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                         pr_warning("Dead loop on netdevice %s, fix it urgently!\n",
  91                                    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                         pr_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         int quota = weight_p;
 193         int work = 0;
 194
 195         while (qdisc_restart(q)) {
 196                 work++;
 197                 /*
 198                  * Ordered by possible occurrence: Postpone processing if
 199                  * 1. we've exceeded packet quota
 200                  * 2. another process needs the CPU;
 201                  */
 202                 if (work >= quota || need_resched()) {
 203                         __netif_schedule(q);
 204                         break;
 205                 }
 206         }
 207
 208         qdisc_run_end(q);
 209 }
 210
 211 unsigned long dev_trans_start(struct net_device *dev)
 212 {
 213         unsigned long val, res = dev->trans_start;
 214         unsigned int i;
 215
 216         for (i = 0; i < dev->num_tx_queues; i++) {
 217                 val = netdev_get_tx_queue(dev, i)->trans_start;
 218                 if (val && time_after(val, res))
 219                         res = val;
 220         }
 221         dev->trans_start = res;
 222         return res;
 223 }
 224 EXPORT_SYMBOL(dev_trans_start);
 225
 226 static void dev_watchdog(unsigned long arg)
 227 {
 228         struct net_device *dev = (struct net_device *)arg;
 229
 230         netif_tx_lock(dev);
 231         if (!qdisc_tx_is_noop(dev)) {
 232                 if (netif_device_present(dev) &&
 233                     netif_running(dev) &&
 234                     netif_carrier_ok(dev)) {
 235                         int some_queue_timedout = 0;
 236                         unsigned int i;
 237                         unsigned long trans_start;
 238
 239                         for (i = 0; i < dev->num_tx_queues; i++) {
 240                                 struct netdev_queue *txq;
 241
 242                                 txq = netdev_get_tx_queue(dev, i);
 243                                 /*
 244                                  * old device drivers set dev->trans_start
 245                                  */
 246                                 trans_start = txq->trans_start ? : dev->trans_start;
 247                                 if (netif_tx_queue_stopped(txq) &&
 248                                     time_after(jiffies, (trans_start +
 249                                                          dev->watchdog_timeo))) {
 250                                         some_queue_timedout = 1;
 251                                         break;
 252                                 }
 253                         }
 254
 255                         if (some_queue_timedout) {
 256                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 257                                        dev->name, netdev_drivername(dev), i);
 258                                 dev->netdev_ops->ndo_tx_timeout(dev);
 259                         }
 260                         if (!mod_timer(&dev->watchdog_timer,
 261                                        round_jiffies(jiffies +
 262                                                      dev->watchdog_timeo)))
 263                                 dev_hold(dev);
 264                 }
 265         }
 266         netif_tx_unlock(dev);
 267
 268         dev_put(dev);
 269 }
 270
 271 void __netdev_watchdog_up(struct net_device *dev)
 272 {
 273         if (dev->netdev_ops->ndo_tx_timeout) {
 274                 if (dev->watchdog_timeo <= 0)
 275                         dev->watchdog_timeo = 5*HZ;
 276                 if (!mod_timer(&dev->watchdog_timer,
 277                                round_jiffies(jiffies + dev->watchdog_timeo)))
 278                         dev_hold(dev);
 279         }
 280 }
 281
 282 static void dev_watchdog_up(struct net_device *dev)
 283 {
 284         __netdev_watchdog_up(dev);
 285 }
 286
 287 static void dev_watchdog_down(struct net_device *dev)
 288 {
 289         netif_tx_lock_bh(dev);
 290         if (del_timer(&dev->watchdog_timer))
 291                 dev_put(dev);
 292         netif_tx_unlock_bh(dev);
 293 }
 294
 295 /**
 296  *      netif_carrier_on - set carrier
 297  *      @dev: network device
 298  *
 299  * Device has detected that carrier.
 300  */
 301 void netif_carrier_on(struct net_device *dev)
 302 {
 303         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 304                 if (dev->reg_state == NETREG_UNINITIALIZED)
 305                         return;
 306                 linkwatch_fire_event(dev);
 307                 if (netif_running(dev))
 308                         __netdev_watchdog_up(dev);
 309         }
 310 }
 311 EXPORT_SYMBOL(netif_carrier_on);
 312
 313 /**
 314  *      netif_carrier_off - clear carrier
 315  *      @dev: network device
 316  *
 317  * Device has detected loss of carrier.
 318  */
 319 void netif_carrier_off(struct net_device *dev)
 320 {
 321         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 322                 if (dev->reg_state == NETREG_UNINITIALIZED)
 323                         return;
 324                 linkwatch_fire_event(dev);
 325         }
 326 }
 327 EXPORT_SYMBOL(netif_carrier_off);
 328
 329 /**
 330  *      netif_notify_peers - notify network peers about existence of @dev
 331  *      @dev: network device
 332  *
 333  * Generate traffic such that interested network peers are aware of
 334  * @dev, such as by generating a gratuitous ARP. This may be used when
 335  * a device wants to inform the rest of the network about some sort of
 336  * reconfiguration such as a failover event or virtual machine
 337  * migration.
 338  */
 339 void netif_notify_peers(struct net_device *dev)
 340 {
 341         rtnl_lock();
 342         call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
 343         rtnl_unlock();
 344 }
 345 EXPORT_SYMBOL(netif_notify_peers);
 346
 347 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 348    under all circumstances. It is difficult to invent anything faster or
 349    cheaper.
 350  */
 351
 352 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 353 {
 354         kfree_skb(skb);
 355         return NET_XMIT_CN;
 356 }
 357
 358 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 359 {
 360         return NULL;
 361 }
 362
 363 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 364         .id             =       "noop",
 365         .priv_size      =       0,
 366         .enqueue        =       noop_enqueue,
 367         .dequeue        =       noop_dequeue,
 368         .peek           =       noop_dequeue,
 369         .owner          =       THIS_MODULE,
 370 };
 371
 372 static struct netdev_queue noop_netdev_queue = {
 373         .qdisc          =       &noop_qdisc,
 374         .qdisc_sleeping =       &noop_qdisc,
 375 };
 376
 377 struct Qdisc noop_qdisc = {
 378         .enqueue        =       noop_enqueue,
 379         .dequeue        =       noop_dequeue,
 380         .flags          =       TCQ_F_BUILTIN,
 381         .ops            =       &noop_qdisc_ops,
 382         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 383         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 384         .dev_queue      =       &noop_netdev_queue,
 385         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 386 };
 387 EXPORT_SYMBOL(noop_qdisc);
 388
 389 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 390         .id             =       "noqueue",
 391         .priv_size      =       0,
 392         .enqueue        =       noop_enqueue,
 393         .dequeue        =       noop_dequeue,
 394         .peek           =       noop_dequeue,
 395         .owner          =       THIS_MODULE,
 396 };
 397
 398 static struct Qdisc noqueue_qdisc;
 399 static struct netdev_queue noqueue_netdev_queue = {
 400         .qdisc          =       &noqueue_qdisc,
 401         .qdisc_sleeping =       &noqueue_qdisc,
 402 };
 403
 404 static struct Qdisc noqueue_qdisc = {
 405         .enqueue        =       NULL,
 406         .dequeue        =       noop_dequeue,
 407         .flags          =       TCQ_F_BUILTIN,
 408         .ops            =       &noqueue_qdisc_ops,
 409         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 410         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 411         .dev_queue      =       &noqueue_netdev_queue,
 412         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 413 };
 414
 415
 416 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 417         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 418 };
 419
 420 /* 3-band FIFO queue: old style, but should be a bit faster than
 421    generic prio+fifo combination.
 422  */
 423
 424 #define PFIFO_FAST_BANDS 3
 425
 426 /*
 427  * Private data for a pfifo_fast scheduler containing:
 428  *      - queues for the three band
 429  *      - bitmap indicating which of the bands contain skbs
 430  */
 431 struct pfifo_fast_priv {
 432         u32 bitmap;
 433         struct sk_buff_head q[PFIFO_FAST_BANDS];
 434 };
 435
 436 /*
 437  * Convert a bitmap to the first band number where an skb is queued, where:
 438  *      bitmap=0 means there are no skbs on any band.
 439  *      bitmap=1 means there is an skb on band 0.
 440  *      bitmap=7 means there are skbs on all 3 bands, etc.
 441  */
 442 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 443
 444 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 445                                              int band)
 446 {
 447         return priv->q + band;
 448 }
 449
 450 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 451 {
 452         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 453                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 454                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 455                 struct sk_buff_head *list = band2list(priv, band);
 456
 457                 priv->bitmap |= (1 << band);
 458                 qdisc->q.qlen++;
 459                 return __qdisc_enqueue_tail(skb, qdisc, list);
 460         }
 461
 462         return qdisc_drop(skb, qdisc);
 463 }
 464
 465 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 466 {
 467         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 468         int band = bitmap2band[priv->bitmap];
 469
 470         if (likely(band >= 0)) {
 471                 struct sk_buff_head *list = band2list(priv, band);
 472                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 473
 474                 qdisc->q.qlen--;
 475                 if (skb_queue_empty(list))
 476                         priv->bitmap &= ~(1 << band);
 477
 478                 return skb;
 479         }
 480
 481         return NULL;
 482 }
 483
 484 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 485 {
 486         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 487         int band = bitmap2band[priv->bitmap];
 488
 489         if (band >= 0) {
 490                 struct sk_buff_head *list = band2list(priv, band);
 491
 492                 return skb_peek(list);
 493         }
 494
 495         return NULL;
 496 }
 497
 498 static void pfifo_fast_reset(struct Qdisc *qdisc)
 499 {
 500         int prio;
 501         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 502
 503         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 504                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 505
 506         priv->bitmap = 0;
 507         qdisc->qstats.backlog = 0;
 508         qdisc->q.qlen = 0;
 509 }
 510
 511 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 512 {
 513         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 514
 515         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 516         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 517         return skb->len;
 518
 519 nla_put_failure:
 520         return -1;
 521 }
 522
 523 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 524 {
 525         int prio;
 526         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 527
 528         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 529                 skb_queue_head_init(band2list(priv, prio));
 530
 531         /* Can by-pass the queue discipline */
 532         qdisc->flags |= TCQ_F_CAN_BYPASS;
 533         return 0;
 534 }
 535
 536 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 537         .id             =       "pfifo_fast",
 538         .priv_size      =       sizeof(struct pfifo_fast_priv),
 539         .enqueue        =       pfifo_fast_enqueue,
 540         .dequeue        =       pfifo_fast_dequeue,
 541         .peek           =       pfifo_fast_peek,
 542         .init           =       pfifo_fast_init,
 543         .reset          =       pfifo_fast_reset,
 544         .dump           =       pfifo_fast_dump,
 545         .owner          =       THIS_MODULE,
 546 };
 547 EXPORT_SYMBOL(pfifo_fast_ops);
 548
 549 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 550                           struct Qdisc_ops *ops)
 551 {
 552         void *p;
 553         struct Qdisc *sch;
 554         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 555         int err = -ENOBUFS;
 556
 557         p = kzalloc_node(size, GFP_KERNEL,
 558                          netdev_queue_numa_node_read(dev_queue));
 559
 560         if (!p)
 561                 goto errout;
 562         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 563         /* if we got non aligned memory, ask more and do alignment ourself */
 564         if (sch != p) {
 565                 kfree(p);
 566                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 567                                  netdev_queue_numa_node_read(dev_queue));
 568                 if (!p)
 569                         goto errout;
 570                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 571                 sch->padded = (char *) sch - (char *) p;
 572         }
 573         INIT_LIST_HEAD(&sch->list);
 574         skb_queue_head_init(&sch->q);
 575         spin_lock_init(&sch->busylock);
 576         sch->ops = ops;
 577         sch->enqueue = ops->enqueue;
 578         sch->dequeue = ops->dequeue;
 579         sch->dev_queue = dev_queue;
 580         dev_hold(qdisc_dev(sch));
 581         atomic_set(&sch->refcnt, 1);
 582
 583         return sch;
 584 errout:
 585         return ERR_PTR(err);
 586 }
 587
 588 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 589                                 struct Qdisc_ops *ops, unsigned int parentid)
 590 {
 591         struct Qdisc *sch;
 592
 593         sch = qdisc_alloc(dev_queue, ops);
 594         if (IS_ERR(sch))
 595                 goto errout;
 596         sch->parent = parentid;
 597
 598         if (!ops->init || ops->init(sch, NULL) == 0)
 599                 return sch;
 600
 601         qdisc_destroy(sch);
 602 errout:
 603         return NULL;
 604 }
 605 EXPORT_SYMBOL(qdisc_create_dflt);
 606
 607 /* Under qdisc_lock(qdisc) and BH! */
 608
 609 void qdisc_reset(struct Qdisc *qdisc)
 610 {
 611         const struct Qdisc_ops *ops = qdisc->ops;
 612
 613         if (ops->reset)
 614                 ops->reset(qdisc);
 615
 616         if (qdisc->gso_skb) {
 617                 kfree_skb(qdisc->gso_skb);
 618                 qdisc->gso_skb = NULL;
 619                 qdisc->q.qlen = 0;
 620         }
 621 }
 622 EXPORT_SYMBOL(qdisc_reset);
 623
 624 static void qdisc_rcu_free(struct rcu_head *head)
 625 {
 626         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 627
 628         kfree((char *) qdisc - qdisc->padded);
 629 }
 630
 631 void qdisc_destroy(struct Qdisc *qdisc)
 632 {
 633         const struct Qdisc_ops  *ops = qdisc->ops;
 634
 635         if (qdisc->flags & TCQ_F_BUILTIN ||
 636             !atomic_dec_and_test(&qdisc->refcnt))
 637                 return;
 638
 639 #ifdef CONFIG_NET_SCHED
 640         qdisc_list_del(qdisc);
 641
 642         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 643 #endif
 644         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 645         if (ops->reset)
 646                 ops->reset(qdisc);
 647         if (ops->destroy)
 648                 ops->destroy(qdisc);
 649
 650         module_put(ops->owner);
 651         dev_put(qdisc_dev(qdisc));
 652
 653         kfree_skb(qdisc->gso_skb);
 654         /*
 655          * gen_estimator est_timer() might access qdisc->q.lock,
 656          * wait a RCU grace period before freeing qdisc.
 657          */
 658         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 659 }
 660 EXPORT_SYMBOL(qdisc_destroy);
 661
 662 /* Attach toplevel qdisc to device queue. */
 663 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 664                               struct Qdisc *qdisc)
 665 {
 666         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 667         spinlock_t *root_lock;
 668
 669         root_lock = qdisc_lock(oqdisc);
 670         spin_lock_bh(root_lock);
 671
 672         /* Prune old scheduler */
 673         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 674                 qdisc_reset(oqdisc);
 675
 676         /* ... and graft new one */
 677         if (qdisc == NULL)
 678                 qdisc = &noop_qdisc;
 679         dev_queue->qdisc_sleeping = qdisc;
 680         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 681
 682         spin_unlock_bh(root_lock);
 683
 684         return oqdisc;
 685 }
 686 EXPORT_SYMBOL(dev_graft_qdisc);
 687
 688 static void attach_one_default_qdisc(struct net_device *dev,
 689                                      struct netdev_queue *dev_queue,
 690                                      void *_unused)
 691 {
 692         struct Qdisc *qdisc = &noqueue_qdisc;
 693
 694         if (dev->tx_queue_len) {
 695                 qdisc = qdisc_create_dflt(dev_queue,
 696                                           &pfifo_fast_ops, TC_H_ROOT);
 697                 if (!qdisc) {
 698                         netdev_info(dev, "activation failed\n");
 699                         return;
 700                 }
 701         }
 702         dev_queue->qdisc_sleeping = qdisc;
 703 }
 704
 705 static void attach_default_qdiscs(struct net_device *dev)
 706 {
 707         struct netdev_queue *txq;
 708         struct Qdisc *qdisc;
 709
 710         txq = netdev_get_tx_queue(dev, 0);
 711
 712         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 713                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 714                 dev->qdisc = txq->qdisc_sleeping;
 715                 atomic_inc(&dev->qdisc->refcnt);
 716         } else {
 717                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 718                 if (qdisc) {
 719                         qdisc->ops->attach(qdisc);
 720                         dev->qdisc = qdisc;
 721                 }
 722         }
 723 }
 724
 725 static void transition_one_qdisc(struct net_device *dev,
 726                                  struct netdev_queue *dev_queue,
 727                                  void *_need_watchdog)
 728 {
 729         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 730         int *need_watchdog_p = _need_watchdog;
 731
 732         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 733                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 734
 735         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 736         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 737                 dev_queue->trans_start = 0;
 738                 *need_watchdog_p = 1;
 739         }
 740 }
 741
 742 void dev_activate(struct net_device *dev)
 743 {
 744         int need_watchdog;
 745
 746         /* No queueing discipline is attached to device;
 747            create default one i.e. pfifo_fast for devices,
 748            which need queueing and noqueue_qdisc for
 749            virtual interfaces
 750          */
 751
 752         if (dev->qdisc == &noop_qdisc)
 753                 attach_default_qdiscs(dev);
 754
 755         if (!netif_carrier_ok(dev))
 756                 /* Delay activation until next carrier-on event */
 757                 return;
 758
 759         need_watchdog = 0;
 760         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 761         if (dev_ingress_queue(dev))
 762                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 763
 764         if (need_watchdog) {
 765                 dev->trans_start = jiffies;
 766                 dev_watchdog_up(dev);
 767         }
 768 }
 769 EXPORT_SYMBOL(dev_activate);
 770
 771 static void dev_deactivate_queue(struct net_device *dev,
 772                                  struct netdev_queue *dev_queue,
 773                                  void *_qdisc_default)
 774 {
 775         struct Qdisc *qdisc_default = _qdisc_default;
 776         struct Qdisc *qdisc;
 777
 778         qdisc = dev_queue->qdisc;
 779         if (qdisc) {
 780                 spin_lock_bh(qdisc_lock(qdisc));
 781
 782                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 783                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 784
 785                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 786                 qdisc_reset(qdisc);
 787
 788                 spin_unlock_bh(qdisc_lock(qdisc));
 789         }
 790 }
 791
 792 static bool some_qdisc_is_busy(struct net_device *dev)
 793 {
 794         unsigned int i;
 795
 796         for (i = 0; i < dev->num_tx_queues; i++) {
 797                 struct netdev_queue *dev_queue;
 798                 spinlock_t *root_lock;
 799                 struct Qdisc *q;
 800                 int val;
 801
 802                 dev_queue = netdev_get_tx_queue(dev, i);
 803                 q = dev_queue->qdisc_sleeping;
 804                 root_lock = qdisc_lock(q);
 805
 806                 spin_lock_bh(root_lock);
 807
 808                 val = (qdisc_is_running(q) ||
 809                        test_bit(__QDISC_STATE_SCHED, &q->state));
 810
 811                 spin_unlock_bh(root_lock);
 812
 813                 if (val)
 814                         return true;
 815         }
 816         return false;
 817 }
 818
 819 /**
 820  *      dev_deactivate_many - deactivate transmissions on several devices
 821  *      @head: list of devices to deactivate
 822  *
 823  *      This function returns only when all outstanding transmissions
 824  *      have completed, unless all devices are in dismantle phase.
 825  */
 826 void dev_deactivate_many(struct list_head *head)
 827 {
 828         struct net_device *dev;
 829         bool sync_needed = false;
 830
 831         list_for_each_entry(dev, head, unreg_list) {
 832                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 833                                          &noop_qdisc);
 834                 if (dev_ingress_queue(dev))
 835                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 836                                              &noop_qdisc);
 837
 838                 dev_watchdog_down(dev);
 839                 sync_needed |= !dev->dismantle;
 840         }
 841
 842         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 843          * This is avoided if all devices are in dismantle phase :
 844          * Caller will call synchronize_net() for us
 845          */
 846         if (sync_needed)
 847                 synchronize_net();
 848
 849         /* Wait for outstanding qdisc_run calls. */
 850         list_for_each_entry(dev, head, unreg_list)
 851                 while (some_qdisc_is_busy(dev))
 852                         yield();
 853 }
 854
 855 void dev_deactivate(struct net_device *dev)
 856 {
 857         LIST_HEAD(single);
 858
 859         list_add(&dev->unreg_list, &single);
 860         dev_deactivate_many(&single);
 861         list_del(&single);
 862 }
 863 EXPORT_SYMBOL(dev_deactivate);
 864
 865 static void dev_init_scheduler_queue(struct net_device *dev,
 866                                      struct netdev_queue *dev_queue,
 867                                      void *_qdisc)
 868 {
 869         struct Qdisc *qdisc = _qdisc;
 870
 871         dev_queue->qdisc = qdisc;
 872         dev_queue->qdisc_sleeping = qdisc;
 873 }
 874
 875 void dev_init_scheduler(struct net_device *dev)
 876 {
 877         dev->qdisc = &noop_qdisc;
 878         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 879         if (dev_ingress_queue(dev))
 880                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 881
 882         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 883 }
 884
 885 static void shutdown_scheduler_queue(struct net_device *dev,
 886                                      struct netdev_queue *dev_queue,
 887                                      void *_qdisc_default)
 888 {
 889         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 890         struct Qdisc *qdisc_default = _qdisc_default;
 891
 892         if (qdisc) {
 893                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 894                 dev_queue->qdisc_sleeping = qdisc_default;
 895
 896                 qdisc_destroy(qdisc);
 897         }
 898 }
 899
 900 void dev_shutdown(struct net_device *dev)
 901 {
 902         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 903         if (dev_ingress_queue(dev))
 904                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 905         qdisc_destroy(dev->qdisc);
 906         dev->qdisc = &noop_qdisc;
 907
 908         WARN_ON(timer_pending(&dev->watchdog_timer));
 909 }