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, <[email protected]>
   6  *              Jamal Hadi Salim, <[email protected]> 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 <net/hotdata.h>
  31 #include <trace/events/qdisc.h>
  32 #include <trace/events/net.h>
  33 #include <net/xfrm.h>
  34
  35 /* Qdisc to use by default */
  36 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  37 EXPORT_SYMBOL(default_qdisc_ops);
  38
  39 static void qdisc_maybe_clear_missed(struct Qdisc *q,
  40                                      const struct netdev_queue *txq)
  41 {
  42         clear_bit(__QDISC_STATE_MISSED, &q->state);
  43
  44         /* Make sure the below netif_xmit_frozen_or_stopped()
  45          * checking happens after clearing STATE_MISSED.
  46          */
  47         smp_mb__after_atomic();
  48
  49         /* Checking netif_xmit_frozen_or_stopped() again to
  50          * make sure STATE_MISSED is set if the STATE_MISSED
  51          * set by netif_tx_wake_queue()'s rescheduling of
  52          * net_tx_action() is cleared by the above clear_bit().
  53          */
  54         if (!netif_xmit_frozen_or_stopped(txq))
  55                 set_bit(__QDISC_STATE_MISSED, &q->state);
  56         else
  57                 set_bit(__QDISC_STATE_DRAINING, &q->state);
  58 }
  59
  60 /* Main transmission queue. */
  61
  62 /* Modifications to data participating in scheduling must be protected with
  63  * qdisc_lock(qdisc) spinlock.
  64  *
  65  * The idea is the following:
  66  * - enqueue, dequeue are serialized via qdisc root lock
  67  * - ingress filtering is also serialized via qdisc root lock
  68  * - updates to tree and tree walking are only done under the rtnl mutex.
  69  */
  70
  71 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  72
  73 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  74 {
  75         const struct netdev_queue *txq = q->dev_queue;
  76         spinlock_t *lock = NULL;
  77         struct sk_buff *skb;
  78
  79         if (q->flags & TCQ_F_NOLOCK) {
  80                 lock = qdisc_lock(q);
  81                 spin_lock(lock);
  82         }
  83
  84         skb = skb_peek(&q->skb_bad_txq);
  85         if (skb) {
  86                 /* check the reason of requeuing without tx lock first */
  87                 txq = skb_get_tx_queue(txq->dev, skb);
  88                 if (!netif_xmit_frozen_or_stopped(txq)) {
  89                         skb = __skb_dequeue(&q->skb_bad_txq);
  90                         if (qdisc_is_percpu_stats(q)) {
  91                                 qdisc_qstats_cpu_backlog_dec(q, skb);
  92                                 qdisc_qstats_cpu_qlen_dec(q);
  93                         } else {
  94                                 qdisc_qstats_backlog_dec(q, skb);
  95                                 q->q.qlen--;
  96                         }
  97                 } else {
  98                         skb = SKB_XOFF_MAGIC;
  99                         qdisc_maybe_clear_missed(q, txq);
 100                 }
 101         }
 102
 103         if (lock)
 104                 spin_unlock(lock);
 105
 106         return skb;
 107 }
 108
 109 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
 110 {
 111         struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
 112
 113         if (unlikely(skb))
 114                 skb = __skb_dequeue_bad_txq(q);
 115
 116         return skb;
 117 }
 118
 119 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
 120                                              struct sk_buff *skb)
 121 {
 122         spinlock_t *lock = NULL;
 123
 124         if (q->flags & TCQ_F_NOLOCK) {
 125                 lock = qdisc_lock(q);
 126                 spin_lock(lock);
 127         }
 128
 129         __skb_queue_tail(&q->skb_bad_txq, skb);
 130
 131         if (qdisc_is_percpu_stats(q)) {
 132                 qdisc_qstats_cpu_backlog_inc(q, skb);
 133                 qdisc_qstats_cpu_qlen_inc(q);
 134         } else {
 135                 qdisc_qstats_backlog_inc(q, skb);
 136                 q->q.qlen++;
 137         }
 138
 139         if (lock)
 140                 spin_unlock(lock);
 141 }
 142
 143 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 144 {
 145         spinlock_t *lock = NULL;
 146
 147         if (q->flags & TCQ_F_NOLOCK) {
 148                 lock = qdisc_lock(q);
 149                 spin_lock(lock);
 150         }
 151
 152         while (skb) {
 153                 struct sk_buff *next = skb->next;
 154
 155                 __skb_queue_tail(&q->gso_skb, skb);
 156
 157                 /* it's still part of the queue */
 158                 if (qdisc_is_percpu_stats(q)) {
 159                         qdisc_qstats_cpu_requeues_inc(q);
 160                         qdisc_qstats_cpu_backlog_inc(q, skb);
 161                         qdisc_qstats_cpu_qlen_inc(q);
 162                 } else {
 163                         q->qstats.requeues++;
 164                         qdisc_qstats_backlog_inc(q, skb);
 165                         q->q.qlen++;
 166                 }
 167
 168                 skb = next;
 169         }
 170
 171         if (lock) {
 172                 spin_unlock(lock);
 173                 set_bit(__QDISC_STATE_MISSED, &q->state);
 174         } else {
 175                 __netif_schedule(q);
 176         }
 177 }
 178
 179 static void try_bulk_dequeue_skb(struct Qdisc *q,
 180                                  struct sk_buff *skb,
 181                                  const struct netdev_queue *txq,
 182                                  int *packets)
 183 {
 184         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 185
 186         while (bytelimit > 0) {
 187                 struct sk_buff *nskb = q->dequeue(q);
 188
 189                 if (!nskb)
 190                         break;
 191
 192                 bytelimit -= nskb->len; /* covers GSO len */
 193                 skb->next = nskb;
 194                 skb = nskb;
 195                 (*packets)++; /* GSO counts as one pkt */
 196         }
 197         skb_mark_not_on_list(skb);
 198 }
 199
 200 /* This variant of try_bulk_dequeue_skb() makes sure
 201  * all skbs in the chain are for the same txq
 202  */
 203 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 204                                       struct sk_buff *skb,
 205                                       int *packets)
 206 {
 207         int mapping = skb_get_queue_mapping(skb);
 208         struct sk_buff *nskb;
 209         int cnt = 0;
 210
 211         do {
 212                 nskb = q->dequeue(q);
 213                 if (!nskb)
 214                         break;
 215                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 216                         qdisc_enqueue_skb_bad_txq(q, nskb);
 217                         break;
 218                 }
 219                 skb->next = nskb;
 220                 skb = nskb;
 221         } while (++cnt < 8);
 222         (*packets) += cnt;
 223         skb_mark_not_on_list(skb);
 224 }
 225
 226 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 227  * A requeued skb (via q->gso_skb) can also be a SKB list.
 228  */
 229 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 230                                    int *packets)
 231 {
 232         const struct netdev_queue *txq = q->dev_queue;
 233         struct sk_buff *skb = NULL;
 234
 235         *packets = 1;
 236         if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 237                 spinlock_t *lock = NULL;
 238
 239                 if (q->flags & TCQ_F_NOLOCK) {
 240                         lock = qdisc_lock(q);
 241                         spin_lock(lock);
 242                 }
 243
 244                 skb = skb_peek(&q->gso_skb);
 245
 246                 /* skb may be null if another cpu pulls gso_skb off in between
 247                  * empty check and lock.
 248                  */
 249                 if (!skb) {
 250                         if (lock)
 251                                 spin_unlock(lock);
 252                         goto validate;
 253                 }
 254
 255                 /* skb in gso_skb were already validated */
 256                 *validate = false;
 257                 if (xfrm_offload(skb))
 258                         *validate = true;
 259                 /* check the reason of requeuing without tx lock first */
 260                 txq = skb_get_tx_queue(txq->dev, skb);
 261                 if (!netif_xmit_frozen_or_stopped(txq)) {
 262                         skb = __skb_dequeue(&q->gso_skb);
 263                         if (qdisc_is_percpu_stats(q)) {
 264                                 qdisc_qstats_cpu_backlog_dec(q, skb);
 265                                 qdisc_qstats_cpu_qlen_dec(q);
 266                         } else {
 267                                 qdisc_qstats_backlog_dec(q, skb);
 268                                 q->q.qlen--;
 269                         }
 270                 } else {
 271                         skb = NULL;
 272                         qdisc_maybe_clear_missed(q, txq);
 273                 }
 274                 if (lock)
 275                         spin_unlock(lock);
 276                 goto trace;
 277         }
 278 validate:
 279         *validate = true;
 280
 281         if ((q->flags & TCQ_F_ONETXQUEUE) &&
 282             netif_xmit_frozen_or_stopped(txq)) {
 283                 qdisc_maybe_clear_missed(q, txq);
 284                 return skb;
 285         }
 286
 287         skb = qdisc_dequeue_skb_bad_txq(q);
 288         if (unlikely(skb)) {
 289                 if (skb == SKB_XOFF_MAGIC)
 290                         return NULL;
 291                 goto bulk;
 292         }
 293         skb = q->dequeue(q);
 294         if (skb) {
 295 bulk:
 296                 if (qdisc_may_bulk(q))
 297                         try_bulk_dequeue_skb(q, skb, txq, packets);
 298                 else
 299                         try_bulk_dequeue_skb_slow(q, skb, packets);
 300         }
 301 trace:
 302         trace_qdisc_dequeue(q, txq, *packets, skb);
 303         return skb;
 304 }
 305
 306 /*
 307  * Transmit possibly several skbs, and handle the return status as
 308  * required. Owning qdisc running bit guarantees that only one CPU
 309  * can execute this function.
 310  *
 311  * Returns to the caller:
 312  *                              false  - hardware queue frozen backoff
 313  *                              true   - feel free to send more pkts
 314  */
 315 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 316                      struct net_device *dev, struct netdev_queue *txq,
 317                      spinlock_t *root_lock, bool validate)
 318 {
 319         int ret = NETDEV_TX_BUSY;
 320         bool again = false;
 321
 322         /* And release qdisc */
 323         if (root_lock)
 324                 spin_unlock(root_lock);
 325
 326         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 327         if (validate)
 328                 skb = validate_xmit_skb_list(skb, dev, &again);
 329
 330 #ifdef CONFIG_XFRM_OFFLOAD
 331         if (unlikely(again)) {
 332                 if (root_lock)
 333                         spin_lock(root_lock);
 334
 335                 dev_requeue_skb(skb, q);
 336                 return false;
 337         }
 338 #endif
 339
 340         if (likely(skb)) {
 341                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 342                 if (!netif_xmit_frozen_or_stopped(txq))
 343                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 344                 else
 345                         qdisc_maybe_clear_missed(q, txq);
 346
 347                 HARD_TX_UNLOCK(dev, txq);
 348         } else {
 349                 if (root_lock)
 350                         spin_lock(root_lock);
 351                 return true;
 352         }
 353
 354         if (root_lock)
 355                 spin_lock(root_lock);
 356
 357         if (!dev_xmit_complete(ret)) {
 358                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 359                 if (unlikely(ret != NETDEV_TX_BUSY))
 360                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 361                                              dev->name, ret, q->q.qlen);
 362
 363                 dev_requeue_skb(skb, q);
 364                 return false;
 365         }
 366
 367         return true;
 368 }
 369
 370 /*
 371  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 372  *
 373  * running seqcount guarantees only one CPU can process
 374  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 375  * this queue.
 376  *
 377  *  netif_tx_lock serializes accesses to device driver.
 378  *
 379  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 380  *  if one is grabbed, another must be free.
 381  *
 382  * Note, that this procedure can be called by a watchdog timer
 383  *
 384  * Returns to the caller:
 385  *                              0  - queue is empty or throttled.
 386  *                              >0 - queue is not empty.
 387  *
 388  */
 389 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 390 {
 391         spinlock_t *root_lock = NULL;
 392         struct netdev_queue *txq;
 393         struct net_device *dev;
 394         struct sk_buff *skb;
 395         bool validate;
 396
 397         /* Dequeue packet */
 398         skb = dequeue_skb(q, &validate, packets);
 399         if (unlikely(!skb))
 400                 return false;
 401
 402         if (!(q->flags & TCQ_F_NOLOCK))
 403                 root_lock = qdisc_lock(q);
 404
 405         dev = qdisc_dev(q);
 406         txq = skb_get_tx_queue(dev, skb);
 407
 408         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 409 }
 410
 411 void __qdisc_run(struct Qdisc *q)
 412 {
 413         int quota = READ_ONCE(net_hotdata.dev_tx_weight);
 414         int packets;
 415
 416         while (qdisc_restart(q, &packets)) {
 417                 quota -= packets;
 418                 if (quota <= 0) {
 419                         if (q->flags & TCQ_F_NOLOCK)
 420                                 set_bit(__QDISC_STATE_MISSED, &q->state);
 421                         else
 422                                 __netif_schedule(q);
 423
 424                         break;
 425                 }
 426         }
 427 }
 428
 429 unsigned long dev_trans_start(struct net_device *dev)
 430 {
 431         unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
 432         unsigned long val;
 433         unsigned int i;
 434
 435         for (i = 1; i < dev->num_tx_queues; i++) {
 436                 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
 437                 if (val && time_after(val, res))
 438                         res = val;
 439         }
 440
 441         return res;
 442 }
 443 EXPORT_SYMBOL(dev_trans_start);
 444
 445 static void netif_freeze_queues(struct net_device *dev)
 446 {
 447         unsigned int i;
 448         int cpu;
 449
 450         cpu = smp_processor_id();
 451         for (i = 0; i < dev->num_tx_queues; i++) {
 452                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 453
 454                 /* We are the only thread of execution doing a
 455                  * freeze, but we have to grab the _xmit_lock in
 456                  * order to synchronize with threads which are in
 457                  * the ->hard_start_xmit() handler and already
 458                  * checked the frozen bit.
 459                  */
 460                 __netif_tx_lock(txq, cpu);
 461                 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
 462                 __netif_tx_unlock(txq);
 463         }
 464 }
 465
 466 void netif_tx_lock(struct net_device *dev)
 467 {
 468         spin_lock(&dev->tx_global_lock);
 469         netif_freeze_queues(dev);
 470 }
 471 EXPORT_SYMBOL(netif_tx_lock);
 472
 473 static void netif_unfreeze_queues(struct net_device *dev)
 474 {
 475         unsigned int i;
 476
 477         for (i = 0; i < dev->num_tx_queues; i++) {
 478                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 479
 480                 /* No need to grab the _xmit_lock here.  If the
 481                  * queue is not stopped for another reason, we
 482                  * force a schedule.
 483                  */
 484                 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
 485                 netif_schedule_queue(txq);
 486         }
 487 }
 488
 489 void netif_tx_unlock(struct net_device *dev)
 490 {
 491         netif_unfreeze_queues(dev);
 492         spin_unlock(&dev->tx_global_lock);
 493 }
 494 EXPORT_SYMBOL(netif_tx_unlock);
 495
 496 static void dev_watchdog(struct timer_list *t)
 497 {
 498         struct net_device *dev = from_timer(dev, t, watchdog_timer);
 499         bool release = true;
 500
 501         spin_lock(&dev->tx_global_lock);
 502         if (!qdisc_tx_is_noop(dev)) {
 503                 if (netif_device_present(dev) &&
 504                     netif_running(dev) &&
 505                     netif_carrier_ok(dev)) {
 506                         unsigned int timedout_ms = 0;
 507                         unsigned int i;
 508                         unsigned long trans_start;
 509
 510                         for (i = 0; i < dev->num_tx_queues; i++) {
 511                                 struct netdev_queue *txq;
 512
 513                                 txq = netdev_get_tx_queue(dev, i);
 514                                 trans_start = READ_ONCE(txq->trans_start);
 515                                 if (netif_xmit_stopped(txq) &&
 516                                     time_after(jiffies, (trans_start +
 517                                                          dev->watchdog_timeo))) {
 518                                         timedout_ms = jiffies_to_msecs(jiffies - trans_start);
 519                                         atomic_long_inc(&txq->trans_timeout);
 520                                         break;
 521                                 }
 522                         }
 523
 524                         if (unlikely(timedout_ms)) {
 525                                 trace_net_dev_xmit_timeout(dev, i);
 526                                 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
 527                                             raw_smp_processor_id(),
 528                                             i, timedout_ms);
 529                                 netif_freeze_queues(dev);
 530                                 dev->netdev_ops->ndo_tx_timeout(dev, i);
 531                                 netif_unfreeze_queues(dev);
 532                         }
 533                         if (!mod_timer(&dev->watchdog_timer,
 534                                        round_jiffies(jiffies +
 535                                                      dev->watchdog_timeo)))
 536                                 release = false;
 537                 }
 538         }
 539         spin_unlock(&dev->tx_global_lock);
 540
 541         if (release)
 542                 netdev_put(dev, &dev->watchdog_dev_tracker);
 543 }
 544
 545 void __netdev_watchdog_up(struct net_device *dev)
 546 {
 547         if (dev->netdev_ops->ndo_tx_timeout) {
 548                 if (dev->watchdog_timeo <= 0)
 549                         dev->watchdog_timeo = 5*HZ;
 550                 if (!mod_timer(&dev->watchdog_timer,
 551                                round_jiffies(jiffies + dev->watchdog_timeo)))
 552                         netdev_hold(dev, &dev->watchdog_dev_tracker,
 553                                     GFP_ATOMIC);
 554         }
 555 }
 556 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 557
 558 static void dev_watchdog_up(struct net_device *dev)
 559 {
 560         __netdev_watchdog_up(dev);
 561 }
 562
 563 static void dev_watchdog_down(struct net_device *dev)
 564 {
 565         netif_tx_lock_bh(dev);
 566         if (del_timer(&dev->watchdog_timer))
 567                 netdev_put(dev, &dev->watchdog_dev_tracker);
 568         netif_tx_unlock_bh(dev);
 569 }
 570
 571 /**
 572  *      netif_carrier_on - set carrier
 573  *      @dev: network device
 574  *
 575  * Device has detected acquisition of carrier.
 576  */
 577 void netif_carrier_on(struct net_device *dev)
 578 {
 579         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 580                 if (dev->reg_state == NETREG_UNINITIALIZED)
 581                         return;
 582                 atomic_inc(&dev->carrier_up_count);
 583                 linkwatch_fire_event(dev);
 584                 if (netif_running(dev))
 585                         __netdev_watchdog_up(dev);
 586         }
 587 }
 588 EXPORT_SYMBOL(netif_carrier_on);
 589
 590 /**
 591  *      netif_carrier_off - clear carrier
 592  *      @dev: network device
 593  *
 594  * Device has detected loss of carrier.
 595  */
 596 void netif_carrier_off(struct net_device *dev)
 597 {
 598         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 599                 if (dev->reg_state == NETREG_UNINITIALIZED)
 600                         return;
 601                 atomic_inc(&dev->carrier_down_count);
 602                 linkwatch_fire_event(dev);
 603         }
 604 }
 605 EXPORT_SYMBOL(netif_carrier_off);
 606
 607 /**
 608  *      netif_carrier_event - report carrier state event
 609  *      @dev: network device
 610  *
 611  * Device has detected a carrier event but the carrier state wasn't changed.
 612  * Use in drivers when querying carrier state asynchronously, to avoid missing
 613  * events (link flaps) if link recovers before it's queried.
 614  */
 615 void netif_carrier_event(struct net_device *dev)
 616 {
 617         if (dev->reg_state == NETREG_UNINITIALIZED)
 618                 return;
 619         atomic_inc(&dev->carrier_up_count);
 620         atomic_inc(&dev->carrier_down_count);
 621         linkwatch_fire_event(dev);
 622 }
 623 EXPORT_SYMBOL_GPL(netif_carrier_event);
 624
 625 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 626    under all circumstances. It is difficult to invent anything faster or
 627    cheaper.
 628  */
 629
 630 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 631                         struct sk_buff **to_free)
 632 {
 633         __qdisc_drop(skb, to_free);
 634         return NET_XMIT_CN;
 635 }
 636
 637 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 638 {
 639         return NULL;
 640 }
 641
 642 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 643         .id             =       "noop",
 644         .priv_size      =       0,
 645         .enqueue        =       noop_enqueue,
 646         .dequeue        =       noop_dequeue,
 647         .peek           =       noop_dequeue,
 648         .owner          =       THIS_MODULE,
 649 };
 650
 651 static struct netdev_queue noop_netdev_queue = {
 652         RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 653         RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
 654 };
 655
 656 struct Qdisc noop_qdisc = {
 657         .enqueue        =       noop_enqueue,
 658         .dequeue        =       noop_dequeue,
 659         .flags          =       TCQ_F_BUILTIN,
 660         .ops            =       &noop_qdisc_ops,
 661         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 662         .dev_queue      =       &noop_netdev_queue,
 663         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 664         .gso_skb = {
 665                 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
 666                 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 667                 .qlen = 0,
 668                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 669         },
 670         .skb_bad_txq = {
 671                 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 672                 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 673                 .qlen = 0,
 674                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 675         },
 676 };
 677 EXPORT_SYMBOL(noop_qdisc);
 678
 679 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 680                         struct netlink_ext_ack *extack)
 681 {
 682         /* register_qdisc() assigns a default of noop_enqueue if unset,
 683          * but __dev_queue_xmit() treats noqueue only as such
 684          * if this is NULL - so clear it here. */
 685         qdisc->enqueue = NULL;
 686         return 0;
 687 }
 688
 689 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 690         .id             =       "noqueue",
 691         .priv_size      =       0,
 692         .init           =       noqueue_init,
 693         .enqueue        =       noop_enqueue,
 694         .dequeue        =       noop_dequeue,
 695         .peek           =       noop_dequeue,
 696         .owner          =       THIS_MODULE,
 697 };
 698
 699 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
 700         1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
 701 };
 702 EXPORT_SYMBOL(sch_default_prio2band);
 703
 704 /* 3-band FIFO queue: old style, but should be a bit faster than
 705    generic prio+fifo combination.
 706  */
 707
 708 #define PFIFO_FAST_BANDS 3
 709
 710 /*
 711  * Private data for a pfifo_fast scheduler containing:
 712  *      - rings for priority bands
 713  */
 714 struct pfifo_fast_priv {
 715         struct skb_array q[PFIFO_FAST_BANDS];
 716 };
 717
 718 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 719                                           int band)
 720 {
 721         return &priv->q[band];
 722 }
 723
 724 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 725                               struct sk_buff **to_free)
 726 {
 727         int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
 728         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 729         struct skb_array *q = band2list(priv, band);
 730         unsigned int pkt_len = qdisc_pkt_len(skb);
 731         int err;
 732
 733         err = skb_array_produce(q, skb);
 734
 735         if (unlikely(err)) {
 736                 if (qdisc_is_percpu_stats(qdisc))
 737                         return qdisc_drop_cpu(skb, qdisc, to_free);
 738                 else
 739                         return qdisc_drop(skb, qdisc, to_free);
 740         }
 741
 742         qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 743         return NET_XMIT_SUCCESS;
 744 }
 745
 746 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 747 {
 748         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 749         struct sk_buff *skb = NULL;
 750         bool need_retry = true;
 751         int band;
 752
 753 retry:
 754         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 755                 struct skb_array *q = band2list(priv, band);
 756
 757                 if (__skb_array_empty(q))
 758                         continue;
 759
 760                 skb = __skb_array_consume(q);
 761         }
 762         if (likely(skb)) {
 763                 qdisc_update_stats_at_dequeue(qdisc, skb);
 764         } else if (need_retry &&
 765                    READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
 766                 /* Delay clearing the STATE_MISSED here to reduce
 767                  * the overhead of the second spin_trylock() in
 768                  * qdisc_run_begin() and __netif_schedule() calling
 769                  * in qdisc_run_end().
 770                  */
 771                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
 772                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
 773
 774                 /* Make sure dequeuing happens after clearing
 775                  * STATE_MISSED.
 776                  */
 777                 smp_mb__after_atomic();
 778
 779                 need_retry = false;
 780
 781                 goto retry;
 782         }
 783
 784         return skb;
 785 }
 786
 787 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 788 {
 789         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 790         struct sk_buff *skb = NULL;
 791         int band;
 792
 793         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 794                 struct skb_array *q = band2list(priv, band);
 795
 796                 skb = __skb_array_peek(q);
 797         }
 798
 799         return skb;
 800 }
 801
 802 static void pfifo_fast_reset(struct Qdisc *qdisc)
 803 {
 804         int i, band;
 805         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 806
 807         for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 808                 struct skb_array *q = band2list(priv, band);
 809                 struct sk_buff *skb;
 810
 811                 /* NULL ring is possible if destroy path is due to a failed
 812                  * skb_array_init() in pfifo_fast_init() case.
 813                  */
 814                 if (!q->ring.queue)
 815                         continue;
 816
 817                 while ((skb = __skb_array_consume(q)) != NULL)
 818                         kfree_skb(skb);
 819         }
 820
 821         if (qdisc_is_percpu_stats(qdisc)) {
 822                 for_each_possible_cpu(i) {
 823                         struct gnet_stats_queue *q;
 824
 825                         q = per_cpu_ptr(qdisc->cpu_qstats, i);
 826                         q->backlog = 0;
 827                         q->qlen = 0;
 828                 }
 829         }
 830 }
 831
 832 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 833 {
 834         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 835
 836         memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
 837         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 838                 goto nla_put_failure;
 839         return skb->len;
 840
 841 nla_put_failure:
 842         return -1;
 843 }
 844
 845 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 846                            struct netlink_ext_ack *extack)
 847 {
 848         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 849         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 850         int prio;
 851
 852         /* guard against zero length rings */
 853         if (!qlen)
 854                 return -EINVAL;
 855
 856         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 857                 struct skb_array *q = band2list(priv, prio);
 858                 int err;
 859
 860                 err = skb_array_init(q, qlen, GFP_KERNEL);
 861                 if (err)
 862                         return -ENOMEM;
 863         }
 864
 865         /* Can by-pass the queue discipline */
 866         qdisc->flags |= TCQ_F_CAN_BYPASS;
 867         return 0;
 868 }
 869
 870 static void pfifo_fast_destroy(struct Qdisc *sch)
 871 {
 872         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 873         int prio;
 874
 875         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 876                 struct skb_array *q = band2list(priv, prio);
 877
 878                 /* NULL ring is possible if destroy path is due to a failed
 879                  * skb_array_init() in pfifo_fast_init() case.
 880                  */
 881                 if (!q->ring.queue)
 882                         continue;
 883                 /* Destroy ring but no need to kfree_skb because a call to
 884                  * pfifo_fast_reset() has already done that work.
 885                  */
 886                 ptr_ring_cleanup(&q->ring, NULL);
 887         }
 888 }
 889
 890 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 891                                           unsigned int new_len)
 892 {
 893         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 894         struct skb_array *bands[PFIFO_FAST_BANDS];
 895         int prio;
 896
 897         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 898                 struct skb_array *q = band2list(priv, prio);
 899
 900                 bands[prio] = q;
 901         }
 902
 903         return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 904                                          GFP_KERNEL);
 905 }
 906
 907 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 908         .id             =       "pfifo_fast",
 909         .priv_size      =       sizeof(struct pfifo_fast_priv),
 910         .enqueue        =       pfifo_fast_enqueue,
 911         .dequeue        =       pfifo_fast_dequeue,
 912         .peek           =       pfifo_fast_peek,
 913         .init           =       pfifo_fast_init,
 914         .destroy        =       pfifo_fast_destroy,
 915         .reset          =       pfifo_fast_reset,
 916         .dump           =       pfifo_fast_dump,
 917         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 918         .owner          =       THIS_MODULE,
 919         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 920 };
 921 EXPORT_SYMBOL(pfifo_fast_ops);
 922
 923 static struct lock_class_key qdisc_tx_busylock;
 924
 925 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 926                           const struct Qdisc_ops *ops,
 927                           struct netlink_ext_ack *extack)
 928 {
 929         struct Qdisc *sch;
 930         unsigned int size = sizeof(*sch) + ops->priv_size;
 931         int err = -ENOBUFS;
 932         struct net_device *dev;
 933
 934         if (!dev_queue) {
 935                 NL_SET_ERR_MSG(extack, "No device queue given");
 936                 err = -EINVAL;
 937                 goto errout;
 938         }
 939
 940         dev = dev_queue->dev;
 941         sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
 942
 943         if (!sch)
 944                 goto errout;
 945         __skb_queue_head_init(&sch->gso_skb);
 946         __skb_queue_head_init(&sch->skb_bad_txq);
 947         gnet_stats_basic_sync_init(&sch->bstats);
 948         spin_lock_init(&sch->q.lock);
 949
 950         if (ops->static_flags & TCQ_F_CPUSTATS) {
 951                 sch->cpu_bstats =
 952                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
 953                 if (!sch->cpu_bstats)
 954                         goto errout1;
 955
 956                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 957                 if (!sch->cpu_qstats) {
 958                         free_percpu(sch->cpu_bstats);
 959                         goto errout1;
 960                 }
 961         }
 962
 963         spin_lock_init(&sch->busylock);
 964         lockdep_set_class(&sch->busylock,
 965                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 966
 967         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 968         spin_lock_init(&sch->seqlock);
 969         lockdep_set_class(&sch->seqlock,
 970                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 971
 972         sch->ops = ops;
 973         sch->flags = ops->static_flags;
 974         sch->enqueue = ops->enqueue;
 975         sch->dequeue = ops->dequeue;
 976         sch->dev_queue = dev_queue;
 977         netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
 978         refcount_set(&sch->refcnt, 1);
 979
 980         return sch;
 981 errout1:
 982         kfree(sch);
 983 errout:
 984         return ERR_PTR(err);
 985 }
 986
 987 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 988                                 const struct Qdisc_ops *ops,
 989                                 unsigned int parentid,
 990                                 struct netlink_ext_ack *extack)
 991 {
 992         struct Qdisc *sch;
 993
 994         if (!try_module_get(ops->owner)) {
 995                 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 996                 return NULL;
 997         }
 998
 999         sch = qdisc_alloc(dev_queue, ops, extack);
1000         if (IS_ERR(sch)) {
1001                 module_put(ops->owner);
1002                 return NULL;
1003         }
1004         sch->parent = parentid;
1005
1006         if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1007                 trace_qdisc_create(ops, dev_queue->dev, parentid);
1008                 return sch;
1009         }
1010
1011         qdisc_put(sch);
1012         return NULL;
1013 }
1014 EXPORT_SYMBOL(qdisc_create_dflt);
1015
1016 /* Under qdisc_lock(qdisc) and BH! */
1017
1018 void qdisc_reset(struct Qdisc *qdisc)
1019 {
1020         const struct Qdisc_ops *ops = qdisc->ops;
1021
1022         trace_qdisc_reset(qdisc);
1023
1024         if (ops->reset)
1025                 ops->reset(qdisc);
1026
1027         __skb_queue_purge(&qdisc->gso_skb);
1028         __skb_queue_purge(&qdisc->skb_bad_txq);
1029
1030         qdisc->q.qlen = 0;
1031         qdisc->qstats.backlog = 0;
1032 }
1033 EXPORT_SYMBOL(qdisc_reset);
1034
1035 void qdisc_free(struct Qdisc *qdisc)
1036 {
1037         if (qdisc_is_percpu_stats(qdisc)) {
1038                 free_percpu(qdisc->cpu_bstats);
1039                 free_percpu(qdisc->cpu_qstats);
1040         }
1041
1042         kfree(qdisc);
1043 }
1044
1045 static void qdisc_free_cb(struct rcu_head *head)
1046 {
1047         struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1048
1049         qdisc_free(q);
1050 }
1051
1052 static void __qdisc_destroy(struct Qdisc *qdisc)
1053 {
1054         const struct Qdisc_ops  *ops = qdisc->ops;
1055         struct net_device *dev = qdisc_dev(qdisc);
1056
1057 #ifdef CONFIG_NET_SCHED
1058         qdisc_hash_del(qdisc);
1059
1060         qdisc_put_stab(rtnl_dereference(qdisc->stab));
1061 #endif
1062         gen_kill_estimator(&qdisc->rate_est);
1063
1064         qdisc_reset(qdisc);
1065
1066
1067         if (ops->destroy)
1068                 ops->destroy(qdisc);
1069
1070         module_put(ops->owner);
1071         netdev_put(dev, &qdisc->dev_tracker);
1072
1073         trace_qdisc_destroy(qdisc);
1074
1075         call_rcu(&qdisc->rcu, qdisc_free_cb);
1076 }
1077
1078 void qdisc_destroy(struct Qdisc *qdisc)
1079 {
1080         if (qdisc->flags & TCQ_F_BUILTIN)
1081                 return;
1082
1083         __qdisc_destroy(qdisc);
1084 }
1085
1086 void qdisc_put(struct Qdisc *qdisc)
1087 {
1088         if (!qdisc)
1089                 return;
1090
1091         if (qdisc->flags & TCQ_F_BUILTIN ||
1092             !refcount_dec_and_test(&qdisc->refcnt))
1093                 return;
1094
1095         __qdisc_destroy(qdisc);
1096 }
1097 EXPORT_SYMBOL(qdisc_put);
1098
1099 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1100  * Intended to be used as optimization, this function only takes rtnl lock if
1101  * qdisc reference counter reached zero.
1102  */
1103
1104 void qdisc_put_unlocked(struct Qdisc *qdisc)
1105 {
1106         if (qdisc->flags & TCQ_F_BUILTIN ||
1107             !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1108                 return;
1109
1110         __qdisc_destroy(qdisc);
1111         rtnl_unlock();
1112 }
1113 EXPORT_SYMBOL(qdisc_put_unlocked);
1114
1115 /* Attach toplevel qdisc to device queue. */
1116 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1117                               struct Qdisc *qdisc)
1118 {
1119         struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1120         spinlock_t *root_lock;
1121
1122         root_lock = qdisc_lock(oqdisc);
1123         spin_lock_bh(root_lock);
1124
1125         /* ... and graft new one */
1126         if (qdisc == NULL)
1127                 qdisc = &noop_qdisc;
1128         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1129         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1130
1131         spin_unlock_bh(root_lock);
1132
1133         return oqdisc;
1134 }
1135 EXPORT_SYMBOL(dev_graft_qdisc);
1136
1137 static void shutdown_scheduler_queue(struct net_device *dev,
1138                                      struct netdev_queue *dev_queue,
1139                                      void *_qdisc_default)
1140 {
1141         struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1142         struct Qdisc *qdisc_default = _qdisc_default;
1143
1144         if (qdisc) {
1145                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1146                 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1147
1148                 qdisc_put(qdisc);
1149         }
1150 }
1151
1152 static void attach_one_default_qdisc(struct net_device *dev,
1153                                      struct netdev_queue *dev_queue,
1154                                      void *_unused)
1155 {
1156         struct Qdisc *qdisc;
1157         const struct Qdisc_ops *ops = default_qdisc_ops;
1158
1159         if (dev->priv_flags & IFF_NO_QUEUE)
1160                 ops = &noqueue_qdisc_ops;
1161         else if(dev->type == ARPHRD_CAN)
1162                 ops = &pfifo_fast_ops;
1163
1164         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1165         if (!qdisc)
1166                 return;
1167
1168         if (!netif_is_multiqueue(dev))
1169                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1170         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1171 }
1172
1173 static void attach_default_qdiscs(struct net_device *dev)
1174 {
1175         struct netdev_queue *txq;
1176         struct Qdisc *qdisc;
1177
1178         txq = netdev_get_tx_queue(dev, 0);
1179
1180         if (!netif_is_multiqueue(dev) ||
1181             dev->priv_flags & IFF_NO_QUEUE) {
1182                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1183                 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1184                 rcu_assign_pointer(dev->qdisc, qdisc);
1185                 qdisc_refcount_inc(qdisc);
1186         } else {
1187                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1188                 if (qdisc) {
1189                         rcu_assign_pointer(dev->qdisc, qdisc);
1190                         qdisc->ops->attach(qdisc);
1191                 }
1192         }
1193         qdisc = rtnl_dereference(dev->qdisc);
1194
1195         /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1196         if (qdisc == &noop_qdisc) {
1197                 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1198                             default_qdisc_ops->id, noqueue_qdisc_ops.id);
1199                 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1200                 dev->priv_flags |= IFF_NO_QUEUE;
1201                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1202                 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1203                 rcu_assign_pointer(dev->qdisc, qdisc);
1204                 qdisc_refcount_inc(qdisc);
1205                 dev->priv_flags ^= IFF_NO_QUEUE;
1206         }
1207
1208 #ifdef CONFIG_NET_SCHED
1209         if (qdisc != &noop_qdisc)
1210                 qdisc_hash_add(qdisc, false);
1211 #endif
1212 }
1213
1214 static void transition_one_qdisc(struct net_device *dev,
1215                                  struct netdev_queue *dev_queue,
1216                                  void *_need_watchdog)
1217 {
1218         struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1219         int *need_watchdog_p = _need_watchdog;
1220
1221         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1222                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1223
1224         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1225         if (need_watchdog_p) {
1226                 WRITE_ONCE(dev_queue->trans_start, 0);
1227                 *need_watchdog_p = 1;
1228         }
1229 }
1230
1231 void dev_activate(struct net_device *dev)
1232 {
1233         int need_watchdog;
1234
1235         /* No queueing discipline is attached to device;
1236          * create default one for devices, which need queueing
1237          * and noqueue_qdisc for virtual interfaces
1238          */
1239
1240         if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1241                 attach_default_qdiscs(dev);
1242
1243         if (!netif_carrier_ok(dev))
1244                 /* Delay activation until next carrier-on event */
1245                 return;
1246
1247         need_watchdog = 0;
1248         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1249         if (dev_ingress_queue(dev))
1250                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1251
1252         if (need_watchdog) {
1253                 netif_trans_update(dev);
1254                 dev_watchdog_up(dev);
1255         }
1256 }
1257 EXPORT_SYMBOL(dev_activate);
1258
1259 static void qdisc_deactivate(struct Qdisc *qdisc)
1260 {
1261         if (qdisc->flags & TCQ_F_BUILTIN)
1262                 return;
1263
1264         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1265 }
1266
1267 static void dev_deactivate_queue(struct net_device *dev,
1268                                  struct netdev_queue *dev_queue,
1269                                  void *_qdisc_default)
1270 {
1271         struct Qdisc *qdisc_default = _qdisc_default;
1272         struct Qdisc *qdisc;
1273
1274         qdisc = rtnl_dereference(dev_queue->qdisc);
1275         if (qdisc) {
1276                 qdisc_deactivate(qdisc);
1277                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1278         }
1279 }
1280
1281 static void dev_reset_queue(struct net_device *dev,
1282                             struct netdev_queue *dev_queue,
1283                             void *_unused)
1284 {
1285         struct Qdisc *qdisc;
1286         bool nolock;
1287
1288         qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1289         if (!qdisc)
1290                 return;
1291
1292         nolock = qdisc->flags & TCQ_F_NOLOCK;
1293
1294         if (nolock)
1295                 spin_lock_bh(&qdisc->seqlock);
1296         spin_lock_bh(qdisc_lock(qdisc));
1297
1298         qdisc_reset(qdisc);
1299
1300         spin_unlock_bh(qdisc_lock(qdisc));
1301         if (nolock) {
1302                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1303                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1304                 spin_unlock_bh(&qdisc->seqlock);
1305         }
1306 }
1307
1308 static bool some_qdisc_is_busy(struct net_device *dev)
1309 {
1310         unsigned int i;
1311
1312         for (i = 0; i < dev->num_tx_queues; i++) {
1313                 struct netdev_queue *dev_queue;
1314                 spinlock_t *root_lock;
1315                 struct Qdisc *q;
1316                 int val;
1317
1318                 dev_queue = netdev_get_tx_queue(dev, i);
1319                 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1320
1321                 root_lock = qdisc_lock(q);
1322                 spin_lock_bh(root_lock);
1323
1324                 val = (qdisc_is_running(q) ||
1325                        test_bit(__QDISC_STATE_SCHED, &q->state));
1326
1327                 spin_unlock_bh(root_lock);
1328
1329                 if (val)
1330                         return true;
1331         }
1332         return false;
1333 }
1334
1335 /**
1336  *      dev_deactivate_many - deactivate transmissions on several devices
1337  *      @head: list of devices to deactivate
1338  *
1339  *      This function returns only when all outstanding transmissions
1340  *      have completed, unless all devices are in dismantle phase.
1341  */
1342 void dev_deactivate_many(struct list_head *head)
1343 {
1344         struct net_device *dev;
1345
1346         list_for_each_entry(dev, head, close_list) {
1347                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1348                                          &noop_qdisc);
1349                 if (dev_ingress_queue(dev))
1350                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
1351                                              &noop_qdisc);
1352
1353                 dev_watchdog_down(dev);
1354         }
1355
1356         /* Wait for outstanding qdisc-less dev_queue_xmit calls or
1357          * outstanding qdisc enqueuing calls.
1358          * This is avoided if all devices are in dismantle phase :
1359          * Caller will call synchronize_net() for us
1360          */
1361         synchronize_net();
1362
1363         list_for_each_entry(dev, head, close_list) {
1364                 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1365
1366                 if (dev_ingress_queue(dev))
1367                         dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1368         }
1369
1370         /* Wait for outstanding qdisc_run calls. */
1371         list_for_each_entry(dev, head, close_list) {
1372                 while (some_qdisc_is_busy(dev)) {
1373                         /* wait_event() would avoid this sleep-loop but would
1374                          * require expensive checks in the fast paths of packet
1375                          * processing which isn't worth it.
1376                          */
1377                         schedule_timeout_uninterruptible(1);
1378                 }
1379         }
1380 }
1381
1382 void dev_deactivate(struct net_device *dev)
1383 {
1384         LIST_HEAD(single);
1385
1386         list_add(&dev->close_list, &single);
1387         dev_deactivate_many(&single);
1388         list_del(&single);
1389 }
1390 EXPORT_SYMBOL(dev_deactivate);
1391
1392 static int qdisc_change_tx_queue_len(struct net_device *dev,
1393                                      struct netdev_queue *dev_queue)
1394 {
1395         struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1396         const struct Qdisc_ops *ops = qdisc->ops;
1397
1398         if (ops->change_tx_queue_len)
1399                 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1400         return 0;
1401 }
1402
1403 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1404                                   unsigned int new_real_tx)
1405 {
1406         struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1407
1408         if (qdisc->ops->change_real_num_tx)
1409                 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1410 }
1411
1412 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1413 {
1414 #ifdef CONFIG_NET_SCHED
1415         struct net_device *dev = qdisc_dev(sch);
1416         struct Qdisc *qdisc;
1417         unsigned int i;
1418
1419         for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1420                 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1421                 /* Only update the default qdiscs we created,
1422                  * qdiscs with handles are always hashed.
1423                  */
1424                 if (qdisc != &noop_qdisc && !qdisc->handle)
1425                         qdisc_hash_del(qdisc);
1426         }
1427         for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1428                 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1429                 if (qdisc != &noop_qdisc && !qdisc->handle)
1430                         qdisc_hash_add(qdisc, false);
1431         }
1432 #endif
1433 }
1434 EXPORT_SYMBOL(mq_change_real_num_tx);
1435
1436 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1437 {
1438         bool up = dev->flags & IFF_UP;
1439         unsigned int i;
1440         int ret = 0;
1441
1442         if (up)
1443                 dev_deactivate(dev);
1444
1445         for (i = 0; i < dev->num_tx_queues; i++) {
1446                 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1447
1448                 /* TODO: revert changes on a partial failure */
1449                 if (ret)
1450                         break;
1451         }
1452
1453         if (up)
1454                 dev_activate(dev);
1455         return ret;
1456 }
1457
1458 static void dev_init_scheduler_queue(struct net_device *dev,
1459                                      struct netdev_queue *dev_queue,
1460                                      void *_qdisc)
1461 {
1462         struct Qdisc *qdisc = _qdisc;
1463
1464         rcu_assign_pointer(dev_queue->qdisc, qdisc);
1465         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1466 }
1467
1468 void dev_init_scheduler(struct net_device *dev)
1469 {
1470         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1471         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1472         if (dev_ingress_queue(dev))
1473                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1474
1475         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1476 }
1477
1478 void dev_shutdown(struct net_device *dev)
1479 {
1480         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1481         if (dev_ingress_queue(dev))
1482                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1483         qdisc_put(rtnl_dereference(dev->qdisc));
1484         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1485
1486         WARN_ON(timer_pending(&dev->watchdog_timer));
1487 }
1488
1489 /**
1490  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1491  * @rate:   Rate to compute reciprocal division values of
1492  * @mult:   Multiplier for reciprocal division
1493  * @shift:  Shift for reciprocal division
1494  *
1495  * The multiplier and shift for reciprocal division by rate are stored
1496  * in mult and shift.
1497  *
1498  * The deal here is to replace a divide by a reciprocal one
1499  * in fast path (a reciprocal divide is a multiply and a shift)
1500  *
1501  * Normal formula would be :
1502  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1503  *
1504  * We compute mult/shift to use instead :
1505  *  time_in_ns = (len * mult) >> shift;
1506  *
1507  * We try to get the highest possible mult value for accuracy,
1508  * but have to make sure no overflows will ever happen.
1509  *
1510  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1511  */
1512 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1513 {
1514         u64 factor = NSEC_PER_SEC;
1515
1516         *mult = 1;
1517         *shift = 0;
1518
1519         if (rate <= 0)
1520                 return;
1521
1522         for (;;) {
1523                 *mult = div64_u64(factor, rate);
1524                 if (*mult & (1U << 31) || factor & (1ULL << 63))
1525                         break;
1526                 factor <<= 1;
1527                 (*shift)++;
1528         }
1529 }
1530
1531 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1532                                const struct tc_ratespec *conf,
1533                                u64 rate64)
1534 {
1535         memset(r, 0, sizeof(*r));
1536         r->overhead = conf->overhead;
1537         r->mpu = conf->mpu;
1538         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1539         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1540         psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1541 }
1542 EXPORT_SYMBOL(psched_ratecfg_precompute);
1543
1544 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1545 {
1546         r->rate_pkts_ps = pktrate64;
1547         psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1548 }
1549 EXPORT_SYMBOL(psched_ppscfg_precompute);
1550
1551 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1552                           struct tcf_proto *tp_head)
1553 {
1554         /* Protected with chain0->filter_chain_lock.
1555          * Can't access chain directly because tp_head can be NULL.
1556          */
1557         struct mini_Qdisc *miniq_old =
1558                 rcu_dereference_protected(*miniqp->p_miniq, 1);
1559         struct mini_Qdisc *miniq;
1560
1561         if (!tp_head) {
1562                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1563         } else {
1564                 miniq = miniq_old != &miniqp->miniq1 ?
1565                         &miniqp->miniq1 : &miniqp->miniq2;
1566
1567                 /* We need to make sure that readers won't see the miniq
1568                  * we are about to modify. So ensure that at least one RCU
1569                  * grace period has elapsed since the miniq was made
1570                  * inactive.
1571                  */
1572                 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1573                         cond_synchronize_rcu(miniq->rcu_state);
1574                 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1575                         synchronize_rcu_expedited();
1576
1577                 miniq->filter_list = tp_head;
1578                 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1579         }
1580
1581         if (miniq_old)
1582                 /* This is counterpart of the rcu sync above. We need to
1583                  * block potential new user of miniq_old until all readers
1584                  * are not seeing it.
1585                  */
1586                 miniq_old->rcu_state = start_poll_synchronize_rcu();
1587 }
1588 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1589
1590 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1591                                 struct tcf_block *block)
1592 {
1593         miniqp->miniq1.block = block;
1594         miniqp->miniq2.block = block;
1595 }
1596 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1597
1598 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1599                           struct mini_Qdisc __rcu **p_miniq)
1600 {
1601         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1602         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1603         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1604         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1605         miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1606         miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1607         miniqp->p_miniq = p_miniq;
1608 }
1609 EXPORT_SYMBOL(mini_qdisc_pair_init);