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