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         const struct netdev_queue *txq = q->dev_queue;
  57
  58         if (unlikely(skb)) {
  59                 /* check the reason of requeuing without tx lock first */
  60                 txq = netdev_get_tx_queue(txq->dev, skb_get_queue_mapping(skb));
  61                 if (!netif_xmit_frozen_or_stopped(txq)) {
  62                         q->gso_skb = NULL;
  63                         q->q.qlen--;
  64                 } else
  65                         skb = NULL;
  66         } else {
  67                 if (!(q->flags & TCQ_F_ONETXQUEUE) || !netif_xmit_frozen_or_stopped(txq))
  68                         skb = q->dequeue(q);
  69         }
  70
  71         return skb;
  72 }
  73
  74 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  75                                            struct netdev_queue *dev_queue,
  76                                            struct Qdisc *q)
  77 {
  78         int ret;
  79
  80         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  81                 /*
  82                  * Same CPU holding the lock. It may be a transient
  83                  * configuration error, when hard_start_xmit() recurses. We
  84                  * detect it by checking xmit owner and drop the packet when
  85                  * deadloop is detected. Return OK to try the next skb.
  86                  */
  87                 kfree_skb(skb);
  88                 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
  89                                      dev_queue->dev->name);
  90                 ret = qdisc_qlen(q);
  91         } else {
  92                 /*
  93                  * Another cpu is holding lock, requeue & delay xmits for
  94                  * some time.
  95                  */
  96                 __this_cpu_inc(softnet_data.cpu_collision);
  97                 ret = dev_requeue_skb(skb, q);
  98         }
  99
 100         return ret;
 101 }
 102
 103 /*
 104  * Transmit one skb, and handle the return status as required. Holding the
 105  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 106  * function.
 107  *
 108  * Returns to the caller:
 109  *                              0  - queue is empty or throttled.
 110  *                              >0 - queue is not empty.
 111  */
 112 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 113                     struct net_device *dev, struct netdev_queue *txq,
 114                     spinlock_t *root_lock)
 115 {
 116         int ret = NETDEV_TX_BUSY;
 117
 118         /* And release qdisc */
 119         spin_unlock(root_lock);
 120
 121         HARD_TX_LOCK(dev, txq, smp_processor_id());
 122         if (!netif_xmit_frozen_or_stopped(txq))
 123                 ret = dev_hard_start_xmit(skb, dev, txq);
 124
 125         HARD_TX_UNLOCK(dev, txq);
 126
 127         spin_lock(root_lock);
 128
 129         if (dev_xmit_complete(ret)) {
 130                 /* Driver sent out skb successfully or skb was consumed */
 131                 ret = qdisc_qlen(q);
 132         } else if (ret == NETDEV_TX_LOCKED) {
 133                 /* Driver try lock failed */
 134                 ret = handle_dev_cpu_collision(skb, txq, q);
 135         } else {
 136                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 137                 if (unlikely(ret != NETDEV_TX_BUSY))
 138                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 139                                              dev->name, ret, q->q.qlen);
 140
 141                 ret = dev_requeue_skb(skb, q);
 142         }
 143
 144         if (ret && netif_xmit_frozen_or_stopped(txq))
 145                 ret = 0;
 146
 147         return ret;
 148 }
 149
 150 /*
 151  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 152  *
 153  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 154  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 155  * this queue.
 156  *
 157  *  netif_tx_lock serializes accesses to device driver.
 158  *
 159  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 160  *  if one is grabbed, another must be free.
 161  *
 162  * Note, that this procedure can be called by a watchdog timer
 163  *
 164  * Returns to the caller:
 165  *                              0  - queue is empty or throttled.
 166  *                              >0 - queue is not empty.
 167  *
 168  */
 169 static inline int qdisc_restart(struct Qdisc *q)
 170 {
 171         struct netdev_queue *txq;
 172         struct net_device *dev;
 173         spinlock_t *root_lock;
 174         struct sk_buff *skb;
 175
 176         /* Dequeue packet */
 177         skb = dequeue_skb(q);
 178         if (unlikely(!skb))
 179                 return 0;
 180         WARN_ON_ONCE(skb_dst_is_noref(skb));
 181         root_lock = qdisc_lock(q);
 182         dev = qdisc_dev(q);
 183         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 184
 185         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 186 }
 187
 188 void __qdisc_run(struct Qdisc *q)
 189 {
 190         int quota = weight_p;
 191
 192         while (qdisc_restart(q)) {
 193                 /*
 194                  * Ordered by possible occurrence: Postpone processing if
 195                  * 1. we've exceeded packet quota
 196                  * 2. another process needs the CPU;
 197                  */
 198                 if (--quota <= 0 || need_resched()) {
 199                         __netif_schedule(q);
 200                         break;
 201                 }
 202         }
 203
 204         qdisc_run_end(q);
 205 }
 206
 207 unsigned long dev_trans_start(struct net_device *dev)
 208 {
 209         unsigned long val, res = dev->trans_start;
 210         unsigned int i;
 211
 212         for (i = 0; i < dev->num_tx_queues; i++) {
 213                 val = netdev_get_tx_queue(dev, i)->trans_start;
 214                 if (val && time_after(val, res))
 215                         res = val;
 216         }
 217         dev->trans_start = res;
 218         return res;
 219 }
 220 EXPORT_SYMBOL(dev_trans_start);
 221
 222 static void dev_watchdog(unsigned long arg)
 223 {
 224         struct net_device *dev = (struct net_device *)arg;
 225
 226         netif_tx_lock(dev);
 227         if (!qdisc_tx_is_noop(dev)) {
 228                 if (netif_device_present(dev) &&
 229                     netif_running(dev) &&
 230                     netif_carrier_ok(dev)) {
 231                         int some_queue_timedout = 0;
 232                         unsigned int i;
 233                         unsigned long trans_start;
 234
 235                         for (i = 0; i < dev->num_tx_queues; i++) {
 236                                 struct netdev_queue *txq;
 237
 238                                 txq = netdev_get_tx_queue(dev, i);
 239                                 /*
 240                                  * old device drivers set dev->trans_start
 241                                  */
 242                                 trans_start = txq->trans_start ? : dev->trans_start;
 243                                 if (netif_xmit_stopped(txq) &&
 244                                     time_after(jiffies, (trans_start +
 245                                                          dev->watchdog_timeo))) {
 246                                         some_queue_timedout = 1;
 247                                         txq->trans_timeout++;
 248                                         break;
 249                                 }
 250                         }
 251
 252                         if (some_queue_timedout) {
 253                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 254                                        dev->name, netdev_drivername(dev), i);
 255                                 dev->netdev_ops->ndo_tx_timeout(dev);
 256                         }
 257                         if (!mod_timer(&dev->watchdog_timer,
 258                                        round_jiffies(jiffies +
 259                                                      dev->watchdog_timeo)))
 260                                 dev_hold(dev);
 261                 }
 262         }
 263         netif_tx_unlock(dev);
 264
 265         dev_put(dev);
 266 }
 267
 268 void __netdev_watchdog_up(struct net_device *dev)
 269 {
 270         if (dev->netdev_ops->ndo_tx_timeout) {
 271                 if (dev->watchdog_timeo <= 0)
 272                         dev->watchdog_timeo = 5*HZ;
 273                 if (!mod_timer(&dev->watchdog_timer,
 274                                round_jiffies(jiffies + dev->watchdog_timeo)))
 275                         dev_hold(dev);
 276         }
 277 }
 278
 279 static void dev_watchdog_up(struct net_device *dev)
 280 {
 281         __netdev_watchdog_up(dev);
 282 }
 283
 284 static void dev_watchdog_down(struct net_device *dev)
 285 {
 286         netif_tx_lock_bh(dev);
 287         if (del_timer(&dev->watchdog_timer))
 288                 dev_put(dev);
 289         netif_tx_unlock_bh(dev);
 290 }
 291
 292 /**
 293  *      netif_carrier_on - set carrier
 294  *      @dev: network device
 295  *
 296  * Device has detected that carrier.
 297  */
 298 void netif_carrier_on(struct net_device *dev)
 299 {
 300         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 301                 if (dev->reg_state == NETREG_UNINITIALIZED)
 302                         return;
 303                 linkwatch_fire_event(dev);
 304                 if (netif_running(dev))
 305                         __netdev_watchdog_up(dev);
 306         }
 307 }
 308 EXPORT_SYMBOL(netif_carrier_on);
 309
 310 /**
 311  *      netif_carrier_off - clear carrier
 312  *      @dev: network device
 313  *
 314  * Device has detected loss of carrier.
 315  */
 316 void netif_carrier_off(struct net_device *dev)
 317 {
 318         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 319                 if (dev->reg_state == NETREG_UNINITIALIZED)
 320                         return;
 321                 linkwatch_fire_event(dev);
 322         }
 323 }
 324 EXPORT_SYMBOL(netif_carrier_off);
 325
 326 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 327    under all circumstances. It is difficult to invent anything faster or
 328    cheaper.
 329  */
 330
 331 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 332 {
 333         kfree_skb(skb);
 334         return NET_XMIT_CN;
 335 }
 336
 337 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 338 {
 339         return NULL;
 340 }
 341
 342 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 343         .id             =       "noop",
 344         .priv_size      =       0,
 345         .enqueue        =       noop_enqueue,
 346         .dequeue        =       noop_dequeue,
 347         .peek           =       noop_dequeue,
 348         .owner          =       THIS_MODULE,
 349 };
 350
 351 static struct netdev_queue noop_netdev_queue = {
 352         .qdisc          =       &noop_qdisc,
 353         .qdisc_sleeping =       &noop_qdisc,
 354 };
 355
 356 struct Qdisc noop_qdisc = {
 357         .enqueue        =       noop_enqueue,
 358         .dequeue        =       noop_dequeue,
 359         .flags          =       TCQ_F_BUILTIN,
 360         .ops            =       &noop_qdisc_ops,
 361         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 362         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 363         .dev_queue      =       &noop_netdev_queue,
 364         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 365 };
 366 EXPORT_SYMBOL(noop_qdisc);
 367
 368 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 369         .id             =       "noqueue",
 370         .priv_size      =       0,
 371         .enqueue        =       noop_enqueue,
 372         .dequeue        =       noop_dequeue,
 373         .peek           =       noop_dequeue,
 374         .owner          =       THIS_MODULE,
 375 };
 376
 377 static struct Qdisc noqueue_qdisc;
 378 static struct netdev_queue noqueue_netdev_queue = {
 379         .qdisc          =       &noqueue_qdisc,
 380         .qdisc_sleeping =       &noqueue_qdisc,
 381 };
 382
 383 static struct Qdisc noqueue_qdisc = {
 384         .enqueue        =       NULL,
 385         .dequeue        =       noop_dequeue,
 386         .flags          =       TCQ_F_BUILTIN,
 387         .ops            =       &noqueue_qdisc_ops,
 388         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 389         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 390         .dev_queue      =       &noqueue_netdev_queue,
 391         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 392 };
 393
 394
 395 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 396         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 397 };
 398
 399 /* 3-band FIFO queue: old style, but should be a bit faster than
 400    generic prio+fifo combination.
 401  */
 402
 403 #define PFIFO_FAST_BANDS 3
 404
 405 /*
 406  * Private data for a pfifo_fast scheduler containing:
 407  *      - queues for the three band
 408  *      - bitmap indicating which of the bands contain skbs
 409  */
 410 struct pfifo_fast_priv {
 411         u32 bitmap;
 412         struct sk_buff_head q[PFIFO_FAST_BANDS];
 413 };
 414
 415 /*
 416  * Convert a bitmap to the first band number where an skb is queued, where:
 417  *      bitmap=0 means there are no skbs on any band.
 418  *      bitmap=1 means there is an skb on band 0.
 419  *      bitmap=7 means there are skbs on all 3 bands, etc.
 420  */
 421 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 422
 423 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 424                                              int band)
 425 {
 426         return priv->q + band;
 427 }
 428
 429 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 430 {
 431         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 432                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 433                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 434                 struct sk_buff_head *list = band2list(priv, band);
 435
 436                 priv->bitmap |= (1 << band);
 437                 qdisc->q.qlen++;
 438                 return __qdisc_enqueue_tail(skb, qdisc, list);
 439         }
 440
 441         return qdisc_drop(skb, qdisc);
 442 }
 443
 444 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 445 {
 446         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 447         int band = bitmap2band[priv->bitmap];
 448
 449         if (likely(band >= 0)) {
 450                 struct sk_buff_head *list = band2list(priv, band);
 451                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 452
 453                 qdisc->q.qlen--;
 454                 if (skb_queue_empty(list))
 455                         priv->bitmap &= ~(1 << band);
 456
 457                 return skb;
 458         }
 459
 460         return NULL;
 461 }
 462
 463 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 464 {
 465         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 466         int band = bitmap2band[priv->bitmap];
 467
 468         if (band >= 0) {
 469                 struct sk_buff_head *list = band2list(priv, band);
 470
 471                 return skb_peek(list);
 472         }
 473
 474         return NULL;
 475 }
 476
 477 static void pfifo_fast_reset(struct Qdisc *qdisc)
 478 {
 479         int prio;
 480         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 481
 482         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 483                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 484
 485         priv->bitmap = 0;
 486         qdisc->qstats.backlog = 0;
 487         qdisc->q.qlen = 0;
 488 }
 489
 490 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 491 {
 492         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 493
 494         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 495         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 496                 goto nla_put_failure;
 497         return skb->len;
 498
 499 nla_put_failure:
 500         return -1;
 501 }
 502
 503 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 504 {
 505         int prio;
 506         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 507
 508         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 509                 skb_queue_head_init(band2list(priv, prio));
 510
 511         /* Can by-pass the queue discipline */
 512         qdisc->flags |= TCQ_F_CAN_BYPASS;
 513         return 0;
 514 }
 515
 516 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 517         .id             =       "pfifo_fast",
 518         .priv_size      =       sizeof(struct pfifo_fast_priv),
 519         .enqueue        =       pfifo_fast_enqueue,
 520         .dequeue        =       pfifo_fast_dequeue,
 521         .peek           =       pfifo_fast_peek,
 522         .init           =       pfifo_fast_init,
 523         .reset          =       pfifo_fast_reset,
 524         .dump           =       pfifo_fast_dump,
 525         .owner          =       THIS_MODULE,
 526 };
 527 EXPORT_SYMBOL(pfifo_fast_ops);
 528
 529 static struct lock_class_key qdisc_tx_busylock;
 530
 531 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 532                           struct Qdisc_ops *ops)
 533 {
 534         void *p;
 535         struct Qdisc *sch;
 536         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 537         int err = -ENOBUFS;
 538         struct net_device *dev = dev_queue->dev;
 539
 540         p = kzalloc_node(size, GFP_KERNEL,
 541                          netdev_queue_numa_node_read(dev_queue));
 542
 543         if (!p)
 544                 goto errout;
 545         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 546         /* if we got non aligned memory, ask more and do alignment ourself */
 547         if (sch != p) {
 548                 kfree(p);
 549                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 550                                  netdev_queue_numa_node_read(dev_queue));
 551                 if (!p)
 552                         goto errout;
 553                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 554                 sch->padded = (char *) sch - (char *) p;
 555         }
 556         INIT_LIST_HEAD(&sch->list);
 557         skb_queue_head_init(&sch->q);
 558
 559         spin_lock_init(&sch->busylock);
 560         lockdep_set_class(&sch->busylock,
 561                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 562
 563         sch->ops = ops;
 564         sch->enqueue = ops->enqueue;
 565         sch->dequeue = ops->dequeue;
 566         sch->dev_queue = dev_queue;
 567         dev_hold(dev);
 568         atomic_set(&sch->refcnt, 1);
 569
 570         return sch;
 571 errout:
 572         return ERR_PTR(err);
 573 }
 574
 575 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 576                                 struct Qdisc_ops *ops, unsigned int parentid)
 577 {
 578         struct Qdisc *sch;
 579
 580         sch = qdisc_alloc(dev_queue, ops);
 581         if (IS_ERR(sch))
 582                 goto errout;
 583         sch->parent = parentid;
 584
 585         if (!ops->init || ops->init(sch, NULL) == 0)
 586                 return sch;
 587
 588         qdisc_destroy(sch);
 589 errout:
 590         return NULL;
 591 }
 592 EXPORT_SYMBOL(qdisc_create_dflt);
 593
 594 /* Under qdisc_lock(qdisc) and BH! */
 595
 596 void qdisc_reset(struct Qdisc *qdisc)
 597 {
 598         const struct Qdisc_ops *ops = qdisc->ops;
 599
 600         if (ops->reset)
 601                 ops->reset(qdisc);
 602
 603         if (qdisc->gso_skb) {
 604                 kfree_skb(qdisc->gso_skb);
 605                 qdisc->gso_skb = NULL;
 606                 qdisc->q.qlen = 0;
 607         }
 608 }
 609 EXPORT_SYMBOL(qdisc_reset);
 610
 611 static void qdisc_rcu_free(struct rcu_head *head)
 612 {
 613         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 614
 615         kfree((char *) qdisc - qdisc->padded);
 616 }
 617
 618 void qdisc_destroy(struct Qdisc *qdisc)
 619 {
 620         const struct Qdisc_ops  *ops = qdisc->ops;
 621
 622         if (qdisc->flags & TCQ_F_BUILTIN ||
 623             !atomic_dec_and_test(&qdisc->refcnt))
 624                 return;
 625
 626 #ifdef CONFIG_NET_SCHED
 627         qdisc_list_del(qdisc);
 628
 629         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 630 #endif
 631         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 632         if (ops->reset)
 633                 ops->reset(qdisc);
 634         if (ops->destroy)
 635                 ops->destroy(qdisc);
 636
 637         module_put(ops->owner);
 638         dev_put(qdisc_dev(qdisc));
 639
 640         kfree_skb(qdisc->gso_skb);
 641         /*
 642          * gen_estimator est_timer() might access qdisc->q.lock,
 643          * wait a RCU grace period before freeing qdisc.
 644          */
 645         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 646 }
 647 EXPORT_SYMBOL(qdisc_destroy);
 648
 649 /* Attach toplevel qdisc to device queue. */
 650 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 651                               struct Qdisc *qdisc)
 652 {
 653         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 654         spinlock_t *root_lock;
 655
 656         root_lock = qdisc_lock(oqdisc);
 657         spin_lock_bh(root_lock);
 658
 659         /* Prune old scheduler */
 660         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 661                 qdisc_reset(oqdisc);
 662
 663         /* ... and graft new one */
 664         if (qdisc == NULL)
 665                 qdisc = &noop_qdisc;
 666         dev_queue->qdisc_sleeping = qdisc;
 667         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 668
 669         spin_unlock_bh(root_lock);
 670
 671         return oqdisc;
 672 }
 673 EXPORT_SYMBOL(dev_graft_qdisc);
 674
 675 static void attach_one_default_qdisc(struct net_device *dev,
 676                                      struct netdev_queue *dev_queue,
 677                                      void *_unused)
 678 {
 679         struct Qdisc *qdisc = &noqueue_qdisc;
 680
 681         if (dev->tx_queue_len) {
 682                 qdisc = qdisc_create_dflt(dev_queue,
 683                                           &pfifo_fast_ops, TC_H_ROOT);
 684                 if (!qdisc) {
 685                         netdev_info(dev, "activation failed\n");
 686                         return;
 687                 }
 688                 if (!netif_is_multiqueue(dev))
 689                         qdisc->flags |= TCQ_F_ONETXQUEUE;
 690         }
 691         dev_queue->qdisc_sleeping = qdisc;
 692 }
 693
 694 static void attach_default_qdiscs(struct net_device *dev)
 695 {
 696         struct netdev_queue *txq;
 697         struct Qdisc *qdisc;
 698
 699         txq = netdev_get_tx_queue(dev, 0);
 700
 701         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 702                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 703                 dev->qdisc = txq->qdisc_sleeping;
 704                 atomic_inc(&dev->qdisc->refcnt);
 705         } else {
 706                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 707                 if (qdisc) {
 708                         qdisc->ops->attach(qdisc);
 709                         dev->qdisc = qdisc;
 710                 }
 711         }
 712 }
 713
 714 static void transition_one_qdisc(struct net_device *dev,
 715                                  struct netdev_queue *dev_queue,
 716                                  void *_need_watchdog)
 717 {
 718         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 719         int *need_watchdog_p = _need_watchdog;
 720
 721         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 722                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 723
 724         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 725         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 726                 dev_queue->trans_start = 0;
 727                 *need_watchdog_p = 1;
 728         }
 729 }
 730
 731 void dev_activate(struct net_device *dev)
 732 {
 733         int need_watchdog;
 734
 735         /* No queueing discipline is attached to device;
 736            create default one i.e. pfifo_fast for devices,
 737            which need queueing and noqueue_qdisc for
 738            virtual interfaces
 739          */
 740
 741         if (dev->qdisc == &noop_qdisc)
 742                 attach_default_qdiscs(dev);
 743
 744         if (!netif_carrier_ok(dev))
 745                 /* Delay activation until next carrier-on event */
 746                 return;
 747
 748         need_watchdog = 0;
 749         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 750         if (dev_ingress_queue(dev))
 751                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 752
 753         if (need_watchdog) {
 754                 dev->trans_start = jiffies;
 755                 dev_watchdog_up(dev);
 756         }
 757 }
 758 EXPORT_SYMBOL(dev_activate);
 759
 760 static void dev_deactivate_queue(struct net_device *dev,
 761                                  struct netdev_queue *dev_queue,
 762                                  void *_qdisc_default)
 763 {
 764         struct Qdisc *qdisc_default = _qdisc_default;
 765         struct Qdisc *qdisc;
 766
 767         qdisc = dev_queue->qdisc;
 768         if (qdisc) {
 769                 spin_lock_bh(qdisc_lock(qdisc));
 770
 771                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 772                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 773
 774                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 775                 qdisc_reset(qdisc);
 776
 777                 spin_unlock_bh(qdisc_lock(qdisc));
 778         }
 779 }
 780
 781 static bool some_qdisc_is_busy(struct net_device *dev)
 782 {
 783         unsigned int i;
 784
 785         for (i = 0; i < dev->num_tx_queues; i++) {
 786                 struct netdev_queue *dev_queue;
 787                 spinlock_t *root_lock;
 788                 struct Qdisc *q;
 789                 int val;
 790
 791                 dev_queue = netdev_get_tx_queue(dev, i);
 792                 q = dev_queue->qdisc_sleeping;
 793                 root_lock = qdisc_lock(q);
 794
 795                 spin_lock_bh(root_lock);
 796
 797                 val = (qdisc_is_running(q) ||
 798                        test_bit(__QDISC_STATE_SCHED, &q->state));
 799
 800                 spin_unlock_bh(root_lock);
 801
 802                 if (val)
 803                         return true;
 804         }
 805         return false;
 806 }
 807
 808 /**
 809  *      dev_deactivate_many - deactivate transmissions on several devices
 810  *      @head: list of devices to deactivate
 811  *
 812  *      This function returns only when all outstanding transmissions
 813  *      have completed, unless all devices are in dismantle phase.
 814  */
 815 void dev_deactivate_many(struct list_head *head)
 816 {
 817         struct net_device *dev;
 818         bool sync_needed = false;
 819
 820         list_for_each_entry(dev, head, unreg_list) {
 821                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 822                                          &noop_qdisc);
 823                 if (dev_ingress_queue(dev))
 824                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 825                                              &noop_qdisc);
 826
 827                 dev_watchdog_down(dev);
 828                 sync_needed |= !dev->dismantle;
 829         }
 830
 831         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 832          * This is avoided if all devices are in dismantle phase :
 833          * Caller will call synchronize_net() for us
 834          */
 835         if (sync_needed)
 836                 synchronize_net();
 837
 838         /* Wait for outstanding qdisc_run calls. */
 839         list_for_each_entry(dev, head, unreg_list)
 840                 while (some_qdisc_is_busy(dev))
 841                         yield();
 842 }
 843
 844 void dev_deactivate(struct net_device *dev)
 845 {
 846         LIST_HEAD(single);
 847
 848         list_add(&dev->unreg_list, &single);
 849         dev_deactivate_many(&single);
 850         list_del(&single);
 851 }
 852 EXPORT_SYMBOL(dev_deactivate);
 853
 854 static void dev_init_scheduler_queue(struct net_device *dev,
 855                                      struct netdev_queue *dev_queue,
 856                                      void *_qdisc)
 857 {
 858         struct Qdisc *qdisc = _qdisc;
 859
 860         dev_queue->qdisc = qdisc;
 861         dev_queue->qdisc_sleeping = qdisc;
 862 }
 863
 864 void dev_init_scheduler(struct net_device *dev)
 865 {
 866         dev->qdisc = &noop_qdisc;
 867         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 868         if (dev_ingress_queue(dev))
 869                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 870
 871         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 872 }
 873
 874 static void shutdown_scheduler_queue(struct net_device *dev,
 875                                      struct netdev_queue *dev_queue,
 876                                      void *_qdisc_default)
 877 {
 878         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 879         struct Qdisc *qdisc_default = _qdisc_default;
 880
 881         if (qdisc) {
 882                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 883                 dev_queue->qdisc_sleeping = qdisc_default;
 884
 885                 qdisc_destroy(qdisc);
 886         }
 887 }
 888
 889 void dev_shutdown(struct net_device *dev)
 890 {
 891         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 892         if (dev_ingress_queue(dev))
 893                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 894         qdisc_destroy(dev->qdisc);
 895         dev->qdisc = &noop_qdisc;
 896
 897         WARN_ON(timer_pending(&dev->watchdog_timer));
 898 }