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