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