net/sched/sch_generic.c

   1 /*
   2  * net/sched/sch_generic.c      Generic packet scheduler routines.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
  11  *              - Ingress support
  12  */
  13
  14 #include <linux/bitops.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/string.h>
  20 #include <linux/errno.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/rtnetlink.h>
  24 #include <linux/init.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/list.h>
  27 #include <net/pkt_sched.h>
  28
  29 /* Main transmission queue. */
  30
  31 /* Modifications to data participating in scheduling must be protected with
  32  * qdisc_lock(qdisc) spinlock.
  33  *
  34  * The idea is the following:
  35  * - enqueue, dequeue are serialized via qdisc root lock
  36  * - ingress filtering is also serialized via qdisc root lock
  37  * - updates to tree and tree walking are only done under the rtnl mutex.
  38  */
  39
  40 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  41 {
  42         q->gso_skb = skb;
  43         q->qstats.requeues++;
  44         q->q.qlen++;    /* it's still part of the queue */
  45         __netif_schedule(q);
  46
  47         return 0;
  48 }
  49
  50 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  51 {
  52         struct sk_buff *skb = q->gso_skb;
  53
  54         if (unlikely(skb)) {
  55                 struct net_device *dev = qdisc_dev(q);
  56                 struct netdev_queue *txq;
  57
  58                 /* check the reason of requeuing without tx lock first */
  59                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
  60                 if (!netif_tx_queue_stopped(txq) &&
  61                     !netif_tx_queue_frozen(txq)) {
  62                         q->gso_skb = NULL;
  63                         q->q.qlen--;
  64                 } else
  65                         skb = NULL;
  66         } else {
  67                 skb = q->dequeue(q);
  68         }
  69
  70         return skb;
  71 }
  72
  73 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  74                                            struct netdev_queue *dev_queue,
  75                                            struct Qdisc *q)
  76 {
  77         int ret;
  78
  79         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  80                 /*
  81                  * Same CPU holding the lock. It may be a transient
  82                  * configuration error, when hard_start_xmit() recurses. We
  83                  * detect it by checking xmit owner and drop the packet when
  84                  * deadloop is detected. Return OK to try the next skb.
  85                  */
  86                 kfree_skb(skb);
  87                 if (net_ratelimit())
  88                         printk(KERN_WARNING "Dead loop on netdevice %s, "
  89                                "fix it urgently!\n", dev_queue->dev->name);
  90                 ret = qdisc_qlen(q);
  91         } else {
  92                 /*
  93                  * Another cpu is holding lock, requeue & delay xmits for
  94                  * some time.
  95                  */
  96                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
  97                 ret = dev_requeue_skb(skb, q);
  98         }
  99
 100         return ret;
 101 }
 102
 103 /*
 104  * Transmit one skb, and handle the return status as required. Holding the
 105  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 106  * function.
 107  *
 108  * Returns to the caller:
 109  *                              0  - queue is empty or throttled.
 110  *                              >0 - queue is not empty.
 111  */
 112 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 113                     struct net_device *dev, struct netdev_queue *txq,
 114                     spinlock_t *root_lock)
 115 {
 116         int ret = NETDEV_TX_BUSY;
 117
 118         /* And release qdisc */
 119         spin_unlock(root_lock);
 120
 121         HARD_TX_LOCK(dev, txq, smp_processor_id());
 122         if (!netif_tx_queue_stopped(txq) &&
 123             !netif_tx_queue_frozen(txq))
 124                 ret = dev_hard_start_xmit(skb, dev, txq);
 125         HARD_TX_UNLOCK(dev, txq);
 126
 127         spin_lock(root_lock);
 128
 129         switch (ret) {
 130         case NETDEV_TX_OK:
 131                 /* Driver sent out skb successfully */
 132                 ret = qdisc_qlen(q);
 133                 break;
 134
 135         case NETDEV_TX_LOCKED:
 136                 /* Driver try lock failed */
 137                 ret = handle_dev_cpu_collision(skb, txq, q);
 138                 break;
 139
 140         default:
 141                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 142                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 143                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 144                                dev->name, ret, q->q.qlen);
 145
 146                 ret = dev_requeue_skb(skb, q);
 147                 break;
 148         }
 149
 150         if (ret && (netif_tx_queue_stopped(txq) ||
 151                     netif_tx_queue_frozen(txq)))
 152                 ret = 0;
 153
 154         return ret;
 155 }
 156
 157 /*
 158  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 159  *
 160  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 161  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 162  * this queue.
 163  *
 164  *  netif_tx_lock serializes accesses to device driver.
 165  *
 166  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 167  *  if one is grabbed, another must be free.
 168  *
 169  * Note, that this procedure can be called by a watchdog timer
 170  *
 171  * Returns to the caller:
 172  *                              0  - queue is empty or throttled.
 173  *                              >0 - queue is not empty.
 174  *
 175  */
 176 static inline int qdisc_restart(struct Qdisc *q)
 177 {
 178         struct netdev_queue *txq;
 179         struct net_device *dev;
 180         spinlock_t *root_lock;
 181         struct sk_buff *skb;
 182
 183         /* Dequeue packet */
 184         skb = dequeue_skb(q);
 185         if (unlikely(!skb))
 186                 return 0;
 187
 188         root_lock = qdisc_lock(q);
 189         dev = qdisc_dev(q);
 190         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 191
 192         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 193 }
 194
 195 void __qdisc_run(struct Qdisc *q)
 196 {
 197         unsigned long start_time = jiffies;
 198
 199         while (qdisc_restart(q)) {
 200                 /*
 201                  * Postpone processing if
 202                  * 1. another process needs the CPU;
 203                  * 2. we've been doing it for too long.
 204                  */
 205                 if (need_resched() || jiffies != start_time) {
 206                         __netif_schedule(q);
 207                         break;
 208                 }
 209         }
 210
 211         clear_bit(__QDISC_STATE_RUNNING, &q->state);
 212 }
 213
 214 unsigned long dev_trans_start(struct net_device *dev)
 215 {
 216         unsigned long val, res = dev->trans_start;
 217         unsigned int i;
 218
 219         for (i = 0; i < dev->num_tx_queues; i++) {
 220                 val = netdev_get_tx_queue(dev, i)->trans_start;
 221                 if (val && time_after(val, res))
 222                         res = val;
 223         }
 224         dev->trans_start = res;
 225         return res;
 226 }
 227 EXPORT_SYMBOL(dev_trans_start);
 228
 229 static void dev_watchdog(unsigned long arg)
 230 {
 231         struct net_device *dev = (struct net_device *)arg;
 232
 233         netif_tx_lock(dev);
 234         if (!qdisc_tx_is_noop(dev)) {
 235                 if (netif_device_present(dev) &&
 236                     netif_running(dev) &&
 237                     netif_carrier_ok(dev)) {
 238                         int some_queue_timedout = 0;
 239                         unsigned int i;
 240                         unsigned long trans_start;
 241
 242                         for (i = 0; i < dev->num_tx_queues; i++) {
 243                                 struct netdev_queue *txq;
 244
 245                                 txq = netdev_get_tx_queue(dev, i);
 246                                 /*
 247                                  * old device drivers set dev->trans_start
 248                                  */
 249                                 trans_start = txq->trans_start ? : dev->trans_start;
 250                                 if (netif_tx_queue_stopped(txq) &&
 251                                     time_after(jiffies, (trans_start +
 252                                                          dev->watchdog_timeo))) {
 253                                         some_queue_timedout = 1;
 254                                         break;
 255                                 }
 256                         }
 257
 258                         if (some_queue_timedout) {
 259                                 char drivername[64];
 260                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 261                                        dev->name, netdev_drivername(dev, drivername, 64), i);
 262                                 dev->netdev_ops->ndo_tx_timeout(dev);
 263                         }
 264                         if (!mod_timer(&dev->watchdog_timer,
 265                                        round_jiffies(jiffies +
 266                                                      dev->watchdog_timeo)))
 267                                 dev_hold(dev);
 268                 }
 269         }
 270         netif_tx_unlock(dev);
 271
 272         dev_put(dev);
 273 }
 274
 275 void __netdev_watchdog_up(struct net_device *dev)
 276 {
 277         if (dev->netdev_ops->ndo_tx_timeout) {
 278                 if (dev->watchdog_timeo <= 0)
 279                         dev->watchdog_timeo = 5*HZ;
 280                 if (!mod_timer(&dev->watchdog_timer,
 281                                round_jiffies(jiffies + dev->watchdog_timeo)))
 282                         dev_hold(dev);
 283         }
 284 }
 285
 286 static void dev_watchdog_up(struct net_device *dev)
 287 {
 288         __netdev_watchdog_up(dev);
 289 }
 290
 291 static void dev_watchdog_down(struct net_device *dev)
 292 {
 293         netif_tx_lock_bh(dev);
 294         if (del_timer(&dev->watchdog_timer))
 295                 dev_put(dev);
 296         netif_tx_unlock_bh(dev);
 297 }
 298
 299 /**
 300  *      netif_carrier_on - set carrier
 301  *      @dev: network device
 302  *
 303  * Device has detected that carrier.
 304  */
 305 void netif_carrier_on(struct net_device *dev)
 306 {
 307         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 308                 if (dev->reg_state == NETREG_UNINITIALIZED)
 309                         return;
 310                 linkwatch_fire_event(dev);
 311                 if (netif_running(dev))
 312                         __netdev_watchdog_up(dev);
 313         }
 314 }
 315 EXPORT_SYMBOL(netif_carrier_on);
 316
 317 /**
 318  *      netif_carrier_off - clear carrier
 319  *      @dev: network device
 320  *
 321  * Device has detected loss of carrier.
 322  */
 323 void netif_carrier_off(struct net_device *dev)
 324 {
 325         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 326                 if (dev->reg_state == NETREG_UNINITIALIZED)
 327                         return;
 328                 linkwatch_fire_event(dev);
 329         }
 330 }
 331 EXPORT_SYMBOL(netif_carrier_off);
 332
 333 /**
 334  *      netif_notify_peers - notify network peers about existence of @dev
 335  *      @dev: network device
 336  *
 337  * Generate traffic such that interested network peers are aware of
 338  * @dev, such as by generating a gratuitous ARP. This may be used when
 339  * a device wants to inform the rest of the network about some sort of
 340  * reconfiguration such as a failover event or virtual machine
 341  * migration.
 342  */
 343 void netif_notify_peers(struct net_device *dev)
 344 {
 345         rtnl_lock();
 346         call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
 347         rtnl_unlock();
 348 }
 349 EXPORT_SYMBOL(netif_notify_peers);
 350
 351 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 352    under all circumstances. It is difficult to invent anything faster or
 353    cheaper.
 354  */
 355
 356 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 357 {
 358         kfree_skb(skb);
 359         return NET_XMIT_CN;
 360 }
 361
 362 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 363 {
 364         return NULL;
 365 }
 366
 367 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 368         .id             =       "noop",
 369         .priv_size      =       0,
 370         .enqueue        =       noop_enqueue,
 371         .dequeue        =       noop_dequeue,
 372         .peek           =       noop_dequeue,
 373         .owner          =       THIS_MODULE,
 374 };
 375
 376 static struct netdev_queue noop_netdev_queue = {
 377         .qdisc          =       &noop_qdisc,
 378         .qdisc_sleeping =       &noop_qdisc,
 379 };
 380
 381 struct Qdisc noop_qdisc = {
 382         .enqueue        =       noop_enqueue,
 383         .dequeue        =       noop_dequeue,
 384         .flags          =       TCQ_F_BUILTIN,
 385         .ops            =       &noop_qdisc_ops,
 386         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 387         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 388         .dev_queue      =       &noop_netdev_queue,
 389 };
 390 EXPORT_SYMBOL(noop_qdisc);
 391
 392 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 393         .id             =       "noqueue",
 394         .priv_size      =       0,
 395         .enqueue        =       noop_enqueue,
 396         .dequeue        =       noop_dequeue,
 397         .peek           =       noop_dequeue,
 398         .owner          =       THIS_MODULE,
 399 };
 400
 401 static struct Qdisc noqueue_qdisc;
 402 static struct netdev_queue noqueue_netdev_queue = {
 403         .qdisc          =       &noqueue_qdisc,
 404         .qdisc_sleeping =       &noqueue_qdisc,
 405 };
 406
 407 static struct Qdisc noqueue_qdisc = {
 408         .enqueue        =       NULL,
 409         .dequeue        =       noop_dequeue,
 410         .flags          =       TCQ_F_BUILTIN,
 411         .ops            =       &noqueue_qdisc_ops,
 412         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 413         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 414         .dev_queue      =       &noqueue_netdev_queue,
 415 };
 416
 417
 418 static const u8 prio2band[TC_PRIO_MAX+1] =
 419         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 420
 421 /* 3-band FIFO queue: old style, but should be a bit faster than
 422    generic prio+fifo combination.
 423  */
 424
 425 #define PFIFO_FAST_BANDS 3
 426
 427 /*
 428  * Private data for a pfifo_fast scheduler containing:
 429  *      - queues for the three band
 430  *      - bitmap indicating which of the bands contain skbs
 431  */
 432 struct pfifo_fast_priv {
 433         u32 bitmap;
 434         struct sk_buff_head q[PFIFO_FAST_BANDS];
 435 };
 436
 437 /*
 438  * Convert a bitmap to the first band number where an skb is queued, where:
 439  *      bitmap=0 means there are no skbs on any band.
 440  *      bitmap=1 means there is an skb on band 0.
 441  *      bitmap=7 means there are skbs on all 3 bands, etc.
 442  */
 443 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 444
 445 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 446                                              int band)
 447 {
 448         return priv->q + band;
 449 }
 450
 451 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 452 {
 453         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 454                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 455                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 456                 struct sk_buff_head *list = band2list(priv, band);
 457
 458                 priv->bitmap |= (1 << band);
 459                 qdisc->q.qlen++;
 460                 return __qdisc_enqueue_tail(skb, qdisc, list);
 461         }
 462
 463         return qdisc_drop(skb, qdisc);
 464 }
 465
 466 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 467 {
 468         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 469         int band = bitmap2band[priv->bitmap];
 470
 471         if (likely(band >= 0)) {
 472                 struct sk_buff_head *list = band2list(priv, band);
 473                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 474
 475                 qdisc->q.qlen--;
 476                 if (skb_queue_empty(list))
 477                         priv->bitmap &= ~(1 << band);
 478
 479                 return skb;
 480         }
 481
 482         return NULL;
 483 }
 484
 485 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
 486 {
 487         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 488         int band = bitmap2band[priv->bitmap];
 489
 490         if (band >= 0) {
 491                 struct sk_buff_head *list = band2list(priv, band);
 492
 493                 return skb_peek(list);
 494         }
 495
 496         return NULL;
 497 }
 498
 499 static void pfifo_fast_reset(struct Qdisc* qdisc)
 500 {
 501         int prio;
 502         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 503
 504         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 505                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 506
 507         priv->bitmap = 0;
 508         qdisc->qstats.backlog = 0;
 509         qdisc->q.qlen = 0;
 510 }
 511
 512 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 513 {
 514         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 515
 516         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 517         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 518         return skb->len;
 519
 520 nla_put_failure:
 521         return -1;
 522 }
 523
 524 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 525 {
 526         int prio;
 527         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 528
 529         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 530                 skb_queue_head_init(band2list(priv, prio));
 531
 532         return 0;
 533 }
 534
 535 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 536         .id             =       "pfifo_fast",
 537         .priv_size      =       sizeof(struct pfifo_fast_priv),
 538         .enqueue        =       pfifo_fast_enqueue,
 539         .dequeue        =       pfifo_fast_dequeue,
 540         .peek           =       pfifo_fast_peek,
 541         .init           =       pfifo_fast_init,
 542         .reset          =       pfifo_fast_reset,
 543         .dump           =       pfifo_fast_dump,
 544         .owner          =       THIS_MODULE,
 545 };
 546
 547 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 548                           struct Qdisc_ops *ops)
 549 {
 550         void *p;
 551         struct Qdisc *sch;
 552         unsigned int size;
 553         int err = -ENOBUFS;
 554
 555         /* ensure that the Qdisc and the private data are 32-byte aligned */
 556         size = QDISC_ALIGN(sizeof(*sch));
 557         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 558
 559         p = kzalloc(size, GFP_KERNEL);
 560         if (!p)
 561                 goto errout;
 562         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 563         sch->padded = (char *) sch - (char *) p;
 564
 565         INIT_LIST_HEAD(&sch->list);
 566         skb_queue_head_init(&sch->q);
 567         sch->ops = ops;
 568         sch->enqueue = ops->enqueue;
 569         sch->dequeue = ops->dequeue;
 570         sch->dev_queue = dev_queue;
 571         dev_hold(qdisc_dev(sch));
 572         atomic_set(&sch->refcnt, 1);
 573
 574         return sch;
 575 errout:
 576         return ERR_PTR(err);
 577 }
 578
 579 struct Qdisc * qdisc_create_dflt(struct net_device *dev,
 580                                  struct netdev_queue *dev_queue,
 581                                  struct Qdisc_ops *ops,
 582                                  unsigned int parentid)
 583 {
 584         struct Qdisc *sch;
 585
 586         sch = qdisc_alloc(dev_queue, ops);
 587         if (IS_ERR(sch))
 588                 goto errout;
 589         sch->parent = parentid;
 590
 591         if (!ops->init || ops->init(sch, NULL) == 0)
 592                 return sch;
 593
 594         qdisc_destroy(sch);
 595 errout:
 596         return NULL;
 597 }
 598 EXPORT_SYMBOL(qdisc_create_dflt);
 599
 600 /* Under qdisc_lock(qdisc) and BH! */
 601
 602 void qdisc_reset(struct Qdisc *qdisc)
 603 {
 604         const struct Qdisc_ops *ops = qdisc->ops;
 605
 606         if (ops->reset)
 607                 ops->reset(qdisc);
 608
 609         if (qdisc->gso_skb) {
 610                 kfree_skb(qdisc->gso_skb);
 611                 qdisc->gso_skb = NULL;
 612                 qdisc->q.qlen = 0;
 613         }
 614 }
 615 EXPORT_SYMBOL(qdisc_reset);
 616
 617 void qdisc_destroy(struct Qdisc *qdisc)
 618 {
 619         const struct Qdisc_ops  *ops = qdisc->ops;
 620
 621         if (qdisc->flags & TCQ_F_BUILTIN ||
 622             !atomic_dec_and_test(&qdisc->refcnt))
 623                 return;
 624
 625 #ifdef CONFIG_NET_SCHED
 626         qdisc_list_del(qdisc);
 627
 628         qdisc_put_stab(qdisc->stab);
 629 #endif
 630         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 631         if (ops->reset)
 632                 ops->reset(qdisc);
 633         if (ops->destroy)
 634                 ops->destroy(qdisc);
 635
 636         module_put(ops->owner);
 637         dev_put(qdisc_dev(qdisc));
 638
 639         kfree_skb(qdisc->gso_skb);
 640         kfree((char *) qdisc - qdisc->padded);
 641 }
 642 EXPORT_SYMBOL(qdisc_destroy);
 643
 644 /* Attach toplevel qdisc to device queue. */
 645 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 646                               struct Qdisc *qdisc)
 647 {
 648         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 649         spinlock_t *root_lock;
 650
 651         root_lock = qdisc_lock(oqdisc);
 652         spin_lock_bh(root_lock);
 653
 654         /* Prune old scheduler */
 655         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 656                 qdisc_reset(oqdisc);
 657
 658         /* ... and graft new one */
 659         if (qdisc == NULL)
 660                 qdisc = &noop_qdisc;
 661         dev_queue->qdisc_sleeping = qdisc;
 662         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 663
 664         spin_unlock_bh(root_lock);
 665
 666         return oqdisc;
 667 }
 668
 669 static void attach_one_default_qdisc(struct net_device *dev,
 670                                      struct netdev_queue *dev_queue,
 671                                      void *_unused)
 672 {
 673         struct Qdisc *qdisc;
 674
 675         if (dev->tx_queue_len) {
 676                 qdisc = qdisc_create_dflt(dev, dev_queue,
 677                                           &pfifo_fast_ops, TC_H_ROOT);
 678                 if (!qdisc) {
 679                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 680                         return;
 681                 }
 682
 683                 /* Can by-pass the queue discipline for default qdisc */
 684                 qdisc->flags |= TCQ_F_CAN_BYPASS;
 685         } else {
 686                 qdisc =  &noqueue_qdisc;
 687         }
 688         dev_queue->qdisc_sleeping = qdisc;
 689 }
 690
 691 static void attach_default_qdiscs(struct net_device *dev)
 692 {
 693         struct netdev_queue *txq;
 694         struct Qdisc *qdisc;
 695
 696         txq = netdev_get_tx_queue(dev, 0);
 697
 698         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 699                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 700                 dev->qdisc = txq->qdisc_sleeping;
 701                 atomic_inc(&dev->qdisc->refcnt);
 702         } else {
 703                 qdisc = qdisc_create_dflt(dev, txq, &mq_qdisc_ops, TC_H_ROOT);
 704                 if (qdisc) {
 705                         qdisc->ops->attach(qdisc);
 706                         dev->qdisc = qdisc;
 707                 }
 708         }
 709 }
 710
 711 static void transition_one_qdisc(struct net_device *dev,
 712                                  struct netdev_queue *dev_queue,
 713                                  void *_need_watchdog)
 714 {
 715         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 716         int *need_watchdog_p = _need_watchdog;
 717
 718         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 719                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 720
 721         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 722         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 723                 dev_queue->trans_start = 0;
 724                 *need_watchdog_p = 1;
 725         }
 726 }
 727
 728 void dev_activate(struct net_device *dev)
 729 {
 730         int need_watchdog;
 731
 732         /* No queueing discipline is attached to device;
 733            create default one i.e. pfifo_fast for devices,
 734            which need queueing and noqueue_qdisc for
 735            virtual interfaces
 736          */
 737
 738         if (dev->qdisc == &noop_qdisc)
 739                 attach_default_qdiscs(dev);
 740
 741         if (!netif_carrier_ok(dev))
 742                 /* Delay activation until next carrier-on event */
 743                 return;
 744
 745         need_watchdog = 0;
 746         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 747         transition_one_qdisc(dev, &dev->rx_queue, NULL);
 748
 749         if (need_watchdog) {
 750                 dev->trans_start = jiffies;
 751                 dev_watchdog_up(dev);
 752         }
 753 }
 754
 755 static void dev_deactivate_queue(struct net_device *dev,
 756                                  struct netdev_queue *dev_queue,
 757                                  void *_qdisc_default)
 758 {
 759         struct Qdisc *qdisc_default = _qdisc_default;
 760         struct Qdisc *qdisc;
 761
 762         qdisc = dev_queue->qdisc;
 763         if (qdisc) {
 764                 spin_lock_bh(qdisc_lock(qdisc));
 765
 766                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 767                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 768
 769                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 770                 qdisc_reset(qdisc);
 771
 772                 spin_unlock_bh(qdisc_lock(qdisc));
 773         }
 774 }
 775
 776 static bool some_qdisc_is_busy(struct net_device *dev)
 777 {
 778         unsigned int i;
 779
 780         for (i = 0; i < dev->num_tx_queues; i++) {
 781                 struct netdev_queue *dev_queue;
 782                 spinlock_t *root_lock;
 783                 struct Qdisc *q;
 784                 int val;
 785
 786                 dev_queue = netdev_get_tx_queue(dev, i);
 787                 q = dev_queue->qdisc_sleeping;
 788                 root_lock = qdisc_lock(q);
 789
 790                 spin_lock_bh(root_lock);
 791
 792                 val = (test_bit(__QDISC_STATE_RUNNING, &q->state) ||
 793                        test_bit(__QDISC_STATE_SCHED, &q->state));
 794
 795                 spin_unlock_bh(root_lock);
 796
 797                 if (val)
 798                         return true;
 799         }
 800         return false;
 801 }
 802
 803 void dev_deactivate(struct net_device *dev)
 804 {
 805         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 806         dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc);
 807
 808         dev_watchdog_down(dev);
 809
 810         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 811         synchronize_rcu();
 812
 813         /* Wait for outstanding qdisc_run calls. */
 814         while (some_qdisc_is_busy(dev))
 815                 yield();
 816 }
 817
 818 static void dev_init_scheduler_queue(struct net_device *dev,
 819                                      struct netdev_queue *dev_queue,
 820                                      void *_qdisc)
 821 {
 822         struct Qdisc *qdisc = _qdisc;
 823
 824         dev_queue->qdisc = qdisc;
 825         dev_queue->qdisc_sleeping = qdisc;
 826 }
 827
 828 void dev_init_scheduler(struct net_device *dev)
 829 {
 830         dev->qdisc = &noop_qdisc;
 831         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 832         dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 833
 834         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 835 }
 836
 837 static void shutdown_scheduler_queue(struct net_device *dev,
 838                                      struct netdev_queue *dev_queue,
 839                                      void *_qdisc_default)
 840 {
 841         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 842         struct Qdisc *qdisc_default = _qdisc_default;
 843
 844         if (qdisc) {
 845                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 846                 dev_queue->qdisc_sleeping = qdisc_default;
 847
 848                 qdisc_destroy(qdisc);
 849         }
 850 }
 851
 852 void dev_shutdown(struct net_device *dev)
 853 {
 854         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 855         shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc);
 856         qdisc_destroy(dev->qdisc);
 857         dev->qdisc = &noop_qdisc;
 858
 859         WARN_ON(timer_pending(&dev->watchdog_timer));
 860 }