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. Owning running seqcount 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  * running seqcount 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 = netdev_get_tx_queue(dev, 0)->trans_start;
 231         for (i = 1; 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
 237         return res;
 238 }
 239 EXPORT_SYMBOL(dev_trans_start);
 240
 241 static void dev_watchdog(unsigned long arg)
 242 {
 243         struct net_device *dev = (struct net_device *)arg;
 244
 245         netif_tx_lock(dev);
 246         if (!qdisc_tx_is_noop(dev)) {
 247                 if (netif_device_present(dev) &&
 248                     netif_running(dev) &&
 249                     netif_carrier_ok(dev)) {
 250                         int some_queue_timedout = 0;
 251                         unsigned int i;
 252                         unsigned long trans_start;
 253
 254                         for (i = 0; i < dev->num_tx_queues; i++) {
 255                                 struct netdev_queue *txq;
 256
 257                                 txq = netdev_get_tx_queue(dev, i);
 258                                 trans_start = txq->trans_start;
 259                                 if (netif_xmit_stopped(txq) &&
 260                                     time_after(jiffies, (trans_start +
 261                                                          dev->watchdog_timeo))) {
 262                                         some_queue_timedout = 1;
 263                                         txq->trans_timeout++;
 264                                         break;
 265                                 }
 266                         }
 267
 268                         if (some_queue_timedout) {
 269                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 270                                        dev->name, netdev_drivername(dev), i);
 271                                 dev->netdev_ops->ndo_tx_timeout(dev);
 272                         }
 273                         if (!mod_timer(&dev->watchdog_timer,
 274                                        round_jiffies(jiffies +
 275                                                      dev->watchdog_timeo)))
 276                                 dev_hold(dev);
 277                 }
 278         }
 279         netif_tx_unlock(dev);
 280
 281         dev_put(dev);
 282 }
 283
 284 void __netdev_watchdog_up(struct net_device *dev)
 285 {
 286         if (dev->netdev_ops->ndo_tx_timeout) {
 287                 if (dev->watchdog_timeo <= 0)
 288                         dev->watchdog_timeo = 5*HZ;
 289                 if (!mod_timer(&dev->watchdog_timer,
 290                                round_jiffies(jiffies + dev->watchdog_timeo)))
 291                         dev_hold(dev);
 292         }
 293 }
 294
 295 static void dev_watchdog_up(struct net_device *dev)
 296 {
 297         __netdev_watchdog_up(dev);
 298 }
 299
 300 static void dev_watchdog_down(struct net_device *dev)
 301 {
 302         netif_tx_lock_bh(dev);
 303         if (del_timer(&dev->watchdog_timer))
 304                 dev_put(dev);
 305         netif_tx_unlock_bh(dev);
 306 }
 307
 308 /**
 309  *      netif_carrier_on - set carrier
 310  *      @dev: network device
 311  *
 312  * Device has detected that carrier.
 313  */
 314 void netif_carrier_on(struct net_device *dev)
 315 {
 316         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 317                 if (dev->reg_state == NETREG_UNINITIALIZED)
 318                         return;
 319                 atomic_inc(&dev->carrier_changes);
 320                 linkwatch_fire_event(dev);
 321                 if (netif_running(dev))
 322                         __netdev_watchdog_up(dev);
 323         }
 324 }
 325 EXPORT_SYMBOL(netif_carrier_on);
 326
 327 /**
 328  *      netif_carrier_off - clear carrier
 329  *      @dev: network device
 330  *
 331  * Device has detected loss of carrier.
 332  */
 333 void netif_carrier_off(struct net_device *dev)
 334 {
 335         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 336                 if (dev->reg_state == NETREG_UNINITIALIZED)
 337                         return;
 338                 atomic_inc(&dev->carrier_changes);
 339                 linkwatch_fire_event(dev);
 340         }
 341 }
 342 EXPORT_SYMBOL(netif_carrier_off);
 343
 344 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 345    under all circumstances. It is difficult to invent anything faster or
 346    cheaper.
 347  */
 348
 349 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 350 {
 351         kfree_skb(skb);
 352         return NET_XMIT_CN;
 353 }
 354
 355 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 356 {
 357         return NULL;
 358 }
 359
 360 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 361         .id             =       "noop",
 362         .priv_size      =       0,
 363         .enqueue        =       noop_enqueue,
 364         .dequeue        =       noop_dequeue,
 365         .peek           =       noop_dequeue,
 366         .owner          =       THIS_MODULE,
 367 };
 368
 369 static struct netdev_queue noop_netdev_queue = {
 370         .qdisc          =       &noop_qdisc,
 371         .qdisc_sleeping =       &noop_qdisc,
 372 };
 373
 374 struct Qdisc noop_qdisc = {
 375         .enqueue        =       noop_enqueue,
 376         .dequeue        =       noop_dequeue,
 377         .flags          =       TCQ_F_BUILTIN,
 378         .ops            =       &noop_qdisc_ops,
 379         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 380         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 381         .dev_queue      =       &noop_netdev_queue,
 382         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 383         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 384 };
 385 EXPORT_SYMBOL(noop_qdisc);
 386
 387 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 388 {
 389         /* register_qdisc() assigns a default of noop_enqueue if unset,
 390          * but __dev_queue_xmit() treats noqueue only as such
 391          * if this is NULL - so clear it here. */
 392         qdisc->enqueue = NULL;
 393         return 0;
 394 }
 395
 396 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 397         .id             =       "noqueue",
 398         .priv_size      =       0,
 399         .init           =       noqueue_init,
 400         .enqueue        =       noop_enqueue,
 401         .dequeue        =       noop_dequeue,
 402         .peek           =       noop_dequeue,
 403         .owner          =       THIS_MODULE,
 404 };
 405
 406 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 407         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 408 };
 409
 410 /* 3-band FIFO queue: old style, but should be a bit faster than
 411    generic prio+fifo combination.
 412  */
 413
 414 #define PFIFO_FAST_BANDS 3
 415
 416 /*
 417  * Private data for a pfifo_fast scheduler containing:
 418  *      - queues for the three band
 419  *      - bitmap indicating which of the bands contain skbs
 420  */
 421 struct pfifo_fast_priv {
 422         u32 bitmap;
 423         struct sk_buff_head q[PFIFO_FAST_BANDS];
 424 };
 425
 426 /*
 427  * Convert a bitmap to the first band number where an skb is queued, where:
 428  *      bitmap=0 means there are no skbs on any band.
 429  *      bitmap=1 means there is an skb on band 0.
 430  *      bitmap=7 means there are skbs on all 3 bands, etc.
 431  */
 432 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 433
 434 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 435                                              int band)
 436 {
 437         return priv->q + band;
 438 }
 439
 440 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 441 {
 442         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 443                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 444                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 445                 struct sk_buff_head *list = band2list(priv, band);
 446
 447                 priv->bitmap |= (1 << band);
 448                 qdisc->q.qlen++;
 449                 return __qdisc_enqueue_tail(skb, qdisc, list);
 450         }
 451
 452         return qdisc_drop(skb, qdisc);
 453 }
 454
 455 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 456 {
 457         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 458         int band = bitmap2band[priv->bitmap];
 459
 460         if (likely(band >= 0)) {
 461                 struct sk_buff_head *list = band2list(priv, band);
 462                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 463
 464                 qdisc->q.qlen--;
 465                 if (skb_queue_empty(list))
 466                         priv->bitmap &= ~(1 << band);
 467
 468                 return skb;
 469         }
 470
 471         return NULL;
 472 }
 473
 474 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 475 {
 476         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 477         int band = bitmap2band[priv->bitmap];
 478
 479         if (band >= 0) {
 480                 struct sk_buff_head *list = band2list(priv, band);
 481
 482                 return skb_peek(list);
 483         }
 484
 485         return NULL;
 486 }
 487
 488 static void pfifo_fast_reset(struct Qdisc *qdisc)
 489 {
 490         int prio;
 491         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 492
 493         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 494                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 495
 496         priv->bitmap = 0;
 497         qdisc->qstats.backlog = 0;
 498         qdisc->q.qlen = 0;
 499 }
 500
 501 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 502 {
 503         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 504
 505         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 506         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 507                 goto nla_put_failure;
 508         return skb->len;
 509
 510 nla_put_failure:
 511         return -1;
 512 }
 513
 514 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 515 {
 516         int prio;
 517         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 518
 519         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 520                 __skb_queue_head_init(band2list(priv, prio));
 521
 522         /* Can by-pass the queue discipline */
 523         qdisc->flags |= TCQ_F_CAN_BYPASS;
 524         return 0;
 525 }
 526
 527 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 528         .id             =       "pfifo_fast",
 529         .priv_size      =       sizeof(struct pfifo_fast_priv),
 530         .enqueue        =       pfifo_fast_enqueue,
 531         .dequeue        =       pfifo_fast_dequeue,
 532         .peek           =       pfifo_fast_peek,
 533         .init           =       pfifo_fast_init,
 534         .reset          =       pfifo_fast_reset,
 535         .dump           =       pfifo_fast_dump,
 536         .owner          =       THIS_MODULE,
 537 };
 538 EXPORT_SYMBOL(pfifo_fast_ops);
 539
 540 static struct lock_class_key qdisc_tx_busylock;
 541 static struct lock_class_key qdisc_running_key;
 542
 543 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 544                           const struct Qdisc_ops *ops)
 545 {
 546         void *p;
 547         struct Qdisc *sch;
 548         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 549         int err = -ENOBUFS;
 550         struct net_device *dev = dev_queue->dev;
 551
 552         p = kzalloc_node(size, GFP_KERNEL,
 553                          netdev_queue_numa_node_read(dev_queue));
 554
 555         if (!p)
 556                 goto errout;
 557         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 558         /* if we got non aligned memory, ask more and do alignment ourself */
 559         if (sch != p) {
 560                 kfree(p);
 561                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 562                                  netdev_queue_numa_node_read(dev_queue));
 563                 if (!p)
 564                         goto errout;
 565                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 566                 sch->padded = (char *) sch - (char *) p;
 567         }
 568         INIT_LIST_HEAD(&sch->list);
 569         skb_queue_head_init(&sch->q);
 570
 571         spin_lock_init(&sch->busylock);
 572         lockdep_set_class(&sch->busylock,
 573                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 574
 575         seqcount_init(&sch->running);
 576         lockdep_set_class(&sch->running,
 577                           dev->qdisc_running_key ?: &qdisc_running_key);
 578
 579         sch->ops = ops;
 580         sch->enqueue = ops->enqueue;
 581         sch->dequeue = ops->dequeue;
 582         sch->dev_queue = dev_queue;
 583         dev_hold(dev);
 584         atomic_set(&sch->refcnt, 1);
 585
 586         return sch;
 587 errout:
 588         return ERR_PTR(err);
 589 }
 590
 591 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 592                                 const struct Qdisc_ops *ops,
 593                                 unsigned int parentid)
 594 {
 595         struct Qdisc *sch;
 596
 597         if (!try_module_get(ops->owner))
 598                 goto errout;
 599
 600         sch = qdisc_alloc(dev_queue, ops);
 601         if (IS_ERR(sch))
 602                 goto errout;
 603         sch->parent = parentid;
 604
 605         if (!ops->init || ops->init(sch, NULL) == 0)
 606                 return sch;
 607
 608         qdisc_destroy(sch);
 609 errout:
 610         return NULL;
 611 }
 612 EXPORT_SYMBOL(qdisc_create_dflt);
 613
 614 /* Under qdisc_lock(qdisc) and BH! */
 615
 616 void qdisc_reset(struct Qdisc *qdisc)
 617 {
 618         const struct Qdisc_ops *ops = qdisc->ops;
 619
 620         if (ops->reset)
 621                 ops->reset(qdisc);
 622
 623         if (qdisc->gso_skb) {
 624                 kfree_skb_list(qdisc->gso_skb);
 625                 qdisc->gso_skb = NULL;
 626                 qdisc->q.qlen = 0;
 627         }
 628 }
 629 EXPORT_SYMBOL(qdisc_reset);
 630
 631 static void qdisc_rcu_free(struct rcu_head *head)
 632 {
 633         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 634
 635         if (qdisc_is_percpu_stats(qdisc)) {
 636                 free_percpu(qdisc->cpu_bstats);
 637                 free_percpu(qdisc->cpu_qstats);
 638         }
 639
 640         kfree((char *) qdisc - qdisc->padded);
 641 }
 642
 643 void qdisc_destroy(struct Qdisc *qdisc)
 644 {
 645         const struct Qdisc_ops  *ops = qdisc->ops;
 646
 647         if (qdisc->flags & TCQ_F_BUILTIN ||
 648             !atomic_dec_and_test(&qdisc->refcnt))
 649                 return;
 650
 651 #ifdef CONFIG_NET_SCHED
 652         qdisc_list_del(qdisc);
 653
 654         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 655 #endif
 656         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 657         if (ops->reset)
 658                 ops->reset(qdisc);
 659         if (ops->destroy)
 660                 ops->destroy(qdisc);
 661
 662         module_put(ops->owner);
 663         dev_put(qdisc_dev(qdisc));
 664
 665         kfree_skb_list(qdisc->gso_skb);
 666         /*
 667          * gen_estimator est_timer() might access qdisc->q.lock,
 668          * wait a RCU grace period before freeing qdisc.
 669          */
 670         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 671 }
 672 EXPORT_SYMBOL(qdisc_destroy);
 673
 674 /* Attach toplevel qdisc to device queue. */
 675 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 676                               struct Qdisc *qdisc)
 677 {
 678         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 679         spinlock_t *root_lock;
 680
 681         root_lock = qdisc_lock(oqdisc);
 682         spin_lock_bh(root_lock);
 683
 684         /* Prune old scheduler */
 685         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 686                 qdisc_reset(oqdisc);
 687
 688         /* ... and graft new one */
 689         if (qdisc == NULL)
 690                 qdisc = &noop_qdisc;
 691         dev_queue->qdisc_sleeping = qdisc;
 692         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 693
 694         spin_unlock_bh(root_lock);
 695
 696         return oqdisc;
 697 }
 698 EXPORT_SYMBOL(dev_graft_qdisc);
 699
 700 static void attach_one_default_qdisc(struct net_device *dev,
 701                                      struct netdev_queue *dev_queue,
 702                                      void *_unused)
 703 {
 704         struct Qdisc *qdisc;
 705         const struct Qdisc_ops *ops = default_qdisc_ops;
 706
 707         if (dev->priv_flags & IFF_NO_QUEUE)
 708                 ops = &noqueue_qdisc_ops;
 709
 710         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 711         if (!qdisc) {
 712                 netdev_info(dev, "activation failed\n");
 713                 return;
 714         }
 715         if (!netif_is_multiqueue(dev))
 716                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 717         dev_queue->qdisc_sleeping = qdisc;
 718 }
 719
 720 static void attach_default_qdiscs(struct net_device *dev)
 721 {
 722         struct netdev_queue *txq;
 723         struct Qdisc *qdisc;
 724
 725         txq = netdev_get_tx_queue(dev, 0);
 726
 727         if (!netif_is_multiqueue(dev) ||
 728             dev->priv_flags & IFF_NO_QUEUE) {
 729                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 730                 dev->qdisc = txq->qdisc_sleeping;
 731                 atomic_inc(&dev->qdisc->refcnt);
 732         } else {
 733                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 734                 if (qdisc) {
 735                         dev->qdisc = qdisc;
 736                         qdisc->ops->attach(qdisc);
 737                 }
 738         }
 739 }
 740
 741 static void transition_one_qdisc(struct net_device *dev,
 742                                  struct netdev_queue *dev_queue,
 743                                  void *_need_watchdog)
 744 {
 745         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 746         int *need_watchdog_p = _need_watchdog;
 747
 748         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 749                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 750
 751         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 752         if (need_watchdog_p) {
 753                 dev_queue->trans_start = 0;
 754                 *need_watchdog_p = 1;
 755         }
 756 }
 757
 758 void dev_activate(struct net_device *dev)
 759 {
 760         int need_watchdog;
 761
 762         /* No queueing discipline is attached to device;
 763          * create default one for devices, which need queueing
 764          * and noqueue_qdisc for virtual interfaces
 765          */
 766
 767         if (dev->qdisc == &noop_qdisc)
 768                 attach_default_qdiscs(dev);
 769
 770         if (!netif_carrier_ok(dev))
 771                 /* Delay activation until next carrier-on event */
 772                 return;
 773
 774         need_watchdog = 0;
 775         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 776         if (dev_ingress_queue(dev))
 777                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 778
 779         if (need_watchdog) {
 780                 netif_trans_update(dev);
 781                 dev_watchdog_up(dev);
 782         }
 783 }
 784 EXPORT_SYMBOL(dev_activate);
 785
 786 static void dev_deactivate_queue(struct net_device *dev,
 787                                  struct netdev_queue *dev_queue,
 788                                  void *_qdisc_default)
 789 {
 790         struct Qdisc *qdisc_default = _qdisc_default;
 791         struct Qdisc *qdisc;
 792
 793         qdisc = rtnl_dereference(dev_queue->qdisc);
 794         if (qdisc) {
 795                 spin_lock_bh(qdisc_lock(qdisc));
 796
 797                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 798                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 799
 800                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 801                 qdisc_reset(qdisc);
 802
 803                 spin_unlock_bh(qdisc_lock(qdisc));
 804         }
 805 }
 806
 807 static bool some_qdisc_is_busy(struct net_device *dev)
 808 {
 809         unsigned int i;
 810
 811         for (i = 0; i < dev->num_tx_queues; i++) {
 812                 struct netdev_queue *dev_queue;
 813                 spinlock_t *root_lock;
 814                 struct Qdisc *q;
 815                 int val;
 816
 817                 dev_queue = netdev_get_tx_queue(dev, i);
 818                 q = dev_queue->qdisc_sleeping;
 819                 root_lock = qdisc_lock(q);
 820
 821                 spin_lock_bh(root_lock);
 822
 823                 val = (qdisc_is_running(q) ||
 824                        test_bit(__QDISC_STATE_SCHED, &q->state));
 825
 826                 spin_unlock_bh(root_lock);
 827
 828                 if (val)
 829                         return true;
 830         }
 831         return false;
 832 }
 833
 834 /**
 835  *      dev_deactivate_many - deactivate transmissions on several devices
 836  *      @head: list of devices to deactivate
 837  *
 838  *      This function returns only when all outstanding transmissions
 839  *      have completed, unless all devices are in dismantle phase.
 840  */
 841 void dev_deactivate_many(struct list_head *head)
 842 {
 843         struct net_device *dev;
 844         bool sync_needed = false;
 845
 846         list_for_each_entry(dev, head, close_list) {
 847                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 848                                          &noop_qdisc);
 849                 if (dev_ingress_queue(dev))
 850                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 851                                              &noop_qdisc);
 852
 853                 dev_watchdog_down(dev);
 854                 sync_needed |= !dev->dismantle;
 855         }
 856
 857         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 858          * This is avoided if all devices are in dismantle phase :
 859          * Caller will call synchronize_net() for us
 860          */
 861         if (sync_needed)
 862                 synchronize_net();
 863
 864         /* Wait for outstanding qdisc_run calls. */
 865         list_for_each_entry(dev, head, close_list)
 866                 while (some_qdisc_is_busy(dev))
 867                         yield();
 868 }
 869
 870 void dev_deactivate(struct net_device *dev)
 871 {
 872         LIST_HEAD(single);
 873
 874         list_add(&dev->close_list, &single);
 875         dev_deactivate_many(&single);
 876         list_del(&single);
 877 }
 878 EXPORT_SYMBOL(dev_deactivate);
 879
 880 static void dev_init_scheduler_queue(struct net_device *dev,
 881                                      struct netdev_queue *dev_queue,
 882                                      void *_qdisc)
 883 {
 884         struct Qdisc *qdisc = _qdisc;
 885
 886         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 887         dev_queue->qdisc_sleeping = qdisc;
 888 }
 889
 890 void dev_init_scheduler(struct net_device *dev)
 891 {
 892         dev->qdisc = &noop_qdisc;
 893         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 894         if (dev_ingress_queue(dev))
 895                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 896
 897         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 898 }
 899
 900 static void shutdown_scheduler_queue(struct net_device *dev,
 901                                      struct netdev_queue *dev_queue,
 902                                      void *_qdisc_default)
 903 {
 904         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 905         struct Qdisc *qdisc_default = _qdisc_default;
 906
 907         if (qdisc) {
 908                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 909                 dev_queue->qdisc_sleeping = qdisc_default;
 910
 911                 qdisc_destroy(qdisc);
 912         }
 913 }
 914
 915 void dev_shutdown(struct net_device *dev)
 916 {
 917         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 918         if (dev_ingress_queue(dev))
 919                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 920         qdisc_destroy(dev->qdisc);
 921         dev->qdisc = &noop_qdisc;
 922
 923         WARN_ON(timer_pending(&dev->watchdog_timer));
 924 }
 925
 926 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 927                                const struct tc_ratespec *conf,
 928                                u64 rate64)
 929 {
 930         memset(r, 0, sizeof(*r));
 931         r->overhead = conf->overhead;
 932         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
 933         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
 934         r->mult = 1;
 935         /*
 936          * The deal here is to replace a divide by a reciprocal one
 937          * in fast path (a reciprocal divide is a multiply and a shift)
 938          *
 939          * Normal formula would be :
 940          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
 941          *
 942          * We compute mult/shift to use instead :
 943          *  time_in_ns = (len * mult) >> shift;
 944          *
 945          * We try to get the highest possible mult value for accuracy,
 946          * but have to make sure no overflows will ever happen.
 947          */
 948         if (r->rate_bytes_ps > 0) {
 949                 u64 factor = NSEC_PER_SEC;
 950
 951                 for (;;) {
 952                         r->mult = div64_u64(factor, r->rate_bytes_ps);
 953                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
 954                                 break;
 955                         factor <<= 1;
 956                         r->shift++;
 957                 }
 958         }
 959 }
 960 EXPORT_SYMBOL(psched_ratecfg_precompute);