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/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/sched.h>
19#include <linux/string.h>
20#include <linux/errno.h>
21#include <linux/netdevice.h>
22#include <linux/skbuff.h>
23#include <linux/rtnetlink.h>
24#include <linux/init.h>
25#include <linux/rcupdate.h>
26#include <linux/list.h>
27#include <linux/slab.h>
28#include <linux/if_vlan.h>
29#include <net/sch_generic.h>
30#include <net/pkt_sched.h>
31#include <net/dst.h>
32
33/* Qdisc to use by default */
34const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
35EXPORT_SYMBOL(default_qdisc_ops);
36
37/* Main transmission queue. */
38
39/* Modifications to data participating in scheduling must be protected with
40 * qdisc_lock(qdisc) spinlock.
41 *
42 * The idea is the following:
43 * - enqueue, dequeue are serialized via qdisc root lock
44 * - ingress filtering is also serialized via qdisc root lock
45 * - updates to tree and tree walking are only done under the rtnl mutex.
46 */
47
48static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
49{
50 q->gso_skb = skb;
51 q->qstats.requeues++;
52 qdisc_qstats_backlog_inc(q, skb);
53 q->q.qlen++; /* it's still part of the queue */
54 __netif_schedule(q);
55
56 return 0;
57}
58
59static void try_bulk_dequeue_skb(struct Qdisc *q,
60 struct sk_buff *skb,
61 const struct netdev_queue *txq,
62 int *packets)
63{
64 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
65
66 while (bytelimit > 0) {
67 struct sk_buff *nskb = q->dequeue(q);
68
69 if (!nskb)
70 break;
71
72 bytelimit -= nskb->len; /* covers GSO len */
73 skb->next = nskb;
74 skb = nskb;
75 (*packets)++; /* GSO counts as one pkt */
76 }
77 skb->next = NULL;
78}
79
80/* This variant of try_bulk_dequeue_skb() makes sure
81 * all skbs in the chain are for the same txq
82 */
83static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
84 struct sk_buff *skb,
85 int *packets)
86{
87 int mapping = skb_get_queue_mapping(skb);
88 struct sk_buff *nskb;
89 int cnt = 0;
90
91 do {
92 nskb = q->dequeue(q);
93 if (!nskb)
94 break;
95 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
96 q->skb_bad_txq = nskb;
97 qdisc_qstats_backlog_inc(q, nskb);
98 q->q.qlen++;
99 break;
100 }
101 skb->next = nskb;
102 skb = nskb;
103 } while (++cnt < 8);
104 (*packets) += cnt;
105 skb->next = NULL;
106}
107
108/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
109 * A requeued skb (via q->gso_skb) can also be a SKB list.
110 */
111static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
112 int *packets)
113{
114 struct sk_buff *skb = q->gso_skb;
115 const struct netdev_queue *txq = q->dev_queue;
116
117 *packets = 1;
118 if (unlikely(skb)) {
119 /* skb in gso_skb were already validated */
120 *validate = false;
121 /* check the reason of requeuing without tx lock first */
122 txq = skb_get_tx_queue(txq->dev, skb);
123 if (!netif_xmit_frozen_or_stopped(txq)) {
124 q->gso_skb = NULL;
125 qdisc_qstats_backlog_dec(q, skb);
126 q->q.qlen--;
127 } else
128 skb = NULL;
129 return skb;
130 }
131 *validate = true;
132 skb = q->skb_bad_txq;
133 if (unlikely(skb)) {
134 /* check the reason of requeuing without tx lock first */
135 txq = skb_get_tx_queue(txq->dev, skb);
136 if (!netif_xmit_frozen_or_stopped(txq)) {
137 q->skb_bad_txq = NULL;
138 qdisc_qstats_backlog_dec(q, skb);
139 q->q.qlen--;
140 goto bulk;
141 }
142 return NULL;
143 }
144 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
145 !netif_xmit_frozen_or_stopped(txq))
146 skb = q->dequeue(q);
147 if (skb) {
148bulk:
149 if (qdisc_may_bulk(q))
150 try_bulk_dequeue_skb(q, skb, txq, packets);
151 else
152 try_bulk_dequeue_skb_slow(q, skb, packets);
153 }
154 return skb;
155}
156
157/*
158 * Transmit possibly several skbs, and handle the return status as
159 * required. Owning running seqcount bit guarantees that
160 * only one CPU can execute this function.
161 *
162 * Returns to the caller:
163 * 0 - queue is empty or throttled.
164 * >0 - queue is not empty.
165 */
166int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
167 struct net_device *dev, struct netdev_queue *txq,
168 spinlock_t *root_lock, bool validate)
169{
170 int ret = NETDEV_TX_BUSY;
171
172 /* And release qdisc */
173 spin_unlock(root_lock);
174
175 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
176 if (validate)
177 skb = validate_xmit_skb_list(skb, dev);
178
179 if (likely(skb)) {
180 HARD_TX_LOCK(dev, txq, smp_processor_id());
181 if (!netif_xmit_frozen_or_stopped(txq))
182 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
183
184 HARD_TX_UNLOCK(dev, txq);
185 } else {
186 spin_lock(root_lock);
187 return qdisc_qlen(q);
188 }
189 spin_lock(root_lock);
190
191 if (dev_xmit_complete(ret)) {
192 /* Driver sent out skb successfully or skb was consumed */
193 ret = qdisc_qlen(q);
194 } else {
195 /* Driver returned NETDEV_TX_BUSY - requeue skb */
196 if (unlikely(ret != NETDEV_TX_BUSY))
197 net_warn_ratelimited("BUG %s code %d qlen %d\n",
198 dev->name, ret, q->q.qlen);
199
200 ret = dev_requeue_skb(skb, q);
201 }
202
203 if (ret && netif_xmit_frozen_or_stopped(txq))
204 ret = 0;
205
206 return ret;
207}
208
209/*
210 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
211 *
212 * running seqcount guarantees only one CPU can process
213 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
214 * this queue.
215 *
216 * netif_tx_lock serializes accesses to device driver.
217 *
218 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
219 * if one is grabbed, another must be free.
220 *
221 * Note, that this procedure can be called by a watchdog timer
222 *
223 * Returns to the caller:
224 * 0 - queue is empty or throttled.
225 * >0 - queue is not empty.
226 *
227 */
228static inline int qdisc_restart(struct Qdisc *q, int *packets)
229{
230 struct netdev_queue *txq;
231 struct net_device *dev;
232 spinlock_t *root_lock;
233 struct sk_buff *skb;
234 bool validate;
235
236 /* Dequeue packet */
237 skb = dequeue_skb(q, &validate, packets);
238 if (unlikely(!skb))
239 return 0;
240
241 root_lock = qdisc_lock(q);
242 dev = qdisc_dev(q);
243 txq = skb_get_tx_queue(dev, skb);
244
245 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
246}
247
248void __qdisc_run(struct Qdisc *q)
249{
250 int quota = weight_p;
251 int packets;
252
253 while (qdisc_restart(q, &packets)) {
254 /*
255 * Ordered by possible occurrence: Postpone processing if
256 * 1. we've exceeded packet quota
257 * 2. another process needs the CPU;
258 */
259 quota -= packets;
260 if (quota <= 0 || need_resched()) {
261 __netif_schedule(q);
262 break;
263 }
264 }
265
266 qdisc_run_end(q);
267}
268
269unsigned long dev_trans_start(struct net_device *dev)
270{
271 unsigned long val, res;
272 unsigned int i;
273
274 if (is_vlan_dev(dev))
275 dev = vlan_dev_real_dev(dev);
276 res = netdev_get_tx_queue(dev, 0)->trans_start;
277 for (i = 1; i < dev->num_tx_queues; i++) {
278 val = netdev_get_tx_queue(dev, i)->trans_start;
279 if (val && time_after(val, res))
280 res = val;
281 }
282
283 return res;
284}
285EXPORT_SYMBOL(dev_trans_start);
286
287static void dev_watchdog(unsigned long arg)
288{
289 struct net_device *dev = (struct net_device *)arg;
290
291 netif_tx_lock(dev);
292 if (!qdisc_tx_is_noop(dev)) {
293 if (netif_device_present(dev) &&
294 netif_running(dev) &&
295 netif_carrier_ok(dev)) {
296 int some_queue_timedout = 0;
297 unsigned int i;
298 unsigned long trans_start;
299
300 for (i = 0; i < dev->num_tx_queues; i++) {
301 struct netdev_queue *txq;
302
303 txq = netdev_get_tx_queue(dev, i);
304 trans_start = txq->trans_start;
305 if (netif_xmit_stopped(txq) &&
306 time_after(jiffies, (trans_start +
307 dev->watchdog_timeo))) {
308 some_queue_timedout = 1;
309 txq->trans_timeout++;
310 break;
311 }
312 }
313
314 if (some_queue_timedout) {
315 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
316 dev->name, netdev_drivername(dev), i);
317 dev->netdev_ops->ndo_tx_timeout(dev);
318 }
319 if (!mod_timer(&dev->watchdog_timer,
320 round_jiffies(jiffies +
321 dev->watchdog_timeo)))
322 dev_hold(dev);
323 }
324 }
325 netif_tx_unlock(dev);
326
327 dev_put(dev);
328}
329
330void __netdev_watchdog_up(struct net_device *dev)
331{
332 if (dev->netdev_ops->ndo_tx_timeout) {
333 if (dev->watchdog_timeo <= 0)
334 dev->watchdog_timeo = 5*HZ;
335 if (!mod_timer(&dev->watchdog_timer,
336 round_jiffies(jiffies + dev->watchdog_timeo)))
337 dev_hold(dev);
338 }
339}
340
341static void dev_watchdog_up(struct net_device *dev)
342{
343 __netdev_watchdog_up(dev);
344}
345
346static void dev_watchdog_down(struct net_device *dev)
347{
348 netif_tx_lock_bh(dev);
349 if (del_timer(&dev->watchdog_timer))
350 dev_put(dev);
351 netif_tx_unlock_bh(dev);
352}
353
354/**
355 * netif_carrier_on - set carrier
356 * @dev: network device
357 *
358 * Device has detected that carrier.
359 */
360void netif_carrier_on(struct net_device *dev)
361{
362 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
363 if (dev->reg_state == NETREG_UNINITIALIZED)
364 return;
365 atomic_inc(&dev->carrier_changes);
366 linkwatch_fire_event(dev);
367 if (netif_running(dev))
368 __netdev_watchdog_up(dev);
369 }
370}
371EXPORT_SYMBOL(netif_carrier_on);
372
373/**
374 * netif_carrier_off - clear carrier
375 * @dev: network device
376 *
377 * Device has detected loss of carrier.
378 */
379void netif_carrier_off(struct net_device *dev)
380{
381 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
382 if (dev->reg_state == NETREG_UNINITIALIZED)
383 return;
384 atomic_inc(&dev->carrier_changes);
385 linkwatch_fire_event(dev);
386 }
387}
388EXPORT_SYMBOL(netif_carrier_off);
389
390/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
391 under all circumstances. It is difficult to invent anything faster or
392 cheaper.
393 */
394
395static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
396 struct sk_buff **to_free)
397{
398 __qdisc_drop(skb, to_free);
399 return NET_XMIT_CN;
400}
401
402static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
403{
404 return NULL;
405}
406
407struct Qdisc_ops noop_qdisc_ops __read_mostly = {
408 .id = "noop",
409 .priv_size = 0,
410 .enqueue = noop_enqueue,
411 .dequeue = noop_dequeue,
412 .peek = noop_dequeue,
413 .owner = THIS_MODULE,
414};
415
416static struct netdev_queue noop_netdev_queue = {
417 .qdisc = &noop_qdisc,
418 .qdisc_sleeping = &noop_qdisc,
419};
420
421struct Qdisc noop_qdisc = {
422 .enqueue = noop_enqueue,
423 .dequeue = noop_dequeue,
424 .flags = TCQ_F_BUILTIN,
425 .ops = &noop_qdisc_ops,
426 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
427 .dev_queue = &noop_netdev_queue,
428 .running = SEQCNT_ZERO(noop_qdisc.running),
429 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
430};
431EXPORT_SYMBOL(noop_qdisc);
432
433static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
434{
435 /* register_qdisc() assigns a default of noop_enqueue if unset,
436 * but __dev_queue_xmit() treats noqueue only as such
437 * if this is NULL - so clear it here. */
438 qdisc->enqueue = NULL;
439 return 0;
440}
441
442struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
443 .id = "noqueue",
444 .priv_size = 0,
445 .init = noqueue_init,
446 .enqueue = noop_enqueue,
447 .dequeue = noop_dequeue,
448 .peek = noop_dequeue,
449 .owner = THIS_MODULE,
450};
451
452static const u8 prio2band[TC_PRIO_MAX + 1] = {
453 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
454};
455
456/* 3-band FIFO queue: old style, but should be a bit faster than
457 generic prio+fifo combination.
458 */
459
460#define PFIFO_FAST_BANDS 3
461
462/*
463 * Private data for a pfifo_fast scheduler containing:
464 * - queues for the three band
465 * - bitmap indicating which of the bands contain skbs
466 */
467struct pfifo_fast_priv {
468 u32 bitmap;
469 struct qdisc_skb_head q[PFIFO_FAST_BANDS];
470};
471
472/*
473 * Convert a bitmap to the first band number where an skb is queued, where:
474 * bitmap=0 means there are no skbs on any band.
475 * bitmap=1 means there is an skb on band 0.
476 * bitmap=7 means there are skbs on all 3 bands, etc.
477 */
478static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
479
480static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
481 int band)
482{
483 return priv->q + band;
484}
485
486static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
487 struct sk_buff **to_free)
488{
489 if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
490 int band = prio2band[skb->priority & TC_PRIO_MAX];
491 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
492 struct qdisc_skb_head *list = band2list(priv, band);
493
494 priv->bitmap |= (1 << band);
495 qdisc->q.qlen++;
496 return __qdisc_enqueue_tail(skb, qdisc, list);
497 }
498
499 return qdisc_drop(skb, qdisc, to_free);
500}
501
502static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
503{
504 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
505 int band = bitmap2band[priv->bitmap];
506
507 if (likely(band >= 0)) {
508 struct qdisc_skb_head *qh = band2list(priv, band);
509 struct sk_buff *skb = __qdisc_dequeue_head(qh);
510
511 if (likely(skb != NULL)) {
512 qdisc_qstats_backlog_dec(qdisc, skb);
513 qdisc_bstats_update(qdisc, skb);
514 }
515
516 qdisc->q.qlen--;
517 if (qh->qlen == 0)
518 priv->bitmap &= ~(1 << band);
519
520 return skb;
521 }
522
523 return NULL;
524}
525
526static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
527{
528 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
529 int band = bitmap2band[priv->bitmap];
530
531 if (band >= 0) {
532 struct qdisc_skb_head *qh = band2list(priv, band);
533
534 return qh->head;
535 }
536
537 return NULL;
538}
539
540static void pfifo_fast_reset(struct Qdisc *qdisc)
541{
542 int prio;
543 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
544
545 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
546 __qdisc_reset_queue(band2list(priv, prio));
547
548 priv->bitmap = 0;
549 qdisc->qstats.backlog = 0;
550 qdisc->q.qlen = 0;
551}
552
553static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
554{
555 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
556
557 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
558 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
559 goto nla_put_failure;
560 return skb->len;
561
562nla_put_failure:
563 return -1;
564}
565
566static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
567{
568 int prio;
569 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
570
571 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
572 qdisc_skb_head_init(band2list(priv, prio));
573
574 /* Can by-pass the queue discipline */
575 qdisc->flags |= TCQ_F_CAN_BYPASS;
576 return 0;
577}
578
579struct Qdisc_ops pfifo_fast_ops __read_mostly = {
580 .id = "pfifo_fast",
581 .priv_size = sizeof(struct pfifo_fast_priv),
582 .enqueue = pfifo_fast_enqueue,
583 .dequeue = pfifo_fast_dequeue,
584 .peek = pfifo_fast_peek,
585 .init = pfifo_fast_init,
586 .reset = pfifo_fast_reset,
587 .dump = pfifo_fast_dump,
588 .owner = THIS_MODULE,
589};
590EXPORT_SYMBOL(pfifo_fast_ops);
591
592static struct lock_class_key qdisc_tx_busylock;
593static struct lock_class_key qdisc_running_key;
594
595struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
596 const struct Qdisc_ops *ops)
597{
598 void *p;
599 struct Qdisc *sch;
600 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
601 int err = -ENOBUFS;
602 struct net_device *dev = dev_queue->dev;
603
604 p = kzalloc_node(size, GFP_KERNEL,
605 netdev_queue_numa_node_read(dev_queue));
606
607 if (!p)
608 goto errout;
609 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
610 /* if we got non aligned memory, ask more and do alignment ourself */
611 if (sch != p) {
612 kfree(p);
613 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
614 netdev_queue_numa_node_read(dev_queue));
615 if (!p)
616 goto errout;
617 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
618 sch->padded = (char *) sch - (char *) p;
619 }
620 qdisc_skb_head_init(&sch->q);
621 spin_lock_init(&sch->q.lock);
622
623 spin_lock_init(&sch->busylock);
624 lockdep_set_class(&sch->busylock,
625 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
626
627 seqcount_init(&sch->running);
628 lockdep_set_class(&sch->running,
629 dev->qdisc_running_key ?: &qdisc_running_key);
630
631 sch->ops = ops;
632 sch->enqueue = ops->enqueue;
633 sch->dequeue = ops->dequeue;
634 sch->dev_queue = dev_queue;
635 dev_hold(dev);
636 atomic_set(&sch->refcnt, 1);
637
638 return sch;
639errout:
640 return ERR_PTR(err);
641}
642
643struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
644 const struct Qdisc_ops *ops,
645 unsigned int parentid)
646{
647 struct Qdisc *sch;
648
649 if (!try_module_get(ops->owner))
650 return NULL;
651
652 sch = qdisc_alloc(dev_queue, ops);
653 if (IS_ERR(sch)) {
654 module_put(ops->owner);
655 return NULL;
656 }
657 sch->parent = parentid;
658
659 if (!ops->init || ops->init(sch, NULL) == 0)
660 return sch;
661
662 qdisc_destroy(sch);
663 return NULL;
664}
665EXPORT_SYMBOL(qdisc_create_dflt);
666
667/* Under qdisc_lock(qdisc) and BH! */
668
669void qdisc_reset(struct Qdisc *qdisc)
670{
671 const struct Qdisc_ops *ops = qdisc->ops;
672
673 if (ops->reset)
674 ops->reset(qdisc);
675
676 kfree_skb(qdisc->skb_bad_txq);
677 qdisc->skb_bad_txq = NULL;
678
679 if (qdisc->gso_skb) {
680 kfree_skb_list(qdisc->gso_skb);
681 qdisc->gso_skb = NULL;
682 }
683 qdisc->q.qlen = 0;
684}
685EXPORT_SYMBOL(qdisc_reset);
686
687static void qdisc_rcu_free(struct rcu_head *head)
688{
689 struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
690
691 if (qdisc_is_percpu_stats(qdisc)) {
692 free_percpu(qdisc->cpu_bstats);
693 free_percpu(qdisc->cpu_qstats);
694 }
695
696 kfree((char *) qdisc - qdisc->padded);
697}
698
699void qdisc_destroy(struct Qdisc *qdisc)
700{
701 const struct Qdisc_ops *ops = qdisc->ops;
702
703 if (qdisc->flags & TCQ_F_BUILTIN ||
704 !atomic_dec_and_test(&qdisc->refcnt))
705 return;
706
707#ifdef CONFIG_NET_SCHED
708 qdisc_hash_del(qdisc);
709
710 qdisc_put_stab(rtnl_dereference(qdisc->stab));
711#endif
712 gen_kill_estimator(&qdisc->rate_est);
713 if (ops->reset)
714 ops->reset(qdisc);
715 if (ops->destroy)
716 ops->destroy(qdisc);
717
718 module_put(ops->owner);
719 dev_put(qdisc_dev(qdisc));
720
721 kfree_skb_list(qdisc->gso_skb);
722 kfree_skb(qdisc->skb_bad_txq);
723 /*
724 * gen_estimator est_timer() might access qdisc->q.lock,
725 * wait a RCU grace period before freeing qdisc.
726 */
727 call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
728}
729EXPORT_SYMBOL(qdisc_destroy);
730
731/* Attach toplevel qdisc to device queue. */
732struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
733 struct Qdisc *qdisc)
734{
735 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
736 spinlock_t *root_lock;
737
738 root_lock = qdisc_lock(oqdisc);
739 spin_lock_bh(root_lock);
740
741 /* Prune old scheduler */
742 if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
743 qdisc_reset(oqdisc);
744
745 /* ... and graft new one */
746 if (qdisc == NULL)
747 qdisc = &noop_qdisc;
748 dev_queue->qdisc_sleeping = qdisc;
749 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
750
751 spin_unlock_bh(root_lock);
752
753 return oqdisc;
754}
755EXPORT_SYMBOL(dev_graft_qdisc);
756
757static void attach_one_default_qdisc(struct net_device *dev,
758 struct netdev_queue *dev_queue,
759 void *_unused)
760{
761 struct Qdisc *qdisc;
762 const struct Qdisc_ops *ops = default_qdisc_ops;
763
764 if (dev->priv_flags & IFF_NO_QUEUE)
765 ops = &noqueue_qdisc_ops;
766
767 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
768 if (!qdisc) {
769 netdev_info(dev, "activation failed\n");
770 return;
771 }
772 if (!netif_is_multiqueue(dev))
773 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
774 dev_queue->qdisc_sleeping = qdisc;
775}
776
777static void attach_default_qdiscs(struct net_device *dev)
778{
779 struct netdev_queue *txq;
780 struct Qdisc *qdisc;
781
782 txq = netdev_get_tx_queue(dev, 0);
783
784 if (!netif_is_multiqueue(dev) ||
785 dev->priv_flags & IFF_NO_QUEUE) {
786 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
787 dev->qdisc = txq->qdisc_sleeping;
788 atomic_inc(&dev->qdisc->refcnt);
789 } else {
790 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
791 if (qdisc) {
792 dev->qdisc = qdisc;
793 qdisc->ops->attach(qdisc);
794 }
795 }
796#ifdef CONFIG_NET_SCHED
797 if (dev->qdisc)
798 qdisc_hash_add(dev->qdisc);
799#endif
800}
801
802static void transition_one_qdisc(struct net_device *dev,
803 struct netdev_queue *dev_queue,
804 void *_need_watchdog)
805{
806 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
807 int *need_watchdog_p = _need_watchdog;
808
809 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
810 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
811
812 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
813 if (need_watchdog_p) {
814 dev_queue->trans_start = 0;
815 *need_watchdog_p = 1;
816 }
817}
818
819void dev_activate(struct net_device *dev)
820{
821 int need_watchdog;
822
823 /* No queueing discipline is attached to device;
824 * create default one for devices, which need queueing
825 * and noqueue_qdisc for virtual interfaces
826 */
827
828 if (dev->qdisc == &noop_qdisc)
829 attach_default_qdiscs(dev);
830
831 if (!netif_carrier_ok(dev))
832 /* Delay activation until next carrier-on event */
833 return;
834
835 need_watchdog = 0;
836 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
837 if (dev_ingress_queue(dev))
838 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
839
840 if (need_watchdog) {
841 netif_trans_update(dev);
842 dev_watchdog_up(dev);
843 }
844}
845EXPORT_SYMBOL(dev_activate);
846
847static void dev_deactivate_queue(struct net_device *dev,
848 struct netdev_queue *dev_queue,
849 void *_qdisc_default)
850{
851 struct Qdisc *qdisc_default = _qdisc_default;
852 struct Qdisc *qdisc;
853
854 qdisc = rtnl_dereference(dev_queue->qdisc);
855 if (qdisc) {
856 spin_lock_bh(qdisc_lock(qdisc));
857
858 if (!(qdisc->flags & TCQ_F_BUILTIN))
859 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
860
861 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
862 qdisc_reset(qdisc);
863
864 spin_unlock_bh(qdisc_lock(qdisc));
865 }
866}
867
868static bool some_qdisc_is_busy(struct net_device *dev)
869{
870 unsigned int i;
871
872 for (i = 0; i < dev->num_tx_queues; i++) {
873 struct netdev_queue *dev_queue;
874 spinlock_t *root_lock;
875 struct Qdisc *q;
876 int val;
877
878 dev_queue = netdev_get_tx_queue(dev, i);
879 q = dev_queue->qdisc_sleeping;
880 root_lock = qdisc_lock(q);
881
882 spin_lock_bh(root_lock);
883
884 val = (qdisc_is_running(q) ||
885 test_bit(__QDISC_STATE_SCHED, &q->state));
886
887 spin_unlock_bh(root_lock);
888
889 if (val)
890 return true;
891 }
892 return false;
893}
894
895/**
896 * dev_deactivate_many - deactivate transmissions on several devices
897 * @head: list of devices to deactivate
898 *
899 * This function returns only when all outstanding transmissions
900 * have completed, unless all devices are in dismantle phase.
901 */
902void dev_deactivate_many(struct list_head *head)
903{
904 struct net_device *dev;
905 bool sync_needed = false;
906
907 list_for_each_entry(dev, head, close_list) {
908 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
909 &noop_qdisc);
910 if (dev_ingress_queue(dev))
911 dev_deactivate_queue(dev, dev_ingress_queue(dev),
912 &noop_qdisc);
913
914 dev_watchdog_down(dev);
915 sync_needed |= !dev->dismantle;
916 }
917
918 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
919 * This is avoided if all devices are in dismantle phase :
920 * Caller will call synchronize_net() for us
921 */
922 if (sync_needed)
923 synchronize_net();
924
925 /* Wait for outstanding qdisc_run calls. */
926 list_for_each_entry(dev, head, close_list)
927 while (some_qdisc_is_busy(dev))
928 yield();
929}
930
931void dev_deactivate(struct net_device *dev)
932{
933 LIST_HEAD(single);
934
935 list_add(&dev->close_list, &single);
936 dev_deactivate_many(&single);
937 list_del(&single);
938}
939EXPORT_SYMBOL(dev_deactivate);
940
941static void dev_init_scheduler_queue(struct net_device *dev,
942 struct netdev_queue *dev_queue,
943 void *_qdisc)
944{
945 struct Qdisc *qdisc = _qdisc;
946
947 rcu_assign_pointer(dev_queue->qdisc, qdisc);
948 dev_queue->qdisc_sleeping = qdisc;
949}
950
951void dev_init_scheduler(struct net_device *dev)
952{
953 dev->qdisc = &noop_qdisc;
954 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
955 if (dev_ingress_queue(dev))
956 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
957
958 setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
959}
960
961static void shutdown_scheduler_queue(struct net_device *dev,
962 struct netdev_queue *dev_queue,
963 void *_qdisc_default)
964{
965 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
966 struct Qdisc *qdisc_default = _qdisc_default;
967
968 if (qdisc) {
969 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
970 dev_queue->qdisc_sleeping = qdisc_default;
971
972 qdisc_destroy(qdisc);
973 }
974}
975
976void dev_shutdown(struct net_device *dev)
977{
978 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
979 if (dev_ingress_queue(dev))
980 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
981 qdisc_destroy(dev->qdisc);
982 dev->qdisc = &noop_qdisc;
983
984 WARN_ON(timer_pending(&dev->watchdog_timer));
985}
986
987void psched_ratecfg_precompute(struct psched_ratecfg *r,
988 const struct tc_ratespec *conf,
989 u64 rate64)
990{
991 memset(r, 0, sizeof(*r));
992 r->overhead = conf->overhead;
993 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
994 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
995 r->mult = 1;
996 /*
997 * The deal here is to replace a divide by a reciprocal one
998 * in fast path (a reciprocal divide is a multiply and a shift)
999 *
1000 * Normal formula would be :
1001 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1002 *
1003 * We compute mult/shift to use instead :
1004 * time_in_ns = (len * mult) >> shift;
1005 *
1006 * We try to get the highest possible mult value for accuracy,
1007 * but have to make sure no overflows will ever happen.
1008 */
1009 if (r->rate_bytes_ps > 0) {
1010 u64 factor = NSEC_PER_SEC;
1011
1012 for (;;) {
1013 r->mult = div64_u64(factor, r->rate_bytes_ps);
1014 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1015 break;
1016 factor <<= 1;
1017 r->shift++;
1018 }
1019 }
1020}
1021EXPORT_SYMBOL(psched_ratecfg_precompute);