SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
 126                 /* an error implies that the skb was consumed */
 127                 if (ret < 0)
 128                         ret = NETDEV_TX_OK;
 129                 /* all NET_XMIT codes map to NETDEV_TX_OK */
 130                 ret &= ~NET_XMIT_MASK;
 131         }
 132         HARD_TX_UNLOCK(dev, txq);
 133
 134         spin_lock(root_lock);
 135
 136         switch (ret) {
 137         case NETDEV_TX_OK:
 138                 /* Driver sent out skb successfully */
 139                 ret = qdisc_qlen(q);
 140                 break;
 141
 142         case NETDEV_TX_LOCKED:
 143                 /* Driver try lock failed */
 144                 ret = handle_dev_cpu_collision(skb, txq, q);
 145                 break;
 146
 147         default:
 148                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 149                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 150                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 151                                dev->name, ret, q->q.qlen);
 152
 153                 ret = dev_requeue_skb(skb, q);
 154                 break;
 155         }
 156
 157         if (ret && (netif_tx_queue_stopped(txq) ||
 158                     netif_tx_queue_frozen(txq)))
 159                 ret = 0;
 160
 161         return ret;
 162 }
 163
 164 /*
 165  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 166  *
 167  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 168  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 169  * this queue.
 170  *
 171  *  netif_tx_lock serializes accesses to device driver.
 172  *
 173  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 174  *  if one is grabbed, another must be free.
 175  *
 176  * Note, that this procedure can be called by a watchdog timer
 177  *
 178  * Returns to the caller:
 179  *                              0  - queue is empty or throttled.
 180  *                              >0 - queue is not empty.
 181  *
 182  */
 183 static inline int qdisc_restart(struct Qdisc *q)
 184 {
 185         struct netdev_queue *txq;
 186         struct net_device *dev;
 187         spinlock_t *root_lock;
 188         struct sk_buff *skb;
 189
 190         /* Dequeue packet */
 191         skb = dequeue_skb(q);
 192         if (unlikely(!skb))
 193                 return 0;
 194
 195         root_lock = qdisc_lock(q);
 196         dev = qdisc_dev(q);
 197         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 198
 199         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 200 }
 201
 202 void __qdisc_run(struct Qdisc *q)
 203 {
 204         unsigned long start_time = jiffies;
 205
 206         while (qdisc_restart(q)) {
 207                 /*
 208                  * Postpone processing if
 209                  * 1. another process needs the CPU;
 210                  * 2. we've been doing it for too long.
 211                  */
 212                 if (need_resched() || jiffies != start_time) {
 213                         __netif_schedule(q);
 214                         break;
 215                 }
 216         }
 217
 218         clear_bit(__QDISC_STATE_RUNNING, &q->state);
 219 }
 220
 221 unsigned long dev_trans_start(struct net_device *dev)
 222 {
 223         unsigned long val, res = dev->trans_start;
 224         unsigned int i;
 225
 226         for (i = 0; i < dev->num_tx_queues; i++) {
 227                 val = netdev_get_tx_queue(dev, i)->trans_start;
 228                 if (val && time_after(val, res))
 229                         res = val;
 230         }
 231         dev->trans_start = res;
 232         return res;
 233 }
 234 EXPORT_SYMBOL(dev_trans_start);
 235
 236 static void dev_watchdog(unsigned long arg)
 237 {
 238         struct net_device *dev = (struct net_device *)arg;
 239
 240         netif_tx_lock(dev);
 241         if (!qdisc_tx_is_noop(dev)) {
 242                 if (netif_device_present(dev) &&
 243                     netif_running(dev) &&
 244                     netif_carrier_ok(dev)) {
 245                         int some_queue_timedout = 0;
 246                         unsigned int i;
 247                         unsigned long trans_start;
 248
 249                         for (i = 0; i < dev->num_tx_queues; i++) {
 250                                 struct netdev_queue *txq;
 251
 252                                 txq = netdev_get_tx_queue(dev, i);
 253                                 /*
 254                                  * old device drivers set dev->trans_start
 255                                  */
 256                                 trans_start = txq->trans_start ? : dev->trans_start;
 257                                 if (netif_tx_queue_stopped(txq) &&
 258                                     time_after(jiffies, (trans_start +
 259                                                          dev->watchdog_timeo))) {
 260                                         some_queue_timedout = 1;
 261                                         break;
 262                                 }
 263                         }
 264
 265                         if (some_queue_timedout) {
 266                                 char drivername[64];
 267                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 268                                        dev->name, netdev_drivername(dev, drivername, 64), i);
 269                                 dev->netdev_ops->ndo_tx_timeout(dev);
 270                         }
 271                         if (!mod_timer(&dev->watchdog_timer,
 272                                        round_jiffies(jiffies +
 273                                                      dev->watchdog_timeo)))
 274                                 dev_hold(dev);
 275                 }
 276         }
 277         netif_tx_unlock(dev);
 278
 279         dev_put(dev);
 280 }
 281
 282 void __netdev_watchdog_up(struct net_device *dev)
 283 {
 284         if (dev->netdev_ops->ndo_tx_timeout) {
 285                 if (dev->watchdog_timeo <= 0)
 286                         dev->watchdog_timeo = 5*HZ;
 287                 if (!mod_timer(&dev->watchdog_timer,
 288                                round_jiffies(jiffies + dev->watchdog_timeo)))
 289                         dev_hold(dev);
 290         }
 291 }
 292
 293 static void dev_watchdog_up(struct net_device *dev)
 294 {
 295         __netdev_watchdog_up(dev);
 296 }
 297
 298 static void dev_watchdog_down(struct net_device *dev)
 299 {
 300         netif_tx_lock_bh(dev);
 301         if (del_timer(&dev->watchdog_timer))
 302                 dev_put(dev);
 303         netif_tx_unlock_bh(dev);
 304 }
 305
 306 /**
 307  *      netif_carrier_on - set carrier
 308  *      @dev: network device
 309  *
 310  * Device has detected that carrier.
 311  */
 312 void netif_carrier_on(struct net_device *dev)
 313 {
 314         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 315                 if (dev->reg_state == NETREG_UNINITIALIZED)
 316                         return;
 317                 linkwatch_fire_event(dev);
 318                 if (netif_running(dev))
 319                         __netdev_watchdog_up(dev);
 320         }
 321 }
 322 EXPORT_SYMBOL(netif_carrier_on);
 323
 324 /**
 325  *      netif_carrier_off - clear carrier
 326  *      @dev: network device
 327  *
 328  * Device has detected loss of carrier.
 329  */
 330 void netif_carrier_off(struct net_device *dev)
 331 {
 332         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 333                 if (dev->reg_state == NETREG_UNINITIALIZED)
 334                         return;
 335                 linkwatch_fire_event(dev);
 336         }
 337 }
 338 EXPORT_SYMBOL(netif_carrier_off);
 339
 340 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 341    under all circumstances. It is difficult to invent anything faster or
 342    cheaper.
 343  */
 344
 345 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 346 {
 347         kfree_skb(skb);
 348         return NET_XMIT_CN;
 349 }
 350
 351 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 352 {
 353         return NULL;
 354 }
 355
 356 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 357         .id             =       "noop",
 358         .priv_size      =       0,
 359         .enqueue        =       noop_enqueue,
 360         .dequeue        =       noop_dequeue,
 361         .peek           =       noop_dequeue,
 362         .owner          =       THIS_MODULE,
 363 };
 364
 365 static struct netdev_queue noop_netdev_queue = {
 366         .qdisc          =       &noop_qdisc,
 367         .qdisc_sleeping =       &noop_qdisc,
 368 };
 369
 370 struct Qdisc noop_qdisc = {
 371         .enqueue        =       noop_enqueue,
 372         .dequeue        =       noop_dequeue,
 373         .flags          =       TCQ_F_BUILTIN,
 374         .ops            =       &noop_qdisc_ops,
 375         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 376         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 377         .dev_queue      =       &noop_netdev_queue,
 378 };
 379 EXPORT_SYMBOL(noop_qdisc);
 380
 381 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 382         .id             =       "noqueue",
 383         .priv_size      =       0,
 384         .enqueue        =       noop_enqueue,
 385         .dequeue        =       noop_dequeue,
 386         .peek           =       noop_dequeue,
 387         .owner          =       THIS_MODULE,
 388 };
 389
 390 static struct Qdisc noqueue_qdisc;
 391 static struct netdev_queue noqueue_netdev_queue = {
 392         .qdisc          =       &noqueue_qdisc,
 393         .qdisc_sleeping =       &noqueue_qdisc,
 394 };
 395
 396 static struct Qdisc noqueue_qdisc = {
 397         .enqueue        =       NULL,
 398         .dequeue        =       noop_dequeue,
 399         .flags          =       TCQ_F_BUILTIN,
 400         .ops            =       &noqueue_qdisc_ops,
 401         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 402         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 403         .dev_queue      =       &noqueue_netdev_queue,
 404 };
 405
 406
 407 static const u8 prio2band[TC_PRIO_MAX+1] =
 408         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 409
 410 /* 3-band FIFO queue: old style, but should be a bit faster than
 411    generic prio+fifo combination.
 412  */
 413
 414 #define PFIFO_FAST_BANDS 3
 415
 416 /*
 417  * Private data for a pfifo_fast scheduler containing:
 418  *      - queues for the three band
 419  *      - bitmap indicating which of the bands contain skbs
 420  */
 421 struct pfifo_fast_priv {
 422         u32 bitmap;
 423         struct sk_buff_head q[PFIFO_FAST_BANDS];
 424 };
 425
 426 /*
 427  * Convert a bitmap to the first band number where an skb is queued, where:
 428  *      bitmap=0 means there are no skbs on any band.
 429  *      bitmap=1 means there is an skb on band 0.
 430  *      bitmap=7 means there are skbs on all 3 bands, etc.
 431  */
 432 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 433
 434 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 435                                              int band)
 436 {
 437         return priv->q + band;
 438 }
 439
 440 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 441 {
 442         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 443                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 444                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 445                 struct sk_buff_head *list = band2list(priv, band);
 446
 447                 priv->bitmap |= (1 << band);
 448                 qdisc->q.qlen++;
 449                 return __qdisc_enqueue_tail(skb, qdisc, list);
 450         }
 451
 452         return qdisc_drop(skb, qdisc);
 453 }
 454
 455 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 456 {
 457         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 458         int band = bitmap2band[priv->bitmap];
 459
 460         if (likely(band >= 0)) {
 461                 struct sk_buff_head *list = band2list(priv, band);
 462                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 463
 464                 qdisc->q.qlen--;
 465                 if (skb_queue_empty(list))
 466                         priv->bitmap &= ~(1 << band);
 467
 468                 return skb;
 469         }
 470
 471         return NULL;
 472 }
 473
 474 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
 475 {
 476         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 477         int band = bitmap2band[priv->bitmap];
 478
 479         if (band >= 0) {
 480                 struct sk_buff_head *list = band2list(priv, band);
 481
 482                 return skb_peek(list);
 483         }
 484
 485         return NULL;
 486 }
 487
 488 static void pfifo_fast_reset(struct Qdisc* qdisc)
 489 {
 490         int prio;
 491         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 492
 493         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 494                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 495
 496         priv->bitmap = 0;
 497         qdisc->qstats.backlog = 0;
 498         qdisc->q.qlen = 0;
 499 }
 500
 501 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 502 {
 503         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 504
 505         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 506         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 507         return skb->len;
 508
 509 nla_put_failure:
 510         return -1;
 511 }
 512
 513 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 514 {
 515         int prio;
 516         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 517
 518         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 519                 skb_queue_head_init(band2list(priv, prio));
 520
 521         return 0;
 522 }
 523
 524 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 525         .id             =       "pfifo_fast",
 526         .priv_size      =       sizeof(struct pfifo_fast_priv),
 527         .enqueue        =       pfifo_fast_enqueue,
 528         .dequeue        =       pfifo_fast_dequeue,
 529         .peek           =       pfifo_fast_peek,
 530         .init           =       pfifo_fast_init,
 531         .reset          =       pfifo_fast_reset,
 532         .dump           =       pfifo_fast_dump,
 533         .owner          =       THIS_MODULE,
 534 };
 535
 536 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 537                           struct Qdisc_ops *ops)
 538 {
 539         void *p;
 540         struct Qdisc *sch;
 541         unsigned int size;
 542         int err = -ENOBUFS;
 543
 544         /* ensure that the Qdisc and the private data are 32-byte aligned */
 545         size = QDISC_ALIGN(sizeof(*sch));
 546         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 547
 548         p = kzalloc(size, GFP_KERNEL);
 549         if (!p)
 550                 goto errout;
 551         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 552         sch->padded = (char *) sch - (char *) p;
 553
 554         INIT_LIST_HEAD(&sch->list);
 555         skb_queue_head_init(&sch->q);
 556         sch->ops = ops;
 557         sch->enqueue = ops->enqueue;
 558         sch->dequeue = ops->dequeue;
 559         sch->dev_queue = dev_queue;
 560         dev_hold(qdisc_dev(sch));
 561         atomic_set(&sch->refcnt, 1);
 562
 563         return sch;
 564 errout:
 565         return ERR_PTR(err);
 566 }
 567
 568 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
 569                                  struct netdev_queue *dev_queue,
 570                                  struct Qdisc_ops *ops,
 571                                  unsigned int parentid)
 572 {
 573         struct Qdisc *sch;
 574
 575         sch = qdisc_alloc(dev_queue, ops);
 576         if (IS_ERR(sch))
 577                 goto errout;
 578         sch->parent = parentid;
 579
 580         if (!ops->init || ops->init(sch, NULL) == 0)
 581                 return sch;
 582
 583         qdisc_destroy(sch);
 584 errout:
 585         return NULL;
 586 }
 587 EXPORT_SYMBOL(qdisc_create_dflt);
 588
 589 /* Under qdisc_lock(qdisc) and BH! */
 590
 591 void qdisc_reset(struct Qdisc *qdisc)
 592 {
 593         const struct Qdisc_ops *ops = qdisc->ops;
 594
 595         if (ops->reset)
 596                 ops->reset(qdisc);
 597
 598         if (qdisc->gso_skb) {
 599                 kfree_skb(qdisc->gso_skb);
 600                 qdisc->gso_skb = NULL;
 601                 qdisc->q.qlen = 0;
 602         }
 603 }
 604 EXPORT_SYMBOL(qdisc_reset);
 605
 606 void qdisc_destroy(struct Qdisc *qdisc)
 607 {
 608         const struct Qdisc_ops  *ops = qdisc->ops;
 609
 610         if (qdisc->flags & TCQ_F_BUILTIN ||
 611             !atomic_dec_and_test(&qdisc->refcnt))
 612                 return;
 613
 614 #ifdef CONFIG_NET_SCHED
 615         qdisc_list_del(qdisc);
 616
 617         qdisc_put_stab(qdisc->stab);
 618 #endif
 619         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 620         if (ops->reset)
 621                 ops->reset(qdisc);
 622         if (ops->destroy)
 623                 ops->destroy(qdisc);
 624
 625         module_put(ops->owner);
 626         dev_put(qdisc_dev(qdisc));
 627
 628         kfree_skb(qdisc->gso_skb);
 629         kfree((char *) qdisc - qdisc->padded);
 630 }
 631 EXPORT_SYMBOL(qdisc_destroy);
 632
 633 /* Attach toplevel qdisc to device queue. */
 634 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 635                               struct Qdisc *qdisc)
 636 {
 637         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 638         spinlock_t *root_lock;
 639
 640         root_lock = qdisc_lock(oqdisc);
 641         spin_lock_bh(root_lock);
 642
 643         /* Prune old scheduler */
 644         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 645                 qdisc_reset(oqdisc);
 646
 647         /* ... and graft new one */
 648         if (qdisc == NULL)
 649                 qdisc = &noop_qdisc;
 650         dev_queue->qdisc_sleeping = qdisc;
 651         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 652
 653         spin_unlock_bh(root_lock);
 654
 655         return oqdisc;
 656 }
 657
 658 static void attach_one_default_qdisc(struct net_device *dev,
 659                                      struct netdev_queue *dev_queue,
 660                                      void *_unused)
 661 {
 662         struct Qdisc *qdisc;
 663
 664         if (dev->tx_queue_len) {
 665                 qdisc = qdisc_create_dflt(dev, dev_queue,
 666                                           &pfifo_fast_ops, TC_H_ROOT);
 667                 if (!qdisc) {
 668                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 669                         return;
 670                 }
 671
 672                 /* Can by-pass the queue discipline for default qdisc */
 673                 qdisc->flags |= TCQ_F_CAN_BYPASS;
 674         } else {
 675                 qdisc =  &noqueue_qdisc;
 676         }
 677         dev_queue->qdisc_sleeping = qdisc;
 678 }
 679
 680 static void attach_default_qdiscs(struct net_device *dev)
 681 {
 682         struct netdev_queue *txq;
 683         struct Qdisc *qdisc;
 684
 685         txq = netdev_get_tx_queue(dev, 0);
 686
 687         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 688                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 689                 dev->qdisc = txq->qdisc_sleeping;
 690                 atomic_inc(&dev->qdisc->refcnt);
 691         } else {
 692                 qdisc = qdisc_create_dflt(dev, txq, &mq_qdisc_ops, TC_H_ROOT);
 693                 if (qdisc) {
 694                         qdisc->ops->attach(qdisc);
 695                         dev->qdisc = qdisc;
 696                 }
 697         }
 698 }
 699
 700 static void transition_one_qdisc(struct net_device *dev,
 701                                  struct netdev_queue *dev_queue,
 702                                  void *_need_watchdog)
 703 {
 704         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 705         int *need_watchdog_p = _need_watchdog;
 706
 707         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 708                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 709
 710         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 711         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 712                 dev_queue->trans_start = 0;
 713                 *need_watchdog_p = 1;
 714         }
 715 }
 716
 717 void dev_activate(struct net_device *dev)
 718 {
 719         int need_watchdog;
 720
 721         /* No queueing discipline is attached to device;
 722            create default one i.e. pfifo_fast for devices,
 723            which need queueing and noqueue_qdisc for
 724            virtual interfaces
 725          */
 726
 727         if (dev->qdisc == &noop_qdisc)
 728                 attach_default_qdiscs(dev);
 729
 730         if (!netif_carrier_ok(dev))
 731                 /* Delay activation until next carrier-on event */
 732                 return;
 733
 734         need_watchdog = 0;
 735         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 736         transition_one_qdisc(dev, &dev->rx_queue, NULL);
 737
 738         if (need_watchdog) {
 739                 dev->trans_start = jiffies;
 740                 dev_watchdog_up(dev);
 741         }
 742 }
 743
 744 static void dev_deactivate_queue(struct net_device *dev,
 745                                  struct netdev_queue *dev_queue,
 746                                  void *_qdisc_default)
 747 {
 748         struct Qdisc *qdisc_default = _qdisc_default;
 749         struct Qdisc *qdisc;
 750
 751         qdisc = dev_queue->qdisc;
 752         if (qdisc) {
 753                 spin_lock_bh(qdisc_lock(qdisc));
 754
 755                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 756                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 757
 758                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 759                 qdisc_reset(qdisc);
 760
 761                 spin_unlock_bh(qdisc_lock(qdisc));
 762         }
 763 }
 764
 765 static bool some_qdisc_is_busy(struct net_device *dev)
 766 {
 767         unsigned int i;
 768
 769         for (i = 0; i < dev->num_tx_queues; i++) {
 770                 struct netdev_queue *dev_queue;
 771                 spinlock_t *root_lock;
 772                 struct Qdisc *q;
 773                 int val;
 774
 775                 dev_queue = netdev_get_tx_queue(dev, i);
 776                 q = dev_queue->qdisc_sleeping;
 777                 root_lock = qdisc_lock(q);
 778
 779                 spin_lock_bh(root_lock);
 780
 781                 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
 782                        test_bit(__QDISC_STATE_SCHED, &q->state));
 783
 784                 spin_unlock_bh(root_lock);
 785
 786                 if (val)
 787                         return true;
 788         }
 789         return false;
 790 }
 791
 792 void dev_deactivate(struct net_device *dev)
 793 {
 794         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 795         dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
 796
 797         dev_watchdog_down(dev);
 798
 799         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 800         synchronize_rcu();
 801
 802         /* Wait for outstanding qdisc_run calls. */
 803         while (some_qdisc_is_busy(dev))
 804                 yield();
 805 }
 806
 807 static void dev_init_scheduler_queue(struct net_device *dev,
 808                                      struct netdev_queue *dev_queue,
 809                                      void *_qdisc)
 810 {
 811         struct Qdisc *qdisc = _qdisc;
 812
 813         dev_queue->qdisc = qdisc;
 814         dev_queue->qdisc_sleeping = qdisc;
 815 }
 816
 817 void dev_init_scheduler(struct net_device *dev)
 818 {
 819         dev->qdisc = &noop_qdisc;
 820         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 821         dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 822
 823         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 824 }
 825
 826 static void shutdown_scheduler_queue(struct net_device *dev,
 827                                      struct netdev_queue *dev_queue,
 828                                      void *_qdisc_default)
 829 {
 830         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 831         struct Qdisc *qdisc_default = _qdisc_default;
 832
 833         if (qdisc) {
 834                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 835                 dev_queue->qdisc_sleeping = qdisc_default;
 836
 837                 qdisc_destroy(qdisc);
 838         }
 839 }
 840
 841 void dev_shutdown(struct net_device *dev)
 842 {
 843         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 844         shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 845         qdisc_destroy(dev->qdisc);
 846         dev->qdisc = &noop_qdisc;
 847
 848         WARN_ON(timer_pending(&dev->watchdog_timer));
 849 }