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