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