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