SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/sched/sch_generic.c

   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 <net/pkt_sched.h>
  28
  29 /* Main transmission queue. */
  30
  31 /* Modifications to data participating in scheduling must be protected with
  32  * queue->lock spinlock.
  33  *
  34  * The idea is the following:
  35  * - enqueue, dequeue are serialized via top level device
  36  *   spinlock queue->lock.
  37  * - ingress filtering is serialized via top level device
  38  *   spinlock dev->rx_queue.lock.
  39  * - updates to tree and tree walking are only done under the rtnl mutex.
  40  */
  41
  42 void qdisc_lock_tree(struct net_device *dev)
  43         __acquires(dev->rx_queue.lock)
  44 {
  45         unsigned int i;
  46
  47         local_bh_disable();
  48         for (i = 0; i < dev->num_tx_queues; i++) {
  49                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
  50                 spin_lock(&txq->lock);
  51         }
  52         spin_lock(&dev->rx_queue.lock);
  53 }
  54 EXPORT_SYMBOL(qdisc_lock_tree);
  55
  56 void qdisc_unlock_tree(struct net_device *dev)
  57         __releases(dev->rx_queue.lock)
  58 {
  59         unsigned int i;
  60
  61         spin_unlock(&dev->rx_queue.lock);
  62         for (i = 0; i < dev->num_tx_queues; i++) {
  63                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
  64                 spin_unlock(&txq->lock);
  65         }
  66         local_bh_enable();
  67 }
  68 EXPORT_SYMBOL(qdisc_unlock_tree);
  69
  70 static inline int qdisc_qlen(struct Qdisc *q)
  71 {
  72         return q->q.qlen;
  73 }
  74
  75 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  76 {
  77         if (unlikely(skb->next))
  78                 q->gso_skb = skb;
  79         else
  80                 q->ops->requeue(skb, q);
  81
  82         __netif_schedule(q);
  83         return 0;
  84 }
  85
  86 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  87 {
  88         struct sk_buff *skb;
  89
  90         if ((skb = q->gso_skb))
  91                 q->gso_skb = NULL;
  92         else
  93                 skb = q->dequeue(q);
  94
  95         return skb;
  96 }
  97
  98 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  99                                            struct netdev_queue *dev_queue,
 100                                            struct Qdisc *q)
 101 {
 102         int ret;
 103
 104         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
 105                 /*
 106                  * Same CPU holding the lock. It may be a transient
 107                  * configuration error, when hard_start_xmit() recurses. We
 108                  * detect it by checking xmit owner and drop the packet when
 109                  * deadloop is detected. Return OK to try the next skb.
 110                  */
 111                 kfree_skb(skb);
 112                 if (net_ratelimit())
 113                         printk(KERN_WARNING "Dead loop on netdevice %s, "
 114                                "fix it urgently!\n", dev_queue->dev->name);
 115                 ret = qdisc_qlen(q);
 116         } else {
 117                 /*
 118                  * Another cpu is holding lock, requeue & delay xmits for
 119                  * some time.
 120                  */
 121                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
 122                 ret = dev_requeue_skb(skb, q);
 123         }
 124
 125         return ret;
 126 }
 127
 128 /*
 129  * NOTE: Called under queue->lock with locally disabled BH.
 130  *
 131  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 132  * this qdisc at a time. queue->lock serializes queue accesses for
 133  * this queue AND txq->qdisc pointer itself.
 134  *
 135  *  netif_tx_lock serializes accesses to device driver.
 136  *
 137  *  queue->lock and netif_tx_lock are mutually exclusive,
 138  *  if one is grabbed, another must be free.
 139  *
 140  * Note, that this procedure can be called by a watchdog timer
 141  *
 142  * Returns to the caller:
 143  *                              0  - queue is empty or throttled.
 144  *                              >0 - queue is not empty.
 145  *
 146  */
 147 static inline int qdisc_restart(struct Qdisc *q)
 148 {
 149         struct netdev_queue *txq;
 150         int ret = NETDEV_TX_BUSY;
 151         struct net_device *dev;
 152         spinlock_t *root_lock;
 153         struct sk_buff *skb;
 154
 155         /* Dequeue packet */
 156         if (unlikely((skb = dequeue_skb(q)) == NULL))
 157                 return 0;
 158
 159         root_lock = qdisc_root_lock(q);
 160
 161         /* And release qdisc */
 162         spin_unlock(root_lock);
 163
 164         dev = qdisc_dev(q);
 165         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 166
 167         HARD_TX_LOCK(dev, txq, smp_processor_id());
 168         if (!netif_subqueue_stopped(dev, skb))
 169                 ret = dev_hard_start_xmit(skb, dev, txq);
 170         HARD_TX_UNLOCK(dev, txq);
 171
 172         spin_lock(root_lock);
 173
 174         switch (ret) {
 175         case NETDEV_TX_OK:
 176                 /* Driver sent out skb successfully */
 177                 ret = qdisc_qlen(q);
 178                 break;
 179
 180         case NETDEV_TX_LOCKED:
 181                 /* Driver try lock failed */
 182                 ret = handle_dev_cpu_collision(skb, txq, q);
 183                 break;
 184
 185         default:
 186                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 187                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 188                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 189                                dev->name, ret, q->q.qlen);
 190
 191                 ret = dev_requeue_skb(skb, q);
 192                 break;
 193         }
 194
 195         if (ret && netif_tx_queue_stopped(txq))
 196                 ret = 0;
 197
 198         return ret;
 199 }
 200
 201 void __qdisc_run(struct Qdisc *q)
 202 {
 203         unsigned long start_time = jiffies;
 204
 205         while (qdisc_restart(q)) {
 206                 /*
 207                  * Postpone processing if
 208                  * 1. another process needs the CPU;
 209                  * 2. we've been doing it for too long.
 210                  */
 211                 if (need_resched() || jiffies != start_time) {
 212                         __netif_schedule(q);
 213                         break;
 214                 }
 215         }
 216
 217         clear_bit(__QDISC_STATE_RUNNING, &q->state);
 218 }
 219
 220 static void dev_watchdog(unsigned long arg)
 221 {
 222         struct net_device *dev = (struct net_device *)arg;
 223
 224         netif_tx_lock(dev);
 225         if (!qdisc_tx_is_noop(dev)) {
 226                 if (netif_device_present(dev) &&
 227                     netif_running(dev) &&
 228                     netif_carrier_ok(dev)) {
 229                         int some_queue_stopped = 0;
 230                         unsigned int i;
 231
 232                         for (i = 0; i < dev->num_tx_queues; i++) {
 233                                 struct netdev_queue *txq;
 234
 235                                 txq = netdev_get_tx_queue(dev, i);
 236                                 if (netif_tx_queue_stopped(txq)) {
 237                                         some_queue_stopped = 1;
 238                                         break;
 239                                 }
 240                         }
 241
 242                         if (some_queue_stopped &&
 243                             time_after(jiffies, (dev->trans_start +
 244                                                  dev->watchdog_timeo))) {
 245                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: "
 246                                        "transmit timed out\n",
 247                                        dev->name);
 248                                 dev->tx_timeout(dev);
 249                                 WARN_ON_ONCE(1);
 250                         }
 251                         if (!mod_timer(&dev->watchdog_timer,
 252                                        round_jiffies(jiffies +
 253                                                      dev->watchdog_timeo)))
 254                                 dev_hold(dev);
 255                 }
 256         }
 257         netif_tx_unlock(dev);
 258
 259         dev_put(dev);
 260 }
 261
 262 void __netdev_watchdog_up(struct net_device *dev)
 263 {
 264         if (dev->tx_timeout) {
 265                 if (dev->watchdog_timeo <= 0)
 266                         dev->watchdog_timeo = 5*HZ;
 267                 if (!mod_timer(&dev->watchdog_timer,
 268                                round_jiffies(jiffies + dev->watchdog_timeo)))
 269                         dev_hold(dev);
 270         }
 271 }
 272
 273 static void dev_watchdog_up(struct net_device *dev)
 274 {
 275         __netdev_watchdog_up(dev);
 276 }
 277
 278 static void dev_watchdog_down(struct net_device *dev)
 279 {
 280         netif_tx_lock_bh(dev);
 281         if (del_timer(&dev->watchdog_timer))
 282                 dev_put(dev);
 283         netif_tx_unlock_bh(dev);
 284 }
 285
 286 /**
 287  *      netif_carrier_on - set carrier
 288  *      @dev: network device
 289  *
 290  * Device has detected that carrier.
 291  */
 292 void netif_carrier_on(struct net_device *dev)
 293 {
 294         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 295                 linkwatch_fire_event(dev);
 296                 if (netif_running(dev))
 297                         __netdev_watchdog_up(dev);
 298         }
 299 }
 300 EXPORT_SYMBOL(netif_carrier_on);
 301
 302 /**
 303  *      netif_carrier_off - clear carrier
 304  *      @dev: network device
 305  *
 306  * Device has detected loss of carrier.
 307  */
 308 void netif_carrier_off(struct net_device *dev)
 309 {
 310         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 311                 linkwatch_fire_event(dev);
 312 }
 313 EXPORT_SYMBOL(netif_carrier_off);
 314
 315 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 316    under all circumstances. It is difficult to invent anything faster or
 317    cheaper.
 318  */
 319
 320 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 321 {
 322         kfree_skb(skb);
 323         return NET_XMIT_CN;
 324 }
 325
 326 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 327 {
 328         return NULL;
 329 }
 330
 331 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 332 {
 333         if (net_ratelimit())
 334                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 335                        skb->dev->name);
 336         kfree_skb(skb);
 337         return NET_XMIT_CN;
 338 }
 339
 340 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 341         .id             =       "noop",
 342         .priv_size      =       0,
 343         .enqueue        =       noop_enqueue,
 344         .dequeue        =       noop_dequeue,
 345         .requeue        =       noop_requeue,
 346         .owner          =       THIS_MODULE,
 347 };
 348
 349 static struct netdev_queue noop_netdev_queue = {
 350         .lock           =       __SPIN_LOCK_UNLOCKED(noop_netdev_queue.lock),
 351         .qdisc          =       &noop_qdisc,
 352 };
 353
 354 struct Qdisc noop_qdisc = {
 355         .enqueue        =       noop_enqueue,
 356         .dequeue        =       noop_dequeue,
 357         .flags          =       TCQ_F_BUILTIN,
 358         .ops            =       &noop_qdisc_ops,
 359         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 360         .dev_queue      =       &noop_netdev_queue,
 361 };
 362 EXPORT_SYMBOL(noop_qdisc);
 363
 364 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 365         .id             =       "noqueue",
 366         .priv_size      =       0,
 367         .enqueue        =       noop_enqueue,
 368         .dequeue        =       noop_dequeue,
 369         .requeue        =       noop_requeue,
 370         .owner          =       THIS_MODULE,
 371 };
 372
 373 static struct Qdisc noqueue_qdisc = {
 374         .enqueue        =       NULL,
 375         .dequeue        =       noop_dequeue,
 376         .flags          =       TCQ_F_BUILTIN,
 377         .ops            =       &noqueue_qdisc_ops,
 378         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 379 };
 380
 381
 382 static const u8 prio2band[TC_PRIO_MAX+1] =
 383         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 384
 385 /* 3-band FIFO queue: old style, but should be a bit faster than
 386    generic prio+fifo combination.
 387  */
 388
 389 #define PFIFO_FAST_BANDS 3
 390
 391 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 392                                              struct Qdisc *qdisc)
 393 {
 394         struct sk_buff_head *list = qdisc_priv(qdisc);
 395         return list + prio2band[skb->priority & TC_PRIO_MAX];
 396 }
 397
 398 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 399 {
 400         struct sk_buff_head *list = prio2list(skb, qdisc);
 401
 402         if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
 403                 qdisc->q.qlen++;
 404                 return __qdisc_enqueue_tail(skb, qdisc, list);
 405         }
 406
 407         return qdisc_drop(skb, qdisc);
 408 }
 409
 410 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 411 {
 412         int prio;
 413         struct sk_buff_head *list = qdisc_priv(qdisc);
 414
 415         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 416                 if (!skb_queue_empty(list + prio)) {
 417                         qdisc->q.qlen--;
 418                         return __qdisc_dequeue_head(qdisc, list + prio);
 419                 }
 420         }
 421
 422         return NULL;
 423 }
 424
 425 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 426 {
 427         qdisc->q.qlen++;
 428         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 429 }
 430
 431 static void pfifo_fast_reset(struct Qdisc* qdisc)
 432 {
 433         int prio;
 434         struct sk_buff_head *list = qdisc_priv(qdisc);
 435
 436         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 437                 __qdisc_reset_queue(qdisc, list + prio);
 438
 439         qdisc->qstats.backlog = 0;
 440         qdisc->q.qlen = 0;
 441 }
 442
 443 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 444 {
 445         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 446
 447         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 448         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 449         return skb->len;
 450
 451 nla_put_failure:
 452         return -1;
 453 }
 454
 455 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 456 {
 457         int prio;
 458         struct sk_buff_head *list = qdisc_priv(qdisc);
 459
 460         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 461                 skb_queue_head_init(list + prio);
 462
 463         return 0;
 464 }
 465
 466 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 467         .id             =       "pfifo_fast",
 468         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 469         .enqueue        =       pfifo_fast_enqueue,
 470         .dequeue        =       pfifo_fast_dequeue,
 471         .requeue        =       pfifo_fast_requeue,
 472         .init           =       pfifo_fast_init,
 473         .reset          =       pfifo_fast_reset,
 474         .dump           =       pfifo_fast_dump,
 475         .owner          =       THIS_MODULE,
 476 };
 477
 478 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 479                           struct Qdisc_ops *ops)
 480 {
 481         void *p;
 482         struct Qdisc *sch;
 483         unsigned int size;
 484         int err = -ENOBUFS;
 485
 486         /* ensure that the Qdisc and the private data are 32-byte aligned */
 487         size = QDISC_ALIGN(sizeof(*sch));
 488         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 489
 490         p = kzalloc(size, GFP_KERNEL);
 491         if (!p)
 492                 goto errout;
 493         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 494         sch->padded = (char *) sch - (char *) p;
 495
 496         INIT_LIST_HEAD(&sch->list);
 497         skb_queue_head_init(&sch->q);
 498         sch->ops = ops;
 499         sch->enqueue = ops->enqueue;
 500         sch->dequeue = ops->dequeue;
 501         sch->dev_queue = dev_queue;
 502         dev_hold(qdisc_dev(sch));
 503         atomic_set(&sch->refcnt, 1);
 504
 505         return sch;
 506 errout:
 507         return ERR_PTR(err);
 508 }
 509
 510 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
 511                                  struct netdev_queue *dev_queue,
 512                                  struct Qdisc_ops *ops,
 513                                  unsigned int parentid)
 514 {
 515         struct Qdisc *sch;
 516
 517         sch = qdisc_alloc(dev_queue, ops);
 518         if (IS_ERR(sch))
 519                 goto errout;
 520         sch->parent = parentid;
 521
 522         if (!ops->init || ops->init(sch, NULL) == 0)
 523                 return sch;
 524
 525         qdisc_destroy(sch);
 526 errout:
 527         return NULL;
 528 }
 529 EXPORT_SYMBOL(qdisc_create_dflt);
 530
 531 /* Under queue->lock and BH! */
 532
 533 void qdisc_reset(struct Qdisc *qdisc)
 534 {
 535         const struct Qdisc_ops *ops = qdisc->ops;
 536
 537         if (ops->reset)
 538                 ops->reset(qdisc);
 539 }
 540 EXPORT_SYMBOL(qdisc_reset);
 541
 542 /* this is the rcu callback function to clean up a qdisc when there
 543  * are no further references to it */
 544
 545 static void __qdisc_destroy(struct rcu_head *head)
 546 {
 547         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 548         const struct Qdisc_ops  *ops = qdisc->ops;
 549
 550         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 551         if (ops->reset)
 552                 ops->reset(qdisc);
 553         if (ops->destroy)
 554                 ops->destroy(qdisc);
 555
 556         module_put(ops->owner);
 557         dev_put(qdisc_dev(qdisc));
 558
 559         kfree((char *) qdisc - qdisc->padded);
 560 }
 561
 562 /* Under queue->lock and BH! */
 563
 564 void qdisc_destroy(struct Qdisc *qdisc)
 565 {
 566         struct net_device *dev = qdisc_dev(qdisc);
 567
 568         if (qdisc->flags & TCQ_F_BUILTIN ||
 569             !atomic_dec_and_test(&qdisc->refcnt))
 570                 return;
 571
 572         spin_lock_bh(&dev->qdisc_list_lock);
 573         list_del(&qdisc->list);
 574         spin_unlock_bh(&dev->qdisc_list_lock);
 575
 576         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 577 }
 578 EXPORT_SYMBOL(qdisc_destroy);
 579
 580 static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
 581 {
 582         unsigned int i;
 583
 584         for (i = 0; i < dev->num_tx_queues; i++) {
 585                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 586
 587                 if (txq->qdisc_sleeping != &noop_qdisc)
 588                         return false;
 589         }
 590         return true;
 591 }
 592
 593 static void attach_one_default_qdisc(struct net_device *dev,
 594                                      struct netdev_queue *dev_queue,
 595                                      void *_unused)
 596 {
 597         struct Qdisc *qdisc;
 598
 599         if (dev->tx_queue_len) {
 600                 qdisc = qdisc_create_dflt(dev, dev_queue,
 601                                           &pfifo_fast_ops, TC_H_ROOT);
 602                 if (!qdisc) {
 603                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 604                         return;
 605                 }
 606                 spin_lock_bh(&dev->qdisc_list_lock);
 607                 list_add_tail(&qdisc->list, &dev->qdisc_list);
 608                 spin_unlock_bh(&dev->qdisc_list_lock);
 609         } else {
 610                 qdisc =  &noqueue_qdisc;
 611         }
 612         dev_queue->qdisc_sleeping = qdisc;
 613 }
 614
 615 static void transition_one_qdisc(struct net_device *dev,
 616                                  struct netdev_queue *dev_queue,
 617                                  void *_need_watchdog)
 618 {
 619         int *need_watchdog_p = _need_watchdog;
 620
 621         spin_lock_bh(&dev_queue->lock);
 622         rcu_assign_pointer(dev_queue->qdisc, dev_queue->qdisc_sleeping);
 623         if (dev_queue->qdisc != &noqueue_qdisc)
 624                 *need_watchdog_p = 1;
 625         spin_unlock_bh(&dev_queue->lock);
 626 }
 627
 628 void dev_activate(struct net_device *dev)
 629 {
 630         int need_watchdog;
 631
 632         /* No queueing discipline is attached to device;
 633            create default one i.e. pfifo_fast for devices,
 634            which need queueing and noqueue_qdisc for
 635            virtual interfaces
 636          */
 637
 638         if (dev_all_qdisc_sleeping_noop(dev))
 639                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 640
 641         if (!netif_carrier_ok(dev))
 642                 /* Delay activation until next carrier-on event */
 643                 return;
 644
 645         need_watchdog = 0;
 646         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 647
 648         if (need_watchdog) {
 649                 dev->trans_start = jiffies;
 650                 dev_watchdog_up(dev);
 651         }
 652 }
 653
 654 static void dev_deactivate_queue(struct net_device *dev,
 655                                  struct netdev_queue *dev_queue,
 656                                  void *_qdisc_default)
 657 {
 658         struct Qdisc *qdisc_default = _qdisc_default;
 659         struct sk_buff *skb = NULL;
 660         struct Qdisc *qdisc;
 661
 662         spin_lock_bh(&dev_queue->lock);
 663
 664         qdisc = dev_queue->qdisc;
 665         if (qdisc) {
 666                 dev_queue->qdisc = qdisc_default;
 667                 qdisc_reset(qdisc);
 668
 669                 skb = qdisc->gso_skb;
 670                 qdisc->gso_skb = NULL;
 671         }
 672
 673         spin_unlock_bh(&dev_queue->lock);
 674
 675         kfree_skb(skb);
 676 }
 677
 678 static bool some_qdisc_is_running(struct net_device *dev, int lock)
 679 {
 680         unsigned int i;
 681
 682         for (i = 0; i < dev->num_tx_queues; i++) {
 683                 struct netdev_queue *dev_queue;
 684                 spinlock_t *root_lock;
 685                 struct Qdisc *q;
 686                 int val;
 687
 688                 dev_queue = netdev_get_tx_queue(dev, i);
 689                 q = dev_queue->qdisc;
 690                 root_lock = qdisc_root_lock(q);
 691
 692                 if (lock)
 693                         spin_lock_bh(root_lock);
 694
 695                 val = test_bit(__QDISC_STATE_RUNNING, &q->state);
 696
 697                 if (lock)
 698                         spin_unlock_bh(root_lock);
 699
 700                 if (val)
 701                         return true;
 702         }
 703         return false;
 704 }
 705
 706 void dev_deactivate(struct net_device *dev)
 707 {
 708         bool running;
 709
 710         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 711
 712         dev_watchdog_down(dev);
 713
 714         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 715         synchronize_rcu();
 716
 717         /* Wait for outstanding qdisc_run calls. */
 718         do {
 719                 while (some_qdisc_is_running(dev, 0))
 720                         yield();
 721
 722                 /*
 723                  * Double-check inside queue lock to ensure that all effects
 724                  * of the queue run are visible when we return.
 725                  */
 726                 running = some_qdisc_is_running(dev, 1);
 727
 728                 /*
 729                  * The running flag should never be set at this point because
 730                  * we've already set dev->qdisc to noop_qdisc *inside* the same
 731                  * pair of spin locks.  That is, if any qdisc_run starts after
 732                  * our initial test it should see the noop_qdisc and then
 733                  * clear the RUNNING bit before dropping the queue lock.  So
 734                  * if it is set here then we've found a bug.
 735                  */
 736         } while (WARN_ON_ONCE(running));
 737 }
 738
 739 static void dev_init_scheduler_queue(struct net_device *dev,
 740                                      struct netdev_queue *dev_queue,
 741                                      void *_qdisc)
 742 {
 743         struct Qdisc *qdisc = _qdisc;
 744
 745         dev_queue->qdisc = qdisc;
 746         dev_queue->qdisc_sleeping = qdisc;
 747 }
 748
 749 void dev_init_scheduler(struct net_device *dev)
 750 {
 751         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 752         dev_init_scheduler_queue(dev, &dev->rx_queue, NULL);
 753
 754         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 755 }
 756
 757 static void shutdown_scheduler_queue(struct net_device *dev,
 758                                      struct netdev_queue *dev_queue,
 759                                      void *_qdisc_default)
 760 {
 761         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 762         struct Qdisc *qdisc_default = _qdisc_default;
 763
 764         if (qdisc) {
 765                 spinlock_t *root_lock = qdisc_root_lock(qdisc);
 766
 767                 dev_queue->qdisc = qdisc_default;
 768                 dev_queue->qdisc_sleeping = qdisc_default;
 769
 770                 spin_lock(root_lock);
 771                 qdisc_destroy(qdisc);
 772                 spin_unlock(root_lock);
 773         }
 774 }
 775
 776 void dev_shutdown(struct net_device *dev)
 777 {
 778         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 779         shutdown_scheduler_queue(dev, &dev->rx_queue, NULL);
 780         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 781 }