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 <linux/if_vlan.h>
  29 #include <net/sch_generic.h>
  30 #include <net/pkt_sched.h>
  31 #include <net/dst.h>
  32
  33 /* Qdisc to use by default */
  34 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  35 EXPORT_SYMBOL(default_qdisc_ops);
  36
  37 /* Main transmission queue. */
  38
  39 /* Modifications to data participating in scheduling must be protected with
  40  * qdisc_lock(qdisc) spinlock.
  41  *
  42  * The idea is the following:
  43  * - enqueue, dequeue are serialized via qdisc root lock
  44  * - ingress filtering is also serialized via qdisc root lock
  45  * - updates to tree and tree walking are only done under the rtnl mutex.
  46  */
  47
  48 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  49 {
  50         q->gso_skb = skb;
  51         q->qstats.requeues++;
  52         q->q.qlen++;    /* it's still part of the queue */
  53         __netif_schedule(q);
  54
  55         return 0;
  56 }
  57
  58 static void try_bulk_dequeue_skb(struct Qdisc *q,
  59                                  struct sk_buff *skb,
  60                                  const struct netdev_queue *txq,
  61                                  int *packets)
  62 {
  63         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  64
  65         while (bytelimit > 0) {
  66                 struct sk_buff *nskb = q->dequeue(q);
  67
  68                 if (!nskb)
  69                         break;
  70
  71                 bytelimit -= nskb->len; /* covers GSO len */
  72                 skb->next = nskb;
  73                 skb = nskb;
  74                 (*packets)++; /* GSO counts as one pkt */
  75         }
  76         skb->next = NULL;
  77 }
  78
  79 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
  80  * A requeued skb (via q->gso_skb) can also be a SKB list.
  81  */
  82 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
  83                                    int *packets)
  84 {
  85         struct sk_buff *skb = q->gso_skb;
  86         const struct netdev_queue *txq = q->dev_queue;
  87
  88         *packets = 1;
  89         *validate = true;
  90         if (unlikely(skb)) {
  91                 /* check the reason of requeuing without tx lock first */
  92                 txq = skb_get_tx_queue(txq->dev, skb);
  93                 if (!netif_xmit_frozen_or_stopped(txq)) {
  94                         q->gso_skb = NULL;
  95                         q->q.qlen--;
  96                 } else
  97                         skb = NULL;
  98                 /* skb in gso_skb were already validated */
  99                 *validate = false;
 100         } else {
 101                 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
 102                     !netif_xmit_frozen_or_stopped(txq)) {
 103                         skb = q->dequeue(q);
 104                         if (skb && qdisc_may_bulk(q))
 105                                 try_bulk_dequeue_skb(q, skb, txq, packets);
 106                 }
 107         }
 108         return skb;
 109 }
 110
 111 /*
 112  * Transmit possibly several skbs, and handle the return status as
 113  * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
 114  * only one CPU can execute this function.
 115  *
 116  * Returns to the caller:
 117  *                              0  - queue is empty or throttled.
 118  *                              >0 - queue is not empty.
 119  */
 120 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 121                     struct net_device *dev, struct netdev_queue *txq,
 122                     spinlock_t *root_lock, bool validate)
 123 {
 124         int ret = NETDEV_TX_BUSY;
 125
 126         /* And release qdisc */
 127         spin_unlock(root_lock);
 128
 129         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 130         if (validate)
 131                 skb = validate_xmit_skb_list(skb, dev);
 132
 133         if (likely(skb)) {
 134                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 135                 if (!netif_xmit_frozen_or_stopped(txq))
 136                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 137
 138                 HARD_TX_UNLOCK(dev, txq);
 139         } else {
 140                 spin_lock(root_lock);
 141                 return qdisc_qlen(q);
 142         }
 143         spin_lock(root_lock);
 144
 145         if (dev_xmit_complete(ret)) {
 146                 /* Driver sent out skb successfully or skb was consumed */
 147                 ret = qdisc_qlen(q);
 148         } else {
 149                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 150                 if (unlikely(ret != NETDEV_TX_BUSY))
 151                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 152                                              dev->name, ret, q->q.qlen);
 153
 154                 ret = dev_requeue_skb(skb, q);
 155         }
 156
 157         if (ret && netif_xmit_frozen_or_stopped(txq))
 158                 ret = 0;
 159
 160         return ret;
 161 }
 162
 163 /*
 164  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 165  *
 166  * __QDISC___STATE_RUNNING guarantees only one CPU can process
 167  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 168  * this queue.
 169  *
 170  *  netif_tx_lock serializes accesses to device driver.
 171  *
 172  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 173  *  if one is grabbed, another must be free.
 174  *
 175  * Note, that this procedure can be called by a watchdog timer
 176  *
 177  * Returns to the caller:
 178  *                              0  - queue is empty or throttled.
 179  *                              >0 - queue is not empty.
 180  *
 181  */
 182 static inline int qdisc_restart(struct Qdisc *q, int *packets)
 183 {
 184         struct netdev_queue *txq;
 185         struct net_device *dev;
 186         spinlock_t *root_lock;
 187         struct sk_buff *skb;
 188         bool validate;
 189
 190         /* Dequeue packet */
 191         skb = dequeue_skb(q, &validate, packets);
 192         if (unlikely(!skb))
 193                 return 0;
 194
 195         root_lock = qdisc_lock(q);
 196         dev = qdisc_dev(q);
 197         txq = skb_get_tx_queue(dev, skb);
 198
 199         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 200 }
 201
 202 void __qdisc_run(struct Qdisc *q)
 203 {
 204         int quota = weight_p;
 205         int packets;
 206
 207         while (qdisc_restart(q, &packets)) {
 208                 /*
 209                  * Ordered by possible occurrence: Postpone processing if
 210                  * 1. we've exceeded packet quota
 211                  * 2. another process needs the CPU;
 212                  */
 213                 quota -= packets;
 214                 if (quota <= 0 || need_resched()) {
 215                         __netif_schedule(q);
 216                         break;
 217                 }
 218         }
 219
 220         qdisc_run_end(q);
 221 }
 222
 223 unsigned long dev_trans_start(struct net_device *dev)
 224 {
 225         unsigned long val, res;
 226         unsigned int i;
 227
 228         if (is_vlan_dev(dev))
 229                 dev = vlan_dev_real_dev(dev);
 230         res = dev->trans_start;
 231         for (i = 0; i < dev->num_tx_queues; i++) {
 232                 val = netdev_get_tx_queue(dev, i)->trans_start;
 233                 if (val && time_after(val, res))
 234                         res = val;
 235         }
 236         dev->trans_start = res;
 237
 238         return res;
 239 }
 240 EXPORT_SYMBOL(dev_trans_start);
 241
 242 static void dev_watchdog(unsigned long arg)
 243 {
 244         struct net_device *dev = (struct net_device *)arg;
 245
 246         netif_tx_lock(dev);
 247         if (!qdisc_tx_is_noop(dev)) {
 248                 if (netif_device_present(dev) &&
 249                     netif_running(dev) &&
 250                     netif_carrier_ok(dev)) {
 251                         int some_queue_timedout = 0;
 252                         unsigned int i;
 253                         unsigned long trans_start;
 254
 255                         for (i = 0; i < dev->num_tx_queues; i++) {
 256                                 struct netdev_queue *txq;
 257
 258                                 txq = netdev_get_tx_queue(dev, i);
 259                                 /*
 260                                  * old device drivers set dev->trans_start
 261                                  */
 262                                 trans_start = txq->trans_start ? : dev->trans_start;
 263                                 if (netif_xmit_stopped(txq) &&
 264                                     time_after(jiffies, (trans_start +
 265                                                          dev->watchdog_timeo))) {
 266                                         some_queue_timedout = 1;
 267                                         txq->trans_timeout++;
 268                                         break;
 269                                 }
 270                         }
 271
 272                         if (some_queue_timedout) {
 273                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 274                                        dev->name, netdev_drivername(dev), i);
 275                                 dev->netdev_ops->ndo_tx_timeout(dev);
 276                         }
 277                         if (!mod_timer(&dev->watchdog_timer,
 278                                        round_jiffies(jiffies +
 279                                                      dev->watchdog_timeo)))
 280                                 dev_hold(dev);
 281                 }
 282         }
 283         netif_tx_unlock(dev);
 284
 285         dev_put(dev);
 286 }
 287
 288 void __netdev_watchdog_up(struct net_device *dev)
 289 {
 290         if (dev->netdev_ops->ndo_tx_timeout) {
 291                 if (dev->watchdog_timeo <= 0)
 292                         dev->watchdog_timeo = 5*HZ;
 293                 if (!mod_timer(&dev->watchdog_timer,
 294                                round_jiffies(jiffies + dev->watchdog_timeo)))
 295                         dev_hold(dev);
 296         }
 297 }
 298
 299 static void dev_watchdog_up(struct net_device *dev)
 300 {
 301         __netdev_watchdog_up(dev);
 302 }
 303
 304 static void dev_watchdog_down(struct net_device *dev)
 305 {
 306         netif_tx_lock_bh(dev);
 307         if (del_timer(&dev->watchdog_timer))
 308                 dev_put(dev);
 309         netif_tx_unlock_bh(dev);
 310 }
 311
 312 /**
 313  *      netif_carrier_on - set carrier
 314  *      @dev: network device
 315  *
 316  * Device has detected that carrier.
 317  */
 318 void netif_carrier_on(struct net_device *dev)
 319 {
 320         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 321                 if (dev->reg_state == NETREG_UNINITIALIZED)
 322                         return;
 323                 atomic_inc(&dev->carrier_changes);
 324                 linkwatch_fire_event(dev);
 325                 if (netif_running(dev))
 326                         __netdev_watchdog_up(dev);
 327         }
 328 }
 329 EXPORT_SYMBOL(netif_carrier_on);
 330
 331 /**
 332  *      netif_carrier_off - clear carrier
 333  *      @dev: network device
 334  *
 335  * Device has detected loss of carrier.
 336  */
 337 void netif_carrier_off(struct net_device *dev)
 338 {
 339         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 340                 if (dev->reg_state == NETREG_UNINITIALIZED)
 341                         return;
 342                 atomic_inc(&dev->carrier_changes);
 343                 linkwatch_fire_event(dev);
 344         }
 345 }
 346 EXPORT_SYMBOL(netif_carrier_off);
 347
 348 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 349    under all circumstances. It is difficult to invent anything faster or
 350    cheaper.
 351  */
 352
 353 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 354 {
 355         kfree_skb(skb);
 356         return NET_XMIT_CN;
 357 }
 358
 359 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 360 {
 361         return NULL;
 362 }
 363
 364 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 365         .id             =       "noop",
 366         .priv_size      =       0,
 367         .enqueue        =       noop_enqueue,
 368         .dequeue        =       noop_dequeue,
 369         .peek           =       noop_dequeue,
 370         .owner          =       THIS_MODULE,
 371 };
 372
 373 static struct netdev_queue noop_netdev_queue = {
 374         .qdisc          =       &noop_qdisc,
 375         .qdisc_sleeping =       &noop_qdisc,
 376 };
 377
 378 struct Qdisc noop_qdisc = {
 379         .enqueue        =       noop_enqueue,
 380         .dequeue        =       noop_dequeue,
 381         .flags          =       TCQ_F_BUILTIN,
 382         .ops            =       &noop_qdisc_ops,
 383         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 384         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 385         .dev_queue      =       &noop_netdev_queue,
 386         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 387 };
 388 EXPORT_SYMBOL(noop_qdisc);
 389
 390 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 391 {
 392         /* register_qdisc() assigns a default of noop_enqueue if unset,
 393          * but __dev_queue_xmit() treats noqueue only as such
 394          * if this is NULL - so clear it here. */
 395         qdisc->enqueue = NULL;
 396         return 0;
 397 }
 398
 399 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 400         .id             =       "noqueue",
 401         .priv_size      =       0,
 402         .init           =       noqueue_init,
 403         .enqueue        =       noop_enqueue,
 404         .dequeue        =       noop_dequeue,
 405         .peek           =       noop_dequeue,
 406         .owner          =       THIS_MODULE,
 407 };
 408
 409 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 410         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 411 };
 412
 413 /* 3-band FIFO queue: old style, but should be a bit faster than
 414    generic prio+fifo combination.
 415  */
 416
 417 #define PFIFO_FAST_BANDS 3
 418
 419 /*
 420  * Private data for a pfifo_fast scheduler containing:
 421  *      - queues for the three band
 422  *      - bitmap indicating which of the bands contain skbs
 423  */
 424 struct pfifo_fast_priv {
 425         u32 bitmap;
 426         struct sk_buff_head q[PFIFO_FAST_BANDS];
 427 };
 428
 429 /*
 430  * Convert a bitmap to the first band number where an skb is queued, where:
 431  *      bitmap=0 means there are no skbs on any band.
 432  *      bitmap=1 means there is an skb on band 0.
 433  *      bitmap=7 means there are skbs on all 3 bands, etc.
 434  */
 435 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 436
 437 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 438                                              int band)
 439 {
 440         return priv->q + band;
 441 }
 442
 443 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 444 {
 445         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 446                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 447                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 448                 struct sk_buff_head *list = band2list(priv, band);
 449
 450                 priv->bitmap |= (1 << band);
 451                 qdisc->q.qlen++;
 452                 return __qdisc_enqueue_tail(skb, qdisc, list);
 453         }
 454
 455         return qdisc_drop(skb, qdisc);
 456 }
 457
 458 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 459 {
 460         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 461         int band = bitmap2band[priv->bitmap];
 462
 463         if (likely(band >= 0)) {
 464                 struct sk_buff_head *list = band2list(priv, band);
 465                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 466
 467                 qdisc->q.qlen--;
 468                 if (skb_queue_empty(list))
 469                         priv->bitmap &= ~(1 << band);
 470
 471                 return skb;
 472         }
 473
 474         return NULL;
 475 }
 476
 477 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 478 {
 479         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 480         int band = bitmap2band[priv->bitmap];
 481
 482         if (band >= 0) {
 483                 struct sk_buff_head *list = band2list(priv, band);
 484
 485                 return skb_peek(list);
 486         }
 487
 488         return NULL;
 489 }
 490
 491 static void pfifo_fast_reset(struct Qdisc *qdisc)
 492 {
 493         int prio;
 494         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 495
 496         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 497                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 498
 499         priv->bitmap = 0;
 500         qdisc->qstats.backlog = 0;
 501         qdisc->q.qlen = 0;
 502 }
 503
 504 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 505 {
 506         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 507
 508         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 509         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 510                 goto nla_put_failure;
 511         return skb->len;
 512
 513 nla_put_failure:
 514         return -1;
 515 }
 516
 517 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 518 {
 519         int prio;
 520         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 521
 522         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 523                 __skb_queue_head_init(band2list(priv, prio));
 524
 525         /* Can by-pass the queue discipline */
 526         qdisc->flags |= TCQ_F_CAN_BYPASS;
 527         return 0;
 528 }
 529
 530 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 531         .id             =       "pfifo_fast",
 532         .priv_size      =       sizeof(struct pfifo_fast_priv),
 533         .enqueue        =       pfifo_fast_enqueue,
 534         .dequeue        =       pfifo_fast_dequeue,
 535         .peek           =       pfifo_fast_peek,
 536         .init           =       pfifo_fast_init,
 537         .reset          =       pfifo_fast_reset,
 538         .dump           =       pfifo_fast_dump,
 539         .owner          =       THIS_MODULE,
 540 };
 541 EXPORT_SYMBOL(pfifo_fast_ops);
 542
 543 static struct lock_class_key qdisc_tx_busylock;
 544
 545 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 546                           const struct Qdisc_ops *ops)
 547 {
 548         void *p;
 549         struct Qdisc *sch;
 550         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 551         int err = -ENOBUFS;
 552         struct net_device *dev = dev_queue->dev;
 553
 554         p = kzalloc_node(size, GFP_KERNEL,
 555                          netdev_queue_numa_node_read(dev_queue));
 556
 557         if (!p)
 558                 goto errout;
 559         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 560         /* if we got non aligned memory, ask more and do alignment ourself */
 561         if (sch != p) {
 562                 kfree(p);
 563                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 564                                  netdev_queue_numa_node_read(dev_queue));
 565                 if (!p)
 566                         goto errout;
 567                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 568                 sch->padded = (char *) sch - (char *) p;
 569         }
 570         INIT_LIST_HEAD(&sch->list);
 571         skb_queue_head_init(&sch->q);
 572
 573         spin_lock_init(&sch->busylock);
 574         lockdep_set_class(&sch->busylock,
 575                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 576
 577         sch->ops = ops;
 578         sch->enqueue = ops->enqueue;
 579         sch->dequeue = ops->dequeue;
 580         sch->dev_queue = dev_queue;
 581         dev_hold(dev);
 582         atomic_set(&sch->refcnt, 1);
 583
 584         return sch;
 585 errout:
 586         return ERR_PTR(err);
 587 }
 588
 589 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 590                                 const struct Qdisc_ops *ops,
 591                                 unsigned int parentid)
 592 {
 593         struct Qdisc *sch;
 594
 595         if (!try_module_get(ops->owner))
 596                 goto errout;
 597
 598         sch = qdisc_alloc(dev_queue, ops);
 599         if (IS_ERR(sch))
 600                 goto errout;
 601         sch->parent = parentid;
 602
 603         if (!ops->init || ops->init(sch, NULL) == 0)
 604                 return sch;
 605
 606         qdisc_destroy(sch);
 607 errout:
 608         return NULL;
 609 }
 610 EXPORT_SYMBOL(qdisc_create_dflt);
 611
 612 /* Under qdisc_lock(qdisc) and BH! */
 613
 614 void qdisc_reset(struct Qdisc *qdisc)
 615 {
 616         const struct Qdisc_ops *ops = qdisc->ops;
 617
 618         if (ops->reset)
 619                 ops->reset(qdisc);
 620
 621         if (qdisc->gso_skb) {
 622                 kfree_skb_list(qdisc->gso_skb);
 623                 qdisc->gso_skb = NULL;
 624                 qdisc->q.qlen = 0;
 625         }
 626 }
 627 EXPORT_SYMBOL(qdisc_reset);
 628
 629 static void qdisc_rcu_free(struct rcu_head *head)
 630 {
 631         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 632
 633         if (qdisc_is_percpu_stats(qdisc)) {
 634                 free_percpu(qdisc->cpu_bstats);
 635                 free_percpu(qdisc->cpu_qstats);
 636         }
 637
 638         kfree((char *) qdisc - qdisc->padded);
 639 }
 640
 641 void qdisc_destroy(struct Qdisc *qdisc)
 642 {
 643         const struct Qdisc_ops  *ops = qdisc->ops;
 644
 645         if (qdisc->flags & TCQ_F_BUILTIN ||
 646             !atomic_dec_and_test(&qdisc->refcnt))
 647                 return;
 648
 649 #ifdef CONFIG_NET_SCHED
 650         qdisc_list_del(qdisc);
 651
 652         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 653 #endif
 654         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 655         if (ops->reset)
 656                 ops->reset(qdisc);
 657         if (ops->destroy)
 658                 ops->destroy(qdisc);
 659
 660         module_put(ops->owner);
 661         dev_put(qdisc_dev(qdisc));
 662
 663         kfree_skb_list(qdisc->gso_skb);
 664         /*
 665          * gen_estimator est_timer() might access qdisc->q.lock,
 666          * wait a RCU grace period before freeing qdisc.
 667          */
 668         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 669 }
 670 EXPORT_SYMBOL(qdisc_destroy);
 671
 672 /* Attach toplevel qdisc to device queue. */
 673 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 674                               struct Qdisc *qdisc)
 675 {
 676         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 677         spinlock_t *root_lock;
 678
 679         root_lock = qdisc_lock(oqdisc);
 680         spin_lock_bh(root_lock);
 681
 682         /* Prune old scheduler */
 683         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 684                 qdisc_reset(oqdisc);
 685
 686         /* ... and graft new one */
 687         if (qdisc == NULL)
 688                 qdisc = &noop_qdisc;
 689         dev_queue->qdisc_sleeping = qdisc;
 690         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 691
 692         spin_unlock_bh(root_lock);
 693
 694         return oqdisc;
 695 }
 696 EXPORT_SYMBOL(dev_graft_qdisc);
 697
 698 static void attach_one_default_qdisc(struct net_device *dev,
 699                                      struct netdev_queue *dev_queue,
 700                                      void *_unused)
 701 {
 702         struct Qdisc *qdisc;
 703         const struct Qdisc_ops *ops = default_qdisc_ops;
 704
 705         if (dev->priv_flags & IFF_NO_QUEUE)
 706                 ops = &noqueue_qdisc_ops;
 707
 708         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 709         if (!qdisc) {
 710                 netdev_info(dev, "activation failed\n");
 711                 return;
 712         }
 713         if (!netif_is_multiqueue(dev))
 714                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 715         dev_queue->qdisc_sleeping = qdisc;
 716 }
 717
 718 static void attach_default_qdiscs(struct net_device *dev)
 719 {
 720         struct netdev_queue *txq;
 721         struct Qdisc *qdisc;
 722
 723         txq = netdev_get_tx_queue(dev, 0);
 724
 725         if (!netif_is_multiqueue(dev) ||
 726             dev->priv_flags & IFF_NO_QUEUE) {
 727                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 728                 dev->qdisc = txq->qdisc_sleeping;
 729                 atomic_inc(&dev->qdisc->refcnt);
 730         } else {
 731                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 732                 if (qdisc) {
 733                         dev->qdisc = qdisc;
 734                         qdisc->ops->attach(qdisc);
 735                 }
 736         }
 737 }
 738
 739 static void transition_one_qdisc(struct net_device *dev,
 740                                  struct netdev_queue *dev_queue,
 741                                  void *_need_watchdog)
 742 {
 743         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 744         int *need_watchdog_p = _need_watchdog;
 745
 746         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 747                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 748
 749         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 750         if (need_watchdog_p) {
 751                 dev_queue->trans_start = 0;
 752                 *need_watchdog_p = 1;
 753         }
 754 }
 755
 756 void dev_activate(struct net_device *dev)
 757 {
 758         int need_watchdog;
 759
 760         /* No queueing discipline is attached to device;
 761          * create default one for devices, which need queueing
 762          * and noqueue_qdisc for virtual interfaces
 763          */
 764
 765         if (dev->qdisc == &noop_qdisc)
 766                 attach_default_qdiscs(dev);
 767
 768         if (!netif_carrier_ok(dev))
 769                 /* Delay activation until next carrier-on event */
 770                 return;
 771
 772         need_watchdog = 0;
 773         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 774         if (dev_ingress_queue(dev))
 775                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 776
 777         if (need_watchdog) {
 778                 netif_trans_update(dev);
 779                 dev_watchdog_up(dev);
 780         }
 781 }
 782 EXPORT_SYMBOL(dev_activate);
 783
 784 static void dev_deactivate_queue(struct net_device *dev,
 785                                  struct netdev_queue *dev_queue,
 786                                  void *_qdisc_default)
 787 {
 788         struct Qdisc *qdisc_default = _qdisc_default;
 789         struct Qdisc *qdisc;
 790
 791         qdisc = rtnl_dereference(dev_queue->qdisc);
 792         if (qdisc) {
 793                 spin_lock_bh(qdisc_lock(qdisc));
 794
 795                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 796                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 797
 798                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 799                 qdisc_reset(qdisc);
 800
 801                 spin_unlock_bh(qdisc_lock(qdisc));
 802         }
 803 }
 804
 805 static bool some_qdisc_is_busy(struct net_device *dev)
 806 {
 807         unsigned int i;
 808
 809         for (i = 0; i < dev->num_tx_queues; i++) {
 810                 struct netdev_queue *dev_queue;
 811                 spinlock_t *root_lock;
 812                 struct Qdisc *q;
 813                 int val;
 814
 815                 dev_queue = netdev_get_tx_queue(dev, i);
 816                 q = dev_queue->qdisc_sleeping;
 817                 root_lock = qdisc_lock(q);
 818
 819                 spin_lock_bh(root_lock);
 820
 821                 val = (qdisc_is_running(q) ||
 822                        test_bit(__QDISC_STATE_SCHED, &q->state));
 823
 824                 spin_unlock_bh(root_lock);
 825
 826                 if (val)
 827                         return true;
 828         }
 829         return false;
 830 }
 831
 832 /**
 833  *      dev_deactivate_many - deactivate transmissions on several devices
 834  *      @head: list of devices to deactivate
 835  *
 836  *      This function returns only when all outstanding transmissions
 837  *      have completed, unless all devices are in dismantle phase.
 838  */
 839 void dev_deactivate_many(struct list_head *head)
 840 {
 841         struct net_device *dev;
 842         bool sync_needed = false;
 843
 844         list_for_each_entry(dev, head, close_list) {
 845                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 846                                          &noop_qdisc);
 847                 if (dev_ingress_queue(dev))
 848                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 849                                              &noop_qdisc);
 850
 851                 dev_watchdog_down(dev);
 852                 sync_needed |= !dev->dismantle;
 853         }
 854
 855         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 856          * This is avoided if all devices are in dismantle phase :
 857          * Caller will call synchronize_net() for us
 858          */
 859         if (sync_needed)
 860                 synchronize_net();
 861
 862         /* Wait for outstanding qdisc_run calls. */
 863         list_for_each_entry(dev, head, close_list)
 864                 while (some_qdisc_is_busy(dev))
 865                         yield();
 866 }
 867
 868 void dev_deactivate(struct net_device *dev)
 869 {
 870         LIST_HEAD(single);
 871
 872         list_add(&dev->close_list, &single);
 873         dev_deactivate_many(&single);
 874         list_del(&single);
 875 }
 876 EXPORT_SYMBOL(dev_deactivate);
 877
 878 static void dev_init_scheduler_queue(struct net_device *dev,
 879                                      struct netdev_queue *dev_queue,
 880                                      void *_qdisc)
 881 {
 882         struct Qdisc *qdisc = _qdisc;
 883
 884         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 885         dev_queue->qdisc_sleeping = qdisc;
 886 }
 887
 888 void dev_init_scheduler(struct net_device *dev)
 889 {
 890         dev->qdisc = &noop_qdisc;
 891         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 892         if (dev_ingress_queue(dev))
 893                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 894
 895         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 896 }
 897
 898 static void shutdown_scheduler_queue(struct net_device *dev,
 899                                      struct netdev_queue *dev_queue,
 900                                      void *_qdisc_default)
 901 {
 902         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 903         struct Qdisc *qdisc_default = _qdisc_default;
 904
 905         if (qdisc) {
 906                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 907                 dev_queue->qdisc_sleeping = qdisc_default;
 908
 909                 qdisc_destroy(qdisc);
 910         }
 911 }
 912
 913 void dev_shutdown(struct net_device *dev)
 914 {
 915         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 916         if (dev_ingress_queue(dev))
 917                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 918         qdisc_destroy(dev->qdisc);
 919         dev->qdisc = &noop_qdisc;
 920
 921         WARN_ON(timer_pending(&dev->watchdog_timer));
 922 }
 923
 924 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 925                                const struct tc_ratespec *conf,
 926                                u64 rate64)
 927 {
 928         memset(r, 0, sizeof(*r));
 929         r->overhead = conf->overhead;
 930         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
 931         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
 932         r->mult = 1;
 933         /*
 934          * The deal here is to replace a divide by a reciprocal one
 935          * in fast path (a reciprocal divide is a multiply and a shift)
 936          *
 937          * Normal formula would be :
 938          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
 939          *
 940          * We compute mult/shift to use instead :
 941          *  time_in_ns = (len * mult) >> shift;
 942          *
 943          * We try to get the highest possible mult value for accuracy,
 944          * but have to make sure no overflows will ever happen.
 945          */
 946         if (r->rate_bytes_ps > 0) {
 947                 u64 factor = NSEC_PER_SEC;
 948
 949                 for (;;) {
 950                         r->mult = div64_u64(factor, r->rate_bytes_ps);
 951                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
 952                                 break;
 953                         factor <<= 1;
 954                         r->shift++;
 955                 }
 956         }
 957 }
 958 EXPORT_SYMBOL(psched_ratecfg_precompute);