Loading...
1/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/sched.h>
19#include <linux/string.h>
20#include <linux/errno.h>
21#include <linux/netdevice.h>
22#include <linux/skbuff.h>
23#include <linux/rtnetlink.h>
24#include <linux/init.h>
25#include <linux/rcupdate.h>
26#include <linux/list.h>
27#include <linux/slab.h>
28#include <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 skb_dst_force(skb);
51 q->gso_skb = skb;
52 q->qstats.requeues++;
53 q->q.qlen++; /* it's still part of the queue */
54 __netif_schedule(q);
55
56 return 0;
57}
58
59static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
60{
61 struct sk_buff *skb = q->gso_skb;
62 const struct netdev_queue *txq = q->dev_queue;
63
64 if (unlikely(skb)) {
65 /* check the reason of requeuing without tx lock first */
66 txq = netdev_get_tx_queue(txq->dev, skb_get_queue_mapping(skb));
67 if (!netif_xmit_frozen_or_stopped(txq)) {
68 q->gso_skb = NULL;
69 q->q.qlen--;
70 } else
71 skb = NULL;
72 } else {
73 if (!(q->flags & TCQ_F_ONETXQUEUE) || !netif_xmit_frozen_or_stopped(txq))
74 skb = q->dequeue(q);
75 }
76
77 return skb;
78}
79
80static inline int handle_dev_cpu_collision(struct sk_buff *skb,
81 struct netdev_queue *dev_queue,
82 struct Qdisc *q)
83{
84 int ret;
85
86 if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
87 /*
88 * Same CPU holding the lock. It may be a transient
89 * configuration error, when hard_start_xmit() recurses. We
90 * detect it by checking xmit owner and drop the packet when
91 * deadloop is detected. Return OK to try the next skb.
92 */
93 kfree_skb(skb);
94 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
95 dev_queue->dev->name);
96 ret = qdisc_qlen(q);
97 } else {
98 /*
99 * Another cpu is holding lock, requeue & delay xmits for
100 * some time.
101 */
102 __this_cpu_inc(softnet_data.cpu_collision);
103 ret = dev_requeue_skb(skb, q);
104 }
105
106 return ret;
107}
108
109/*
110 * Transmit one skb, and handle the return status as required. Holding the
111 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
112 * function.
113 *
114 * Returns to the caller:
115 * 0 - queue is empty or throttled.
116 * >0 - queue is not empty.
117 */
118int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
119 struct net_device *dev, struct netdev_queue *txq,
120 spinlock_t *root_lock)
121{
122 int ret = NETDEV_TX_BUSY;
123
124 /* And release qdisc */
125 spin_unlock(root_lock);
126
127 HARD_TX_LOCK(dev, txq, smp_processor_id());
128 if (!netif_xmit_frozen_or_stopped(txq))
129 ret = dev_hard_start_xmit(skb, dev, txq);
130
131 HARD_TX_UNLOCK(dev, txq);
132
133 spin_lock(root_lock);
134
135 if (dev_xmit_complete(ret)) {
136 /* Driver sent out skb successfully or skb was consumed */
137 ret = qdisc_qlen(q);
138 } else if (ret == NETDEV_TX_LOCKED) {
139 /* Driver try lock failed */
140 ret = handle_dev_cpu_collision(skb, txq, q);
141 } else {
142 /* Driver returned NETDEV_TX_BUSY - requeue skb */
143 if (unlikely(ret != NETDEV_TX_BUSY))
144 net_warn_ratelimited("BUG %s code %d qlen %d\n",
145 dev->name, ret, q->q.qlen);
146
147 ret = dev_requeue_skb(skb, q);
148 }
149
150 if (ret && netif_xmit_frozen_or_stopped(txq))
151 ret = 0;
152
153 return ret;
154}
155
156/*
157 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
158 *
159 * __QDISC_STATE_RUNNING guarantees only one CPU can process
160 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
161 * this queue.
162 *
163 * netif_tx_lock serializes accesses to device driver.
164 *
165 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
166 * if one is grabbed, another must be free.
167 *
168 * Note, that this procedure can be called by a watchdog timer
169 *
170 * Returns to the caller:
171 * 0 - queue is empty or throttled.
172 * >0 - queue is not empty.
173 *
174 */
175static inline int qdisc_restart(struct Qdisc *q)
176{
177 struct netdev_queue *txq;
178 struct net_device *dev;
179 spinlock_t *root_lock;
180 struct sk_buff *skb;
181
182 /* Dequeue packet */
183 skb = dequeue_skb(q);
184 if (unlikely(!skb))
185 return 0;
186 WARN_ON_ONCE(skb_dst_is_noref(skb));
187 root_lock = qdisc_lock(q);
188 dev = qdisc_dev(q);
189 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
190
191 return sch_direct_xmit(skb, q, dev, txq, root_lock);
192}
193
194void __qdisc_run(struct Qdisc *q)
195{
196 int quota = weight_p;
197
198 while (qdisc_restart(q)) {
199 /*
200 * Ordered by possible occurrence: Postpone processing if
201 * 1. we've exceeded packet quota
202 * 2. another process needs the CPU;
203 */
204 if (--quota <= 0 || need_resched()) {
205 __netif_schedule(q);
206 break;
207 }
208 }
209
210 qdisc_run_end(q);
211}
212
213unsigned long dev_trans_start(struct net_device *dev)
214{
215 unsigned long val, res;
216 unsigned int i;
217
218 if (is_vlan_dev(dev))
219 dev = vlan_dev_real_dev(dev);
220 res = dev->trans_start;
221 for (i = 0; i < dev->num_tx_queues; i++) {
222 val = netdev_get_tx_queue(dev, i)->trans_start;
223 if (val && time_after(val, res))
224 res = val;
225 }
226 dev->trans_start = res;
227
228 return res;
229}
230EXPORT_SYMBOL(dev_trans_start);
231
232static void dev_watchdog(unsigned long arg)
233{
234 struct net_device *dev = (struct net_device *)arg;
235
236 netif_tx_lock(dev);
237 if (!qdisc_tx_is_noop(dev)) {
238 if (netif_device_present(dev) &&
239 netif_running(dev) &&
240 netif_carrier_ok(dev)) {
241 int some_queue_timedout = 0;
242 unsigned int i;
243 unsigned long trans_start;
244
245 for (i = 0; i < dev->num_tx_queues; i++) {
246 struct netdev_queue *txq;
247
248 txq = netdev_get_tx_queue(dev, i);
249 /*
250 * old device drivers set dev->trans_start
251 */
252 trans_start = txq->trans_start ? : dev->trans_start;
253 if (netif_xmit_stopped(txq) &&
254 time_after(jiffies, (trans_start +
255 dev->watchdog_timeo))) {
256 some_queue_timedout = 1;
257 txq->trans_timeout++;
258 break;
259 }
260 }
261
262 if (some_queue_timedout) {
263 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
264 dev->name, netdev_drivername(dev), i);
265 dev->netdev_ops->ndo_tx_timeout(dev);
266 }
267 if (!mod_timer(&dev->watchdog_timer,
268 round_jiffies(jiffies +
269 dev->watchdog_timeo)))
270 dev_hold(dev);
271 }
272 }
273 netif_tx_unlock(dev);
274
275 dev_put(dev);
276}
277
278void __netdev_watchdog_up(struct net_device *dev)
279{
280 if (dev->netdev_ops->ndo_tx_timeout) {
281 if (dev->watchdog_timeo <= 0)
282 dev->watchdog_timeo = 5*HZ;
283 if (!mod_timer(&dev->watchdog_timer,
284 round_jiffies(jiffies + dev->watchdog_timeo)))
285 dev_hold(dev);
286 }
287}
288
289static void dev_watchdog_up(struct net_device *dev)
290{
291 __netdev_watchdog_up(dev);
292}
293
294static void dev_watchdog_down(struct net_device *dev)
295{
296 netif_tx_lock_bh(dev);
297 if (del_timer(&dev->watchdog_timer))
298 dev_put(dev);
299 netif_tx_unlock_bh(dev);
300}
301
302/**
303 * netif_carrier_on - set carrier
304 * @dev: network device
305 *
306 * Device has detected that carrier.
307 */
308void netif_carrier_on(struct net_device *dev)
309{
310 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
311 if (dev->reg_state == NETREG_UNINITIALIZED)
312 return;
313 atomic_inc(&dev->carrier_changes);
314 linkwatch_fire_event(dev);
315 if (netif_running(dev))
316 __netdev_watchdog_up(dev);
317 }
318}
319EXPORT_SYMBOL(netif_carrier_on);
320
321/**
322 * netif_carrier_off - clear carrier
323 * @dev: network device
324 *
325 * Device has detected loss of carrier.
326 */
327void netif_carrier_off(struct net_device *dev)
328{
329 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
330 if (dev->reg_state == NETREG_UNINITIALIZED)
331 return;
332 atomic_inc(&dev->carrier_changes);
333 linkwatch_fire_event(dev);
334 }
335}
336EXPORT_SYMBOL(netif_carrier_off);
337
338/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
339 under all circumstances. It is difficult to invent anything faster or
340 cheaper.
341 */
342
343static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
344{
345 kfree_skb(skb);
346 return NET_XMIT_CN;
347}
348
349static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
350{
351 return NULL;
352}
353
354struct Qdisc_ops noop_qdisc_ops __read_mostly = {
355 .id = "noop",
356 .priv_size = 0,
357 .enqueue = noop_enqueue,
358 .dequeue = noop_dequeue,
359 .peek = noop_dequeue,
360 .owner = THIS_MODULE,
361};
362
363static struct netdev_queue noop_netdev_queue = {
364 .qdisc = &noop_qdisc,
365 .qdisc_sleeping = &noop_qdisc,
366};
367
368struct Qdisc noop_qdisc = {
369 .enqueue = noop_enqueue,
370 .dequeue = noop_dequeue,
371 .flags = TCQ_F_BUILTIN,
372 .ops = &noop_qdisc_ops,
373 .list = LIST_HEAD_INIT(noop_qdisc.list),
374 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
375 .dev_queue = &noop_netdev_queue,
376 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
377};
378EXPORT_SYMBOL(noop_qdisc);
379
380static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
381 .id = "noqueue",
382 .priv_size = 0,
383 .enqueue = noop_enqueue,
384 .dequeue = noop_dequeue,
385 .peek = noop_dequeue,
386 .owner = THIS_MODULE,
387};
388
389static struct Qdisc noqueue_qdisc;
390static struct netdev_queue noqueue_netdev_queue = {
391 .qdisc = &noqueue_qdisc,
392 .qdisc_sleeping = &noqueue_qdisc,
393};
394
395static struct Qdisc noqueue_qdisc = {
396 .enqueue = NULL,
397 .dequeue = noop_dequeue,
398 .flags = TCQ_F_BUILTIN,
399 .ops = &noqueue_qdisc_ops,
400 .list = LIST_HEAD_INIT(noqueue_qdisc.list),
401 .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
402 .dev_queue = &noqueue_netdev_queue,
403 .busylock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
404};
405
406
407static const u8 prio2band[TC_PRIO_MAX + 1] = {
408 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
409};
410
411/* 3-band FIFO queue: old style, but should be a bit faster than
412 generic prio+fifo combination.
413 */
414
415#define PFIFO_FAST_BANDS 3
416
417/*
418 * Private data for a pfifo_fast scheduler containing:
419 * - queues for the three band
420 * - bitmap indicating which of the bands contain skbs
421 */
422struct pfifo_fast_priv {
423 u32 bitmap;
424 struct sk_buff_head q[PFIFO_FAST_BANDS];
425};
426
427/*
428 * Convert a bitmap to the first band number where an skb is queued, where:
429 * bitmap=0 means there are no skbs on any band.
430 * bitmap=1 means there is an skb on band 0.
431 * bitmap=7 means there are skbs on all 3 bands, etc.
432 */
433static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
434
435static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
436 int band)
437{
438 return priv->q + band;
439}
440
441static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
442{
443 if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
444 int band = prio2band[skb->priority & TC_PRIO_MAX];
445 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
446 struct sk_buff_head *list = band2list(priv, band);
447
448 priv->bitmap |= (1 << band);
449 qdisc->q.qlen++;
450 return __qdisc_enqueue_tail(skb, qdisc, list);
451 }
452
453 return qdisc_drop(skb, qdisc);
454}
455
456static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
457{
458 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
459 int band = bitmap2band[priv->bitmap];
460
461 if (likely(band >= 0)) {
462 struct sk_buff_head *list = band2list(priv, band);
463 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
464
465 qdisc->q.qlen--;
466 if (skb_queue_empty(list))
467 priv->bitmap &= ~(1 << band);
468
469 return skb;
470 }
471
472 return NULL;
473}
474
475static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
476{
477 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
478 int band = bitmap2band[priv->bitmap];
479
480 if (band >= 0) {
481 struct sk_buff_head *list = band2list(priv, band);
482
483 return skb_peek(list);
484 }
485
486 return NULL;
487}
488
489static void pfifo_fast_reset(struct Qdisc *qdisc)
490{
491 int prio;
492 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
493
494 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
495 __qdisc_reset_queue(qdisc, band2list(priv, prio));
496
497 priv->bitmap = 0;
498 qdisc->qstats.backlog = 0;
499 qdisc->q.qlen = 0;
500}
501
502static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
503{
504 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
505
506 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
507 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
508 goto nla_put_failure;
509 return skb->len;
510
511nla_put_failure:
512 return -1;
513}
514
515static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
516{
517 int prio;
518 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
519
520 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
521 skb_queue_head_init(band2list(priv, prio));
522
523 /* Can by-pass the queue discipline */
524 qdisc->flags |= TCQ_F_CAN_BYPASS;
525 return 0;
526}
527
528struct Qdisc_ops pfifo_fast_ops __read_mostly = {
529 .id = "pfifo_fast",
530 .priv_size = sizeof(struct pfifo_fast_priv),
531 .enqueue = pfifo_fast_enqueue,
532 .dequeue = pfifo_fast_dequeue,
533 .peek = pfifo_fast_peek,
534 .init = pfifo_fast_init,
535 .reset = pfifo_fast_reset,
536 .dump = pfifo_fast_dump,
537 .owner = THIS_MODULE,
538};
539
540static struct lock_class_key qdisc_tx_busylock;
541
542struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
543 const struct Qdisc_ops *ops)
544{
545 void *p;
546 struct Qdisc *sch;
547 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
548 int err = -ENOBUFS;
549 struct net_device *dev = dev_queue->dev;
550
551 p = kzalloc_node(size, GFP_KERNEL,
552 netdev_queue_numa_node_read(dev_queue));
553
554 if (!p)
555 goto errout;
556 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
557 /* if we got non aligned memory, ask more and do alignment ourself */
558 if (sch != p) {
559 kfree(p);
560 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
561 netdev_queue_numa_node_read(dev_queue));
562 if (!p)
563 goto errout;
564 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
565 sch->padded = (char *) sch - (char *) p;
566 }
567 INIT_LIST_HEAD(&sch->list);
568 skb_queue_head_init(&sch->q);
569
570 spin_lock_init(&sch->busylock);
571 lockdep_set_class(&sch->busylock,
572 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
573
574 sch->ops = ops;
575 sch->enqueue = ops->enqueue;
576 sch->dequeue = ops->dequeue;
577 sch->dev_queue = dev_queue;
578 dev_hold(dev);
579 atomic_set(&sch->refcnt, 1);
580
581 return sch;
582errout:
583 return ERR_PTR(err);
584}
585
586struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
587 const struct Qdisc_ops *ops,
588 unsigned int parentid)
589{
590 struct Qdisc *sch;
591
592 if (!try_module_get(ops->owner))
593 goto errout;
594
595 sch = qdisc_alloc(dev_queue, ops);
596 if (IS_ERR(sch))
597 goto errout;
598 sch->parent = parentid;
599
600 if (!ops->init || ops->init(sch, NULL) == 0)
601 return sch;
602
603 qdisc_destroy(sch);
604errout:
605 return NULL;
606}
607EXPORT_SYMBOL(qdisc_create_dflt);
608
609/* Under qdisc_lock(qdisc) and BH! */
610
611void qdisc_reset(struct Qdisc *qdisc)
612{
613 const struct Qdisc_ops *ops = qdisc->ops;
614
615 if (ops->reset)
616 ops->reset(qdisc);
617
618 if (qdisc->gso_skb) {
619 kfree_skb(qdisc->gso_skb);
620 qdisc->gso_skb = NULL;
621 qdisc->q.qlen = 0;
622 }
623}
624EXPORT_SYMBOL(qdisc_reset);
625
626static void qdisc_rcu_free(struct rcu_head *head)
627{
628 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
629
630 kfree((char *) qdisc - qdisc->padded);
631}
632
633void qdisc_destroy(struct Qdisc *qdisc)
634{
635 const struct Qdisc_ops *ops = qdisc->ops;
636
637 if (qdisc->flags & TCQ_F_BUILTIN ||
638 !atomic_dec_and_test(&qdisc->refcnt))
639 return;
640
641#ifdef CONFIG_NET_SCHED
642 qdisc_list_del(qdisc);
643
644 qdisc_put_stab(rtnl_dereference(qdisc->stab));
645#endif
646 gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
647 if (ops->reset)
648 ops->reset(qdisc);
649 if (ops->destroy)
650 ops->destroy(qdisc);
651
652 module_put(ops->owner);
653 dev_put(qdisc_dev(qdisc));
654
655 kfree_skb(qdisc->gso_skb);
656 /*
657 * gen_estimator est_timer() might access qdisc->q.lock,
658 * wait a RCU grace period before freeing qdisc.
659 */
660 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
661}
662EXPORT_SYMBOL(qdisc_destroy);
663
664/* Attach toplevel qdisc to device queue. */
665struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
666 struct Qdisc *qdisc)
667{
668 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
669 spinlock_t *root_lock;
670
671 root_lock = qdisc_lock(oqdisc);
672 spin_lock_bh(root_lock);
673
674 /* Prune old scheduler */
675 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
676 qdisc_reset(oqdisc);
677
678 /* ... and graft new one */
679 if (qdisc == NULL)
680 qdisc = &noop_qdisc;
681 dev_queue->qdisc_sleeping = qdisc;
682 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
683
684 spin_unlock_bh(root_lock);
685
686 return oqdisc;
687}
688EXPORT_SYMBOL(dev_graft_qdisc);
689
690static void attach_one_default_qdisc(struct net_device *dev,
691 struct netdev_queue *dev_queue,
692 void *_unused)
693{
694 struct Qdisc *qdisc = &noqueue_qdisc;
695
696 if (dev->tx_queue_len) {
697 qdisc = qdisc_create_dflt(dev_queue,
698 default_qdisc_ops, TC_H_ROOT);
699 if (!qdisc) {
700 netdev_info(dev, "activation failed\n");
701 return;
702 }
703 if (!netif_is_multiqueue(dev))
704 qdisc->flags |= TCQ_F_ONETXQUEUE;
705 }
706 dev_queue->qdisc_sleeping = qdisc;
707}
708
709static void attach_default_qdiscs(struct net_device *dev)
710{
711 struct netdev_queue *txq;
712 struct Qdisc *qdisc;
713
714 txq = netdev_get_tx_queue(dev, 0);
715
716 if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
717 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
718 dev->qdisc = txq->qdisc_sleeping;
719 atomic_inc(&dev->qdisc->refcnt);
720 } else {
721 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
722 if (qdisc) {
723 dev->qdisc = qdisc;
724 qdisc->ops->attach(qdisc);
725 }
726 }
727}
728
729static void transition_one_qdisc(struct net_device *dev,
730 struct netdev_queue *dev_queue,
731 void *_need_watchdog)
732{
733 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
734 int *need_watchdog_p = _need_watchdog;
735
736 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
737 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
738
739 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
740 if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
741 dev_queue->trans_start = 0;
742 *need_watchdog_p = 1;
743 }
744}
745
746void dev_activate(struct net_device *dev)
747{
748 int need_watchdog;
749
750 /* No queueing discipline is attached to device;
751 * create default one for devices, which need queueing
752 * and noqueue_qdisc for virtual interfaces
753 */
754
755 if (dev->qdisc == &noop_qdisc)
756 attach_default_qdiscs(dev);
757
758 if (!netif_carrier_ok(dev))
759 /* Delay activation until next carrier-on event */
760 return;
761
762 need_watchdog = 0;
763 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
764 if (dev_ingress_queue(dev))
765 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
766
767 if (need_watchdog) {
768 dev->trans_start = jiffies;
769 dev_watchdog_up(dev);
770 }
771}
772EXPORT_SYMBOL(dev_activate);
773
774static void dev_deactivate_queue(struct net_device *dev,
775 struct netdev_queue *dev_queue,
776 void *_qdisc_default)
777{
778 struct Qdisc *qdisc_default = _qdisc_default;
779 struct Qdisc *qdisc;
780
781 qdisc = dev_queue->qdisc;
782 if (qdisc) {
783 spin_lock_bh(qdisc_lock(qdisc));
784
785 if (!(qdisc->flags & TCQ_F_BUILTIN))
786 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
787
788 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
789 qdisc_reset(qdisc);
790
791 spin_unlock_bh(qdisc_lock(qdisc));
792 }
793}
794
795static bool some_qdisc_is_busy(struct net_device *dev)
796{
797 unsigned int i;
798
799 for (i = 0; i < dev->num_tx_queues; i++) {
800 struct netdev_queue *dev_queue;
801 spinlock_t *root_lock;
802 struct Qdisc *q;
803 int val;
804
805 dev_queue = netdev_get_tx_queue(dev, i);
806 q = dev_queue->qdisc_sleeping;
807 root_lock = qdisc_lock(q);
808
809 spin_lock_bh(root_lock);
810
811 val = (qdisc_is_running(q) ||
812 test_bit(__QDISC_STATE_SCHED, &q->state));
813
814 spin_unlock_bh(root_lock);
815
816 if (val)
817 return true;
818 }
819 return false;
820}
821
822/**
823 * dev_deactivate_many - deactivate transmissions on several devices
824 * @head: list of devices to deactivate
825 *
826 * This function returns only when all outstanding transmissions
827 * have completed, unless all devices are in dismantle phase.
828 */
829void dev_deactivate_many(struct list_head *head)
830{
831 struct net_device *dev;
832 bool sync_needed = false;
833
834 list_for_each_entry(dev, head, close_list) {
835 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
836 &noop_qdisc);
837 if (dev_ingress_queue(dev))
838 dev_deactivate_queue(dev, dev_ingress_queue(dev),
839 &noop_qdisc);
840
841 dev_watchdog_down(dev);
842 sync_needed |= !dev->dismantle;
843 }
844
845 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
846 * This is avoided if all devices are in dismantle phase :
847 * Caller will call synchronize_net() for us
848 */
849 if (sync_needed)
850 synchronize_net();
851
852 /* Wait for outstanding qdisc_run calls. */
853 list_for_each_entry(dev, head, close_list)
854 while (some_qdisc_is_busy(dev))
855 yield();
856}
857
858void dev_deactivate(struct net_device *dev)
859{
860 LIST_HEAD(single);
861
862 list_add(&dev->close_list, &single);
863 dev_deactivate_many(&single);
864 list_del(&single);
865}
866EXPORT_SYMBOL(dev_deactivate);
867
868static void dev_init_scheduler_queue(struct net_device *dev,
869 struct netdev_queue *dev_queue,
870 void *_qdisc)
871{
872 struct Qdisc *qdisc = _qdisc;
873
874 dev_queue->qdisc = qdisc;
875 dev_queue->qdisc_sleeping = qdisc;
876}
877
878void dev_init_scheduler(struct net_device *dev)
879{
880 dev->qdisc = &noop_qdisc;
881 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
882 if (dev_ingress_queue(dev))
883 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
884
885 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
886}
887
888static void shutdown_scheduler_queue(struct net_device *dev,
889 struct netdev_queue *dev_queue,
890 void *_qdisc_default)
891{
892 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
893 struct Qdisc *qdisc_default = _qdisc_default;
894
895 if (qdisc) {
896 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
897 dev_queue->qdisc_sleeping = qdisc_default;
898
899 qdisc_destroy(qdisc);
900 }
901}
902
903void dev_shutdown(struct net_device *dev)
904{
905 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
906 if (dev_ingress_queue(dev))
907 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
908 qdisc_destroy(dev->qdisc);
909 dev->qdisc = &noop_qdisc;
910
911 WARN_ON(timer_pending(&dev->watchdog_timer));
912}
913
914void psched_ratecfg_precompute(struct psched_ratecfg *r,
915 const struct tc_ratespec *conf,
916 u64 rate64)
917{
918 memset(r, 0, sizeof(*r));
919 r->overhead = conf->overhead;
920 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
921 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
922 r->mult = 1;
923 /*
924 * The deal here is to replace a divide by a reciprocal one
925 * in fast path (a reciprocal divide is a multiply and a shift)
926 *
927 * Normal formula would be :
928 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
929 *
930 * We compute mult/shift to use instead :
931 * time_in_ns = (len * mult) >> shift;
932 *
933 * We try to get the highest possible mult value for accuracy,
934 * but have to make sure no overflows will ever happen.
935 */
936 if (r->rate_bytes_ps > 0) {
937 u64 factor = NSEC_PER_SEC;
938
939 for (;;) {
940 r->mult = div64_u64(factor, r->rate_bytes_ps);
941 if (r->mult & (1U << 31) || factor & (1ULL << 63))
942 break;
943 factor <<= 1;
944 r->shift++;
945 }
946 }
947}
948EXPORT_SYMBOL(psched_ratecfg_precompute);
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 qdisc running bit guarantees that only one CPU
308 * 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 = READ_ONCE(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 res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
431 unsigned long val;
432 unsigned int i;
433
434 for (i = 1; i < dev->num_tx_queues; i++) {
435 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
436 if (val && time_after(val, res))
437 res = val;
438 }
439
440 return res;
441}
442EXPORT_SYMBOL(dev_trans_start);
443
444static void netif_freeze_queues(struct net_device *dev)
445{
446 unsigned int i;
447 int cpu;
448
449 cpu = smp_processor_id();
450 for (i = 0; i < dev->num_tx_queues; i++) {
451 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
452
453 /* We are the only thread of execution doing a
454 * freeze, but we have to grab the _xmit_lock in
455 * order to synchronize with threads which are in
456 * the ->hard_start_xmit() handler and already
457 * checked the frozen bit.
458 */
459 __netif_tx_lock(txq, cpu);
460 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
461 __netif_tx_unlock(txq);
462 }
463}
464
465void netif_tx_lock(struct net_device *dev)
466{
467 spin_lock(&dev->tx_global_lock);
468 netif_freeze_queues(dev);
469}
470EXPORT_SYMBOL(netif_tx_lock);
471
472static void netif_unfreeze_queues(struct net_device *dev)
473{
474 unsigned int i;
475
476 for (i = 0; i < dev->num_tx_queues; i++) {
477 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
478
479 /* No need to grab the _xmit_lock here. If the
480 * queue is not stopped for another reason, we
481 * force a schedule.
482 */
483 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
484 netif_schedule_queue(txq);
485 }
486}
487
488void netif_tx_unlock(struct net_device *dev)
489{
490 netif_unfreeze_queues(dev);
491 spin_unlock(&dev->tx_global_lock);
492}
493EXPORT_SYMBOL(netif_tx_unlock);
494
495static void dev_watchdog(struct timer_list *t)
496{
497 struct net_device *dev = from_timer(dev, t, watchdog_timer);
498 bool release = true;
499
500 spin_lock(&dev->tx_global_lock);
501 if (!qdisc_tx_is_noop(dev)) {
502 if (netif_device_present(dev) &&
503 netif_running(dev) &&
504 netif_carrier_ok(dev)) {
505 unsigned int timedout_ms = 0;
506 unsigned int i;
507 unsigned long trans_start;
508
509 for (i = 0; i < dev->num_tx_queues; i++) {
510 struct netdev_queue *txq;
511
512 txq = netdev_get_tx_queue(dev, i);
513 trans_start = READ_ONCE(txq->trans_start);
514 if (netif_xmit_stopped(txq) &&
515 time_after(jiffies, (trans_start +
516 dev->watchdog_timeo))) {
517 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
518 atomic_long_inc(&txq->trans_timeout);
519 break;
520 }
521 }
522
523 if (unlikely(timedout_ms)) {
524 trace_net_dev_xmit_timeout(dev, i);
525 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
526 raw_smp_processor_id(),
527 i, timedout_ms);
528 netif_freeze_queues(dev);
529 dev->netdev_ops->ndo_tx_timeout(dev, i);
530 netif_unfreeze_queues(dev);
531 }
532 if (!mod_timer(&dev->watchdog_timer,
533 round_jiffies(jiffies +
534 dev->watchdog_timeo)))
535 release = false;
536 }
537 }
538 spin_unlock(&dev->tx_global_lock);
539
540 if (release)
541 netdev_put(dev, &dev->watchdog_dev_tracker);
542}
543
544void __netdev_watchdog_up(struct net_device *dev)
545{
546 if (dev->netdev_ops->ndo_tx_timeout) {
547 if (dev->watchdog_timeo <= 0)
548 dev->watchdog_timeo = 5*HZ;
549 if (!mod_timer(&dev->watchdog_timer,
550 round_jiffies(jiffies + dev->watchdog_timeo)))
551 netdev_hold(dev, &dev->watchdog_dev_tracker,
552 GFP_ATOMIC);
553 }
554}
555EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
556
557static void dev_watchdog_up(struct net_device *dev)
558{
559 __netdev_watchdog_up(dev);
560}
561
562static void dev_watchdog_down(struct net_device *dev)
563{
564 netif_tx_lock_bh(dev);
565 if (del_timer(&dev->watchdog_timer))
566 netdev_put(dev, &dev->watchdog_dev_tracker);
567 netif_tx_unlock_bh(dev);
568}
569
570/**
571 * netif_carrier_on - set carrier
572 * @dev: network device
573 *
574 * Device has detected acquisition of carrier.
575 */
576void netif_carrier_on(struct net_device *dev)
577{
578 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
579 if (dev->reg_state == NETREG_UNINITIALIZED)
580 return;
581 atomic_inc(&dev->carrier_up_count);
582 linkwatch_fire_event(dev);
583 if (netif_running(dev))
584 __netdev_watchdog_up(dev);
585 }
586}
587EXPORT_SYMBOL(netif_carrier_on);
588
589/**
590 * netif_carrier_off - clear carrier
591 * @dev: network device
592 *
593 * Device has detected loss of carrier.
594 */
595void netif_carrier_off(struct net_device *dev)
596{
597 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
598 if (dev->reg_state == NETREG_UNINITIALIZED)
599 return;
600 atomic_inc(&dev->carrier_down_count);
601 linkwatch_fire_event(dev);
602 }
603}
604EXPORT_SYMBOL(netif_carrier_off);
605
606/**
607 * netif_carrier_event - report carrier state event
608 * @dev: network device
609 *
610 * Device has detected a carrier event but the carrier state wasn't changed.
611 * Use in drivers when querying carrier state asynchronously, to avoid missing
612 * events (link flaps) if link recovers before it's queried.
613 */
614void netif_carrier_event(struct net_device *dev)
615{
616 if (dev->reg_state == NETREG_UNINITIALIZED)
617 return;
618 atomic_inc(&dev->carrier_up_count);
619 atomic_inc(&dev->carrier_down_count);
620 linkwatch_fire_event(dev);
621}
622EXPORT_SYMBOL_GPL(netif_carrier_event);
623
624/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
625 under all circumstances. It is difficult to invent anything faster or
626 cheaper.
627 */
628
629static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
630 struct sk_buff **to_free)
631{
632 __qdisc_drop(skb, to_free);
633 return NET_XMIT_CN;
634}
635
636static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
637{
638 return NULL;
639}
640
641struct Qdisc_ops noop_qdisc_ops __read_mostly = {
642 .id = "noop",
643 .priv_size = 0,
644 .enqueue = noop_enqueue,
645 .dequeue = noop_dequeue,
646 .peek = noop_dequeue,
647 .owner = THIS_MODULE,
648};
649
650static struct netdev_queue noop_netdev_queue = {
651 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
652 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
653};
654
655struct Qdisc noop_qdisc = {
656 .enqueue = noop_enqueue,
657 .dequeue = noop_dequeue,
658 .flags = TCQ_F_BUILTIN,
659 .ops = &noop_qdisc_ops,
660 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
661 .dev_queue = &noop_netdev_queue,
662 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
663 .gso_skb = {
664 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
665 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
666 .qlen = 0,
667 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
668 },
669 .skb_bad_txq = {
670 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
671 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
672 .qlen = 0,
673 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
674 },
675};
676EXPORT_SYMBOL(noop_qdisc);
677
678static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
679 struct netlink_ext_ack *extack)
680{
681 /* register_qdisc() assigns a default of noop_enqueue if unset,
682 * but __dev_queue_xmit() treats noqueue only as such
683 * if this is NULL - so clear it here. */
684 qdisc->enqueue = NULL;
685 return 0;
686}
687
688struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
689 .id = "noqueue",
690 .priv_size = 0,
691 .init = noqueue_init,
692 .enqueue = noop_enqueue,
693 .dequeue = noop_dequeue,
694 .peek = noop_dequeue,
695 .owner = THIS_MODULE,
696};
697
698const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
699 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
700};
701EXPORT_SYMBOL(sch_default_prio2band);
702
703/* 3-band FIFO queue: old style, but should be a bit faster than
704 generic prio+fifo combination.
705 */
706
707#define PFIFO_FAST_BANDS 3
708
709/*
710 * Private data for a pfifo_fast scheduler containing:
711 * - rings for priority bands
712 */
713struct pfifo_fast_priv {
714 struct skb_array q[PFIFO_FAST_BANDS];
715};
716
717static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
718 int band)
719{
720 return &priv->q[band];
721}
722
723static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
724 struct sk_buff **to_free)
725{
726 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
727 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
728 struct skb_array *q = band2list(priv, band);
729 unsigned int pkt_len = qdisc_pkt_len(skb);
730 int err;
731
732 err = skb_array_produce(q, skb);
733
734 if (unlikely(err)) {
735 if (qdisc_is_percpu_stats(qdisc))
736 return qdisc_drop_cpu(skb, qdisc, to_free);
737 else
738 return qdisc_drop(skb, qdisc, to_free);
739 }
740
741 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
742 return NET_XMIT_SUCCESS;
743}
744
745static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
746{
747 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
748 struct sk_buff *skb = NULL;
749 bool need_retry = true;
750 int band;
751
752retry:
753 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
754 struct skb_array *q = band2list(priv, band);
755
756 if (__skb_array_empty(q))
757 continue;
758
759 skb = __skb_array_consume(q);
760 }
761 if (likely(skb)) {
762 qdisc_update_stats_at_dequeue(qdisc, skb);
763 } else if (need_retry &&
764 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
765 /* Delay clearing the STATE_MISSED here to reduce
766 * the overhead of the second spin_trylock() in
767 * qdisc_run_begin() and __netif_schedule() calling
768 * in qdisc_run_end().
769 */
770 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
771 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
772
773 /* Make sure dequeuing happens after clearing
774 * STATE_MISSED.
775 */
776 smp_mb__after_atomic();
777
778 need_retry = false;
779
780 goto retry;
781 }
782
783 return skb;
784}
785
786static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
787{
788 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
789 struct sk_buff *skb = NULL;
790 int band;
791
792 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
793 struct skb_array *q = band2list(priv, band);
794
795 skb = __skb_array_peek(q);
796 }
797
798 return skb;
799}
800
801static void pfifo_fast_reset(struct Qdisc *qdisc)
802{
803 int i, band;
804 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
805
806 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
807 struct skb_array *q = band2list(priv, band);
808 struct sk_buff *skb;
809
810 /* NULL ring is possible if destroy path is due to a failed
811 * skb_array_init() in pfifo_fast_init() case.
812 */
813 if (!q->ring.queue)
814 continue;
815
816 while ((skb = __skb_array_consume(q)) != NULL)
817 kfree_skb(skb);
818 }
819
820 if (qdisc_is_percpu_stats(qdisc)) {
821 for_each_possible_cpu(i) {
822 struct gnet_stats_queue *q;
823
824 q = per_cpu_ptr(qdisc->cpu_qstats, i);
825 q->backlog = 0;
826 q->qlen = 0;
827 }
828 }
829}
830
831static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
832{
833 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
834
835 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
836 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
837 goto nla_put_failure;
838 return skb->len;
839
840nla_put_failure:
841 return -1;
842}
843
844static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
845 struct netlink_ext_ack *extack)
846{
847 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
848 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
849 int prio;
850
851 /* guard against zero length rings */
852 if (!qlen)
853 return -EINVAL;
854
855 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
856 struct skb_array *q = band2list(priv, prio);
857 int err;
858
859 err = skb_array_init(q, qlen, GFP_KERNEL);
860 if (err)
861 return -ENOMEM;
862 }
863
864 /* Can by-pass the queue discipline */
865 qdisc->flags |= TCQ_F_CAN_BYPASS;
866 return 0;
867}
868
869static void pfifo_fast_destroy(struct Qdisc *sch)
870{
871 struct pfifo_fast_priv *priv = qdisc_priv(sch);
872 int prio;
873
874 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
875 struct skb_array *q = band2list(priv, prio);
876
877 /* NULL ring is possible if destroy path is due to a failed
878 * skb_array_init() in pfifo_fast_init() case.
879 */
880 if (!q->ring.queue)
881 continue;
882 /* Destroy ring but no need to kfree_skb because a call to
883 * pfifo_fast_reset() has already done that work.
884 */
885 ptr_ring_cleanup(&q->ring, NULL);
886 }
887}
888
889static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
890 unsigned int new_len)
891{
892 struct pfifo_fast_priv *priv = qdisc_priv(sch);
893 struct skb_array *bands[PFIFO_FAST_BANDS];
894 int prio;
895
896 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
897 struct skb_array *q = band2list(priv, prio);
898
899 bands[prio] = q;
900 }
901
902 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
903 GFP_KERNEL);
904}
905
906struct Qdisc_ops pfifo_fast_ops __read_mostly = {
907 .id = "pfifo_fast",
908 .priv_size = sizeof(struct pfifo_fast_priv),
909 .enqueue = pfifo_fast_enqueue,
910 .dequeue = pfifo_fast_dequeue,
911 .peek = pfifo_fast_peek,
912 .init = pfifo_fast_init,
913 .destroy = pfifo_fast_destroy,
914 .reset = pfifo_fast_reset,
915 .dump = pfifo_fast_dump,
916 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
917 .owner = THIS_MODULE,
918 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
919};
920EXPORT_SYMBOL(pfifo_fast_ops);
921
922static struct lock_class_key qdisc_tx_busylock;
923
924struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
925 const struct Qdisc_ops *ops,
926 struct netlink_ext_ack *extack)
927{
928 struct Qdisc *sch;
929 unsigned int size = sizeof(*sch) + ops->priv_size;
930 int err = -ENOBUFS;
931 struct net_device *dev;
932
933 if (!dev_queue) {
934 NL_SET_ERR_MSG(extack, "No device queue given");
935 err = -EINVAL;
936 goto errout;
937 }
938
939 dev = dev_queue->dev;
940 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
941
942 if (!sch)
943 goto errout;
944 __skb_queue_head_init(&sch->gso_skb);
945 __skb_queue_head_init(&sch->skb_bad_txq);
946 gnet_stats_basic_sync_init(&sch->bstats);
947 spin_lock_init(&sch->q.lock);
948
949 if (ops->static_flags & TCQ_F_CPUSTATS) {
950 sch->cpu_bstats =
951 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
952 if (!sch->cpu_bstats)
953 goto errout1;
954
955 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
956 if (!sch->cpu_qstats) {
957 free_percpu(sch->cpu_bstats);
958 goto errout1;
959 }
960 }
961
962 spin_lock_init(&sch->busylock);
963 lockdep_set_class(&sch->busylock,
964 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
965
966 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
967 spin_lock_init(&sch->seqlock);
968 lockdep_set_class(&sch->seqlock,
969 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
970
971 sch->ops = ops;
972 sch->flags = ops->static_flags;
973 sch->enqueue = ops->enqueue;
974 sch->dequeue = ops->dequeue;
975 sch->dev_queue = dev_queue;
976 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
977 refcount_set(&sch->refcnt, 1);
978
979 return sch;
980errout1:
981 kfree(sch);
982errout:
983 return ERR_PTR(err);
984}
985
986struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
987 const struct Qdisc_ops *ops,
988 unsigned int parentid,
989 struct netlink_ext_ack *extack)
990{
991 struct Qdisc *sch;
992
993 if (!try_module_get(ops->owner)) {
994 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
995 return NULL;
996 }
997
998 sch = qdisc_alloc(dev_queue, ops, extack);
999 if (IS_ERR(sch)) {
1000 module_put(ops->owner);
1001 return NULL;
1002 }
1003 sch->parent = parentid;
1004
1005 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1006 trace_qdisc_create(ops, dev_queue->dev, parentid);
1007 return sch;
1008 }
1009
1010 qdisc_put(sch);
1011 return NULL;
1012}
1013EXPORT_SYMBOL(qdisc_create_dflt);
1014
1015/* Under qdisc_lock(qdisc) and BH! */
1016
1017void qdisc_reset(struct Qdisc *qdisc)
1018{
1019 const struct Qdisc_ops *ops = qdisc->ops;
1020
1021 trace_qdisc_reset(qdisc);
1022
1023 if (ops->reset)
1024 ops->reset(qdisc);
1025
1026 __skb_queue_purge(&qdisc->gso_skb);
1027 __skb_queue_purge(&qdisc->skb_bad_txq);
1028
1029 qdisc->q.qlen = 0;
1030 qdisc->qstats.backlog = 0;
1031}
1032EXPORT_SYMBOL(qdisc_reset);
1033
1034void qdisc_free(struct Qdisc *qdisc)
1035{
1036 if (qdisc_is_percpu_stats(qdisc)) {
1037 free_percpu(qdisc->cpu_bstats);
1038 free_percpu(qdisc->cpu_qstats);
1039 }
1040
1041 kfree(qdisc);
1042}
1043
1044static void qdisc_free_cb(struct rcu_head *head)
1045{
1046 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1047
1048 qdisc_free(q);
1049}
1050
1051static void __qdisc_destroy(struct Qdisc *qdisc)
1052{
1053 const struct Qdisc_ops *ops = qdisc->ops;
1054 struct net_device *dev = qdisc_dev(qdisc);
1055
1056#ifdef CONFIG_NET_SCHED
1057 qdisc_hash_del(qdisc);
1058
1059 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1060#endif
1061 gen_kill_estimator(&qdisc->rate_est);
1062
1063 qdisc_reset(qdisc);
1064
1065
1066 if (ops->destroy)
1067 ops->destroy(qdisc);
1068
1069 module_put(ops->owner);
1070 netdev_put(dev, &qdisc->dev_tracker);
1071
1072 trace_qdisc_destroy(qdisc);
1073
1074 call_rcu(&qdisc->rcu, qdisc_free_cb);
1075}
1076
1077void qdisc_destroy(struct Qdisc *qdisc)
1078{
1079 if (qdisc->flags & TCQ_F_BUILTIN)
1080 return;
1081
1082 __qdisc_destroy(qdisc);
1083}
1084
1085void qdisc_put(struct Qdisc *qdisc)
1086{
1087 if (!qdisc)
1088 return;
1089
1090 if (qdisc->flags & TCQ_F_BUILTIN ||
1091 !refcount_dec_and_test(&qdisc->refcnt))
1092 return;
1093
1094 __qdisc_destroy(qdisc);
1095}
1096EXPORT_SYMBOL(qdisc_put);
1097
1098/* Version of qdisc_put() that is called with rtnl mutex unlocked.
1099 * Intended to be used as optimization, this function only takes rtnl lock if
1100 * qdisc reference counter reached zero.
1101 */
1102
1103void qdisc_put_unlocked(struct Qdisc *qdisc)
1104{
1105 if (qdisc->flags & TCQ_F_BUILTIN ||
1106 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1107 return;
1108
1109 __qdisc_destroy(qdisc);
1110 rtnl_unlock();
1111}
1112EXPORT_SYMBOL(qdisc_put_unlocked);
1113
1114/* Attach toplevel qdisc to device queue. */
1115struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1116 struct Qdisc *qdisc)
1117{
1118 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1119 spinlock_t *root_lock;
1120
1121 root_lock = qdisc_lock(oqdisc);
1122 spin_lock_bh(root_lock);
1123
1124 /* ... and graft new one */
1125 if (qdisc == NULL)
1126 qdisc = &noop_qdisc;
1127 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1128 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1129
1130 spin_unlock_bh(root_lock);
1131
1132 return oqdisc;
1133}
1134EXPORT_SYMBOL(dev_graft_qdisc);
1135
1136static void shutdown_scheduler_queue(struct net_device *dev,
1137 struct netdev_queue *dev_queue,
1138 void *_qdisc_default)
1139{
1140 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1141 struct Qdisc *qdisc_default = _qdisc_default;
1142
1143 if (qdisc) {
1144 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1145 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1146
1147 qdisc_put(qdisc);
1148 }
1149}
1150
1151static void attach_one_default_qdisc(struct net_device *dev,
1152 struct netdev_queue *dev_queue,
1153 void *_unused)
1154{
1155 struct Qdisc *qdisc;
1156 const struct Qdisc_ops *ops = default_qdisc_ops;
1157
1158 if (dev->priv_flags & IFF_NO_QUEUE)
1159 ops = &noqueue_qdisc_ops;
1160 else if(dev->type == ARPHRD_CAN)
1161 ops = &pfifo_fast_ops;
1162
1163 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1164 if (!qdisc)
1165 return;
1166
1167 if (!netif_is_multiqueue(dev))
1168 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1169 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1170}
1171
1172static void attach_default_qdiscs(struct net_device *dev)
1173{
1174 struct netdev_queue *txq;
1175 struct Qdisc *qdisc;
1176
1177 txq = netdev_get_tx_queue(dev, 0);
1178
1179 if (!netif_is_multiqueue(dev) ||
1180 dev->priv_flags & IFF_NO_QUEUE) {
1181 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1182 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1183 rcu_assign_pointer(dev->qdisc, qdisc);
1184 qdisc_refcount_inc(qdisc);
1185 } else {
1186 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1187 if (qdisc) {
1188 rcu_assign_pointer(dev->qdisc, qdisc);
1189 qdisc->ops->attach(qdisc);
1190 }
1191 }
1192 qdisc = rtnl_dereference(dev->qdisc);
1193
1194 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1195 if (qdisc == &noop_qdisc) {
1196 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1197 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1198 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1199 dev->priv_flags |= IFF_NO_QUEUE;
1200 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1201 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1202 rcu_assign_pointer(dev->qdisc, qdisc);
1203 qdisc_refcount_inc(qdisc);
1204 dev->priv_flags ^= IFF_NO_QUEUE;
1205 }
1206
1207#ifdef CONFIG_NET_SCHED
1208 if (qdisc != &noop_qdisc)
1209 qdisc_hash_add(qdisc, false);
1210#endif
1211}
1212
1213static void transition_one_qdisc(struct net_device *dev,
1214 struct netdev_queue *dev_queue,
1215 void *_need_watchdog)
1216{
1217 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1218 int *need_watchdog_p = _need_watchdog;
1219
1220 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1221 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1222
1223 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1224 if (need_watchdog_p) {
1225 WRITE_ONCE(dev_queue->trans_start, 0);
1226 *need_watchdog_p = 1;
1227 }
1228}
1229
1230void dev_activate(struct net_device *dev)
1231{
1232 int need_watchdog;
1233
1234 /* No queueing discipline is attached to device;
1235 * create default one for devices, which need queueing
1236 * and noqueue_qdisc for virtual interfaces
1237 */
1238
1239 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1240 attach_default_qdiscs(dev);
1241
1242 if (!netif_carrier_ok(dev))
1243 /* Delay activation until next carrier-on event */
1244 return;
1245
1246 need_watchdog = 0;
1247 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1248 if (dev_ingress_queue(dev))
1249 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1250
1251 if (need_watchdog) {
1252 netif_trans_update(dev);
1253 dev_watchdog_up(dev);
1254 }
1255}
1256EXPORT_SYMBOL(dev_activate);
1257
1258static void qdisc_deactivate(struct Qdisc *qdisc)
1259{
1260 if (qdisc->flags & TCQ_F_BUILTIN)
1261 return;
1262
1263 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1264}
1265
1266static void dev_deactivate_queue(struct net_device *dev,
1267 struct netdev_queue *dev_queue,
1268 void *_qdisc_default)
1269{
1270 struct Qdisc *qdisc_default = _qdisc_default;
1271 struct Qdisc *qdisc;
1272
1273 qdisc = rtnl_dereference(dev_queue->qdisc);
1274 if (qdisc) {
1275 qdisc_deactivate(qdisc);
1276 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1277 }
1278}
1279
1280static void dev_reset_queue(struct net_device *dev,
1281 struct netdev_queue *dev_queue,
1282 void *_unused)
1283{
1284 struct Qdisc *qdisc;
1285 bool nolock;
1286
1287 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1288 if (!qdisc)
1289 return;
1290
1291 nolock = qdisc->flags & TCQ_F_NOLOCK;
1292
1293 if (nolock)
1294 spin_lock_bh(&qdisc->seqlock);
1295 spin_lock_bh(qdisc_lock(qdisc));
1296
1297 qdisc_reset(qdisc);
1298
1299 spin_unlock_bh(qdisc_lock(qdisc));
1300 if (nolock) {
1301 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1302 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1303 spin_unlock_bh(&qdisc->seqlock);
1304 }
1305}
1306
1307static bool some_qdisc_is_busy(struct net_device *dev)
1308{
1309 unsigned int i;
1310
1311 for (i = 0; i < dev->num_tx_queues; i++) {
1312 struct netdev_queue *dev_queue;
1313 spinlock_t *root_lock;
1314 struct Qdisc *q;
1315 int val;
1316
1317 dev_queue = netdev_get_tx_queue(dev, i);
1318 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1319
1320 root_lock = qdisc_lock(q);
1321 spin_lock_bh(root_lock);
1322
1323 val = (qdisc_is_running(q) ||
1324 test_bit(__QDISC_STATE_SCHED, &q->state));
1325
1326 spin_unlock_bh(root_lock);
1327
1328 if (val)
1329 return true;
1330 }
1331 return false;
1332}
1333
1334/**
1335 * dev_deactivate_many - deactivate transmissions on several devices
1336 * @head: list of devices to deactivate
1337 *
1338 * This function returns only when all outstanding transmissions
1339 * have completed, unless all devices are in dismantle phase.
1340 */
1341void dev_deactivate_many(struct list_head *head)
1342{
1343 struct net_device *dev;
1344
1345 list_for_each_entry(dev, head, close_list) {
1346 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1347 &noop_qdisc);
1348 if (dev_ingress_queue(dev))
1349 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1350 &noop_qdisc);
1351
1352 dev_watchdog_down(dev);
1353 }
1354
1355 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1356 * outstanding qdisc enqueuing calls.
1357 * This is avoided if all devices are in dismantle phase :
1358 * Caller will call synchronize_net() for us
1359 */
1360 synchronize_net();
1361
1362 list_for_each_entry(dev, head, close_list) {
1363 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1364
1365 if (dev_ingress_queue(dev))
1366 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1367 }
1368
1369 /* Wait for outstanding qdisc_run calls. */
1370 list_for_each_entry(dev, head, close_list) {
1371 while (some_qdisc_is_busy(dev)) {
1372 /* wait_event() would avoid this sleep-loop but would
1373 * require expensive checks in the fast paths of packet
1374 * processing which isn't worth it.
1375 */
1376 schedule_timeout_uninterruptible(1);
1377 }
1378 }
1379}
1380
1381void dev_deactivate(struct net_device *dev)
1382{
1383 LIST_HEAD(single);
1384
1385 list_add(&dev->close_list, &single);
1386 dev_deactivate_many(&single);
1387 list_del(&single);
1388}
1389EXPORT_SYMBOL(dev_deactivate);
1390
1391static int qdisc_change_tx_queue_len(struct net_device *dev,
1392 struct netdev_queue *dev_queue)
1393{
1394 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1395 const struct Qdisc_ops *ops = qdisc->ops;
1396
1397 if (ops->change_tx_queue_len)
1398 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1399 return 0;
1400}
1401
1402void dev_qdisc_change_real_num_tx(struct net_device *dev,
1403 unsigned int new_real_tx)
1404{
1405 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1406
1407 if (qdisc->ops->change_real_num_tx)
1408 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1409}
1410
1411void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1412{
1413#ifdef CONFIG_NET_SCHED
1414 struct net_device *dev = qdisc_dev(sch);
1415 struct Qdisc *qdisc;
1416 unsigned int i;
1417
1418 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1419 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1420 /* Only update the default qdiscs we created,
1421 * qdiscs with handles are always hashed.
1422 */
1423 if (qdisc != &noop_qdisc && !qdisc->handle)
1424 qdisc_hash_del(qdisc);
1425 }
1426 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1427 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1428 if (qdisc != &noop_qdisc && !qdisc->handle)
1429 qdisc_hash_add(qdisc, false);
1430 }
1431#endif
1432}
1433EXPORT_SYMBOL(mq_change_real_num_tx);
1434
1435int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1436{
1437 bool up = dev->flags & IFF_UP;
1438 unsigned int i;
1439 int ret = 0;
1440
1441 if (up)
1442 dev_deactivate(dev);
1443
1444 for (i = 0; i < dev->num_tx_queues; i++) {
1445 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1446
1447 /* TODO: revert changes on a partial failure */
1448 if (ret)
1449 break;
1450 }
1451
1452 if (up)
1453 dev_activate(dev);
1454 return ret;
1455}
1456
1457static void dev_init_scheduler_queue(struct net_device *dev,
1458 struct netdev_queue *dev_queue,
1459 void *_qdisc)
1460{
1461 struct Qdisc *qdisc = _qdisc;
1462
1463 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1464 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1465}
1466
1467void dev_init_scheduler(struct net_device *dev)
1468{
1469 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1470 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1471 if (dev_ingress_queue(dev))
1472 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1473
1474 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1475}
1476
1477void dev_shutdown(struct net_device *dev)
1478{
1479 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1480 if (dev_ingress_queue(dev))
1481 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1482 qdisc_put(rtnl_dereference(dev->qdisc));
1483 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1484
1485 WARN_ON(timer_pending(&dev->watchdog_timer));
1486}
1487
1488/**
1489 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1490 * @rate: Rate to compute reciprocal division values of
1491 * @mult: Multiplier for reciprocal division
1492 * @shift: Shift for reciprocal division
1493 *
1494 * The multiplier and shift for reciprocal division by rate are stored
1495 * in mult and shift.
1496 *
1497 * The deal here is to replace a divide by a reciprocal one
1498 * in fast path (a reciprocal divide is a multiply and a shift)
1499 *
1500 * Normal formula would be :
1501 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1502 *
1503 * We compute mult/shift to use instead :
1504 * time_in_ns = (len * mult) >> shift;
1505 *
1506 * We try to get the highest possible mult value for accuracy,
1507 * but have to make sure no overflows will ever happen.
1508 *
1509 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1510 */
1511static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1512{
1513 u64 factor = NSEC_PER_SEC;
1514
1515 *mult = 1;
1516 *shift = 0;
1517
1518 if (rate <= 0)
1519 return;
1520
1521 for (;;) {
1522 *mult = div64_u64(factor, rate);
1523 if (*mult & (1U << 31) || factor & (1ULL << 63))
1524 break;
1525 factor <<= 1;
1526 (*shift)++;
1527 }
1528}
1529
1530void psched_ratecfg_precompute(struct psched_ratecfg *r,
1531 const struct tc_ratespec *conf,
1532 u64 rate64)
1533{
1534 memset(r, 0, sizeof(*r));
1535 r->overhead = conf->overhead;
1536 r->mpu = conf->mpu;
1537 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1538 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1539 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1540}
1541EXPORT_SYMBOL(psched_ratecfg_precompute);
1542
1543void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1544{
1545 r->rate_pkts_ps = pktrate64;
1546 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1547}
1548EXPORT_SYMBOL(psched_ppscfg_precompute);
1549
1550void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1551 struct tcf_proto *tp_head)
1552{
1553 /* Protected with chain0->filter_chain_lock.
1554 * Can't access chain directly because tp_head can be NULL.
1555 */
1556 struct mini_Qdisc *miniq_old =
1557 rcu_dereference_protected(*miniqp->p_miniq, 1);
1558 struct mini_Qdisc *miniq;
1559
1560 if (!tp_head) {
1561 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1562 } else {
1563 miniq = miniq_old != &miniqp->miniq1 ?
1564 &miniqp->miniq1 : &miniqp->miniq2;
1565
1566 /* We need to make sure that readers won't see the miniq
1567 * we are about to modify. So ensure that at least one RCU
1568 * grace period has elapsed since the miniq was made
1569 * inactive.
1570 */
1571 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1572 cond_synchronize_rcu(miniq->rcu_state);
1573 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1574 synchronize_rcu_expedited();
1575
1576 miniq->filter_list = tp_head;
1577 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1578 }
1579
1580 if (miniq_old)
1581 /* This is counterpart of the rcu sync above. We need to
1582 * block potential new user of miniq_old until all readers
1583 * are not seeing it.
1584 */
1585 miniq_old->rcu_state = start_poll_synchronize_rcu();
1586}
1587EXPORT_SYMBOL(mini_qdisc_pair_swap);
1588
1589void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1590 struct tcf_block *block)
1591{
1592 miniqp->miniq1.block = block;
1593 miniqp->miniq2.block = block;
1594}
1595EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1596
1597void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1598 struct mini_Qdisc __rcu **p_miniq)
1599{
1600 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1601 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1602 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1603 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1604 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1605 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1606 miniqp->p_miniq = p_miniq;
1607}
1608EXPORT_SYMBOL(mini_qdisc_pair_init);