
858 .peek = pfifo_fast_peek,
859 .init = pfifo_fast_init,
860 .destroy = pfifo_fast_destroy,
861 .reset = pfifo_fast_reset,
862 .dump = pfifo_fast_dump,
863 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
864 .owner = THIS_MODULE,
865 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
866};
867EXPORT_SYMBOL(pfifo_fast_ops);
868
869static struct lock_class_key qdisc_tx_busylock;
870static struct lock_class_key qdisc_running_key;
871
872struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
873 const struct Qdisc_ops *ops,
874 struct netlink_ext_ack *extack)
875{
876 struct Qdisc *sch;
877 unsigned int size = sizeof(*sch) + ops->priv_size;
878 int err = -ENOBUFS;
879 struct net_device *dev;
880
881 if (!dev_queue) {
882 NL_SET_ERR_MSG(extack, "No device queue given");
883 err = -EINVAL;
884 goto errout;
885 }
886
887 dev = dev_queue->dev;
888 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
889
890 if (!sch)
891 goto errout;
892 __skb_queue_head_init(&sch->gso_skb);
893 __skb_queue_head_init(&sch->skb_bad_txq);
894 qdisc_skb_head_init(&sch->q);
895 spin_lock_init(&sch->q.lock);
896
897 if (ops->static_flags & TCQ_F_CPUSTATS) {
898 sch->cpu_bstats =
899 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
900 if (!sch->cpu_bstats)
901 goto errout1;
902
903 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
904 if (!sch->cpu_qstats) {
905 free_percpu(sch->cpu_bstats);
906 goto errout1;
907 }
908 }
909
910 spin_lock_init(&sch->busylock);
911 lockdep_set_class(&sch->busylock,
912 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
913
914 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
915 spin_lock_init(&sch->seqlock);
916 lockdep_set_class(&sch->seqlock,
917 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
918
919 seqcount_init(&sch->running);
920 lockdep_set_class(&sch->running,
921 dev->qdisc_running_key ?: &qdisc_running_key);
922
923 sch->ops = ops;
924 sch->flags = ops->static_flags;
925 sch->enqueue = ops->enqueue;
926 sch->dequeue = ops->dequeue;
927 sch->dev_queue = dev_queue;
928 dev_hold(dev);
929 refcount_set(&sch->refcnt, 1);
930
931 return sch;
932errout1:
933 kfree(sch);
934errout:
935 return ERR_PTR(err);
936}
937
938struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
939 const struct Qdisc_ops *ops,
940 unsigned int parentid,
941 struct netlink_ext_ack *extack)
942{
943 struct Qdisc *sch;
944
945 if (!try_module_get(ops->owner)) {
946 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
947 return NULL;
948 }
949
950 sch = qdisc_alloc(dev_queue, ops, extack);
951 if (IS_ERR(sch)) {
952 module_put(ops->owner);
953 return NULL;
954 }
955 sch->parent = parentid;
956
957 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
958 trace_qdisc_create(ops, dev_queue->dev, parentid);
959 return sch;
960 }
961
962 qdisc_put(sch);
963 return NULL;
964}
965EXPORT_SYMBOL(qdisc_create_dflt);
966
967/* Under qdisc_lock(qdisc) and BH! */
968
969void qdisc_reset(struct Qdisc *qdisc)
970{
971 const struct Qdisc_ops *ops = qdisc->ops;
972 struct sk_buff *skb, *tmp;
973
974 trace_qdisc_reset(qdisc);
975
976 if (ops->reset)
977 ops->reset(qdisc);
978
979 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
980 __skb_unlink(skb, &qdisc->gso_skb);
981 kfree_skb_list(skb);
982 }
983
984 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
985 __skb_unlink(skb, &qdisc->skb_bad_txq);
986 kfree_skb_list(skb);
987 }
988
989 qdisc->q.qlen = 0;
990 qdisc->qstats.backlog = 0;
991}
992EXPORT_SYMBOL(qdisc_reset);
993
994void qdisc_free(struct Qdisc *qdisc)
995{
996 if (qdisc_is_percpu_stats(qdisc)) {
997 free_percpu(qdisc->cpu_bstats);
998 free_percpu(qdisc->cpu_qstats);
999 }
1000
1001 kfree(qdisc);
1002}
1003
1004static void qdisc_free_cb(struct rcu_head *head)
1005{
1006 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1007
1008 qdisc_free(q);
1009}
1010
1011static void qdisc_destroy(struct Qdisc *qdisc)
1012{
1013 const struct Qdisc_ops *ops = qdisc->ops;
1014
1015#ifdef CONFIG_NET_SCHED
1016 qdisc_hash_del(qdisc);
1017
1018 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1019#endif
1020 gen_kill_estimator(&qdisc->rate_est);
1021
1022 qdisc_reset(qdisc);
1023
1024 if (ops->destroy)
1025 ops->destroy(qdisc);
1026
1027 module_put(ops->owner);
1028 dev_put(qdisc_dev(qdisc));
1029
1030 trace_qdisc_destroy(qdisc);
1031
1032 call_rcu(&qdisc->rcu, qdisc_free_cb);
1033}
1034
1035void qdisc_put(struct Qdisc *qdisc)
1036{
1037 if (!qdisc)
1038 return;
1039
1040 if (qdisc->flags & TCQ_F_BUILTIN ||
1041 !refcount_dec_and_test(&qdisc->refcnt))
1042 return;
1043
1044 qdisc_destroy(qdisc);
1045}
1046EXPORT_SYMBOL(qdisc_put);
1047
1048/* Version of qdisc_put() that is called with rtnl mutex unlocked.
1049 * Intended to be used as optimization, this function only takes rtnl lock if
1050 * qdisc reference counter reached zero.
1051 */
1052
1053void qdisc_put_unlocked(struct Qdisc *qdisc)
1054{
1055 if (qdisc->flags & TCQ_F_BUILTIN ||
1056 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1057 return;
1058
1059 qdisc_destroy(qdisc);
1060 rtnl_unlock();
1061}
1062EXPORT_SYMBOL(qdisc_put_unlocked);
1063
1064/* Attach toplevel qdisc to device queue. */
1065struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1066 struct Qdisc *qdisc)
1067{
1068 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1069 spinlock_t *root_lock;
1070
1071 root_lock = qdisc_lock(oqdisc);
1072 spin_lock_bh(root_lock);
1073
1074 /* ... and graft new one */
1075 if (qdisc == NULL)
1076 qdisc = &noop_qdisc;
1077 dev_queue->qdisc_sleeping = qdisc;
1078 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1079
1080 spin_unlock_bh(root_lock);
1081
1082 return oqdisc;
1083}
1084EXPORT_SYMBOL(dev_graft_qdisc);
1085
1086static void attach_one_default_qdisc(struct net_device *dev,
1087 struct netdev_queue *dev_queue,
1088 void *_unused)
1089{
1090 struct Qdisc *qdisc;
1091 const struct Qdisc_ops *ops = default_qdisc_ops;
1092
1093 if (dev->priv_flags & IFF_NO_QUEUE)
1094 ops = &noqueue_qdisc_ops;
1095 else if(dev->type == ARPHRD_CAN)
1096 ops = &pfifo_fast_ops;
1097
1098 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1099 if (!qdisc)
1100 return;
1101
1102 if (!netif_is_multiqueue(dev))
1103 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1104 dev_queue->qdisc_sleeping = qdisc;
1105}
1106
1107static void attach_default_qdiscs(struct net_device *dev)
1108{
1109 struct netdev_queue *txq;
1110 struct Qdisc *qdisc;
1111
1112 txq = netdev_get_tx_queue(dev, 0);
1113
1114 if (!netif_is_multiqueue(dev) ||
1115 dev->priv_flags & IFF_NO_QUEUE) {
1116 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1117 dev->qdisc = txq->qdisc_sleeping;
1118 qdisc_refcount_inc(dev->qdisc);
1119 } else {
1120 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1121 if (qdisc) {
1122 dev->qdisc = qdisc;
1123 qdisc->ops->attach(qdisc);
1124 }
1125 }
1126
1127 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1128 if (dev->qdisc == &noop_qdisc) {
1129 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1130 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1131 dev->priv_flags |= IFF_NO_QUEUE;
1132 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1133 dev->qdisc = txq->qdisc_sleeping;
1134 qdisc_refcount_inc(dev->qdisc);
1135 dev->priv_flags ^= IFF_NO_QUEUE;
1136 }
1137
1138#ifdef CONFIG_NET_SCHED
1139 if (dev->qdisc != &noop_qdisc)
1140 qdisc_hash_add(dev->qdisc, false);
1141#endif
1142}
1143
1144static void transition_one_qdisc(struct net_device *dev,
1145 struct netdev_queue *dev_queue,
1146 void *_need_watchdog)
1147{
1148 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1149 int *need_watchdog_p = _need_watchdog;
1150
1151 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1152 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1153
1154 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1155 if (need_watchdog_p) {
1156 dev_queue->trans_start = 0;
1157 *need_watchdog_p = 1;
1158 }
1159}
1160
1161void dev_activate(struct net_device *dev)
1162{
1163 int need_watchdog;
1164
1165 /* No queueing discipline is attached to device;
1166 * create default one for devices, which need queueing
1167 * and noqueue_qdisc for virtual interfaces
1168 */
1169
1170 if (dev->qdisc == &noop_qdisc)
1171 attach_default_qdiscs(dev);
1172
1173 if (!netif_carrier_ok(dev))
1174 /* Delay activation until next carrier-on event */
1175 return;
1176
1177 need_watchdog = 0;
1178 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1179 if (dev_ingress_queue(dev))
1180 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1181
1182 if (need_watchdog) {
1183 netif_trans_update(dev);
1184 dev_watchdog_up(dev);
1185 }
1186}
1187EXPORT_SYMBOL(dev_activate);
1188
1189static void qdisc_deactivate(struct Qdisc *qdisc)
1190{
1191 if (qdisc->flags & TCQ_F_BUILTIN)
1192 return;
1193
1194 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1195}
1196
1197static void dev_deactivate_queue(struct net_device *dev,
1198 struct netdev_queue *dev_queue,
1199 void *_qdisc_default)
1200{
1201 struct Qdisc *qdisc_default = _qdisc_default;
1202 struct Qdisc *qdisc;
1203
1204 qdisc = rtnl_dereference(dev_queue->qdisc);
1205 if (qdisc) {
1206 qdisc_deactivate(qdisc);
1207 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1208 }
1209}
1210
1211static void dev_reset_queue(struct net_device *dev,
1212 struct netdev_queue *dev_queue,
1213 void *_unused)
1214{
1215 struct Qdisc *qdisc;
1216 bool nolock;
1217
1218 qdisc = dev_queue->qdisc_sleeping;
1219 if (!qdisc)
1220 return;
1221
1222 nolock = qdisc->flags & TCQ_F_NOLOCK;
1223
1224 if (nolock)
1225 spin_lock_bh(&qdisc->seqlock);
1226 spin_lock_bh(qdisc_lock(qdisc));
1227
1228 qdisc_reset(qdisc);
1229
1230 spin_unlock_bh(qdisc_lock(qdisc));
1231 if (nolock) {
1232 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1233 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1234 spin_unlock_bh(&qdisc->seqlock);
1235 }
1236}
1237
1238static bool some_qdisc_is_busy(struct net_device *dev)
1239{
1240 unsigned int i;
1241
1242 for (i = 0; i < dev->num_tx_queues; i++) {
1243 struct netdev_queue *dev_queue;
1244 spinlock_t *root_lock;
1245 struct Qdisc *q;
1246 int val;
1247
1248 dev_queue = netdev_get_tx_queue(dev, i);
1249 q = dev_queue->qdisc_sleeping;
1250
1251 root_lock = qdisc_lock(q);
1252 spin_lock_bh(root_lock);
1253
1254 val = (qdisc_is_running(q) ||
1255 test_bit(__QDISC_STATE_SCHED, &q->state));
1256
1257 spin_unlock_bh(root_lock);
1258
1259 if (val)
1260 return true;
1261 }
1262 return false;
1263}
1264
1265/**
1266 * dev_deactivate_many - deactivate transmissions on several devices
1267 * @head: list of devices to deactivate
1268 *
1269 * This function returns only when all outstanding transmissions
1270 * have completed, unless all devices are in dismantle phase.
1271 */
1272void dev_deactivate_many(struct list_head *head)
1273{
1274 struct net_device *dev;
1275
1276 list_for_each_entry(dev, head, close_list) {
1277 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1278 &noop_qdisc);
1279 if (dev_ingress_queue(dev))
1280 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1281 &noop_qdisc);
1282
1283 dev_watchdog_down(dev);
1284 }
1285
1286 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1287 * outstanding qdisc enqueuing calls.
1288 * This is avoided if all devices are in dismantle phase :
1289 * Caller will call synchronize_net() for us
1290 */
1291 synchronize_net();
1292
1293 list_for_each_entry(dev, head, close_list) {
1294 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1295
1296 if (dev_ingress_queue(dev))
1297 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1298 }
1299
1300 /* Wait for outstanding qdisc_run calls. */
1301 list_for_each_entry(dev, head, close_list) {
1302 while (some_qdisc_is_busy(dev)) {
1303 /* wait_event() would avoid this sleep-loop but would
1304 * require expensive checks in the fast paths of packet
1305 * processing which isn't worth it.
1306 */
1307 schedule_timeout_uninterruptible(1);
1308 }
1309 }
1310}
1311
1312void dev_deactivate(struct net_device *dev)
1313{
1314 LIST_HEAD(single);
1315
1316 list_add(&dev->close_list, &single);
1317 dev_deactivate_many(&single);
1318 list_del(&single);
1319}
1320EXPORT_SYMBOL(dev_deactivate);
1321
1322static int qdisc_change_tx_queue_len(struct net_device *dev,
1323 struct netdev_queue *dev_queue)
1324{
1325 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1326 const struct Qdisc_ops *ops = qdisc->ops;
1327
1328 if (ops->change_tx_queue_len)
1329 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1330 return 0;
1331}
1332
1333int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1334{
1335 bool up = dev->flags & IFF_UP;
1336 unsigned int i;
1337 int ret = 0;
1338
1339 if (up)
1340 dev_deactivate(dev);
1341
1342 for (i = 0; i < dev->num_tx_queues; i++) {
1343 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1344
1345 /* TODO: revert changes on a partial failure */
1346 if (ret)
1347 break;
1348 }
1349
1350 if (up)
1351 dev_activate(dev);
1352 return ret;
1353}
1354
1355static void dev_init_scheduler_queue(struct net_device *dev,
1356 struct netdev_queue *dev_queue,
1357 void *_qdisc)
1358{
1359 struct Qdisc *qdisc = _qdisc;
1360
1361 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1362 dev_queue->qdisc_sleeping = qdisc;
1363}
1364
1365void dev_init_scheduler(struct net_device *dev)
1366{
1367 dev->qdisc = &noop_qdisc;
1368 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1369 if (dev_ingress_queue(dev))
1370 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1371
1372 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1373}
1374
1375static void shutdown_scheduler_queue(struct net_device *dev,
1376 struct netdev_queue *dev_queue,
1377 void *_qdisc_default)
1378{
1379 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1380 struct Qdisc *qdisc_default = _qdisc_default;
1381
1382 if (qdisc) {
1383 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1384 dev_queue->qdisc_sleeping = qdisc_default;
1385
1386 qdisc_put(qdisc);
1387 }
1388}
1389
1390void dev_shutdown(struct net_device *dev)
1391{
1392 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1393 if (dev_ingress_queue(dev))
1394 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1395 qdisc_put(dev->qdisc);
1396 dev->qdisc = &noop_qdisc;
1397
1398 WARN_ON(timer_pending(&dev->watchdog_timer));
1399}
1400
1401/**
1402 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1403 * @rate: Rate to compute reciprocal division values of
1404 * @mult: Multiplier for reciprocal division
1405 * @shift: Shift for reciprocal division
1406 *
1407 * The multiplier and shift for reciprocal division by rate are stored
1408 * in mult and shift.
1409 *
1410 * The deal here is to replace a divide by a reciprocal one
1411 * in fast path (a reciprocal divide is a multiply and a shift)
1412 *
1413 * Normal formula would be :
1414 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1415 *
1416 * We compute mult/shift to use instead :
1417 * time_in_ns = (len * mult) >> shift;
1418 *
1419 * We try to get the highest possible mult value for accuracy,
1420 * but have to make sure no overflows will ever happen.
1421 *
1422 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1423 */
1424static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1425{
1426 u64 factor = NSEC_PER_SEC;
1427
1428 *mult = 1;
1429 *shift = 0;
1430
1431 if (rate <= 0)
1432 return;
1433
1434 for (;;) {
1435 *mult = div64_u64(factor, rate);
1436 if (*mult & (1U << 31) || factor & (1ULL << 63))
1437 break;
1438 factor <<= 1;
1439 (*shift)++;
1440 }
1441}
1442
1443void psched_ratecfg_precompute(struct psched_ratecfg *r,
1444 const struct tc_ratespec *conf,
1445 u64 rate64)
1446{
1447 memset(r, 0, sizeof(*r));
1448 r->overhead = conf->overhead;
1449 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1450 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1451 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1452}
1453EXPORT_SYMBOL(psched_ratecfg_precompute);
1454
1455void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1456{
1457 r->rate_pkts_ps = pktrate64;
1458 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1459}
1460EXPORT_SYMBOL(psched_ppscfg_precompute);
1461
1462static void mini_qdisc_rcu_func(struct rcu_head *head)
1463{
1464}
1465
1466void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1467 struct tcf_proto *tp_head)
1468{
1469 /* Protected with chain0->filter_chain_lock.
1470 * Can't access chain directly because tp_head can be NULL.
1471 */
1472 struct mini_Qdisc *miniq_old =
1473 rcu_dereference_protected(*miniqp->p_miniq, 1);
1474 struct mini_Qdisc *miniq;
1475
1476 if (!tp_head) {
1477 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1478 /* Wait for flying RCU callback before it is freed. */
1479 rcu_barrier();
1480 return;
1481 }
1482
1483 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1484 &miniqp->miniq1 : &miniqp->miniq2;
1485
1486 /* We need to make sure that readers won't see the miniq
1487 * we are about to modify. So wait until previous call_rcu callback
1488 * is done.
1489 */
1490 rcu_barrier();
1491 miniq->filter_list = tp_head;
1492 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1493
1494 if (miniq_old)
1495 /* This is counterpart of the rcu barriers above. We need to
1496 * block potential new user of miniq_old until all readers
1497 * are not seeing it.
1498 */
1499 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1500}
1501EXPORT_SYMBOL(mini_qdisc_pair_swap);
1502
1503void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1504 struct tcf_block *block)
1505{
1506 miniqp->miniq1.block = block;
1507 miniqp->miniq2.block = block;
1508}
1509EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1510
1511void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1512 struct mini_Qdisc __rcu **p_miniq)
1513{
1514 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1515 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1516 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1517 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1518 miniqp->p_miniq = p_miniq;
1519}
1520EXPORT_SYMBOL(mini_qdisc_pair_init);
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 <net/hotdata.h>
31#include <trace/events/qdisc.h>
32#include <trace/events/net.h>
33#include <net/xfrm.h>
34
35/* Qdisc to use by default */
36const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
37EXPORT_SYMBOL(default_qdisc_ops);
38
39static void qdisc_maybe_clear_missed(struct Qdisc *q,
40 const struct netdev_queue *txq)
41{
42 clear_bit(__QDISC_STATE_MISSED, &q->state);
43
44 /* Make sure the below netif_xmit_frozen_or_stopped()
45 * checking happens after clearing STATE_MISSED.
46 */
47 smp_mb__after_atomic();
48
49 /* Checking netif_xmit_frozen_or_stopped() again to
50 * make sure STATE_MISSED is set if the STATE_MISSED
51 * set by netif_tx_wake_queue()'s rescheduling of
52 * net_tx_action() is cleared by the above clear_bit().
53 */
54 if (!netif_xmit_frozen_or_stopped(txq))
55 set_bit(__QDISC_STATE_MISSED, &q->state);
56 else
57 set_bit(__QDISC_STATE_DRAINING, &q->state);
58}
59
60/* Main transmission queue. */
61
62/* Modifications to data participating in scheduling must be protected with
63 * qdisc_lock(qdisc) spinlock.
64 *
65 * The idea is the following:
66 * - enqueue, dequeue are serialized via qdisc root lock
67 * - ingress filtering is also serialized via qdisc root lock
68 * - updates to tree and tree walking are only done under the rtnl mutex.
69 */
70
71#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
72
73static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
74{
75 const struct netdev_queue *txq = q->dev_queue;
76 spinlock_t *lock = NULL;
77 struct sk_buff *skb;
78
79 if (q->flags & TCQ_F_NOLOCK) {
80 lock = qdisc_lock(q);
81 spin_lock(lock);
82 }
83
84 skb = skb_peek(&q->skb_bad_txq);
85 if (skb) {
86 /* check the reason of requeuing without tx lock first */
87 txq = skb_get_tx_queue(txq->dev, skb);
88 if (!netif_xmit_frozen_or_stopped(txq)) {
89 skb = __skb_dequeue(&q->skb_bad_txq);
90 if (qdisc_is_percpu_stats(q)) {
91 qdisc_qstats_cpu_backlog_dec(q, skb);
92 qdisc_qstats_cpu_qlen_dec(q);
93 } else {
94 qdisc_qstats_backlog_dec(q, skb);
95 q->q.qlen--;
96 }
97 } else {
98 skb = SKB_XOFF_MAGIC;
99 qdisc_maybe_clear_missed(q, txq);
100 }
101 }
102
103 if (lock)
104 spin_unlock(lock);
105
106 return skb;
107}
108
109static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
110{
111 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
112
113 if (unlikely(skb))
114 skb = __skb_dequeue_bad_txq(q);
115
116 return skb;
117}
118
119static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
120 struct sk_buff *skb)
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 __skb_queue_tail(&q->skb_bad_txq, skb);
130
131 if (qdisc_is_percpu_stats(q)) {
132 qdisc_qstats_cpu_backlog_inc(q, skb);
133 qdisc_qstats_cpu_qlen_inc(q);
134 } else {
135 qdisc_qstats_backlog_inc(q, skb);
136 q->q.qlen++;
137 }
138
139 if (lock)
140 spin_unlock(lock);
141}
142
143static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
144{
145 spinlock_t *lock = NULL;
146
147 if (q->flags & TCQ_F_NOLOCK) {
148 lock = qdisc_lock(q);
149 spin_lock(lock);
150 }
151
152 while (skb) {
153 struct sk_buff *next = skb->next;
154
155 __skb_queue_tail(&q->gso_skb, skb);
156
157 /* it's still part of the queue */
158 if (qdisc_is_percpu_stats(q)) {
159 qdisc_qstats_cpu_requeues_inc(q);
160 qdisc_qstats_cpu_backlog_inc(q, skb);
161 qdisc_qstats_cpu_qlen_inc(q);
162 } else {
163 q->qstats.requeues++;
164 qdisc_qstats_backlog_inc(q, skb);
165 q->q.qlen++;
166 }
167
168 skb = next;
169 }
170
171 if (lock) {
172 spin_unlock(lock);
173 set_bit(__QDISC_STATE_MISSED, &q->state);
174 } else {
175 __netif_schedule(q);
176 }
177}
178
179static void try_bulk_dequeue_skb(struct Qdisc *q,
180 struct sk_buff *skb,
181 const struct netdev_queue *txq,
182 int *packets)
183{
184 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
185
186 while (bytelimit > 0) {
187 struct sk_buff *nskb = q->dequeue(q);
188
189 if (!nskb)
190 break;
191
192 bytelimit -= nskb->len; /* covers GSO len */
193 skb->next = nskb;
194 skb = nskb;
195 (*packets)++; /* GSO counts as one pkt */
196 }
197 skb_mark_not_on_list(skb);
198}
199
200/* This variant of try_bulk_dequeue_skb() makes sure
201 * all skbs in the chain are for the same txq
202 */
203static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
204 struct sk_buff *skb,
205 int *packets)
206{
207 int mapping = skb_get_queue_mapping(skb);
208 struct sk_buff *nskb;
209 int cnt = 0;
210
211 do {
212 nskb = q->dequeue(q);
213 if (!nskb)
214 break;
215 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
216 qdisc_enqueue_skb_bad_txq(q, nskb);
217 break;
218 }
219 skb->next = nskb;
220 skb = nskb;
221 } while (++cnt < 8);
222 (*packets) += cnt;
223 skb_mark_not_on_list(skb);
224}
225
226/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
227 * A requeued skb (via q->gso_skb) can also be a SKB list.
228 */
229static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
230 int *packets)
231{
232 const struct netdev_queue *txq = q->dev_queue;
233 struct sk_buff *skb = NULL;
234
235 *packets = 1;
236 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
237 spinlock_t *lock = NULL;
238
239 if (q->flags & TCQ_F_NOLOCK) {
240 lock = qdisc_lock(q);
241 spin_lock(lock);
242 }
243
244 skb = skb_peek(&q->gso_skb);
245
246 /* skb may be null if another cpu pulls gso_skb off in between
247 * empty check and lock.
248 */
249 if (!skb) {
250 if (lock)
251 spin_unlock(lock);
252 goto validate;
253 }
254
255 /* skb in gso_skb were already validated */
256 *validate = false;
257 if (xfrm_offload(skb))
258 *validate = true;
259 /* check the reason of requeuing without tx lock first */
260 txq = skb_get_tx_queue(txq->dev, skb);
261 if (!netif_xmit_frozen_or_stopped(txq)) {
262 skb = __skb_dequeue(&q->gso_skb);
263 if (qdisc_is_percpu_stats(q)) {
264 qdisc_qstats_cpu_backlog_dec(q, skb);
265 qdisc_qstats_cpu_qlen_dec(q);
266 } else {
267 qdisc_qstats_backlog_dec(q, skb);
268 q->q.qlen--;
269 }
270 } else {
271 skb = NULL;
272 qdisc_maybe_clear_missed(q, txq);
273 }
274 if (lock)
275 spin_unlock(lock);
276 goto trace;
277 }
278validate:
279 *validate = true;
280
281 if ((q->flags & TCQ_F_ONETXQUEUE) &&
282 netif_xmit_frozen_or_stopped(txq)) {
283 qdisc_maybe_clear_missed(q, txq);
284 return skb;
285 }
286
287 skb = qdisc_dequeue_skb_bad_txq(q);
288 if (unlikely(skb)) {
289 if (skb == SKB_XOFF_MAGIC)
290 return NULL;
291 goto bulk;
292 }
293 skb = q->dequeue(q);
294 if (skb) {
295bulk:
296 if (qdisc_may_bulk(q))
297 try_bulk_dequeue_skb(q, skb, txq, packets);
298 else
299 try_bulk_dequeue_skb_slow(q, skb, packets);
300 }
301trace:
302 trace_qdisc_dequeue(q, txq, *packets, skb);
303 return skb;
304}
305
306/*
307 * Transmit possibly several skbs, and handle the return status as
308 * required. Owning qdisc running bit guarantees that only one CPU
309 * can execute this function.
310 *
311 * Returns to the caller:
312 * false - hardware queue frozen backoff
313 * true - feel free to send more pkts
314 */
315bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
316 struct net_device *dev, struct netdev_queue *txq,
317 spinlock_t *root_lock, bool validate)
318{
319 int ret = NETDEV_TX_BUSY;
320 bool again = false;
321
322 /* And release qdisc */
323 if (root_lock)
324 spin_unlock(root_lock);
325
326 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
327 if (validate)
328 skb = validate_xmit_skb_list(skb, dev, &again);
329
330#ifdef CONFIG_XFRM_OFFLOAD
331 if (unlikely(again)) {
332 if (root_lock)
333 spin_lock(root_lock);
334
335 dev_requeue_skb(skb, q);
336 return false;
337 }
338#endif
339
340 if (likely(skb)) {
341 HARD_TX_LOCK(dev, txq, smp_processor_id());
342 if (!netif_xmit_frozen_or_stopped(txq))
343 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
344 else
345 qdisc_maybe_clear_missed(q, txq);
346
347 HARD_TX_UNLOCK(dev, txq);
348 } else {
349 if (root_lock)
350 spin_lock(root_lock);
351 return true;
352 }
353
354 if (root_lock)
355 spin_lock(root_lock);
356
357 if (!dev_xmit_complete(ret)) {
358 /* Driver returned NETDEV_TX_BUSY - requeue skb */
359 if (unlikely(ret != NETDEV_TX_BUSY))
360 net_warn_ratelimited("BUG %s code %d qlen %d\n",
361 dev->name, ret, q->q.qlen);
362
363 dev_requeue_skb(skb, q);
364 return false;
365 }
366
367 return true;
368}
369
370/*
371 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
372 *
373 * running seqcount guarantees only one CPU can process
374 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
375 * this queue.
376 *
377 * netif_tx_lock serializes accesses to device driver.
378 *
379 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
380 * if one is grabbed, another must be free.
381 *
382 * Note, that this procedure can be called by a watchdog timer
383 *
384 * Returns to the caller:
385 * 0 - queue is empty or throttled.
386 * >0 - queue is not empty.
387 *
388 */
389static inline bool qdisc_restart(struct Qdisc *q, int *packets)
390{
391 spinlock_t *root_lock = NULL;
392 struct netdev_queue *txq;
393 struct net_device *dev;
394 struct sk_buff *skb;
395 bool validate;
396
397 /* Dequeue packet */
398 skb = dequeue_skb(q, &validate, packets);
399 if (unlikely(!skb))
400 return false;
401
402 if (!(q->flags & TCQ_F_NOLOCK))
403 root_lock = qdisc_lock(q);
404
405 dev = qdisc_dev(q);
406 txq = skb_get_tx_queue(dev, skb);
407
408 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
409}
410
411void __qdisc_run(struct Qdisc *q)
412{
413 int quota = READ_ONCE(net_hotdata.dev_tx_weight);
414 int packets;
415
416 while (qdisc_restart(q, &packets)) {
417 quota -= packets;
418 if (quota <= 0) {
419 if (q->flags & TCQ_F_NOLOCK)
420 set_bit(__QDISC_STATE_MISSED, &q->state);
421 else
422 __netif_schedule(q);
423
424 break;
425 }
426 }
427}
428
429unsigned long dev_trans_start(struct net_device *dev)
430{
431 unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
432 unsigned long val;
433 unsigned int i;
434
435 for (i = 1; i < dev->num_tx_queues; i++) {
436 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
437 if (val && time_after(val, res))
438 res = val;
439 }
440
441 return res;
442}
443EXPORT_SYMBOL(dev_trans_start);
444
445static void netif_freeze_queues(struct net_device *dev)
446{
447 unsigned int i;
448 int cpu;
449
450 cpu = smp_processor_id();
451 for (i = 0; i < dev->num_tx_queues; i++) {
452 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
453
454 /* We are the only thread of execution doing a
455 * freeze, but we have to grab the _xmit_lock in
456 * order to synchronize with threads which are in
457 * the ->hard_start_xmit() handler and already
458 * checked the frozen bit.
459 */
460 __netif_tx_lock(txq, cpu);
461 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
462 __netif_tx_unlock(txq);
463 }
464}
465
466void netif_tx_lock(struct net_device *dev)
467{
468 spin_lock(&dev->tx_global_lock);
469 netif_freeze_queues(dev);
470}
471EXPORT_SYMBOL(netif_tx_lock);
472
473static void netif_unfreeze_queues(struct net_device *dev)
474{
475 unsigned int i;
476
477 for (i = 0; i < dev->num_tx_queues; i++) {
478 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
479
480 /* No need to grab the _xmit_lock here. If the
481 * queue is not stopped for another reason, we
482 * force a schedule.
483 */
484 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
485 netif_schedule_queue(txq);
486 }
487}
488
489void netif_tx_unlock(struct net_device *dev)
490{
491 netif_unfreeze_queues(dev);
492 spin_unlock(&dev->tx_global_lock);
493}
494EXPORT_SYMBOL(netif_tx_unlock);
495
496static void dev_watchdog(struct timer_list *t)
497{
498 struct net_device *dev = from_timer(dev, t, watchdog_timer);
499 bool release = true;
500
501 spin_lock(&dev->tx_global_lock);
502 if (!qdisc_tx_is_noop(dev)) {
503 if (netif_device_present(dev) &&
504 netif_running(dev) &&
505 netif_carrier_ok(dev)) {
506 unsigned int timedout_ms = 0;
507 unsigned int i;
508 unsigned long trans_start;
509 unsigned long oldest_start = jiffies;
510
511 for (i = 0; i < dev->num_tx_queues; i++) {
512 struct netdev_queue *txq;
513
514 txq = netdev_get_tx_queue(dev, i);
515 if (!netif_xmit_stopped(txq))
516 continue;
517
518 /* Paired with WRITE_ONCE() + smp_mb...() in
519 * netdev_tx_sent_queue() and netif_tx_stop_queue().
520 */
521 smp_mb();
522 trans_start = READ_ONCE(txq->trans_start);
523
524 if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
525 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
526 atomic_long_inc(&txq->trans_timeout);
527 break;
528 }
529 if (time_after(oldest_start, trans_start))
530 oldest_start = trans_start;
531 }
532
533 if (unlikely(timedout_ms)) {
534 trace_net_dev_xmit_timeout(dev, i);
535 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
536 raw_smp_processor_id(),
537 i, timedout_ms);
538 netif_freeze_queues(dev);
539 dev->netdev_ops->ndo_tx_timeout(dev, i);
540 netif_unfreeze_queues(dev);
541 }
542 if (!mod_timer(&dev->watchdog_timer,
543 round_jiffies(oldest_start +
544 dev->watchdog_timeo)))
545 release = false;
546 }
547 }
548 spin_unlock(&dev->tx_global_lock);
549
550 if (release)
551 netdev_put(dev, &dev->watchdog_dev_tracker);
552}
553
554void __netdev_watchdog_up(struct net_device *dev)
555{
556 if (dev->netdev_ops->ndo_tx_timeout) {
557 if (dev->watchdog_timeo <= 0)
558 dev->watchdog_timeo = 5*HZ;
559 if (!mod_timer(&dev->watchdog_timer,
560 round_jiffies(jiffies + dev->watchdog_timeo)))
561 netdev_hold(dev, &dev->watchdog_dev_tracker,
562 GFP_ATOMIC);
563 }
564}
565EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
566
567static void dev_watchdog_up(struct net_device *dev)
568{
569 __netdev_watchdog_up(dev);
570}
571
572static void dev_watchdog_down(struct net_device *dev)
573{
574 netif_tx_lock_bh(dev);
575 if (del_timer(&dev->watchdog_timer))
576 netdev_put(dev, &dev->watchdog_dev_tracker);
577 netif_tx_unlock_bh(dev);
578}
579
580/**
581 * netif_carrier_on - set carrier
582 * @dev: network device
583 *
584 * Device has detected acquisition of carrier.
585 */
586void netif_carrier_on(struct net_device *dev)
587{
588 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
589 if (dev->reg_state == NETREG_UNINITIALIZED)
590 return;
591 atomic_inc(&dev->carrier_up_count);
592 linkwatch_fire_event(dev);
593 if (netif_running(dev))
594 __netdev_watchdog_up(dev);
595 }
596}
597EXPORT_SYMBOL(netif_carrier_on);
598
599/**
600 * netif_carrier_off - clear carrier
601 * @dev: network device
602 *
603 * Device has detected loss of carrier.
604 */
605void netif_carrier_off(struct net_device *dev)
606{
607 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
608 if (dev->reg_state == NETREG_UNINITIALIZED)
609 return;
610 atomic_inc(&dev->carrier_down_count);
611 linkwatch_fire_event(dev);
612 }
613}
614EXPORT_SYMBOL(netif_carrier_off);
615
616/**
617 * netif_carrier_event - report carrier state event
618 * @dev: network device
619 *
620 * Device has detected a carrier event but the carrier state wasn't changed.
621 * Use in drivers when querying carrier state asynchronously, to avoid missing
622 * events (link flaps) if link recovers before it's queried.
623 */
624void netif_carrier_event(struct net_device *dev)
625{
626 if (dev->reg_state == NETREG_UNINITIALIZED)
627 return;
628 atomic_inc(&dev->carrier_up_count);
629 atomic_inc(&dev->carrier_down_count);
630 linkwatch_fire_event(dev);
631}
632EXPORT_SYMBOL_GPL(netif_carrier_event);
633
634/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
635 under all circumstances. It is difficult to invent anything faster or
636 cheaper.
637 */
638
639static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
640 struct sk_buff **to_free)
641{
642 dev_core_stats_tx_dropped_inc(skb->dev);
643 __qdisc_drop(skb, to_free);
644 return NET_XMIT_CN;
645}
646
647static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
648{
649 return NULL;
650}
651
652struct Qdisc_ops noop_qdisc_ops __read_mostly = {
653 .id = "noop",
654 .priv_size = 0,
655 .enqueue = noop_enqueue,
656 .dequeue = noop_dequeue,
657 .peek = noop_dequeue,
658 .owner = THIS_MODULE,
659};
660
661static struct netdev_queue noop_netdev_queue = {
662 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
663 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
664};
665
666struct Qdisc noop_qdisc = {
667 .enqueue = noop_enqueue,
668 .dequeue = noop_dequeue,
669 .flags = TCQ_F_BUILTIN,
670 .ops = &noop_qdisc_ops,
671 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
672 .dev_queue = &noop_netdev_queue,
673 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
674 .gso_skb = {
675 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
676 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
677 .qlen = 0,
678 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
679 },
680 .skb_bad_txq = {
681 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
682 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
683 .qlen = 0,
684 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
685 },
686 .owner = -1,
687};
688EXPORT_SYMBOL(noop_qdisc);
689
690static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
691 struct netlink_ext_ack *extack)
692{
693 /* register_qdisc() assigns a default of noop_enqueue if unset,
694 * but __dev_queue_xmit() treats noqueue only as such
695 * if this is NULL - so clear it here. */
696 qdisc->enqueue = NULL;
697 return 0;
698}
699
700struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
701 .id = "noqueue",
702 .priv_size = 0,
703 .init = noqueue_init,
704 .enqueue = noop_enqueue,
705 .dequeue = noop_dequeue,
706 .peek = noop_dequeue,
707 .owner = THIS_MODULE,
708};
709
710const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
711 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
712};
713EXPORT_SYMBOL(sch_default_prio2band);
714
715/* 3-band FIFO queue: old style, but should be a bit faster than
716 generic prio+fifo combination.
717 */
718
719#define PFIFO_FAST_BANDS 3
720
721/*
722 * Private data for a pfifo_fast scheduler containing:
723 * - rings for priority bands
724 */
725struct pfifo_fast_priv {
726 struct skb_array q[PFIFO_FAST_BANDS];
727};
728
729static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
730 int band)
731{
732 return &priv->q[band];
733}
734
735static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
736 struct sk_buff **to_free)
737{
738 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
739 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
740 struct skb_array *q = band2list(priv, band);
741 unsigned int pkt_len = qdisc_pkt_len(skb);
742 int err;
743
744 err = skb_array_produce(q, skb);
745
746 if (unlikely(err)) {
747 if (qdisc_is_percpu_stats(qdisc))
748 return qdisc_drop_cpu(skb, qdisc, to_free);
749 else
750 return qdisc_drop(skb, qdisc, to_free);
751 }
752
753 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
754 return NET_XMIT_SUCCESS;
755}
756
757static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
758{
759 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
760 struct sk_buff *skb = NULL;
761 bool need_retry = true;
762 int band;
763
764retry:
765 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
766 struct skb_array *q = band2list(priv, band);
767
768 if (__skb_array_empty(q))
769 continue;
770
771 skb = __skb_array_consume(q);
772 }
773 if (likely(skb)) {
774 qdisc_update_stats_at_dequeue(qdisc, skb);
775 } else if (need_retry &&
776 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
777 /* Delay clearing the STATE_MISSED here to reduce
778 * the overhead of the second spin_trylock() in
779 * qdisc_run_begin() and __netif_schedule() calling
780 * in qdisc_run_end().
781 */
782 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
783 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
784
785 /* Make sure dequeuing happens after clearing
786 * STATE_MISSED.
787 */
788 smp_mb__after_atomic();
789
790 need_retry = false;
791
792 goto retry;
793 }
794
795 return skb;
796}
797
798static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
799{
800 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
801 struct sk_buff *skb = NULL;
802 int band;
803
804 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
805 struct skb_array *q = band2list(priv, band);
806
807 skb = __skb_array_peek(q);
808 }
809
810 return skb;
811}
812
813static void pfifo_fast_reset(struct Qdisc *qdisc)
814{
815 int i, band;
816 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
817
818 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
819 struct skb_array *q = band2list(priv, band);
820 struct sk_buff *skb;
821
822 /* NULL ring is possible if destroy path is due to a failed
823 * skb_array_init() in pfifo_fast_init() case.
824 */
825 if (!q->ring.queue)
826 continue;
827
828 while ((skb = __skb_array_consume(q)) != NULL)
829 kfree_skb(skb);
830 }
831
832 if (qdisc_is_percpu_stats(qdisc)) {
833 for_each_possible_cpu(i) {
834 struct gnet_stats_queue *q;
835
836 q = per_cpu_ptr(qdisc->cpu_qstats, i);
837 q->backlog = 0;
838 q->qlen = 0;
839 }
840 }
841}
842
843static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
844{
845 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
846
847 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
848 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
849 goto nla_put_failure;
850 return skb->len;
851
852nla_put_failure:
853 return -1;
854}
855
856static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
857 struct netlink_ext_ack *extack)
858{
859 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
860 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
861 int prio;
862
863 /* guard against zero length rings */
864 if (!qlen)
865 return -EINVAL;
866
867 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
868 struct skb_array *q = band2list(priv, prio);
869 int err;
870
871 err = skb_array_init(q, qlen, GFP_KERNEL);
872 if (err)
873 return -ENOMEM;
874 }
875
876 /* Can by-pass the queue discipline */
877 qdisc->flags |= TCQ_F_CAN_BYPASS;
878 return 0;
879}
880
881static void pfifo_fast_destroy(struct Qdisc *sch)
882{
883 struct pfifo_fast_priv *priv = qdisc_priv(sch);
884 int prio;
885
886 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
887 struct skb_array *q = band2list(priv, prio);
888
889 /* NULL ring is possible if destroy path is due to a failed
890 * skb_array_init() in pfifo_fast_init() case.
891 */
892 if (!q->ring.queue)
893 continue;
894 /* Destroy ring but no need to kfree_skb because a call to
895 * pfifo_fast_reset() has already done that work.
896 */
897 ptr_ring_cleanup(&q->ring, NULL);
898 }
899}
900
901static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
902 unsigned int new_len)
903{
904 struct pfifo_fast_priv *priv = qdisc_priv(sch);
905 struct skb_array *bands[PFIFO_FAST_BANDS];
906 int prio;
907
908 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
909 struct skb_array *q = band2list(priv, prio);
910
911 bands[prio] = q;
912 }
913
914 return skb_array_resize_multiple_bh(bands, PFIFO_FAST_BANDS, new_len,
915 GFP_KERNEL);
916}
917
918struct Qdisc_ops pfifo_fast_ops __read_mostly = {
919 .id = "pfifo_fast",
920 .priv_size = sizeof(struct pfifo_fast_priv),
921 .enqueue = pfifo_fast_enqueue,
922 .dequeue = pfifo_fast_dequeue,
923 .peek = pfifo_fast_peek,
924 .init = pfifo_fast_init,
925 .destroy = pfifo_fast_destroy,
926 .reset = pfifo_fast_reset,
927 .dump = pfifo_fast_dump,
928 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
929 .owner = THIS_MODULE,
930 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
931};
932EXPORT_SYMBOL(pfifo_fast_ops);
933
934static struct lock_class_key qdisc_tx_busylock;
935
936struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
937 const struct Qdisc_ops *ops,
938 struct netlink_ext_ack *extack)
939{
940 struct Qdisc *sch;
941 unsigned int size = sizeof(*sch) + ops->priv_size;
942 int err = -ENOBUFS;
943 struct net_device *dev;
944
945 if (!dev_queue) {
946 NL_SET_ERR_MSG(extack, "No device queue given");
947 err = -EINVAL;
948 goto errout;
949 }
950
951 dev = dev_queue->dev;
952 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
953
954 if (!sch)
955 goto errout;
956 __skb_queue_head_init(&sch->gso_skb);
957 __skb_queue_head_init(&sch->skb_bad_txq);
958 gnet_stats_basic_sync_init(&sch->bstats);
959 lockdep_register_key(&sch->root_lock_key);
960 spin_lock_init(&sch->q.lock);
961 lockdep_set_class(&sch->q.lock, &sch->root_lock_key);
962
963 if (ops->static_flags & TCQ_F_CPUSTATS) {
964 sch->cpu_bstats =
965 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
966 if (!sch->cpu_bstats)
967 goto errout1;
968
969 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
970 if (!sch->cpu_qstats) {
971 free_percpu(sch->cpu_bstats);
972 goto errout1;
973 }
974 }
975
976 spin_lock_init(&sch->busylock);
977 lockdep_set_class(&sch->busylock,
978 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
979
980 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
981 spin_lock_init(&sch->seqlock);
982 lockdep_set_class(&sch->seqlock,
983 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
984
985 sch->ops = ops;
986 sch->flags = ops->static_flags;
987 sch->enqueue = ops->enqueue;
988 sch->dequeue = ops->dequeue;
989 sch->dev_queue = dev_queue;
990 sch->owner = -1;
991 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
992 refcount_set(&sch->refcnt, 1);
993
994 return sch;
995errout1:
996 lockdep_unregister_key(&sch->root_lock_key);
997 kfree(sch);
998errout:
999 return ERR_PTR(err);
1000}
1001
1002struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
1003 const struct Qdisc_ops *ops,
1004 unsigned int parentid,
1005 struct netlink_ext_ack *extack)
1006{
1007 struct Qdisc *sch;
1008
1009 if (!try_module_get(ops->owner)) {
1010 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1011 return NULL;
1012 }
1013
1014 sch = qdisc_alloc(dev_queue, ops, extack);
1015 if (IS_ERR(sch)) {
1016 module_put(ops->owner);
1017 return NULL;
1018 }
1019 sch->parent = parentid;
1020
1021 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1022 trace_qdisc_create(ops, dev_queue->dev, parentid);
1023 return sch;
1024 }
1025
1026 qdisc_put(sch);
1027 return NULL;
1028}
1029EXPORT_SYMBOL(qdisc_create_dflt);
1030
1031/* Under qdisc_lock(qdisc) and BH! */
1032
1033void qdisc_reset(struct Qdisc *qdisc)
1034{
1035 const struct Qdisc_ops *ops = qdisc->ops;
1036
1037 trace_qdisc_reset(qdisc);
1038
1039 if (ops->reset)
1040 ops->reset(qdisc);
1041
1042 __skb_queue_purge(&qdisc->gso_skb);
1043 __skb_queue_purge(&qdisc->skb_bad_txq);
1044
1045 qdisc->q.qlen = 0;
1046 qdisc->qstats.backlog = 0;
1047}
1048EXPORT_SYMBOL(qdisc_reset);
1049
1050void qdisc_free(struct Qdisc *qdisc)
1051{
1052 if (qdisc_is_percpu_stats(qdisc)) {
1053 free_percpu(qdisc->cpu_bstats);
1054 free_percpu(qdisc->cpu_qstats);
1055 }
1056
1057 kfree(qdisc);
1058}
1059
1060static void qdisc_free_cb(struct rcu_head *head)
1061{
1062 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1063
1064 qdisc_free(q);
1065}
1066
1067static void __qdisc_destroy(struct Qdisc *qdisc)
1068{
1069 const struct Qdisc_ops *ops = qdisc->ops;
1070 struct net_device *dev = qdisc_dev(qdisc);
1071
1072#ifdef CONFIG_NET_SCHED
1073 qdisc_hash_del(qdisc);
1074
1075 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1076#endif
1077 gen_kill_estimator(&qdisc->rate_est);
1078
1079 qdisc_reset(qdisc);
1080
1081
1082 if (ops->destroy)
1083 ops->destroy(qdisc);
1084
1085 lockdep_unregister_key(&qdisc->root_lock_key);
1086 module_put(ops->owner);
1087 netdev_put(dev, &qdisc->dev_tracker);
1088
1089 trace_qdisc_destroy(qdisc);
1090
1091 call_rcu(&qdisc->rcu, qdisc_free_cb);
1092}
1093
1094void qdisc_destroy(struct Qdisc *qdisc)
1095{
1096 if (qdisc->flags & TCQ_F_BUILTIN)
1097 return;
1098
1099 __qdisc_destroy(qdisc);
1100}
1101
1102void qdisc_put(struct Qdisc *qdisc)
1103{
1104 if (!qdisc)
1105 return;
1106
1107 if (qdisc->flags & TCQ_F_BUILTIN ||
1108 !refcount_dec_and_test(&qdisc->refcnt))
1109 return;
1110
1111 __qdisc_destroy(qdisc);
1112}
1113EXPORT_SYMBOL(qdisc_put);
1114
1115/* Version of qdisc_put() that is called with rtnl mutex unlocked.
1116 * Intended to be used as optimization, this function only takes rtnl lock if
1117 * qdisc reference counter reached zero.
1118 */
1119
1120void qdisc_put_unlocked(struct Qdisc *qdisc)
1121{
1122 if (qdisc->flags & TCQ_F_BUILTIN ||
1123 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1124 return;
1125
1126 __qdisc_destroy(qdisc);
1127 rtnl_unlock();
1128}
1129EXPORT_SYMBOL(qdisc_put_unlocked);
1130
1131/* Attach toplevel qdisc to device queue. */
1132struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1133 struct Qdisc *qdisc)
1134{
1135 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1136 spinlock_t *root_lock;
1137
1138 root_lock = qdisc_lock(oqdisc);
1139 spin_lock_bh(root_lock);
1140
1141 /* ... and graft new one */
1142 if (qdisc == NULL)
1143 qdisc = &noop_qdisc;
1144 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1145 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1146
1147 spin_unlock_bh(root_lock);
1148
1149 return oqdisc;
1150}
1151EXPORT_SYMBOL(dev_graft_qdisc);
1152
1153static void shutdown_scheduler_queue(struct net_device *dev,
1154 struct netdev_queue *dev_queue,
1155 void *_qdisc_default)
1156{
1157 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1158 struct Qdisc *qdisc_default = _qdisc_default;
1159
1160 if (qdisc) {
1161 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1162 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1163
1164 qdisc_put(qdisc);
1165 }
1166}
1167
1168static void attach_one_default_qdisc(struct net_device *dev,
1169 struct netdev_queue *dev_queue,
1170 void *_unused)
1171{
1172 struct Qdisc *qdisc;
1173 const struct Qdisc_ops *ops = default_qdisc_ops;
1174
1175 if (dev->priv_flags & IFF_NO_QUEUE)
1176 ops = &noqueue_qdisc_ops;
1177 else if(dev->type == ARPHRD_CAN)
1178 ops = &pfifo_fast_ops;
1179
1180 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1181 if (!qdisc)
1182 return;
1183
1184 if (!netif_is_multiqueue(dev))
1185 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1186 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1187}
1188
1189static void attach_default_qdiscs(struct net_device *dev)
1190{
1191 struct netdev_queue *txq;
1192 struct Qdisc *qdisc;
1193
1194 txq = netdev_get_tx_queue(dev, 0);
1195
1196 if (!netif_is_multiqueue(dev) ||
1197 dev->priv_flags & IFF_NO_QUEUE) {
1198 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1199 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1200 rcu_assign_pointer(dev->qdisc, qdisc);
1201 qdisc_refcount_inc(qdisc);
1202 } else {
1203 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1204 if (qdisc) {
1205 rcu_assign_pointer(dev->qdisc, qdisc);
1206 qdisc->ops->attach(qdisc);
1207 }
1208 }
1209 qdisc = rtnl_dereference(dev->qdisc);
1210
1211 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1212 if (qdisc == &noop_qdisc) {
1213 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1214 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1215 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1216 dev->priv_flags |= IFF_NO_QUEUE;
1217 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1218 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1219 rcu_assign_pointer(dev->qdisc, qdisc);
1220 qdisc_refcount_inc(qdisc);
1221 dev->priv_flags ^= IFF_NO_QUEUE;
1222 }
1223
1224#ifdef CONFIG_NET_SCHED
1225 if (qdisc != &noop_qdisc)
1226 qdisc_hash_add(qdisc, false);
1227#endif
1228}
1229
1230static void transition_one_qdisc(struct net_device *dev,
1231 struct netdev_queue *dev_queue,
1232 void *_need_watchdog)
1233{
1234 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1235 int *need_watchdog_p = _need_watchdog;
1236
1237 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1238 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1239
1240 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1241 if (need_watchdog_p) {
1242 WRITE_ONCE(dev_queue->trans_start, 0);
1243 *need_watchdog_p = 1;
1244 }
1245}
1246
1247void dev_activate(struct net_device *dev)
1248{
1249 int need_watchdog;
1250
1251 /* No queueing discipline is attached to device;
1252 * create default one for devices, which need queueing
1253 * and noqueue_qdisc for virtual interfaces
1254 */
1255
1256 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1257 attach_default_qdiscs(dev);
1258
1259 if (!netif_carrier_ok(dev))
1260 /* Delay activation until next carrier-on event */
1261 return;
1262
1263 need_watchdog = 0;
1264 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1265 if (dev_ingress_queue(dev))
1266 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1267
1268 if (need_watchdog) {
1269 netif_trans_update(dev);
1270 dev_watchdog_up(dev);
1271 }
1272}
1273EXPORT_SYMBOL(dev_activate);
1274
1275static void qdisc_deactivate(struct Qdisc *qdisc)
1276{
1277 if (qdisc->flags & TCQ_F_BUILTIN)
1278 return;
1279
1280 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1281}
1282
1283static void dev_deactivate_queue(struct net_device *dev,
1284 struct netdev_queue *dev_queue,
1285 void *_qdisc_default)
1286{
1287 struct Qdisc *qdisc_default = _qdisc_default;
1288 struct Qdisc *qdisc;
1289
1290 qdisc = rtnl_dereference(dev_queue->qdisc);
1291 if (qdisc) {
1292 qdisc_deactivate(qdisc);
1293 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1294 }
1295}
1296
1297static void dev_reset_queue(struct net_device *dev,
1298 struct netdev_queue *dev_queue,
1299 void *_unused)
1300{
1301 struct Qdisc *qdisc;
1302 bool nolock;
1303
1304 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1305 if (!qdisc)
1306 return;
1307
1308 nolock = qdisc->flags & TCQ_F_NOLOCK;
1309
1310 if (nolock)
1311 spin_lock_bh(&qdisc->seqlock);
1312 spin_lock_bh(qdisc_lock(qdisc));
1313
1314 qdisc_reset(qdisc);
1315
1316 spin_unlock_bh(qdisc_lock(qdisc));
1317 if (nolock) {
1318 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1319 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1320 spin_unlock_bh(&qdisc->seqlock);
1321 }
1322}
1323
1324static bool some_qdisc_is_busy(struct net_device *dev)
1325{
1326 unsigned int i;
1327
1328 for (i = 0; i < dev->num_tx_queues; i++) {
1329 struct netdev_queue *dev_queue;
1330 spinlock_t *root_lock;
1331 struct Qdisc *q;
1332 int val;
1333
1334 dev_queue = netdev_get_tx_queue(dev, i);
1335 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1336
1337 root_lock = qdisc_lock(q);
1338 spin_lock_bh(root_lock);
1339
1340 val = (qdisc_is_running(q) ||
1341 test_bit(__QDISC_STATE_SCHED, &q->state));
1342
1343 spin_unlock_bh(root_lock);
1344
1345 if (val)
1346 return true;
1347 }
1348 return false;
1349}
1350
1351/**
1352 * dev_deactivate_many - deactivate transmissions on several devices
1353 * @head: list of devices to deactivate
1354 *
1355 * This function returns only when all outstanding transmissions
1356 * have completed, unless all devices are in dismantle phase.
1357 */
1358void dev_deactivate_many(struct list_head *head)
1359{
1360 struct net_device *dev;
1361
1362 list_for_each_entry(dev, head, close_list) {
1363 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1364 &noop_qdisc);
1365 if (dev_ingress_queue(dev))
1366 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1367 &noop_qdisc);
1368
1369 dev_watchdog_down(dev);
1370 }
1371
1372 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1373 * outstanding qdisc enqueuing calls.
1374 * This is avoided if all devices are in dismantle phase :
1375 * Caller will call synchronize_net() for us
1376 */
1377 synchronize_net();
1378
1379 list_for_each_entry(dev, head, close_list) {
1380 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1381
1382 if (dev_ingress_queue(dev))
1383 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1384 }
1385
1386 /* Wait for outstanding qdisc_run calls. */
1387 list_for_each_entry(dev, head, close_list) {
1388 while (some_qdisc_is_busy(dev)) {
1389 /* wait_event() would avoid this sleep-loop but would
1390 * require expensive checks in the fast paths of packet
1391 * processing which isn't worth it.
1392 */
1393 schedule_timeout_uninterruptible(1);
1394 }
1395 }
1396}
1397
1398void dev_deactivate(struct net_device *dev)
1399{
1400 LIST_HEAD(single);
1401
1402 list_add(&dev->close_list, &single);
1403 dev_deactivate_many(&single);
1404 list_del(&single);
1405}
1406EXPORT_SYMBOL(dev_deactivate);
1407
1408static int qdisc_change_tx_queue_len(struct net_device *dev,
1409 struct netdev_queue *dev_queue)
1410{
1411 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1412 const struct Qdisc_ops *ops = qdisc->ops;
1413
1414 if (ops->change_tx_queue_len)
1415 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1416 return 0;
1417}
1418
1419void dev_qdisc_change_real_num_tx(struct net_device *dev,
1420 unsigned int new_real_tx)
1421{
1422 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1423
1424 if (qdisc->ops->change_real_num_tx)
1425 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1426}
1427
1428void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1429{
1430#ifdef CONFIG_NET_SCHED
1431 struct net_device *dev = qdisc_dev(sch);
1432 struct Qdisc *qdisc;
1433 unsigned int i;
1434
1435 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1436 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1437 /* Only update the default qdiscs we created,
1438 * qdiscs with handles are always hashed.
1439 */
1440 if (qdisc != &noop_qdisc && !qdisc->handle)
1441 qdisc_hash_del(qdisc);
1442 }
1443 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1444 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1445 if (qdisc != &noop_qdisc && !qdisc->handle)
1446 qdisc_hash_add(qdisc, false);
1447 }
1448#endif
1449}
1450EXPORT_SYMBOL(mq_change_real_num_tx);
1451
1452int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1453{
1454 bool up = dev->flags & IFF_UP;
1455 unsigned int i;
1456 int ret = 0;
1457
1458 if (up)
1459 dev_deactivate(dev);
1460
1461 for (i = 0; i < dev->num_tx_queues; i++) {
1462 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1463
1464 /* TODO: revert changes on a partial failure */
1465 if (ret)
1466 break;
1467 }
1468
1469 if (up)
1470 dev_activate(dev);
1471 return ret;
1472}
1473
1474static void dev_init_scheduler_queue(struct net_device *dev,
1475 struct netdev_queue *dev_queue,
1476 void *_qdisc)
1477{
1478 struct Qdisc *qdisc = _qdisc;
1479
1480 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1481 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1482}
1483
1484void dev_init_scheduler(struct net_device *dev)
1485{
1486 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1487 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1488 if (dev_ingress_queue(dev))
1489 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1490
1491 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1492}
1493
1494void dev_shutdown(struct net_device *dev)
1495{
1496 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1497 if (dev_ingress_queue(dev))
1498 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1499 qdisc_put(rtnl_dereference(dev->qdisc));
1500 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1501
1502 WARN_ON(timer_pending(&dev->watchdog_timer));
1503}
1504
1505/**
1506 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1507 * @rate: Rate to compute reciprocal division values of
1508 * @mult: Multiplier for reciprocal division
1509 * @shift: Shift for reciprocal division
1510 *
1511 * The multiplier and shift for reciprocal division by rate are stored
1512 * in mult and shift.
1513 *
1514 * The deal here is to replace a divide by a reciprocal one
1515 * in fast path (a reciprocal divide is a multiply and a shift)
1516 *
1517 * Normal formula would be :
1518 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1519 *
1520 * We compute mult/shift to use instead :
1521 * time_in_ns = (len * mult) >> shift;
1522 *
1523 * We try to get the highest possible mult value for accuracy,
1524 * but have to make sure no overflows will ever happen.
1525 *
1526 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1527 */
1528static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1529{
1530 u64 factor = NSEC_PER_SEC;
1531
1532 *mult = 1;
1533 *shift = 0;
1534
1535 if (rate <= 0)
1536 return;
1537
1538 for (;;) {
1539 *mult = div64_u64(factor, rate);
1540 if (*mult & (1U << 31) || factor & (1ULL << 63))
1541 break;
1542 factor <<= 1;
1543 (*shift)++;
1544 }
1545}
1546
1547void psched_ratecfg_precompute(struct psched_ratecfg *r,
1548 const struct tc_ratespec *conf,
1549 u64 rate64)
1550{
1551 memset(r, 0, sizeof(*r));
1552 r->overhead = conf->overhead;
1553 r->mpu = conf->mpu;
1554 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1555 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1556 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1557}
1558EXPORT_SYMBOL(psched_ratecfg_precompute);
1559
1560void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1561{
1562 r->rate_pkts_ps = pktrate64;
1563 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1564}
1565EXPORT_SYMBOL(psched_ppscfg_precompute);
1566
1567void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1568 struct tcf_proto *tp_head)
1569{
1570 /* Protected with chain0->filter_chain_lock.
1571 * Can't access chain directly because tp_head can be NULL.
1572 */
1573 struct mini_Qdisc *miniq_old =
1574 rcu_dereference_protected(*miniqp->p_miniq, 1);
1575 struct mini_Qdisc *miniq;
1576
1577 if (!tp_head) {
1578 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1579 } else {
1580 miniq = miniq_old != &miniqp->miniq1 ?
1581 &miniqp->miniq1 : &miniqp->miniq2;
1582
1583 /* We need to make sure that readers won't see the miniq
1584 * we are about to modify. So ensure that at least one RCU
1585 * grace period has elapsed since the miniq was made
1586 * inactive.
1587 */
1588 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1589 cond_synchronize_rcu(miniq->rcu_state);
1590 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1591 synchronize_rcu_expedited();
1592
1593 miniq->filter_list = tp_head;
1594 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1595 }
1596
1597 if (miniq_old)
1598 /* This is counterpart of the rcu sync above. We need to
1599 * block potential new user of miniq_old until all readers
1600 * are not seeing it.
1601 */
1602 miniq_old->rcu_state = start_poll_synchronize_rcu();
1603}
1604EXPORT_SYMBOL(mini_qdisc_pair_swap);
1605
1606void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1607 struct tcf_block *block)
1608{
1609 miniqp->miniq1.block = block;
1610 miniqp->miniq2.block = block;
1611}
1612EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1613
1614void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1615 struct mini_Qdisc __rcu **p_miniq)
1616{
1617 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1618 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1619 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1620 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1621 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1622 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1623 miniqp->p_miniq = p_miniq;
1624}
1625EXPORT_SYMBOL(mini_qdisc_pair_init);