shrink cor_conn.source.sock
[cor.git] / net / sched / sch_generic.c
blob5ab696efca951313372c3dfda0c2da52e97f4975
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * net/sched/sch_generic.c Generic packet scheduler routines.
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
7 * - Ingress support
8 */
10 #include <linux/bitops.h>
11 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/sched.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/netdevice.h>
18 #include <linux/skbuff.h>
19 #include <linux/rtnetlink.h>
20 #include <linux/init.h>
21 #include <linux/rcupdate.h>
22 #include <linux/list.h>
23 #include <linux/slab.h>
24 #include <linux/if_vlan.h>
25 #include <linux/skb_array.h>
26 #include <linux/if_macvlan.h>
27 #include <net/sch_generic.h>
28 #include <net/pkt_sched.h>
29 #include <net/dst.h>
30 #include <trace/events/qdisc.h>
31 #include <trace/events/net.h>
32 #include <net/xfrm.h>
34 /* Qdisc to use by default */
35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36 EXPORT_SYMBOL(default_qdisc_ops);
38 /* Main transmission queue. */
40 /* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
49 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
51 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53 const struct netdev_queue *txq = q->dev_queue;
54 spinlock_t *lock = NULL;
55 struct sk_buff *skb;
57 if (q->flags & TCQ_F_NOLOCK) {
58 lock = qdisc_lock(q);
59 spin_lock(lock);
62 skb = skb_peek(&q->skb_bad_txq);
63 if (skb) {
64 /* check the reason of requeuing without tx lock first */
65 txq = skb_get_tx_queue(txq->dev, skb);
66 if (!netif_xmit_frozen_or_stopped(txq)) {
67 skb = __skb_dequeue(&q->skb_bad_txq);
68 if (qdisc_is_percpu_stats(q)) {
69 qdisc_qstats_cpu_backlog_dec(q, skb);
70 qdisc_qstats_cpu_qlen_dec(q);
71 } else {
72 qdisc_qstats_backlog_dec(q, skb);
73 q->q.qlen--;
75 } else {
76 skb = SKB_XOFF_MAGIC;
80 if (lock)
81 spin_unlock(lock);
83 return skb;
86 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
90 if (unlikely(skb))
91 skb = __skb_dequeue_bad_txq(q);
93 return skb;
96 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
97 struct sk_buff *skb)
99 spinlock_t *lock = NULL;
101 if (q->flags & TCQ_F_NOLOCK) {
102 lock = qdisc_lock(q);
103 spin_lock(lock);
106 __skb_queue_tail(&q->skb_bad_txq, skb);
108 if (qdisc_is_percpu_stats(q)) {
109 qdisc_qstats_cpu_backlog_inc(q, skb);
110 qdisc_qstats_cpu_qlen_inc(q);
111 } else {
112 qdisc_qstats_backlog_inc(q, skb);
113 q->q.qlen++;
116 if (lock)
117 spin_unlock(lock);
120 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
122 spinlock_t *lock = NULL;
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
126 spin_lock(lock);
129 while (skb) {
130 struct sk_buff *next = skb->next;
132 __skb_queue_tail(&q->gso_skb, skb);
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q)) {
136 qdisc_qstats_cpu_requeues_inc(q);
137 qdisc_qstats_cpu_backlog_inc(q, skb);
138 qdisc_qstats_cpu_qlen_inc(q);
139 } else {
140 q->qstats.requeues++;
141 qdisc_qstats_backlog_inc(q, skb);
142 q->q.qlen++;
145 skb = next;
147 if (lock)
148 spin_unlock(lock);
149 __netif_schedule(q);
152 static void try_bulk_dequeue_skb(struct Qdisc *q,
153 struct sk_buff *skb,
154 const struct netdev_queue *txq,
155 int *packets)
157 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
159 while (bytelimit > 0) {
160 struct sk_buff *nskb = q->dequeue(q);
162 if (!nskb)
163 break;
165 bytelimit -= nskb->len; /* covers GSO len */
166 skb->next = nskb;
167 skb = nskb;
168 (*packets)++; /* GSO counts as one pkt */
170 skb_mark_not_on_list(skb);
173 /* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
176 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
177 struct sk_buff *skb,
178 int *packets)
180 int mapping = skb_get_queue_mapping(skb);
181 struct sk_buff *nskb;
182 int cnt = 0;
184 do {
185 nskb = q->dequeue(q);
186 if (!nskb)
187 break;
188 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
189 qdisc_enqueue_skb_bad_txq(q, nskb);
190 break;
192 skb->next = nskb;
193 skb = nskb;
194 } while (++cnt < 8);
195 (*packets) += cnt;
196 skb_mark_not_on_list(skb);
199 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
202 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
203 int *packets)
205 const struct netdev_queue *txq = q->dev_queue;
206 struct sk_buff *skb = NULL;
208 *packets = 1;
209 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
210 spinlock_t *lock = NULL;
212 if (q->flags & TCQ_F_NOLOCK) {
213 lock = qdisc_lock(q);
214 spin_lock(lock);
217 skb = skb_peek(&q->gso_skb);
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
222 if (!skb) {
223 if (lock)
224 spin_unlock(lock);
225 goto validate;
228 /* skb in gso_skb were already validated */
229 *validate = false;
230 if (xfrm_offload(skb))
231 *validate = true;
232 /* check the reason of requeuing without tx lock first */
233 txq = skb_get_tx_queue(txq->dev, skb);
234 if (!netif_xmit_frozen_or_stopped(txq)) {
235 skb = __skb_dequeue(&q->gso_skb);
236 if (qdisc_is_percpu_stats(q)) {
237 qdisc_qstats_cpu_backlog_dec(q, skb);
238 qdisc_qstats_cpu_qlen_dec(q);
239 } else {
240 qdisc_qstats_backlog_dec(q, skb);
241 q->q.qlen--;
243 } else {
244 skb = NULL;
246 if (lock)
247 spin_unlock(lock);
248 goto trace;
250 validate:
251 *validate = true;
253 if ((q->flags & TCQ_F_ONETXQUEUE) &&
254 netif_xmit_frozen_or_stopped(txq))
255 return skb;
257 skb = qdisc_dequeue_skb_bad_txq(q);
258 if (unlikely(skb)) {
259 if (skb == SKB_XOFF_MAGIC)
260 return NULL;
261 goto bulk;
263 skb = q->dequeue(q);
264 if (skb) {
265 bulk:
266 if (qdisc_may_bulk(q))
267 try_bulk_dequeue_skb(q, skb, txq, packets);
268 else
269 try_bulk_dequeue_skb_slow(q, skb, packets);
271 trace:
272 trace_qdisc_dequeue(q, txq, *packets, skb);
273 return skb;
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
285 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
286 struct net_device *dev, struct netdev_queue *txq,
287 spinlock_t *root_lock, bool validate)
289 int ret = NETDEV_TX_BUSY;
290 bool again = false;
292 /* And release qdisc */
293 if (root_lock)
294 spin_unlock(root_lock);
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
297 if (validate)
298 skb = validate_xmit_skb_list(skb, dev, &again);
300 #ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again)) {
302 if (root_lock)
303 spin_lock(root_lock);
305 dev_requeue_skb(skb, q);
306 return false;
308 #endif
310 if (likely(skb)) {
311 HARD_TX_LOCK(dev, txq, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq))
313 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
315 HARD_TX_UNLOCK(dev, txq);
316 } else {
317 if (root_lock)
318 spin_lock(root_lock);
319 return true;
322 if (root_lock)
323 spin_lock(root_lock);
325 if (!dev_xmit_complete(ret)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret != NETDEV_TX_BUSY))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev->name, ret, q->q.qlen);
331 dev_requeue_skb(skb, q);
332 return false;
335 return true;
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
343 * this queue.
345 * netif_tx_lock serializes accesses to device driver.
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
350 * Note, that this procedure can be called by a watchdog timer
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
357 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
359 spinlock_t *root_lock = NULL;
360 struct netdev_queue *txq;
361 struct net_device *dev;
362 struct sk_buff *skb;
363 bool validate;
365 /* Dequeue packet */
366 skb = dequeue_skb(q, &validate, packets);
367 if (unlikely(!skb))
368 return false;
370 if (!(q->flags & TCQ_F_NOLOCK))
371 root_lock = qdisc_lock(q);
373 dev = qdisc_dev(q);
374 txq = skb_get_tx_queue(dev, skb);
376 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
379 void __qdisc_run(struct Qdisc *q)
381 int quota = dev_tx_weight;
382 int packets;
384 while (qdisc_restart(q, &packets)) {
385 quota -= packets;
386 if (quota <= 0) {
387 __netif_schedule(q);
388 break;
393 unsigned long dev_trans_start(struct net_device *dev)
395 unsigned long val, res;
396 unsigned int i;
398 if (is_vlan_dev(dev))
399 dev = vlan_dev_real_dev(dev);
400 else if (netif_is_macvlan(dev))
401 dev = macvlan_dev_real_dev(dev);
402 res = netdev_get_tx_queue(dev, 0)->trans_start;
403 for (i = 1; i < dev->num_tx_queues; i++) {
404 val = netdev_get_tx_queue(dev, i)->trans_start;
405 if (val && time_after(val, res))
406 res = val;
409 return res;
411 EXPORT_SYMBOL(dev_trans_start);
413 static void dev_watchdog(struct timer_list *t)
415 struct net_device *dev = from_timer(dev, t, watchdog_timer);
417 netif_tx_lock(dev);
418 if (!qdisc_tx_is_noop(dev)) {
419 if (netif_device_present(dev) &&
420 netif_running(dev) &&
421 netif_carrier_ok(dev)) {
422 int some_queue_timedout = 0;
423 unsigned int i;
424 unsigned long trans_start;
426 for (i = 0; i < dev->num_tx_queues; i++) {
427 struct netdev_queue *txq;
429 txq = netdev_get_tx_queue(dev, i);
430 trans_start = txq->trans_start;
431 if (netif_xmit_stopped(txq) &&
432 time_after(jiffies, (trans_start +
433 dev->watchdog_timeo))) {
434 some_queue_timedout = 1;
435 txq->trans_timeout++;
436 break;
440 if (some_queue_timedout) {
441 trace_net_dev_xmit_timeout(dev, i);
442 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
443 dev->name, netdev_drivername(dev), i);
444 dev->netdev_ops->ndo_tx_timeout(dev);
446 if (!mod_timer(&dev->watchdog_timer,
447 round_jiffies(jiffies +
448 dev->watchdog_timeo)))
449 dev_hold(dev);
452 netif_tx_unlock(dev);
454 dev_put(dev);
457 void __netdev_watchdog_up(struct net_device *dev)
459 if (dev->netdev_ops->ndo_tx_timeout) {
460 if (dev->watchdog_timeo <= 0)
461 dev->watchdog_timeo = 5*HZ;
462 if (!mod_timer(&dev->watchdog_timer,
463 round_jiffies(jiffies + dev->watchdog_timeo)))
464 dev_hold(dev);
468 static void dev_watchdog_up(struct net_device *dev)
470 __netdev_watchdog_up(dev);
473 static void dev_watchdog_down(struct net_device *dev)
475 netif_tx_lock_bh(dev);
476 if (del_timer(&dev->watchdog_timer))
477 dev_put(dev);
478 netif_tx_unlock_bh(dev);
482 * netif_carrier_on - set carrier
483 * @dev: network device
485 * Device has detected acquisition of carrier.
487 void netif_carrier_on(struct net_device *dev)
489 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
490 if (dev->reg_state == NETREG_UNINITIALIZED)
491 return;
492 atomic_inc(&dev->carrier_up_count);
493 linkwatch_fire_event(dev);
494 if (netif_running(dev))
495 __netdev_watchdog_up(dev);
498 EXPORT_SYMBOL(netif_carrier_on);
501 * netif_carrier_off - clear carrier
502 * @dev: network device
504 * Device has detected loss of carrier.
506 void netif_carrier_off(struct net_device *dev)
508 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
509 if (dev->reg_state == NETREG_UNINITIALIZED)
510 return;
511 atomic_inc(&dev->carrier_down_count);
512 linkwatch_fire_event(dev);
515 EXPORT_SYMBOL(netif_carrier_off);
517 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
518 under all circumstances. It is difficult to invent anything faster or
519 cheaper.
522 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
523 struct sk_buff **to_free)
525 __qdisc_drop(skb, to_free);
526 return NET_XMIT_CN;
529 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
531 return NULL;
534 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
535 .id = "noop",
536 .priv_size = 0,
537 .enqueue = noop_enqueue,
538 .dequeue = noop_dequeue,
539 .peek = noop_dequeue,
540 .owner = THIS_MODULE,
543 static struct netdev_queue noop_netdev_queue = {
544 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
545 .qdisc_sleeping = &noop_qdisc,
548 struct Qdisc noop_qdisc = {
549 .enqueue = noop_enqueue,
550 .dequeue = noop_dequeue,
551 .flags = TCQ_F_BUILTIN,
552 .ops = &noop_qdisc_ops,
553 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
554 .dev_queue = &noop_netdev_queue,
555 .running = SEQCNT_ZERO(noop_qdisc.running),
556 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
557 .gso_skb = {
558 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
559 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
560 .qlen = 0,
561 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
563 .skb_bad_txq = {
564 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
565 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
566 .qlen = 0,
567 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
570 EXPORT_SYMBOL(noop_qdisc);
572 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
573 struct netlink_ext_ack *extack)
575 /* register_qdisc() assigns a default of noop_enqueue if unset,
576 * but __dev_queue_xmit() treats noqueue only as such
577 * if this is NULL - so clear it here. */
578 qdisc->enqueue = NULL;
579 return 0;
582 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
583 .id = "noqueue",
584 .priv_size = 0,
585 .init = noqueue_init,
586 .enqueue = noop_enqueue,
587 .dequeue = noop_dequeue,
588 .peek = noop_dequeue,
589 .owner = THIS_MODULE,
592 static const u8 prio2band[TC_PRIO_MAX + 1] = {
593 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
596 /* 3-band FIFO queue: old style, but should be a bit faster than
597 generic prio+fifo combination.
600 #define PFIFO_FAST_BANDS 3
603 * Private data for a pfifo_fast scheduler containing:
604 * - rings for priority bands
606 struct pfifo_fast_priv {
607 struct skb_array q[PFIFO_FAST_BANDS];
610 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
611 int band)
613 return &priv->q[band];
616 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
617 struct sk_buff **to_free)
619 int band = prio2band[skb->priority & TC_PRIO_MAX];
620 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
621 struct skb_array *q = band2list(priv, band);
622 unsigned int pkt_len = qdisc_pkt_len(skb);
623 int err;
625 err = skb_array_produce(q, skb);
627 if (unlikely(err)) {
628 if (qdisc_is_percpu_stats(qdisc))
629 return qdisc_drop_cpu(skb, qdisc, to_free);
630 else
631 return qdisc_drop(skb, qdisc, to_free);
634 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
635 return NET_XMIT_SUCCESS;
638 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
640 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
641 struct sk_buff *skb = NULL;
642 int band;
644 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
645 struct skb_array *q = band2list(priv, band);
647 if (__skb_array_empty(q))
648 continue;
650 skb = __skb_array_consume(q);
652 if (likely(skb)) {
653 qdisc_update_stats_at_dequeue(qdisc, skb);
654 } else {
655 WRITE_ONCE(qdisc->empty, true);
658 return skb;
661 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
663 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
664 struct sk_buff *skb = NULL;
665 int band;
667 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
668 struct skb_array *q = band2list(priv, band);
670 skb = __skb_array_peek(q);
673 return skb;
676 static void pfifo_fast_reset(struct Qdisc *qdisc)
678 int i, band;
679 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
681 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
682 struct skb_array *q = band2list(priv, band);
683 struct sk_buff *skb;
685 /* NULL ring is possible if destroy path is due to a failed
686 * skb_array_init() in pfifo_fast_init() case.
688 if (!q->ring.queue)
689 continue;
691 while ((skb = __skb_array_consume(q)) != NULL)
692 kfree_skb(skb);
695 if (qdisc_is_percpu_stats(qdisc)) {
696 for_each_possible_cpu(i) {
697 struct gnet_stats_queue *q;
699 q = per_cpu_ptr(qdisc->cpu_qstats, i);
700 q->backlog = 0;
701 q->qlen = 0;
706 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
708 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
710 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
711 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
712 goto nla_put_failure;
713 return skb->len;
715 nla_put_failure:
716 return -1;
719 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
720 struct netlink_ext_ack *extack)
722 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
723 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
724 int prio;
726 /* guard against zero length rings */
727 if (!qlen)
728 return -EINVAL;
730 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
731 struct skb_array *q = band2list(priv, prio);
732 int err;
734 err = skb_array_init(q, qlen, GFP_KERNEL);
735 if (err)
736 return -ENOMEM;
739 /* Can by-pass the queue discipline */
740 qdisc->flags |= TCQ_F_CAN_BYPASS;
741 return 0;
744 static void pfifo_fast_destroy(struct Qdisc *sch)
746 struct pfifo_fast_priv *priv = qdisc_priv(sch);
747 int prio;
749 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
750 struct skb_array *q = band2list(priv, prio);
752 /* NULL ring is possible if destroy path is due to a failed
753 * skb_array_init() in pfifo_fast_init() case.
755 if (!q->ring.queue)
756 continue;
757 /* Destroy ring but no need to kfree_skb because a call to
758 * pfifo_fast_reset() has already done that work.
760 ptr_ring_cleanup(&q->ring, NULL);
764 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
765 unsigned int new_len)
767 struct pfifo_fast_priv *priv = qdisc_priv(sch);
768 struct skb_array *bands[PFIFO_FAST_BANDS];
769 int prio;
771 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
772 struct skb_array *q = band2list(priv, prio);
774 bands[prio] = q;
777 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
778 GFP_KERNEL);
781 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
782 .id = "pfifo_fast",
783 .priv_size = sizeof(struct pfifo_fast_priv),
784 .enqueue = pfifo_fast_enqueue,
785 .dequeue = pfifo_fast_dequeue,
786 .peek = pfifo_fast_peek,
787 .init = pfifo_fast_init,
788 .destroy = pfifo_fast_destroy,
789 .reset = pfifo_fast_reset,
790 .dump = pfifo_fast_dump,
791 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
792 .owner = THIS_MODULE,
793 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
795 EXPORT_SYMBOL(pfifo_fast_ops);
797 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
798 const struct Qdisc_ops *ops,
799 struct netlink_ext_ack *extack)
801 void *p;
802 struct Qdisc *sch;
803 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
804 int err = -ENOBUFS;
805 struct net_device *dev;
807 if (!dev_queue) {
808 NL_SET_ERR_MSG(extack, "No device queue given");
809 err = -EINVAL;
810 goto errout;
813 dev = dev_queue->dev;
814 p = kzalloc_node(size, GFP_KERNEL,
815 netdev_queue_numa_node_read(dev_queue));
817 if (!p)
818 goto errout;
819 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
820 /* if we got non aligned memory, ask more and do alignment ourself */
821 if (sch != p) {
822 kfree(p);
823 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
824 netdev_queue_numa_node_read(dev_queue));
825 if (!p)
826 goto errout;
827 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
828 sch->padded = (char *) sch - (char *) p;
830 __skb_queue_head_init(&sch->gso_skb);
831 __skb_queue_head_init(&sch->skb_bad_txq);
832 qdisc_skb_head_init(&sch->q);
833 spin_lock_init(&sch->q.lock);
835 if (ops->static_flags & TCQ_F_CPUSTATS) {
836 sch->cpu_bstats =
837 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
838 if (!sch->cpu_bstats)
839 goto errout1;
841 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
842 if (!sch->cpu_qstats) {
843 free_percpu(sch->cpu_bstats);
844 goto errout1;
848 spin_lock_init(&sch->busylock);
849 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
850 spin_lock_init(&sch->seqlock);
851 seqcount_init(&sch->running);
853 sch->ops = ops;
854 sch->flags = ops->static_flags;
855 sch->enqueue = ops->enqueue;
856 sch->dequeue = ops->dequeue;
857 sch->dev_queue = dev_queue;
858 sch->empty = true;
859 dev_hold(dev);
860 refcount_set(&sch->refcnt, 1);
862 if (sch != &noop_qdisc) {
863 lockdep_set_class(&sch->busylock, &dev->qdisc_tx_busylock_key);
864 lockdep_set_class(&sch->seqlock, &dev->qdisc_tx_busylock_key);
865 lockdep_set_class(&sch->running, &dev->qdisc_running_key);
868 return sch;
869 errout1:
870 kfree(p);
871 errout:
872 return ERR_PTR(err);
875 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
876 const struct Qdisc_ops *ops,
877 unsigned int parentid,
878 struct netlink_ext_ack *extack)
880 struct Qdisc *sch;
882 if (!try_module_get(ops->owner)) {
883 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
884 return NULL;
887 sch = qdisc_alloc(dev_queue, ops, extack);
888 if (IS_ERR(sch)) {
889 module_put(ops->owner);
890 return NULL;
892 sch->parent = parentid;
894 if (!ops->init || ops->init(sch, NULL, extack) == 0)
895 return sch;
897 qdisc_put(sch);
898 return NULL;
900 EXPORT_SYMBOL(qdisc_create_dflt);
902 /* Under qdisc_lock(qdisc) and BH! */
904 void qdisc_reset(struct Qdisc *qdisc)
906 const struct Qdisc_ops *ops = qdisc->ops;
907 struct sk_buff *skb, *tmp;
909 if (ops->reset)
910 ops->reset(qdisc);
912 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
913 __skb_unlink(skb, &qdisc->gso_skb);
914 kfree_skb_list(skb);
917 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
918 __skb_unlink(skb, &qdisc->skb_bad_txq);
919 kfree_skb_list(skb);
922 qdisc->q.qlen = 0;
923 qdisc->qstats.backlog = 0;
925 EXPORT_SYMBOL(qdisc_reset);
927 void qdisc_free(struct Qdisc *qdisc)
929 if (qdisc_is_percpu_stats(qdisc)) {
930 free_percpu(qdisc->cpu_bstats);
931 free_percpu(qdisc->cpu_qstats);
934 kfree((char *) qdisc - qdisc->padded);
937 static void qdisc_free_cb(struct rcu_head *head)
939 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
941 qdisc_free(q);
944 static void qdisc_destroy(struct Qdisc *qdisc)
946 const struct Qdisc_ops *ops = qdisc->ops;
947 struct sk_buff *skb, *tmp;
949 #ifdef CONFIG_NET_SCHED
950 qdisc_hash_del(qdisc);
952 qdisc_put_stab(rtnl_dereference(qdisc->stab));
953 #endif
954 gen_kill_estimator(&qdisc->rate_est);
955 if (ops->reset)
956 ops->reset(qdisc);
957 if (ops->destroy)
958 ops->destroy(qdisc);
960 module_put(ops->owner);
961 dev_put(qdisc_dev(qdisc));
963 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
964 __skb_unlink(skb, &qdisc->gso_skb);
965 kfree_skb_list(skb);
968 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
969 __skb_unlink(skb, &qdisc->skb_bad_txq);
970 kfree_skb_list(skb);
973 call_rcu(&qdisc->rcu, qdisc_free_cb);
976 void qdisc_put(struct Qdisc *qdisc)
978 if (!qdisc)
979 return;
981 if (qdisc->flags & TCQ_F_BUILTIN ||
982 !refcount_dec_and_test(&qdisc->refcnt))
983 return;
985 qdisc_destroy(qdisc);
987 EXPORT_SYMBOL(qdisc_put);
989 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
990 * Intended to be used as optimization, this function only takes rtnl lock if
991 * qdisc reference counter reached zero.
994 void qdisc_put_unlocked(struct Qdisc *qdisc)
996 if (qdisc->flags & TCQ_F_BUILTIN ||
997 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
998 return;
1000 qdisc_destroy(qdisc);
1001 rtnl_unlock();
1003 EXPORT_SYMBOL(qdisc_put_unlocked);
1005 /* Attach toplevel qdisc to device queue. */
1006 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1007 struct Qdisc *qdisc)
1009 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1010 spinlock_t *root_lock;
1012 root_lock = qdisc_lock(oqdisc);
1013 spin_lock_bh(root_lock);
1015 /* ... and graft new one */
1016 if (qdisc == NULL)
1017 qdisc = &noop_qdisc;
1018 dev_queue->qdisc_sleeping = qdisc;
1019 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1021 spin_unlock_bh(root_lock);
1023 return oqdisc;
1025 EXPORT_SYMBOL(dev_graft_qdisc);
1027 static void attach_one_default_qdisc(struct net_device *dev,
1028 struct netdev_queue *dev_queue,
1029 void *_unused)
1031 struct Qdisc *qdisc;
1032 const struct Qdisc_ops *ops = default_qdisc_ops;
1034 if (dev->priv_flags & IFF_NO_QUEUE)
1035 ops = &noqueue_qdisc_ops;
1036 else if(dev->type == ARPHRD_CAN)
1037 ops = &pfifo_fast_ops;
1039 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1040 if (!qdisc) {
1041 netdev_info(dev, "activation failed\n");
1042 return;
1044 if (!netif_is_multiqueue(dev))
1045 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1046 dev_queue->qdisc_sleeping = qdisc;
1049 static void attach_default_qdiscs(struct net_device *dev)
1051 struct netdev_queue *txq;
1052 struct Qdisc *qdisc;
1054 txq = netdev_get_tx_queue(dev, 0);
1056 if (!netif_is_multiqueue(dev) ||
1057 dev->priv_flags & IFF_NO_QUEUE) {
1058 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1059 dev->qdisc = txq->qdisc_sleeping;
1060 qdisc_refcount_inc(dev->qdisc);
1061 } else {
1062 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1063 if (qdisc) {
1064 dev->qdisc = qdisc;
1065 qdisc->ops->attach(qdisc);
1068 #ifdef CONFIG_NET_SCHED
1069 if (dev->qdisc != &noop_qdisc)
1070 qdisc_hash_add(dev->qdisc, false);
1071 #endif
1074 static void transition_one_qdisc(struct net_device *dev,
1075 struct netdev_queue *dev_queue,
1076 void *_need_watchdog)
1078 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1079 int *need_watchdog_p = _need_watchdog;
1081 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1082 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1084 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1085 if (need_watchdog_p) {
1086 dev_queue->trans_start = 0;
1087 *need_watchdog_p = 1;
1091 void dev_activate(struct net_device *dev)
1093 int need_watchdog;
1095 /* No queueing discipline is attached to device;
1096 * create default one for devices, which need queueing
1097 * and noqueue_qdisc for virtual interfaces
1100 if (dev->qdisc == &noop_qdisc)
1101 attach_default_qdiscs(dev);
1103 if (!netif_carrier_ok(dev))
1104 /* Delay activation until next carrier-on event */
1105 return;
1107 need_watchdog = 0;
1108 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1109 if (dev_ingress_queue(dev))
1110 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1112 if (need_watchdog) {
1113 netif_trans_update(dev);
1114 dev_watchdog_up(dev);
1117 EXPORT_SYMBOL(dev_activate);
1119 static void dev_deactivate_queue(struct net_device *dev,
1120 struct netdev_queue *dev_queue,
1121 void *_qdisc_default)
1123 struct Qdisc *qdisc_default = _qdisc_default;
1124 struct Qdisc *qdisc;
1126 qdisc = rtnl_dereference(dev_queue->qdisc);
1127 if (qdisc) {
1128 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1130 if (nolock)
1131 spin_lock_bh(&qdisc->seqlock);
1132 spin_lock_bh(qdisc_lock(qdisc));
1134 if (!(qdisc->flags & TCQ_F_BUILTIN))
1135 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1137 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1138 qdisc_reset(qdisc);
1140 spin_unlock_bh(qdisc_lock(qdisc));
1141 if (nolock)
1142 spin_unlock_bh(&qdisc->seqlock);
1146 static bool some_qdisc_is_busy(struct net_device *dev)
1148 unsigned int i;
1150 for (i = 0; i < dev->num_tx_queues; i++) {
1151 struct netdev_queue *dev_queue;
1152 spinlock_t *root_lock;
1153 struct Qdisc *q;
1154 int val;
1156 dev_queue = netdev_get_tx_queue(dev, i);
1157 q = dev_queue->qdisc_sleeping;
1159 root_lock = qdisc_lock(q);
1160 spin_lock_bh(root_lock);
1162 val = (qdisc_is_running(q) ||
1163 test_bit(__QDISC_STATE_SCHED, &q->state));
1165 spin_unlock_bh(root_lock);
1167 if (val)
1168 return true;
1170 return false;
1173 static void dev_qdisc_reset(struct net_device *dev,
1174 struct netdev_queue *dev_queue,
1175 void *none)
1177 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1179 if (qdisc)
1180 qdisc_reset(qdisc);
1184 * dev_deactivate_many - deactivate transmissions on several devices
1185 * @head: list of devices to deactivate
1187 * This function returns only when all outstanding transmissions
1188 * have completed, unless all devices are in dismantle phase.
1190 void dev_deactivate_many(struct list_head *head)
1192 struct net_device *dev;
1194 list_for_each_entry(dev, head, close_list) {
1195 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1196 &noop_qdisc);
1197 if (dev_ingress_queue(dev))
1198 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1199 &noop_qdisc);
1201 dev_watchdog_down(dev);
1204 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1205 * This is avoided if all devices are in dismantle phase :
1206 * Caller will call synchronize_net() for us
1208 synchronize_net();
1210 /* Wait for outstanding qdisc_run calls. */
1211 list_for_each_entry(dev, head, close_list) {
1212 while (some_qdisc_is_busy(dev)) {
1213 /* wait_event() would avoid this sleep-loop but would
1214 * require expensive checks in the fast paths of packet
1215 * processing which isn't worth it.
1217 schedule_timeout_uninterruptible(1);
1219 /* The new qdisc is assigned at this point so we can safely
1220 * unwind stale skb lists and qdisc statistics
1222 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1223 if (dev_ingress_queue(dev))
1224 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1228 void dev_deactivate(struct net_device *dev)
1230 LIST_HEAD(single);
1232 list_add(&dev->close_list, &single);
1233 dev_deactivate_many(&single);
1234 list_del(&single);
1236 EXPORT_SYMBOL(dev_deactivate);
1238 static int qdisc_change_tx_queue_len(struct net_device *dev,
1239 struct netdev_queue *dev_queue)
1241 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1242 const struct Qdisc_ops *ops = qdisc->ops;
1244 if (ops->change_tx_queue_len)
1245 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1246 return 0;
1249 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1251 bool up = dev->flags & IFF_UP;
1252 unsigned int i;
1253 int ret = 0;
1255 if (up)
1256 dev_deactivate(dev);
1258 for (i = 0; i < dev->num_tx_queues; i++) {
1259 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1261 /* TODO: revert changes on a partial failure */
1262 if (ret)
1263 break;
1266 if (up)
1267 dev_activate(dev);
1268 return ret;
1271 static void dev_init_scheduler_queue(struct net_device *dev,
1272 struct netdev_queue *dev_queue,
1273 void *_qdisc)
1275 struct Qdisc *qdisc = _qdisc;
1277 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1278 dev_queue->qdisc_sleeping = qdisc;
1281 void dev_init_scheduler(struct net_device *dev)
1283 dev->qdisc = &noop_qdisc;
1284 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1288 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1291 static void shutdown_scheduler_queue(struct net_device *dev,
1292 struct netdev_queue *dev_queue,
1293 void *_qdisc_default)
1295 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1296 struct Qdisc *qdisc_default = _qdisc_default;
1298 if (qdisc) {
1299 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1300 dev_queue->qdisc_sleeping = qdisc_default;
1302 qdisc_put(qdisc);
1306 void dev_shutdown(struct net_device *dev)
1308 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1309 if (dev_ingress_queue(dev))
1310 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311 qdisc_put(dev->qdisc);
1312 dev->qdisc = &noop_qdisc;
1314 WARN_ON(timer_pending(&dev->watchdog_timer));
1317 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1318 const struct tc_ratespec *conf,
1319 u64 rate64)
1321 memset(r, 0, sizeof(*r));
1322 r->overhead = conf->overhead;
1323 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1324 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1325 r->mult = 1;
1327 * The deal here is to replace a divide by a reciprocal one
1328 * in fast path (a reciprocal divide is a multiply and a shift)
1330 * Normal formula would be :
1331 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1333 * We compute mult/shift to use instead :
1334 * time_in_ns = (len * mult) >> shift;
1336 * We try to get the highest possible mult value for accuracy,
1337 * but have to make sure no overflows will ever happen.
1339 if (r->rate_bytes_ps > 0) {
1340 u64 factor = NSEC_PER_SEC;
1342 for (;;) {
1343 r->mult = div64_u64(factor, r->rate_bytes_ps);
1344 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1345 break;
1346 factor <<= 1;
1347 r->shift++;
1351 EXPORT_SYMBOL(psched_ratecfg_precompute);
1353 static void mini_qdisc_rcu_func(struct rcu_head *head)
1357 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1358 struct tcf_proto *tp_head)
1360 /* Protected with chain0->filter_chain_lock.
1361 * Can't access chain directly because tp_head can be NULL.
1363 struct mini_Qdisc *miniq_old =
1364 rcu_dereference_protected(*miniqp->p_miniq, 1);
1365 struct mini_Qdisc *miniq;
1367 if (!tp_head) {
1368 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1369 /* Wait for flying RCU callback before it is freed. */
1370 rcu_barrier();
1371 return;
1374 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1375 &miniqp->miniq1 : &miniqp->miniq2;
1377 /* We need to make sure that readers won't see the miniq
1378 * we are about to modify. So wait until previous call_rcu callback
1379 * is done.
1381 rcu_barrier();
1382 miniq->filter_list = tp_head;
1383 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1385 if (miniq_old)
1386 /* This is counterpart of the rcu barriers above. We need to
1387 * block potential new user of miniq_old until all readers
1388 * are not seeing it.
1390 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1392 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1394 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1395 struct mini_Qdisc __rcu **p_miniq)
1397 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1398 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1399 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1400 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1401 miniqp->p_miniq = p_miniq;
1403 EXPORT_SYMBOL(mini_qdisc_pair_init);