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