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