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