net/sched/sch_generic.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/sch_generic.c      Generic packet scheduler routines.
   4  *
   5  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   6  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
   7  *              - Ingress support
   8  */
   9
  10 #include <linux/bitops.h>
  11 #include <linux/module.h>
  12 #include <linux/types.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/string.h>
  16 #include <linux/errno.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/rtnetlink.h>
  20 #include <linux/init.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/list.h>
  23 #include <linux/slab.h>
  24 #include <linux/if_vlan.h>
  25 #include <linux/skb_array.h>
  26 #include <linux/if_macvlan.h>
  27 #include <net/sch_generic.h>
  28 #include <net/pkt_sched.h>
  29 #include <net/dst.h>
  30 #include <trace/events/qdisc.h>
  31 #include <trace/events/net.h>
  32 #include <net/xfrm.h>
  33
  34 /* Qdisc to use by default */
  35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  36 EXPORT_SYMBOL(default_qdisc_ops);
  37
  38 /* Main transmission queue. */
  39
  40 /* Modifications to data participating in scheduling must be protected with
  41  * qdisc_lock(qdisc) spinlock.
  42  *
  43  * The idea is the following:
  44  * - enqueue, dequeue are serialized via qdisc root lock
  45  * - ingress filtering is also serialized via qdisc root lock
  46  * - updates to tree and tree walking are only done under the rtnl mutex.
  47  */
  48
  49 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  50
  51 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  52 {
  53         const struct netdev_queue *txq = q->dev_queue;
  54         spinlock_t *lock = NULL;
  55         struct sk_buff *skb;
  56
  57         if (q->flags & TCQ_F_NOLOCK) {
  58                 lock = qdisc_lock(q);
  59                 spin_lock(lock);
  60         }
  61
  62         skb = skb_peek(&q->skb_bad_txq);
  63         if (skb) {
  64                 /* check the reason of requeuing without tx lock first */
  65                 txq = skb_get_tx_queue(txq->dev, skb);
  66                 if (!netif_xmit_frozen_or_stopped(txq)) {
  67                         skb = __skb_dequeue(&q->skb_bad_txq);
  68                         if (qdisc_is_percpu_stats(q)) {
  69                                 qdisc_qstats_cpu_backlog_dec(q, skb);
  70                                 qdisc_qstats_cpu_qlen_dec(q);
  71                         } else {
  72                                 qdisc_qstats_backlog_dec(q, skb);
  73                                 q->q.qlen--;
  74                         }
  75                 } else {
  76                         skb = SKB_XOFF_MAGIC;
  77                 }
  78         }
  79
  80         if (lock)
  81                 spin_unlock(lock);
  82
  83         return skb;
  84 }
  85
  86 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
  87 {
  88         struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
  89
  90         if (unlikely(skb))
  91                 skb = __skb_dequeue_bad_txq(q);
  92
  93         return skb;
  94 }
  95
  96 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
  97                                              struct sk_buff *skb)
  98 {
  99         spinlock_t *lock = NULL;
 100
 101         if (q->flags & TCQ_F_NOLOCK) {
 102                 lock = qdisc_lock(q);
 103                 spin_lock(lock);
 104         }
 105
 106         __skb_queue_tail(&q->skb_bad_txq, skb);
 107
 108         if (qdisc_is_percpu_stats(q)) {
 109                 qdisc_qstats_cpu_backlog_inc(q, skb);
 110                 qdisc_qstats_cpu_qlen_inc(q);
 111         } else {
 112                 qdisc_qstats_backlog_inc(q, skb);
 113                 q->q.qlen++;
 114         }
 115
 116         if (lock)
 117                 spin_unlock(lock);
 118 }
 119
 120 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 121 {
 122         spinlock_t *lock = NULL;
 123
 124         if (q->flags & TCQ_F_NOLOCK) {
 125                 lock = qdisc_lock(q);
 126                 spin_lock(lock);
 127         }
 128
 129         while (skb) {
 130                 struct sk_buff *next = skb->next;
 131
 132                 __skb_queue_tail(&q->gso_skb, skb);
 133
 134                 /* it's still part of the queue */
 135                 if (qdisc_is_percpu_stats(q)) {
 136                         qdisc_qstats_cpu_requeues_inc(q);
 137                         qdisc_qstats_cpu_backlog_inc(q, skb);
 138                         qdisc_qstats_cpu_qlen_inc(q);
 139                 } else {
 140                         q->qstats.requeues++;
 141                         qdisc_qstats_backlog_inc(q, skb);
 142                         q->q.qlen++;
 143                 }
 144
 145                 skb = next;
 146         }
 147         if (lock)
 148                 spin_unlock(lock);
 149         __netif_schedule(q);
 150 }
 151
 152 static void try_bulk_dequeue_skb(struct Qdisc *q,
 153                                  struct sk_buff *skb,
 154                                  const struct netdev_queue *txq,
 155                                  int *packets)
 156 {
 157         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 158
 159         while (bytelimit > 0) {
 160                 struct sk_buff *nskb = q->dequeue(q);
 161
 162                 if (!nskb)
 163                         break;
 164
 165                 bytelimit -= nskb->len; /* covers GSO len */
 166                 skb->next = nskb;
 167                 skb = nskb;
 168                 (*packets)++; /* GSO counts as one pkt */
 169         }
 170         skb_mark_not_on_list(skb);
 171 }
 172
 173 /* This variant of try_bulk_dequeue_skb() makes sure
 174  * all skbs in the chain are for the same txq
 175  */
 176 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 177                                       struct sk_buff *skb,
 178                                       int *packets)
 179 {
 180         int mapping = skb_get_queue_mapping(skb);
 181         struct sk_buff *nskb;
 182         int cnt = 0;
 183
 184         do {
 185                 nskb = q->dequeue(q);
 186                 if (!nskb)
 187                         break;
 188                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 189                         qdisc_enqueue_skb_bad_txq(q, nskb);
 190                         break;
 191                 }
 192                 skb->next = nskb;
 193                 skb = nskb;
 194         } while (++cnt < 8);
 195         (*packets) += cnt;
 196         skb_mark_not_on_list(skb);
 197 }
 198
 199 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 200  * A requeued skb (via q->gso_skb) can also be a SKB list.
 201  */
 202 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 203                                    int *packets)
 204 {
 205         const struct netdev_queue *txq = q->dev_queue;
 206         struct sk_buff *skb = NULL;
 207
 208         *packets = 1;
 209         if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 210                 spinlock_t *lock = NULL;
 211
 212                 if (q->flags & TCQ_F_NOLOCK) {
 213                         lock = qdisc_lock(q);
 214                         spin_lock(lock);
 215                 }
 216
 217                 skb = skb_peek(&q->gso_skb);
 218
 219                 /* skb may be null if another cpu pulls gso_skb off in between
 220                  * empty check and lock.
 221                  */
 222                 if (!skb) {
 223                         if (lock)
 224                                 spin_unlock(lock);
 225                         goto validate;
 226                 }
 227
 228                 /* skb in gso_skb were already validated */
 229                 *validate = false;
 230                 if (xfrm_offload(skb))
 231                         *validate = true;
 232                 /* check the reason of requeuing without tx lock first */
 233                 txq = skb_get_tx_queue(txq->dev, skb);
 234                 if (!netif_xmit_frozen_or_stopped(txq)) {
 235                         skb = __skb_dequeue(&q->gso_skb);
 236                         if (qdisc_is_percpu_stats(q)) {
 237                                 qdisc_qstats_cpu_backlog_dec(q, skb);
 238                                 qdisc_qstats_cpu_qlen_dec(q);
 239                         } else {
 240                                 qdisc_qstats_backlog_dec(q, skb);
 241                                 q->q.qlen--;
 242                         }
 243                 } else {
 244                         skb = NULL;
 245                 }
 246                 if (lock)
 247                         spin_unlock(lock);
 248                 goto trace;
 249         }
 250 validate:
 251         *validate = true;
 252
 253         if ((q->flags & TCQ_F_ONETXQUEUE) &&
 254             netif_xmit_frozen_or_stopped(txq))
 255                 return skb;
 256
 257         skb = qdisc_dequeue_skb_bad_txq(q);
 258         if (unlikely(skb)) {
 259                 if (skb == SKB_XOFF_MAGIC)
 260                         return NULL;
 261                 goto bulk;
 262         }
 263         skb = q->dequeue(q);
 264         if (skb) {
 265 bulk:
 266                 if (qdisc_may_bulk(q))
 267                         try_bulk_dequeue_skb(q, skb, txq, packets);
 268                 else
 269                         try_bulk_dequeue_skb_slow(q, skb, packets);
 270         }
 271 trace:
 272         trace_qdisc_dequeue(q, txq, *packets, skb);
 273         return skb;
 274 }
 275
 276 /*
 277  * Transmit possibly several skbs, and handle the return status as
 278  * required. Owning running seqcount bit guarantees that
 279  * only one CPU can execute this function.
 280  *
 281  * Returns to the caller:
 282  *                              false  - hardware queue frozen backoff
 283  *                              true   - feel free to send more pkts
 284  */
 285 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 286                      struct net_device *dev, struct netdev_queue *txq,
 287                      spinlock_t *root_lock, bool validate)
 288 {
 289         int ret = NETDEV_TX_BUSY;
 290         bool again = false;
 291
 292         /* And release qdisc */
 293         if (root_lock)
 294                 spin_unlock(root_lock);
 295
 296         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 297         if (validate)
 298                 skb = validate_xmit_skb_list(skb, dev, &again);
 299
 300 #ifdef CONFIG_XFRM_OFFLOAD
 301         if (unlikely(again)) {
 302                 if (root_lock)
 303                         spin_lock(root_lock);
 304
 305                 dev_requeue_skb(skb, q);
 306                 return false;
 307         }
 308 #endif
 309
 310         if (likely(skb)) {
 311                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 312                 if (!netif_xmit_frozen_or_stopped(txq))
 313                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 314
 315                 HARD_TX_UNLOCK(dev, txq);
 316         } else {
 317                 if (root_lock)
 318                         spin_lock(root_lock);
 319                 return true;
 320         }
 321
 322         if (root_lock)
 323                 spin_lock(root_lock);
 324
 325         if (!dev_xmit_complete(ret)) {
 326                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 327                 if (unlikely(ret != NETDEV_TX_BUSY))
 328                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 329                                              dev->name, ret, q->q.qlen);
 330
 331                 dev_requeue_skb(skb, q);
 332                 return false;
 333         }
 334
 335         return true;
 336 }
 337
 338 /*
 339  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 340  *
 341  * running seqcount guarantees only one CPU can process
 342  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 343  * this queue.
 344  *
 345  *  netif_tx_lock serializes accesses to device driver.
 346  *
 347  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 348  *  if one is grabbed, another must be free.
 349  *
 350  * Note, that this procedure can be called by a watchdog timer
 351  *
 352  * Returns to the caller:
 353  *                              0  - queue is empty or throttled.
 354  *                              >0 - queue is not empty.
 355  *
 356  */
 357 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 358 {
 359         spinlock_t *root_lock = NULL;
 360         struct netdev_queue *txq;
 361         struct net_device *dev;
 362         struct sk_buff *skb;
 363         bool validate;
 364
 365         /* Dequeue packet */
 366         skb = dequeue_skb(q, &validate, packets);
 367         if (unlikely(!skb))
 368                 return false;
 369
 370         if (!(q->flags & TCQ_F_NOLOCK))
 371                 root_lock = qdisc_lock(q);
 372
 373         dev = qdisc_dev(q);
 374         txq = skb_get_tx_queue(dev, skb);
 375
 376         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 377 }
 378
 379 void __qdisc_run(struct Qdisc *q)
 380 {
 381         int quota = dev_tx_weight;
 382         int packets;
 383
 384         while (qdisc_restart(q, &packets)) {
 385                 quota -= packets;
 386                 if (quota <= 0) {
 387                         __netif_schedule(q);
 388                         break;
 389                 }
 390         }
 391 }
 392
 393 unsigned long dev_trans_start(struct net_device *dev)
 394 {
 395         unsigned long val, res;
 396         unsigned int i;
 397
 398         if (is_vlan_dev(dev))
 399                 dev = vlan_dev_real_dev(dev);
 400         else if (netif_is_macvlan(dev))
 401                 dev = macvlan_dev_real_dev(dev);
 402         res = netdev_get_tx_queue(dev, 0)->trans_start;
 403         for (i = 1; i < dev->num_tx_queues; i++) {
 404                 val = netdev_get_tx_queue(dev, i)->trans_start;
 405                 if (val && time_after(val, res))
 406                         res = val;
 407         }
 408
 409         return res;
 410 }
 411 EXPORT_SYMBOL(dev_trans_start);
 412
 413 static void dev_watchdog(struct timer_list *t)
 414 {
 415         struct net_device *dev = from_timer(dev, t, watchdog_timer);
 416
 417         netif_tx_lock(dev);
 418         if (!qdisc_tx_is_noop(dev)) {
 419                 if (netif_device_present(dev) &&
 420                     netif_running(dev) &&
 421                     netif_carrier_ok(dev)) {
 422                         int some_queue_timedout = 0;
 423                         unsigned int i;
 424                         unsigned long trans_start;
 425
 426                         for (i = 0; i < dev->num_tx_queues; i++) {
 427                                 struct netdev_queue *txq;
 428
 429                                 txq = netdev_get_tx_queue(dev, i);
 430                                 trans_start = txq->trans_start;
 431                                 if (netif_xmit_stopped(txq) &&
 432                                     time_after(jiffies, (trans_start +
 433                                                          dev->watchdog_timeo))) {
 434                                         some_queue_timedout = 1;
 435                                         txq->trans_timeout++;
 436                                         break;
 437                                 }
 438                         }
 439
 440                         if (some_queue_timedout) {
 441                                 trace_net_dev_xmit_timeout(dev, i);
 442                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 443                                        dev->name, netdev_drivername(dev), i);
 444                                 dev->netdev_ops->ndo_tx_timeout(dev, i);
 445                         }
 446                         if (!mod_timer(&dev->watchdog_timer,
 447                                        round_jiffies(jiffies +
 448                                                      dev->watchdog_timeo)))
 449                                 dev_hold(dev);
 450                 }
 451         }
 452         netif_tx_unlock(dev);
 453
 454         dev_put(dev);
 455 }
 456
 457 void __netdev_watchdog_up(struct net_device *dev)
 458 {
 459         if (dev->netdev_ops->ndo_tx_timeout) {
 460                 if (dev->watchdog_timeo <= 0)
 461                         dev->watchdog_timeo = 5*HZ;
 462                 if (!mod_timer(&dev->watchdog_timer,
 463                                round_jiffies(jiffies + dev->watchdog_timeo)))
 464                         dev_hold(dev);
 465         }
 466 }
 467 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 468
 469 static void dev_watchdog_up(struct net_device *dev)
 470 {
 471         __netdev_watchdog_up(dev);
 472 }
 473
 474 static void dev_watchdog_down(struct net_device *dev)
 475 {
 476         netif_tx_lock_bh(dev);
 477         if (del_timer(&dev->watchdog_timer))
 478                 dev_put(dev);
 479         netif_tx_unlock_bh(dev);
 480 }
 481
 482 /**
 483  *      netif_carrier_on - set carrier
 484  *      @dev: network device
 485  *
 486  * Device has detected acquisition of carrier.
 487  */
 488 void netif_carrier_on(struct net_device *dev)
 489 {
 490         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 491                 if (dev->reg_state == NETREG_UNINITIALIZED)
 492                         return;
 493                 atomic_inc(&dev->carrier_up_count);
 494                 linkwatch_fire_event(dev);
 495                 if (netif_running(dev))
 496                         __netdev_watchdog_up(dev);
 497         }
 498 }
 499 EXPORT_SYMBOL(netif_carrier_on);
 500
 501 /**
 502  *      netif_carrier_off - clear carrier
 503  *      @dev: network device
 504  *
 505  * Device has detected loss of carrier.
 506  */
 507 void netif_carrier_off(struct net_device *dev)
 508 {
 509         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 510                 if (dev->reg_state == NETREG_UNINITIALIZED)
 511                         return;
 512                 atomic_inc(&dev->carrier_down_count);
 513                 linkwatch_fire_event(dev);
 514         }
 515 }
 516 EXPORT_SYMBOL(netif_carrier_off);
 517
 518 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 519    under all circumstances. It is difficult to invent anything faster or
 520    cheaper.
 521  */
 522
 523 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 524                         struct sk_buff **to_free)
 525 {
 526         __qdisc_drop(skb, to_free);
 527         return NET_XMIT_CN;
 528 }
 529
 530 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 531 {
 532         return NULL;
 533 }
 534
 535 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 536         .id             =       "noop",
 537         .priv_size      =       0,
 538         .enqueue        =       noop_enqueue,
 539         .dequeue        =       noop_dequeue,
 540         .peek           =       noop_dequeue,
 541         .owner          =       THIS_MODULE,
 542 };
 543
 544 static struct netdev_queue noop_netdev_queue = {
 545         RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 546         .qdisc_sleeping =       &noop_qdisc,
 547 };
 548
 549 struct Qdisc noop_qdisc = {
 550         .enqueue        =       noop_enqueue,
 551         .dequeue        =       noop_dequeue,
 552         .flags          =       TCQ_F_BUILTIN,
 553         .ops            =       &noop_qdisc_ops,
 554         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 555         .dev_queue      =       &noop_netdev_queue,
 556         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 557         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 558         .gso_skb = {
 559                 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
 560                 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 561                 .qlen = 0,
 562                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 563         },
 564         .skb_bad_txq = {
 565                 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 566                 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 567                 .qlen = 0,
 568                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 569         },
 570 };
 571 EXPORT_SYMBOL(noop_qdisc);
 572
 573 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 574                         struct netlink_ext_ack *extack)
 575 {
 576         /* register_qdisc() assigns a default of noop_enqueue if unset,
 577          * but __dev_queue_xmit() treats noqueue only as such
 578          * if this is NULL - so clear it here. */
 579         qdisc->enqueue = NULL;
 580         return 0;
 581 }
 582
 583 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 584         .id             =       "noqueue",
 585         .priv_size      =       0,
 586         .init           =       noqueue_init,
 587         .enqueue        =       noop_enqueue,
 588         .dequeue        =       noop_dequeue,
 589         .peek           =       noop_dequeue,
 590         .owner          =       THIS_MODULE,
 591 };
 592
 593 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 594         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 595 };
 596
 597 /* 3-band FIFO queue: old style, but should be a bit faster than
 598    generic prio+fifo combination.
 599  */
 600
 601 #define PFIFO_FAST_BANDS 3
 602
 603 /*
 604  * Private data for a pfifo_fast scheduler containing:
 605  *      - rings for priority bands
 606  */
 607 struct pfifo_fast_priv {
 608         struct skb_array q[PFIFO_FAST_BANDS];
 609 };
 610
 611 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 612                                           int band)
 613 {
 614         return &priv->q[band];
 615 }
 616
 617 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 618                               struct sk_buff **to_free)
 619 {
 620         int band = prio2band[skb->priority & TC_PRIO_MAX];
 621         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 622         struct skb_array *q = band2list(priv, band);
 623         unsigned int pkt_len = qdisc_pkt_len(skb);
 624         int err;
 625
 626         err = skb_array_produce(q, skb);
 627
 628         if (unlikely(err)) {
 629                 if (qdisc_is_percpu_stats(qdisc))
 630                         return qdisc_drop_cpu(skb, qdisc, to_free);
 631                 else
 632                         return qdisc_drop(skb, qdisc, to_free);
 633         }
 634
 635         qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 636         return NET_XMIT_SUCCESS;
 637 }
 638
 639 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 640 {
 641         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 642         struct sk_buff *skb = NULL;
 643         int band;
 644
 645         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 646                 struct skb_array *q = band2list(priv, band);
 647
 648                 if (__skb_array_empty(q))
 649                         continue;
 650
 651                 skb = __skb_array_consume(q);
 652         }
 653         if (likely(skb)) {
 654                 qdisc_update_stats_at_dequeue(qdisc, skb);
 655         } else {
 656                 WRITE_ONCE(qdisc->empty, true);
 657         }
 658
 659         return skb;
 660 }
 661
 662 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 663 {
 664         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 665         struct sk_buff *skb = NULL;
 666         int band;
 667
 668         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 669                 struct skb_array *q = band2list(priv, band);
 670
 671                 skb = __skb_array_peek(q);
 672         }
 673
 674         return skb;
 675 }
 676
 677 static void pfifo_fast_reset(struct Qdisc *qdisc)
 678 {
 679         int i, band;
 680         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 681
 682         for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 683                 struct skb_array *q = band2list(priv, band);
 684                 struct sk_buff *skb;
 685
 686                 /* NULL ring is possible if destroy path is due to a failed
 687                  * skb_array_init() in pfifo_fast_init() case.
 688                  */
 689                 if (!q->ring.queue)
 690                         continue;
 691
 692                 while ((skb = __skb_array_consume(q)) != NULL)
 693                         kfree_skb(skb);
 694         }
 695
 696         if (qdisc_is_percpu_stats(qdisc)) {
 697                 for_each_possible_cpu(i) {
 698                         struct gnet_stats_queue *q;
 699
 700                         q = per_cpu_ptr(qdisc->cpu_qstats, i);
 701                         q->backlog = 0;
 702                         q->qlen = 0;
 703                 }
 704         }
 705 }
 706
 707 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 708 {
 709         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 710
 711         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 712         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 713                 goto nla_put_failure;
 714         return skb->len;
 715
 716 nla_put_failure:
 717         return -1;
 718 }
 719
 720 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 721                            struct netlink_ext_ack *extack)
 722 {
 723         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 724         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 725         int prio;
 726
 727         /* guard against zero length rings */
 728         if (!qlen)
 729                 return -EINVAL;
 730
 731         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 732                 struct skb_array *q = band2list(priv, prio);
 733                 int err;
 734
 735                 err = skb_array_init(q, qlen, GFP_KERNEL);
 736                 if (err)
 737                         return -ENOMEM;
 738         }
 739
 740         /* Can by-pass the queue discipline */
 741         qdisc->flags |= TCQ_F_CAN_BYPASS;
 742         return 0;
 743 }
 744
 745 static void pfifo_fast_destroy(struct Qdisc *sch)
 746 {
 747         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 748         int prio;
 749
 750         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 751                 struct skb_array *q = band2list(priv, prio);
 752
 753                 /* NULL ring is possible if destroy path is due to a failed
 754                  * skb_array_init() in pfifo_fast_init() case.
 755                  */
 756                 if (!q->ring.queue)
 757                         continue;
 758                 /* Destroy ring but no need to kfree_skb because a call to
 759                  * pfifo_fast_reset() has already done that work.
 760                  */
 761                 ptr_ring_cleanup(&q->ring, NULL);
 762         }
 763 }
 764
 765 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 766                                           unsigned int new_len)
 767 {
 768         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 769         struct skb_array *bands[PFIFO_FAST_BANDS];
 770         int prio;
 771
 772         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 773                 struct skb_array *q = band2list(priv, prio);
 774
 775                 bands[prio] = q;
 776         }
 777
 778         return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 779                                          GFP_KERNEL);
 780 }
 781
 782 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 783         .id             =       "pfifo_fast",
 784         .priv_size      =       sizeof(struct pfifo_fast_priv),
 785         .enqueue        =       pfifo_fast_enqueue,
 786         .dequeue        =       pfifo_fast_dequeue,
 787         .peek           =       pfifo_fast_peek,
 788         .init           =       pfifo_fast_init,
 789         .destroy        =       pfifo_fast_destroy,
 790         .reset          =       pfifo_fast_reset,
 791         .dump           =       pfifo_fast_dump,
 792         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 793         .owner          =       THIS_MODULE,
 794         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 795 };
 796 EXPORT_SYMBOL(pfifo_fast_ops);
 797
 798 static struct lock_class_key qdisc_tx_busylock;
 799 static struct lock_class_key qdisc_running_key;
 800
 801 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 802                           const struct Qdisc_ops *ops,
 803                           struct netlink_ext_ack *extack)
 804 {
 805         void *p;
 806         struct Qdisc *sch;
 807         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 808         int err = -ENOBUFS;
 809         struct net_device *dev;
 810
 811         if (!dev_queue) {
 812                 NL_SET_ERR_MSG(extack, "No device queue given");
 813                 err = -EINVAL;
 814                 goto errout;
 815         }
 816
 817         dev = dev_queue->dev;
 818         p = kzalloc_node(size, GFP_KERNEL,
 819                          netdev_queue_numa_node_read(dev_queue));
 820
 821         if (!p)
 822                 goto errout;
 823         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 824         /* if we got non aligned memory, ask more and do alignment ourself */
 825         if (sch != p) {
 826                 kfree(p);
 827                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 828                                  netdev_queue_numa_node_read(dev_queue));
 829                 if (!p)
 830                         goto errout;
 831                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 832                 sch->padded = (char *) sch - (char *) p;
 833         }
 834         __skb_queue_head_init(&sch->gso_skb);
 835         __skb_queue_head_init(&sch->skb_bad_txq);
 836         qdisc_skb_head_init(&sch->q);
 837         spin_lock_init(&sch->q.lock);
 838
 839         if (ops->static_flags & TCQ_F_CPUSTATS) {
 840                 sch->cpu_bstats =
 841                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 842                 if (!sch->cpu_bstats)
 843                         goto errout1;
 844
 845                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 846                 if (!sch->cpu_qstats) {
 847                         free_percpu(sch->cpu_bstats);
 848                         goto errout1;
 849                 }
 850         }
 851
 852         spin_lock_init(&sch->busylock);
 853         lockdep_set_class(&sch->busylock,
 854                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 855
 856         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 857         spin_lock_init(&sch->seqlock);
 858         lockdep_set_class(&sch->busylock,
 859                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 860
 861         seqcount_init(&sch->running);
 862         lockdep_set_class(&sch->running,
 863                           dev->qdisc_running_key ?: &qdisc_running_key);
 864
 865         sch->ops = ops;
 866         sch->flags = ops->static_flags;
 867         sch->enqueue = ops->enqueue;
 868         sch->dequeue = ops->dequeue;
 869         sch->dev_queue = dev_queue;
 870         sch->empty = true;
 871         dev_hold(dev);
 872         refcount_set(&sch->refcnt, 1);
 873
 874         return sch;
 875 errout1:
 876         kfree(p);
 877 errout:
 878         return ERR_PTR(err);
 879 }
 880
 881 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 882                                 const struct Qdisc_ops *ops,
 883                                 unsigned int parentid,
 884                                 struct netlink_ext_ack *extack)
 885 {
 886         struct Qdisc *sch;
 887
 888         if (!try_module_get(ops->owner)) {
 889                 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 890                 return NULL;
 891         }
 892
 893         sch = qdisc_alloc(dev_queue, ops, extack);
 894         if (IS_ERR(sch)) {
 895                 module_put(ops->owner);
 896                 return NULL;
 897         }
 898         sch->parent = parentid;
 899
 900         if (!ops->init || ops->init(sch, NULL, extack) == 0) {
 901                 trace_qdisc_create(ops, dev_queue->dev, parentid);
 902                 return sch;
 903         }
 904
 905         qdisc_put(sch);
 906         return NULL;
 907 }
 908 EXPORT_SYMBOL(qdisc_create_dflt);
 909
 910 /* Under qdisc_lock(qdisc) and BH! */
 911
 912 void qdisc_reset(struct Qdisc *qdisc)
 913 {
 914         const struct Qdisc_ops *ops = qdisc->ops;
 915         struct sk_buff *skb, *tmp;
 916
 917         trace_qdisc_reset(qdisc);
 918
 919         if (ops->reset)
 920                 ops->reset(qdisc);
 921
 922         skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 923                 __skb_unlink(skb, &qdisc->gso_skb);
 924                 kfree_skb_list(skb);
 925         }
 926
 927         skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 928                 __skb_unlink(skb, &qdisc->skb_bad_txq);
 929                 kfree_skb_list(skb);
 930         }
 931
 932         qdisc->q.qlen = 0;
 933         qdisc->qstats.backlog = 0;
 934 }
 935 EXPORT_SYMBOL(qdisc_reset);
 936
 937 void qdisc_free(struct Qdisc *qdisc)
 938 {
 939         if (qdisc_is_percpu_stats(qdisc)) {
 940                 free_percpu(qdisc->cpu_bstats);
 941                 free_percpu(qdisc->cpu_qstats);
 942         }
 943
 944         kfree((char *) qdisc - qdisc->padded);
 945 }
 946
 947 static void qdisc_free_cb(struct rcu_head *head)
 948 {
 949         struct Qdisc *q = container_of(head, struct Qdisc, rcu);
 950
 951         qdisc_free(q);
 952 }
 953
 954 static void qdisc_destroy(struct Qdisc *qdisc)
 955 {
 956         const struct Qdisc_ops  *ops = qdisc->ops;
 957
 958 #ifdef CONFIG_NET_SCHED
 959         qdisc_hash_del(qdisc);
 960
 961         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 962 #endif
 963         gen_kill_estimator(&qdisc->rate_est);
 964
 965         qdisc_reset(qdisc);
 966
 967         if (ops->destroy)
 968                 ops->destroy(qdisc);
 969
 970         module_put(ops->owner);
 971         dev_put(qdisc_dev(qdisc));
 972
 973         trace_qdisc_destroy(qdisc);
 974
 975         call_rcu(&qdisc->rcu, qdisc_free_cb);
 976 }
 977
 978 void qdisc_put(struct Qdisc *qdisc)
 979 {
 980         if (!qdisc)
 981                 return;
 982
 983         if (qdisc->flags & TCQ_F_BUILTIN ||
 984             !refcount_dec_and_test(&qdisc->refcnt))
 985                 return;
 986
 987         qdisc_destroy(qdisc);
 988 }
 989 EXPORT_SYMBOL(qdisc_put);
 990
 991 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
 992  * Intended to be used as optimization, this function only takes rtnl lock if
 993  * qdisc reference counter reached zero.
 994  */
 995
 996 void qdisc_put_unlocked(struct Qdisc *qdisc)
 997 {
 998         if (qdisc->flags & TCQ_F_BUILTIN ||
 999             !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1000                 return;
1001
1002         qdisc_destroy(qdisc);
1003         rtnl_unlock();
1004 }
1005 EXPORT_SYMBOL(qdisc_put_unlocked);
1006
1007 /* Attach toplevel qdisc to device queue. */
1008 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1009                               struct Qdisc *qdisc)
1010 {
1011         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1012         spinlock_t *root_lock;
1013
1014         root_lock = qdisc_lock(oqdisc);
1015         spin_lock_bh(root_lock);
1016
1017         /* ... and graft new one */
1018         if (qdisc == NULL)
1019                 qdisc = &noop_qdisc;
1020         dev_queue->qdisc_sleeping = qdisc;
1021         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1022
1023         spin_unlock_bh(root_lock);
1024
1025         return oqdisc;
1026 }
1027 EXPORT_SYMBOL(dev_graft_qdisc);
1028
1029 static void attach_one_default_qdisc(struct net_device *dev,
1030                                      struct netdev_queue *dev_queue,
1031                                      void *_unused)
1032 {
1033         struct Qdisc *qdisc;
1034         const struct Qdisc_ops *ops = default_qdisc_ops;
1035
1036         if (dev->priv_flags & IFF_NO_QUEUE)
1037                 ops = &noqueue_qdisc_ops;
1038         else if(dev->type == ARPHRD_CAN)
1039                 ops = &pfifo_fast_ops;
1040
1041         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1042         if (!qdisc)
1043                 return;
1044
1045         if (!netif_is_multiqueue(dev))
1046                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1047         dev_queue->qdisc_sleeping = qdisc;
1048 }
1049
1050 static void attach_default_qdiscs(struct net_device *dev)
1051 {
1052         struct netdev_queue *txq;
1053         struct Qdisc *qdisc;
1054
1055         txq = netdev_get_tx_queue(dev, 0);
1056
1057         if (!netif_is_multiqueue(dev) ||
1058             dev->priv_flags & IFF_NO_QUEUE) {
1059                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1060                 dev->qdisc = txq->qdisc_sleeping;
1061                 qdisc_refcount_inc(dev->qdisc);
1062         } else {
1063                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1064                 if (qdisc) {
1065                         dev->qdisc = qdisc;
1066                         qdisc->ops->attach(qdisc);
1067                 }
1068         }
1069
1070         /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1071         if (dev->qdisc == &noop_qdisc) {
1072                 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1073                             default_qdisc_ops->id, noqueue_qdisc_ops.id);
1074                 dev->priv_flags |= IFF_NO_QUEUE;
1075                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1076                 dev->qdisc = txq->qdisc_sleeping;
1077                 qdisc_refcount_inc(dev->qdisc);
1078                 dev->priv_flags ^= IFF_NO_QUEUE;
1079         }
1080
1081 #ifdef CONFIG_NET_SCHED
1082         if (dev->qdisc != &noop_qdisc)
1083                 qdisc_hash_add(dev->qdisc, false);
1084 #endif
1085 }
1086
1087 static void transition_one_qdisc(struct net_device *dev,
1088                                  struct netdev_queue *dev_queue,
1089                                  void *_need_watchdog)
1090 {
1091         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1092         int *need_watchdog_p = _need_watchdog;
1093
1094         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1095                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1096
1097         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1098         if (need_watchdog_p) {
1099                 dev_queue->trans_start = 0;
1100                 *need_watchdog_p = 1;
1101         }
1102 }
1103
1104 void dev_activate(struct net_device *dev)
1105 {
1106         int need_watchdog;
1107
1108         /* No queueing discipline is attached to device;
1109          * create default one for devices, which need queueing
1110          * and noqueue_qdisc for virtual interfaces
1111          */
1112
1113         if (dev->qdisc == &noop_qdisc)
1114                 attach_default_qdiscs(dev);
1115
1116         if (!netif_carrier_ok(dev))
1117                 /* Delay activation until next carrier-on event */
1118                 return;
1119
1120         need_watchdog = 0;
1121         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1122         if (dev_ingress_queue(dev))
1123                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1124
1125         if (need_watchdog) {
1126                 netif_trans_update(dev);
1127                 dev_watchdog_up(dev);
1128         }
1129 }
1130 EXPORT_SYMBOL(dev_activate);
1131
1132 static void qdisc_deactivate(struct Qdisc *qdisc)
1133 {
1134         bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1135
1136         if (qdisc->flags & TCQ_F_BUILTIN)
1137                 return;
1138         if (test_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state))
1139                 return;
1140
1141         if (nolock)
1142                 spin_lock_bh(&qdisc->seqlock);
1143         spin_lock_bh(qdisc_lock(qdisc));
1144
1145         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1146
1147         qdisc_reset(qdisc);
1148
1149         spin_unlock_bh(qdisc_lock(qdisc));
1150         if (nolock)
1151                 spin_unlock_bh(&qdisc->seqlock);
1152 }
1153
1154 static void dev_deactivate_queue(struct net_device *dev,
1155                                  struct netdev_queue *dev_queue,
1156                                  void *_qdisc_default)
1157 {
1158         struct Qdisc *qdisc_default = _qdisc_default;
1159         struct Qdisc *qdisc;
1160
1161         qdisc = rtnl_dereference(dev_queue->qdisc);
1162         if (qdisc) {
1163                 qdisc_deactivate(qdisc);
1164                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1165         }
1166 }
1167
1168 static bool some_qdisc_is_busy(struct net_device *dev)
1169 {
1170         unsigned int i;
1171
1172         for (i = 0; i < dev->num_tx_queues; i++) {
1173                 struct netdev_queue *dev_queue;
1174                 spinlock_t *root_lock;
1175                 struct Qdisc *q;
1176                 int val;
1177
1178                 dev_queue = netdev_get_tx_queue(dev, i);
1179                 q = dev_queue->qdisc_sleeping;
1180
1181                 root_lock = qdisc_lock(q);
1182                 spin_lock_bh(root_lock);
1183
1184                 val = (qdisc_is_running(q) ||
1185                        test_bit(__QDISC_STATE_SCHED, &q->state));
1186
1187                 spin_unlock_bh(root_lock);
1188
1189                 if (val)
1190                         return true;
1191         }
1192         return false;
1193 }
1194
1195 /**
1196  *      dev_deactivate_many - deactivate transmissions on several devices
1197  *      @head: list of devices to deactivate
1198  *
1199  *      This function returns only when all outstanding transmissions
1200  *      have completed, unless all devices are in dismantle phase.
1201  */
1202 void dev_deactivate_many(struct list_head *head)
1203 {
1204         struct net_device *dev;
1205
1206         list_for_each_entry(dev, head, close_list) {
1207                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1208                                          &noop_qdisc);
1209                 if (dev_ingress_queue(dev))
1210                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
1211                                              &noop_qdisc);
1212
1213                 dev_watchdog_down(dev);
1214         }
1215
1216         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1217          * This is avoided if all devices are in dismantle phase :
1218          * Caller will call synchronize_net() for us
1219          */
1220         synchronize_net();
1221
1222         /* Wait for outstanding qdisc_run calls. */
1223         list_for_each_entry(dev, head, close_list) {
1224                 while (some_qdisc_is_busy(dev)) {
1225                         /* wait_event() would avoid this sleep-loop but would
1226                          * require expensive checks in the fast paths of packet
1227                          * processing which isn't worth it.
1228                          */
1229                         schedule_timeout_uninterruptible(1);
1230                 }
1231         }
1232 }
1233
1234 void dev_deactivate(struct net_device *dev)
1235 {
1236         LIST_HEAD(single);
1237
1238         list_add(&dev->close_list, &single);
1239         dev_deactivate_many(&single);
1240         list_del(&single);
1241 }
1242 EXPORT_SYMBOL(dev_deactivate);
1243
1244 static int qdisc_change_tx_queue_len(struct net_device *dev,
1245                                      struct netdev_queue *dev_queue)
1246 {
1247         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1248         const struct Qdisc_ops *ops = qdisc->ops;
1249
1250         if (ops->change_tx_queue_len)
1251                 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1252         return 0;
1253 }
1254
1255 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1256 {
1257         bool up = dev->flags & IFF_UP;
1258         unsigned int i;
1259         int ret = 0;
1260
1261         if (up)
1262                 dev_deactivate(dev);
1263
1264         for (i = 0; i < dev->num_tx_queues; i++) {
1265                 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1266
1267                 /* TODO: revert changes on a partial failure */
1268                 if (ret)
1269                         break;
1270         }
1271
1272         if (up)
1273                 dev_activate(dev);
1274         return ret;
1275 }
1276
1277 static void dev_init_scheduler_queue(struct net_device *dev,
1278                                      struct netdev_queue *dev_queue,
1279                                      void *_qdisc)
1280 {
1281         struct Qdisc *qdisc = _qdisc;
1282
1283         rcu_assign_pointer(dev_queue->qdisc, qdisc);
1284         dev_queue->qdisc_sleeping = qdisc;
1285 }
1286
1287 void dev_init_scheduler(struct net_device *dev)
1288 {
1289         dev->qdisc = &noop_qdisc;
1290         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1291         if (dev_ingress_queue(dev))
1292                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1293
1294         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1295 }
1296
1297 static void shutdown_scheduler_queue(struct net_device *dev,
1298                                      struct netdev_queue *dev_queue,
1299                                      void *_qdisc_default)
1300 {
1301         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1302         struct Qdisc *qdisc_default = _qdisc_default;
1303
1304         if (qdisc) {
1305                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1306                 dev_queue->qdisc_sleeping = qdisc_default;
1307
1308                 qdisc_put(qdisc);
1309         }
1310 }
1311
1312 void dev_shutdown(struct net_device *dev)
1313 {
1314         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1315         if (dev_ingress_queue(dev))
1316                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1317         qdisc_put(dev->qdisc);
1318         dev->qdisc = &noop_qdisc;
1319
1320         WARN_ON(timer_pending(&dev->watchdog_timer));
1321 }
1322
1323 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1324                                const struct tc_ratespec *conf,
1325                                u64 rate64)
1326 {
1327         memset(r, 0, sizeof(*r));
1328         r->overhead = conf->overhead;
1329         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1330         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1331         r->mult = 1;
1332         /*
1333          * The deal here is to replace a divide by a reciprocal one
1334          * in fast path (a reciprocal divide is a multiply and a shift)
1335          *
1336          * Normal formula would be :
1337          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1338          *
1339          * We compute mult/shift to use instead :
1340          *  time_in_ns = (len * mult) >> shift;
1341          *
1342          * We try to get the highest possible mult value for accuracy,
1343          * but have to make sure no overflows will ever happen.
1344          */
1345         if (r->rate_bytes_ps > 0) {
1346                 u64 factor = NSEC_PER_SEC;
1347
1348                 for (;;) {
1349                         r->mult = div64_u64(factor, r->rate_bytes_ps);
1350                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
1351                                 break;
1352                         factor <<= 1;
1353                         r->shift++;
1354                 }
1355         }
1356 }
1357 EXPORT_SYMBOL(psched_ratecfg_precompute);
1358
1359 static void mini_qdisc_rcu_func(struct rcu_head *head)
1360 {
1361 }
1362
1363 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1364                           struct tcf_proto *tp_head)
1365 {
1366         /* Protected with chain0->filter_chain_lock.
1367          * Can't access chain directly because tp_head can be NULL.
1368          */
1369         struct mini_Qdisc *miniq_old =
1370                 rcu_dereference_protected(*miniqp->p_miniq, 1);
1371         struct mini_Qdisc *miniq;
1372
1373         if (!tp_head) {
1374                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1375                 /* Wait for flying RCU callback before it is freed. */
1376                 rcu_barrier();
1377                 return;
1378         }
1379
1380         miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1381                 &miniqp->miniq1 : &miniqp->miniq2;
1382
1383         /* We need to make sure that readers won't see the miniq
1384          * we are about to modify. So wait until previous call_rcu callback
1385          * is done.
1386          */
1387         rcu_barrier();
1388         miniq->filter_list = tp_head;
1389         rcu_assign_pointer(*miniqp->p_miniq, miniq);
1390
1391         if (miniq_old)
1392                 /* This is counterpart of the rcu barriers above. We need to
1393                  * block potential new user of miniq_old until all readers
1394                  * are not seeing it.
1395                  */
1396                 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1397 }
1398 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1399
1400 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1401                                 struct tcf_block *block)
1402 {
1403         miniqp->miniq1.block = block;
1404         miniqp->miniq2.block = block;
1405 }
1406 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1407
1408 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1409                           struct mini_Qdisc __rcu **p_miniq)
1410 {
1411         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1412         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1413         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1414         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1415         miniqp->p_miniq = p_miniq;
1416 }
1417 EXPORT_SYMBOL(mini_qdisc_pair_init);