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         qdisc_qstats_backlog_inc(q, skb);
  53         q->q.qlen++;    /* it's still part of the queue */
  54         __netif_schedule(q);
  55
  56         return 0;
  57 }
  58
  59 static void try_bulk_dequeue_skb(struct Qdisc *q,
  60                                  struct sk_buff *skb,
  61                                  const struct netdev_queue *txq,
  62                                  int *packets)
  63 {
  64         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  65
  66         while (bytelimit > 0) {
  67                 struct sk_buff *nskb = q->dequeue(q);
  68
  69                 if (!nskb)
  70                         break;
  71
  72                 bytelimit -= nskb->len; /* covers GSO len */
  73                 skb->next = nskb;
  74                 skb = nskb;
  75                 (*packets)++; /* GSO counts as one pkt */
  76         }
  77         skb->next = NULL;
  78 }
  79
  80 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
  81  * A requeued skb (via q->gso_skb) can also be a SKB list.
  82  */
  83 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
  84                                    int *packets)
  85 {
  86         struct sk_buff *skb = q->gso_skb;
  87         const struct netdev_queue *txq = q->dev_queue;
  88
  89         *packets = 1;
  90         *validate = true;
  91         if (unlikely(skb)) {
  92                 /* check the reason of requeuing without tx lock first */
  93                 txq = skb_get_tx_queue(txq->dev, skb);
  94                 if (!netif_xmit_frozen_or_stopped(txq)) {
  95                         q->gso_skb = NULL;
  96                         qdisc_qstats_backlog_dec(q, skb);
  97                         q->q.qlen--;
  98                 } else
  99                         skb = NULL;
 100                 /* skb in gso_skb were already validated */
 101                 *validate = false;
 102         } else {
 103                 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
 104                     !netif_xmit_frozen_or_stopped(txq)) {
 105                         skb = q->dequeue(q);
 106                         if (skb && qdisc_may_bulk(q))
 107                                 try_bulk_dequeue_skb(q, skb, txq, packets);
 108                 }
 109         }
 110         return skb;
 111 }
 112
 113 /*
 114  * Transmit possibly several skbs, and handle the return status as
 115  * required. Owning running seqcount bit guarantees that
 116  * only one CPU can execute this function.
 117  *
 118  * Returns to the caller:
 119  *                              0  - queue is empty or throttled.
 120  *                              >0 - queue is not empty.
 121  */
 122 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 123                     struct net_device *dev, struct netdev_queue *txq,
 124                     spinlock_t *root_lock, bool validate)
 125 {
 126         int ret = NETDEV_TX_BUSY;
 127
 128         /* And release qdisc */
 129         spin_unlock(root_lock);
 130
 131         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 132         if (validate)
 133                 skb = validate_xmit_skb_list(skb, dev);
 134
 135         if (likely(skb)) {
 136                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 137                 if (!netif_xmit_frozen_or_stopped(txq))
 138                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 139
 140                 HARD_TX_UNLOCK(dev, txq);
 141         } else {
 142                 spin_lock(root_lock);
 143                 return qdisc_qlen(q);
 144         }
 145         spin_lock(root_lock);
 146
 147         if (dev_xmit_complete(ret)) {
 148                 /* Driver sent out skb successfully or skb was consumed */
 149                 ret = qdisc_qlen(q);
 150         } else {
 151                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 152                 if (unlikely(ret != NETDEV_TX_BUSY))
 153                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 154                                              dev->name, ret, q->q.qlen);
 155
 156                 ret = dev_requeue_skb(skb, q);
 157         }
 158
 159         if (ret && netif_xmit_frozen_or_stopped(txq))
 160                 ret = 0;
 161
 162         return ret;
 163 }
 164
 165 /*
 166  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 167  *
 168  * running seqcount guarantees only one CPU can process
 169  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 170  * this queue.
 171  *
 172  *  netif_tx_lock serializes accesses to device driver.
 173  *
 174  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 175  *  if one is grabbed, another must be free.
 176  *
 177  * Note, that this procedure can be called by a watchdog timer
 178  *
 179  * Returns to the caller:
 180  *                              0  - queue is empty or throttled.
 181  *                              >0 - queue is not empty.
 182  *
 183  */
 184 static inline int qdisc_restart(struct Qdisc *q, int *packets)
 185 {
 186         struct netdev_queue *txq;
 187         struct net_device *dev;
 188         spinlock_t *root_lock;
 189         struct sk_buff *skb;
 190         bool validate;
 191
 192         /* Dequeue packet */
 193         skb = dequeue_skb(q, &validate, packets);
 194         if (unlikely(!skb))
 195                 return 0;
 196
 197         root_lock = qdisc_lock(q);
 198         dev = qdisc_dev(q);
 199         txq = skb_get_tx_queue(dev, skb);
 200
 201         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 202 }
 203
 204 void __qdisc_run(struct Qdisc *q)
 205 {
 206         int quota = weight_p;
 207         int packets;
 208
 209         while (qdisc_restart(q, &packets)) {
 210                 /*
 211                  * Ordered by possible occurrence: Postpone processing if
 212                  * 1. we've exceeded packet quota
 213                  * 2. another process needs the CPU;
 214                  */
 215                 quota -= packets;
 216                 if (quota <= 0 || need_resched()) {
 217                         __netif_schedule(q);
 218                         break;
 219                 }
 220         }
 221
 222         qdisc_run_end(q);
 223 }
 224
 225 unsigned long dev_trans_start(struct net_device *dev)
 226 {
 227         unsigned long val, res;
 228         unsigned int i;
 229
 230         if (is_vlan_dev(dev))
 231                 dev = vlan_dev_real_dev(dev);
 232         res = netdev_get_tx_queue(dev, 0)->trans_start;
 233         for (i = 1; i < dev->num_tx_queues; i++) {
 234                 val = netdev_get_tx_queue(dev, i)->trans_start;
 235                 if (val && time_after(val, res))
 236                         res = val;
 237         }
 238
 239         return res;
 240 }
 241 EXPORT_SYMBOL(dev_trans_start);
 242
 243 static void dev_watchdog(unsigned long arg)
 244 {
 245         struct net_device *dev = (struct net_device *)arg;
 246
 247         netif_tx_lock(dev);
 248         if (!qdisc_tx_is_noop(dev)) {
 249                 if (netif_device_present(dev) &&
 250                     netif_running(dev) &&
 251                     netif_carrier_ok(dev)) {
 252                         int some_queue_timedout = 0;
 253                         unsigned int i;
 254                         unsigned long trans_start;
 255
 256                         for (i = 0; i < dev->num_tx_queues; i++) {
 257                                 struct netdev_queue *txq;
 258
 259                                 txq = netdev_get_tx_queue(dev, i);
 260                                 trans_start = txq->trans_start;
 261                                 if (netif_xmit_stopped(txq) &&
 262                                     time_after(jiffies, (trans_start +
 263                                                          dev->watchdog_timeo))) {
 264                                         some_queue_timedout = 1;
 265                                         txq->trans_timeout++;
 266                                         break;
 267                                 }
 268                         }
 269
 270                         if (some_queue_timedout) {
 271                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 272                                        dev->name, netdev_drivername(dev), i);
 273                                 dev->netdev_ops->ndo_tx_timeout(dev);
 274                         }
 275                         if (!mod_timer(&dev->watchdog_timer,
 276                                        round_jiffies(jiffies +
 277                                                      dev->watchdog_timeo)))
 278                                 dev_hold(dev);
 279                 }
 280         }
 281         netif_tx_unlock(dev);
 282
 283         dev_put(dev);
 284 }
 285
 286 void __netdev_watchdog_up(struct net_device *dev)
 287 {
 288         if (dev->netdev_ops->ndo_tx_timeout) {
 289                 if (dev->watchdog_timeo <= 0)
 290                         dev->watchdog_timeo = 5*HZ;
 291                 if (!mod_timer(&dev->watchdog_timer,
 292                                round_jiffies(jiffies + dev->watchdog_timeo)))
 293                         dev_hold(dev);
 294         }
 295 }
 296
 297 static void dev_watchdog_up(struct net_device *dev)
 298 {
 299         __netdev_watchdog_up(dev);
 300 }
 301
 302 static void dev_watchdog_down(struct net_device *dev)
 303 {
 304         netif_tx_lock_bh(dev);
 305         if (del_timer(&dev->watchdog_timer))
 306                 dev_put(dev);
 307         netif_tx_unlock_bh(dev);
 308 }
 309
 310 /**
 311  *      netif_carrier_on - set carrier
 312  *      @dev: network device
 313  *
 314  * Device has detected that carrier.
 315  */
 316 void netif_carrier_on(struct net_device *dev)
 317 {
 318         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 319                 if (dev->reg_state == NETREG_UNINITIALIZED)
 320                         return;
 321                 atomic_inc(&dev->carrier_changes);
 322                 linkwatch_fire_event(dev);
 323                 if (netif_running(dev))
 324                         __netdev_watchdog_up(dev);
 325         }
 326 }
 327 EXPORT_SYMBOL(netif_carrier_on);
 328
 329 /**
 330  *      netif_carrier_off - clear carrier
 331  *      @dev: network device
 332  *
 333  * Device has detected loss of carrier.
 334  */
 335 void netif_carrier_off(struct net_device *dev)
 336 {
 337         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 338                 if (dev->reg_state == NETREG_UNINITIALIZED)
 339                         return;
 340                 atomic_inc(&dev->carrier_changes);
 341                 linkwatch_fire_event(dev);
 342         }
 343 }
 344 EXPORT_SYMBOL(netif_carrier_off);
 345
 346 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 347    under all circumstances. It is difficult to invent anything faster or
 348    cheaper.
 349  */
 350
 351 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 352                         struct sk_buff **to_free)
 353 {
 354         __qdisc_drop(skb, to_free);
 355         return NET_XMIT_CN;
 356 }
 357
 358 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 359 {
 360         return NULL;
 361 }
 362
 363 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 364         .id             =       "noop",
 365         .priv_size      =       0,
 366         .enqueue        =       noop_enqueue,
 367         .dequeue        =       noop_dequeue,
 368         .peek           =       noop_dequeue,
 369         .owner          =       THIS_MODULE,
 370 };
 371
 372 static struct netdev_queue noop_netdev_queue = {
 373         .qdisc          =       &noop_qdisc,
 374         .qdisc_sleeping =       &noop_qdisc,
 375 };
 376
 377 struct Qdisc noop_qdisc = {
 378         .enqueue        =       noop_enqueue,
 379         .dequeue        =       noop_dequeue,
 380         .flags          =       TCQ_F_BUILTIN,
 381         .ops            =       &noop_qdisc_ops,
 382         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 383         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 384         .dev_queue      =       &noop_netdev_queue,
 385         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 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                               struct sk_buff **to_free)
 445 {
 446         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 447                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 448                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 449                 struct sk_buff_head *list = band2list(priv, band);
 450
 451                 priv->bitmap |= (1 << band);
 452                 qdisc->q.qlen++;
 453                 return __qdisc_enqueue_tail(skb, qdisc, list);
 454         }
 455
 456         return qdisc_drop(skb, qdisc, to_free);
 457 }
 458
 459 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 460 {
 461         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 462         int band = bitmap2band[priv->bitmap];
 463
 464         if (likely(band >= 0)) {
 465                 struct sk_buff_head *list = band2list(priv, band);
 466                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 467
 468                 qdisc->q.qlen--;
 469                 if (skb_queue_empty(list))
 470                         priv->bitmap &= ~(1 << band);
 471
 472                 return skb;
 473         }
 474
 475         return NULL;
 476 }
 477
 478 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 479 {
 480         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 481         int band = bitmap2band[priv->bitmap];
 482
 483         if (band >= 0) {
 484                 struct sk_buff_head *list = band2list(priv, band);
 485
 486                 return skb_peek(list);
 487         }
 488
 489         return NULL;
 490 }
 491
 492 static void pfifo_fast_reset(struct Qdisc *qdisc)
 493 {
 494         int prio;
 495         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 496
 497         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 498                 __qdisc_reset_queue(band2list(priv, prio));
 499
 500         priv->bitmap = 0;
 501         qdisc->qstats.backlog = 0;
 502         qdisc->q.qlen = 0;
 503 }
 504
 505 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 506 {
 507         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 508
 509         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 510         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 511                 goto nla_put_failure;
 512         return skb->len;
 513
 514 nla_put_failure:
 515         return -1;
 516 }
 517
 518 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 519 {
 520         int prio;
 521         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 522
 523         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 524                 __skb_queue_head_init(band2list(priv, prio));
 525
 526         /* Can by-pass the queue discipline */
 527         qdisc->flags |= TCQ_F_CAN_BYPASS;
 528         return 0;
 529 }
 530
 531 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 532         .id             =       "pfifo_fast",
 533         .priv_size      =       sizeof(struct pfifo_fast_priv),
 534         .enqueue        =       pfifo_fast_enqueue,
 535         .dequeue        =       pfifo_fast_dequeue,
 536         .peek           =       pfifo_fast_peek,
 537         .init           =       pfifo_fast_init,
 538         .reset          =       pfifo_fast_reset,
 539         .dump           =       pfifo_fast_dump,
 540         .owner          =       THIS_MODULE,
 541 };
 542 EXPORT_SYMBOL(pfifo_fast_ops);
 543
 544 static struct lock_class_key qdisc_tx_busylock;
 545 static struct lock_class_key qdisc_running_key;
 546
 547 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 548                           const struct Qdisc_ops *ops)
 549 {
 550         void *p;
 551         struct Qdisc *sch;
 552         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 553         int err = -ENOBUFS;
 554         struct net_device *dev = dev_queue->dev;
 555
 556         p = kzalloc_node(size, GFP_KERNEL,
 557                          netdev_queue_numa_node_read(dev_queue));
 558
 559         if (!p)
 560                 goto errout;
 561         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 562         /* if we got non aligned memory, ask more and do alignment ourself */
 563         if (sch != p) {
 564                 kfree(p);
 565                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 566                                  netdev_queue_numa_node_read(dev_queue));
 567                 if (!p)
 568                         goto errout;
 569                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 570                 sch->padded = (char *) sch - (char *) p;
 571         }
 572         INIT_LIST_HEAD(&sch->list);
 573         skb_queue_head_init(&sch->q);
 574
 575         spin_lock_init(&sch->busylock);
 576         lockdep_set_class(&sch->busylock,
 577                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 578
 579         seqcount_init(&sch->running);
 580         lockdep_set_class(&sch->running,
 581                           dev->qdisc_running_key ?: &qdisc_running_key);
 582
 583         sch->ops = ops;
 584         sch->enqueue = ops->enqueue;
 585         sch->dequeue = ops->dequeue;
 586         sch->dev_queue = dev_queue;
 587         dev_hold(dev);
 588         atomic_set(&sch->refcnt, 1);
 589
 590         return sch;
 591 errout:
 592         return ERR_PTR(err);
 593 }
 594
 595 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 596                                 const struct Qdisc_ops *ops,
 597                                 unsigned int parentid)
 598 {
 599         struct Qdisc *sch;
 600
 601         if (!try_module_get(ops->owner))
 602                 goto errout;
 603
 604         sch = qdisc_alloc(dev_queue, ops);
 605         if (IS_ERR(sch))
 606                 goto errout;
 607         sch->parent = parentid;
 608
 609         if (!ops->init || ops->init(sch, NULL) == 0)
 610                 return sch;
 611
 612         qdisc_destroy(sch);
 613 errout:
 614         return NULL;
 615 }
 616 EXPORT_SYMBOL(qdisc_create_dflt);
 617
 618 /* Under qdisc_lock(qdisc) and BH! */
 619
 620 void qdisc_reset(struct Qdisc *qdisc)
 621 {
 622         const struct Qdisc_ops *ops = qdisc->ops;
 623
 624         if (ops->reset)
 625                 ops->reset(qdisc);
 626
 627         if (qdisc->gso_skb) {
 628                 kfree_skb_list(qdisc->gso_skb);
 629                 qdisc->gso_skb = NULL;
 630                 qdisc->q.qlen = 0;
 631         }
 632 }
 633 EXPORT_SYMBOL(qdisc_reset);
 634
 635 static void qdisc_rcu_free(struct rcu_head *head)
 636 {
 637         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 638
 639         if (qdisc_is_percpu_stats(qdisc)) {
 640                 free_percpu(qdisc->cpu_bstats);
 641                 free_percpu(qdisc->cpu_qstats);
 642         }
 643
 644         kfree((char *) qdisc - qdisc->padded);
 645 }
 646
 647 void qdisc_destroy(struct Qdisc *qdisc)
 648 {
 649         const struct Qdisc_ops  *ops = qdisc->ops;
 650
 651         if (qdisc->flags & TCQ_F_BUILTIN ||
 652             !atomic_dec_and_test(&qdisc->refcnt))
 653                 return;
 654
 655 #ifdef CONFIG_NET_SCHED
 656         qdisc_list_del(qdisc);
 657
 658         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 659 #endif
 660         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 661         if (ops->reset)
 662                 ops->reset(qdisc);
 663         if (ops->destroy)
 664                 ops->destroy(qdisc);
 665
 666         module_put(ops->owner);
 667         dev_put(qdisc_dev(qdisc));
 668
 669         kfree_skb_list(qdisc->gso_skb);
 670         /*
 671          * gen_estimator est_timer() might access qdisc->q.lock,
 672          * wait a RCU grace period before freeing qdisc.
 673          */
 674         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 675 }
 676 EXPORT_SYMBOL(qdisc_destroy);
 677
 678 /* Attach toplevel qdisc to device queue. */
 679 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 680                               struct Qdisc *qdisc)
 681 {
 682         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 683         spinlock_t *root_lock;
 684
 685         root_lock = qdisc_lock(oqdisc);
 686         spin_lock_bh(root_lock);
 687
 688         /* Prune old scheduler */
 689         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 690                 qdisc_reset(oqdisc);
 691
 692         /* ... and graft new one */
 693         if (qdisc == NULL)
 694                 qdisc = &noop_qdisc;
 695         dev_queue->qdisc_sleeping = qdisc;
 696         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 697
 698         spin_unlock_bh(root_lock);
 699
 700         return oqdisc;
 701 }
 702 EXPORT_SYMBOL(dev_graft_qdisc);
 703
 704 static void attach_one_default_qdisc(struct net_device *dev,
 705                                      struct netdev_queue *dev_queue,
 706                                      void *_unused)
 707 {
 708         struct Qdisc *qdisc;
 709         const struct Qdisc_ops *ops = default_qdisc_ops;
 710
 711         if (dev->priv_flags & IFF_NO_QUEUE)
 712                 ops = &noqueue_qdisc_ops;
 713
 714         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 715         if (!qdisc) {
 716                 netdev_info(dev, "activation failed\n");
 717                 return;
 718         }
 719         if (!netif_is_multiqueue(dev))
 720                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 721         dev_queue->qdisc_sleeping = qdisc;
 722 }
 723
 724 static void attach_default_qdiscs(struct net_device *dev)
 725 {
 726         struct netdev_queue *txq;
 727         struct Qdisc *qdisc;
 728
 729         txq = netdev_get_tx_queue(dev, 0);
 730
 731         if (!netif_is_multiqueue(dev) ||
 732             dev->priv_flags & IFF_NO_QUEUE) {
 733                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 734                 dev->qdisc = txq->qdisc_sleeping;
 735                 atomic_inc(&dev->qdisc->refcnt);
 736         } else {
 737                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 738                 if (qdisc) {
 739                         dev->qdisc = qdisc;
 740                         qdisc->ops->attach(qdisc);
 741                 }
 742         }
 743 }
 744
 745 static void transition_one_qdisc(struct net_device *dev,
 746                                  struct netdev_queue *dev_queue,
 747                                  void *_need_watchdog)
 748 {
 749         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 750         int *need_watchdog_p = _need_watchdog;
 751
 752         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 753                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 754
 755         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 756         if (need_watchdog_p) {
 757                 dev_queue->trans_start = 0;
 758                 *need_watchdog_p = 1;
 759         }
 760 }
 761
 762 void dev_activate(struct net_device *dev)
 763 {
 764         int need_watchdog;
 765
 766         /* No queueing discipline is attached to device;
 767          * create default one for devices, which need queueing
 768          * and noqueue_qdisc for virtual interfaces
 769          */
 770
 771         if (dev->qdisc == &noop_qdisc)
 772                 attach_default_qdiscs(dev);
 773
 774         if (!netif_carrier_ok(dev))
 775                 /* Delay activation until next carrier-on event */
 776                 return;
 777
 778         need_watchdog = 0;
 779         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 780         if (dev_ingress_queue(dev))
 781                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 782
 783         if (need_watchdog) {
 784                 netif_trans_update(dev);
 785                 dev_watchdog_up(dev);
 786         }
 787 }
 788 EXPORT_SYMBOL(dev_activate);
 789
 790 static void dev_deactivate_queue(struct net_device *dev,
 791                                  struct netdev_queue *dev_queue,
 792                                  void *_qdisc_default)
 793 {
 794         struct Qdisc *qdisc_default = _qdisc_default;
 795         struct Qdisc *qdisc;
 796
 797         qdisc = rtnl_dereference(dev_queue->qdisc);
 798         if (qdisc) {
 799                 spin_lock_bh(qdisc_lock(qdisc));
 800
 801                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 802                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 803
 804                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 805                 qdisc_reset(qdisc);
 806
 807                 spin_unlock_bh(qdisc_lock(qdisc));
 808         }
 809 }
 810
 811 static bool some_qdisc_is_busy(struct net_device *dev)
 812 {
 813         unsigned int i;
 814
 815         for (i = 0; i < dev->num_tx_queues; i++) {
 816                 struct netdev_queue *dev_queue;
 817                 spinlock_t *root_lock;
 818                 struct Qdisc *q;
 819                 int val;
 820
 821                 dev_queue = netdev_get_tx_queue(dev, i);
 822                 q = dev_queue->qdisc_sleeping;
 823                 root_lock = qdisc_lock(q);
 824
 825                 spin_lock_bh(root_lock);
 826
 827                 val = (qdisc_is_running(q) ||
 828                        test_bit(__QDISC_STATE_SCHED, &q->state));
 829
 830                 spin_unlock_bh(root_lock);
 831
 832                 if (val)
 833                         return true;
 834         }
 835         return false;
 836 }
 837
 838 /**
 839  *      dev_deactivate_many - deactivate transmissions on several devices
 840  *      @head: list of devices to deactivate
 841  *
 842  *      This function returns only when all outstanding transmissions
 843  *      have completed, unless all devices are in dismantle phase.
 844  */
 845 void dev_deactivate_many(struct list_head *head)
 846 {
 847         struct net_device *dev;
 848         bool sync_needed = false;
 849
 850         list_for_each_entry(dev, head, close_list) {
 851                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 852                                          &noop_qdisc);
 853                 if (dev_ingress_queue(dev))
 854                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 855                                              &noop_qdisc);
 856
 857                 dev_watchdog_down(dev);
 858                 sync_needed |= !dev->dismantle;
 859         }
 860
 861         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 862          * This is avoided if all devices are in dismantle phase :
 863          * Caller will call synchronize_net() for us
 864          */
 865         if (sync_needed)
 866                 synchronize_net();
 867
 868         /* Wait for outstanding qdisc_run calls. */
 869         list_for_each_entry(dev, head, close_list)
 870                 while (some_qdisc_is_busy(dev))
 871                         yield();
 872 }
 873
 874 void dev_deactivate(struct net_device *dev)
 875 {
 876         LIST_HEAD(single);
 877
 878         list_add(&dev->close_list, &single);
 879         dev_deactivate_many(&single);
 880         list_del(&single);
 881 }
 882 EXPORT_SYMBOL(dev_deactivate);
 883
 884 static void dev_init_scheduler_queue(struct net_device *dev,
 885                                      struct netdev_queue *dev_queue,
 886                                      void *_qdisc)
 887 {
 888         struct Qdisc *qdisc = _qdisc;
 889
 890         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 891         dev_queue->qdisc_sleeping = qdisc;
 892 }
 893
 894 void dev_init_scheduler(struct net_device *dev)
 895 {
 896         dev->qdisc = &noop_qdisc;
 897         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 898         if (dev_ingress_queue(dev))
 899                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 900
 901         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 902 }
 903
 904 static void shutdown_scheduler_queue(struct net_device *dev,
 905                                      struct netdev_queue *dev_queue,
 906                                      void *_qdisc_default)
 907 {
 908         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 909         struct Qdisc *qdisc_default = _qdisc_default;
 910
 911         if (qdisc) {
 912                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 913                 dev_queue->qdisc_sleeping = qdisc_default;
 914
 915                 qdisc_destroy(qdisc);
 916         }
 917 }
 918
 919 void dev_shutdown(struct net_device *dev)
 920 {
 921         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 922         if (dev_ingress_queue(dev))
 923                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 924         qdisc_destroy(dev->qdisc);
 925         dev->qdisc = &noop_qdisc;
 926
 927         WARN_ON(timer_pending(&dev->watchdog_timer));
 928 }
 929
 930 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 931                                const struct tc_ratespec *conf,
 932                                u64 rate64)
 933 {
 934         memset(r, 0, sizeof(*r));
 935         r->overhead = conf->overhead;
 936         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
 937         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
 938         r->mult = 1;
 939         /*
 940          * The deal here is to replace a divide by a reciprocal one
 941          * in fast path (a reciprocal divide is a multiply and a shift)
 942          *
 943          * Normal formula would be :
 944          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
 945          *
 946          * We compute mult/shift to use instead :
 947          *  time_in_ns = (len * mult) >> shift;
 948          *
 949          * We try to get the highest possible mult value for accuracy,
 950          * but have to make sure no overflows will ever happen.
 951          */
 952         if (r->rate_bytes_ps > 0) {
 953                 u64 factor = NSEC_PER_SEC;
 954
 955                 for (;;) {
 956                         r->mult = div64_u64(factor, r->rate_bytes_ps);
 957                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
 958                                 break;
 959                         factor <<= 1;
 960                         r->shift++;
 961                 }
 962         }
 963 }
 964 EXPORT_SYMBOL(psched_ratecfg_precompute);