~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/net/sched/sch_generic.c

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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 <linux/if_macvlan.h>
 30 #include <net/sch_generic.h>
 31 #include <net/pkt_sched.h>
 32 #include <net/dst.h>
 33 #include <trace/events/qdisc.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 
 39 /* Main transmission queue. */
 40 
 41 /* Modifications to data participating in scheduling must be protected with
 42  * qdisc_lock(qdisc) spinlock.
 43  *
 44  * The idea is the following:
 45  * - enqueue, dequeue are serialized via qdisc root lock
 46  * - ingress filtering is also serialized via qdisc root lock
 47  * - updates to tree and tree walking are only done under the rtnl mutex.
 48  */
 49 
 50 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 51 {
 52         q->gso_skb = skb;
 53         q->qstats.requeues++;
 54         qdisc_qstats_backlog_inc(q, skb);
 55         q->q.qlen++;    /* it's still part of the queue */
 56         __netif_schedule(q);
 57 
 58         return 0;
 59 }
 60 
 61 static void try_bulk_dequeue_skb(struct Qdisc *q,
 62                                  struct sk_buff *skb,
 63                                  const struct netdev_queue *txq,
 64                                  int *packets)
 65 {
 66         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 67 
 68         while (bytelimit > 0) {
 69                 struct sk_buff *nskb = q->dequeue(q);
 70 
 71                 if (!nskb)
 72                         break;
 73 
 74                 bytelimit -= nskb->len; /* covers GSO len */
 75                 skb->next = nskb;
 76                 skb = nskb;
 77                 (*packets)++; /* GSO counts as one pkt */
 78         }
 79         skb->next = NULL;
 80 }
 81 
 82 /* This variant of try_bulk_dequeue_skb() makes sure
 83  * all skbs in the chain are for the same txq
 84  */
 85 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 86                                       struct sk_buff *skb,
 87                                       int *packets)
 88 {
 89         int mapping = skb_get_queue_mapping(skb);
 90         struct sk_buff *nskb;
 91         int cnt = 0;
 92 
 93         do {
 94                 nskb = q->dequeue(q);
 95                 if (!nskb)
 96                         break;
 97                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 98                         q->skb_bad_txq = nskb;
 99                         qdisc_qstats_backlog_inc(q, nskb);
100                         q->q.qlen++;
101                         break;
102                 }
103                 skb->next = nskb;
104                 skb = nskb;
105         } while (++cnt < 8);
106         (*packets) += cnt;
107         skb->next = NULL;
108 }
109 
110 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
111  * A requeued skb (via q->gso_skb) can also be a SKB list.
112  */
113 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
114                                    int *packets)
115 {
116         struct sk_buff *skb = q->gso_skb;
117         const struct netdev_queue *txq = q->dev_queue;
118 
119         *packets = 1;
120         if (unlikely(skb)) {
121                 /* skb in gso_skb were already validated */
122                 *validate = false;
123                 /* check the reason of requeuing without tx lock first */
124                 txq = skb_get_tx_queue(txq->dev, skb);
125                 if (!netif_xmit_frozen_or_stopped(txq)) {
126                         q->gso_skb = NULL;
127                         qdisc_qstats_backlog_dec(q, skb);
128                         q->q.qlen--;
129                 } else
130                         skb = NULL;
131                 goto trace;
132         }
133         *validate = true;
134         skb = q->skb_bad_txq;
135         if (unlikely(skb)) {
136                 /* check the reason of requeuing without tx lock first */
137                 txq = skb_get_tx_queue(txq->dev, skb);
138                 if (!netif_xmit_frozen_or_stopped(txq)) {
139                         q->skb_bad_txq = NULL;
140                         qdisc_qstats_backlog_dec(q, skb);
141                         q->q.qlen--;
142                         goto bulk;
143                 }
144                 skb = NULL;
145                 goto trace;
146         }
147         if (!(q->flags & TCQ_F_ONETXQUEUE) ||
148             !netif_xmit_frozen_or_stopped(txq))
149                 skb = q->dequeue(q);
150         if (skb) {
151 bulk:
152                 if (qdisc_may_bulk(q))
153                         try_bulk_dequeue_skb(q, skb, txq, packets);
154                 else
155                         try_bulk_dequeue_skb_slow(q, skb, packets);
156         }
157 trace:
158         trace_qdisc_dequeue(q, txq, *packets, skb);
159         return skb;
160 }
161 
162 /*
163  * Transmit possibly several skbs, and handle the return status as
164  * required. Owning running seqcount bit guarantees that
165  * only one CPU can execute this function.
166  *
167  * Returns to the caller:
168  *                              0  - queue is empty or throttled.
169  *                              >0 - queue is not empty.
170  */
171 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
172                     struct net_device *dev, struct netdev_queue *txq,
173                     spinlock_t *root_lock, bool validate)
174 {
175         int ret = NETDEV_TX_BUSY;
176 
177         /* And release qdisc */
178         spin_unlock(root_lock);
179 
180         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
181         if (validate)
182                 skb = validate_xmit_skb_list(skb, dev);
183 
184         if (likely(skb)) {
185                 HARD_TX_LOCK(dev, txq, smp_processor_id());
186                 if (!netif_xmit_frozen_or_stopped(txq))
187                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
188 
189                 HARD_TX_UNLOCK(dev, txq);
190         } else {
191                 spin_lock(root_lock);
192                 return qdisc_qlen(q);
193         }
194         spin_lock(root_lock);
195 
196         if (dev_xmit_complete(ret)) {
197                 /* Driver sent out skb successfully or skb was consumed */
198                 ret = qdisc_qlen(q);
199         } else {
200                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
201                 if (unlikely(ret != NETDEV_TX_BUSY))
202                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
203                                              dev->name, ret, q->q.qlen);
204 
205                 ret = dev_requeue_skb(skb, q);
206         }
207 
208         if (ret && netif_xmit_frozen_or_stopped(txq))
209                 ret = 0;
210 
211         return ret;
212 }
213 
214 /*
215  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
216  *
217  * running seqcount guarantees only one CPU can process
218  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
219  * this queue.
220  *
221  *  netif_tx_lock serializes accesses to device driver.
222  *
223  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
224  *  if one is grabbed, another must be free.
225  *
226  * Note, that this procedure can be called by a watchdog timer
227  *
228  * Returns to the caller:
229  *                              0  - queue is empty or throttled.
230  *                              >0 - queue is not empty.
231  *
232  */
233 static inline int qdisc_restart(struct Qdisc *q, int *packets)
234 {
235         struct netdev_queue *txq;
236         struct net_device *dev;
237         spinlock_t *root_lock;
238         struct sk_buff *skb;
239         bool validate;
240 
241         /* Dequeue packet */
242         skb = dequeue_skb(q, &validate, packets);
243         if (unlikely(!skb))
244                 return 0;
245 
246         root_lock = qdisc_lock(q);
247         dev = qdisc_dev(q);
248         txq = skb_get_tx_queue(dev, skb);
249 
250         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
251 }
252 
253 void __qdisc_run(struct Qdisc *q)
254 {
255         int quota = dev_tx_weight;
256         int packets;
257 
258         while (qdisc_restart(q, &packets)) {
259                 /*
260                  * Ordered by possible occurrence: Postpone processing if
261                  * 1. we've exceeded packet quota
262                  * 2. another process needs the CPU;
263                  */
264                 quota -= packets;
265                 if (quota <= 0 || need_resched()) {
266                         __netif_schedule(q);
267                         break;
268                 }
269         }
270 
271         qdisc_run_end(q);
272 }
273 
274 unsigned long dev_trans_start(struct net_device *dev)
275 {
276         unsigned long val, res;
277         unsigned int i;
278 
279         if (is_vlan_dev(dev))
280                 dev = vlan_dev_real_dev(dev);
281         else if (netif_is_macvlan(dev))
282                 dev = macvlan_dev_real_dev(dev);
283         res = netdev_get_tx_queue(dev, 0)->trans_start;
284         for (i = 1; i < dev->num_tx_queues; i++) {
285                 val = netdev_get_tx_queue(dev, i)->trans_start;
286                 if (val && time_after(val, res))
287                         res = val;
288         }
289 
290         return res;
291 }
292 EXPORT_SYMBOL(dev_trans_start);
293 
294 static void dev_watchdog(struct timer_list *t)
295 {
296         struct net_device *dev = from_timer(dev, t, watchdog_timer);
297 
298         netif_tx_lock(dev);
299         if (!qdisc_tx_is_noop(dev)) {
300                 if (netif_device_present(dev) &&
301                     netif_running(dev) &&
302                     netif_carrier_ok(dev)) {
303                         int some_queue_timedout = 0;
304                         unsigned int i;
305                         unsigned long trans_start;
306 
307                         for (i = 0; i < dev->num_tx_queues; i++) {
308                                 struct netdev_queue *txq;
309 
310                                 txq = netdev_get_tx_queue(dev, i);
311                                 trans_start = txq->trans_start;
312                                 if (netif_xmit_stopped(txq) &&
313                                     time_after(jiffies, (trans_start +
314                                                          dev->watchdog_timeo))) {
315                                         some_queue_timedout = 1;
316                                         txq->trans_timeout++;
317                                         break;
318                                 }
319                         }
320 
321                         if (some_queue_timedout) {
322                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
323                                        dev->name, netdev_drivername(dev), i);
324                                 dev->netdev_ops->ndo_tx_timeout(dev);
325                         }
326                         if (!mod_timer(&dev->watchdog_timer,
327                                        round_jiffies(jiffies +
328                                                      dev->watchdog_timeo)))
329                                 dev_hold(dev);
330                 }
331         }
332         netif_tx_unlock(dev);
333 
334         dev_put(dev);
335 }
336 
337 void __netdev_watchdog_up(struct net_device *dev)
338 {
339         if (dev->netdev_ops->ndo_tx_timeout) {
340                 if (dev->watchdog_timeo <= 0)
341                         dev->watchdog_timeo = 5*HZ;
342                 if (!mod_timer(&dev->watchdog_timer,
343                                round_jiffies(jiffies + dev->watchdog_timeo)))
344                         dev_hold(dev);
345         }
346 }
347 
348 static void dev_watchdog_up(struct net_device *dev)
349 {
350         __netdev_watchdog_up(dev);
351 }
352 
353 static void dev_watchdog_down(struct net_device *dev)
354 {
355         netif_tx_lock_bh(dev);
356         if (del_timer(&dev->watchdog_timer))
357                 dev_put(dev);
358         netif_tx_unlock_bh(dev);
359 }
360 
361 /**
362  *      netif_carrier_on - set carrier
363  *      @dev: network device
364  *
365  * Device has detected that carrier.
366  */
367 void netif_carrier_on(struct net_device *dev)
368 {
369         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
370                 if (dev->reg_state == NETREG_UNINITIALIZED)
371                         return;
372                 atomic_inc(&dev->carrier_changes);
373                 linkwatch_fire_event(dev);
374                 if (netif_running(dev))
375                         __netdev_watchdog_up(dev);
376         }
377 }
378 EXPORT_SYMBOL(netif_carrier_on);
379 
380 /**
381  *      netif_carrier_off - clear carrier
382  *      @dev: network device
383  *
384  * Device has detected loss of carrier.
385  */
386 void netif_carrier_off(struct net_device *dev)
387 {
388         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
389                 if (dev->reg_state == NETREG_UNINITIALIZED)
390                         return;
391                 atomic_inc(&dev->carrier_changes);
392                 linkwatch_fire_event(dev);
393         }
394 }
395 EXPORT_SYMBOL(netif_carrier_off);
396 
397 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
398    under all circumstances. It is difficult to invent anything faster or
399    cheaper.
400  */
401 
402 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
403                         struct sk_buff **to_free)
404 {
405         __qdisc_drop(skb, to_free);
406         return NET_XMIT_CN;
407 }
408 
409 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
410 {
411         return NULL;
412 }
413 
414 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
415         .id             =       "noop",
416         .priv_size      =       0,
417         .enqueue        =       noop_enqueue,
418         .dequeue        =       noop_dequeue,
419         .peek           =       noop_dequeue,
420         .owner          =       THIS_MODULE,
421 };
422 
423 static struct netdev_queue noop_netdev_queue = {
424         .qdisc          =       &noop_qdisc,
425         .qdisc_sleeping =       &noop_qdisc,
426 };
427 
428 struct Qdisc noop_qdisc = {
429         .enqueue        =       noop_enqueue,
430         .dequeue        =       noop_dequeue,
431         .flags          =       TCQ_F_BUILTIN,
432         .ops            =       &noop_qdisc_ops,
433         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
434         .dev_queue      =       &noop_netdev_queue,
435         .running        =       SEQCNT_ZERO(noop_qdisc.running),
436         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
437 };
438 EXPORT_SYMBOL(noop_qdisc);
439 
440 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
441 {
442         /* register_qdisc() assigns a default of noop_enqueue if unset,
443          * but __dev_queue_xmit() treats noqueue only as such
444          * if this is NULL - so clear it here. */
445         qdisc->enqueue = NULL;
446         return 0;
447 }
448 
449 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
450         .id             =       "noqueue",
451         .priv_size      =       0,
452         .init           =       noqueue_init,
453         .enqueue        =       noop_enqueue,
454         .dequeue        =       noop_dequeue,
455         .peek           =       noop_dequeue,
456         .owner          =       THIS_MODULE,
457 };
458 
459 static const u8 prio2band[TC_PRIO_MAX + 1] = {
460         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
461 };
462 
463 /* 3-band FIFO queue: old style, but should be a bit faster than
464    generic prio+fifo combination.
465  */
466 
467 #define PFIFO_FAST_BANDS 3
468 
469 /*
470  * Private data for a pfifo_fast scheduler containing:
471  *      - queues for the three band
472  *      - bitmap indicating which of the bands contain skbs
473  */
474 struct pfifo_fast_priv {
475         u32 bitmap;
476         struct qdisc_skb_head q[PFIFO_FAST_BANDS];
477 };
478 
479 /*
480  * Convert a bitmap to the first band number where an skb is queued, where:
481  *      bitmap=0 means there are no skbs on any band.
482  *      bitmap=1 means there is an skb on band 0.
483  *      bitmap=7 means there are skbs on all 3 bands, etc.
484  */
485 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
486 
487 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
488                                              int band)
489 {
490         return priv->q + band;
491 }
492 
493 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
494                               struct sk_buff **to_free)
495 {
496         if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
497                 int band = prio2band[skb->priority & TC_PRIO_MAX];
498                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
499                 struct qdisc_skb_head *list = band2list(priv, band);
500 
501                 priv->bitmap |= (1 << band);
502                 qdisc->q.qlen++;
503                 return __qdisc_enqueue_tail(skb, qdisc, list);
504         }
505 
506         return qdisc_drop(skb, qdisc, to_free);
507 }
508 
509 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
510 {
511         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
512         int band = bitmap2band[priv->bitmap];
513 
514         if (likely(band >= 0)) {
515                 struct qdisc_skb_head *qh = band2list(priv, band);
516                 struct sk_buff *skb = __qdisc_dequeue_head(qh);
517 
518                 if (likely(skb != NULL)) {
519                         qdisc_qstats_backlog_dec(qdisc, skb);
520                         qdisc_bstats_update(qdisc, skb);
521                 }
522 
523                 qdisc->q.qlen--;
524                 if (qh->qlen == 0)
525                         priv->bitmap &= ~(1 << band);
526 
527                 return skb;
528         }
529 
530         return NULL;
531 }
532 
533 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
534 {
535         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
536         int band = bitmap2band[priv->bitmap];
537 
538         if (band >= 0) {
539                 struct qdisc_skb_head *qh = band2list(priv, band);
540 
541                 return qh->head;
542         }
543 
544         return NULL;
545 }
546 
547 static void pfifo_fast_reset(struct Qdisc *qdisc)
548 {
549         int prio;
550         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
551 
552         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
553                 __qdisc_reset_queue(band2list(priv, prio));
554 
555         priv->bitmap = 0;
556         qdisc->qstats.backlog = 0;
557         qdisc->q.qlen = 0;
558 }
559 
560 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
561 {
562         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
563 
564         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
565         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
566                 goto nla_put_failure;
567         return skb->len;
568 
569 nla_put_failure:
570         return -1;
571 }
572 
573 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
574 {
575         int prio;
576         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
577 
578         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
579                 qdisc_skb_head_init(band2list(priv, prio));
580 
581         /* Can by-pass the queue discipline */
582         qdisc->flags |= TCQ_F_CAN_BYPASS;
583         return 0;
584 }
585 
586 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
587         .id             =       "pfifo_fast",
588         .priv_size      =       sizeof(struct pfifo_fast_priv),
589         .enqueue        =       pfifo_fast_enqueue,
590         .dequeue        =       pfifo_fast_dequeue,
591         .peek           =       pfifo_fast_peek,
592         .init           =       pfifo_fast_init,
593         .reset          =       pfifo_fast_reset,
594         .dump           =       pfifo_fast_dump,
595         .owner          =       THIS_MODULE,
596 };
597 EXPORT_SYMBOL(pfifo_fast_ops);
598 
599 static struct lock_class_key qdisc_tx_busylock;
600 static struct lock_class_key qdisc_running_key;
601 
602 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
603                           const struct Qdisc_ops *ops)
604 {
605         void *p;
606         struct Qdisc *sch;
607         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
608         int err = -ENOBUFS;
609         struct net_device *dev;
610 
611         if (!dev_queue) {
612                 err = -EINVAL;
613                 goto errout;
614         }
615 
616         dev = dev_queue->dev;
617         p = kzalloc_node(size, GFP_KERNEL,
618                          netdev_queue_numa_node_read(dev_queue));
619 
620         if (!p)
621                 goto errout;
622         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
623         /* if we got non aligned memory, ask more and do alignment ourself */
624         if (sch != p) {
625                 kfree(p);
626                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
627                                  netdev_queue_numa_node_read(dev_queue));
628                 if (!p)
629                         goto errout;
630                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
631                 sch->padded = (char *) sch - (char *) p;
632         }
633         qdisc_skb_head_init(&sch->q);
634         spin_lock_init(&sch->q.lock);
635 
636         if (ops->static_flags & TCQ_F_CPUSTATS) {
637                 sch->cpu_bstats =
638                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
639                 if (!sch->cpu_bstats)
640                         goto errout1;
641 
642                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
643                 if (!sch->cpu_qstats) {
644                         free_percpu(sch->cpu_bstats);
645                         goto errout1;
646                 }
647         }
648 
649         spin_lock_init(&sch->busylock);
650         lockdep_set_class(&sch->busylock,
651                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
652 
653         seqcount_init(&sch->running);
654         lockdep_set_class(&sch->running,
655                           dev->qdisc_running_key ?: &qdisc_running_key);
656 
657         sch->ops = ops;
658         sch->flags = ops->static_flags;
659         sch->enqueue = ops->enqueue;
660         sch->dequeue = ops->dequeue;
661         sch->dev_queue = dev_queue;
662         dev_hold(dev);
663         refcount_set(&sch->refcnt, 1);
664 
665         return sch;
666 errout1:
667         kfree(p);
668 errout:
669         return ERR_PTR(err);
670 }
671 
672 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
673                                 const struct Qdisc_ops *ops,
674                                 unsigned int parentid)
675 {
676         struct Qdisc *sch;
677 
678         if (!try_module_get(ops->owner))
679                 return NULL;
680 
681         sch = qdisc_alloc(dev_queue, ops);
682         if (IS_ERR(sch)) {
683                 module_put(ops->owner);
684                 return NULL;
685         }
686         sch->parent = parentid;
687 
688         if (!ops->init || ops->init(sch, NULL) == 0)
689                 return sch;
690 
691         qdisc_destroy(sch);
692         return NULL;
693 }
694 EXPORT_SYMBOL(qdisc_create_dflt);
695 
696 /* Under qdisc_lock(qdisc) and BH! */
697 
698 void qdisc_reset(struct Qdisc *qdisc)
699 {
700         const struct Qdisc_ops *ops = qdisc->ops;
701 
702         if (ops->reset)
703                 ops->reset(qdisc);
704 
705         kfree_skb(qdisc->skb_bad_txq);
706         qdisc->skb_bad_txq = NULL;
707 
708         if (qdisc->gso_skb) {
709                 kfree_skb_list(qdisc->gso_skb);
710                 qdisc->gso_skb = NULL;
711         }
712         qdisc->q.qlen = 0;
713         qdisc->qstats.backlog = 0;
714 }
715 EXPORT_SYMBOL(qdisc_reset);
716 
717 void qdisc_free(struct Qdisc *qdisc)
718 {
719         if (qdisc_is_percpu_stats(qdisc)) {
720                 free_percpu(qdisc->cpu_bstats);
721                 free_percpu(qdisc->cpu_qstats);
722         }
723 
724         kfree((char *) qdisc - qdisc->padded);
725 }
726 
727 void qdisc_destroy(struct Qdisc *qdisc)
728 {
729         const struct Qdisc_ops  *ops = qdisc->ops;
730 
731         if (qdisc->flags & TCQ_F_BUILTIN ||
732             !refcount_dec_and_test(&qdisc->refcnt))
733                 return;
734 
735 #ifdef CONFIG_NET_SCHED
736         qdisc_hash_del(qdisc);
737 
738         qdisc_put_stab(rtnl_dereference(qdisc->stab));
739 #endif
740         gen_kill_estimator(&qdisc->rate_est);
741         if (ops->reset)
742                 ops->reset(qdisc);
743         if (ops->destroy)
744                 ops->destroy(qdisc);
745 
746         module_put(ops->owner);
747         dev_put(qdisc_dev(qdisc));
748 
749         kfree_skb_list(qdisc->gso_skb);
750         kfree_skb(qdisc->skb_bad_txq);
751         qdisc_free(qdisc);
752 }
753 EXPORT_SYMBOL(qdisc_destroy);
754 
755 /* Attach toplevel qdisc to device queue. */
756 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
757                               struct Qdisc *qdisc)
758 {
759         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
760         spinlock_t *root_lock;
761 
762         root_lock = qdisc_lock(oqdisc);
763         spin_lock_bh(root_lock);
764 
765         /* ... and graft new one */
766         if (qdisc == NULL)
767                 qdisc = &noop_qdisc;
768         dev_queue->qdisc_sleeping = qdisc;
769         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
770 
771         spin_unlock_bh(root_lock);
772 
773         return oqdisc;
774 }
775 EXPORT_SYMBOL(dev_graft_qdisc);
776 
777 static void attach_one_default_qdisc(struct net_device *dev,
778                                      struct netdev_queue *dev_queue,
779                                      void *_unused)
780 {
781         struct Qdisc *qdisc;
782         const struct Qdisc_ops *ops = default_qdisc_ops;
783 
784         if (dev->priv_flags & IFF_NO_QUEUE)
785                 ops = &noqueue_qdisc_ops;
786 
787         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
788         if (!qdisc) {
789                 netdev_info(dev, "activation failed\n");
790                 return;
791         }
792         if (!netif_is_multiqueue(dev))
793                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
794         dev_queue->qdisc_sleeping = qdisc;
795 }
796 
797 static void attach_default_qdiscs(struct net_device *dev)
798 {
799         struct netdev_queue *txq;
800         struct Qdisc *qdisc;
801 
802         txq = netdev_get_tx_queue(dev, 0);
803 
804         if (!netif_is_multiqueue(dev) ||
805             dev->priv_flags & IFF_NO_QUEUE) {
806                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
807                 dev->qdisc = txq->qdisc_sleeping;
808                 qdisc_refcount_inc(dev->qdisc);
809         } else {
810                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
811                 if (qdisc) {
812                         dev->qdisc = qdisc;
813                         qdisc->ops->attach(qdisc);
814                 }
815         }
816 #ifdef CONFIG_NET_SCHED
817         if (dev->qdisc != &noop_qdisc)
818                 qdisc_hash_add(dev->qdisc, false);
819 #endif
820 }
821 
822 static void transition_one_qdisc(struct net_device *dev,
823                                  struct netdev_queue *dev_queue,
824                                  void *_need_watchdog)
825 {
826         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
827         int *need_watchdog_p = _need_watchdog;
828 
829         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
830                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
831 
832         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
833         if (need_watchdog_p) {
834                 dev_queue->trans_start = 0;
835                 *need_watchdog_p = 1;
836         }
837 }
838 
839 void dev_activate(struct net_device *dev)
840 {
841         int need_watchdog;
842 
843         /* No queueing discipline is attached to device;
844          * create default one for devices, which need queueing
845          * and noqueue_qdisc for virtual interfaces
846          */
847 
848         if (dev->qdisc == &noop_qdisc)
849                 attach_default_qdiscs(dev);
850 
851         if (!netif_carrier_ok(dev))
852                 /* Delay activation until next carrier-on event */
853                 return;
854 
855         need_watchdog = 0;
856         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
857         if (dev_ingress_queue(dev))
858                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
859 
860         if (need_watchdog) {
861                 netif_trans_update(dev);
862                 dev_watchdog_up(dev);
863         }
864 }
865 EXPORT_SYMBOL(dev_activate);
866 
867 static void dev_deactivate_queue(struct net_device *dev,
868                                  struct netdev_queue *dev_queue,
869                                  void *_qdisc_default)
870 {
871         struct Qdisc *qdisc_default = _qdisc_default;
872         struct Qdisc *qdisc;
873 
874         qdisc = rtnl_dereference(dev_queue->qdisc);
875         if (qdisc) {
876                 spin_lock_bh(qdisc_lock(qdisc));
877 
878                 if (!(qdisc->flags & TCQ_F_BUILTIN))
879                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
880 
881                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
882                 qdisc_reset(qdisc);
883 
884                 spin_unlock_bh(qdisc_lock(qdisc));
885         }
886 }
887 
888 static bool some_qdisc_is_busy(struct net_device *dev)
889 {
890         unsigned int i;
891 
892         for (i = 0; i < dev->num_tx_queues; i++) {
893                 struct netdev_queue *dev_queue;
894                 spinlock_t *root_lock;
895                 struct Qdisc *q;
896                 int val;
897 
898                 dev_queue = netdev_get_tx_queue(dev, i);
899                 q = dev_queue->qdisc_sleeping;
900                 root_lock = qdisc_lock(q);
901 
902                 spin_lock_bh(root_lock);
903 
904                 val = (qdisc_is_running(q) ||
905                        test_bit(__QDISC_STATE_SCHED, &q->state));
906 
907                 spin_unlock_bh(root_lock);
908 
909                 if (val)
910                         return true;
911         }
912         return false;
913 }
914 
915 static void dev_qdisc_reset(struct net_device *dev,
916                             struct netdev_queue *dev_queue,
917                             void *none)
918 {
919         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
920 
921         if (qdisc)
922                 qdisc_reset(qdisc);
923 }
924 
925 /**
926  *      dev_deactivate_many - deactivate transmissions on several devices
927  *      @head: list of devices to deactivate
928  *
929  *      This function returns only when all outstanding transmissions
930  *      have completed, unless all devices are in dismantle phase.
931  */
932 void dev_deactivate_many(struct list_head *head)
933 {
934         struct net_device *dev;
935 
936         list_for_each_entry(dev, head, close_list) {
937                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
938                                          &noop_qdisc);
939                 if (dev_ingress_queue(dev))
940                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
941                                              &noop_qdisc);
942 
943                 dev_watchdog_down(dev);
944         }
945 
946         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
947          * This is avoided if all devices are in dismantle phase :
948          * Caller will call synchronize_net() for us
949          */
950         synchronize_net();
951 
952         /* Wait for outstanding qdisc_run calls. */
953         list_for_each_entry(dev, head, close_list) {
954                 while (some_qdisc_is_busy(dev))
955                         yield();
956                 /* The new qdisc is assigned at this point so we can safely
957                  * unwind stale skb lists and qdisc statistics
958                  */
959                 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
960                 if (dev_ingress_queue(dev))
961                         dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
962         }
963 }
964 
965 void dev_deactivate(struct net_device *dev)
966 {
967         LIST_HEAD(single);
968 
969         list_add(&dev->close_list, &single);
970         dev_deactivate_many(&single);
971         list_del(&single);
972 }
973 EXPORT_SYMBOL(dev_deactivate);
974 
975 static void dev_init_scheduler_queue(struct net_device *dev,
976                                      struct netdev_queue *dev_queue,
977                                      void *_qdisc)
978 {
979         struct Qdisc *qdisc = _qdisc;
980 
981         rcu_assign_pointer(dev_queue->qdisc, qdisc);
982         dev_queue->qdisc_sleeping = qdisc;
983 }
984 
985 void dev_init_scheduler(struct net_device *dev)
986 {
987         dev->qdisc = &noop_qdisc;
988         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
989         if (dev_ingress_queue(dev))
990                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
991 
992         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
993 }
994 
995 static void shutdown_scheduler_queue(struct net_device *dev,
996                                      struct netdev_queue *dev_queue,
997                                      void *_qdisc_default)
998 {
999         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1000         struct Qdisc *qdisc_default = _qdisc_default;
1001 
1002         if (qdisc) {
1003                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1004                 dev_queue->qdisc_sleeping = qdisc_default;
1005 
1006                 qdisc_destroy(qdisc);
1007         }
1008 }
1009 
1010 void dev_shutdown(struct net_device *dev)
1011 {
1012         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1013         if (dev_ingress_queue(dev))
1014                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1015         qdisc_destroy(dev->qdisc);
1016         dev->qdisc = &noop_qdisc;
1017 
1018         WARN_ON(timer_pending(&dev->watchdog_timer));
1019 }
1020 
1021 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1022                                const struct tc_ratespec *conf,
1023                                u64 rate64)
1024 {
1025         memset(r, 0, sizeof(*r));
1026         r->overhead = conf->overhead;
1027         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1028         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1029         r->mult = 1;
1030         /*
1031          * The deal here is to replace a divide by a reciprocal one
1032          * in fast path (a reciprocal divide is a multiply and a shift)
1033          *
1034          * Normal formula would be :
1035          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1036          *
1037          * We compute mult/shift to use instead :
1038          *  time_in_ns = (len * mult) >> shift;
1039          *
1040          * We try to get the highest possible mult value for accuracy,
1041          * but have to make sure no overflows will ever happen.
1042          */
1043         if (r->rate_bytes_ps > 0) {
1044                 u64 factor = NSEC_PER_SEC;
1045 
1046                 for (;;) {
1047                         r->mult = div64_u64(factor, r->rate_bytes_ps);
1048                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
1049                                 break;
1050                         factor <<= 1;
1051                         r->shift++;
1052                 }
1053         }
1054 }
1055 EXPORT_SYMBOL(psched_ratecfg_precompute);
1056 
1057 static void mini_qdisc_rcu_func(struct rcu_head *head)
1058 {
1059 }
1060 
1061 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1062                           struct tcf_proto *tp_head)
1063 {
1064         struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1065         struct mini_Qdisc *miniq;
1066 
1067         if (!tp_head) {
1068                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1069                 /* Wait for flying RCU callback before it is freed. */
1070                 rcu_barrier_bh();
1071                 return;
1072         }
1073 
1074         miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1075                 &miniqp->miniq1 : &miniqp->miniq2;
1076 
1077         /* We need to make sure that readers won't see the miniq
1078          * we are about to modify. So wait until previous call_rcu_bh callback
1079          * is done.
1080          */
1081         rcu_barrier_bh();
1082         miniq->filter_list = tp_head;
1083         rcu_assign_pointer(*miniqp->p_miniq, miniq);
1084 
1085         if (miniq_old)
1086                 /* This is counterpart of the rcu barriers above. We need to
1087                  * block potential new user of miniq_old until all readers
1088                  * are not seeing it.
1089                  */
1090                 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1091 }
1092 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1093 
1094 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1095                           struct mini_Qdisc __rcu **p_miniq)
1096 {
1097         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1098         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1099         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1100         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1101         miniqp->p_miniq = p_miniq;
1102 }
1103 EXPORT_SYMBOL(mini_qdisc_pair_init);
1104 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp