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