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