net/sched/sch_generic.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/sch_generic.c      Generic packet scheduler routines.
   4  *
   5  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   6  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
   7  *              - Ingress support
   8  */
   9
  10 #include <linux/bitops.h>
  11 #include <linux/module.h>
  12 #include <linux/types.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/string.h>
  16 #include <linux/errno.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/rtnetlink.h>
  20 #include <linux/init.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/list.h>
  23 #include <linux/slab.h>
  24 #include <linux/if_vlan.h>
  25 #include <linux/skb_array.h>
  26 #include <linux/if_macvlan.h>
  27 #include <net/sch_generic.h>
  28 #include <net/pkt_sched.h>
  29 #include <net/dst.h>
  30 #include <trace/events/qdisc.h>
  31 #include <trace/events/net.h>
  32 #include <net/xfrm.h>
  33
  34 /* Qdisc to use by default */
  35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  36 EXPORT_SYMBOL(default_qdisc_ops);
  37
  38 static void qdisc_maybe_clear_missed(struct Qdisc *q,
  39                                      const struct netdev_queue *txq)
  40 {
  41         clear_bit(__QDISC_STATE_MISSED, &q->state);
  42
  43         /* Make sure the below netif_xmit_frozen_or_stopped()
  44          * checking happens after clearing STATE_MISSED.
  45          */
  46         smp_mb__after_atomic();
  47
  48         /* Checking netif_xmit_frozen_or_stopped() again to
  49          * make sure STATE_MISSED is set if the STATE_MISSED
  50          * set by netif_tx_wake_queue()'s rescheduling of
  51          * net_tx_action() is cleared by the above clear_bit().
  52          */
  53         if (!netif_xmit_frozen_or_stopped(txq))
  54                 set_bit(__QDISC_STATE_MISSED, &q->state);
  55         else
  56                 set_bit(__QDISC_STATE_DRAINING, &q->state);
  57 }
  58
  59 /* Main transmission queue. */
  60
  61 /* Modifications to data participating in scheduling must be protected with
  62  * qdisc_lock(qdisc) spinlock.
  63  *
  64  * The idea is the following:
  65  * - enqueue, dequeue are serialized via qdisc root lock
  66  * - ingress filtering is also serialized via qdisc root lock
  67  * - updates to tree and tree walking are only done under the rtnl mutex.
  68  */
  69
  70 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  71
  72 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  73 {
  74         const struct netdev_queue *txq = q->dev_queue;
  75         spinlock_t *lock = NULL;
  76         struct sk_buff *skb;
  77
  78         if (q->flags & TCQ_F_NOLOCK) {
  79                 lock = qdisc_lock(q);
  80                 spin_lock(lock);
  81         }
  82
  83         skb = skb_peek(&q->skb_bad_txq);
  84         if (skb) {
  85                 /* check the reason of requeuing without tx lock first */
  86                 txq = skb_get_tx_queue(txq->dev, skb);
  87                 if (!netif_xmit_frozen_or_stopped(txq)) {
  88                         skb = __skb_dequeue(&q->skb_bad_txq);
  89                         if (qdisc_is_percpu_stats(q)) {
  90                                 qdisc_qstats_cpu_backlog_dec(q, skb);
  91                                 qdisc_qstats_cpu_qlen_dec(q);
  92                         } else {
  93                                 qdisc_qstats_backlog_dec(q, skb);
  94                                 q->q.qlen--;
  95                         }
  96                 } else {
  97                         skb = SKB_XOFF_MAGIC;
  98                         qdisc_maybe_clear_missed(q, txq);
  99                 }
 100         }
 101
 102         if (lock)
 103                 spin_unlock(lock);
 104
 105         return skb;
 106 }
 107
 108 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
 109 {
 110         struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
 111
 112         if (unlikely(skb))
 113                 skb = __skb_dequeue_bad_txq(q);
 114
 115         return skb;
 116 }
 117
 118 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
 119                                              struct sk_buff *skb)
 120 {
 121         spinlock_t *lock = NULL;
 122
 123         if (q->flags & TCQ_F_NOLOCK) {
 124                 lock = qdisc_lock(q);
 125                 spin_lock(lock);
 126         }
 127
 128         __skb_queue_tail(&q->skb_bad_txq, skb);
 129
 130         if (qdisc_is_percpu_stats(q)) {
 131                 qdisc_qstats_cpu_backlog_inc(q, skb);
 132                 qdisc_qstats_cpu_qlen_inc(q);
 133         } else {
 134                 qdisc_qstats_backlog_inc(q, skb);
 135                 q->q.qlen++;
 136         }
 137
 138         if (lock)
 139                 spin_unlock(lock);
 140 }
 141
 142 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 143 {
 144         spinlock_t *lock = NULL;
 145
 146         if (q->flags & TCQ_F_NOLOCK) {
 147                 lock = qdisc_lock(q);
 148                 spin_lock(lock);
 149         }
 150
 151         while (skb) {
 152                 struct sk_buff *next = skb->next;
 153
 154                 __skb_queue_tail(&q->gso_skb, skb);
 155
 156                 /* it's still part of the queue */
 157                 if (qdisc_is_percpu_stats(q)) {
 158                         qdisc_qstats_cpu_requeues_inc(q);
 159                         qdisc_qstats_cpu_backlog_inc(q, skb);
 160                         qdisc_qstats_cpu_qlen_inc(q);
 161                 } else {
 162                         q->qstats.requeues++;
 163                         qdisc_qstats_backlog_inc(q, skb);
 164                         q->q.qlen++;
 165                 }
 166
 167                 skb = next;
 168         }
 169
 170         if (lock) {
 171                 spin_unlock(lock);
 172                 set_bit(__QDISC_STATE_MISSED, &q->state);
 173         } else {
 174                 __netif_schedule(q);
 175         }
 176 }
 177
 178 static void try_bulk_dequeue_skb(struct Qdisc *q,
 179                                  struct sk_buff *skb,
 180                                  const struct netdev_queue *txq,
 181                                  int *packets)
 182 {
 183         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 184
 185         while (bytelimit > 0) {
 186                 struct sk_buff *nskb = q->dequeue(q);
 187
 188                 if (!nskb)
 189                         break;
 190
 191                 bytelimit -= nskb->len; /* covers GSO len */
 192                 skb->next = nskb;
 193                 skb = nskb;
 194                 (*packets)++; /* GSO counts as one pkt */
 195         }
 196         skb_mark_not_on_list(skb);
 197 }
 198
 199 /* This variant of try_bulk_dequeue_skb() makes sure
 200  * all skbs in the chain are for the same txq
 201  */
 202 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 203                                       struct sk_buff *skb,
 204                                       int *packets)
 205 {
 206         int mapping = skb_get_queue_mapping(skb);
 207         struct sk_buff *nskb;
 208         int cnt = 0;
 209
 210         do {
 211                 nskb = q->dequeue(q);
 212                 if (!nskb)
 213                         break;
 214                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 215                         qdisc_enqueue_skb_bad_txq(q, nskb);
 216                         break;
 217                 }
 218                 skb->next = nskb;
 219                 skb = nskb;
 220         } while (++cnt < 8);
 221         (*packets) += cnt;
 222         skb_mark_not_on_list(skb);
 223 }
 224
 225 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 226  * A requeued skb (via q->gso_skb) can also be a SKB list.
 227  */
 228 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 229                                    int *packets)
 230 {
 231         const struct netdev_queue *txq = q->dev_queue;
 232         struct sk_buff *skb = NULL;
 233
 234         *packets = 1;
 235         if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 236                 spinlock_t *lock = NULL;
 237
 238                 if (q->flags & TCQ_F_NOLOCK) {
 239                         lock = qdisc_lock(q);
 240                         spin_lock(lock);
 241                 }
 242
 243                 skb = skb_peek(&q->gso_skb);
 244
 245                 /* skb may be null if another cpu pulls gso_skb off in between
 246                  * empty check and lock.
 247                  */
 248                 if (!skb) {
 249                         if (lock)
 250                                 spin_unlock(lock);
 251                         goto validate;
 252                 }
 253
 254                 /* skb in gso_skb were already validated */
 255                 *validate = false;
 256                 if (xfrm_offload(skb))
 257                         *validate = true;
 258                 /* check the reason of requeuing without tx lock first */
 259                 txq = skb_get_tx_queue(txq->dev, skb);
 260                 if (!netif_xmit_frozen_or_stopped(txq)) {
 261                         skb = __skb_dequeue(&q->gso_skb);
 262                         if (qdisc_is_percpu_stats(q)) {
 263                                 qdisc_qstats_cpu_backlog_dec(q, skb);
 264                                 qdisc_qstats_cpu_qlen_dec(q);
 265                         } else {
 266                                 qdisc_qstats_backlog_dec(q, skb);
 267                                 q->q.qlen--;
 268                         }
 269                 } else {
 270                         skb = NULL;
 271                         qdisc_maybe_clear_missed(q, txq);
 272                 }
 273                 if (lock)
 274                         spin_unlock(lock);
 275                 goto trace;
 276         }
 277 validate:
 278         *validate = true;
 279
 280         if ((q->flags & TCQ_F_ONETXQUEUE) &&
 281             netif_xmit_frozen_or_stopped(txq)) {
 282                 qdisc_maybe_clear_missed(q, txq);
 283                 return skb;
 284         }
 285
 286         skb = qdisc_dequeue_skb_bad_txq(q);
 287         if (unlikely(skb)) {
 288                 if (skb == SKB_XOFF_MAGIC)
 289                         return NULL;
 290                 goto bulk;
 291         }
 292         skb = q->dequeue(q);
 293         if (skb) {
 294 bulk:
 295                 if (qdisc_may_bulk(q))
 296                         try_bulk_dequeue_skb(q, skb, txq, packets);
 297                 else
 298                         try_bulk_dequeue_skb_slow(q, skb, packets);
 299         }
 300 trace:
 301         trace_qdisc_dequeue(q, txq, *packets, skb);
 302         return skb;
 303 }
 304
 305 /*
 306  * Transmit possibly several skbs, and handle the return status as
 307  * required. Owning running seqcount bit guarantees that
 308  * only one CPU can execute this function.
 309  *
 310  * Returns to the caller:
 311  *                              false  - hardware queue frozen backoff
 312  *                              true   - feel free to send more pkts
 313  */
 314 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 315                      struct net_device *dev, struct netdev_queue *txq,
 316                      spinlock_t *root_lock, bool validate)
 317 {
 318         int ret = NETDEV_TX_BUSY;
 319         bool again = false;
 320
 321         /* And release qdisc */
 322         if (root_lock)
 323                 spin_unlock(root_lock);
 324
 325         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 326         if (validate)
 327                 skb = validate_xmit_skb_list(skb, dev, &again);
 328
 329 #ifdef CONFIG_XFRM_OFFLOAD
 330         if (unlikely(again)) {
 331                 if (root_lock)
 332                         spin_lock(root_lock);
 333
 334                 dev_requeue_skb(skb, q);
 335                 return false;
 336         }
 337 #endif
 338
 339         if (likely(skb)) {
 340                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 341                 if (!netif_xmit_frozen_or_stopped(txq))
 342                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 343                 else
 344                         qdisc_maybe_clear_missed(q, txq);
 345
 346                 HARD_TX_UNLOCK(dev, txq);
 347         } else {
 348                 if (root_lock)
 349                         spin_lock(root_lock);
 350                 return true;
 351         }
 352
 353         if (root_lock)
 354                 spin_lock(root_lock);
 355
 356         if (!dev_xmit_complete(ret)) {
 357                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 358                 if (unlikely(ret != NETDEV_TX_BUSY))
 359                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 360                                              dev->name, ret, q->q.qlen);
 361
 362                 dev_requeue_skb(skb, q);
 363                 return false;
 364         }
 365
 366         return true;
 367 }
 368
 369 /*
 370  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 371  *
 372  * running seqcount guarantees only one CPU can process
 373  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 374  * this queue.
 375  *
 376  *  netif_tx_lock serializes accesses to device driver.
 377  *
 378  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 379  *  if one is grabbed, another must be free.
 380  *
 381  * Note, that this procedure can be called by a watchdog timer
 382  *
 383  * Returns to the caller:
 384  *                              0  - queue is empty or throttled.
 385  *                              >0 - queue is not empty.
 386  *
 387  */
 388 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 389 {
 390         spinlock_t *root_lock = NULL;
 391         struct netdev_queue *txq;
 392         struct net_device *dev;
 393         struct sk_buff *skb;
 394         bool validate;
 395
 396         /* Dequeue packet */
 397         skb = dequeue_skb(q, &validate, packets);
 398         if (unlikely(!skb))
 399                 return false;
 400
 401         if (!(q->flags & TCQ_F_NOLOCK))
 402                 root_lock = qdisc_lock(q);
 403
 404         dev = qdisc_dev(q);
 405         txq = skb_get_tx_queue(dev, skb);
 406
 407         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 408 }
 409
 410 void __qdisc_run(struct Qdisc *q)
 411 {
 412         int quota = READ_ONCE(dev_tx_weight);
 413         int packets;
 414
 415         while (qdisc_restart(q, &packets)) {
 416                 quota -= packets;
 417                 if (quota <= 0) {
 418                         if (q->flags & TCQ_F_NOLOCK)
 419                                 set_bit(__QDISC_STATE_MISSED, &q->state);
 420                         else
 421                                 __netif_schedule(q);
 422
 423                         break;
 424                 }
 425         }
 426 }
 427
 428 unsigned long dev_trans_start(struct net_device *dev)
 429 {
 430         unsigned long val, res;
 431         unsigned int i;
 432
 433         if (is_vlan_dev(dev))
 434                 dev = vlan_dev_real_dev(dev);
 435         else if (netif_is_macvlan(dev))
 436                 dev = macvlan_dev_real_dev(dev);
 437         res = netdev_get_tx_queue(dev, 0)->trans_start;
 438         for (i = 1; i < dev->num_tx_queues; i++) {
 439                 val = netdev_get_tx_queue(dev, i)->trans_start;
 440                 if (val && time_after(val, res))
 441                         res = val;
 442         }
 443
 444         return res;
 445 }
 446 EXPORT_SYMBOL(dev_trans_start);
 447
 448 static void dev_watchdog(struct timer_list *t)
 449 {
 450         struct net_device *dev = from_timer(dev, t, watchdog_timer);
 451
 452         netif_tx_lock(dev);
 453         if (!qdisc_tx_is_noop(dev)) {
 454                 if (netif_device_present(dev) &&
 455                     netif_running(dev) &&
 456                     netif_carrier_ok(dev)) {
 457                         int some_queue_timedout = 0;
 458                         unsigned int i;
 459                         unsigned long trans_start;
 460
 461                         for (i = 0; i < dev->num_tx_queues; i++) {
 462                                 struct netdev_queue *txq;
 463
 464                                 txq = netdev_get_tx_queue(dev, i);
 465                                 trans_start = txq->trans_start;
 466                                 if (netif_xmit_stopped(txq) &&
 467                                     time_after(jiffies, (trans_start +
 468                                                          dev->watchdog_timeo))) {
 469                                         some_queue_timedout = 1;
 470                                         txq->trans_timeout++;
 471                                         break;
 472                                 }
 473                         }
 474
 475                         if (some_queue_timedout) {
 476                                 trace_net_dev_xmit_timeout(dev, i);
 477                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 478                                        dev->name, netdev_drivername(dev), i);
 479                                 dev->netdev_ops->ndo_tx_timeout(dev, i);
 480                         }
 481                         if (!mod_timer(&dev->watchdog_timer,
 482                                        round_jiffies(jiffies +
 483                                                      dev->watchdog_timeo)))
 484                                 dev_hold(dev);
 485                 }
 486         }
 487         netif_tx_unlock(dev);
 488
 489         dev_put(dev);
 490 }
 491
 492 void __netdev_watchdog_up(struct net_device *dev)
 493 {
 494         if (dev->netdev_ops->ndo_tx_timeout) {
 495                 if (dev->watchdog_timeo <= 0)
 496                         dev->watchdog_timeo = 5*HZ;
 497                 if (!mod_timer(&dev->watchdog_timer,
 498                                round_jiffies(jiffies + dev->watchdog_timeo)))
 499                         dev_hold(dev);
 500         }
 501 }
 502 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 503
 504 static void dev_watchdog_up(struct net_device *dev)
 505 {
 506         __netdev_watchdog_up(dev);
 507 }
 508
 509 static void dev_watchdog_down(struct net_device *dev)
 510 {
 511         netif_tx_lock_bh(dev);
 512         if (del_timer(&dev->watchdog_timer))
 513                 dev_put(dev);
 514         netif_tx_unlock_bh(dev);
 515 }
 516
 517 /**
 518  *      netif_carrier_on - set carrier
 519  *      @dev: network device
 520  *
 521  * Device has detected acquisition of carrier.
 522  */
 523 void netif_carrier_on(struct net_device *dev)
 524 {
 525         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 526                 if (dev->reg_state == NETREG_UNINITIALIZED)
 527                         return;
 528                 atomic_inc(&dev->carrier_up_count);
 529                 linkwatch_fire_event(dev);
 530                 if (netif_running(dev))
 531                         __netdev_watchdog_up(dev);
 532         }
 533 }
 534 EXPORT_SYMBOL(netif_carrier_on);
 535
 536 /**
 537  *      netif_carrier_off - clear carrier
 538  *      @dev: network device
 539  *
 540  * Device has detected loss of carrier.
 541  */
 542 void netif_carrier_off(struct net_device *dev)
 543 {
 544         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 545                 if (dev->reg_state == NETREG_UNINITIALIZED)
 546                         return;
 547                 atomic_inc(&dev->carrier_down_count);
 548                 linkwatch_fire_event(dev);
 549         }
 550 }
 551 EXPORT_SYMBOL(netif_carrier_off);
 552
 553 /**
 554  *      netif_carrier_event - report carrier state event
 555  *      @dev: network device
 556  *
 557  * Device has detected a carrier event but the carrier state wasn't changed.
 558  * Use in drivers when querying carrier state asynchronously, to avoid missing
 559  * events (link flaps) if link recovers before it's queried.
 560  */
 561 void netif_carrier_event(struct net_device *dev)
 562 {
 563         if (dev->reg_state == NETREG_UNINITIALIZED)
 564                 return;
 565         atomic_inc(&dev->carrier_up_count);
 566         atomic_inc(&dev->carrier_down_count);
 567         linkwatch_fire_event(dev);
 568 }
 569 EXPORT_SYMBOL_GPL(netif_carrier_event);
 570
 571 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 572    under all circumstances. It is difficult to invent anything faster or
 573    cheaper.
 574  */
 575
 576 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 577                         struct sk_buff **to_free)
 578 {
 579         __qdisc_drop(skb, to_free);
 580         return NET_XMIT_CN;
 581 }
 582
 583 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 584 {
 585         return NULL;
 586 }
 587
 588 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 589         .id             =       "noop",
 590         .priv_size      =       0,
 591         .enqueue        =       noop_enqueue,
 592         .dequeue        =       noop_dequeue,
 593         .peek           =       noop_dequeue,
 594         .owner          =       THIS_MODULE,
 595 };
 596
 597 static struct netdev_queue noop_netdev_queue = {
 598         RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 599         .qdisc_sleeping =       &noop_qdisc,
 600 };
 601
 602 struct Qdisc noop_qdisc = {
 603         .enqueue        =       noop_enqueue,
 604         .dequeue        =       noop_dequeue,
 605         .flags          =       TCQ_F_BUILTIN,
 606         .ops            =       &noop_qdisc_ops,
 607         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 608         .dev_queue      =       &noop_netdev_queue,
 609         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 610         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 611         .gso_skb = {
 612                 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
 613                 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 614                 .qlen = 0,
 615                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 616         },
 617         .skb_bad_txq = {
 618                 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 619                 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 620                 .qlen = 0,
 621                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 622         },
 623 };
 624 EXPORT_SYMBOL(noop_qdisc);
 625
 626 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 627                         struct netlink_ext_ack *extack)
 628 {
 629         /* register_qdisc() assigns a default of noop_enqueue if unset,
 630          * but __dev_queue_xmit() treats noqueue only as such
 631          * if this is NULL - so clear it here. */
 632         qdisc->enqueue = NULL;
 633         return 0;
 634 }
 635
 636 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 637         .id             =       "noqueue",
 638         .priv_size      =       0,
 639         .init           =       noqueue_init,
 640         .enqueue        =       noop_enqueue,
 641         .dequeue        =       noop_dequeue,
 642         .peek           =       noop_dequeue,
 643         .owner          =       THIS_MODULE,
 644 };
 645
 646 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 647         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 648 };
 649
 650 /* 3-band FIFO queue: old style, but should be a bit faster than
 651    generic prio+fifo combination.
 652  */
 653
 654 #define PFIFO_FAST_BANDS 3
 655
 656 /*
 657  * Private data for a pfifo_fast scheduler containing:
 658  *      - rings for priority bands
 659  */
 660 struct pfifo_fast_priv {
 661         struct skb_array q[PFIFO_FAST_BANDS];
 662 };
 663
 664 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 665                                           int band)
 666 {
 667         return &priv->q[band];
 668 }
 669
 670 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 671                               struct sk_buff **to_free)
 672 {
 673         int band = prio2band[skb->priority & TC_PRIO_MAX];
 674         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 675         struct skb_array *q = band2list(priv, band);
 676         unsigned int pkt_len = qdisc_pkt_len(skb);
 677         int err;
 678
 679         err = skb_array_produce(q, skb);
 680
 681         if (unlikely(err)) {
 682                 if (qdisc_is_percpu_stats(qdisc))
 683                         return qdisc_drop_cpu(skb, qdisc, to_free);
 684                 else
 685                         return qdisc_drop(skb, qdisc, to_free);
 686         }
 687
 688         qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 689         return NET_XMIT_SUCCESS;
 690 }
 691
 692 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 693 {
 694         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 695         struct sk_buff *skb = NULL;
 696         bool need_retry = true;
 697         int band;
 698
 699 retry:
 700         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 701                 struct skb_array *q = band2list(priv, band);
 702
 703                 if (__skb_array_empty(q))
 704                         continue;
 705
 706                 skb = __skb_array_consume(q);
 707         }
 708         if (likely(skb)) {
 709                 qdisc_update_stats_at_dequeue(qdisc, skb);
 710         } else if (need_retry &&
 711                    READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
 712                 /* Delay clearing the STATE_MISSED here to reduce
 713                  * the overhead of the second spin_trylock() in
 714                  * qdisc_run_begin() and __netif_schedule() calling
 715                  * in qdisc_run_end().
 716                  */
 717                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
 718                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
 719
 720                 /* Make sure dequeuing happens after clearing
 721                  * STATE_MISSED.
 722                  */
 723                 smp_mb__after_atomic();
 724
 725                 need_retry = false;
 726
 727                 goto retry;
 728         }
 729
 730         return skb;
 731 }
 732
 733 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 734 {
 735         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 736         struct sk_buff *skb = NULL;
 737         int band;
 738
 739         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 740                 struct skb_array *q = band2list(priv, band);
 741
 742                 skb = __skb_array_peek(q);
 743         }
 744
 745         return skb;
 746 }
 747
 748 static void pfifo_fast_reset(struct Qdisc *qdisc)
 749 {
 750         int i, band;
 751         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 752
 753         for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 754                 struct skb_array *q = band2list(priv, band);
 755                 struct sk_buff *skb;
 756
 757                 /* NULL ring is possible if destroy path is due to a failed
 758                  * skb_array_init() in pfifo_fast_init() case.
 759                  */
 760                 if (!q->ring.queue)
 761                         continue;
 762
 763                 while ((skb = __skb_array_consume(q)) != NULL)
 764                         kfree_skb(skb);
 765         }
 766
 767         if (qdisc_is_percpu_stats(qdisc)) {
 768                 for_each_possible_cpu(i) {
 769                         struct gnet_stats_queue *q;
 770
 771                         q = per_cpu_ptr(qdisc->cpu_qstats, i);
 772                         q->backlog = 0;
 773                         q->qlen = 0;
 774                 }
 775         }
 776 }
 777
 778 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 779 {
 780         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 781
 782         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 783         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 784                 goto nla_put_failure;
 785         return skb->len;
 786
 787 nla_put_failure:
 788         return -1;
 789 }
 790
 791 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 792                            struct netlink_ext_ack *extack)
 793 {
 794         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 795         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 796         int prio;
 797
 798         /* guard against zero length rings */
 799         if (!qlen)
 800                 return -EINVAL;
 801
 802         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 803                 struct skb_array *q = band2list(priv, prio);
 804                 int err;
 805
 806                 err = skb_array_init(q, qlen, GFP_KERNEL);
 807                 if (err)
 808                         return -ENOMEM;
 809         }
 810
 811         /* Can by-pass the queue discipline */
 812         qdisc->flags |= TCQ_F_CAN_BYPASS;
 813         return 0;
 814 }
 815
 816 static void pfifo_fast_destroy(struct Qdisc *sch)
 817 {
 818         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 819         int prio;
 820
 821         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 822                 struct skb_array *q = band2list(priv, prio);
 823
 824                 /* NULL ring is possible if destroy path is due to a failed
 825                  * skb_array_init() in pfifo_fast_init() case.
 826                  */
 827                 if (!q->ring.queue)
 828                         continue;
 829                 /* Destroy ring but no need to kfree_skb because a call to
 830                  * pfifo_fast_reset() has already done that work.
 831                  */
 832                 ptr_ring_cleanup(&q->ring, NULL);
 833         }
 834 }
 835
 836 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 837                                           unsigned int new_len)
 838 {
 839         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 840         struct skb_array *bands[PFIFO_FAST_BANDS];
 841         int prio;
 842
 843         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 844                 struct skb_array *q = band2list(priv, prio);
 845
 846                 bands[prio] = q;
 847         }
 848
 849         return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 850                                          GFP_KERNEL);
 851 }
 852
 853 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 854         .id             =       "pfifo_fast",
 855         .priv_size      =       sizeof(struct pfifo_fast_priv),
 856         .enqueue        =       pfifo_fast_enqueue,
 857         .dequeue        =       pfifo_fast_dequeue,
 858         .peek           =       pfifo_fast_peek,
 859         .init           =       pfifo_fast_init,
 860         .destroy        =       pfifo_fast_destroy,
 861         .reset          =       pfifo_fast_reset,
 862         .dump           =       pfifo_fast_dump,
 863         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 864         .owner          =       THIS_MODULE,
 865         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 866 };
 867 EXPORT_SYMBOL(pfifo_fast_ops);
 868
 869 static struct lock_class_key qdisc_tx_busylock;
 870 static struct lock_class_key qdisc_running_key;
 871
 872 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 873                           const struct Qdisc_ops *ops,
 874                           struct netlink_ext_ack *extack)
 875 {
 876         struct Qdisc *sch;
 877         unsigned int size = sizeof(*sch) + ops->priv_size;
 878         int err = -ENOBUFS;
 879         struct net_device *dev;
 880
 881         if (!dev_queue) {
 882                 NL_SET_ERR_MSG(extack, "No device queue given");
 883                 err = -EINVAL;
 884                 goto errout;
 885         }
 886
 887         dev = dev_queue->dev;
 888         sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
 889
 890         if (!sch)
 891                 goto errout;
 892         __skb_queue_head_init(&sch->gso_skb);
 893         __skb_queue_head_init(&sch->skb_bad_txq);
 894         qdisc_skb_head_init(&sch->q);
 895         spin_lock_init(&sch->q.lock);
 896
 897         if (ops->static_flags & TCQ_F_CPUSTATS) {
 898                 sch->cpu_bstats =
 899                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 900                 if (!sch->cpu_bstats)
 901                         goto errout1;
 902
 903                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 904                 if (!sch->cpu_qstats) {
 905                         free_percpu(sch->cpu_bstats);
 906                         goto errout1;
 907                 }
 908         }
 909
 910         spin_lock_init(&sch->busylock);
 911         lockdep_set_class(&sch->busylock,
 912                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 913
 914         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 915         spin_lock_init(&sch->seqlock);
 916         lockdep_set_class(&sch->seqlock,
 917                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 918
 919         seqcount_init(&sch->running);
 920         lockdep_set_class(&sch->running,
 921                           dev->qdisc_running_key ?: &qdisc_running_key);
 922
 923         sch->ops = ops;
 924         sch->flags = ops->static_flags;
 925         sch->enqueue = ops->enqueue;
 926         sch->dequeue = ops->dequeue;
 927         sch->dev_queue = dev_queue;
 928         dev_hold(dev);
 929         refcount_set(&sch->refcnt, 1);
 930
 931         return sch;
 932 errout1:
 933         kfree(sch);
 934 errout:
 935         return ERR_PTR(err);
 936 }
 937
 938 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 939                                 const struct Qdisc_ops *ops,
 940                                 unsigned int parentid,
 941                                 struct netlink_ext_ack *extack)
 942 {
 943         struct Qdisc *sch;
 944
 945         if (!try_module_get(ops->owner)) {
 946                 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 947                 return NULL;
 948         }
 949
 950         sch = qdisc_alloc(dev_queue, ops, extack);
 951         if (IS_ERR(sch)) {
 952                 module_put(ops->owner);
 953                 return NULL;
 954         }
 955         sch->parent = parentid;
 956
 957         if (!ops->init || ops->init(sch, NULL, extack) == 0) {
 958                 trace_qdisc_create(ops, dev_queue->dev, parentid);
 959                 return sch;
 960         }
 961
 962         qdisc_put(sch);
 963         return NULL;
 964 }
 965 EXPORT_SYMBOL(qdisc_create_dflt);
 966
 967 /* Under qdisc_lock(qdisc) and BH! */
 968
 969 void qdisc_reset(struct Qdisc *qdisc)
 970 {
 971         const struct Qdisc_ops *ops = qdisc->ops;
 972         struct sk_buff *skb, *tmp;
 973
 974         trace_qdisc_reset(qdisc);
 975
 976         if (ops->reset)
 977                 ops->reset(qdisc);
 978
 979         skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 980                 __skb_unlink(skb, &qdisc->gso_skb);
 981                 kfree_skb_list(skb);
 982         }
 983
 984         skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 985                 __skb_unlink(skb, &qdisc->skb_bad_txq);
 986                 kfree_skb_list(skb);
 987         }
 988
 989         qdisc->q.qlen = 0;
 990         qdisc->qstats.backlog = 0;
 991 }
 992 EXPORT_SYMBOL(qdisc_reset);
 993
 994 void qdisc_free(struct Qdisc *qdisc)
 995 {
 996         if (qdisc_is_percpu_stats(qdisc)) {
 997                 free_percpu(qdisc->cpu_bstats);
 998                 free_percpu(qdisc->cpu_qstats);
 999         }
1000
1001         kfree(qdisc);
1002 }
1003
1004 static void qdisc_free_cb(struct rcu_head *head)
1005 {
1006         struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1007
1008         qdisc_free(q);
1009 }
1010
1011 static void qdisc_destroy(struct Qdisc *qdisc)
1012 {
1013         const struct Qdisc_ops  *ops = qdisc->ops;
1014
1015 #ifdef CONFIG_NET_SCHED
1016         qdisc_hash_del(qdisc);
1017
1018         qdisc_put_stab(rtnl_dereference(qdisc->stab));
1019 #endif
1020         gen_kill_estimator(&qdisc->rate_est);
1021
1022         qdisc_reset(qdisc);
1023
1024         if (ops->destroy)
1025                 ops->destroy(qdisc);
1026
1027         module_put(ops->owner);
1028         dev_put(qdisc_dev(qdisc));
1029
1030         trace_qdisc_destroy(qdisc);
1031
1032         call_rcu(&qdisc->rcu, qdisc_free_cb);
1033 }
1034
1035 void qdisc_put(struct Qdisc *qdisc)
1036 {
1037         if (!qdisc)
1038                 return;
1039
1040         if (qdisc->flags & TCQ_F_BUILTIN ||
1041             !refcount_dec_and_test(&qdisc->refcnt))
1042                 return;
1043
1044         qdisc_destroy(qdisc);
1045 }
1046 EXPORT_SYMBOL(qdisc_put);
1047
1048 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1049  * Intended to be used as optimization, this function only takes rtnl lock if
1050  * qdisc reference counter reached zero.
1051  */
1052
1053 void qdisc_put_unlocked(struct Qdisc *qdisc)
1054 {
1055         if (qdisc->flags & TCQ_F_BUILTIN ||
1056             !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1057                 return;
1058
1059         qdisc_destroy(qdisc);
1060         rtnl_unlock();
1061 }
1062 EXPORT_SYMBOL(qdisc_put_unlocked);
1063
1064 /* Attach toplevel qdisc to device queue. */
1065 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1066                               struct Qdisc *qdisc)
1067 {
1068         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1069         spinlock_t *root_lock;
1070
1071         root_lock = qdisc_lock(oqdisc);
1072         spin_lock_bh(root_lock);
1073
1074         /* ... and graft new one */
1075         if (qdisc == NULL)
1076                 qdisc = &noop_qdisc;
1077         dev_queue->qdisc_sleeping = qdisc;
1078         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1079
1080         spin_unlock_bh(root_lock);
1081
1082         return oqdisc;
1083 }
1084 EXPORT_SYMBOL(dev_graft_qdisc);
1085
1086 static void shutdown_scheduler_queue(struct net_device *dev,
1087                                      struct netdev_queue *dev_queue,
1088                                      void *_qdisc_default)
1089 {
1090         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1091         struct Qdisc *qdisc_default = _qdisc_default;
1092
1093         if (qdisc) {
1094                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1095                 dev_queue->qdisc_sleeping = qdisc_default;
1096
1097                 qdisc_put(qdisc);
1098         }
1099 }
1100
1101 static void attach_one_default_qdisc(struct net_device *dev,
1102                                      struct netdev_queue *dev_queue,
1103                                      void *_unused)
1104 {
1105         struct Qdisc *qdisc;
1106         const struct Qdisc_ops *ops = default_qdisc_ops;
1107
1108         if (dev->priv_flags & IFF_NO_QUEUE)
1109                 ops = &noqueue_qdisc_ops;
1110         else if(dev->type == ARPHRD_CAN)
1111                 ops = &pfifo_fast_ops;
1112
1113         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1114         if (!qdisc)
1115                 return;
1116
1117         if (!netif_is_multiqueue(dev))
1118                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1119         dev_queue->qdisc_sleeping = qdisc;
1120 }
1121
1122 static void attach_default_qdiscs(struct net_device *dev)
1123 {
1124         struct netdev_queue *txq;
1125         struct Qdisc *qdisc;
1126
1127         txq = netdev_get_tx_queue(dev, 0);
1128
1129         if (!netif_is_multiqueue(dev) ||
1130             dev->priv_flags & IFF_NO_QUEUE) {
1131                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1132                 qdisc = txq->qdisc_sleeping;
1133                 rcu_assign_pointer(dev->qdisc, qdisc);
1134                 qdisc_refcount_inc(qdisc);
1135         } else {
1136                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1137                 if (qdisc) {
1138                         rcu_assign_pointer(dev->qdisc, qdisc);
1139                         qdisc->ops->attach(qdisc);
1140                 }
1141         }
1142         qdisc = rtnl_dereference(dev->qdisc);
1143
1144         /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1145         if (qdisc == &noop_qdisc) {
1146                 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1147                             default_qdisc_ops->id, noqueue_qdisc_ops.id);
1148                 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1149                 dev->priv_flags |= IFF_NO_QUEUE;
1150                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1151                 qdisc = txq->qdisc_sleeping;
1152                 rcu_assign_pointer(dev->qdisc, qdisc);
1153                 qdisc_refcount_inc(qdisc);
1154                 dev->priv_flags ^= IFF_NO_QUEUE;
1155         }
1156
1157 #ifdef CONFIG_NET_SCHED
1158         if (qdisc != &noop_qdisc)
1159                 qdisc_hash_add(qdisc, false);
1160 #endif
1161 }
1162
1163 static void transition_one_qdisc(struct net_device *dev,
1164                                  struct netdev_queue *dev_queue,
1165                                  void *_need_watchdog)
1166 {
1167         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1168         int *need_watchdog_p = _need_watchdog;
1169
1170         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1171                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1172
1173         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1174         if (need_watchdog_p) {
1175                 dev_queue->trans_start = 0;
1176                 *need_watchdog_p = 1;
1177         }
1178 }
1179
1180 void dev_activate(struct net_device *dev)
1181 {
1182         int need_watchdog;
1183
1184         /* No queueing discipline is attached to device;
1185          * create default one for devices, which need queueing
1186          * and noqueue_qdisc for virtual interfaces
1187          */
1188
1189         if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1190                 attach_default_qdiscs(dev);
1191
1192         if (!netif_carrier_ok(dev))
1193                 /* Delay activation until next carrier-on event */
1194                 return;
1195
1196         need_watchdog = 0;
1197         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1198         if (dev_ingress_queue(dev))
1199                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1200
1201         if (need_watchdog) {
1202                 netif_trans_update(dev);
1203                 dev_watchdog_up(dev);
1204         }
1205 }
1206 EXPORT_SYMBOL(dev_activate);
1207
1208 static void qdisc_deactivate(struct Qdisc *qdisc)
1209 {
1210         if (qdisc->flags & TCQ_F_BUILTIN)
1211                 return;
1212
1213         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1214 }
1215
1216 static void dev_deactivate_queue(struct net_device *dev,
1217                                  struct netdev_queue *dev_queue,
1218                                  void *_qdisc_default)
1219 {
1220         struct Qdisc *qdisc_default = _qdisc_default;
1221         struct Qdisc *qdisc;
1222
1223         qdisc = rtnl_dereference(dev_queue->qdisc);
1224         if (qdisc) {
1225                 qdisc_deactivate(qdisc);
1226                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1227         }
1228 }
1229
1230 static void dev_reset_queue(struct net_device *dev,
1231                             struct netdev_queue *dev_queue,
1232                             void *_unused)
1233 {
1234         struct Qdisc *qdisc;
1235         bool nolock;
1236
1237         qdisc = dev_queue->qdisc_sleeping;
1238         if (!qdisc)
1239                 return;
1240
1241         nolock = qdisc->flags & TCQ_F_NOLOCK;
1242
1243         if (nolock)
1244                 spin_lock_bh(&qdisc->seqlock);
1245         spin_lock_bh(qdisc_lock(qdisc));
1246
1247         qdisc_reset(qdisc);
1248
1249         spin_unlock_bh(qdisc_lock(qdisc));
1250         if (nolock) {
1251                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1252                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1253                 spin_unlock_bh(&qdisc->seqlock);
1254         }
1255 }
1256
1257 static bool some_qdisc_is_busy(struct net_device *dev)
1258 {
1259         unsigned int i;
1260
1261         for (i = 0; i < dev->num_tx_queues; i++) {
1262                 struct netdev_queue *dev_queue;
1263                 spinlock_t *root_lock;
1264                 struct Qdisc *q;
1265                 int val;
1266
1267                 dev_queue = netdev_get_tx_queue(dev, i);
1268                 q = dev_queue->qdisc_sleeping;
1269
1270                 root_lock = qdisc_lock(q);
1271                 spin_lock_bh(root_lock);
1272
1273                 val = (qdisc_is_running(q) ||
1274                        test_bit(__QDISC_STATE_SCHED, &q->state));
1275
1276                 spin_unlock_bh(root_lock);
1277
1278                 if (val)
1279                         return true;
1280         }
1281         return false;
1282 }
1283
1284 /**
1285  *      dev_deactivate_many - deactivate transmissions on several devices
1286  *      @head: list of devices to deactivate
1287  *
1288  *      This function returns only when all outstanding transmissions
1289  *      have completed, unless all devices are in dismantle phase.
1290  */
1291 void dev_deactivate_many(struct list_head *head)
1292 {
1293         struct net_device *dev;
1294
1295         list_for_each_entry(dev, head, close_list) {
1296                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1297                                          &noop_qdisc);
1298                 if (dev_ingress_queue(dev))
1299                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
1300                                              &noop_qdisc);
1301
1302                 dev_watchdog_down(dev);
1303         }
1304
1305         /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1306          * outstanding qdisc enqueuing calls.
1307          * This is avoided if all devices are in dismantle phase :
1308          * Caller will call synchronize_net() for us
1309          */
1310         synchronize_net();
1311
1312         list_for_each_entry(dev, head, close_list) {
1313                 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1314
1315                 if (dev_ingress_queue(dev))
1316                         dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1317         }
1318
1319         /* Wait for outstanding qdisc_run calls. */
1320         list_for_each_entry(dev, head, close_list) {
1321                 while (some_qdisc_is_busy(dev)) {
1322                         /* wait_event() would avoid this sleep-loop but would
1323                          * require expensive checks in the fast paths of packet
1324                          * processing which isn't worth it.
1325                          */
1326                         schedule_timeout_uninterruptible(1);
1327                 }
1328         }
1329 }
1330
1331 void dev_deactivate(struct net_device *dev)
1332 {
1333         LIST_HEAD(single);
1334
1335         list_add(&dev->close_list, &single);
1336         dev_deactivate_many(&single);
1337         list_del(&single);
1338 }
1339 EXPORT_SYMBOL(dev_deactivate);
1340
1341 static int qdisc_change_tx_queue_len(struct net_device *dev,
1342                                      struct netdev_queue *dev_queue)
1343 {
1344         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1345         const struct Qdisc_ops *ops = qdisc->ops;
1346
1347         if (ops->change_tx_queue_len)
1348                 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1349         return 0;
1350 }
1351
1352 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1353                                   unsigned int new_real_tx)
1354 {
1355         struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1356
1357         if (qdisc->ops->change_real_num_tx)
1358                 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1359 }
1360
1361 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1362 {
1363         bool up = dev->flags & IFF_UP;
1364         unsigned int i;
1365         int ret = 0;
1366
1367         if (up)
1368                 dev_deactivate(dev);
1369
1370         for (i = 0; i < dev->num_tx_queues; i++) {
1371                 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1372
1373                 /* TODO: revert changes on a partial failure */
1374                 if (ret)
1375                         break;
1376         }
1377
1378         if (up)
1379                 dev_activate(dev);
1380         return ret;
1381 }
1382
1383 static void dev_init_scheduler_queue(struct net_device *dev,
1384                                      struct netdev_queue *dev_queue,
1385                                      void *_qdisc)
1386 {
1387         struct Qdisc *qdisc = _qdisc;
1388
1389         rcu_assign_pointer(dev_queue->qdisc, qdisc);
1390         dev_queue->qdisc_sleeping = qdisc;
1391 }
1392
1393 void dev_init_scheduler(struct net_device *dev)
1394 {
1395         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1396         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1397         if (dev_ingress_queue(dev))
1398                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1399
1400         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1401 }
1402
1403 void dev_shutdown(struct net_device *dev)
1404 {
1405         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1406         if (dev_ingress_queue(dev))
1407                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1408         qdisc_put(rtnl_dereference(dev->qdisc));
1409         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1410
1411         WARN_ON(timer_pending(&dev->watchdog_timer));
1412 }
1413
1414 /**
1415  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1416  * @rate:   Rate to compute reciprocal division values of
1417  * @mult:   Multiplier for reciprocal division
1418  * @shift:  Shift for reciprocal division
1419  *
1420  * The multiplier and shift for reciprocal division by rate are stored
1421  * in mult and shift.
1422  *
1423  * The deal here is to replace a divide by a reciprocal one
1424  * in fast path (a reciprocal divide is a multiply and a shift)
1425  *
1426  * Normal formula would be :
1427  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1428  *
1429  * We compute mult/shift to use instead :
1430  *  time_in_ns = (len * mult) >> shift;
1431  *
1432  * We try to get the highest possible mult value for accuracy,
1433  * but have to make sure no overflows will ever happen.
1434  *
1435  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1436  */
1437 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1438 {
1439         u64 factor = NSEC_PER_SEC;
1440
1441         *mult = 1;
1442         *shift = 0;
1443
1444         if (rate <= 0)
1445                 return;
1446
1447         for (;;) {
1448                 *mult = div64_u64(factor, rate);
1449                 if (*mult & (1U << 31) || factor & (1ULL << 63))
1450                         break;
1451                 factor <<= 1;
1452                 (*shift)++;
1453         }
1454 }
1455
1456 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1457                                const struct tc_ratespec *conf,
1458                                u64 rate64)
1459 {
1460         memset(r, 0, sizeof(*r));
1461         r->overhead = conf->overhead;
1462         r->mpu = conf->mpu;
1463         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1464         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1465         psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1466 }
1467 EXPORT_SYMBOL(psched_ratecfg_precompute);
1468
1469 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1470 {
1471         r->rate_pkts_ps = pktrate64;
1472         psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1473 }
1474 EXPORT_SYMBOL(psched_ppscfg_precompute);
1475
1476 static void mini_qdisc_rcu_func(struct rcu_head *head)
1477 {
1478 }
1479
1480 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1481                           struct tcf_proto *tp_head)
1482 {
1483         /* Protected with chain0->filter_chain_lock.
1484          * Can't access chain directly because tp_head can be NULL.
1485          */
1486         struct mini_Qdisc *miniq_old =
1487                 rcu_dereference_protected(*miniqp->p_miniq, 1);
1488         struct mini_Qdisc *miniq;
1489
1490         if (!tp_head) {
1491                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1492                 /* Wait for flying RCU callback before it is freed. */
1493                 rcu_barrier();
1494                 return;
1495         }
1496
1497         miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1498                 &miniqp->miniq1 : &miniqp->miniq2;
1499
1500         /* We need to make sure that readers won't see the miniq
1501          * we are about to modify. So wait until previous call_rcu callback
1502          * is done.
1503          */
1504         rcu_barrier();
1505         miniq->filter_list = tp_head;
1506         rcu_assign_pointer(*miniqp->p_miniq, miniq);
1507
1508         if (miniq_old)
1509                 /* This is counterpart of the rcu barriers above. We need to
1510                  * block potential new user of miniq_old until all readers
1511                  * are not seeing it.
1512                  */
1513                 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1514 }
1515 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1516
1517 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1518                                 struct tcf_block *block)
1519 {
1520         miniqp->miniq1.block = block;
1521         miniqp->miniq2.block = block;
1522 }
1523 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1524
1525 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1526                           struct mini_Qdisc __rcu **p_miniq)
1527 {
1528         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1529         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1530         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1531         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1532         miniqp->p_miniq = p_miniq;
1533 }
1534 EXPORT_SYMBOL(mini_qdisc_pair_init);