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_atomic_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_atomic_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_atomic_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_atomic_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 acquisition of 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         RCU_POINTER_INITIALIZER(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_atomic_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_atomic_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         }
 718 }
 719
 720 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 721 {
 722         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 723
 724         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 725         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 726                 goto nla_put_failure;
 727         return skb->len;
 728
 729 nla_put_failure:
 730         return -1;
 731 }
 732
 733 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 734                            struct netlink_ext_ack *extack)
 735 {
 736         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 737         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 738         int prio;
 739
 740         /* guard against zero length rings */
 741         if (!qlen)
 742                 return -EINVAL;
 743
 744         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 745                 struct skb_array *q = band2list(priv, prio);
 746                 int err;
 747
 748                 err = skb_array_init(q, qlen, GFP_KERNEL);
 749                 if (err)
 750                         return -ENOMEM;
 751         }
 752
 753         /* Can by-pass the queue discipline */
 754         qdisc->flags |= TCQ_F_CAN_BYPASS;
 755         return 0;
 756 }
 757
 758 static void pfifo_fast_destroy(struct Qdisc *sch)
 759 {
 760         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 761         int prio;
 762
 763         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 764                 struct skb_array *q = band2list(priv, prio);
 765
 766                 /* NULL ring is possible if destroy path is due to a failed
 767                  * skb_array_init() in pfifo_fast_init() case.
 768                  */
 769                 if (!q->ring.queue)
 770                         continue;
 771                 /* Destroy ring but no need to kfree_skb because a call to
 772                  * pfifo_fast_reset() has already done that work.
 773                  */
 774                 ptr_ring_cleanup(&q->ring, NULL);
 775         }
 776 }
 777
 778 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 779                                           unsigned int new_len)
 780 {
 781         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 782         struct skb_array *bands[PFIFO_FAST_BANDS];
 783         int prio;
 784
 785         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 786                 struct skb_array *q = band2list(priv, prio);
 787
 788                 bands[prio] = q;
 789         }
 790
 791         return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 792                                          GFP_KERNEL);
 793 }
 794
 795 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 796         .id             =       "pfifo_fast",
 797         .priv_size      =       sizeof(struct pfifo_fast_priv),
 798         .enqueue        =       pfifo_fast_enqueue,
 799         .dequeue        =       pfifo_fast_dequeue,
 800         .peek           =       pfifo_fast_peek,
 801         .init           =       pfifo_fast_init,
 802         .destroy        =       pfifo_fast_destroy,
 803         .reset          =       pfifo_fast_reset,
 804         .dump           =       pfifo_fast_dump,
 805         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 806         .owner          =       THIS_MODULE,
 807         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 808 };
 809 EXPORT_SYMBOL(pfifo_fast_ops);
 810
 811 static struct lock_class_key qdisc_tx_busylock;
 812 static struct lock_class_key qdisc_running_key;
 813
 814 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 815                           const struct Qdisc_ops *ops,
 816                           struct netlink_ext_ack *extack)
 817 {
 818         void *p;
 819         struct Qdisc *sch;
 820         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 821         int err = -ENOBUFS;
 822         struct net_device *dev;
 823
 824         if (!dev_queue) {
 825                 NL_SET_ERR_MSG(extack, "No device queue given");
 826                 err = -EINVAL;
 827                 goto errout;
 828         }
 829
 830         dev = dev_queue->dev;
 831         p = kzalloc_node(size, GFP_KERNEL,
 832                          netdev_queue_numa_node_read(dev_queue));
 833
 834         if (!p)
 835                 goto errout;
 836         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 837         /* if we got non aligned memory, ask more and do alignment ourself */
 838         if (sch != p) {
 839                 kfree(p);
 840                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 841                                  netdev_queue_numa_node_read(dev_queue));
 842                 if (!p)
 843                         goto errout;
 844                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 845                 sch->padded = (char *) sch - (char *) p;
 846         }
 847         __skb_queue_head_init(&sch->gso_skb);
 848         __skb_queue_head_init(&sch->skb_bad_txq);
 849         qdisc_skb_head_init(&sch->q);
 850         spin_lock_init(&sch->q.lock);
 851
 852         if (ops->static_flags & TCQ_F_CPUSTATS) {
 853                 sch->cpu_bstats =
 854                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 855                 if (!sch->cpu_bstats)
 856                         goto errout1;
 857
 858                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 859                 if (!sch->cpu_qstats) {
 860                         free_percpu(sch->cpu_bstats);
 861                         goto errout1;
 862                 }
 863         }
 864
 865         spin_lock_init(&sch->busylock);
 866         lockdep_set_class(&sch->busylock,
 867                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 868
 869         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 870         spin_lock_init(&sch->seqlock);
 871         lockdep_set_class(&sch->busylock,
 872                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 873
 874         seqcount_init(&sch->running);
 875         lockdep_set_class(&sch->running,
 876                           dev->qdisc_running_key ?: &qdisc_running_key);
 877
 878         sch->ops = ops;
 879         sch->flags = ops->static_flags;
 880         sch->enqueue = ops->enqueue;
 881         sch->dequeue = ops->dequeue;
 882         sch->dev_queue = dev_queue;
 883         dev_hold(dev);
 884         refcount_set(&sch->refcnt, 1);
 885
 886         return sch;
 887 errout1:
 888         kfree(p);
 889 errout:
 890         return ERR_PTR(err);
 891 }
 892
 893 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 894                                 const struct Qdisc_ops *ops,
 895                                 unsigned int parentid,
 896                                 struct netlink_ext_ack *extack)
 897 {
 898         struct Qdisc *sch;
 899
 900         if (!try_module_get(ops->owner)) {
 901                 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 902                 return NULL;
 903         }
 904
 905         sch = qdisc_alloc(dev_queue, ops, extack);
 906         if (IS_ERR(sch)) {
 907                 module_put(ops->owner);
 908                 return NULL;
 909         }
 910         sch->parent = parentid;
 911
 912         if (!ops->init || ops->init(sch, NULL, extack) == 0)
 913                 return sch;
 914
 915         qdisc_put(sch);
 916         return NULL;
 917 }
 918 EXPORT_SYMBOL(qdisc_create_dflt);
 919
 920 /* Under qdisc_lock(qdisc) and BH! */
 921
 922 void qdisc_reset(struct Qdisc *qdisc)
 923 {
 924         const struct Qdisc_ops *ops = qdisc->ops;
 925         struct sk_buff *skb, *tmp;
 926
 927         if (ops->reset)
 928                 ops->reset(qdisc);
 929
 930         skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 931                 __skb_unlink(skb, &qdisc->gso_skb);
 932                 kfree_skb_list(skb);
 933         }
 934
 935         skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 936                 __skb_unlink(skb, &qdisc->skb_bad_txq);
 937                 kfree_skb_list(skb);
 938         }
 939
 940         qdisc->q.qlen = 0;
 941         qdisc->qstats.backlog = 0;
 942 }
 943 EXPORT_SYMBOL(qdisc_reset);
 944
 945 void qdisc_free(struct Qdisc *qdisc)
 946 {
 947         if (qdisc_is_percpu_stats(qdisc)) {
 948                 free_percpu(qdisc->cpu_bstats);
 949                 free_percpu(qdisc->cpu_qstats);
 950         }
 951
 952         kfree((char *) qdisc - qdisc->padded);
 953 }
 954
 955 static void qdisc_free_cb(struct rcu_head *head)
 956 {
 957         struct Qdisc *q = container_of(head, struct Qdisc, rcu);
 958
 959         qdisc_free(q);
 960 }
 961
 962 static void qdisc_destroy(struct Qdisc *qdisc)
 963 {
 964         const struct Qdisc_ops  *ops = qdisc->ops;
 965         struct sk_buff *skb, *tmp;
 966
 967 #ifdef CONFIG_NET_SCHED
 968         qdisc_hash_del(qdisc);
 969
 970         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 971 #endif
 972         gen_kill_estimator(&qdisc->rate_est);
 973         if (ops->reset)
 974                 ops->reset(qdisc);
 975         if (ops->destroy)
 976                 ops->destroy(qdisc);
 977
 978         module_put(ops->owner);
 979         dev_put(qdisc_dev(qdisc));
 980
 981         skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 982                 __skb_unlink(skb, &qdisc->gso_skb);
 983                 kfree_skb_list(skb);
 984         }
 985
 986         skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 987                 __skb_unlink(skb, &qdisc->skb_bad_txq);
 988                 kfree_skb_list(skb);
 989         }
 990
 991         call_rcu(&qdisc->rcu, qdisc_free_cb);
 992 }
 993
 994 void qdisc_put(struct Qdisc *qdisc)
 995 {
 996         if (qdisc->flags & TCQ_F_BUILTIN ||
 997             !refcount_dec_and_test(&qdisc->refcnt))
 998                 return;
 999
1000         qdisc_destroy(qdisc);
1001 }
1002 EXPORT_SYMBOL(qdisc_put);
1003
1004 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1005  * Intended to be used as optimization, this function only takes rtnl lock if
1006  * qdisc reference counter reached zero.
1007  */
1008
1009 void qdisc_put_unlocked(struct Qdisc *qdisc)
1010 {
1011         if (qdisc->flags & TCQ_F_BUILTIN ||
1012             !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1013                 return;
1014
1015         qdisc_destroy(qdisc);
1016         rtnl_unlock();
1017 }
1018 EXPORT_SYMBOL(qdisc_put_unlocked);
1019
1020 /* Attach toplevel qdisc to device queue. */
1021 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1022                               struct Qdisc *qdisc)
1023 {
1024         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1025         spinlock_t *root_lock;
1026
1027         root_lock = qdisc_lock(oqdisc);
1028         spin_lock_bh(root_lock);
1029
1030         /* ... and graft new one */
1031         if (qdisc == NULL)
1032                 qdisc = &noop_qdisc;
1033         dev_queue->qdisc_sleeping = qdisc;
1034         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1035
1036         spin_unlock_bh(root_lock);
1037
1038         return oqdisc;
1039 }
1040 EXPORT_SYMBOL(dev_graft_qdisc);
1041
1042 static void attach_one_default_qdisc(struct net_device *dev,
1043                                      struct netdev_queue *dev_queue,
1044                                      void *_unused)
1045 {
1046         struct Qdisc *qdisc;
1047         const struct Qdisc_ops *ops = default_qdisc_ops;
1048
1049         if (dev->priv_flags & IFF_NO_QUEUE)
1050                 ops = &noqueue_qdisc_ops;
1051
1052         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1053         if (!qdisc) {
1054                 netdev_info(dev, "activation failed\n");
1055                 return;
1056         }
1057         if (!netif_is_multiqueue(dev))
1058                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1059         dev_queue->qdisc_sleeping = qdisc;
1060 }
1061
1062 static void attach_default_qdiscs(struct net_device *dev)
1063 {
1064         struct netdev_queue *txq;
1065         struct Qdisc *qdisc;
1066
1067         txq = netdev_get_tx_queue(dev, 0);
1068
1069         if (!netif_is_multiqueue(dev) ||
1070             dev->priv_flags & IFF_NO_QUEUE) {
1071                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1072                 dev->qdisc = txq->qdisc_sleeping;
1073                 qdisc_refcount_inc(dev->qdisc);
1074         } else {
1075                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1076                 if (qdisc) {
1077                         dev->qdisc = qdisc;
1078                         qdisc->ops->attach(qdisc);
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 dev_deactivate_queue(struct net_device *dev,
1133                                  struct netdev_queue *dev_queue,
1134                                  void *_qdisc_default)
1135 {
1136         struct Qdisc *qdisc_default = _qdisc_default;
1137         struct Qdisc *qdisc;
1138
1139         qdisc = rtnl_dereference(dev_queue->qdisc);
1140         if (qdisc) {
1141                 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1142
1143                 if (nolock)
1144                         spin_lock_bh(&qdisc->seqlock);
1145                 spin_lock_bh(qdisc_lock(qdisc));
1146
1147                 if (!(qdisc->flags & TCQ_F_BUILTIN))
1148                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1149
1150                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1151                 qdisc_reset(qdisc);
1152
1153                 spin_unlock_bh(qdisc_lock(qdisc));
1154                 if (nolock)
1155                         spin_unlock_bh(&qdisc->seqlock);
1156         }
1157 }
1158
1159 static bool some_qdisc_is_busy(struct net_device *dev)
1160 {
1161         unsigned int i;
1162
1163         for (i = 0; i < dev->num_tx_queues; i++) {
1164                 struct netdev_queue *dev_queue;
1165                 spinlock_t *root_lock;
1166                 struct Qdisc *q;
1167                 int val;
1168
1169                 dev_queue = netdev_get_tx_queue(dev, i);
1170                 q = dev_queue->qdisc_sleeping;
1171
1172                 root_lock = qdisc_lock(q);
1173                 spin_lock_bh(root_lock);
1174
1175                 val = (qdisc_is_running(q) ||
1176                        test_bit(__QDISC_STATE_SCHED, &q->state));
1177
1178                 spin_unlock_bh(root_lock);
1179
1180                 if (val)
1181                         return true;
1182         }
1183         return false;
1184 }
1185
1186 static void dev_qdisc_reset(struct net_device *dev,
1187                             struct netdev_queue *dev_queue,
1188                             void *none)
1189 {
1190         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1191
1192         if (qdisc)
1193                 qdisc_reset(qdisc);
1194 }
1195
1196 /**
1197  *      dev_deactivate_many - deactivate transmissions on several devices
1198  *      @head: list of devices to deactivate
1199  *
1200  *      This function returns only when all outstanding transmissions
1201  *      have completed, unless all devices are in dismantle phase.
1202  */
1203 void dev_deactivate_many(struct list_head *head)
1204 {
1205         struct net_device *dev;
1206
1207         list_for_each_entry(dev, head, close_list) {
1208                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1209                                          &noop_qdisc);
1210                 if (dev_ingress_queue(dev))
1211                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
1212                                              &noop_qdisc);
1213
1214                 dev_watchdog_down(dev);
1215         }
1216
1217         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1218          * This is avoided if all devices are in dismantle phase :
1219          * Caller will call synchronize_net() for us
1220          */
1221         synchronize_net();
1222
1223         /* Wait for outstanding qdisc_run calls. */
1224         list_for_each_entry(dev, head, close_list) {
1225                 while (some_qdisc_is_busy(dev))
1226                         yield();
1227                 /* The new qdisc is assigned at this point so we can safely
1228                  * unwind stale skb lists and qdisc statistics
1229                  */
1230                 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1231                 if (dev_ingress_queue(dev))
1232                         dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1233         }
1234 }
1235
1236 void dev_deactivate(struct net_device *dev)
1237 {
1238         LIST_HEAD(single);
1239
1240         list_add(&dev->close_list, &single);
1241         dev_deactivate_many(&single);
1242         list_del(&single);
1243 }
1244 EXPORT_SYMBOL(dev_deactivate);
1245
1246 static int qdisc_change_tx_queue_len(struct net_device *dev,
1247                                      struct netdev_queue *dev_queue)
1248 {
1249         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1250         const struct Qdisc_ops *ops = qdisc->ops;
1251
1252         if (ops->change_tx_queue_len)
1253                 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1254         return 0;
1255 }
1256
1257 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1258 {
1259         bool up = dev->flags & IFF_UP;
1260         unsigned int i;
1261         int ret = 0;
1262
1263         if (up)
1264                 dev_deactivate(dev);
1265
1266         for (i = 0; i < dev->num_tx_queues; i++) {
1267                 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1268
1269                 /* TODO: revert changes on a partial failure */
1270                 if (ret)
1271                         break;
1272         }
1273
1274         if (up)
1275                 dev_activate(dev);
1276         return ret;
1277 }
1278
1279 static void dev_init_scheduler_queue(struct net_device *dev,
1280                                      struct netdev_queue *dev_queue,
1281                                      void *_qdisc)
1282 {
1283         struct Qdisc *qdisc = _qdisc;
1284
1285         rcu_assign_pointer(dev_queue->qdisc, qdisc);
1286         dev_queue->qdisc_sleeping = qdisc;
1287 }
1288
1289 void dev_init_scheduler(struct net_device *dev)
1290 {
1291         dev->qdisc = &noop_qdisc;
1292         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1293         if (dev_ingress_queue(dev))
1294                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1295
1296         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1297 }
1298
1299 static void shutdown_scheduler_queue(struct net_device *dev,
1300                                      struct netdev_queue *dev_queue,
1301                                      void *_qdisc_default)
1302 {
1303         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1304         struct Qdisc *qdisc_default = _qdisc_default;
1305
1306         if (qdisc) {
1307                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1308                 dev_queue->qdisc_sleeping = qdisc_default;
1309
1310                 qdisc_put(qdisc);
1311         }
1312 }
1313
1314 void dev_shutdown(struct net_device *dev)
1315 {
1316         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1317         if (dev_ingress_queue(dev))
1318                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1319         qdisc_put(dev->qdisc);
1320         dev->qdisc = &noop_qdisc;
1321
1322         WARN_ON(timer_pending(&dev->watchdog_timer));
1323 }
1324
1325 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1326                                const struct tc_ratespec *conf,
1327                                u64 rate64)
1328 {
1329         memset(r, 0, sizeof(*r));
1330         r->overhead = conf->overhead;
1331         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1332         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1333         r->mult = 1;
1334         /*
1335          * The deal here is to replace a divide by a reciprocal one
1336          * in fast path (a reciprocal divide is a multiply and a shift)
1337          *
1338          * Normal formula would be :
1339          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1340          *
1341          * We compute mult/shift to use instead :
1342          *  time_in_ns = (len * mult) >> shift;
1343          *
1344          * We try to get the highest possible mult value for accuracy,
1345          * but have to make sure no overflows will ever happen.
1346          */
1347         if (r->rate_bytes_ps > 0) {
1348                 u64 factor = NSEC_PER_SEC;
1349
1350                 for (;;) {
1351                         r->mult = div64_u64(factor, r->rate_bytes_ps);
1352                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
1353                                 break;
1354                         factor <<= 1;
1355                         r->shift++;
1356                 }
1357         }
1358 }
1359 EXPORT_SYMBOL(psched_ratecfg_precompute);
1360
1361 static void mini_qdisc_rcu_func(struct rcu_head *head)
1362 {
1363 }
1364
1365 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1366                           struct tcf_proto *tp_head)
1367 {
1368         /* Protected with chain0->filter_chain_lock.
1369          * Can't access chain directly because tp_head can be NULL.
1370          */
1371         struct mini_Qdisc *miniq_old =
1372                 rcu_dereference_protected(*miniqp->p_miniq, 1);
1373         struct mini_Qdisc *miniq;
1374
1375         if (!tp_head) {
1376                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1377                 /* Wait for flying RCU callback before it is freed. */
1378                 rcu_barrier();
1379                 return;
1380         }
1381
1382         miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1383                 &miniqp->miniq1 : &miniqp->miniq2;
1384
1385         /* We need to make sure that readers won't see the miniq
1386          * we are about to modify. So wait until previous call_rcu callback
1387          * is done.
1388          */
1389         rcu_barrier();
1390         miniq->filter_list = tp_head;
1391         rcu_assign_pointer(*miniqp->p_miniq, miniq);
1392
1393         if (miniq_old)
1394                 /* This is counterpart of the rcu barriers above. We need to
1395                  * block potential new user of miniq_old until all readers
1396                  * are not seeing it.
1397                  */
1398                 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1399 }
1400 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1401
1402 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1403                           struct mini_Qdisc __rcu **p_miniq)
1404 {
1405         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1406         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1407         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1408         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1409         miniqp->p_miniq = p_miniq;
1410 }
1411 EXPORT_SYMBOL(mini_qdisc_pair_init);