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 <linux/slab.h>
  28 #include <net/pkt_sched.h>
  29 #include <net/dst.h>
  30
  31 /* Main transmission queue. */
  32
  33 /* Modifications to data participating in scheduling must be protected with
  34  * qdisc_lock(qdisc) spinlock.
  35  *
  36  * The idea is the following:
  37  * - enqueue, dequeue are serialized via qdisc root lock
  38  * - ingress filtering is also serialized via qdisc root lock
  39  * - updates to tree and tree walking are only done under the rtnl mutex.
  40  */
  41
  42 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  43 {
  44         skb_dst_force(skb);
  45         q->gso_skb = skb;
  46         q->qstats.requeues++;
  47         q->q.qlen++;    /* it's still part of the queue */
  48         __netif_schedule(q);
  49
  50         return 0;
  51 }
  52
  53 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  54 {
  55         struct sk_buff *skb = q->gso_skb;
  56
  57         if (unlikely(skb)) {
  58                 struct net_device *dev = qdisc_dev(q);
  59                 struct netdev_queue *txq;
  60
  61                 /* check the reason of requeuing without tx lock first */
  62                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
  63                 if (!netif_tx_queue_stopped(txq) &&
  64                     !netif_tx_queue_frozen(txq)) {
  65                         q->gso_skb = NULL;
  66                         q->q.qlen--;
  67                 } else
  68                         skb = NULL;
  69         } else {
  70                 skb = q->dequeue(q);
  71         }
  72
  73         return skb;
  74 }
  75
  76 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  77                                            struct netdev_queue *dev_queue,
  78                                            struct Qdisc *q)
  79 {
  80         int ret;
  81
  82         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  83                 /*
  84                  * Same CPU holding the lock. It may be a transient
  85                  * configuration error, when hard_start_xmit() recurses. We
  86                  * detect it by checking xmit owner and drop the packet when
  87                  * deadloop is detected. Return OK to try the next skb.
  88                  */
  89                 kfree_skb(skb);
  90                 if (net_ratelimit())
  91                         printk(KERN_WARNING "Dead loop on netdevice %s, "
  92                                "fix it urgently!\n", dev_queue->dev->name);
  93                 ret = qdisc_qlen(q);
  94         } else {
  95                 /*
  96                  * Another cpu is holding lock, requeue & delay xmits for
  97                  * some time.
  98                  */
  99                 __get_cpu_var(softnet_data).cpu_collision++;
 100                 ret = dev_requeue_skb(skb, q);
 101         }
 102
 103         return ret;
 104 }
 105
 106 /*
 107  * Transmit one skb, and handle the return status as required. Holding the
 108  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 109  * function.
 110  *
 111  * Returns to the caller:
 112  *                              0  - queue is empty or throttled.
 113  *                              >0 - queue is not empty.
 114  */
 115 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 116                     struct net_device *dev, struct netdev_queue *txq,
 117                     spinlock_t *root_lock)
 118 {
 119         int ret = NETDEV_TX_BUSY;
 120
 121         /* And release qdisc */
 122         spin_unlock(root_lock);
 123
 124         HARD_TX_LOCK(dev, txq, smp_processor_id());
 125         if (!netif_tx_queue_stopped(txq) && !netif_tx_queue_frozen(txq))
 126                 ret = dev_hard_start_xmit(skb, dev, txq);
 127
 128         HARD_TX_UNLOCK(dev, txq);
 129
 130         spin_lock(root_lock);
 131
 132         if (dev_xmit_complete(ret)) {
 133                 /* Driver sent out skb successfully or skb was consumed */
 134                 ret = qdisc_qlen(q);
 135         } else if (ret == NETDEV_TX_LOCKED) {
 136                 /* Driver try lock failed */
 137                 ret = handle_dev_cpu_collision(skb, txq, q);
 138         } else {
 139                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 140                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 141                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 142                                dev->name, ret, q->q.qlen);
 143
 144                 ret = dev_requeue_skb(skb, q);
 145         }
 146
 147         if (ret && (netif_tx_queue_stopped(txq) ||
 148                     netif_tx_queue_frozen(txq)))
 149                 ret = 0;
 150
 151         return ret;
 152 }
 153
 154 /*
 155  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 156  *
 157  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 158  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 159  * this queue.
 160  *
 161  *  netif_tx_lock serializes accesses to device driver.
 162  *
 163  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 164  *  if one is grabbed, another must be free.
 165  *
 166  * Note, that this procedure can be called by a watchdog timer
 167  *
 168  * Returns to the caller:
 169  *                              0  - queue is empty or throttled.
 170  *                              >0 - queue is not empty.
 171  *
 172  */
 173 static inline int qdisc_restart(struct Qdisc *q)
 174 {
 175         struct netdev_queue *txq;
 176         struct net_device *dev;
 177         spinlock_t *root_lock;
 178         struct sk_buff *skb;
 179
 180         /* Dequeue packet */
 181         skb = dequeue_skb(q);
 182         if (unlikely(!skb))
 183                 return 0;
 184         WARN_ON_ONCE(skb_dst_is_noref(skb));
 185         root_lock = qdisc_lock(q);
 186         dev = qdisc_dev(q);
 187         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 188
 189         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 190 }
 191
 192 void __qdisc_run(struct Qdisc *q)
 193 {
 194         unsigned long start_time = jiffies;
 195
 196         while (qdisc_restart(q)) {
 197                 /*
 198                  * Postpone processing if
 199                  * 1. another process needs the CPU;
 200                  * 2. we've been doing it for too long.
 201                  */
 202                 if (need_resched() || jiffies != start_time) {
 203                         __netif_schedule(q);
 204                         break;
 205                 }
 206         }
 207
 208         clear_bit(__QDISC_STATE_RUNNING, &q->state);
 209 }
 210
 211 unsigned long dev_trans_start(struct net_device *dev)
 212 {
 213         unsigned long val, res = dev->trans_start;
 214         unsigned int i;
 215
 216         for (i = 0; i < dev->num_tx_queues; i++) {
 217                 val = netdev_get_tx_queue(dev, i)->trans_start;
 218                 if (val && time_after(val, res))
 219                         res = val;
 220         }
 221         dev->trans_start = res;
 222         return res;
 223 }
 224 EXPORT_SYMBOL(dev_trans_start);
 225
 226 static void dev_watchdog(unsigned long arg)
 227 {
 228         struct net_device *dev = (struct net_device *)arg;
 229
 230         netif_tx_lock(dev);
 231         if (!qdisc_tx_is_noop(dev)) {
 232                 if (netif_device_present(dev) &&
 233                     netif_running(dev) &&
 234                     netif_carrier_ok(dev)) {
 235                         int some_queue_timedout = 0;
 236                         unsigned int i;
 237                         unsigned long trans_start;
 238
 239                         for (i = 0; i < dev->num_tx_queues; i++) {
 240                                 struct netdev_queue *txq;
 241
 242                                 txq = netdev_get_tx_queue(dev, i);
 243                                 /*
 244                                  * old device drivers set dev->trans_start
 245                                  */
 246                                 trans_start = txq->trans_start ? : dev->trans_start;
 247                                 if (netif_tx_queue_stopped(txq) &&
 248                                     time_after(jiffies, (trans_start +
 249                                                          dev->watchdog_timeo))) {
 250                                         some_queue_timedout = 1;
 251                                         break;
 252                                 }
 253                         }
 254
 255                         if (some_queue_timedout) {
 256                                 char drivername[64];
 257                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 258                                        dev->name, netdev_drivername(dev, drivername, 64), i);
 259                                 dev->netdev_ops->ndo_tx_timeout(dev);
 260                         }
 261                         if (!mod_timer(&dev->watchdog_timer,
 262                                        round_jiffies(jiffies +
 263                                                      dev->watchdog_timeo)))
 264                                 dev_hold(dev);
 265                 }
 266         }
 267         netif_tx_unlock(dev);
 268
 269         dev_put(dev);
 270 }
 271
 272 void __netdev_watchdog_up(struct net_device *dev)
 273 {
 274         if (dev->netdev_ops->ndo_tx_timeout) {
 275                 if (dev->watchdog_timeo <= 0)
 276                         dev->watchdog_timeo = 5*HZ;
 277                 if (!mod_timer(&dev->watchdog_timer,
 278                                round_jiffies(jiffies + dev->watchdog_timeo)))
 279                         dev_hold(dev);
 280         }
 281 }
 282
 283 static void dev_watchdog_up(struct net_device *dev)
 284 {
 285         __netdev_watchdog_up(dev);
 286 }
 287
 288 static void dev_watchdog_down(struct net_device *dev)
 289 {
 290         netif_tx_lock_bh(dev);
 291         if (del_timer(&dev->watchdog_timer))
 292                 dev_put(dev);
 293         netif_tx_unlock_bh(dev);
 294 }
 295
 296 /**
 297  *      netif_carrier_on - set carrier
 298  *      @dev: network device
 299  *
 300  * Device has detected that carrier.
 301  */
 302 void netif_carrier_on(struct net_device *dev)
 303 {
 304         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 305                 if (dev->reg_state == NETREG_UNINITIALIZED)
 306                         return;
 307                 linkwatch_fire_event(dev);
 308                 if (netif_running(dev))
 309                         __netdev_watchdog_up(dev);
 310         }
 311 }
 312 EXPORT_SYMBOL(netif_carrier_on);
 313
 314 /**
 315  *      netif_carrier_off - clear carrier
 316  *      @dev: network device
 317  *
 318  * Device has detected loss of carrier.
 319  */
 320 void netif_carrier_off(struct net_device *dev)
 321 {
 322         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 323                 if (dev->reg_state == NETREG_UNINITIALIZED)
 324                         return;
 325                 linkwatch_fire_event(dev);
 326         }
 327 }
 328 EXPORT_SYMBOL(netif_carrier_off);
 329
 330 /**
 331  *      netif_notify_peers - notify network peers about existence of @dev
 332  *      @dev: network device
 333  *
 334  * Generate traffic such that interested network peers are aware of
 335  * @dev, such as by generating a gratuitous ARP. This may be used when
 336  * a device wants to inform the rest of the network about some sort of
 337  * reconfiguration such as a failover event or virtual machine
 338  * migration.
 339  */
 340 void netif_notify_peers(struct net_device *dev)
 341 {
 342         rtnl_lock();
 343         call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
 344         rtnl_unlock();
 345 }
 346 EXPORT_SYMBOL(netif_notify_peers);
 347
 348 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 349    under all circumstances. It is difficult to invent anything faster or
 350    cheaper.
 351  */
 352
 353 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 354 {
 355         kfree_skb(skb);
 356         return NET_XMIT_CN;
 357 }
 358
 359 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 360 {
 361         return NULL;
 362 }
 363
 364 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 365         .id             =       "noop",
 366         .priv_size      =       0,
 367         .enqueue        =       noop_enqueue,
 368         .dequeue        =       noop_dequeue,
 369         .peek           =       noop_dequeue,
 370         .owner          =       THIS_MODULE,
 371 };
 372
 373 static struct netdev_queue noop_netdev_queue = {
 374         .qdisc          =       &noop_qdisc,
 375         .qdisc_sleeping =       &noop_qdisc,
 376 };
 377
 378 struct Qdisc noop_qdisc = {
 379         .enqueue        =       noop_enqueue,
 380         .dequeue        =       noop_dequeue,
 381         .flags          =       TCQ_F_BUILTIN,
 382         .ops            =       &noop_qdisc_ops,
 383         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 384         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 385         .dev_queue      =       &noop_netdev_queue,
 386 };
 387 EXPORT_SYMBOL(noop_qdisc);
 388
 389 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 390         .id             =       "noqueue",
 391         .priv_size      =       0,
 392         .enqueue        =       noop_enqueue,
 393         .dequeue        =       noop_dequeue,
 394         .peek           =       noop_dequeue,
 395         .owner          =       THIS_MODULE,
 396 };
 397
 398 static struct Qdisc noqueue_qdisc;
 399 static struct netdev_queue noqueue_netdev_queue = {
 400         .qdisc          =       &noqueue_qdisc,
 401         .qdisc_sleeping =       &noqueue_qdisc,
 402 };
 403
 404 static struct Qdisc noqueue_qdisc = {
 405         .enqueue        =       NULL,
 406         .dequeue        =       noop_dequeue,
 407         .flags          =       TCQ_F_BUILTIN,
 408         .ops            =       &noqueue_qdisc_ops,
 409         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 410         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 411         .dev_queue      =       &noqueue_netdev_queue,
 412 };
 413
 414
 415 static const u8 prio2band[TC_PRIO_MAX+1] =
 416         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 417
 418 /* 3-band FIFO queue: old style, but should be a bit faster than
 419    generic prio+fifo combination.
 420  */
 421
 422 #define PFIFO_FAST_BANDS 3
 423
 424 /*
 425  * Private data for a pfifo_fast scheduler containing:
 426  *      - queues for the three band
 427  *      - bitmap indicating which of the bands contain skbs
 428  */
 429 struct pfifo_fast_priv {
 430         u32 bitmap;
 431         struct sk_buff_head q[PFIFO_FAST_BANDS];
 432 };
 433
 434 /*
 435  * Convert a bitmap to the first band number where an skb is queued, where:
 436  *      bitmap=0 means there are no skbs on any band.
 437  *      bitmap=1 means there is an skb on band 0.
 438  *      bitmap=7 means there are skbs on all 3 bands, etc.
 439  */
 440 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 441
 442 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 443                                              int band)
 444 {
 445         return priv->q + band;
 446 }
 447
 448 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 449 {
 450         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 451                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 452                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 453                 struct sk_buff_head *list = band2list(priv, band);
 454
 455                 priv->bitmap |= (1 << band);
 456                 qdisc->q.qlen++;
 457                 return __qdisc_enqueue_tail(skb, qdisc, list);
 458         }
 459
 460         return qdisc_drop(skb, qdisc);
 461 }
 462
 463 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 464 {
 465         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 466         int band = bitmap2band[priv->bitmap];
 467
 468         if (likely(band >= 0)) {
 469                 struct sk_buff_head *list = band2list(priv, band);
 470                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 471
 472                 qdisc->q.qlen--;
 473                 if (skb_queue_empty(list))
 474                         priv->bitmap &= ~(1 << band);
 475
 476                 return skb;
 477         }
 478
 479         return NULL;
 480 }
 481
 482 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
 483 {
 484         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 485         int band = bitmap2band[priv->bitmap];
 486
 487         if (band >= 0) {
 488                 struct sk_buff_head *list = band2list(priv, band);
 489
 490                 return skb_peek(list);
 491         }
 492
 493         return NULL;
 494 }
 495
 496 static void pfifo_fast_reset(struct Qdisc* qdisc)
 497 {
 498         int prio;
 499         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 500
 501         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 502                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 503
 504         priv->bitmap = 0;
 505         qdisc->qstats.backlog = 0;
 506         qdisc->q.qlen = 0;
 507 }
 508
 509 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 510 {
 511         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 512
 513         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 514         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 515         return skb->len;
 516
 517 nla_put_failure:
 518         return -1;
 519 }
 520
 521 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 522 {
 523         int prio;
 524         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 525
 526         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 527                 skb_queue_head_init(band2list(priv, prio));
 528
 529         return 0;
 530 }
 531
 532 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 533         .id             =       "pfifo_fast",
 534         .priv_size      =       sizeof(struct pfifo_fast_priv),
 535         .enqueue        =       pfifo_fast_enqueue,
 536         .dequeue        =       pfifo_fast_dequeue,
 537         .peek           =       pfifo_fast_peek,
 538         .init           =       pfifo_fast_init,
 539         .reset          =       pfifo_fast_reset,
 540         .dump           =       pfifo_fast_dump,
 541         .owner          =       THIS_MODULE,
 542 };
 543
 544 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 545                           struct Qdisc_ops *ops)
 546 {
 547         void *p;
 548         struct Qdisc *sch;
 549         unsigned int size;
 550         int err = -ENOBUFS;
 551
 552         /* ensure that the Qdisc and the private data are 64-byte aligned */
 553         size = QDISC_ALIGN(sizeof(*sch));
 554         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 555
 556         p = kzalloc(size, GFP_KERNEL);
 557         if (!p)
 558                 goto errout;
 559         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 560         sch->padded = (char *) sch - (char *) p;
 561
 562         INIT_LIST_HEAD(&sch->list);
 563         skb_queue_head_init(&sch->q);
 564         sch->ops = ops;
 565         sch->enqueue = ops->enqueue;
 566         sch->dequeue = ops->dequeue;
 567         sch->dev_queue = dev_queue;
 568         dev_hold(qdisc_dev(sch));
 569         atomic_set(&sch->refcnt, 1);
 570
 571         return sch;
 572 errout:
 573         return ERR_PTR(err);
 574 }
 575
 576 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
 577                                  struct netdev_queue *dev_queue,
 578                                  struct Qdisc_ops *ops,
 579                                  unsigned int parentid)
 580 {
 581         struct Qdisc *sch;
 582
 583         sch = qdisc_alloc(dev_queue, ops);
 584         if (IS_ERR(sch))
 585                 goto errout;
 586         sch->parent = parentid;
 587
 588         if (!ops->init || ops->init(sch, NULL) == 0)
 589                 return sch;
 590
 591         qdisc_destroy(sch);
 592 errout:
 593         return NULL;
 594 }
 595 EXPORT_SYMBOL(qdisc_create_dflt);
 596
 597 /* Under qdisc_lock(qdisc) and BH! */
 598
 599 void qdisc_reset(struct Qdisc *qdisc)
 600 {
 601         const struct Qdisc_ops *ops = qdisc->ops;
 602
 603         if (ops->reset)
 604                 ops->reset(qdisc);
 605
 606         if (qdisc->gso_skb) {
 607                 kfree_skb(qdisc->gso_skb);
 608                 qdisc->gso_skb = NULL;
 609                 qdisc->q.qlen = 0;
 610         }
 611 }
 612 EXPORT_SYMBOL(qdisc_reset);
 613
 614 static void qdisc_rcu_free(struct rcu_head *head)
 615 {
 616         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 617
 618         kfree((char *) qdisc - qdisc->padded);
 619 }
 620
 621 void qdisc_destroy(struct Qdisc *qdisc)
 622 {
 623         const struct Qdisc_ops  *ops = qdisc->ops;
 624
 625         if (qdisc->flags & TCQ_F_BUILTIN ||
 626             !atomic_dec_and_test(&qdisc->refcnt))
 627                 return;
 628
 629 #ifdef CONFIG_NET_SCHED
 630         qdisc_list_del(qdisc);
 631
 632         qdisc_put_stab(qdisc->stab);
 633 #endif
 634         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 635         if (ops->reset)
 636                 ops->reset(qdisc);
 637         if (ops->destroy)
 638                 ops->destroy(qdisc);
 639
 640         module_put(ops->owner);
 641         dev_put(qdisc_dev(qdisc));
 642
 643         kfree_skb(qdisc->gso_skb);
 644         /*
 645          * gen_estimator est_timer() might access qdisc->q.lock,
 646          * wait a RCU grace period before freeing qdisc.
 647          */
 648         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 649 }
 650 EXPORT_SYMBOL(qdisc_destroy);
 651
 652 /* Attach toplevel qdisc to device queue. */
 653 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 654                               struct Qdisc *qdisc)
 655 {
 656         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 657         spinlock_t *root_lock;
 658
 659         root_lock = qdisc_lock(oqdisc);
 660         spin_lock_bh(root_lock);
 661
 662         /* Prune old scheduler */
 663         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 664                 qdisc_reset(oqdisc);
 665
 666         /* ... and graft new one */
 667         if (qdisc == NULL)
 668                 qdisc = &noop_qdisc;
 669         dev_queue->qdisc_sleeping = qdisc;
 670         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 671
 672         spin_unlock_bh(root_lock);
 673
 674         return oqdisc;
 675 }
 676
 677 static void attach_one_default_qdisc(struct net_device *dev,
 678                                      struct netdev_queue *dev_queue,
 679                                      void *_unused)
 680 {
 681         struct Qdisc *qdisc;
 682
 683         if (dev->tx_queue_len) {
 684                 qdisc = qdisc_create_dflt(dev, dev_queue,
 685                                           &pfifo_fast_ops, TC_H_ROOT);
 686                 if (!qdisc) {
 687                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 688                         return;
 689                 }
 690
 691                 /* Can by-pass the queue discipline for default qdisc */
 692                 qdisc->flags |= TCQ_F_CAN_BYPASS;
 693         } else {
 694                 qdisc =  &noqueue_qdisc;
 695         }
 696         dev_queue->qdisc_sleeping = qdisc;
 697 }
 698
 699 static void attach_default_qdiscs(struct net_device *dev)
 700 {
 701         struct netdev_queue *txq;
 702         struct Qdisc *qdisc;
 703
 704         txq = netdev_get_tx_queue(dev, 0);
 705
 706         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 707                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 708                 dev->qdisc = txq->qdisc_sleeping;
 709                 atomic_inc(&dev->qdisc->refcnt);
 710         } else {
 711                 qdisc = qdisc_create_dflt(dev, txq, &mq_qdisc_ops, TC_H_ROOT);
 712                 if (qdisc) {
 713                         qdisc->ops->attach(qdisc);
 714                         dev->qdisc = qdisc;
 715                 }
 716         }
 717 }
 718
 719 static void transition_one_qdisc(struct net_device *dev,
 720                                  struct netdev_queue *dev_queue,
 721                                  void *_need_watchdog)
 722 {
 723         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 724         int *need_watchdog_p = _need_watchdog;
 725
 726         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 727                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 728
 729         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 730         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 731                 dev_queue->trans_start = 0;
 732                 *need_watchdog_p = 1;
 733         }
 734 }
 735
 736 void dev_activate(struct net_device *dev)
 737 {
 738         int need_watchdog;
 739
 740         /* No queueing discipline is attached to device;
 741            create default one i.e. pfifo_fast for devices,
 742            which need queueing and noqueue_qdisc for
 743            virtual interfaces
 744          */
 745
 746         if (dev->qdisc == &noop_qdisc)
 747                 attach_default_qdiscs(dev);
 748
 749         if (!netif_carrier_ok(dev))
 750                 /* Delay activation until next carrier-on event */
 751                 return;
 752
 753         need_watchdog = 0;
 754         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 755         transition_one_qdisc(dev, &dev->rx_queue, NULL);
 756
 757         if (need_watchdog) {
 758                 dev->trans_start = jiffies;
 759                 dev_watchdog_up(dev);
 760         }
 761 }
 762
 763 static void dev_deactivate_queue(struct net_device *dev,
 764                                  struct netdev_queue *dev_queue,
 765                                  void *_qdisc_default)
 766 {
 767         struct Qdisc *qdisc_default = _qdisc_default;
 768         struct Qdisc *qdisc;
 769
 770         qdisc = dev_queue->qdisc;
 771         if (qdisc) {
 772                 spin_lock_bh(qdisc_lock(qdisc));
 773
 774                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 775                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 776
 777                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 778                 qdisc_reset(qdisc);
 779
 780                 spin_unlock_bh(qdisc_lock(qdisc));
 781         }
 782 }
 783
 784 static bool some_qdisc_is_busy(struct net_device *dev)
 785 {
 786         unsigned int i;
 787
 788         for (i = 0; i < dev->num_tx_queues; i++) {
 789                 struct netdev_queue *dev_queue;
 790                 spinlock_t *root_lock;
 791                 struct Qdisc *q;
 792                 int val;
 793
 794                 dev_queue = netdev_get_tx_queue(dev, i);
 795                 q = dev_queue->qdisc_sleeping;
 796                 root_lock = qdisc_lock(q);
 797
 798                 spin_lock_bh(root_lock);
 799
 800                 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
 801                        test_bit(__QDISC_STATE_SCHED, &q->state));
 802
 803                 spin_unlock_bh(root_lock);
 804
 805                 if (val)
 806                         return true;
 807         }
 808         return false;
 809 }
 810
 811 void dev_deactivate(struct net_device *dev)
 812 {
 813         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 814         dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
 815
 816         dev_watchdog_down(dev);
 817
 818         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 819         synchronize_rcu();
 820
 821         /* Wait for outstanding qdisc_run calls. */
 822         while (some_qdisc_is_busy(dev))
 823                 yield();
 824 }
 825
 826 static void dev_init_scheduler_queue(struct net_device *dev,
 827                                      struct netdev_queue *dev_queue,
 828                                      void *_qdisc)
 829 {
 830         struct Qdisc *qdisc = _qdisc;
 831
 832         dev_queue->qdisc = qdisc;
 833         dev_queue->qdisc_sleeping = qdisc;
 834 }
 835
 836 void dev_init_scheduler(struct net_device *dev)
 837 {
 838         dev->qdisc = &noop_qdisc;
 839         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 840         dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 841
 842         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 843 }
 844
 845 static void shutdown_scheduler_queue(struct net_device *dev,
 846                                      struct netdev_queue *dev_queue,
 847                                      void *_qdisc_default)
 848 {
 849         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 850         struct Qdisc *qdisc_default = _qdisc_default;
 851
 852         if (qdisc) {
 853                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 854                 dev_queue->qdisc_sleeping = qdisc_default;
 855
 856                 qdisc_destroy(qdisc);
 857         }
 858 }
 859
 860 void dev_shutdown(struct net_device *dev)
 861 {
 862         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 863         shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 864         qdisc_destroy(dev->qdisc);
 865         dev->qdisc = &noop_qdisc;
 866
 867         WARN_ON(timer_pending(&dev->watchdog_timer));
 868 }