Pileus Git - ~andy/linux/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 <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_xmit_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                 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
  90                                      dev_queue->dev->name);
  91                 ret = qdisc_qlen(q);
  92         } else {
  93                 /*
  94                  * Another cpu is holding lock, requeue & delay xmits for
  95                  * some time.
  96                  */
  97                 __this_cpu_inc(softnet_data.cpu_collision);
  98                 ret = dev_requeue_skb(skb, q);
  99         }
 100
 101         return ret;
 102 }
 103
 104 /*
 105  * Transmit one skb, and handle the return status as required. Holding the
 106  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 107  * function.
 108  *
 109  * Returns to the caller:
 110  *                              0  - queue is empty or throttled.
 111  *                              >0 - queue is not empty.
 112  */
 113 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 114                     struct net_device *dev, struct netdev_queue *txq,
 115                     spinlock_t *root_lock)
 116 {
 117         int ret = NETDEV_TX_BUSY;
 118
 119         /* And release qdisc */
 120         spin_unlock(root_lock);
 121
 122         HARD_TX_LOCK(dev, txq, smp_processor_id());
 123         if (!netif_xmit_frozen_or_stopped(txq))
 124                 ret = dev_hard_start_xmit(skb, dev, txq);
 125
 126         HARD_TX_UNLOCK(dev, txq);
 127
 128         spin_lock(root_lock);
 129
 130         if (dev_xmit_complete(ret)) {
 131                 /* Driver sent out skb successfully or skb was consumed */
 132                 ret = qdisc_qlen(q);
 133         } else if (ret == NETDEV_TX_LOCKED) {
 134                 /* Driver try lock failed */
 135                 ret = handle_dev_cpu_collision(skb, txq, q);
 136         } else {
 137                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 138                 if (unlikely(ret != NETDEV_TX_BUSY))
 139                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 140                                              dev->name, ret, q->q.qlen);
 141
 142                 ret = dev_requeue_skb(skb, q);
 143         }
 144
 145         if (ret && netif_xmit_frozen_or_stopped(txq))
 146                 ret = 0;
 147
 148         return ret;
 149 }
 150
 151 /*
 152  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 153  *
 154  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 155  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 156  * this queue.
 157  *
 158  *  netif_tx_lock serializes accesses to device driver.
 159  *
 160  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 161  *  if one is grabbed, another must be free.
 162  *
 163  * Note, that this procedure can be called by a watchdog timer
 164  *
 165  * Returns to the caller:
 166  *                              0  - queue is empty or throttled.
 167  *                              >0 - queue is not empty.
 168  *
 169  */
 170 static inline int qdisc_restart(struct Qdisc *q)
 171 {
 172         struct netdev_queue *txq;
 173         struct net_device *dev;
 174         spinlock_t *root_lock;
 175         struct sk_buff *skb;
 176
 177         /* Dequeue packet */
 178         skb = dequeue_skb(q);
 179         if (unlikely(!skb))
 180                 return 0;
 181         WARN_ON_ONCE(skb_dst_is_noref(skb));
 182         root_lock = qdisc_lock(q);
 183         dev = qdisc_dev(q);
 184         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 185
 186         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 187 }
 188
 189 void __qdisc_run(struct Qdisc *q)
 190 {
 191         int quota = weight_p;
 192
 193         while (qdisc_restart(q)) {
 194                 /*
 195                  * Ordered by possible occurrence: Postpone processing if
 196                  * 1. we've exceeded packet quota
 197                  * 2. another process needs the CPU;
 198                  */
 199                 if (--quota <= 0 || need_resched()) {
 200                         __netif_schedule(q);
 201                         break;
 202                 }
 203         }
 204
 205         qdisc_run_end(q);
 206 }
 207
 208 unsigned long dev_trans_start(struct net_device *dev)
 209 {
 210         unsigned long val, res = dev->trans_start;
 211         unsigned int i;
 212
 213         for (i = 0; i < dev->num_tx_queues; i++) {
 214                 val = netdev_get_tx_queue(dev, i)->trans_start;
 215                 if (val && time_after(val, res))
 216                         res = val;
 217         }
 218         dev->trans_start = res;
 219         return res;
 220 }
 221 EXPORT_SYMBOL(dev_trans_start);
 222
 223 static void dev_watchdog(unsigned long arg)
 224 {
 225         struct net_device *dev = (struct net_device *)arg;
 226
 227         netif_tx_lock(dev);
 228         if (!qdisc_tx_is_noop(dev)) {
 229                 if (netif_device_present(dev) &&
 230                     netif_running(dev) &&
 231                     netif_carrier_ok(dev)) {
 232                         int some_queue_timedout = 0;
 233                         unsigned int i;
 234                         unsigned long trans_start;
 235
 236                         for (i = 0; i < dev->num_tx_queues; i++) {
 237                                 struct netdev_queue *txq;
 238
 239                                 txq = netdev_get_tx_queue(dev, i);
 240                                 /*
 241                                  * old device drivers set dev->trans_start
 242                                  */
 243                                 trans_start = txq->trans_start ? : dev->trans_start;
 244                                 if (netif_xmit_stopped(txq) &&
 245                                     time_after(jiffies, (trans_start +
 246                                                          dev->watchdog_timeo))) {
 247                                         some_queue_timedout = 1;
 248                                         txq->trans_timeout++;
 249                                         break;
 250                                 }
 251                         }
 252
 253                         if (some_queue_timedout) {
 254                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 255                                        dev->name, netdev_drivername(dev), i);
 256                                 dev->netdev_ops->ndo_tx_timeout(dev);
 257                         }
 258                         if (!mod_timer(&dev->watchdog_timer,
 259                                        round_jiffies(jiffies +
 260                                                      dev->watchdog_timeo)))
 261                                 dev_hold(dev);
 262                 }
 263         }
 264         netif_tx_unlock(dev);
 265
 266         dev_put(dev);
 267 }
 268
 269 void __netdev_watchdog_up(struct net_device *dev)
 270 {
 271         if (dev->netdev_ops->ndo_tx_timeout) {
 272                 if (dev->watchdog_timeo <= 0)
 273                         dev->watchdog_timeo = 5*HZ;
 274                 if (!mod_timer(&dev->watchdog_timer,
 275                                round_jiffies(jiffies + dev->watchdog_timeo)))
 276                         dev_hold(dev);
 277         }
 278 }
 279
 280 static void dev_watchdog_up(struct net_device *dev)
 281 {
 282         __netdev_watchdog_up(dev);
 283 }
 284
 285 static void dev_watchdog_down(struct net_device *dev)
 286 {
 287         netif_tx_lock_bh(dev);
 288         if (del_timer(&dev->watchdog_timer))
 289                 dev_put(dev);
 290         netif_tx_unlock_bh(dev);
 291 }
 292
 293 /**
 294  *      netif_carrier_on - set carrier
 295  *      @dev: network device
 296  *
 297  * Device has detected that carrier.
 298  */
 299 void netif_carrier_on(struct net_device *dev)
 300 {
 301         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 302                 if (dev->reg_state == NETREG_UNINITIALIZED)
 303                         return;
 304                 linkwatch_fire_event(dev);
 305                 if (netif_running(dev))
 306                         __netdev_watchdog_up(dev);
 307         }
 308 }
 309 EXPORT_SYMBOL(netif_carrier_on);
 310
 311 /**
 312  *      netif_carrier_off - clear carrier
 313  *      @dev: network device
 314  *
 315  * Device has detected loss of carrier.
 316  */
 317 void netif_carrier_off(struct net_device *dev)
 318 {
 319         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 320                 if (dev->reg_state == NETREG_UNINITIALIZED)
 321                         return;
 322                 linkwatch_fire_event(dev);
 323         }
 324 }
 325 EXPORT_SYMBOL(netif_carrier_off);
 326
 327 /**
 328  *      netif_notify_peers - notify network peers about existence of @dev
 329  *      @dev: network device
 330  *
 331  * Generate traffic such that interested network peers are aware of
 332  * @dev, such as by generating a gratuitous ARP. This may be used when
 333  * a device wants to inform the rest of the network about some sort of
 334  * reconfiguration such as a failover event or virtual machine
 335  * migration.
 336  */
 337 void netif_notify_peers(struct net_device *dev)
 338 {
 339         rtnl_lock();
 340         call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
 341         rtnl_unlock();
 342 }
 343 EXPORT_SYMBOL(netif_notify_peers);
 344
 345 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 346    under all circumstances. It is difficult to invent anything faster or
 347    cheaper.
 348  */
 349
 350 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 351 {
 352         kfree_skb(skb);
 353         return NET_XMIT_CN;
 354 }
 355
 356 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 357 {
 358         return NULL;
 359 }
 360
 361 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 362         .id             =       "noop",
 363         .priv_size      =       0,
 364         .enqueue        =       noop_enqueue,
 365         .dequeue        =       noop_dequeue,
 366         .peek           =       noop_dequeue,
 367         .owner          =       THIS_MODULE,
 368 };
 369
 370 static struct netdev_queue noop_netdev_queue = {
 371         .qdisc          =       &noop_qdisc,
 372         .qdisc_sleeping =       &noop_qdisc,
 373 };
 374
 375 struct Qdisc noop_qdisc = {
 376         .enqueue        =       noop_enqueue,
 377         .dequeue        =       noop_dequeue,
 378         .flags          =       TCQ_F_BUILTIN,
 379         .ops            =       &noop_qdisc_ops,
 380         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 381         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 382         .dev_queue      =       &noop_netdev_queue,
 383         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 384 };
 385 EXPORT_SYMBOL(noop_qdisc);
 386
 387 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 388         .id             =       "noqueue",
 389         .priv_size      =       0,
 390         .enqueue        =       noop_enqueue,
 391         .dequeue        =       noop_dequeue,
 392         .peek           =       noop_dequeue,
 393         .owner          =       THIS_MODULE,
 394 };
 395
 396 static struct Qdisc noqueue_qdisc;
 397 static struct netdev_queue noqueue_netdev_queue = {
 398         .qdisc          =       &noqueue_qdisc,
 399         .qdisc_sleeping =       &noqueue_qdisc,
 400 };
 401
 402 static struct Qdisc noqueue_qdisc = {
 403         .enqueue        =       NULL,
 404         .dequeue        =       noop_dequeue,
 405         .flags          =       TCQ_F_BUILTIN,
 406         .ops            =       &noqueue_qdisc_ops,
 407         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 408         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 409         .dev_queue      =       &noqueue_netdev_queue,
 410         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 411 };
 412
 413
 414 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 415         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 416 };
 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         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 515                 goto nla_put_failure;
 516         return skb->len;
 517
 518 nla_put_failure:
 519         return -1;
 520 }
 521
 522 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 523 {
 524         int prio;
 525         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 526
 527         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 528                 skb_queue_head_init(band2list(priv, prio));
 529
 530         /* Can by-pass the queue discipline */
 531         qdisc->flags |= TCQ_F_CAN_BYPASS;
 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 EXPORT_SYMBOL(pfifo_fast_ops);
 547
 548 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 549                           struct Qdisc_ops *ops)
 550 {
 551         void *p;
 552         struct Qdisc *sch;
 553         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 554         int err = -ENOBUFS;
 555
 556         p = kzalloc_node(size, GFP_KERNEL,
 557                          netdev_queue_numa_node_read(dev_queue));
 558
 559         if (!p)
 560                 goto errout;
 561         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 562         /* if we got non aligned memory, ask more and do alignment ourself */
 563         if (sch != p) {
 564                 kfree(p);
 565                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 566                                  netdev_queue_numa_node_read(dev_queue));
 567                 if (!p)
 568                         goto errout;
 569                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 570                 sch->padded = (char *) sch - (char *) p;
 571         }
 572         INIT_LIST_HEAD(&sch->list);
 573         skb_queue_head_init(&sch->q);
 574         spin_lock_init(&sch->busylock);
 575         sch->ops = ops;
 576         sch->enqueue = ops->enqueue;
 577         sch->dequeue = ops->dequeue;
 578         sch->dev_queue = dev_queue;
 579         dev_hold(qdisc_dev(sch));
 580         atomic_set(&sch->refcnt, 1);
 581
 582         return sch;
 583 errout:
 584         return ERR_PTR(err);
 585 }
 586
 587 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 588                                 struct Qdisc_ops *ops, unsigned int parentid)
 589 {
 590         struct Qdisc *sch;
 591
 592         sch = qdisc_alloc(dev_queue, ops);
 593         if (IS_ERR(sch))
 594                 goto errout;
 595         sch->parent = parentid;
 596
 597         if (!ops->init || ops->init(sch, NULL) == 0)
 598                 return sch;
 599
 600         qdisc_destroy(sch);
 601 errout:
 602         return NULL;
 603 }
 604 EXPORT_SYMBOL(qdisc_create_dflt);
 605
 606 /* Under qdisc_lock(qdisc) and BH! */
 607
 608 void qdisc_reset(struct Qdisc *qdisc)
 609 {
 610         const struct Qdisc_ops *ops = qdisc->ops;
 611
 612         if (ops->reset)
 613                 ops->reset(qdisc);
 614
 615         if (qdisc->gso_skb) {
 616                 kfree_skb(qdisc->gso_skb);
 617                 qdisc->gso_skb = NULL;
 618                 qdisc->q.qlen = 0;
 619         }
 620 }
 621 EXPORT_SYMBOL(qdisc_reset);
 622
 623 static void qdisc_rcu_free(struct rcu_head *head)
 624 {
 625         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 626
 627         kfree((char *) qdisc - qdisc->padded);
 628 }
 629
 630 void qdisc_destroy(struct Qdisc *qdisc)
 631 {
 632         const struct Qdisc_ops  *ops = qdisc->ops;
 633
 634         if (qdisc->flags & TCQ_F_BUILTIN ||
 635             !atomic_dec_and_test(&qdisc->refcnt))
 636                 return;
 637
 638 #ifdef CONFIG_NET_SCHED
 639         qdisc_list_del(qdisc);
 640
 641         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 642 #endif
 643         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 644         if (ops->reset)
 645                 ops->reset(qdisc);
 646         if (ops->destroy)
 647                 ops->destroy(qdisc);
 648
 649         module_put(ops->owner);
 650         dev_put(qdisc_dev(qdisc));
 651
 652         kfree_skb(qdisc->gso_skb);
 653         /*
 654          * gen_estimator est_timer() might access qdisc->q.lock,
 655          * wait a RCU grace period before freeing qdisc.
 656          */
 657         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 658 }
 659 EXPORT_SYMBOL(qdisc_destroy);
 660
 661 /* Attach toplevel qdisc to device queue. */
 662 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 663                               struct Qdisc *qdisc)
 664 {
 665         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 666         spinlock_t *root_lock;
 667
 668         root_lock = qdisc_lock(oqdisc);
 669         spin_lock_bh(root_lock);
 670
 671         /* Prune old scheduler */
 672         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 673                 qdisc_reset(oqdisc);
 674
 675         /* ... and graft new one */
 676         if (qdisc == NULL)
 677                 qdisc = &noop_qdisc;
 678         dev_queue->qdisc_sleeping = qdisc;
 679         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 680
 681         spin_unlock_bh(root_lock);
 682
 683         return oqdisc;
 684 }
 685 EXPORT_SYMBOL(dev_graft_qdisc);
 686
 687 static void attach_one_default_qdisc(struct net_device *dev,
 688                                      struct netdev_queue *dev_queue,
 689                                      void *_unused)
 690 {
 691         struct Qdisc *qdisc = &noqueue_qdisc;
 692
 693         if (dev->tx_queue_len) {
 694                 qdisc = qdisc_create_dflt(dev_queue,
 695                                           &pfifo_fast_ops, TC_H_ROOT);
 696                 if (!qdisc) {
 697                         netdev_info(dev, "activation failed\n");
 698                         return;
 699                 }
 700         }
 701         dev_queue->qdisc_sleeping = qdisc;
 702 }
 703
 704 static void attach_default_qdiscs(struct net_device *dev)
 705 {
 706         struct netdev_queue *txq;
 707         struct Qdisc *qdisc;
 708
 709         txq = netdev_get_tx_queue(dev, 0);
 710
 711         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 712                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 713                 dev->qdisc = txq->qdisc_sleeping;
 714                 atomic_inc(&dev->qdisc->refcnt);
 715         } else {
 716                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 717                 if (qdisc) {
 718                         qdisc->ops->attach(qdisc);
 719                         dev->qdisc = qdisc;
 720                 }
 721         }
 722 }
 723
 724 static void transition_one_qdisc(struct net_device *dev,
 725                                  struct netdev_queue *dev_queue,
 726                                  void *_need_watchdog)
 727 {
 728         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 729         int *need_watchdog_p = _need_watchdog;
 730
 731         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 732                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 733
 734         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 735         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 736                 dev_queue->trans_start = 0;
 737                 *need_watchdog_p = 1;
 738         }
 739 }
 740
 741 void dev_activate(struct net_device *dev)
 742 {
 743         int need_watchdog;
 744
 745         /* No queueing discipline is attached to device;
 746            create default one i.e. pfifo_fast for devices,
 747            which need queueing and noqueue_qdisc for
 748            virtual interfaces
 749          */
 750
 751         if (dev->qdisc == &noop_qdisc)
 752                 attach_default_qdiscs(dev);
 753
 754         if (!netif_carrier_ok(dev))
 755                 /* Delay activation until next carrier-on event */
 756                 return;
 757
 758         need_watchdog = 0;
 759         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 760         if (dev_ingress_queue(dev))
 761                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 762
 763         if (need_watchdog) {
 764                 dev->trans_start = jiffies;
 765                 dev_watchdog_up(dev);
 766         }
 767 }
 768 EXPORT_SYMBOL(dev_activate);
 769
 770 static void dev_deactivate_queue(struct net_device *dev,
 771                                  struct netdev_queue *dev_queue,
 772                                  void *_qdisc_default)
 773 {
 774         struct Qdisc *qdisc_default = _qdisc_default;
 775         struct Qdisc *qdisc;
 776
 777         qdisc = dev_queue->qdisc;
 778         if (qdisc) {
 779                 spin_lock_bh(qdisc_lock(qdisc));
 780
 781                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 782                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 783
 784                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 785                 qdisc_reset(qdisc);
 786
 787                 spin_unlock_bh(qdisc_lock(qdisc));
 788         }
 789 }
 790
 791 static bool some_qdisc_is_busy(struct net_device *dev)
 792 {
 793         unsigned int i;
 794
 795         for (i = 0; i < dev->num_tx_queues; i++) {
 796                 struct netdev_queue *dev_queue;
 797                 spinlock_t *root_lock;
 798                 struct Qdisc *q;
 799                 int val;
 800
 801                 dev_queue = netdev_get_tx_queue(dev, i);
 802                 q = dev_queue->qdisc_sleeping;
 803                 root_lock = qdisc_lock(q);
 804
 805                 spin_lock_bh(root_lock);
 806
 807                 val = (qdisc_is_running(q) ||
 808                        test_bit(__QDISC_STATE_SCHED, &q->state));
 809
 810                 spin_unlock_bh(root_lock);
 811
 812                 if (val)
 813                         return true;
 814         }
 815         return false;
 816 }
 817
 818 /**
 819  *      dev_deactivate_many - deactivate transmissions on several devices
 820  *      @head: list of devices to deactivate
 821  *
 822  *      This function returns only when all outstanding transmissions
 823  *      have completed, unless all devices are in dismantle phase.
 824  */
 825 void dev_deactivate_many(struct list_head *head)
 826 {
 827         struct net_device *dev;
 828         bool sync_needed = false;
 829
 830         list_for_each_entry(dev, head, unreg_list) {
 831                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 832                                          &noop_qdisc);
 833                 if (dev_ingress_queue(dev))
 834                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 835                                              &noop_qdisc);
 836
 837                 dev_watchdog_down(dev);
 838                 sync_needed |= !dev->dismantle;
 839         }
 840
 841         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 842          * This is avoided if all devices are in dismantle phase :
 843          * Caller will call synchronize_net() for us
 844          */
 845         if (sync_needed)
 846                 synchronize_net();
 847
 848         /* Wait for outstanding qdisc_run calls. */
 849         list_for_each_entry(dev, head, unreg_list)
 850                 while (some_qdisc_is_busy(dev))
 851                         yield();
 852 }
 853
 854 void dev_deactivate(struct net_device *dev)
 855 {
 856         LIST_HEAD(single);
 857
 858         list_add(&dev->unreg_list, &single);
 859         dev_deactivate_many(&single);
 860         list_del(&single);
 861 }
 862 EXPORT_SYMBOL(dev_deactivate);
 863
 864 static void dev_init_scheduler_queue(struct net_device *dev,
 865                                      struct netdev_queue *dev_queue,
 866                                      void *_qdisc)
 867 {
 868         struct Qdisc *qdisc = _qdisc;
 869
 870         dev_queue->qdisc = qdisc;
 871         dev_queue->qdisc_sleeping = qdisc;
 872 }
 873
 874 void dev_init_scheduler(struct net_device *dev)
 875 {
 876         dev->qdisc = &noop_qdisc;
 877         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 878         if (dev_ingress_queue(dev))
 879                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 880
 881         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 882 }
 883
 884 static void shutdown_scheduler_queue(struct net_device *dev,
 885                                      struct netdev_queue *dev_queue,
 886                                      void *_qdisc_default)
 887 {
 888         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 889         struct Qdisc *qdisc_default = _qdisc_default;
 890
 891         if (qdisc) {
 892                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 893                 dev_queue->qdisc_sleeping = qdisc_default;
 894
 895                 qdisc_destroy(qdisc);
 896         }
 897 }
 898
 899 void dev_shutdown(struct net_device *dev)
 900 {
 901         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 902         if (dev_ingress_queue(dev))
 903                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 904         qdisc_destroy(dev->qdisc);
 905         dev->qdisc = &noop_qdisc;
 906
 907         WARN_ON(timer_pending(&dev->watchdog_timer));
 908 }