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 <linux/if_macvlan.h>
  30 #include <net/sch_generic.h>
  31 #include <net/pkt_sched.h>
  32 #include <net/dst.h>
  33 #include <trace/events/qdisc.h>
  34
  35 /* Qdisc to use by default */
  36 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  37 EXPORT_SYMBOL(default_qdisc_ops);
  38
  39 /* Main transmission queue. */
  40
  41 /* Modifications to data participating in scheduling must be protected with
  42  * qdisc_lock(qdisc) spinlock.
  43  *
  44  * The idea is the following:
  45  * - enqueue, dequeue are serialized via qdisc root lock
  46  * - ingress filtering is also serialized via qdisc root lock
  47  * - updates to tree and tree walking are only done under the rtnl mutex.
  48  */
  49
  50 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  51 {
  52         q->gso_skb = skb;
  53         q->qstats.requeues++;
  54         qdisc_qstats_backlog_inc(q, skb);
  55         q->q.qlen++;    /* it's still part of the queue */
  56         __netif_schedule(q);
  57
  58         return 0;
  59 }
  60
  61 static void try_bulk_dequeue_skb(struct Qdisc *q,
  62                                  struct sk_buff *skb,
  63                                  const struct netdev_queue *txq,
  64                                  int *packets)
  65 {
  66         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  67
  68         while (bytelimit > 0) {
  69                 struct sk_buff *nskb = q->dequeue(q);
  70
  71                 if (!nskb)
  72                         break;
  73
  74                 bytelimit -= nskb->len; /* covers GSO len */
  75                 skb->next = nskb;
  76                 skb = nskb;
  77                 (*packets)++; /* GSO counts as one pkt */
  78         }
  79         skb->next = NULL;
  80 }
  81
  82 /* This variant of try_bulk_dequeue_skb() makes sure
  83  * all skbs in the chain are for the same txq
  84  */
  85 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
  86                                       struct sk_buff *skb,
  87                                       int *packets)
  88 {
  89         int mapping = skb_get_queue_mapping(skb);
  90         struct sk_buff *nskb;
  91         int cnt = 0;
  92
  93         do {
  94                 nskb = q->dequeue(q);
  95                 if (!nskb)
  96                         break;
  97                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
  98                         q->skb_bad_txq = nskb;
  99                         qdisc_qstats_backlog_inc(q, nskb);
 100                         q->q.qlen++;
 101                         break;
 102                 }
 103                 skb->next = nskb;
 104                 skb = nskb;
 105         } while (++cnt < 8);
 106         (*packets) += cnt;
 107         skb->next = NULL;
 108 }
 109
 110 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 111  * A requeued skb (via q->gso_skb) can also be a SKB list.
 112  */
 113 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 114                                    int *packets)
 115 {
 116         struct sk_buff *skb = q->gso_skb;
 117         const struct netdev_queue *txq = q->dev_queue;
 118
 119         *packets = 1;
 120         if (unlikely(skb)) {
 121                 /* skb in gso_skb were already validated */
 122                 *validate = false;
 123                 /* check the reason of requeuing without tx lock first */
 124                 txq = skb_get_tx_queue(txq->dev, skb);
 125                 if (!netif_xmit_frozen_or_stopped(txq)) {
 126                         q->gso_skb = NULL;
 127                         qdisc_qstats_backlog_dec(q, skb);
 128                         q->q.qlen--;
 129                 } else
 130                         skb = NULL;
 131                 goto trace;
 132         }
 133         *validate = true;
 134         skb = q->skb_bad_txq;
 135         if (unlikely(skb)) {
 136                 /* check the reason of requeuing without tx lock first */
 137                 txq = skb_get_tx_queue(txq->dev, skb);
 138                 if (!netif_xmit_frozen_or_stopped(txq)) {
 139                         q->skb_bad_txq = NULL;
 140                         qdisc_qstats_backlog_dec(q, skb);
 141                         q->q.qlen--;
 142                         goto bulk;
 143                 }
 144                 skb = NULL;
 145                 goto trace;
 146         }
 147         if (!(q->flags & TCQ_F_ONETXQUEUE) ||
 148             !netif_xmit_frozen_or_stopped(txq))
 149                 skb = q->dequeue(q);
 150         if (skb) {
 151 bulk:
 152                 if (qdisc_may_bulk(q))
 153                         try_bulk_dequeue_skb(q, skb, txq, packets);
 154                 else
 155                         try_bulk_dequeue_skb_slow(q, skb, packets);
 156         }
 157 trace:
 158         trace_qdisc_dequeue(q, txq, *packets, skb);
 159         return skb;
 160 }
 161
 162 /*
 163  * Transmit possibly several skbs, and handle the return status as
 164  * required. Owning running seqcount bit guarantees that
 165  * only one CPU can execute this function.
 166  *
 167  * Returns to the caller:
 168  *                              0  - queue is empty or throttled.
 169  *                              >0 - queue is not empty.
 170  */
 171 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 172                     struct net_device *dev, struct netdev_queue *txq,
 173                     spinlock_t *root_lock, bool validate)
 174 {
 175         int ret = NETDEV_TX_BUSY;
 176
 177         /* And release qdisc */
 178         spin_unlock(root_lock);
 179
 180         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 181         if (validate)
 182                 skb = validate_xmit_skb_list(skb, dev);
 183
 184         if (likely(skb)) {
 185                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 186                 if (!netif_xmit_frozen_or_stopped(txq))
 187                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 188
 189                 HARD_TX_UNLOCK(dev, txq);
 190         } else {
 191                 spin_lock(root_lock);
 192                 return qdisc_qlen(q);
 193         }
 194         spin_lock(root_lock);
 195
 196         if (dev_xmit_complete(ret)) {
 197                 /* Driver sent out skb successfully or skb was consumed */
 198                 ret = qdisc_qlen(q);
 199         } else {
 200                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 201                 if (unlikely(ret != NETDEV_TX_BUSY))
 202                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 203                                              dev->name, ret, q->q.qlen);
 204
 205                 ret = dev_requeue_skb(skb, q);
 206         }
 207
 208         if (ret && netif_xmit_frozen_or_stopped(txq))
 209                 ret = 0;
 210
 211         return ret;
 212 }
 213
 214 /*
 215  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 216  *
 217  * running seqcount guarantees only one CPU can process
 218  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 219  * this queue.
 220  *
 221  *  netif_tx_lock serializes accesses to device driver.
 222  *
 223  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 224  *  if one is grabbed, another must be free.
 225  *
 226  * Note, that this procedure can be called by a watchdog timer
 227  *
 228  * Returns to the caller:
 229  *                              0  - queue is empty or throttled.
 230  *                              >0 - queue is not empty.
 231  *
 232  */
 233 static inline int qdisc_restart(struct Qdisc *q, int *packets)
 234 {
 235         struct netdev_queue *txq;
 236         struct net_device *dev;
 237         spinlock_t *root_lock;
 238         struct sk_buff *skb;
 239         bool validate;
 240
 241         /* Dequeue packet */
 242         skb = dequeue_skb(q, &validate, packets);
 243         if (unlikely(!skb))
 244                 return 0;
 245
 246         root_lock = qdisc_lock(q);
 247         dev = qdisc_dev(q);
 248         txq = skb_get_tx_queue(dev, skb);
 249
 250         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 251 }
 252
 253 void __qdisc_run(struct Qdisc *q)
 254 {
 255         int quota = dev_tx_weight;
 256         int packets;
 257
 258         while (qdisc_restart(q, &packets)) {
 259                 /*
 260                  * Ordered by possible occurrence: Postpone processing if
 261                  * 1. we've exceeded packet quota
 262                  * 2. another process needs the CPU;
 263                  */
 264                 quota -= packets;
 265                 if (quota <= 0 || need_resched()) {
 266                         __netif_schedule(q);
 267                         break;
 268                 }
 269         }
 270
 271         qdisc_run_end(q);
 272 }
 273
 274 unsigned long dev_trans_start(struct net_device *dev)
 275 {
 276         unsigned long val, res;
 277         unsigned int i;
 278
 279         if (is_vlan_dev(dev))
 280                 dev = vlan_dev_real_dev(dev);
 281         else if (netif_is_macvlan(dev))
 282                 dev = macvlan_dev_real_dev(dev);
 283         res = netdev_get_tx_queue(dev, 0)->trans_start;
 284         for (i = 1; i < dev->num_tx_queues; i++) {
 285                 val = netdev_get_tx_queue(dev, i)->trans_start;
 286                 if (val && time_after(val, res))
 287                         res = val;
 288         }
 289
 290         return res;
 291 }
 292 EXPORT_SYMBOL(dev_trans_start);
 293
 294 static void dev_watchdog(struct timer_list *t)
 295 {
 296         struct net_device *dev = from_timer(dev, t, watchdog_timer);
 297
 298         netif_tx_lock(dev);
 299         if (!qdisc_tx_is_noop(dev)) {
 300                 if (netif_device_present(dev) &&
 301                     netif_running(dev) &&
 302                     netif_carrier_ok(dev)) {
 303                         int some_queue_timedout = 0;
 304                         unsigned int i;
 305                         unsigned long trans_start;
 306
 307                         for (i = 0; i < dev->num_tx_queues; i++) {
 308                                 struct netdev_queue *txq;
 309
 310                                 txq = netdev_get_tx_queue(dev, i);
 311                                 trans_start = txq->trans_start;
 312                                 if (netif_xmit_stopped(txq) &&
 313                                     time_after(jiffies, (trans_start +
 314                                                          dev->watchdog_timeo))) {
 315                                         some_queue_timedout = 1;
 316                                         txq->trans_timeout++;
 317                                         break;
 318                                 }
 319                         }
 320
 321                         if (some_queue_timedout) {
 322                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 323                                        dev->name, netdev_drivername(dev), i);
 324                                 dev->netdev_ops->ndo_tx_timeout(dev);
 325                         }
 326                         if (!mod_timer(&dev->watchdog_timer,
 327                                        round_jiffies(jiffies +
 328                                                      dev->watchdog_timeo)))
 329                                 dev_hold(dev);
 330                 }
 331         }
 332         netif_tx_unlock(dev);
 333
 334         dev_put(dev);
 335 }
 336
 337 void __netdev_watchdog_up(struct net_device *dev)
 338 {
 339         if (dev->netdev_ops->ndo_tx_timeout) {
 340                 if (dev->watchdog_timeo <= 0)
 341                         dev->watchdog_timeo = 5*HZ;
 342                 if (!mod_timer(&dev->watchdog_timer,
 343                                round_jiffies(jiffies + dev->watchdog_timeo)))
 344                         dev_hold(dev);
 345         }
 346 }
 347
 348 static void dev_watchdog_up(struct net_device *dev)
 349 {
 350         __netdev_watchdog_up(dev);
 351 }
 352
 353 static void dev_watchdog_down(struct net_device *dev)
 354 {
 355         netif_tx_lock_bh(dev);
 356         if (del_timer(&dev->watchdog_timer))
 357                 dev_put(dev);
 358         netif_tx_unlock_bh(dev);
 359 }
 360
 361 /**
 362  *      netif_carrier_on - set carrier
 363  *      @dev: network device
 364  *
 365  * Device has detected that carrier.
 366  */
 367 void netif_carrier_on(struct net_device *dev)
 368 {
 369         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 370                 if (dev->reg_state == NETREG_UNINITIALIZED)
 371                         return;
 372                 atomic_inc(&dev->carrier_changes);
 373                 linkwatch_fire_event(dev);
 374                 if (netif_running(dev))
 375                         __netdev_watchdog_up(dev);
 376         }
 377 }
 378 EXPORT_SYMBOL(netif_carrier_on);
 379
 380 /**
 381  *      netif_carrier_off - clear carrier
 382  *      @dev: network device
 383  *
 384  * Device has detected loss of carrier.
 385  */
 386 void netif_carrier_off(struct net_device *dev)
 387 {
 388         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 389                 if (dev->reg_state == NETREG_UNINITIALIZED)
 390                         return;
 391                 atomic_inc(&dev->carrier_changes);
 392                 linkwatch_fire_event(dev);
 393         }
 394 }
 395 EXPORT_SYMBOL(netif_carrier_off);
 396
 397 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 398    under all circumstances. It is difficult to invent anything faster or
 399    cheaper.
 400  */
 401
 402 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 403                         struct sk_buff **to_free)
 404 {
 405         __qdisc_drop(skb, to_free);
 406         return NET_XMIT_CN;
 407 }
 408
 409 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 410 {
 411         return NULL;
 412 }
 413
 414 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 415         .id             =       "noop",
 416         .priv_size      =       0,
 417         .enqueue        =       noop_enqueue,
 418         .dequeue        =       noop_dequeue,
 419         .peek           =       noop_dequeue,
 420         .owner          =       THIS_MODULE,
 421 };
 422
 423 static struct netdev_queue noop_netdev_queue = {
 424         .qdisc          =       &noop_qdisc,
 425         .qdisc_sleeping =       &noop_qdisc,
 426 };
 427
 428 struct Qdisc noop_qdisc = {
 429         .enqueue        =       noop_enqueue,
 430         .dequeue        =       noop_dequeue,
 431         .flags          =       TCQ_F_BUILTIN,
 432         .ops            =       &noop_qdisc_ops,
 433         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 434         .dev_queue      =       &noop_netdev_queue,
 435         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 436         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 437 };
 438 EXPORT_SYMBOL(noop_qdisc);
 439
 440 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 441 {
 442         /* register_qdisc() assigns a default of noop_enqueue if unset,
 443          * but __dev_queue_xmit() treats noqueue only as such
 444          * if this is NULL - so clear it here. */
 445         qdisc->enqueue = NULL;
 446         return 0;
 447 }
 448
 449 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 450         .id             =       "noqueue",
 451         .priv_size      =       0,
 452         .init           =       noqueue_init,
 453         .enqueue        =       noop_enqueue,
 454         .dequeue        =       noop_dequeue,
 455         .peek           =       noop_dequeue,
 456         .owner          =       THIS_MODULE,
 457 };
 458
 459 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 460         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 461 };
 462
 463 /* 3-band FIFO queue: old style, but should be a bit faster than
 464    generic prio+fifo combination.
 465  */
 466
 467 #define PFIFO_FAST_BANDS 3
 468
 469 /*
 470  * Private data for a pfifo_fast scheduler containing:
 471  *      - queues for the three band
 472  *      - bitmap indicating which of the bands contain skbs
 473  */
 474 struct pfifo_fast_priv {
 475         u32 bitmap;
 476         struct qdisc_skb_head q[PFIFO_FAST_BANDS];
 477 };
 478
 479 /*
 480  * Convert a bitmap to the first band number where an skb is queued, where:
 481  *      bitmap=0 means there are no skbs on any band.
 482  *      bitmap=1 means there is an skb on band 0.
 483  *      bitmap=7 means there are skbs on all 3 bands, etc.
 484  */
 485 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 486
 487 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
 488                                              int band)
 489 {
 490         return priv->q + band;
 491 }
 492
 493 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 494                               struct sk_buff **to_free)
 495 {
 496         if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
 497                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 498                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 499                 struct qdisc_skb_head *list = band2list(priv, band);
 500
 501                 priv->bitmap |= (1 << band);
 502                 qdisc->q.qlen++;
 503                 return __qdisc_enqueue_tail(skb, qdisc, list);
 504         }
 505
 506         return qdisc_drop(skb, qdisc, to_free);
 507 }
 508
 509 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 510 {
 511         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 512         int band = bitmap2band[priv->bitmap];
 513
 514         if (likely(band >= 0)) {
 515                 struct qdisc_skb_head *qh = band2list(priv, band);
 516                 struct sk_buff *skb = __qdisc_dequeue_head(qh);
 517
 518                 if (likely(skb != NULL)) {
 519                         qdisc_qstats_backlog_dec(qdisc, skb);
 520                         qdisc_bstats_update(qdisc, skb);
 521                 }
 522
 523                 qdisc->q.qlen--;
 524                 if (qh->qlen == 0)
 525                         priv->bitmap &= ~(1 << band);
 526
 527                 return skb;
 528         }
 529
 530         return NULL;
 531 }
 532
 533 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 534 {
 535         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 536         int band = bitmap2band[priv->bitmap];
 537
 538         if (band >= 0) {
 539                 struct qdisc_skb_head *qh = band2list(priv, band);
 540
 541                 return qh->head;
 542         }
 543
 544         return NULL;
 545 }
 546
 547 static void pfifo_fast_reset(struct Qdisc *qdisc)
 548 {
 549         int prio;
 550         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 551
 552         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 553                 __qdisc_reset_queue(band2list(priv, prio));
 554
 555         priv->bitmap = 0;
 556         qdisc->qstats.backlog = 0;
 557         qdisc->q.qlen = 0;
 558 }
 559
 560 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 561 {
 562         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 563
 564         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 565         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 566                 goto nla_put_failure;
 567         return skb->len;
 568
 569 nla_put_failure:
 570         return -1;
 571 }
 572
 573 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 574 {
 575         int prio;
 576         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 577
 578         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 579                 qdisc_skb_head_init(band2list(priv, prio));
 580
 581         /* Can by-pass the queue discipline */
 582         qdisc->flags |= TCQ_F_CAN_BYPASS;
 583         return 0;
 584 }
 585
 586 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 587         .id             =       "pfifo_fast",
 588         .priv_size      =       sizeof(struct pfifo_fast_priv),
 589         .enqueue        =       pfifo_fast_enqueue,
 590         .dequeue        =       pfifo_fast_dequeue,
 591         .peek           =       pfifo_fast_peek,
 592         .init           =       pfifo_fast_init,
 593         .reset          =       pfifo_fast_reset,
 594         .dump           =       pfifo_fast_dump,
 595         .owner          =       THIS_MODULE,
 596 };
 597 EXPORT_SYMBOL(pfifo_fast_ops);
 598
 599 static struct lock_class_key qdisc_tx_busylock;
 600 static struct lock_class_key qdisc_running_key;
 601
 602 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 603                           const struct Qdisc_ops *ops)
 604 {
 605         void *p;
 606         struct Qdisc *sch;
 607         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 608         int err = -ENOBUFS;
 609         struct net_device *dev;
 610
 611         if (!dev_queue) {
 612                 err = -EINVAL;
 613                 goto errout;
 614         }
 615
 616         dev = dev_queue->dev;
 617         p = kzalloc_node(size, GFP_KERNEL,
 618                          netdev_queue_numa_node_read(dev_queue));
 619
 620         if (!p)
 621                 goto errout;
 622         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 623         /* if we got non aligned memory, ask more and do alignment ourself */
 624         if (sch != p) {
 625                 kfree(p);
 626                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 627                                  netdev_queue_numa_node_read(dev_queue));
 628                 if (!p)
 629                         goto errout;
 630                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 631                 sch->padded = (char *) sch - (char *) p;
 632         }
 633         qdisc_skb_head_init(&sch->q);
 634         spin_lock_init(&sch->q.lock);
 635
 636         spin_lock_init(&sch->busylock);
 637         lockdep_set_class(&sch->busylock,
 638                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 639
 640         seqcount_init(&sch->running);
 641         lockdep_set_class(&sch->running,
 642                           dev->qdisc_running_key ?: &qdisc_running_key);
 643
 644         sch->ops = ops;
 645         sch->enqueue = ops->enqueue;
 646         sch->dequeue = ops->dequeue;
 647         sch->dev_queue = dev_queue;
 648         dev_hold(dev);
 649         refcount_set(&sch->refcnt, 1);
 650
 651         return sch;
 652 errout:
 653         return ERR_PTR(err);
 654 }
 655
 656 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 657                                 const struct Qdisc_ops *ops,
 658                                 unsigned int parentid)
 659 {
 660         struct Qdisc *sch;
 661
 662         if (!try_module_get(ops->owner))
 663                 return NULL;
 664
 665         sch = qdisc_alloc(dev_queue, ops);
 666         if (IS_ERR(sch)) {
 667                 module_put(ops->owner);
 668                 return NULL;
 669         }
 670         sch->parent = parentid;
 671
 672         if (!ops->init || ops->init(sch, NULL) == 0)
 673                 return sch;
 674
 675         qdisc_destroy(sch);
 676         return NULL;
 677 }
 678 EXPORT_SYMBOL(qdisc_create_dflt);
 679
 680 /* Under qdisc_lock(qdisc) and BH! */
 681
 682 void qdisc_reset(struct Qdisc *qdisc)
 683 {
 684         const struct Qdisc_ops *ops = qdisc->ops;
 685
 686         if (ops->reset)
 687                 ops->reset(qdisc);
 688
 689         kfree_skb(qdisc->skb_bad_txq);
 690         qdisc->skb_bad_txq = NULL;
 691
 692         if (qdisc->gso_skb) {
 693                 kfree_skb_list(qdisc->gso_skb);
 694                 qdisc->gso_skb = NULL;
 695         }
 696         qdisc->q.qlen = 0;
 697         qdisc->qstats.backlog = 0;
 698 }
 699 EXPORT_SYMBOL(qdisc_reset);
 700
 701 static void qdisc_free(struct Qdisc *qdisc)
 702 {
 703         if (qdisc_is_percpu_stats(qdisc)) {
 704                 free_percpu(qdisc->cpu_bstats);
 705                 free_percpu(qdisc->cpu_qstats);
 706         }
 707
 708         kfree((char *) qdisc - qdisc->padded);
 709 }
 710
 711 void qdisc_destroy(struct Qdisc *qdisc)
 712 {
 713         const struct Qdisc_ops  *ops = qdisc->ops;
 714
 715         if (qdisc->flags & TCQ_F_BUILTIN ||
 716             !refcount_dec_and_test(&qdisc->refcnt))
 717                 return;
 718
 719 #ifdef CONFIG_NET_SCHED
 720         qdisc_hash_del(qdisc);
 721
 722         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 723 #endif
 724         gen_kill_estimator(&qdisc->rate_est);
 725         if (ops->reset)
 726                 ops->reset(qdisc);
 727         if (ops->destroy)
 728                 ops->destroy(qdisc);
 729
 730         module_put(ops->owner);
 731         dev_put(qdisc_dev(qdisc));
 732
 733         kfree_skb_list(qdisc->gso_skb);
 734         kfree_skb(qdisc->skb_bad_txq);
 735         qdisc_free(qdisc);
 736 }
 737 EXPORT_SYMBOL(qdisc_destroy);
 738
 739 /* Attach toplevel qdisc to device queue. */
 740 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 741                               struct Qdisc *qdisc)
 742 {
 743         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 744         spinlock_t *root_lock;
 745
 746         root_lock = qdisc_lock(oqdisc);
 747         spin_lock_bh(root_lock);
 748
 749         /* Prune old scheduler */
 750         if (oqdisc && refcount_read(&oqdisc->refcnt) <= 1)
 751                 qdisc_reset(oqdisc);
 752
 753         /* ... and graft new one */
 754         if (qdisc == NULL)
 755                 qdisc = &noop_qdisc;
 756         dev_queue->qdisc_sleeping = qdisc;
 757         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 758
 759         spin_unlock_bh(root_lock);
 760
 761         return oqdisc;
 762 }
 763 EXPORT_SYMBOL(dev_graft_qdisc);
 764
 765 static void attach_one_default_qdisc(struct net_device *dev,
 766                                      struct netdev_queue *dev_queue,
 767                                      void *_unused)
 768 {
 769         struct Qdisc *qdisc;
 770         const struct Qdisc_ops *ops = default_qdisc_ops;
 771
 772         if (dev->priv_flags & IFF_NO_QUEUE)
 773                 ops = &noqueue_qdisc_ops;
 774
 775         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 776         if (!qdisc) {
 777                 netdev_info(dev, "activation failed\n");
 778                 return;
 779         }
 780         if (!netif_is_multiqueue(dev))
 781                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 782         dev_queue->qdisc_sleeping = qdisc;
 783 }
 784
 785 static void attach_default_qdiscs(struct net_device *dev)
 786 {
 787         struct netdev_queue *txq;
 788         struct Qdisc *qdisc;
 789
 790         txq = netdev_get_tx_queue(dev, 0);
 791
 792         if (!netif_is_multiqueue(dev) ||
 793             dev->priv_flags & IFF_NO_QUEUE) {
 794                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 795                 dev->qdisc = txq->qdisc_sleeping;
 796                 qdisc_refcount_inc(dev->qdisc);
 797         } else {
 798                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 799                 if (qdisc) {
 800                         dev->qdisc = qdisc;
 801                         qdisc->ops->attach(qdisc);
 802                 }
 803         }
 804 #ifdef CONFIG_NET_SCHED
 805         if (dev->qdisc != &noop_qdisc)
 806                 qdisc_hash_add(dev->qdisc, false);
 807 #endif
 808 }
 809
 810 static void transition_one_qdisc(struct net_device *dev,
 811                                  struct netdev_queue *dev_queue,
 812                                  void *_need_watchdog)
 813 {
 814         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 815         int *need_watchdog_p = _need_watchdog;
 816
 817         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 818                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 819
 820         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 821         if (need_watchdog_p) {
 822                 dev_queue->trans_start = 0;
 823                 *need_watchdog_p = 1;
 824         }
 825 }
 826
 827 void dev_activate(struct net_device *dev)
 828 {
 829         int need_watchdog;
 830
 831         /* No queueing discipline is attached to device;
 832          * create default one for devices, which need queueing
 833          * and noqueue_qdisc for virtual interfaces
 834          */
 835
 836         if (dev->qdisc == &noop_qdisc)
 837                 attach_default_qdiscs(dev);
 838
 839         if (!netif_carrier_ok(dev))
 840                 /* Delay activation until next carrier-on event */
 841                 return;
 842
 843         need_watchdog = 0;
 844         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 845         if (dev_ingress_queue(dev))
 846                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 847
 848         if (need_watchdog) {
 849                 netif_trans_update(dev);
 850                 dev_watchdog_up(dev);
 851         }
 852 }
 853 EXPORT_SYMBOL(dev_activate);
 854
 855 static void dev_deactivate_queue(struct net_device *dev,
 856                                  struct netdev_queue *dev_queue,
 857                                  void *_qdisc_default)
 858 {
 859         struct Qdisc *qdisc_default = _qdisc_default;
 860         struct Qdisc *qdisc;
 861
 862         qdisc = rtnl_dereference(dev_queue->qdisc);
 863         if (qdisc) {
 864                 spin_lock_bh(qdisc_lock(qdisc));
 865
 866                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 867                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 868
 869                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 870                 qdisc_reset(qdisc);
 871
 872                 spin_unlock_bh(qdisc_lock(qdisc));
 873         }
 874 }
 875
 876 static bool some_qdisc_is_busy(struct net_device *dev)
 877 {
 878         unsigned int i;
 879
 880         for (i = 0; i < dev->num_tx_queues; i++) {
 881                 struct netdev_queue *dev_queue;
 882                 spinlock_t *root_lock;
 883                 struct Qdisc *q;
 884                 int val;
 885
 886                 dev_queue = netdev_get_tx_queue(dev, i);
 887                 q = dev_queue->qdisc_sleeping;
 888                 root_lock = qdisc_lock(q);
 889
 890                 spin_lock_bh(root_lock);
 891
 892                 val = (qdisc_is_running(q) ||
 893                        test_bit(__QDISC_STATE_SCHED, &q->state));
 894
 895                 spin_unlock_bh(root_lock);
 896
 897                 if (val)
 898                         return true;
 899         }
 900         return false;
 901 }
 902
 903 /**
 904  *      dev_deactivate_many - deactivate transmissions on several devices
 905  *      @head: list of devices to deactivate
 906  *
 907  *      This function returns only when all outstanding transmissions
 908  *      have completed, unless all devices are in dismantle phase.
 909  */
 910 void dev_deactivate_many(struct list_head *head)
 911 {
 912         struct net_device *dev;
 913         bool sync_needed = false;
 914
 915         list_for_each_entry(dev, head, close_list) {
 916                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 917                                          &noop_qdisc);
 918                 if (dev_ingress_queue(dev))
 919                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 920                                              &noop_qdisc);
 921
 922                 dev_watchdog_down(dev);
 923                 sync_needed |= !dev->dismantle;
 924         }
 925
 926         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 927          * This is avoided if all devices are in dismantle phase :
 928          * Caller will call synchronize_net() for us
 929          */
 930         if (sync_needed)
 931                 synchronize_net();
 932
 933         /* Wait for outstanding qdisc_run calls. */
 934         list_for_each_entry(dev, head, close_list)
 935                 while (some_qdisc_is_busy(dev))
 936                         yield();
 937 }
 938
 939 void dev_deactivate(struct net_device *dev)
 940 {
 941         LIST_HEAD(single);
 942
 943         list_add(&dev->close_list, &single);
 944         dev_deactivate_many(&single);
 945         list_del(&single);
 946 }
 947 EXPORT_SYMBOL(dev_deactivate);
 948
 949 static void dev_init_scheduler_queue(struct net_device *dev,
 950                                      struct netdev_queue *dev_queue,
 951                                      void *_qdisc)
 952 {
 953         struct Qdisc *qdisc = _qdisc;
 954
 955         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 956         dev_queue->qdisc_sleeping = qdisc;
 957 }
 958
 959 void dev_init_scheduler(struct net_device *dev)
 960 {
 961         dev->qdisc = &noop_qdisc;
 962         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 963         if (dev_ingress_queue(dev))
 964                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 965
 966         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
 967 }
 968
 969 static void shutdown_scheduler_queue(struct net_device *dev,
 970                                      struct netdev_queue *dev_queue,
 971                                      void *_qdisc_default)
 972 {
 973         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 974         struct Qdisc *qdisc_default = _qdisc_default;
 975
 976         if (qdisc) {
 977                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 978                 dev_queue->qdisc_sleeping = qdisc_default;
 979
 980                 qdisc_destroy(qdisc);
 981         }
 982 }
 983
 984 void dev_shutdown(struct net_device *dev)
 985 {
 986         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 987         if (dev_ingress_queue(dev))
 988                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 989         qdisc_destroy(dev->qdisc);
 990         dev->qdisc = &noop_qdisc;
 991
 992         WARN_ON(timer_pending(&dev->watchdog_timer));
 993 }
 994
 995 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 996                                const struct tc_ratespec *conf,
 997                                u64 rate64)
 998 {
 999         memset(r, 0, sizeof(*r));
1000         r->overhead = conf->overhead;
1001         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1002         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1003         r->mult = 1;
1004         /*
1005          * The deal here is to replace a divide by a reciprocal one
1006          * in fast path (a reciprocal divide is a multiply and a shift)
1007          *
1008          * Normal formula would be :
1009          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1010          *
1011          * We compute mult/shift to use instead :
1012          *  time_in_ns = (len * mult) >> shift;
1013          *
1014          * We try to get the highest possible mult value for accuracy,
1015          * but have to make sure no overflows will ever happen.
1016          */
1017         if (r->rate_bytes_ps > 0) {
1018                 u64 factor = NSEC_PER_SEC;
1019
1020                 for (;;) {
1021                         r->mult = div64_u64(factor, r->rate_bytes_ps);
1022                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
1023                                 break;
1024                         factor <<= 1;
1025                         r->shift++;
1026                 }
1027         }
1028 }
1029 EXPORT_SYMBOL(psched_ratecfg_precompute);
1030
1031 static void mini_qdisc_rcu_func(struct rcu_head *head)
1032 {
1033 }
1034
1035 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1036                           struct tcf_proto *tp_head)
1037 {
1038         struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1039         struct mini_Qdisc *miniq;
1040
1041         if (!tp_head) {
1042                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1043                 /* Wait for flying RCU callback before it is freed. */
1044                 rcu_barrier_bh();
1045                 return;
1046         }
1047
1048         miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1049                 &miniqp->miniq1 : &miniqp->miniq2;
1050
1051         /* We need to make sure that readers won't see the miniq
1052          * we are about to modify. So wait until previous call_rcu_bh callback
1053          * is done.
1054          */
1055         rcu_barrier_bh();
1056         miniq->filter_list = tp_head;
1057         rcu_assign_pointer(*miniqp->p_miniq, miniq);
1058
1059         if (miniq_old)
1060                 /* This is counterpart of the rcu barriers above. We need to
1061                  * block potential new user of miniq_old until all readers
1062                  * are not seeing it.
1063                  */
1064                 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1065 }
1066 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1067
1068 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1069                           struct mini_Qdisc __rcu **p_miniq)
1070 {
1071         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1072         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1073         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1074         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1075         miniqp->p_miniq = p_miniq;
1076 }
1077 EXPORT_SYMBOL(mini_qdisc_pair_init);