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 */
36 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
37 EXPORT_SYMBOL(default_qdisc_ops);
38
qdisc_maybe_clear_missed(struct Qdisc * q,const struct netdev_queue * txq)39 static 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
__skb_dequeue_bad_txq(struct Qdisc * q)73 static 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
qdisc_dequeue_skb_bad_txq(struct Qdisc * q)109 static 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
qdisc_enqueue_skb_bad_txq(struct Qdisc * q,struct sk_buff * skb)119 static 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
dev_requeue_skb(struct sk_buff * skb,struct Qdisc * q)143 static 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
try_bulk_dequeue_skb(struct Qdisc * q,struct sk_buff * skb,const struct netdev_queue * txq,int * packets)179 static 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 */
try_bulk_dequeue_skb_slow(struct Qdisc * q,struct sk_buff * skb,int * packets)203 static 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 */
dequeue_skb(struct Qdisc * q,bool * validate,int * packets)229 static 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 }
278 validate:
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) {
295 bulk:
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 }
301 trace:
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 */
sch_direct_xmit(struct sk_buff * skb,struct Qdisc * q,struct net_device * dev,struct netdev_queue * txq,spinlock_t * root_lock,bool validate)315 bool 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 */
qdisc_restart(struct Qdisc * q,int * packets)389 static 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
__qdisc_run(struct Qdisc * q)411 void __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
dev_trans_start(struct net_device * dev)429 unsigned 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 }
443 EXPORT_SYMBOL(dev_trans_start);
444
netif_freeze_queues(struct net_device * dev)445 static 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
netif_tx_lock(struct net_device * dev)466 void netif_tx_lock(struct net_device *dev)
467 {
468 spin_lock(&dev->tx_global_lock);
469 netif_freeze_queues(dev);
470 }
471 EXPORT_SYMBOL(netif_tx_lock);
472
netif_unfreeze_queues(struct net_device * dev)473 static 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
netif_tx_unlock(struct net_device * dev)489 void netif_tx_unlock(struct net_device *dev)
490 {
491 netif_unfreeze_queues(dev);
492 spin_unlock(&dev->tx_global_lock);
493 }
494 EXPORT_SYMBOL(netif_tx_unlock);
495
dev_watchdog(struct timer_list * t)496 static 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 trans_start = READ_ONCE(txq->trans_start);
516 if (!netif_xmit_stopped(txq))
517 continue;
518 if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
519 timedout_ms = jiffies_to_msecs(jiffies - trans_start);
520 atomic_long_inc(&txq->trans_timeout);
521 break;
522 }
523 if (time_after(oldest_start, trans_start))
524 oldest_start = trans_start;
525 }
526
527 if (unlikely(timedout_ms)) {
528 trace_net_dev_xmit_timeout(dev, i);
529 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
530 raw_smp_processor_id(),
531 i, timedout_ms);
532 netif_freeze_queues(dev);
533 dev->netdev_ops->ndo_tx_timeout(dev, i);
534 netif_unfreeze_queues(dev);
535 }
536 if (!mod_timer(&dev->watchdog_timer,
537 round_jiffies(oldest_start +
538 dev->watchdog_timeo)))
539 release = false;
540 }
541 }
542 spin_unlock(&dev->tx_global_lock);
543
544 if (release)
545 netdev_put(dev, &dev->watchdog_dev_tracker);
546 }
547
__netdev_watchdog_up(struct net_device * dev)548 void __netdev_watchdog_up(struct net_device *dev)
549 {
550 if (dev->netdev_ops->ndo_tx_timeout) {
551 if (dev->watchdog_timeo <= 0)
552 dev->watchdog_timeo = 5*HZ;
553 if (!mod_timer(&dev->watchdog_timer,
554 round_jiffies(jiffies + dev->watchdog_timeo)))
555 netdev_hold(dev, &dev->watchdog_dev_tracker,
556 GFP_ATOMIC);
557 }
558 }
559 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
560
dev_watchdog_up(struct net_device * dev)561 static void dev_watchdog_up(struct net_device *dev)
562 {
563 __netdev_watchdog_up(dev);
564 }
565
dev_watchdog_down(struct net_device * dev)566 static void dev_watchdog_down(struct net_device *dev)
567 {
568 netif_tx_lock_bh(dev);
569 if (del_timer(&dev->watchdog_timer))
570 netdev_put(dev, &dev->watchdog_dev_tracker);
571 netif_tx_unlock_bh(dev);
572 }
573
574 /**
575 * netif_carrier_on - set carrier
576 * @dev: network device
577 *
578 * Device has detected acquisition of carrier.
579 */
netif_carrier_on(struct net_device * dev)580 void netif_carrier_on(struct net_device *dev)
581 {
582 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
583 if (dev->reg_state == NETREG_UNINITIALIZED)
584 return;
585 atomic_inc(&dev->carrier_up_count);
586 linkwatch_fire_event(dev);
587 if (netif_running(dev))
588 __netdev_watchdog_up(dev);
589 }
590 }
591 EXPORT_SYMBOL(netif_carrier_on);
592
593 /**
594 * netif_carrier_off - clear carrier
595 * @dev: network device
596 *
597 * Device has detected loss of carrier.
598 */
netif_carrier_off(struct net_device * dev)599 void netif_carrier_off(struct net_device *dev)
600 {
601 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
602 if (dev->reg_state == NETREG_UNINITIALIZED)
603 return;
604 atomic_inc(&dev->carrier_down_count);
605 linkwatch_fire_event(dev);
606 }
607 }
608 EXPORT_SYMBOL(netif_carrier_off);
609
610 /**
611 * netif_carrier_event - report carrier state event
612 * @dev: network device
613 *
614 * Device has detected a carrier event but the carrier state wasn't changed.
615 * Use in drivers when querying carrier state asynchronously, to avoid missing
616 * events (link flaps) if link recovers before it's queried.
617 */
netif_carrier_event(struct net_device * dev)618 void netif_carrier_event(struct net_device *dev)
619 {
620 if (dev->reg_state == NETREG_UNINITIALIZED)
621 return;
622 atomic_inc(&dev->carrier_up_count);
623 atomic_inc(&dev->carrier_down_count);
624 linkwatch_fire_event(dev);
625 }
626 EXPORT_SYMBOL_GPL(netif_carrier_event);
627
628 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
629 under all circumstances. It is difficult to invent anything faster or
630 cheaper.
631 */
632
noop_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)633 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
634 struct sk_buff **to_free)
635 {
636 __qdisc_drop(skb, to_free);
637 return NET_XMIT_CN;
638 }
639
noop_dequeue(struct Qdisc * qdisc)640 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
641 {
642 return NULL;
643 }
644
645 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
646 .id = "noop",
647 .priv_size = 0,
648 .enqueue = noop_enqueue,
649 .dequeue = noop_dequeue,
650 .peek = noop_dequeue,
651 .owner = THIS_MODULE,
652 };
653
654 static struct netdev_queue noop_netdev_queue = {
655 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
656 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
657 };
658
659 struct Qdisc noop_qdisc = {
660 .enqueue = noop_enqueue,
661 .dequeue = noop_dequeue,
662 .flags = TCQ_F_BUILTIN,
663 .ops = &noop_qdisc_ops,
664 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
665 .dev_queue = &noop_netdev_queue,
666 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
667 .gso_skb = {
668 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
669 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
670 .qlen = 0,
671 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
672 },
673 .skb_bad_txq = {
674 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
675 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
676 .qlen = 0,
677 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
678 },
679 .owner = -1,
680 };
681 EXPORT_SYMBOL(noop_qdisc);
682
noqueue_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)683 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
684 struct netlink_ext_ack *extack)
685 {
686 /* register_qdisc() assigns a default of noop_enqueue if unset,
687 * but __dev_queue_xmit() treats noqueue only as such
688 * if this is NULL - so clear it here. */
689 qdisc->enqueue = NULL;
690 return 0;
691 }
692
693 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
694 .id = "noqueue",
695 .priv_size = 0,
696 .init = noqueue_init,
697 .enqueue = noop_enqueue,
698 .dequeue = noop_dequeue,
699 .peek = noop_dequeue,
700 .owner = THIS_MODULE,
701 };
702
703 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
704 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
705 };
706 EXPORT_SYMBOL(sch_default_prio2band);
707
708 /* 3-band FIFO queue: old style, but should be a bit faster than
709 generic prio+fifo combination.
710 */
711
712 #define PFIFO_FAST_BANDS 3
713
714 /*
715 * Private data for a pfifo_fast scheduler containing:
716 * - rings for priority bands
717 */
718 struct pfifo_fast_priv {
719 struct skb_array q[PFIFO_FAST_BANDS];
720 };
721
band2list(struct pfifo_fast_priv * priv,int band)722 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
723 int band)
724 {
725 return &priv->q[band];
726 }
727
pfifo_fast_enqueue(struct sk_buff * skb,struct Qdisc * qdisc,struct sk_buff ** to_free)728 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
729 struct sk_buff **to_free)
730 {
731 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
732 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
733 struct skb_array *q = band2list(priv, band);
734 unsigned int pkt_len = qdisc_pkt_len(skb);
735 int err;
736
737 err = skb_array_produce(q, skb);
738
739 if (unlikely(err)) {
740 if (qdisc_is_percpu_stats(qdisc))
741 return qdisc_drop_cpu(skb, qdisc, to_free);
742 else
743 return qdisc_drop(skb, qdisc, to_free);
744 }
745
746 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
747 return NET_XMIT_SUCCESS;
748 }
749
pfifo_fast_dequeue(struct Qdisc * qdisc)750 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
751 {
752 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
753 struct sk_buff *skb = NULL;
754 bool need_retry = true;
755 int band;
756
757 retry:
758 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
759 struct skb_array *q = band2list(priv, band);
760
761 if (__skb_array_empty(q))
762 continue;
763
764 skb = __skb_array_consume(q);
765 }
766 if (likely(skb)) {
767 qdisc_update_stats_at_dequeue(qdisc, skb);
768 } else if (need_retry &&
769 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
770 /* Delay clearing the STATE_MISSED here to reduce
771 * the overhead of the second spin_trylock() in
772 * qdisc_run_begin() and __netif_schedule() calling
773 * in qdisc_run_end().
774 */
775 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
776 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
777
778 /* Make sure dequeuing happens after clearing
779 * STATE_MISSED.
780 */
781 smp_mb__after_atomic();
782
783 need_retry = false;
784
785 goto retry;
786 }
787
788 return skb;
789 }
790
pfifo_fast_peek(struct Qdisc * qdisc)791 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
792 {
793 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
794 struct sk_buff *skb = NULL;
795 int band;
796
797 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
798 struct skb_array *q = band2list(priv, band);
799
800 skb = __skb_array_peek(q);
801 }
802
803 return skb;
804 }
805
pfifo_fast_reset(struct Qdisc * qdisc)806 static void pfifo_fast_reset(struct Qdisc *qdisc)
807 {
808 int i, band;
809 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
810
811 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
812 struct skb_array *q = band2list(priv, band);
813 struct sk_buff *skb;
814
815 /* NULL ring is possible if destroy path is due to a failed
816 * skb_array_init() in pfifo_fast_init() case.
817 */
818 if (!q->ring.queue)
819 continue;
820
821 while ((skb = __skb_array_consume(q)) != NULL)
822 kfree_skb(skb);
823 }
824
825 if (qdisc_is_percpu_stats(qdisc)) {
826 for_each_possible_cpu(i) {
827 struct gnet_stats_queue *q;
828
829 q = per_cpu_ptr(qdisc->cpu_qstats, i);
830 q->backlog = 0;
831 q->qlen = 0;
832 }
833 }
834 }
835
pfifo_fast_dump(struct Qdisc * qdisc,struct sk_buff * skb)836 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
837 {
838 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
839
840 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
841 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
842 goto nla_put_failure;
843 return skb->len;
844
845 nla_put_failure:
846 return -1;
847 }
848
pfifo_fast_init(struct Qdisc * qdisc,struct nlattr * opt,struct netlink_ext_ack * extack)849 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
850 struct netlink_ext_ack *extack)
851 {
852 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
853 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
854 int prio;
855
856 /* guard against zero length rings */
857 if (!qlen)
858 return -EINVAL;
859
860 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
861 struct skb_array *q = band2list(priv, prio);
862 int err;
863
864 err = skb_array_init(q, qlen, GFP_KERNEL);
865 if (err)
866 return -ENOMEM;
867 }
868
869 /* Can by-pass the queue discipline */
870 qdisc->flags |= TCQ_F_CAN_BYPASS;
871 return 0;
872 }
873
pfifo_fast_destroy(struct Qdisc * sch)874 static void pfifo_fast_destroy(struct Qdisc *sch)
875 {
876 struct pfifo_fast_priv *priv = qdisc_priv(sch);
877 int prio;
878
879 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
880 struct skb_array *q = band2list(priv, prio);
881
882 /* NULL ring is possible if destroy path is due to a failed
883 * skb_array_init() in pfifo_fast_init() case.
884 */
885 if (!q->ring.queue)
886 continue;
887 /* Destroy ring but no need to kfree_skb because a call to
888 * pfifo_fast_reset() has already done that work.
889 */
890 ptr_ring_cleanup(&q->ring, NULL);
891 }
892 }
893
pfifo_fast_change_tx_queue_len(struct Qdisc * sch,unsigned int new_len)894 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
895 unsigned int new_len)
896 {
897 struct pfifo_fast_priv *priv = qdisc_priv(sch);
898 struct skb_array *bands[PFIFO_FAST_BANDS];
899 int prio;
900
901 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
902 struct skb_array *q = band2list(priv, prio);
903
904 bands[prio] = q;
905 }
906
907 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
908 GFP_KERNEL);
909 }
910
911 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
912 .id = "pfifo_fast",
913 .priv_size = sizeof(struct pfifo_fast_priv),
914 .enqueue = pfifo_fast_enqueue,
915 .dequeue = pfifo_fast_dequeue,
916 .peek = pfifo_fast_peek,
917 .init = pfifo_fast_init,
918 .destroy = pfifo_fast_destroy,
919 .reset = pfifo_fast_reset,
920 .dump = pfifo_fast_dump,
921 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
922 .owner = THIS_MODULE,
923 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
924 };
925 EXPORT_SYMBOL(pfifo_fast_ops);
926
927 static struct lock_class_key qdisc_tx_busylock;
928
qdisc_alloc(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,struct netlink_ext_ack * extack)929 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
930 const struct Qdisc_ops *ops,
931 struct netlink_ext_ack *extack)
932 {
933 struct Qdisc *sch;
934 unsigned int size = sizeof(*sch) + ops->priv_size;
935 int err = -ENOBUFS;
936 struct net_device *dev;
937
938 if (!dev_queue) {
939 NL_SET_ERR_MSG(extack, "No device queue given");
940 err = -EINVAL;
941 goto errout;
942 }
943
944 dev = dev_queue->dev;
945 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
946
947 if (!sch)
948 goto errout;
949 __skb_queue_head_init(&sch->gso_skb);
950 __skb_queue_head_init(&sch->skb_bad_txq);
951 gnet_stats_basic_sync_init(&sch->bstats);
952 lockdep_register_key(&sch->root_lock_key);
953 spin_lock_init(&sch->q.lock);
954 lockdep_set_class(&sch->q.lock, &sch->root_lock_key);
955
956 if (ops->static_flags & TCQ_F_CPUSTATS) {
957 sch->cpu_bstats =
958 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
959 if (!sch->cpu_bstats)
960 goto errout1;
961
962 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
963 if (!sch->cpu_qstats) {
964 free_percpu(sch->cpu_bstats);
965 goto errout1;
966 }
967 }
968
969 spin_lock_init(&sch->busylock);
970 lockdep_set_class(&sch->busylock,
971 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
972
973 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
974 spin_lock_init(&sch->seqlock);
975 lockdep_set_class(&sch->seqlock,
976 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
977
978 sch->ops = ops;
979 sch->flags = ops->static_flags;
980 sch->enqueue = ops->enqueue;
981 sch->dequeue = ops->dequeue;
982 sch->dev_queue = dev_queue;
983 sch->owner = -1;
984 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
985 refcount_set(&sch->refcnt, 1);
986
987 return sch;
988 errout1:
989 lockdep_unregister_key(&sch->root_lock_key);
990 kfree(sch);
991 errout:
992 return ERR_PTR(err);
993 }
994
qdisc_create_dflt(struct netdev_queue * dev_queue,const struct Qdisc_ops * ops,unsigned int parentid,struct netlink_ext_ack * extack)995 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
996 const struct Qdisc_ops *ops,
997 unsigned int parentid,
998 struct netlink_ext_ack *extack)
999 {
1000 struct Qdisc *sch;
1001
1002 if (!try_module_get(ops->owner)) {
1003 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1004 return NULL;
1005 }
1006
1007 sch = qdisc_alloc(dev_queue, ops, extack);
1008 if (IS_ERR(sch)) {
1009 module_put(ops->owner);
1010 return NULL;
1011 }
1012 sch->parent = parentid;
1013
1014 if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1015 trace_qdisc_create(ops, dev_queue->dev, parentid);
1016 return sch;
1017 }
1018
1019 qdisc_put(sch);
1020 return NULL;
1021 }
1022 EXPORT_SYMBOL(qdisc_create_dflt);
1023
1024 /* Under qdisc_lock(qdisc) and BH! */
1025
qdisc_reset(struct Qdisc * qdisc)1026 void qdisc_reset(struct Qdisc *qdisc)
1027 {
1028 const struct Qdisc_ops *ops = qdisc->ops;
1029
1030 trace_qdisc_reset(qdisc);
1031
1032 if (ops->reset)
1033 ops->reset(qdisc);
1034
1035 __skb_queue_purge(&qdisc->gso_skb);
1036 __skb_queue_purge(&qdisc->skb_bad_txq);
1037
1038 qdisc->q.qlen = 0;
1039 qdisc->qstats.backlog = 0;
1040 }
1041 EXPORT_SYMBOL(qdisc_reset);
1042
qdisc_free(struct Qdisc * qdisc)1043 void qdisc_free(struct Qdisc *qdisc)
1044 {
1045 if (qdisc_is_percpu_stats(qdisc)) {
1046 free_percpu(qdisc->cpu_bstats);
1047 free_percpu(qdisc->cpu_qstats);
1048 }
1049
1050 kfree(qdisc);
1051 }
1052
qdisc_free_cb(struct rcu_head * head)1053 static void qdisc_free_cb(struct rcu_head *head)
1054 {
1055 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1056
1057 qdisc_free(q);
1058 }
1059
__qdisc_destroy(struct Qdisc * qdisc)1060 static void __qdisc_destroy(struct Qdisc *qdisc)
1061 {
1062 const struct Qdisc_ops *ops = qdisc->ops;
1063 struct net_device *dev = qdisc_dev(qdisc);
1064
1065 #ifdef CONFIG_NET_SCHED
1066 qdisc_hash_del(qdisc);
1067
1068 qdisc_put_stab(rtnl_dereference(qdisc->stab));
1069 #endif
1070 gen_kill_estimator(&qdisc->rate_est);
1071
1072 qdisc_reset(qdisc);
1073
1074
1075 if (ops->destroy)
1076 ops->destroy(qdisc);
1077
1078 lockdep_unregister_key(&qdisc->root_lock_key);
1079 module_put(ops->owner);
1080 netdev_put(dev, &qdisc->dev_tracker);
1081
1082 trace_qdisc_destroy(qdisc);
1083
1084 call_rcu(&qdisc->rcu, qdisc_free_cb);
1085 }
1086
qdisc_destroy(struct Qdisc * qdisc)1087 void qdisc_destroy(struct Qdisc *qdisc)
1088 {
1089 if (qdisc->flags & TCQ_F_BUILTIN)
1090 return;
1091
1092 __qdisc_destroy(qdisc);
1093 }
1094
qdisc_put(struct Qdisc * qdisc)1095 void qdisc_put(struct Qdisc *qdisc)
1096 {
1097 if (!qdisc)
1098 return;
1099
1100 if (qdisc->flags & TCQ_F_BUILTIN ||
1101 !refcount_dec_and_test(&qdisc->refcnt))
1102 return;
1103
1104 __qdisc_destroy(qdisc);
1105 }
1106 EXPORT_SYMBOL(qdisc_put);
1107
1108 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1109 * Intended to be used as optimization, this function only takes rtnl lock if
1110 * qdisc reference counter reached zero.
1111 */
1112
qdisc_put_unlocked(struct Qdisc * qdisc)1113 void qdisc_put_unlocked(struct Qdisc *qdisc)
1114 {
1115 if (qdisc->flags & TCQ_F_BUILTIN ||
1116 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1117 return;
1118
1119 __qdisc_destroy(qdisc);
1120 rtnl_unlock();
1121 }
1122 EXPORT_SYMBOL(qdisc_put_unlocked);
1123
1124 /* Attach toplevel qdisc to device queue. */
dev_graft_qdisc(struct netdev_queue * dev_queue,struct Qdisc * qdisc)1125 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1126 struct Qdisc *qdisc)
1127 {
1128 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1129 spinlock_t *root_lock;
1130
1131 root_lock = qdisc_lock(oqdisc);
1132 spin_lock_bh(root_lock);
1133
1134 /* ... and graft new one */
1135 if (qdisc == NULL)
1136 qdisc = &noop_qdisc;
1137 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1138 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1139
1140 spin_unlock_bh(root_lock);
1141
1142 return oqdisc;
1143 }
1144 EXPORT_SYMBOL(dev_graft_qdisc);
1145
shutdown_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1146 static void shutdown_scheduler_queue(struct net_device *dev,
1147 struct netdev_queue *dev_queue,
1148 void *_qdisc_default)
1149 {
1150 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1151 struct Qdisc *qdisc_default = _qdisc_default;
1152
1153 if (qdisc) {
1154 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1155 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1156
1157 qdisc_put(qdisc);
1158 }
1159 }
1160
attach_one_default_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1161 static void attach_one_default_qdisc(struct net_device *dev,
1162 struct netdev_queue *dev_queue,
1163 void *_unused)
1164 {
1165 struct Qdisc *qdisc;
1166 const struct Qdisc_ops *ops = default_qdisc_ops;
1167
1168 if (dev->priv_flags & IFF_NO_QUEUE)
1169 ops = &noqueue_qdisc_ops;
1170 else if(dev->type == ARPHRD_CAN)
1171 ops = &pfifo_fast_ops;
1172
1173 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1174 if (!qdisc)
1175 return;
1176
1177 if (!netif_is_multiqueue(dev))
1178 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1179 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1180 }
1181
attach_default_qdiscs(struct net_device * dev)1182 static void attach_default_qdiscs(struct net_device *dev)
1183 {
1184 struct netdev_queue *txq;
1185 struct Qdisc *qdisc;
1186
1187 txq = netdev_get_tx_queue(dev, 0);
1188
1189 if (!netif_is_multiqueue(dev) ||
1190 dev->priv_flags & IFF_NO_QUEUE) {
1191 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1192 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1193 rcu_assign_pointer(dev->qdisc, qdisc);
1194 qdisc_refcount_inc(qdisc);
1195 } else {
1196 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1197 if (qdisc) {
1198 rcu_assign_pointer(dev->qdisc, qdisc);
1199 qdisc->ops->attach(qdisc);
1200 }
1201 }
1202 qdisc = rtnl_dereference(dev->qdisc);
1203
1204 /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1205 if (qdisc == &noop_qdisc) {
1206 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1207 default_qdisc_ops->id, noqueue_qdisc_ops.id);
1208 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1209 dev->priv_flags |= IFF_NO_QUEUE;
1210 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1211 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1212 rcu_assign_pointer(dev->qdisc, qdisc);
1213 qdisc_refcount_inc(qdisc);
1214 dev->priv_flags ^= IFF_NO_QUEUE;
1215 }
1216
1217 #ifdef CONFIG_NET_SCHED
1218 if (qdisc != &noop_qdisc)
1219 qdisc_hash_add(qdisc, false);
1220 #endif
1221 }
1222
transition_one_qdisc(struct net_device * dev,struct netdev_queue * dev_queue,void * _need_watchdog)1223 static void transition_one_qdisc(struct net_device *dev,
1224 struct netdev_queue *dev_queue,
1225 void *_need_watchdog)
1226 {
1227 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1228 int *need_watchdog_p = _need_watchdog;
1229
1230 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1231 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1232
1233 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1234 if (need_watchdog_p) {
1235 WRITE_ONCE(dev_queue->trans_start, 0);
1236 *need_watchdog_p = 1;
1237 }
1238 }
1239
dev_activate(struct net_device * dev)1240 void dev_activate(struct net_device *dev)
1241 {
1242 int need_watchdog;
1243
1244 /* No queueing discipline is attached to device;
1245 * create default one for devices, which need queueing
1246 * and noqueue_qdisc for virtual interfaces
1247 */
1248
1249 if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1250 attach_default_qdiscs(dev);
1251
1252 if (!netif_carrier_ok(dev))
1253 /* Delay activation until next carrier-on event */
1254 return;
1255
1256 need_watchdog = 0;
1257 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1258 if (dev_ingress_queue(dev))
1259 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1260
1261 if (need_watchdog) {
1262 netif_trans_update(dev);
1263 dev_watchdog_up(dev);
1264 }
1265 }
1266 EXPORT_SYMBOL(dev_activate);
1267
qdisc_deactivate(struct Qdisc * qdisc)1268 static void qdisc_deactivate(struct Qdisc *qdisc)
1269 {
1270 if (qdisc->flags & TCQ_F_BUILTIN)
1271 return;
1272
1273 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1274 }
1275
dev_deactivate_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc_default)1276 static void dev_deactivate_queue(struct net_device *dev,
1277 struct netdev_queue *dev_queue,
1278 void *_qdisc_default)
1279 {
1280 struct Qdisc *qdisc_default = _qdisc_default;
1281 struct Qdisc *qdisc;
1282
1283 qdisc = rtnl_dereference(dev_queue->qdisc);
1284 if (qdisc) {
1285 qdisc_deactivate(qdisc);
1286 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1287 }
1288 }
1289
dev_reset_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _unused)1290 static void dev_reset_queue(struct net_device *dev,
1291 struct netdev_queue *dev_queue,
1292 void *_unused)
1293 {
1294 struct Qdisc *qdisc;
1295 bool nolock;
1296
1297 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1298 if (!qdisc)
1299 return;
1300
1301 nolock = qdisc->flags & TCQ_F_NOLOCK;
1302
1303 if (nolock)
1304 spin_lock_bh(&qdisc->seqlock);
1305 spin_lock_bh(qdisc_lock(qdisc));
1306
1307 qdisc_reset(qdisc);
1308
1309 spin_unlock_bh(qdisc_lock(qdisc));
1310 if (nolock) {
1311 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1312 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1313 spin_unlock_bh(&qdisc->seqlock);
1314 }
1315 }
1316
some_qdisc_is_busy(struct net_device * dev)1317 static bool some_qdisc_is_busy(struct net_device *dev)
1318 {
1319 unsigned int i;
1320
1321 for (i = 0; i < dev->num_tx_queues; i++) {
1322 struct netdev_queue *dev_queue;
1323 spinlock_t *root_lock;
1324 struct Qdisc *q;
1325 int val;
1326
1327 dev_queue = netdev_get_tx_queue(dev, i);
1328 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1329
1330 root_lock = qdisc_lock(q);
1331 spin_lock_bh(root_lock);
1332
1333 val = (qdisc_is_running(q) ||
1334 test_bit(__QDISC_STATE_SCHED, &q->state));
1335
1336 spin_unlock_bh(root_lock);
1337
1338 if (val)
1339 return true;
1340 }
1341 return false;
1342 }
1343
1344 /**
1345 * dev_deactivate_many - deactivate transmissions on several devices
1346 * @head: list of devices to deactivate
1347 *
1348 * This function returns only when all outstanding transmissions
1349 * have completed, unless all devices are in dismantle phase.
1350 */
dev_deactivate_many(struct list_head * head)1351 void dev_deactivate_many(struct list_head *head)
1352 {
1353 struct net_device *dev;
1354
1355 list_for_each_entry(dev, head, close_list) {
1356 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1357 &noop_qdisc);
1358 if (dev_ingress_queue(dev))
1359 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1360 &noop_qdisc);
1361
1362 dev_watchdog_down(dev);
1363 }
1364
1365 /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1366 * outstanding qdisc enqueuing calls.
1367 * This is avoided if all devices are in dismantle phase :
1368 * Caller will call synchronize_net() for us
1369 */
1370 synchronize_net();
1371
1372 list_for_each_entry(dev, head, close_list) {
1373 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1374
1375 if (dev_ingress_queue(dev))
1376 dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1377 }
1378
1379 /* Wait for outstanding qdisc_run calls. */
1380 list_for_each_entry(dev, head, close_list) {
1381 while (some_qdisc_is_busy(dev)) {
1382 /* wait_event() would avoid this sleep-loop but would
1383 * require expensive checks in the fast paths of packet
1384 * processing which isn't worth it.
1385 */
1386 schedule_timeout_uninterruptible(1);
1387 }
1388 }
1389 }
1390
dev_deactivate(struct net_device * dev)1391 void dev_deactivate(struct net_device *dev)
1392 {
1393 LIST_HEAD(single);
1394
1395 list_add(&dev->close_list, &single);
1396 dev_deactivate_many(&single);
1397 list_del(&single);
1398 }
1399 EXPORT_SYMBOL(dev_deactivate);
1400
qdisc_change_tx_queue_len(struct net_device * dev,struct netdev_queue * dev_queue)1401 static int qdisc_change_tx_queue_len(struct net_device *dev,
1402 struct netdev_queue *dev_queue)
1403 {
1404 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1405 const struct Qdisc_ops *ops = qdisc->ops;
1406
1407 if (ops->change_tx_queue_len)
1408 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1409 return 0;
1410 }
1411
dev_qdisc_change_real_num_tx(struct net_device * dev,unsigned int new_real_tx)1412 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1413 unsigned int new_real_tx)
1414 {
1415 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1416
1417 if (qdisc->ops->change_real_num_tx)
1418 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1419 }
1420
mq_change_real_num_tx(struct Qdisc * sch,unsigned int new_real_tx)1421 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1422 {
1423 #ifdef CONFIG_NET_SCHED
1424 struct net_device *dev = qdisc_dev(sch);
1425 struct Qdisc *qdisc;
1426 unsigned int i;
1427
1428 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1429 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1430 /* Only update the default qdiscs we created,
1431 * qdiscs with handles are always hashed.
1432 */
1433 if (qdisc != &noop_qdisc && !qdisc->handle)
1434 qdisc_hash_del(qdisc);
1435 }
1436 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1437 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1438 if (qdisc != &noop_qdisc && !qdisc->handle)
1439 qdisc_hash_add(qdisc, false);
1440 }
1441 #endif
1442 }
1443 EXPORT_SYMBOL(mq_change_real_num_tx);
1444
dev_qdisc_change_tx_queue_len(struct net_device * dev)1445 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1446 {
1447 bool up = dev->flags & IFF_UP;
1448 unsigned int i;
1449 int ret = 0;
1450
1451 if (up)
1452 dev_deactivate(dev);
1453
1454 for (i = 0; i < dev->num_tx_queues; i++) {
1455 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1456
1457 /* TODO: revert changes on a partial failure */
1458 if (ret)
1459 break;
1460 }
1461
1462 if (up)
1463 dev_activate(dev);
1464 return ret;
1465 }
1466
dev_init_scheduler_queue(struct net_device * dev,struct netdev_queue * dev_queue,void * _qdisc)1467 static void dev_init_scheduler_queue(struct net_device *dev,
1468 struct netdev_queue *dev_queue,
1469 void *_qdisc)
1470 {
1471 struct Qdisc *qdisc = _qdisc;
1472
1473 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1474 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1475 }
1476
dev_init_scheduler(struct net_device * dev)1477 void dev_init_scheduler(struct net_device *dev)
1478 {
1479 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1480 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1481 if (dev_ingress_queue(dev))
1482 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1483
1484 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1485 }
1486
dev_shutdown(struct net_device * dev)1487 void dev_shutdown(struct net_device *dev)
1488 {
1489 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1490 if (dev_ingress_queue(dev))
1491 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1492 qdisc_put(rtnl_dereference(dev->qdisc));
1493 rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1494
1495 WARN_ON(timer_pending(&dev->watchdog_timer));
1496 }
1497
1498 /**
1499 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1500 * @rate: Rate to compute reciprocal division values of
1501 * @mult: Multiplier for reciprocal division
1502 * @shift: Shift for reciprocal division
1503 *
1504 * The multiplier and shift for reciprocal division by rate are stored
1505 * in mult and shift.
1506 *
1507 * The deal here is to replace a divide by a reciprocal one
1508 * in fast path (a reciprocal divide is a multiply and a shift)
1509 *
1510 * Normal formula would be :
1511 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1512 *
1513 * We compute mult/shift to use instead :
1514 * time_in_ns = (len * mult) >> shift;
1515 *
1516 * We try to get the highest possible mult value for accuracy,
1517 * but have to make sure no overflows will ever happen.
1518 *
1519 * reciprocal_value() is not used here it doesn't handle 64-bit values.
1520 */
psched_ratecfg_precompute__(u64 rate,u32 * mult,u8 * shift)1521 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1522 {
1523 u64 factor = NSEC_PER_SEC;
1524
1525 *mult = 1;
1526 *shift = 0;
1527
1528 if (rate <= 0)
1529 return;
1530
1531 for (;;) {
1532 *mult = div64_u64(factor, rate);
1533 if (*mult & (1U << 31) || factor & (1ULL << 63))
1534 break;
1535 factor <<= 1;
1536 (*shift)++;
1537 }
1538 }
1539
psched_ratecfg_precompute(struct psched_ratecfg * r,const struct tc_ratespec * conf,u64 rate64)1540 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1541 const struct tc_ratespec *conf,
1542 u64 rate64)
1543 {
1544 memset(r, 0, sizeof(*r));
1545 r->overhead = conf->overhead;
1546 r->mpu = conf->mpu;
1547 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1548 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1549 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1550 }
1551 EXPORT_SYMBOL(psched_ratecfg_precompute);
1552
psched_ppscfg_precompute(struct psched_pktrate * r,u64 pktrate64)1553 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1554 {
1555 r->rate_pkts_ps = pktrate64;
1556 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1557 }
1558 EXPORT_SYMBOL(psched_ppscfg_precompute);
1559
mini_qdisc_pair_swap(struct mini_Qdisc_pair * miniqp,struct tcf_proto * tp_head)1560 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1561 struct tcf_proto *tp_head)
1562 {
1563 /* Protected with chain0->filter_chain_lock.
1564 * Can't access chain directly because tp_head can be NULL.
1565 */
1566 struct mini_Qdisc *miniq_old =
1567 rcu_dereference_protected(*miniqp->p_miniq, 1);
1568 struct mini_Qdisc *miniq;
1569
1570 if (!tp_head) {
1571 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1572 } else {
1573 miniq = miniq_old != &miniqp->miniq1 ?
1574 &miniqp->miniq1 : &miniqp->miniq2;
1575
1576 /* We need to make sure that readers won't see the miniq
1577 * we are about to modify. So ensure that at least one RCU
1578 * grace period has elapsed since the miniq was made
1579 * inactive.
1580 */
1581 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1582 cond_synchronize_rcu(miniq->rcu_state);
1583 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1584 synchronize_rcu_expedited();
1585
1586 miniq->filter_list = tp_head;
1587 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1588 }
1589
1590 if (miniq_old)
1591 /* This is counterpart of the rcu sync above. We need to
1592 * block potential new user of miniq_old until all readers
1593 * are not seeing it.
1594 */
1595 miniq_old->rcu_state = start_poll_synchronize_rcu();
1596 }
1597 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1598
mini_qdisc_pair_block_init(struct mini_Qdisc_pair * miniqp,struct tcf_block * block)1599 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1600 struct tcf_block *block)
1601 {
1602 miniqp->miniq1.block = block;
1603 miniqp->miniq2.block = block;
1604 }
1605 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1606
mini_qdisc_pair_init(struct mini_Qdisc_pair * miniqp,struct Qdisc * qdisc,struct mini_Qdisc __rcu ** p_miniq)1607 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1608 struct mini_Qdisc __rcu **p_miniq)
1609 {
1610 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1611 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1612 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1613 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1614 miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1615 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1616 miniqp->p_miniq = p_miniq;
1617 }
1618 EXPORT_SYMBOL(mini_qdisc_pair_init);
1619