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