net/sched/sch_generic.c

   1 /*
   2  * net/sched/sch_generic.c      Generic packet scheduler routines.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
  11  *              - Ingress support
  12  */
  13
  14 #include <linux/bitops.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/string.h>
  20 #include <linux/errno.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/rtnetlink.h>
  24 #include <linux/init.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/list.h>
  27 #include <linux/slab.h>
  28 #include <linux/if_vlan.h>
  29 #include <net/sch_generic.h>
  30 #include <net/pkt_sched.h>
  31 #include <net/dst.h>
  32
  33 /* Qdisc to use by default */
  34 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  35 EXPORT_SYMBOL(default_qdisc_ops);
  36
  37 /* Main transmission queue. */
  38
  39 /* Modifications to data participating in scheduling must be protected with
  40  * qdisc_lock(qdisc) spinlock.
  41  *
  42  * The idea is the following:
  43  * - enqueue, dequeue are serialized via qdisc root lock
  44  * - ingress filtering is also serialized via qdisc root lock
  45  * - updates to tree and tree walking are only done under the rtnl mutex.
  46  */
  47
  48 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  49 {
  50         q->gso_skb = skb;
  51         q->qstats.requeues++;
  52         qdisc_qstats_backlog_inc(q, skb);
  53         q->q.qlen++;    /* it's still part of the queue */
  54         __netif_schedule(q);
  55
  56         return 0;
  57 }
  58
  59 static void try_bulk_dequeue_skb(struct Qdisc *q,
  60                                  struct sk_buff *skb,
  61                                  const struct netdev_queue *txq,
  62                                  int *packets)
  63 {
  64         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  65
  66         while (bytelimit > 0) {
  67                 struct sk_buff *nskb = q->dequeue(q);
  68
  69                 if (!nskb)
  70                         break;
  71
  72                 bytelimit -= nskb->len; /* covers GSO len */
  73                 skb->next = nskb;
  74                 skb = nskb;
  75                 (*packets)++; /* GSO counts as one pkt */
  76         }
  77         skb->next = NULL;
  78 }
  79
  80 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
  81  * A requeued skb (via q->gso_skb) can also be a SKB list.
  82  */
  83 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
  84                                    int *packets)
  85 {
  86         struct sk_buff *skb = q->gso_skb;
  87         const struct netdev_queue *txq = q->dev_queue;
  88
  89         *packets = 1;
  90         *validate = true;
  91         if (unlikely(skb)) {
  92                 /* check the reason of requeuing without tx lock first */
  93                 txq = skb_get_tx_queue(txq->dev, skb);
  94                 if (!netif_xmit_frozen_or_stopped(txq)) {
  95                         q->gso_skb = NULL;
  96                         qdisc_qstats_backlog_dec(q, skb);
  97                         q->q.qlen--;
  98                 } else
  99                         skb = NULL;
 100                 /* skb in gso_skb were already validated */
 101                 *validate = false;
 102         } else {
 103                 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
 104                     !netif_xmit_frozen_or_stopped(txq)) {
 105                         skb = q->dequeue(q);
 106                         if (skb && qdisc_may_bulk(q))
 107                                 try_bulk_dequeue_skb(q, skb, txq, packets);
 108                 }
 109         }
 110         return skb;
 111 }
 112
 113 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
 114                                            struct netdev_queue *dev_queue,
 115                                            struct Qdisc *q)
 116 {
 117         int ret;
 118
 119         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
 120                 /*
 121                  * Same CPU holding the lock. It may be a transient
 122                  * configuration error, when hard_start_xmit() recurses. We
 123                  * detect it by checking xmit owner and drop the packet when
 124                  * deadloop is detected. Return OK to try the next skb.
 125                  */
 126                 kfree_skb_list(skb);
 127                 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
 128                                      dev_queue->dev->name);
 129                 ret = qdisc_qlen(q);
 130         } else {
 131                 /*
 132                  * Another cpu is holding lock, requeue & delay xmits for
 133                  * some time.
 134                  */
 135                 __this_cpu_inc(softnet_data.cpu_collision);
 136                 ret = dev_requeue_skb(skb, q);
 137         }
 138
 139         return ret;
 140 }
 141
 142 /*
 143  * Transmit possibly several skbs, and handle the return status as
 144  * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
 145  * only one CPU can execute this function.
 146  *
 147  * Returns to the caller:
 148  *                              0  - queue is empty or throttled.
 149  *                              >0 - queue is not empty.
 150  */
 151 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 152                     struct net_device *dev, struct netdev_queue *txq,
 153                     spinlock_t *root_lock, bool validate)
 154 {
 155         int ret = NETDEV_TX_BUSY;
 156
 157         /* And release qdisc */
 158         spin_unlock(root_lock);
 159
 160         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 161         if (validate)
 162                 skb = validate_xmit_skb_list(skb, dev);
 163
 164         if (likely(skb)) {
 165                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 166                 if (!netif_xmit_frozen_or_stopped(txq))
 167                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 168
 169                 HARD_TX_UNLOCK(dev, txq);
 170         } else {
 171                 spin_lock(root_lock);
 172                 return qdisc_qlen(q);
 173         }
 174         spin_lock(root_lock);
 175
 176         if (dev_xmit_complete(ret)) {
 177                 /* Driver sent out skb successfully or skb was consumed */
 178                 ret = qdisc_qlen(q);
 179         } else if (ret == NETDEV_TX_LOCKED) {
 180                 /* Driver try lock failed */
 181                 ret = handle_dev_cpu_collision(skb, txq, q);
 182         } else {
 183                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 184                 if (unlikely(ret != NETDEV_TX_BUSY))
 185                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 186                                              dev->name, ret, q->q.qlen);
 187
 188                 ret = dev_requeue_skb(skb, q);
 189         }
 190
 191         if (ret && netif_xmit_frozen_or_stopped(txq))
 192                 ret = 0;
 193
 194         return ret;
 195 }
 196
 197 /*
 198  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 199  *
 200  * __QDISC___STATE_RUNNING guarantees only one CPU can process
 201  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 202  * this queue.
 203  *
 204  *  netif_tx_lock serializes accesses to device driver.
 205  *
 206  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 207  *  if one is grabbed, another must be free.
 208  *
 209  * Note, that this procedure can be called by a watchdog timer
 210  *
 211  * Returns to the caller:
 212  *                              0  - queue is empty or throttled.
 213  *                              >0 - queue is not empty.
 214  *
 215  */
 216 static inline int qdisc_restart(struct Qdisc *q, int *packets)
 217 {
 218         struct netdev_queue *txq;
 219         struct net_device *dev;
 220         spinlock_t *root_lock;
 221         struct sk_buff *skb;
 222         bool validate;
 223
 224         /* Dequeue packet */
 225         skb = dequeue_skb(q, &validate, packets);
 226         if (unlikely(!skb))
 227                 return 0;
 228
 229         root_lock = qdisc_lock(q);
 230         dev = qdisc_dev(q);
 231         txq = skb_get_tx_queue(dev, skb);
 232
 233         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 234 }
 235
 236 void __qdisc_run(struct Qdisc *q)
 237 {
 238         int quota = weight_p;
 239         int packets;
 240
 241         while (qdisc_restart(q, &packets)) {
 242                 /*
 243                  * Ordered by possible occurrence: Postpone processing if
 244                  * 1. we've exceeded packet quota
 245                  * 2. another process needs the CPU;
 246                  */
 247                 quota -= packets;
 248                 if (quota <= 0 || need_resched()) {
 249                         __netif_schedule(q);
 250                         break;
 251                 }
 252         }
 253
 254         qdisc_run_end(q);
 255 }
 256
 257 unsigned long dev_trans_start(struct net_device *dev)
 258 {
 259         unsigned long val, res;
 260         unsigned int i;
 261
 262         if (is_vlan_dev(dev))
 263                 dev = vlan_dev_real_dev(dev);
 264         res = dev->trans_start;
 265         for (i = 0; i < dev->num_tx_queues; i++) {
 266                 val = netdev_get_tx_queue(dev, i)->trans_start;
 267                 if (val && time_after(val, res))
 268                         res = val;
 269         }
 270         dev->trans_start = res;
 271
 272         return res;
 273 }
 274 EXPORT_SYMBOL(dev_trans_start);
 275
 276 static void dev_watchdog(unsigned long arg)
 277 {
 278         struct net_device *dev = (struct net_device *)arg;
 279
 280         netif_tx_lock(dev);
 281         if (!qdisc_tx_is_noop(dev)) {
 282                 if (netif_device_present(dev) &&
 283                     netif_running(dev) &&
 284                     netif_carrier_ok(dev)) {
 285                         int some_queue_timedout = 0;
 286                         unsigned int i;
 287                         unsigned long trans_start;
 288
 289                         for (i = 0; i < dev->num_tx_queues; i++) {
 290                                 struct netdev_queue *txq;
 291
 292                                 txq = netdev_get_tx_queue(dev, i);
 293                                 /*
 294                                  * old device drivers set dev->trans_start
 295                                  */
 296                                 trans_start = txq->trans_start ? : dev->trans_start;
 297                                 if (netif_xmit_stopped(txq) &&
 298                                     time_after(jiffies, (trans_start +
 299                                                          dev->watchdog_timeo))) {
 300                                         some_queue_timedout = 1;
 301                                         txq->trans_timeout++;
 302                                         break;
 303                                 }
 304                         }
 305
 306                         if (some_queue_timedout) {
 307                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 308                                        dev->name, netdev_drivername(dev), i);
 309                                 dev->netdev_ops->ndo_tx_timeout(dev);
 310                         }
 311                         if (!mod_timer(&dev->watchdog_timer,
 312                                        round_jiffies(jiffies +
 313                                                      dev->watchdog_timeo)))
 314                                 dev_hold(dev);
 315                 }
 316         }
 317         netif_tx_unlock(dev);
 318
 319         dev_put(dev);
 320 }
 321
 322 void __netdev_watchdog_up(struct net_device *dev)
 323 {
 324         if (dev->netdev_ops->ndo_tx_timeout) {
 325                 if (dev->watchdog_timeo <= 0)
 326                         dev->watchdog_timeo = 5*HZ;
 327                 if (!mod_timer(&dev->watchdog_timer,
 328                                round_jiffies(jiffies + dev->watchdog_timeo)))
 329                         dev_hold(dev);
 330         }
 331 }
 332 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 333
 334 static void dev_watchdog_up(struct net_device *dev)
 335 {
 336         __netdev_watchdog_up(dev);
 337 }
 338
 339 static void dev_watchdog_down(struct net_device *dev)
 340 {
 341         netif_tx_lock_bh(dev);
 342         if (del_timer(&dev->watchdog_timer))
 343                 dev_put(dev);
 344         netif_tx_unlock_bh(dev);
 345 }
 346
 347 /**
 348  *      netif_carrier_on - set carrier
 349  *      @dev: network device
 350  *
 351  * Device has detected that carrier.
 352  */
 353 void netif_carrier_on(struct net_device *dev)
 354 {
 355         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 356                 if (dev->reg_state == NETREG_UNINITIALIZED)
 357                         return;
 358                 atomic_inc(&dev->carrier_changes);
 359                 linkwatch_fire_event(dev);
 360                 if (netif_running(dev))
 361                         __netdev_watchdog_up(dev);
 362         }
 363 }
 364 EXPORT_SYMBOL(netif_carrier_on);
 365
 366 /**
 367  *      netif_carrier_off - clear carrier
 368  *      @dev: network device
 369  *
 370  * Device has detected loss of carrier.
 371  */
 372 void netif_carrier_off(struct net_device *dev)
 373 {
 374         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 375                 if (dev->reg_state == NETREG_UNINITIALIZED)
 376                         return;
 377                 atomic_inc(&dev->carrier_changes);
 378                 linkwatch_fire_event(dev);
 379         }
 380 }
 381 EXPORT_SYMBOL(netif_carrier_off);
 382
 383 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 384    under all circumstances. It is difficult to invent anything faster or
 385    cheaper.
 386  */
 387
 388 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 389 {
 390         kfree_skb(skb);
 391         return NET_XMIT_CN;
 392 }
 393
 394 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 395 {
 396         return NULL;
 397 }
 398
 399 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 400         .id             =       "noop",
 401         .priv_size      =       0,
 402         .enqueue        =       noop_enqueue,
 403         .dequeue        =       noop_dequeue,
 404         .peek           =       noop_dequeue,
 405         .owner          =       THIS_MODULE,
 406 };
 407
 408 static struct netdev_queue noop_netdev_queue = {
 409         .qdisc          =       &noop_qdisc,
 410         .qdisc_sleeping =       &noop_qdisc,
 411 };
 412
 413 struct Qdisc noop_qdisc = {
 414         .enqueue        =       noop_enqueue,
 415         .dequeue        =       noop_dequeue,
 416         .flags          =       TCQ_F_BUILTIN,
 417         .ops            =       &noop_qdisc_ops,
 418         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 419         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 420         .dev_queue      =       &noop_netdev_queue,
 421         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 422 };
 423 EXPORT_SYMBOL(noop_qdisc);
 424
 425 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 426 {
 427         /* register_qdisc() assigns a default of noop_enqueue if unset,
 428          * but __dev_queue_xmit() treats noqueue only as such
 429          * if this is NULL - so clear it here. */
 430         qdisc->enqueue = NULL;
 431         return 0;
 432 }
 433
 434 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 435         .id             =       "noqueue",
 436         .priv_size      =       0,
 437         .init           =       noqueue_init,
 438         .enqueue        =       noop_enqueue,
 439         .dequeue        =       noop_dequeue,
 440         .peek           =       noop_dequeue,
 441         .owner          =       THIS_MODULE,
 442 };
 443
 444 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 445         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 446 };
 447
 448 /* 3-band FIFO queue: old style, but should be a bit faster than
 449    generic prio+fifo combination.
 450  */
 451
 452 #define PFIFO_FAST_BANDS 3
 453
 454 /*
 455  * Private data for a pfifo_fast scheduler containing:
 456  *      - queues for the three band
 457  *      - bitmap indicating which of the bands contain skbs
 458  */
 459 struct pfifo_fast_priv {
 460         u32 bitmap;
 461         struct sk_buff_head q[PFIFO_FAST_BANDS];
 462 };
 463
 464 /*
 465  * Convert a bitmap to the first band number where an skb is queued, where:
 466  *      bitmap=0 means there are no skbs on any band.
 467  *      bitmap=1 means there is an skb on band 0.
 468  *      bitmap=7 means there are skbs on all 3 bands, etc.
 469  */
 470 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 471
 472 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 473                                              int band)
 474 {
 475         return priv->q + band;
 476 }
 477
 478 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 479 {
 480         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 481                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 482                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 483                 struct sk_buff_head *list = band2list(priv, band);
 484
 485                 priv->bitmap |= (1 << band);
 486                 qdisc->q.qlen++;
 487                 return __qdisc_enqueue_tail(skb, qdisc, list);
 488         }
 489
 490         return qdisc_drop(skb, qdisc);
 491 }
 492
 493 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 494 {
 495         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 496         int band = bitmap2band[priv->bitmap];
 497
 498         if (likely(band >= 0)) {
 499                 struct sk_buff_head *list = band2list(priv, band);
 500                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 501
 502                 qdisc->q.qlen--;
 503                 if (skb_queue_empty(list))
 504                         priv->bitmap &= ~(1 << band);
 505
 506                 return skb;
 507         }
 508
 509         return NULL;
 510 }
 511
 512 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 513 {
 514         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 515         int band = bitmap2band[priv->bitmap];
 516
 517         if (band >= 0) {
 518                 struct sk_buff_head *list = band2list(priv, band);
 519
 520                 return skb_peek(list);
 521         }
 522
 523         return NULL;
 524 }
 525
 526 static void pfifo_fast_reset(struct Qdisc *qdisc)
 527 {
 528         int prio;
 529         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 530
 531         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 532                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 533
 534         priv->bitmap = 0;
 535         qdisc->qstats.backlog = 0;
 536         qdisc->q.qlen = 0;
 537 }
 538
 539 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 540 {
 541         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 542
 543         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 544         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 545                 goto nla_put_failure;
 546         return skb->len;
 547
 548 nla_put_failure:
 549         return -1;
 550 }
 551
 552 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 553 {
 554         int prio;
 555         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 556
 557         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 558                 __skb_queue_head_init(band2list(priv, prio));
 559
 560         /* Can by-pass the queue discipline */
 561         qdisc->flags |= TCQ_F_CAN_BYPASS;
 562         return 0;
 563 }
 564
 565 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 566         .id             =       "pfifo_fast",
 567         .priv_size      =       sizeof(struct pfifo_fast_priv),
 568         .enqueue        =       pfifo_fast_enqueue,
 569         .dequeue        =       pfifo_fast_dequeue,
 570         .peek           =       pfifo_fast_peek,
 571         .init           =       pfifo_fast_init,
 572         .reset          =       pfifo_fast_reset,
 573         .dump           =       pfifo_fast_dump,
 574         .owner          =       THIS_MODULE,
 575 };
 576
 577 static struct lock_class_key qdisc_tx_busylock;
 578
 579 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 580                           const struct Qdisc_ops *ops)
 581 {
 582         void *p;
 583         struct Qdisc *sch;
 584         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 585         int err = -ENOBUFS;
 586         struct net_device *dev = dev_queue->dev;
 587
 588         p = kzalloc_node(size, GFP_KERNEL,
 589                          netdev_queue_numa_node_read(dev_queue));
 590
 591         if (!p)
 592                 goto errout;
 593         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 594         /* if we got non aligned memory, ask more and do alignment ourself */
 595         if (sch != p) {
 596                 kfree(p);
 597                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 598                                  netdev_queue_numa_node_read(dev_queue));
 599                 if (!p)
 600                         goto errout;
 601                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 602                 sch->padded = (char *) sch - (char *) p;
 603         }
 604         INIT_LIST_HEAD(&sch->list);
 605         skb_queue_head_init(&sch->q);
 606
 607         spin_lock_init(&sch->busylock);
 608         lockdep_set_class(&sch->busylock,
 609                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 610
 611         sch->ops = ops;
 612         sch->enqueue = ops->enqueue;
 613         sch->dequeue = ops->dequeue;
 614         sch->dev_queue = dev_queue;
 615         dev_hold(dev);
 616         atomic_set(&sch->refcnt, 1);
 617
 618         return sch;
 619 errout:
 620         return ERR_PTR(err);
 621 }
 622
 623 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 624                                 const struct Qdisc_ops *ops,
 625                                 unsigned int parentid)
 626 {
 627         struct Qdisc *sch;
 628
 629         if (!try_module_get(ops->owner))
 630                 return NULL;
 631
 632         sch = qdisc_alloc(dev_queue, ops);
 633         if (IS_ERR(sch)) {
 634                 module_put(ops->owner);
 635                 return NULL;
 636         }
 637         sch->parent = parentid;
 638
 639         if (!ops->init || ops->init(sch, NULL) == 0)
 640                 return sch;
 641
 642         qdisc_destroy(sch);
 643         return NULL;
 644 }
 645 EXPORT_SYMBOL(qdisc_create_dflt);
 646
 647 /* Under qdisc_lock(qdisc) and BH! */
 648
 649 void qdisc_reset(struct Qdisc *qdisc)
 650 {
 651         const struct Qdisc_ops *ops = qdisc->ops;
 652
 653         if (ops->reset)
 654                 ops->reset(qdisc);
 655
 656         if (qdisc->gso_skb) {
 657                 kfree_skb_list(qdisc->gso_skb);
 658                 qdisc->gso_skb = NULL;
 659                 qdisc->q.qlen = 0;
 660         }
 661 }
 662 EXPORT_SYMBOL(qdisc_reset);
 663
 664 static void qdisc_rcu_free(struct rcu_head *head)
 665 {
 666         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 667
 668         if (qdisc_is_percpu_stats(qdisc)) {
 669                 free_percpu(qdisc->cpu_bstats);
 670                 free_percpu(qdisc->cpu_qstats);
 671         }
 672
 673         kfree((char *) qdisc - qdisc->padded);
 674 }
 675
 676 void qdisc_destroy(struct Qdisc *qdisc)
 677 {
 678         const struct Qdisc_ops *ops;
 679
 680         if (!qdisc)
 681                 return;
 682         ops = qdisc->ops;
 683
 684         if (qdisc->flags & TCQ_F_BUILTIN ||
 685             !atomic_dec_and_test(&qdisc->refcnt))
 686                 return;
 687
 688 #ifdef CONFIG_NET_SCHED
 689         qdisc_list_del(qdisc);
 690
 691         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 692 #endif
 693         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 694         if (ops->reset)
 695                 ops->reset(qdisc);
 696         if (ops->destroy)
 697                 ops->destroy(qdisc);
 698
 699         module_put(ops->owner);
 700         dev_put(qdisc_dev(qdisc));
 701
 702         kfree_skb_list(qdisc->gso_skb);
 703         /*
 704          * gen_estimator est_timer() might access qdisc->q.lock,
 705          * wait a RCU grace period before freeing qdisc.
 706          */
 707         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 708 }
 709 EXPORT_SYMBOL(qdisc_destroy);
 710
 711 /* Attach toplevel qdisc to device queue. */
 712 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 713                               struct Qdisc *qdisc)
 714 {
 715         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 716         spinlock_t *root_lock;
 717
 718         root_lock = qdisc_lock(oqdisc);
 719         spin_lock_bh(root_lock);
 720
 721         /* Prune old scheduler */
 722         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 723                 qdisc_reset(oqdisc);
 724
 725         /* ... and graft new one */
 726         if (qdisc == NULL)
 727                 qdisc = &noop_qdisc;
 728         dev_queue->qdisc_sleeping = qdisc;
 729         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 730
 731         spin_unlock_bh(root_lock);
 732
 733         return oqdisc;
 734 }
 735 EXPORT_SYMBOL(dev_graft_qdisc);
 736
 737 static void attach_one_default_qdisc(struct net_device *dev,
 738                                      struct netdev_queue *dev_queue,
 739                                      void *_unused)
 740 {
 741         struct Qdisc *qdisc;
 742         const struct Qdisc_ops *ops = default_qdisc_ops;
 743
 744         if (dev->priv_flags & IFF_NO_QUEUE)
 745                 ops = &noqueue_qdisc_ops;
 746
 747         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 748         if (!qdisc) {
 749                 netdev_info(dev, "activation failed\n");
 750                 return;
 751         }
 752         if (!netif_is_multiqueue(dev))
 753                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 754         dev_queue->qdisc_sleeping = qdisc;
 755 }
 756
 757 static void attach_default_qdiscs(struct net_device *dev)
 758 {
 759         struct netdev_queue *txq;
 760         struct Qdisc *qdisc;
 761
 762         txq = netdev_get_tx_queue(dev, 0);
 763
 764         if (!netif_is_multiqueue(dev) ||
 765             dev->priv_flags & IFF_NO_QUEUE) {
 766                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 767                 dev->qdisc = txq->qdisc_sleeping;
 768                 atomic_inc(&dev->qdisc->refcnt);
 769         } else {
 770                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 771                 if (qdisc) {
 772                         dev->qdisc = qdisc;
 773                         qdisc->ops->attach(qdisc);
 774                 }
 775         }
 776 }
 777
 778 static void transition_one_qdisc(struct net_device *dev,
 779                                  struct netdev_queue *dev_queue,
 780                                  void *_need_watchdog)
 781 {
 782         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 783         int *need_watchdog_p = _need_watchdog;
 784
 785         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 786                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 787
 788         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 789         if (need_watchdog_p) {
 790                 dev_queue->trans_start = 0;
 791                 *need_watchdog_p = 1;
 792         }
 793 }
 794
 795 void dev_activate(struct net_device *dev)
 796 {
 797         int need_watchdog;
 798
 799         /* No queueing discipline is attached to device;
 800          * create default one for devices, which need queueing
 801          * and noqueue_qdisc for virtual interfaces
 802          */
 803
 804         if (dev->qdisc == &noop_qdisc)
 805                 attach_default_qdiscs(dev);
 806
 807         if (!netif_carrier_ok(dev))
 808                 /* Delay activation until next carrier-on event */
 809                 return;
 810
 811         need_watchdog = 0;
 812         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 813         if (dev_ingress_queue(dev))
 814                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 815
 816         if (need_watchdog) {
 817                 dev->trans_start = jiffies;
 818                 dev_watchdog_up(dev);
 819         }
 820 }
 821 EXPORT_SYMBOL(dev_activate);
 822
 823 static void dev_deactivate_queue(struct net_device *dev,
 824                                  struct netdev_queue *dev_queue,
 825                                  void *_qdisc_default)
 826 {
 827         struct Qdisc *qdisc_default = _qdisc_default;
 828         struct Qdisc *qdisc;
 829
 830         qdisc = rtnl_dereference(dev_queue->qdisc);
 831         if (qdisc) {
 832                 spin_lock_bh(qdisc_lock(qdisc));
 833
 834                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 835                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 836
 837                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 838                 qdisc_reset(qdisc);
 839
 840                 spin_unlock_bh(qdisc_lock(qdisc));
 841         }
 842 }
 843
 844 static bool some_qdisc_is_busy(struct net_device *dev)
 845 {
 846         unsigned int i;
 847
 848         for (i = 0; i < dev->num_tx_queues; i++) {
 849                 struct netdev_queue *dev_queue;
 850                 spinlock_t *root_lock;
 851                 struct Qdisc *q;
 852                 int val;
 853
 854                 dev_queue = netdev_get_tx_queue(dev, i);
 855                 q = dev_queue->qdisc_sleeping;
 856                 root_lock = qdisc_lock(q);
 857
 858                 spin_lock_bh(root_lock);
 859
 860                 val = (qdisc_is_running(q) ||
 861                        test_bit(__QDISC_STATE_SCHED, &q->state));
 862
 863                 spin_unlock_bh(root_lock);
 864
 865                 if (val)
 866                         return true;
 867         }
 868         return false;
 869 }
 870
 871 /**
 872  *      dev_deactivate_many - deactivate transmissions on several devices
 873  *      @head: list of devices to deactivate
 874  *
 875  *      This function returns only when all outstanding transmissions
 876  *      have completed, unless all devices are in dismantle phase.
 877  */
 878 void dev_deactivate_many(struct list_head *head)
 879 {
 880         struct net_device *dev;
 881         bool sync_needed = false;
 882
 883         list_for_each_entry(dev, head, close_list) {
 884                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 885                                          &noop_qdisc);
 886                 if (dev_ingress_queue(dev))
 887                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 888                                              &noop_qdisc);
 889
 890                 dev_watchdog_down(dev);
 891                 sync_needed |= !dev->dismantle;
 892         }
 893
 894         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 895          * This is avoided if all devices are in dismantle phase :
 896          * Caller will call synchronize_net() for us
 897          */
 898         if (sync_needed)
 899                 synchronize_net();
 900
 901         /* Wait for outstanding qdisc_run calls. */
 902         list_for_each_entry(dev, head, close_list)
 903                 while (some_qdisc_is_busy(dev))
 904                         yield();
 905 }
 906
 907 void dev_deactivate(struct net_device *dev)
 908 {
 909         LIST_HEAD(single);
 910
 911         list_add(&dev->close_list, &single);
 912         dev_deactivate_many(&single);
 913         list_del(&single);
 914 }
 915 EXPORT_SYMBOL(dev_deactivate);
 916
 917 static void dev_init_scheduler_queue(struct net_device *dev,
 918                                      struct netdev_queue *dev_queue,
 919                                      void *_qdisc)
 920 {
 921         struct Qdisc *qdisc = _qdisc;
 922
 923         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 924         dev_queue->qdisc_sleeping = qdisc;
 925 }
 926
 927 void dev_init_scheduler(struct net_device *dev)
 928 {
 929         dev->qdisc = &noop_qdisc;
 930         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 931         if (dev_ingress_queue(dev))
 932                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 933
 934         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 935 }
 936
 937 static void shutdown_scheduler_queue(struct net_device *dev,
 938                                      struct netdev_queue *dev_queue,
 939                                      void *_qdisc_default)
 940 {
 941         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 942         struct Qdisc *qdisc_default = _qdisc_default;
 943
 944         if (qdisc) {
 945                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 946                 dev_queue->qdisc_sleeping = qdisc_default;
 947
 948                 qdisc_destroy(qdisc);
 949         }
 950 }
 951
 952 void dev_shutdown(struct net_device *dev)
 953 {
 954         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 955         if (dev_ingress_queue(dev))
 956                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 957         qdisc_destroy(dev->qdisc);
 958         dev->qdisc = &noop_qdisc;
 959
 960         WARN_ON(timer_pending(&dev->watchdog_timer));
 961 }
 962
 963 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 964                                const struct tc_ratespec *conf,
 965                                u64 rate64)
 966 {
 967         memset(r, 0, sizeof(*r));
 968         r->overhead = conf->overhead;
 969         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
 970         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
 971         r->mult = 1;
 972         /*
 973          * The deal here is to replace a divide by a reciprocal one
 974          * in fast path (a reciprocal divide is a multiply and a shift)
 975          *
 976          * Normal formula would be :
 977          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
 978          *
 979          * We compute mult/shift to use instead :
 980          *  time_in_ns = (len * mult) >> shift;
 981          *
 982          * We try to get the highest possible mult value for accuracy,
 983          * but have to make sure no overflows will ever happen.
 984          */
 985         if (r->rate_bytes_ps > 0) {
 986                 u64 factor = NSEC_PER_SEC;
 987
 988                 for (;;) {
 989                         r->mult = div64_u64(factor, r->rate_bytes_ps);
 990                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
 991                                 break;
 992                         factor <<= 1;
 993                         r->shift++;
 994                 }
 995         }
 996 }
 997 EXPORT_SYMBOL(psched_ratecfg_precompute);