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 dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  41 {
  42         q->gso_skb = skb;
  43         q->qstats.requeues++;
  44         q->q.qlen++;    /* it's still part of the queue */
  45         __netif_schedule(q);
  46
  47         return 0;
  48 }
  49
  50 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  51 {
  52         struct sk_buff *skb = q->gso_skb;
  53
  54         if (unlikely(skb)) {
  55                 struct net_device *dev = qdisc_dev(q);
  56                 struct netdev_queue *txq;
  57
  58                 /* check the reason of requeuing without tx lock first */
  59                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
  60                 if (!netif_tx_queue_stopped(txq) &&
  61                     !netif_tx_queue_frozen(txq)) {
  62                         q->gso_skb = NULL;
  63                         q->q.qlen--;
  64                 } else
  65                         skb = NULL;
  66         } else {
  67                 skb = q->dequeue(q);
  68         }
  69
  70         return skb;
  71 }
  72
  73 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  74                                            struct netdev_queue *dev_queue,
  75                                            struct Qdisc *q)
  76 {
  77         int ret;
  78
  79         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  80                 /*
  81                  * Same CPU holding the lock. It may be a transient
  82                  * configuration error, when hard_start_xmit() recurses. We
  83                  * detect it by checking xmit owner and drop the packet when
  84                  * deadloop is detected. Return OK to try the next skb.
  85                  */
  86                 kfree_skb(skb);
  87                 if (net_ratelimit())
  88                         printk(KERN_WARNING "Dead loop on netdevice %s, "
  89                                "fix it urgently!\n", dev_queue->dev->name);
  90                 ret = qdisc_qlen(q);
  91         } else {
  92                 /*
  93                  * Another cpu is holding lock, requeue & delay xmits for
  94                  * some time.
  95                  */
  96                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
  97                 ret = dev_requeue_skb(skb, q);
  98         }
  99
 100         return ret;
 101 }
 102
 103 /*
 104  * Transmit one skb, and handle the return status as required. Holding the
 105  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 106  * function.
 107  *
 108  * Returns to the caller:
 109  *                              0  - queue is empty or throttled.
 110  *                              >0 - queue is not empty.
 111  */
 112 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 113                     struct net_device *dev, struct netdev_queue *txq,
 114                     spinlock_t *root_lock)
 115 {
 116         int ret = NETDEV_TX_BUSY;
 117
 118         /* And release qdisc */
 119         spin_unlock(root_lock);
 120
 121         HARD_TX_LOCK(dev, txq, smp_processor_id());
 122         if (!netif_tx_queue_stopped(txq) &&
 123             !netif_tx_queue_frozen(txq))
 124                 ret = dev_hard_start_xmit(skb, dev, txq);
 125         HARD_TX_UNLOCK(dev, txq);
 126
 127         spin_lock(root_lock);
 128
 129         switch (ret) {
 130         case NETDEV_TX_OK:
 131                 /* Driver sent out skb successfully */
 132                 ret = qdisc_qlen(q);
 133                 break;
 134
 135         case NETDEV_TX_LOCKED:
 136                 /* Driver try lock failed */
 137                 ret = handle_dev_cpu_collision(skb, txq, q);
 138                 break;
 139
 140         default:
 141                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 142                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 143                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 144                                dev->name, ret, q->q.qlen);
 145
 146                 ret = dev_requeue_skb(skb, q);
 147                 break;
 148         }
 149
 150         if (ret && (netif_tx_queue_stopped(txq) ||
 151                     netif_tx_queue_frozen(txq)))
 152                 ret = 0;
 153
 154         return ret;
 155 }
 156
 157 /*
 158  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 159  *
 160  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 161  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 162  * this queue.
 163  *
 164  *  netif_tx_lock serializes accesses to device driver.
 165  *
 166  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 167  *  if one is grabbed, another must be free.
 168  *
 169  * Note, that this procedure can be called by a watchdog timer
 170  *
 171  * Returns to the caller:
 172  *                              0  - queue is empty or throttled.
 173  *                              >0 - queue is not empty.
 174  *
 175  */
 176 static inline int qdisc_restart(struct Qdisc *q)
 177 {
 178         struct netdev_queue *txq;
 179         struct net_device *dev;
 180         spinlock_t *root_lock;
 181         struct sk_buff *skb;
 182
 183         /* Dequeue packet */
 184         skb = dequeue_skb(q);
 185         if (unlikely(!skb))
 186                 return 0;
 187
 188         root_lock = qdisc_lock(q);
 189         dev = qdisc_dev(q);
 190         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 191
 192         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 193 }
 194
 195 void __qdisc_run(struct Qdisc *q)
 196 {
 197         unsigned long start_time = jiffies;
 198
 199         while (qdisc_restart(q)) {
 200                 /*
 201                  * Postpone processing if
 202                  * 1. another process needs the CPU;
 203                  * 2. we've been doing it for too long.
 204                  */
 205                 if (need_resched() || jiffies != start_time) {
 206                         __netif_schedule(q);
 207                         break;
 208                 }
 209         }
 210
 211         clear_bit(__QDISC_STATE_RUNNING, &q->state);
 212 }
 213
 214 unsigned long dev_trans_start(struct net_device *dev)
 215 {
 216         unsigned long val, res = dev->trans_start;
 217         unsigned int i;
 218
 219         for (i = 0; i < dev->num_tx_queues; i++) {
 220                 val = netdev_get_tx_queue(dev, i)->trans_start;
 221                 if (val && time_after(val, res))
 222                         res = val;
 223         }
 224         dev->trans_start = res;
 225         return res;
 226 }
 227 EXPORT_SYMBOL(dev_trans_start);
 228
 229 static void dev_watchdog(unsigned long arg)
 230 {
 231         struct net_device *dev = (struct net_device *)arg;
 232
 233         netif_tx_lock(dev);
 234         if (!qdisc_tx_is_noop(dev)) {
 235                 if (netif_device_present(dev) &&
 236                     netif_running(dev) &&
 237                     netif_carrier_ok(dev)) {
 238                         int some_queue_timedout = 0;
 239                         unsigned int i;
 240                         unsigned long trans_start;
 241
 242                         for (i = 0; i < dev->num_tx_queues; i++) {
 243                                 struct netdev_queue *txq;
 244
 245                                 txq = netdev_get_tx_queue(dev, i);
 246                                 /*
 247                                  * old device drivers set dev->trans_start
 248                                  */
 249                                 trans_start = txq->trans_start ? : dev->trans_start;
 250                                 if (netif_tx_queue_stopped(txq) &&
 251                                     time_after(jiffies, (trans_start +
 252                                                          dev->watchdog_timeo))) {
 253                                         some_queue_timedout = 1;
 254                                         break;
 255                                 }
 256                         }
 257
 258                         if (some_queue_timedout) {
 259                                 char drivername[64];
 260                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 261                                        dev->name, netdev_drivername(dev, drivername, 64), i);
 262                                 dev->netdev_ops->ndo_tx_timeout(dev);
 263                         }
 264                         if (!mod_timer(&dev->watchdog_timer,
 265                                        round_jiffies(jiffies +
 266                                                      dev->watchdog_timeo)))
 267                                 dev_hold(dev);
 268                 }
 269         }
 270         netif_tx_unlock(dev);
 271
 272         dev_put(dev);
 273 }
 274
 275 void __netdev_watchdog_up(struct net_device *dev)
 276 {
 277         if (dev->netdev_ops->ndo_tx_timeout) {
 278                 if (dev->watchdog_timeo <= 0)
 279                         dev->watchdog_timeo = 5*HZ;
 280                 if (!mod_timer(&dev->watchdog_timer,
 281                                round_jiffies(jiffies + dev->watchdog_timeo)))
 282                         dev_hold(dev);
 283         }
 284 }
 285
 286 static void dev_watchdog_up(struct net_device *dev)
 287 {
 288         __netdev_watchdog_up(dev);
 289 }
 290
 291 static void dev_watchdog_down(struct net_device *dev)
 292 {
 293         netif_tx_lock_bh(dev);
 294         if (del_timer(&dev->watchdog_timer))
 295                 dev_put(dev);
 296         netif_tx_unlock_bh(dev);
 297 }
 298
 299 /**
 300  *      netif_carrier_on - set carrier
 301  *      @dev: network device
 302  *
 303  * Device has detected that carrier.
 304  */
 305 void netif_carrier_on(struct net_device *dev)
 306 {
 307         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 308                 if (dev->reg_state == NETREG_UNINITIALIZED)
 309                         return;
 310                 linkwatch_fire_event(dev);
 311                 if (netif_running(dev))
 312                         __netdev_watchdog_up(dev);
 313         }
 314 }
 315 EXPORT_SYMBOL(netif_carrier_on);
 316
 317 /**
 318  *      netif_carrier_off - clear carrier
 319  *      @dev: network device
 320  *
 321  * Device has detected loss of carrier.
 322  */
 323 void netif_carrier_off(struct net_device *dev)
 324 {
 325         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 326                 if (dev->reg_state == NETREG_UNINITIALIZED)
 327                         return;
 328                 linkwatch_fire_event(dev);
 329         }
 330 }
 331 EXPORT_SYMBOL(netif_carrier_off);
 332
 333 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 334    under all circumstances. It is difficult to invent anything faster or
 335    cheaper.
 336  */
 337
 338 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 339 {
 340         kfree_skb(skb);
 341         return NET_XMIT_CN;
 342 }
 343
 344 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 345 {
 346         return NULL;
 347 }
 348
 349 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 350         .id             =       "noop",
 351         .priv_size      =       0,
 352         .enqueue        =       noop_enqueue,
 353         .dequeue        =       noop_dequeue,
 354         .peek           =       noop_dequeue,
 355         .owner          =       THIS_MODULE,
 356 };
 357
 358 static struct netdev_queue noop_netdev_queue = {
 359         .qdisc          =       &noop_qdisc,
 360         .qdisc_sleeping =       &noop_qdisc,
 361 };
 362
 363 struct Qdisc noop_qdisc = {
 364         .enqueue        =       noop_enqueue,
 365         .dequeue        =       noop_dequeue,
 366         .flags          =       TCQ_F_BUILTIN,
 367         .ops            =       &noop_qdisc_ops,
 368         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 369         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 370         .dev_queue      =       &noop_netdev_queue,
 371 };
 372 EXPORT_SYMBOL(noop_qdisc);
 373
 374 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 375         .id             =       "noqueue",
 376         .priv_size      =       0,
 377         .enqueue        =       noop_enqueue,
 378         .dequeue        =       noop_dequeue,
 379         .peek           =       noop_dequeue,
 380         .owner          =       THIS_MODULE,
 381 };
 382
 383 static struct Qdisc noqueue_qdisc;
 384 static struct netdev_queue noqueue_netdev_queue = {
 385         .qdisc          =       &noqueue_qdisc,
 386         .qdisc_sleeping =       &noqueue_qdisc,
 387 };
 388
 389 static struct Qdisc noqueue_qdisc = {
 390         .enqueue        =       NULL,
 391         .dequeue        =       noop_dequeue,
 392         .flags          =       TCQ_F_BUILTIN,
 393         .ops            =       &noqueue_qdisc_ops,
 394         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 395         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 396         .dev_queue      =       &noqueue_netdev_queue,
 397 };
 398
 399
 400 static const u8 prio2band[TC_PRIO_MAX+1] =
 401         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 402
 403 /* 3-band FIFO queue: old style, but should be a bit faster than
 404    generic prio+fifo combination.
 405  */
 406
 407 #define PFIFO_FAST_BANDS 3
 408
 409 /*
 410  * Private data for a pfifo_fast scheduler containing:
 411  *      - queues for the three band
 412  *      - bitmap indicating which of the bands contain skbs
 413  */
 414 struct pfifo_fast_priv {
 415         u32 bitmap;
 416         struct sk_buff_head q[PFIFO_FAST_BANDS];
 417 };
 418
 419 /*
 420  * Convert a bitmap to the first band number where an skb is queued, where:
 421  *      bitmap=0 means there are no skbs on any band.
 422  *      bitmap=1 means there is an skb on band 0.
 423  *      bitmap=7 means there are skbs on all 3 bands, etc.
 424  */
 425 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 426
 427 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 428                                              int band)
 429 {
 430         return priv->q + band;
 431 }
 432
 433 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 434 {
 435         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 436                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 437                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 438                 struct sk_buff_head *list = band2list(priv, band);
 439
 440                 priv->bitmap |= (1 << band);
 441                 qdisc->q.qlen++;
 442                 return __qdisc_enqueue_tail(skb, qdisc, list);
 443         }
 444
 445         return qdisc_drop(skb, qdisc);
 446 }
 447
 448 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 449 {
 450         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 451         int band = bitmap2band[priv->bitmap];
 452
 453         if (likely(band >= 0)) {
 454                 struct sk_buff_head *list = band2list(priv, band);
 455                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 456
 457                 qdisc->q.qlen--;
 458                 if (skb_queue_empty(list))
 459                         priv->bitmap &= ~(1 << band);
 460
 461                 return skb;
 462         }
 463
 464         return NULL;
 465 }
 466
 467 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
 468 {
 469         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 470         int band = bitmap2band[priv->bitmap];
 471
 472         if (band >= 0) {
 473                 struct sk_buff_head *list = band2list(priv, band);
 474
 475                 return skb_peek(list);
 476         }
 477
 478         return NULL;
 479 }
 480
 481 static void pfifo_fast_reset(struct Qdisc* qdisc)
 482 {
 483         int prio;
 484         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 485
 486         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 487                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 488
 489         priv->bitmap = 0;
 490         qdisc->qstats.backlog = 0;
 491         qdisc->q.qlen = 0;
 492 }
 493
 494 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 495 {
 496         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 497
 498         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 499         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 500         return skb->len;
 501
 502 nla_put_failure:
 503         return -1;
 504 }
 505
 506 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 507 {
 508         int prio;
 509         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 510
 511         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 512                 skb_queue_head_init(band2list(priv, prio));
 513
 514         return 0;
 515 }
 516
 517 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 518         .id             =       "pfifo_fast",
 519         .priv_size      =       sizeof(struct pfifo_fast_priv),
 520         .enqueue        =       pfifo_fast_enqueue,
 521         .dequeue        =       pfifo_fast_dequeue,
 522         .peek           =       pfifo_fast_peek,
 523         .init           =       pfifo_fast_init,
 524         .reset          =       pfifo_fast_reset,
 525         .dump           =       pfifo_fast_dump,
 526         .owner          =       THIS_MODULE,
 527 };
 528
 529 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 530                           struct Qdisc_ops *ops)
 531 {
 532         void *p;
 533         struct Qdisc *sch;
 534         unsigned int size;
 535         int err = -ENOBUFS;
 536
 537         /* ensure that the Qdisc and the private data are 32-byte aligned */
 538         size = QDISC_ALIGN(sizeof(*sch));
 539         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 540
 541         p = kzalloc(size, GFP_KERNEL);
 542         if (!p)
 543                 goto errout;
 544         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 545         sch->padded = (char *) sch - (char *) p;
 546
 547         INIT_LIST_HEAD(&sch->list);
 548         skb_queue_head_init(&sch->q);
 549         sch->ops = ops;
 550         sch->enqueue = ops->enqueue;
 551         sch->dequeue = ops->dequeue;
 552         sch->dev_queue = dev_queue;
 553         dev_hold(qdisc_dev(sch));
 554         atomic_set(&sch->refcnt, 1);
 555
 556         return sch;
 557 errout:
 558         return ERR_PTR(err);
 559 }
 560
 561 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
 562                                  struct netdev_queue *dev_queue,
 563                                  struct Qdisc_ops *ops,
 564                                  unsigned int parentid)
 565 {
 566         struct Qdisc *sch;
 567
 568         sch = qdisc_alloc(dev_queue, ops);
 569         if (IS_ERR(sch))
 570                 goto errout;
 571         sch->parent = parentid;
 572
 573         if (!ops->init || ops->init(sch, NULL) == 0)
 574                 return sch;
 575
 576         qdisc_destroy(sch);
 577 errout:
 578         return NULL;
 579 }
 580 EXPORT_SYMBOL(qdisc_create_dflt);
 581
 582 /* Under qdisc_lock(qdisc) and BH! */
 583
 584 void qdisc_reset(struct Qdisc *qdisc)
 585 {
 586         const struct Qdisc_ops *ops = qdisc->ops;
 587
 588         if (ops->reset)
 589                 ops->reset(qdisc);
 590
 591         if (qdisc->gso_skb) {
 592                 kfree_skb(qdisc->gso_skb);
 593                 qdisc->gso_skb = NULL;
 594                 qdisc->q.qlen = 0;
 595         }
 596 }
 597 EXPORT_SYMBOL(qdisc_reset);
 598
 599 void qdisc_destroy(struct Qdisc *qdisc)
 600 {
 601         const struct Qdisc_ops  *ops = qdisc->ops;
 602
 603         if (qdisc->flags & TCQ_F_BUILTIN ||
 604             !atomic_dec_and_test(&qdisc->refcnt))
 605                 return;
 606
 607 #ifdef CONFIG_NET_SCHED
 608         qdisc_list_del(qdisc);
 609
 610         qdisc_put_stab(qdisc->stab);
 611 #endif
 612         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 613         if (ops->reset)
 614                 ops->reset(qdisc);
 615         if (ops->destroy)
 616                 ops->destroy(qdisc);
 617
 618         module_put(ops->owner);
 619         dev_put(qdisc_dev(qdisc));
 620
 621         kfree_skb(qdisc->gso_skb);
 622         kfree((char *) qdisc - qdisc->padded);
 623 }
 624 EXPORT_SYMBOL(qdisc_destroy);
 625
 626 /* Attach toplevel qdisc to device queue. */
 627 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 628                               struct Qdisc *qdisc)
 629 {
 630         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 631         spinlock_t *root_lock;
 632
 633         root_lock = qdisc_lock(oqdisc);
 634         spin_lock_bh(root_lock);
 635
 636         /* Prune old scheduler */
 637         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 638                 qdisc_reset(oqdisc);
 639
 640         /* ... and graft new one */
 641         if (qdisc == NULL)
 642                 qdisc = &noop_qdisc;
 643         dev_queue->qdisc_sleeping = qdisc;
 644         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 645
 646         spin_unlock_bh(root_lock);
 647
 648         return oqdisc;
 649 }
 650
 651 static void attach_one_default_qdisc(struct net_device *dev,
 652                                      struct netdev_queue *dev_queue,
 653                                      void *_unused)
 654 {
 655         struct Qdisc *qdisc;
 656
 657         if (dev->tx_queue_len) {
 658                 qdisc = qdisc_create_dflt(dev, dev_queue,
 659                                           &pfifo_fast_ops, TC_H_ROOT);
 660                 if (!qdisc) {
 661                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 662                         return;
 663                 }
 664
 665                 /* Can by-pass the queue discipline for default qdisc */
 666                 qdisc->flags |= TCQ_F_CAN_BYPASS;
 667         } else {
 668                 qdisc =  &noqueue_qdisc;
 669         }
 670         dev_queue->qdisc_sleeping = qdisc;
 671 }
 672
 673 static void attach_default_qdiscs(struct net_device *dev)
 674 {
 675         struct netdev_queue *txq;
 676         struct Qdisc *qdisc;
 677
 678         txq = netdev_get_tx_queue(dev, 0);
 679
 680         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 681                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 682                 dev->qdisc = txq->qdisc_sleeping;
 683                 atomic_inc(&dev->qdisc->refcnt);
 684         } else {
 685                 qdisc = qdisc_create_dflt(dev, txq, &mq_qdisc_ops, TC_H_ROOT);
 686                 if (qdisc) {
 687                         qdisc->ops->attach(qdisc);
 688                         dev->qdisc = qdisc;
 689                 }
 690         }
 691 }
 692
 693 static void transition_one_qdisc(struct net_device *dev,
 694                                  struct netdev_queue *dev_queue,
 695                                  void *_need_watchdog)
 696 {
 697         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 698         int *need_watchdog_p = _need_watchdog;
 699
 700         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 701                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 702
 703         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 704         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 705                 dev_queue->trans_start = 0;
 706                 *need_watchdog_p = 1;
 707         }
 708 }
 709
 710 void dev_activate(struct net_device *dev)
 711 {
 712         int need_watchdog;
 713
 714         /* No queueing discipline is attached to device;
 715            create default one i.e. pfifo_fast for devices,
 716            which need queueing and noqueue_qdisc for
 717            virtual interfaces
 718          */
 719
 720         if (dev->qdisc == &noop_qdisc)
 721                 attach_default_qdiscs(dev);
 722
 723         if (!netif_carrier_ok(dev))
 724                 /* Delay activation until next carrier-on event */
 725                 return;
 726
 727         need_watchdog = 0;
 728         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 729         transition_one_qdisc(dev, &dev->rx_queue, NULL);
 730
 731         if (need_watchdog) {
 732                 dev->trans_start = jiffies;
 733                 dev_watchdog_up(dev);
 734         }
 735 }
 736
 737 static void dev_deactivate_queue(struct net_device *dev,
 738                                  struct netdev_queue *dev_queue,
 739                                  void *_qdisc_default)
 740 {
 741         struct Qdisc *qdisc_default = _qdisc_default;
 742         struct Qdisc *qdisc;
 743
 744         qdisc = dev_queue->qdisc;
 745         if (qdisc) {
 746                 spin_lock_bh(qdisc_lock(qdisc));
 747
 748                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 749                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 750
 751                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 752                 qdisc_reset(qdisc);
 753
 754                 spin_unlock_bh(qdisc_lock(qdisc));
 755         }
 756 }
 757
 758 static bool some_qdisc_is_busy(struct net_device *dev)
 759 {
 760         unsigned int i;
 761
 762         for (i = 0; i < dev->num_tx_queues; i++) {
 763                 struct netdev_queue *dev_queue;
 764                 spinlock_t *root_lock;
 765                 struct Qdisc *q;
 766                 int val;
 767
 768                 dev_queue = netdev_get_tx_queue(dev, i);
 769                 q = dev_queue->qdisc_sleeping;
 770                 root_lock = qdisc_lock(q);
 771
 772                 spin_lock_bh(root_lock);
 773
 774                 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
 775                        test_bit(__QDISC_STATE_SCHED, &q->state));
 776
 777                 spin_unlock_bh(root_lock);
 778
 779                 if (val)
 780                         return true;
 781         }
 782         return false;
 783 }
 784
 785 void dev_deactivate(struct net_device *dev)
 786 {
 787         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 788         dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
 789
 790         dev_watchdog_down(dev);
 791
 792         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 793         synchronize_rcu();
 794
 795         /* Wait for outstanding qdisc_run calls. */
 796         while (some_qdisc_is_busy(dev))
 797                 yield();
 798 }
 799
 800 static void dev_init_scheduler_queue(struct net_device *dev,
 801                                      struct netdev_queue *dev_queue,
 802                                      void *_qdisc)
 803 {
 804         struct Qdisc *qdisc = _qdisc;
 805
 806         dev_queue->qdisc = qdisc;
 807         dev_queue->qdisc_sleeping = qdisc;
 808 }
 809
 810 void dev_init_scheduler(struct net_device *dev)
 811 {
 812         dev->qdisc = &noop_qdisc;
 813         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 814         dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 815
 816         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 817 }
 818
 819 static void shutdown_scheduler_queue(struct net_device *dev,
 820                                      struct netdev_queue *dev_queue,
 821                                      void *_qdisc_default)
 822 {
 823         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 824         struct Qdisc *qdisc_default = _qdisc_default;
 825
 826         if (qdisc) {
 827                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 828                 dev_queue->qdisc_sleeping = qdisc_default;
 829
 830                 qdisc_destroy(qdisc);
 831         }
 832 }
 833
 834 void dev_shutdown(struct net_device *dev)
 835 {
 836         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 837         shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 838         qdisc_destroy(dev->qdisc);
 839         dev->qdisc = &noop_qdisc;
 840
 841         WARN_ON(timer_pending(&dev->watchdog_timer));
 842 }