net/sched/sch_generic.c

   1 /*
   2  * net/sched/sch_generic.c      Generic packet scheduler routines.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
  11  *              - Ingress support
  12  */
  13
  14 #include <linux/bitops.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/string.h>
  20 #include <linux/errno.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/rtnetlink.h>
  24 #include <linux/init.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/list.h>
  27 #include <linux/slab.h>
  28 #include <net/pkt_sched.h>
  29 #include <net/dst.h>
  30
  31 /* Main transmission queue. */
  32
  33 /* Modifications to data participating in scheduling must be protected with
  34  * qdisc_lock(qdisc) spinlock.
  35  *
  36  * The idea is the following:
  37  * - enqueue, dequeue are serialized via qdisc root lock
  38  * - ingress filtering is also serialized via qdisc root lock
  39  * - updates to tree and tree walking are only done under the rtnl mutex.
  40  */
  41
  42 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  43 {
  44         skb_dst_force(skb);
  45         q->gso_skb = skb;
  46         q->qstats.requeues++;
  47         q->q.qlen++;    /* it's still part of the queue */
  48         __netif_schedule(q);
  49
  50         return 0;
  51 }
  52
  53 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  54 {
  55         struct sk_buff *skb = q->gso_skb;
  56
  57         if (unlikely(skb)) {
  58                 struct net_device *dev = qdisc_dev(q);
  59                 struct netdev_queue *txq;
  60
  61                 /* check the reason of requeuing without tx lock first */
  62                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
  63                 if (!netif_tx_queue_frozen_or_stopped(txq)) {
  64                         q->gso_skb = NULL;
  65                         q->q.qlen--;
  66                 } else
  67                         skb = NULL;
  68         } else {
  69                 skb = q->dequeue(q);
  70         }
  71
  72         return skb;
  73 }
  74
  75 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  76                                            struct netdev_queue *dev_queue,
  77                                            struct Qdisc *q)
  78 {
  79         int ret;
  80
  81         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  82                 /*
  83                  * Same CPU holding the lock. It may be a transient
  84                  * configuration error, when hard_start_xmit() recurses. We
  85                  * detect it by checking xmit owner and drop the packet when
  86                  * deadloop is detected. Return OK to try the next skb.
  87                  */
  88                 kfree_skb(skb);
  89                 if (net_ratelimit())
  90                         printk(KERN_WARNING "Dead loop on netdevice %s, "
  91                                "fix it urgently!\n", dev_queue->dev->name);
  92                 ret = qdisc_qlen(q);
  93         } else {
  94                 /*
  95                  * Another cpu is holding lock, requeue & delay xmits for
  96                  * some time.
  97                  */
  98                 __this_cpu_inc(softnet_data.cpu_collision);
  99                 ret = dev_requeue_skb(skb, q);
 100         }
 101
 102         return ret;
 103 }
 104
 105 /*
 106  * Transmit one skb, and handle the return status as required. Holding the
 107  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 108  * function.
 109  *
 110  * Returns to the caller:
 111  *                              0  - queue is empty or throttled.
 112  *                              >0 - queue is not empty.
 113  */
 114 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 115                     struct net_device *dev, struct netdev_queue *txq,
 116                     spinlock_t *root_lock)
 117 {
 118         int ret = NETDEV_TX_BUSY;
 119
 120         /* And release qdisc */
 121         spin_unlock(root_lock);
 122
 123         HARD_TX_LOCK(dev, txq, smp_processor_id());
 124         if (!netif_tx_queue_frozen_or_stopped(txq))
 125                 ret = dev_hard_start_xmit(skb, dev, txq);
 126
 127         HARD_TX_UNLOCK(dev, txq);
 128
 129         spin_lock(root_lock);
 130
 131         if (dev_xmit_complete(ret)) {
 132                 /* Driver sent out skb successfully or skb was consumed */
 133                 ret = qdisc_qlen(q);
 134         } else if (ret == NETDEV_TX_LOCKED) {
 135                 /* Driver try lock failed */
 136                 ret = handle_dev_cpu_collision(skb, txq, q);
 137         } else {
 138                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 139                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 140                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 141                                dev->name, ret, q->q.qlen);
 142
 143                 ret = dev_requeue_skb(skb, q);
 144         }
 145
 146         if (ret && netif_tx_queue_frozen_or_stopped(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         WARN_ON_ONCE(skb_dst_is_noref(skb));
 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         qdisc_run_end(q);
 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 /**
 329  *      netif_notify_peers - notify network peers about existence of @dev
 330  *      @dev: network device
 331  *
 332  * Generate traffic such that interested network peers are aware of
 333  * @dev, such as by generating a gratuitous ARP. This may be used when
 334  * a device wants to inform the rest of the network about some sort of
 335  * reconfiguration such as a failover event or virtual machine
 336  * migration.
 337  */
 338 void netif_notify_peers(struct net_device *dev)
 339 {
 340         rtnl_lock();
 341         call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
 342         rtnl_unlock();
 343 }
 344 EXPORT_SYMBOL(netif_notify_peers);
 345
 346 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 347    under all circumstances. It is difficult to invent anything faster or
 348    cheaper.
 349  */
 350
 351 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 352 {
 353         kfree_skb(skb);
 354         return NET_XMIT_CN;
 355 }
 356
 357 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 358 {
 359         return NULL;
 360 }
 361
 362 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 363         .id             =       "noop",
 364         .priv_size      =       0,
 365         .enqueue        =       noop_enqueue,
 366         .dequeue        =       noop_dequeue,
 367         .peek           =       noop_dequeue,
 368         .owner          =       THIS_MODULE,
 369 };
 370
 371 static struct netdev_queue noop_netdev_queue = {
 372         .qdisc          =       &noop_qdisc,
 373         .qdisc_sleeping =       &noop_qdisc,
 374 };
 375
 376 struct Qdisc noop_qdisc = {
 377         .enqueue        =       noop_enqueue,
 378         .dequeue        =       noop_dequeue,
 379         .flags          =       TCQ_F_BUILTIN,
 380         .ops            =       &noop_qdisc_ops,
 381         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 382         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 383         .dev_queue      =       &noop_netdev_queue,
 384         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 385 };
 386 EXPORT_SYMBOL(noop_qdisc);
 387
 388 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 389         .id             =       "noqueue",
 390         .priv_size      =       0,
 391         .enqueue        =       noop_enqueue,
 392         .dequeue        =       noop_dequeue,
 393         .peek           =       noop_dequeue,
 394         .owner          =       THIS_MODULE,
 395 };
 396
 397 static struct Qdisc noqueue_qdisc;
 398 static struct netdev_queue noqueue_netdev_queue = {
 399         .qdisc          =       &noqueue_qdisc,
 400         .qdisc_sleeping =       &noqueue_qdisc,
 401 };
 402
 403 static struct Qdisc noqueue_qdisc = {
 404         .enqueue        =       NULL,
 405         .dequeue        =       noop_dequeue,
 406         .flags          =       TCQ_F_BUILTIN,
 407         .ops            =       &noqueue_qdisc_ops,
 408         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 409         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 410         .dev_queue      =       &noqueue_netdev_queue,
 411         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 412 };
 413
 414
 415 static const u8 prio2band[TC_PRIO_MAX+1] =
 416         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 417
 418 /* 3-band FIFO queue: old style, but should be a bit faster than
 419    generic prio+fifo combination.
 420  */
 421
 422 #define PFIFO_FAST_BANDS 3
 423
 424 /*
 425  * Private data for a pfifo_fast scheduler containing:
 426  *      - queues for the three band
 427  *      - bitmap indicating which of the bands contain skbs
 428  */
 429 struct pfifo_fast_priv {
 430         u32 bitmap;
 431         struct sk_buff_head q[PFIFO_FAST_BANDS];
 432 };
 433
 434 /*
 435  * Convert a bitmap to the first band number where an skb is queued, where:
 436  *      bitmap=0 means there are no skbs on any band.
 437  *      bitmap=1 means there is an skb on band 0.
 438  *      bitmap=7 means there are skbs on all 3 bands, etc.
 439  */
 440 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 441
 442 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 443                                              int band)
 444 {
 445         return priv->q + band;
 446 }
 447
 448 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 449 {
 450         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 451                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 452                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 453                 struct sk_buff_head *list = band2list(priv, band);
 454
 455                 priv->bitmap |= (1 << band);
 456                 qdisc->q.qlen++;
 457                 return __qdisc_enqueue_tail(skb, qdisc, list);
 458         }
 459
 460         return qdisc_drop(skb, qdisc);
 461 }
 462
 463 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 464 {
 465         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 466         int band = bitmap2band[priv->bitmap];
 467
 468         if (likely(band >= 0)) {
 469                 struct sk_buff_head *list = band2list(priv, band);
 470                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 471
 472                 qdisc->q.qlen--;
 473                 if (skb_queue_empty(list))
 474                         priv->bitmap &= ~(1 << band);
 475
 476                 return skb;
 477         }
 478
 479         return NULL;
 480 }
 481
 482 static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc)
 483 {
 484         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 485         int band = bitmap2band[priv->bitmap];
 486
 487         if (band >= 0) {
 488                 struct sk_buff_head *list = band2list(priv, band);
 489
 490                 return skb_peek(list);
 491         }
 492
 493         return NULL;
 494 }
 495
 496 static void pfifo_fast_reset(struct Qdisc* qdisc)
 497 {
 498         int prio;
 499         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 500
 501         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 502                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 503
 504         priv->bitmap = 0;
 505         qdisc->qstats.backlog = 0;
 506         qdisc->q.qlen = 0;
 507 }
 508
 509 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 510 {
 511         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 512
 513         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 514         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 515         return skb->len;
 516
 517 nla_put_failure:
 518         return -1;
 519 }
 520
 521 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 522 {
 523         int prio;
 524         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 525
 526         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 527                 skb_queue_head_init(band2list(priv, prio));
 528
 529         return 0;
 530 }
 531
 532 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 533         .id             =       "pfifo_fast",
 534         .priv_size      =       sizeof(struct pfifo_fast_priv),
 535         .enqueue        =       pfifo_fast_enqueue,
 536         .dequeue        =       pfifo_fast_dequeue,
 537         .peek           =       pfifo_fast_peek,
 538         .init           =       pfifo_fast_init,
 539         .reset          =       pfifo_fast_reset,
 540         .dump           =       pfifo_fast_dump,
 541         .owner          =       THIS_MODULE,
 542 };
 543
 544 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 545                           struct Qdisc_ops *ops)
 546 {
 547         void *p;
 548         struct Qdisc *sch;
 549         unsigned int size;
 550         int err = -ENOBUFS;
 551
 552         /* ensure that the Qdisc and the private data are 64-byte aligned */
 553         size = QDISC_ALIGN(sizeof(*sch));
 554         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 555
 556         p = kzalloc(size, GFP_KERNEL);
 557         if (!p)
 558                 goto errout;
 559         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 560         sch->padded = (char *) sch - (char *) p;
 561
 562         INIT_LIST_HEAD(&sch->list);
 563         skb_queue_head_init(&sch->q);
 564         spin_lock_init(&sch->busylock);
 565         sch->ops = ops;
 566         sch->enqueue = ops->enqueue;
 567         sch->dequeue = ops->dequeue;
 568         sch->dev_queue = dev_queue;
 569         dev_hold(qdisc_dev(sch));
 570         atomic_set(&sch->refcnt, 1);
 571
 572         return sch;
 573 errout:
 574         return ERR_PTR(err);
 575 }
 576
 577 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 578                                 struct Qdisc_ops *ops, unsigned int parentid)
 579 {
 580         struct Qdisc *sch;
 581
 582         sch = qdisc_alloc(dev_queue, ops);
 583         if (IS_ERR(sch))
 584                 goto errout;
 585         sch->parent = parentid;
 586
 587         if (!ops->init || ops->init(sch, NULL) == 0)
 588                 return sch;
 589
 590         qdisc_destroy(sch);
 591 errout:
 592         return NULL;
 593 }
 594 EXPORT_SYMBOL(qdisc_create_dflt);
 595
 596 /* Under qdisc_lock(qdisc) and BH! */
 597
 598 void qdisc_reset(struct Qdisc *qdisc)
 599 {
 600         const struct Qdisc_ops *ops = qdisc->ops;
 601
 602         if (ops->reset)
 603                 ops->reset(qdisc);
 604
 605         if (qdisc->gso_skb) {
 606                 kfree_skb(qdisc->gso_skb);
 607                 qdisc->gso_skb = NULL;
 608                 qdisc->q.qlen = 0;
 609         }
 610 }
 611 EXPORT_SYMBOL(qdisc_reset);
 612
 613 static void qdisc_rcu_free(struct rcu_head *head)
 614 {
 615         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 616
 617         kfree((char *) qdisc - qdisc->padded);
 618 }
 619
 620 void qdisc_destroy(struct Qdisc *qdisc)
 621 {
 622         const struct Qdisc_ops  *ops = qdisc->ops;
 623
 624         if (qdisc->flags & TCQ_F_BUILTIN ||
 625             !atomic_dec_and_test(&qdisc->refcnt))
 626                 return;
 627
 628 #ifdef CONFIG_NET_SCHED
 629         qdisc_list_del(qdisc);
 630
 631         qdisc_put_stab(qdisc->stab);
 632 #endif
 633         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 634         if (ops->reset)
 635                 ops->reset(qdisc);
 636         if (ops->destroy)
 637                 ops->destroy(qdisc);
 638
 639         module_put(ops->owner);
 640         dev_put(qdisc_dev(qdisc));
 641
 642         kfree_skb(qdisc->gso_skb);
 643         /*
 644          * gen_estimator est_timer() might access qdisc->q.lock,
 645          * wait a RCU grace period before freeing qdisc.
 646          */
 647         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 648 }
 649 EXPORT_SYMBOL(qdisc_destroy);
 650
 651 /* Attach toplevel qdisc to device queue. */
 652 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 653                               struct Qdisc *qdisc)
 654 {
 655         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 656         spinlock_t *root_lock;
 657
 658         root_lock = qdisc_lock(oqdisc);
 659         spin_lock_bh(root_lock);
 660
 661         /* Prune old scheduler */
 662         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 663                 qdisc_reset(oqdisc);
 664
 665         /* ... and graft new one */
 666         if (qdisc == NULL)
 667                 qdisc = &noop_qdisc;
 668         dev_queue->qdisc_sleeping = qdisc;
 669         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 670
 671         spin_unlock_bh(root_lock);
 672
 673         return oqdisc;
 674 }
 675
 676 static void attach_one_default_qdisc(struct net_device *dev,
 677                                      struct netdev_queue *dev_queue,
 678                                      void *_unused)
 679 {
 680         struct Qdisc *qdisc;
 681
 682         if (dev->tx_queue_len) {
 683                 qdisc = qdisc_create_dflt(dev_queue,
 684                                           &pfifo_fast_ops, TC_H_ROOT);
 685                 if (!qdisc) {
 686                         printk(KERN_INFO "%s: activation failed\n", dev->name);
 687                         return;
 688                 }
 689
 690                 /* Can by-pass the queue discipline for default qdisc */
 691                 qdisc->flags |= TCQ_F_CAN_BYPASS;
 692         } else {
 693                 qdisc =  &noqueue_qdisc;
 694         }
 695         dev_queue->qdisc_sleeping = qdisc;
 696 }
 697
 698 static void attach_default_qdiscs(struct net_device *dev)
 699 {
 700         struct netdev_queue *txq;
 701         struct Qdisc *qdisc;
 702
 703         txq = netdev_get_tx_queue(dev, 0);
 704
 705         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 706                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 707                 dev->qdisc = txq->qdisc_sleeping;
 708                 atomic_inc(&dev->qdisc->refcnt);
 709         } else {
 710                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 711                 if (qdisc) {
 712                         qdisc->ops->attach(qdisc);
 713                         dev->qdisc = qdisc;
 714                 }
 715         }
 716 }
 717
 718 static void transition_one_qdisc(struct net_device *dev,
 719                                  struct netdev_queue *dev_queue,
 720                                  void *_need_watchdog)
 721 {
 722         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 723         int *need_watchdog_p = _need_watchdog;
 724
 725         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 726                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 727
 728         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 729         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 730                 dev_queue->trans_start = 0;
 731                 *need_watchdog_p = 1;
 732         }
 733 }
 734
 735 void dev_activate(struct net_device *dev)
 736 {
 737         int need_watchdog;
 738
 739         /* No queueing discipline is attached to device;
 740            create default one i.e. pfifo_fast for devices,
 741            which need queueing and noqueue_qdisc for
 742            virtual interfaces
 743          */
 744
 745         if (dev->qdisc == &noop_qdisc)
 746                 attach_default_qdiscs(dev);
 747
 748         if (!netif_carrier_ok(dev))
 749                 /* Delay activation until next carrier-on event */
 750                 return;
 751
 752         need_watchdog = 0;
 753         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 754         if (dev_ingress_queue(dev))
 755                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 756
 757         if (need_watchdog) {
 758                 dev->trans_start = jiffies;
 759                 dev_watchdog_up(dev);
 760         }
 761 }
 762
 763 static void dev_deactivate_queue(struct net_device *dev,
 764                                  struct netdev_queue *dev_queue,
 765                                  void *_qdisc_default)
 766 {
 767         struct Qdisc *qdisc_default = _qdisc_default;
 768         struct Qdisc *qdisc;
 769
 770         qdisc = dev_queue->qdisc;
 771         if (qdisc) {
 772                 spin_lock_bh(qdisc_lock(qdisc));
 773
 774                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 775                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 776
 777                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 778                 qdisc_reset(qdisc);
 779
 780                 spin_unlock_bh(qdisc_lock(qdisc));
 781         }
 782 }
 783
 784 static bool some_qdisc_is_busy(struct net_device *dev)
 785 {
 786         unsigned int i;
 787
 788         for (i = 0; i < dev->num_tx_queues; i++) {
 789                 struct netdev_queue *dev_queue;
 790                 spinlock_t *root_lock;
 791                 struct Qdisc *q;
 792                 int val;
 793
 794                 dev_queue = netdev_get_tx_queue(dev, i);
 795                 q = dev_queue->qdisc_sleeping;
 796                 root_lock = qdisc_lock(q);
 797
 798                 spin_lock_bh(root_lock);
 799
 800                 val = (qdisc_is_running(q) ||
 801                        test_bit(__QDISC_STATE_SCHED, &q->state));
 802
 803                 spin_unlock_bh(root_lock);
 804
 805                 if (val)
 806                         return true;
 807         }
 808         return false;
 809 }
 810
 811 void dev_deactivate(struct net_device *dev)
 812 {
 813         netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 814         if (dev_ingress_queue(dev))
 815                 dev_deactivate_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 816
 817         dev_watchdog_down(dev);
 818
 819         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 820         synchronize_rcu();
 821
 822         /* Wait for outstanding qdisc_run calls. */
 823         while (some_qdisc_is_busy(dev))
 824                 yield();
 825 }
 826
 827 static void dev_init_scheduler_queue(struct net_device *dev,
 828                                      struct netdev_queue *dev_queue,
 829                                      void *_qdisc)
 830 {
 831         struct Qdisc *qdisc = _qdisc;
 832
 833         dev_queue->qdisc = qdisc;
 834         dev_queue->qdisc_sleeping = qdisc;
 835 }
 836
 837 void dev_init_scheduler(struct net_device *dev)
 838 {
 839         dev->qdisc = &noop_qdisc;
 840         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 841         if (dev_ingress_queue(dev))
 842                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 843
 844         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 845 }
 846
 847 static void shutdown_scheduler_queue(struct net_device *dev,
 848                                      struct netdev_queue *dev_queue,
 849                                      void *_qdisc_default)
 850 {
 851         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 852         struct Qdisc *qdisc_default = _qdisc_default;
 853
 854         if (qdisc) {
 855                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 856                 dev_queue->qdisc_sleeping = qdisc_default;
 857
 858                 qdisc_destroy(qdisc);
 859         }
 860 }
 861
 862 void dev_shutdown(struct net_device *dev)
 863 {
 864         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 865         if (dev_ingress_queue(dev))
 866                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 867         qdisc_destroy(dev->qdisc);
 868         dev->qdisc = &noop_qdisc;
 869
 870         WARN_ON(timer_pending(&dev->watchdog_timer));
 871 }