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 <linux/bpf.h>
  28 #include <net/sch_generic.h>
  29 #include <net/pkt_sched.h>
  30 #include <net/dst.h>
  31 #include <net/hotdata.h>
  32 #include <trace/events/qdisc.h>
  33 #include <trace/events/net.h>
  34 #include <net/xfrm.h>
  35
  36 /* Qdisc to use by default */
  37 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  38 EXPORT_SYMBOL(default_qdisc_ops);
  39
  40 static void qdisc_maybe_clear_missed(struct Qdisc *q,
  41                                      const struct netdev_queue *txq)
  42 {
  43         clear_bit(__QDISC_STATE_MISSED, &q->state);
  44
  45         /* Make sure the below netif_xmit_frozen_or_stopped()
  46          * checking happens after clearing STATE_MISSED.
  47          */
  48         smp_mb__after_atomic();
  49
  50         /* Checking netif_xmit_frozen_or_stopped() again to
  51          * make sure STATE_MISSED is set if the STATE_MISSED
  52          * set by netif_tx_wake_queue()'s rescheduling of
  53          * net_tx_action() is cleared by the above clear_bit().
  54          */
  55         if (!netif_xmit_frozen_or_stopped(txq))
  56                 set_bit(__QDISC_STATE_MISSED, &q->state);
  57         else
  58                 set_bit(__QDISC_STATE_DRAINING, &q->state);
  59 }
  60
  61 /* Main transmission queue. */
  62
  63 /* Modifications to data participating in scheduling must be protected with
  64  * qdisc_lock(qdisc) spinlock.
  65  *
  66  * The idea is the following:
  67  * - enqueue, dequeue are serialized via qdisc root lock
  68  * - ingress filtering is also serialized via qdisc root lock
  69  * - updates to tree and tree walking are only done under the rtnl mutex.
  70  */
  71
  72 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  73
  74 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  75 {
  76         const struct netdev_queue *txq = q->dev_queue;
  77         spinlock_t *lock = NULL;
  78         struct sk_buff *skb;
  79
  80         if (q->flags & TCQ_F_NOLOCK) {
  81                 lock = qdisc_lock(q);
  82                 spin_lock(lock);
  83         }
  84
  85         skb = skb_peek(&q->skb_bad_txq);
  86         if (skb) {
  87                 /* check the reason of requeuing without tx lock first */
  88                 txq = skb_get_tx_queue(txq->dev, skb);
  89                 if (!netif_xmit_frozen_or_stopped(txq)) {
  90                         skb = __skb_dequeue(&q->skb_bad_txq);
  91                         if (qdisc_is_percpu_stats(q)) {
  92                                 qdisc_qstats_cpu_backlog_dec(q, skb);
  93                                 qdisc_qstats_cpu_qlen_dec(q);
  94                         } else {
  95                                 qdisc_qstats_backlog_dec(q, skb);
  96                                 q->q.qlen--;
  97                         }
  98                 } else {
  99                         skb = SKB_XOFF_MAGIC;
 100                         qdisc_maybe_clear_missed(q, txq);
 101                 }
 102         }
 103
 104         if (lock)
 105                 spin_unlock(lock);
 106
 107         return skb;
 108 }
 109
 110 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
 111 {
 112         struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
 113
 114         if (unlikely(skb))
 115                 skb = __skb_dequeue_bad_txq(q);
 116
 117         return skb;
 118 }
 119
 120 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
 121                                              struct sk_buff *skb)
 122 {
 123         spinlock_t *lock = NULL;
 124
 125         if (q->flags & TCQ_F_NOLOCK) {
 126                 lock = qdisc_lock(q);
 127                 spin_lock(lock);
 128         }
 129
 130         __skb_queue_tail(&q->skb_bad_txq, skb);
 131
 132         if (qdisc_is_percpu_stats(q)) {
 133                 qdisc_qstats_cpu_backlog_inc(q, skb);
 134                 qdisc_qstats_cpu_qlen_inc(q);
 135         } else {
 136                 qdisc_qstats_backlog_inc(q, skb);
 137                 q->q.qlen++;
 138         }
 139
 140         if (lock)
 141                 spin_unlock(lock);
 142 }
 143
 144 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 145 {
 146         spinlock_t *lock = NULL;
 147
 148         if (q->flags & TCQ_F_NOLOCK) {
 149                 lock = qdisc_lock(q);
 150                 spin_lock(lock);
 151         }
 152
 153         while (skb) {
 154                 struct sk_buff *next = skb->next;
 155
 156                 __skb_queue_tail(&q->gso_skb, skb);
 157
 158                 /* it's still part of the queue */
 159                 if (qdisc_is_percpu_stats(q)) {
 160                         qdisc_qstats_cpu_requeues_inc(q);
 161                         qdisc_qstats_cpu_backlog_inc(q, skb);
 162                         qdisc_qstats_cpu_qlen_inc(q);
 163                 } else {
 164                         q->qstats.requeues++;
 165                         qdisc_qstats_backlog_inc(q, skb);
 166                         q->q.qlen++;
 167                 }
 168
 169                 skb = next;
 170         }
 171
 172         if (lock) {
 173                 spin_unlock(lock);
 174                 set_bit(__QDISC_STATE_MISSED, &q->state);
 175         } else {
 176                 __netif_schedule(q);
 177         }
 178 }
 179
 180 static void try_bulk_dequeue_skb(struct Qdisc *q,
 181                                  struct sk_buff *skb,
 182                                  const struct netdev_queue *txq,
 183                                  int *packets)
 184 {
 185         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 186
 187         while (bytelimit > 0) {
 188                 struct sk_buff *nskb = q->dequeue(q);
 189
 190                 if (!nskb)
 191                         break;
 192
 193                 bytelimit -= nskb->len; /* covers GSO len */
 194                 skb->next = nskb;
 195                 skb = nskb;
 196                 (*packets)++; /* GSO counts as one pkt */
 197         }
 198         skb_mark_not_on_list(skb);
 199 }
 200
 201 /* This variant of try_bulk_dequeue_skb() makes sure
 202  * all skbs in the chain are for the same txq
 203  */
 204 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 205                                       struct sk_buff *skb,
 206                                       int *packets)
 207 {
 208         int mapping = skb_get_queue_mapping(skb);
 209         struct sk_buff *nskb;
 210         int cnt = 0;
 211
 212         do {
 213                 nskb = q->dequeue(q);
 214                 if (!nskb)
 215                         break;
 216                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 217                         qdisc_enqueue_skb_bad_txq(q, nskb);
 218                         break;
 219                 }
 220                 skb->next = nskb;
 221                 skb = nskb;
 222         } while (++cnt < 8);
 223         (*packets) += cnt;
 224         skb_mark_not_on_list(skb);
 225 }
 226
 227 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 228  * A requeued skb (via q->gso_skb) can also be a SKB list.
 229  */
 230 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 231                                    int *packets)
 232 {
 233         const struct netdev_queue *txq = q->dev_queue;
 234         struct sk_buff *skb = NULL;
 235
 236         *packets = 1;
 237         if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 238                 spinlock_t *lock = NULL;
 239
 240                 if (q->flags & TCQ_F_NOLOCK) {
 241                         lock = qdisc_lock(q);
 242                         spin_lock(lock);
 243                 }
 244
 245                 skb = skb_peek(&q->gso_skb);
 246
 247                 /* skb may be null if another cpu pulls gso_skb off in between
 248                  * empty check and lock.
 249                  */
 250                 if (!skb) {
 251                         if (lock)
 252                                 spin_unlock(lock);
 253                         goto validate;
 254                 }
 255
 256                 /* skb in gso_skb were already validated */
 257                 *validate = false;
 258                 if (xfrm_offload(skb))
 259                         *validate = true;
 260                 /* check the reason of requeuing without tx lock first */
 261                 txq = skb_get_tx_queue(txq->dev, skb);
 262                 if (!netif_xmit_frozen_or_stopped(txq)) {
 263                         skb = __skb_dequeue(&q->gso_skb);
 264                         if (qdisc_is_percpu_stats(q)) {
 265                                 qdisc_qstats_cpu_backlog_dec(q, skb);
 266                                 qdisc_qstats_cpu_qlen_dec(q);
 267                         } else {
 268                                 qdisc_qstats_backlog_dec(q, skb);
 269                                 q->q.qlen--;
 270                         }
 271                 } else {
 272                         skb = NULL;
 273                         qdisc_maybe_clear_missed(q, txq);
 274                 }
 275                 if (lock)
 276                         spin_unlock(lock);
 277                 goto trace;
 278         }
 279 validate:
 280         *validate = true;
 281
 282         if ((q->flags & TCQ_F_ONETXQUEUE) &&
 283             netif_xmit_frozen_or_stopped(txq)) {
 284                 qdisc_maybe_clear_missed(q, txq);
 285                 return skb;
 286         }
 287
 288         skb = qdisc_dequeue_skb_bad_txq(q);
 289         if (unlikely(skb)) {
 290                 if (skb == SKB_XOFF_MAGIC)
 291                         return NULL;
 292                 goto bulk;
 293         }
 294         skb = q->dequeue(q);
 295         if (skb) {
 296 bulk:
 297                 if (qdisc_may_bulk(q))
 298                         try_bulk_dequeue_skb(q, skb, txq, packets);
 299                 else
 300                         try_bulk_dequeue_skb_slow(q, skb, packets);
 301         }
 302 trace:
 303         trace_qdisc_dequeue(q, txq, *packets, skb);
 304         return skb;
 305 }
 306
 307 /*
 308  * Transmit possibly several skbs, and handle the return status as
 309  * required. Owning qdisc running bit guarantees that only one CPU
 310  * can execute this function.
 311  *
 312  * Returns to the caller:
 313  *                              false  - hardware queue frozen backoff
 314  *                              true   - feel free to send more pkts
 315  */
 316 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 317                      struct net_device *dev, struct netdev_queue *txq,
 318                      spinlock_t *root_lock, bool validate)
 319 {
 320         int ret = NETDEV_TX_BUSY;
 321         bool again = false;
 322
 323         /* And release qdisc */
 324         if (root_lock)
 325                 spin_unlock(root_lock);
 326
 327         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 328         if (validate)
 329                 skb = validate_xmit_skb_list(skb, dev, &again);
 330
 331 #ifdef CONFIG_XFRM_OFFLOAD
 332         if (unlikely(again)) {
 333                 if (root_lock)
 334                         spin_lock(root_lock);
 335
 336                 dev_requeue_skb(skb, q);
 337                 return false;
 338         }
 339 #endif
 340
 341         if (likely(skb)) {
 342                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 343                 if (!netif_xmit_frozen_or_stopped(txq))
 344                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 345                 else
 346                         qdisc_maybe_clear_missed(q, txq);
 347
 348                 HARD_TX_UNLOCK(dev, txq);
 349         } else {
 350                 if (root_lock)
 351                         spin_lock(root_lock);
 352                 return true;
 353         }
 354
 355         if (root_lock)
 356                 spin_lock(root_lock);
 357
 358         if (!dev_xmit_complete(ret)) {
 359                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 360                 if (unlikely(ret != NETDEV_TX_BUSY))
 361                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 362                                              dev->name, ret, q->q.qlen);
 363
 364                 dev_requeue_skb(skb, q);
 365                 return false;
 366         }
 367
 368         return true;
 369 }
 370
 371 /*
 372  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 373  *
 374  * running seqcount guarantees only one CPU can process
 375  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 376  * this queue.
 377  *
 378  *  netif_tx_lock serializes accesses to device driver.
 379  *
 380  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 381  *  if one is grabbed, another must be free.
 382  *
 383  * Note, that this procedure can be called by a watchdog timer
 384  *
 385  * Returns to the caller:
 386  *                              0  - queue is empty or throttled.
 387  *                              >0 - queue is not empty.
 388  *
 389  */
 390 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 391 {
 392         spinlock_t *root_lock = NULL;
 393         struct netdev_queue *txq;
 394         struct net_device *dev;
 395         struct sk_buff *skb;
 396         bool validate;
 397
 398         /* Dequeue packet */
 399         skb = dequeue_skb(q, &validate, packets);
 400         if (unlikely(!skb))
 401                 return false;
 402
 403         if (!(q->flags & TCQ_F_NOLOCK))
 404                 root_lock = qdisc_lock(q);
 405
 406         dev = qdisc_dev(q);
 407         txq = skb_get_tx_queue(dev, skb);
 408
 409         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 410 }
 411
 412 void __qdisc_run(struct Qdisc *q)
 413 {
 414         int quota = READ_ONCE(net_hotdata.dev_tx_weight);
 415         int packets;
 416
 417         while (qdisc_restart(q, &packets)) {
 418                 quota -= packets;
 419                 if (quota <= 0) {
 420                         if (q->flags & TCQ_F_NOLOCK)
 421                                 set_bit(__QDISC_STATE_MISSED, &q->state);
 422                         else
 423                                 __netif_schedule(q);
 424
 425                         break;
 426                 }
 427         }
 428 }
 429
 430 unsigned long dev_trans_start(struct net_device *dev)
 431 {
 432         unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
 433         unsigned long val;
 434         unsigned int i;
 435
 436         for (i = 1; i < dev->num_tx_queues; i++) {
 437                 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
 438                 if (val && time_after(val, res))
 439                         res = val;
 440         }
 441
 442         return res;
 443 }
 444 EXPORT_SYMBOL(dev_trans_start);
 445
 446 static void netif_freeze_queues(struct net_device *dev)
 447 {
 448         unsigned int i;
 449         int cpu;
 450
 451         cpu = smp_processor_id();
 452         for (i = 0; i < dev->num_tx_queues; i++) {
 453                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 454
 455                 /* We are the only thread of execution doing a
 456                  * freeze, but we have to grab the _xmit_lock in
 457                  * order to synchronize with threads which are in
 458                  * the ->hard_start_xmit() handler and already
 459                  * checked the frozen bit.
 460                  */
 461                 __netif_tx_lock(txq, cpu);
 462                 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
 463                 __netif_tx_unlock(txq);
 464         }
 465 }
 466
 467 void netif_tx_lock(struct net_device *dev)
 468 {
 469         spin_lock(&dev->tx_global_lock);
 470         netif_freeze_queues(dev);
 471 }
 472 EXPORT_SYMBOL(netif_tx_lock);
 473
 474 static void netif_unfreeze_queues(struct net_device *dev)
 475 {
 476         unsigned int i;
 477
 478         for (i = 0; i < dev->num_tx_queues; i++) {
 479                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 480
 481                 /* No need to grab the _xmit_lock here.  If the
 482                  * queue is not stopped for another reason, we
 483                  * force a schedule.
 484                  */
 485                 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
 486                 netif_schedule_queue(txq);
 487         }
 488 }
 489
 490 void netif_tx_unlock(struct net_device *dev)
 491 {
 492         netif_unfreeze_queues(dev);
 493         spin_unlock(&dev->tx_global_lock);
 494 }
 495 EXPORT_SYMBOL(netif_tx_unlock);
 496
 497 static void dev_watchdog(struct timer_list *t)
 498 {
 499         struct net_device *dev = timer_container_of(dev, t, watchdog_timer);
 500         bool release = true;
 501
 502         spin_lock(&dev->tx_global_lock);
 503         if (!qdisc_tx_is_noop(dev)) {
 504                 if (netif_device_present(dev) &&
 505                     netif_running(dev) &&
 506                     netif_carrier_ok(dev)) {
 507                         unsigned int timedout_ms = 0;
 508                         unsigned int i;
 509                         unsigned long trans_start;
 510                         unsigned long oldest_start = jiffies;
 511
 512                         for (i = 0; i < dev->num_tx_queues; i++) {
 513                                 struct netdev_queue *txq;
 514
 515                                 txq = netdev_get_tx_queue(dev, i);
 516                                 if (!netif_xmit_stopped(txq))
 517                                         continue;
 518
 519                                 /* Paired with WRITE_ONCE() + smp_mb...() in
 520                                  * netdev_tx_sent_queue() and netif_tx_stop_queue().
 521                                  */
 522                                 smp_mb();
 523                                 trans_start = READ_ONCE(txq->trans_start);
 524
 525                                 if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
 526                                         timedout_ms = jiffies_to_msecs(jiffies - trans_start);
 527                                         atomic_long_inc(&txq->trans_timeout);
 528                                         break;
 529                                 }
 530                                 if (time_after(oldest_start, trans_start))
 531                                         oldest_start = trans_start;
 532                         }
 533
 534                         if (unlikely(timedout_ms)) {
 535                                 trace_net_dev_xmit_timeout(dev, i);
 536                                 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
 537                                             raw_smp_processor_id(),
 538                                             i, timedout_ms);
 539                                 netif_freeze_queues(dev);
 540                                 dev->netdev_ops->ndo_tx_timeout(dev, i);
 541                                 netif_unfreeze_queues(dev);
 542                         }
 543                         if (!mod_timer(&dev->watchdog_timer,
 544                                        round_jiffies(oldest_start +
 545                                                      dev->watchdog_timeo)))
 546                                 release = false;
 547                 }
 548         }
 549         spin_unlock(&dev->tx_global_lock);
 550
 551         if (release)
 552                 netdev_put(dev, &dev->watchdog_dev_tracker);
 553 }
 554
 555 void netdev_watchdog_up(struct net_device *dev)
 556 {
 557         if (!dev->netdev_ops->ndo_tx_timeout)
 558                 return;
 559         if (dev->watchdog_timeo <= 0)
 560                 dev->watchdog_timeo = 5*HZ;
 561         if (!mod_timer(&dev->watchdog_timer,
 562                        round_jiffies(jiffies + dev->watchdog_timeo)))
 563                 netdev_hold(dev, &dev->watchdog_dev_tracker,
 564                             GFP_ATOMIC);
 565 }
 566 EXPORT_SYMBOL_GPL(netdev_watchdog_up);
 567
 568 static void netdev_watchdog_down(struct net_device *dev)
 569 {
 570         netif_tx_lock_bh(dev);
 571         if (timer_delete(&dev->watchdog_timer))
 572                 netdev_put(dev, &dev->watchdog_dev_tracker);
 573         netif_tx_unlock_bh(dev);
 574 }
 575
 576 /**
 577  *      netif_carrier_on - set carrier
 578  *      @dev: network device
 579  *
 580  * Device has detected acquisition of carrier.
 581  */
 582 void netif_carrier_on(struct net_device *dev)
 583 {
 584         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 585                 if (dev->reg_state == NETREG_UNINITIALIZED)
 586                         return;
 587                 atomic_inc(&dev->carrier_up_count);
 588                 linkwatch_fire_event(dev);
 589                 if (netif_running(dev))
 590                         netdev_watchdog_up(dev);
 591         }
 592 }
 593 EXPORT_SYMBOL(netif_carrier_on);
 594
 595 /**
 596  *      netif_carrier_off - clear carrier
 597  *      @dev: network device
 598  *
 599  * Device has detected loss of carrier.
 600  */
 601 void netif_carrier_off(struct net_device *dev)
 602 {
 603         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 604                 if (dev->reg_state == NETREG_UNINITIALIZED)
 605                         return;
 606                 atomic_inc(&dev->carrier_down_count);
 607                 linkwatch_fire_event(dev);
 608         }
 609 }
 610 EXPORT_SYMBOL(netif_carrier_off);
 611
 612 /**
 613  *      netif_carrier_event - report carrier state event
 614  *      @dev: network device
 615  *
 616  * Device has detected a carrier event but the carrier state wasn't changed.
 617  * Use in drivers when querying carrier state asynchronously, to avoid missing
 618  * events (link flaps) if link recovers before it's queried.
 619  */
 620 void netif_carrier_event(struct net_device *dev)
 621 {
 622         if (dev->reg_state == NETREG_UNINITIALIZED)
 623                 return;
 624         atomic_inc(&dev->carrier_up_count);
 625         atomic_inc(&dev->carrier_down_count);
 626         linkwatch_fire_event(dev);
 627 }
 628 EXPORT_SYMBOL_GPL(netif_carrier_event);
 629
 630 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 631    under all circumstances. It is difficult to invent anything faster or
 632    cheaper.
 633  */
 634
 635 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 636                         struct sk_buff **to_free)
 637 {
 638         dev_core_stats_tx_dropped_inc(skb->dev);
 639         __qdisc_drop(skb, to_free);
 640         return NET_XMIT_CN;
 641 }
 642
 643 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 644 {
 645         return NULL;
 646 }
 647
 648 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 649         .id             =       "noop",
 650         .priv_size      =       0,
 651         .enqueue        =       noop_enqueue,
 652         .dequeue        =       noop_dequeue,
 653         .peek           =       noop_dequeue,
 654         .owner          =       THIS_MODULE,
 655 };
 656
 657 static struct netdev_queue noop_netdev_queue = {
 658         RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 659         RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
 660 };
 661
 662 struct Qdisc noop_qdisc = {
 663         .enqueue        =       noop_enqueue,
 664         .dequeue        =       noop_dequeue,
 665         .flags          =       TCQ_F_BUILTIN,
 666         .ops            =       &noop_qdisc_ops,
 667         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 668         .dev_queue      =       &noop_netdev_queue,
 669         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 670         .gso_skb = {
 671                 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
 672                 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 673                 .qlen = 0,
 674                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 675         },
 676         .skb_bad_txq = {
 677                 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 678                 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 679                 .qlen = 0,
 680                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 681         },
 682         .owner = -1,
 683 };
 684 EXPORT_SYMBOL(noop_qdisc);
 685
 686 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 687                         struct netlink_ext_ack *extack)
 688 {
 689         /* register_qdisc() assigns a default of noop_enqueue if unset,
 690          * but __dev_queue_xmit() treats noqueue only as such
 691          * if this is NULL - so clear it here. */
 692         qdisc->enqueue = NULL;
 693         return 0;
 694 }
 695
 696 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 697         .id             =       "noqueue",
 698         .priv_size      =       0,
 699         .init           =       noqueue_init,
 700         .enqueue        =       noop_enqueue,
 701         .dequeue        =       noop_dequeue,
 702         .peek           =       noop_dequeue,
 703         .owner          =       THIS_MODULE,
 704 };
 705
 706 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
 707         1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
 708 };
 709 EXPORT_SYMBOL(sch_default_prio2band);
 710
 711 /* 3-band FIFO queue: old style, but should be a bit faster than
 712    generic prio+fifo combination.
 713  */
 714
 715 #define PFIFO_FAST_BANDS 3
 716
 717 /*
 718  * Private data for a pfifo_fast scheduler containing:
 719  *      - rings for priority bands
 720  */
 721 struct pfifo_fast_priv {
 722         struct skb_array q[PFIFO_FAST_BANDS];
 723 };
 724
 725 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 726                                           int band)
 727 {
 728         return &priv->q[band];
 729 }
 730
 731 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 732                               struct sk_buff **to_free)
 733 {
 734         int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
 735         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 736         struct skb_array *q = band2list(priv, band);
 737         unsigned int pkt_len = qdisc_pkt_len(skb);
 738         int err;
 739
 740         err = skb_array_produce(q, skb);
 741
 742         if (unlikely(err)) {
 743                 if (qdisc_is_percpu_stats(qdisc))
 744                         return qdisc_drop_cpu(skb, qdisc, to_free);
 745                 else
 746                         return qdisc_drop(skb, qdisc, to_free);
 747         }
 748
 749         qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 750         return NET_XMIT_SUCCESS;
 751 }
 752
 753 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 754 {
 755         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 756         struct sk_buff *skb = NULL;
 757         bool need_retry = true;
 758         int band;
 759
 760 retry:
 761         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 762                 struct skb_array *q = band2list(priv, band);
 763
 764                 if (__skb_array_empty(q))
 765                         continue;
 766
 767                 skb = __skb_array_consume(q);
 768         }
 769         if (likely(skb)) {
 770                 qdisc_update_stats_at_dequeue(qdisc, skb);
 771         } else if (need_retry &&
 772                    READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
 773                 /* Delay clearing the STATE_MISSED here to reduce
 774                  * the overhead of the second spin_trylock() in
 775                  * qdisc_run_begin() and __netif_schedule() calling
 776                  * in qdisc_run_end().
 777                  */
 778                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
 779                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
 780
 781                 /* Make sure dequeuing happens after clearing
 782                  * STATE_MISSED.
 783                  */
 784                 smp_mb__after_atomic();
 785
 786                 need_retry = false;
 787
 788                 goto retry;
 789         }
 790
 791         return skb;
 792 }
 793
 794 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 795 {
 796         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 797         struct sk_buff *skb = NULL;
 798         int band;
 799
 800         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 801                 struct skb_array *q = band2list(priv, band);
 802
 803                 skb = __skb_array_peek(q);
 804         }
 805
 806         return skb;
 807 }
 808
 809 static void pfifo_fast_reset(struct Qdisc *qdisc)
 810 {
 811         int i, band;
 812         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 813
 814         for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 815                 struct skb_array *q = band2list(priv, band);
 816                 struct sk_buff *skb;
 817
 818                 /* NULL ring is possible if destroy path is due to a failed
 819                  * skb_array_init() in pfifo_fast_init() case.
 820                  */
 821                 if (!q->ring.queue)
 822                         continue;
 823
 824                 while ((skb = __skb_array_consume(q)) != NULL)
 825                         kfree_skb(skb);
 826         }
 827
 828         if (qdisc_is_percpu_stats(qdisc)) {
 829                 for_each_possible_cpu(i) {
 830                         struct gnet_stats_queue *q;
 831
 832                         q = per_cpu_ptr(qdisc->cpu_qstats, i);
 833                         q->backlog = 0;
 834                         q->qlen = 0;
 835                 }
 836         }
 837 }
 838
 839 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 840 {
 841         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 842
 843         memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
 844         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 845                 goto nla_put_failure;
 846         return skb->len;
 847
 848 nla_put_failure:
 849         return -1;
 850 }
 851
 852 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 853                            struct netlink_ext_ack *extack)
 854 {
 855         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 856         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 857         int prio;
 858
 859         /* guard against zero length rings */
 860         if (!qlen)
 861                 return -EINVAL;
 862
 863         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 864                 struct skb_array *q = band2list(priv, prio);
 865                 int err;
 866
 867                 err = skb_array_init(q, qlen, GFP_KERNEL);
 868                 if (err)
 869                         return -ENOMEM;
 870         }
 871
 872         /* Can by-pass the queue discipline */
 873         qdisc->flags |= TCQ_F_CAN_BYPASS;
 874         return 0;
 875 }
 876
 877 static void pfifo_fast_destroy(struct Qdisc *sch)
 878 {
 879         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 880         int prio;
 881
 882         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 883                 struct skb_array *q = band2list(priv, prio);
 884
 885                 /* NULL ring is possible if destroy path is due to a failed
 886                  * skb_array_init() in pfifo_fast_init() case.
 887                  */
 888                 if (!q->ring.queue)
 889                         continue;
 890                 /* Destroy ring but no need to kfree_skb because a call to
 891                  * pfifo_fast_reset() has already done that work.
 892                  */
 893                 ptr_ring_cleanup(&q->ring, NULL);
 894         }
 895 }
 896
 897 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 898                                           unsigned int new_len)
 899 {
 900         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 901         struct skb_array *bands[PFIFO_FAST_BANDS];
 902         int prio;
 903
 904         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 905                 struct skb_array *q = band2list(priv, prio);
 906
 907                 bands[prio] = q;
 908         }
 909
 910         return skb_array_resize_multiple_bh(bands, PFIFO_FAST_BANDS, new_len,
 911                                             GFP_KERNEL);
 912 }
 913
 914 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 915         .id             =       "pfifo_fast",
 916         .priv_size      =       sizeof(struct pfifo_fast_priv),
 917         .enqueue        =       pfifo_fast_enqueue,
 918         .dequeue        =       pfifo_fast_dequeue,
 919         .peek           =       pfifo_fast_peek,
 920         .init           =       pfifo_fast_init,
 921         .destroy        =       pfifo_fast_destroy,
 922         .reset          =       pfifo_fast_reset,
 923         .dump           =       pfifo_fast_dump,
 924         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 925         .owner          =       THIS_MODULE,
 926         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 927 };
 928 EXPORT_SYMBOL(pfifo_fast_ops);
 929
 930 static struct lock_class_key qdisc_tx_busylock;
 931
 932 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 933                           const struct Qdisc_ops *ops,
 934                           struct netlink_ext_ack *extack)
 935 {
 936         struct Qdisc *sch;
 937         unsigned int size = sizeof(*sch) + ops->priv_size;
 938         int err = -ENOBUFS;
 939         struct net_device *dev;
 940
 941         if (!dev_queue) {
 942                 NL_SET_ERR_MSG(extack, "No device queue given");
 943                 err = -EINVAL;
 944                 goto errout;
 945         }
 946
 947         dev = dev_queue->dev;
 948         sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
 949
 950         if (!sch)
 951                 goto errout;
 952         __skb_queue_head_init(&sch->gso_skb);
 953         __skb_queue_head_init(&sch->skb_bad_txq);
 954         gnet_stats_basic_sync_init(&sch->bstats);
 955         lockdep_register_key(&sch->root_lock_key);
 956         spin_lock_init(&sch->q.lock);
 957         lockdep_set_class(&sch->q.lock, &sch->root_lock_key);
 958
 959         if (ops->static_flags & TCQ_F_CPUSTATS) {
 960                 sch->cpu_bstats =
 961                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
 962                 if (!sch->cpu_bstats)
 963                         goto errout1;
 964
 965                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 966                 if (!sch->cpu_qstats) {
 967                         free_percpu(sch->cpu_bstats);
 968                         goto errout1;
 969                 }
 970         }
 971
 972         spin_lock_init(&sch->busylock);
 973         lockdep_set_class(&sch->busylock,
 974                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 975
 976         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 977         spin_lock_init(&sch->seqlock);
 978         lockdep_set_class(&sch->seqlock,
 979                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 980
 981         sch->ops = ops;
 982         sch->flags = ops->static_flags;
 983         sch->enqueue = ops->enqueue;
 984         sch->dequeue = ops->dequeue;
 985         sch->dev_queue = dev_queue;
 986         sch->owner = -1;
 987         netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL);
 988         refcount_set(&sch->refcnt, 1);
 989
 990         return sch;
 991 errout1:
 992         lockdep_unregister_key(&sch->root_lock_key);
 993         kfree(sch);
 994 errout:
 995         return ERR_PTR(err);
 996 }
 997
 998 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 999                                 const struct Qdisc_ops *ops,
1000                                 unsigned int parentid,
1001                                 struct netlink_ext_ack *extack)
1002 {
1003         struct Qdisc *sch;
1004
1005         if (!bpf_try_module_get(ops, ops->owner)) {
1006                 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
1007                 return NULL;
1008         }
1009
1010         sch = qdisc_alloc(dev_queue, ops, extack);
1011         if (IS_ERR(sch)) {
1012                 bpf_module_put(ops, ops->owner);
1013                 return NULL;
1014         }
1015         sch->parent = parentid;
1016
1017         if (!ops->init || ops->init(sch, NULL, extack) == 0) {
1018                 trace_qdisc_create(ops, dev_queue->dev, parentid);
1019                 return sch;
1020         }
1021
1022         qdisc_put(sch);
1023         return NULL;
1024 }
1025 EXPORT_SYMBOL(qdisc_create_dflt);
1026
1027 /* Under qdisc_lock(qdisc) and BH! */
1028
1029 void qdisc_reset(struct Qdisc *qdisc)
1030 {
1031         const struct Qdisc_ops *ops = qdisc->ops;
1032
1033         trace_qdisc_reset(qdisc);
1034
1035         if (ops->reset)
1036                 ops->reset(qdisc);
1037
1038         __skb_queue_purge(&qdisc->gso_skb);
1039         __skb_queue_purge(&qdisc->skb_bad_txq);
1040
1041         qdisc->q.qlen = 0;
1042         qdisc->qstats.backlog = 0;
1043 }
1044 EXPORT_SYMBOL(qdisc_reset);
1045
1046 void qdisc_free(struct Qdisc *qdisc)
1047 {
1048         if (qdisc_is_percpu_stats(qdisc)) {
1049                 free_percpu(qdisc->cpu_bstats);
1050                 free_percpu(qdisc->cpu_qstats);
1051         }
1052
1053         kfree(qdisc);
1054 }
1055
1056 static void qdisc_free_cb(struct rcu_head *head)
1057 {
1058         struct Qdisc *q = container_of(head, struct Qdisc, rcu);
1059
1060         qdisc_free(q);
1061 }
1062
1063 static void __qdisc_destroy(struct Qdisc *qdisc)
1064 {
1065         const struct Qdisc_ops  *ops = qdisc->ops;
1066         struct net_device *dev = qdisc_dev(qdisc);
1067
1068 #ifdef CONFIG_NET_SCHED
1069         qdisc_hash_del(qdisc);
1070
1071         qdisc_put_stab(rtnl_dereference(qdisc->stab));
1072 #endif
1073         gen_kill_estimator(&qdisc->rate_est);
1074
1075         qdisc_reset(qdisc);
1076
1077
1078         if (ops->destroy)
1079                 ops->destroy(qdisc);
1080
1081         lockdep_unregister_key(&qdisc->root_lock_key);
1082         bpf_module_put(ops, ops->owner);
1083         netdev_put(dev, &qdisc->dev_tracker);
1084
1085         trace_qdisc_destroy(qdisc);
1086
1087         call_rcu(&qdisc->rcu, qdisc_free_cb);
1088 }
1089
1090 void qdisc_destroy(struct Qdisc *qdisc)
1091 {
1092         if (qdisc->flags & TCQ_F_BUILTIN)
1093                 return;
1094
1095         __qdisc_destroy(qdisc);
1096 }
1097
1098 void qdisc_put(struct Qdisc *qdisc)
1099 {
1100         if (!qdisc)
1101                 return;
1102
1103         if (qdisc->flags & TCQ_F_BUILTIN ||
1104             !refcount_dec_and_test(&qdisc->refcnt))
1105                 return;
1106
1107         __qdisc_destroy(qdisc);
1108 }
1109 EXPORT_SYMBOL(qdisc_put);
1110
1111 /* Version of qdisc_put() that is called with rtnl mutex unlocked.
1112  * Intended to be used as optimization, this function only takes rtnl lock if
1113  * qdisc reference counter reached zero.
1114  */
1115
1116 void qdisc_put_unlocked(struct Qdisc *qdisc)
1117 {
1118         if (qdisc->flags & TCQ_F_BUILTIN ||
1119             !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1120                 return;
1121
1122         __qdisc_destroy(qdisc);
1123         rtnl_unlock();
1124 }
1125 EXPORT_SYMBOL(qdisc_put_unlocked);
1126
1127 /* Attach toplevel qdisc to device queue. */
1128 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1129                               struct Qdisc *qdisc)
1130 {
1131         struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1132         spinlock_t *root_lock;
1133
1134         root_lock = qdisc_lock(oqdisc);
1135         spin_lock_bh(root_lock);
1136
1137         /* ... and graft new one */
1138         if (qdisc == NULL)
1139                 qdisc = &noop_qdisc;
1140         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1141         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1142
1143         spin_unlock_bh(root_lock);
1144
1145         return oqdisc;
1146 }
1147 EXPORT_SYMBOL(dev_graft_qdisc);
1148
1149 static void shutdown_scheduler_queue(struct net_device *dev,
1150                                      struct netdev_queue *dev_queue,
1151                                      void *_qdisc_default)
1152 {
1153         struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1154         struct Qdisc *qdisc_default = _qdisc_default;
1155
1156         if (qdisc) {
1157                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1158                 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default);
1159
1160                 qdisc_put(qdisc);
1161         }
1162 }
1163
1164 static void attach_one_default_qdisc(struct net_device *dev,
1165                                      struct netdev_queue *dev_queue,
1166                                      void *_unused)
1167 {
1168         struct Qdisc *qdisc;
1169         const struct Qdisc_ops *ops = default_qdisc_ops;
1170
1171         if (dev->priv_flags & IFF_NO_QUEUE)
1172                 ops = &noqueue_qdisc_ops;
1173         else if(dev->type == ARPHRD_CAN)
1174                 ops = &pfifo_fast_ops;
1175
1176         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1177         if (!qdisc)
1178                 return;
1179
1180         if (!netif_is_multiqueue(dev))
1181                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1182         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1183 }
1184
1185 static void attach_default_qdiscs(struct net_device *dev)
1186 {
1187         struct netdev_queue *txq;
1188         struct Qdisc *qdisc;
1189
1190         txq = netdev_get_tx_queue(dev, 0);
1191
1192         if (!netif_is_multiqueue(dev) ||
1193             dev->priv_flags & IFF_NO_QUEUE) {
1194                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1195                 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1196                 rcu_assign_pointer(dev->qdisc, qdisc);
1197                 qdisc_refcount_inc(qdisc);
1198         } else {
1199                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1200                 if (qdisc) {
1201                         rcu_assign_pointer(dev->qdisc, qdisc);
1202                         qdisc->ops->attach(qdisc);
1203                 }
1204         }
1205         qdisc = rtnl_dereference(dev->qdisc);
1206
1207         /* Detect default qdisc setup/init failed and fallback to "noqueue" */
1208         if (qdisc == &noop_qdisc) {
1209                 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n",
1210                             default_qdisc_ops->id, noqueue_qdisc_ops.id);
1211                 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1212                 dev->priv_flags |= IFF_NO_QUEUE;
1213                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1214                 qdisc = rtnl_dereference(txq->qdisc_sleeping);
1215                 rcu_assign_pointer(dev->qdisc, qdisc);
1216                 qdisc_refcount_inc(qdisc);
1217                 dev->priv_flags ^= IFF_NO_QUEUE;
1218         }
1219
1220 #ifdef CONFIG_NET_SCHED
1221         if (qdisc != &noop_qdisc)
1222                 qdisc_hash_add(qdisc, false);
1223 #endif
1224 }
1225
1226 static void transition_one_qdisc(struct net_device *dev,
1227                                  struct netdev_queue *dev_queue,
1228                                  void *_need_watchdog)
1229 {
1230         struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1231         int *need_watchdog_p = _need_watchdog;
1232
1233         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1234                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1235
1236         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1237         if (need_watchdog_p) {
1238                 WRITE_ONCE(dev_queue->trans_start, 0);
1239                 *need_watchdog_p = 1;
1240         }
1241 }
1242
1243 void dev_activate(struct net_device *dev)
1244 {
1245         int need_watchdog;
1246
1247         /* No queueing discipline is attached to device;
1248          * create default one for devices, which need queueing
1249          * and noqueue_qdisc for virtual interfaces
1250          */
1251
1252         if (rtnl_dereference(dev->qdisc) == &noop_qdisc)
1253                 attach_default_qdiscs(dev);
1254
1255         if (!netif_carrier_ok(dev))
1256                 /* Delay activation until next carrier-on event */
1257                 return;
1258
1259         need_watchdog = 0;
1260         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1261         if (dev_ingress_queue(dev))
1262                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1263
1264         if (need_watchdog) {
1265                 netif_trans_update(dev);
1266                 netdev_watchdog_up(dev);
1267         }
1268 }
1269 EXPORT_SYMBOL(dev_activate);
1270
1271 static void qdisc_deactivate(struct Qdisc *qdisc)
1272 {
1273         if (qdisc->flags & TCQ_F_BUILTIN)
1274                 return;
1275
1276         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1277 }
1278
1279 static void dev_deactivate_queue(struct net_device *dev,
1280                                  struct netdev_queue *dev_queue,
1281                                  void *_sync_needed)
1282 {
1283         bool *sync_needed = _sync_needed;
1284         struct Qdisc *qdisc;
1285
1286         qdisc = rtnl_dereference(dev_queue->qdisc);
1287         if (qdisc) {
1288                 if (qdisc->enqueue)
1289                         *sync_needed = true;
1290                 qdisc_deactivate(qdisc);
1291                 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1292         }
1293 }
1294
1295 static void dev_reset_queue(struct net_device *dev,
1296                             struct netdev_queue *dev_queue,
1297                             void *_unused)
1298 {
1299         struct Qdisc *qdisc;
1300         bool nolock;
1301
1302         qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1303         if (!qdisc)
1304                 return;
1305
1306         nolock = qdisc->flags & TCQ_F_NOLOCK;
1307
1308         if (nolock)
1309                 spin_lock_bh(&qdisc->seqlock);
1310         spin_lock_bh(qdisc_lock(qdisc));
1311
1312         qdisc_reset(qdisc);
1313
1314         spin_unlock_bh(qdisc_lock(qdisc));
1315         if (nolock) {
1316                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
1317                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
1318                 spin_unlock_bh(&qdisc->seqlock);
1319         }
1320 }
1321
1322 static bool some_qdisc_is_busy(struct net_device *dev)
1323 {
1324         unsigned int i;
1325
1326         for (i = 0; i < dev->num_tx_queues; i++) {
1327                 struct netdev_queue *dev_queue;
1328                 spinlock_t *root_lock;
1329                 struct Qdisc *q;
1330                 int val;
1331
1332                 dev_queue = netdev_get_tx_queue(dev, i);
1333                 q = rtnl_dereference(dev_queue->qdisc_sleeping);
1334
1335                 root_lock = qdisc_lock(q);
1336                 spin_lock_bh(root_lock);
1337
1338                 val = (qdisc_is_running(q) ||
1339                        test_bit(__QDISC_STATE_SCHED, &q->state));
1340
1341                 spin_unlock_bh(root_lock);
1342
1343                 if (val)
1344                         return true;
1345         }
1346         return false;
1347 }
1348
1349 /**
1350  *      dev_deactivate_many - deactivate transmissions on several devices
1351  *      @head: list of devices to deactivate
1352  *
1353  *      This function returns only when all outstanding transmissions
1354  *      have completed, unless all devices are in dismantle phase.
1355  */
1356 void dev_deactivate_many(struct list_head *head)
1357 {
1358         bool sync_needed = false;
1359         struct net_device *dev;
1360
1361         list_for_each_entry(dev, head, close_list) {
1362                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1363                                          &sync_needed);
1364                 if (dev_ingress_queue(dev))
1365                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
1366                                              &sync_needed);
1367
1368                 netdev_watchdog_down(dev);
1369         }
1370
1371         /* Wait for outstanding qdisc enqueuing calls. */
1372         if (sync_needed)
1373                 synchronize_net();
1374
1375         list_for_each_entry(dev, head, close_list) {
1376                 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
1377
1378                 if (dev_ingress_queue(dev))
1379                         dev_reset_queue(dev, dev_ingress_queue(dev), NULL);
1380         }
1381
1382         /* Wait for outstanding qdisc_run calls. */
1383         list_for_each_entry(dev, head, close_list) {
1384                 while (some_qdisc_is_busy(dev)) {
1385                         /* wait_event() would avoid this sleep-loop but would
1386                          * require expensive checks in the fast paths of packet
1387                          * processing which isn't worth it.
1388                          */
1389                         schedule_timeout_uninterruptible(1);
1390                 }
1391         }
1392 }
1393
1394 void dev_deactivate(struct net_device *dev)
1395 {
1396         LIST_HEAD(single);
1397
1398         list_add(&dev->close_list, &single);
1399         dev_deactivate_many(&single);
1400         list_del(&single);
1401 }
1402 EXPORT_SYMBOL(dev_deactivate);
1403
1404 static int qdisc_change_tx_queue_len(struct net_device *dev,
1405                                      struct netdev_queue *dev_queue)
1406 {
1407         struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping);
1408         const struct Qdisc_ops *ops = qdisc->ops;
1409
1410         if (ops->change_tx_queue_len)
1411                 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1412         return 0;
1413 }
1414
1415 void dev_qdisc_change_real_num_tx(struct net_device *dev,
1416                                   unsigned int new_real_tx)
1417 {
1418         struct Qdisc *qdisc = rtnl_dereference(dev->qdisc);
1419
1420         if (qdisc->ops->change_real_num_tx)
1421                 qdisc->ops->change_real_num_tx(qdisc, new_real_tx);
1422 }
1423
1424 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx)
1425 {
1426 #ifdef CONFIG_NET_SCHED
1427         struct net_device *dev = qdisc_dev(sch);
1428         struct Qdisc *qdisc;
1429         unsigned int i;
1430
1431         for (i = new_real_tx; i < dev->real_num_tx_queues; i++) {
1432                 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1433                 /* Only update the default qdiscs we created,
1434                  * qdiscs with handles are always hashed.
1435                  */
1436                 if (qdisc != &noop_qdisc && !qdisc->handle)
1437                         qdisc_hash_del(qdisc);
1438         }
1439         for (i = dev->real_num_tx_queues; i < new_real_tx; i++) {
1440                 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping);
1441                 if (qdisc != &noop_qdisc && !qdisc->handle)
1442                         qdisc_hash_add(qdisc, false);
1443         }
1444 #endif
1445 }
1446 EXPORT_SYMBOL(mq_change_real_num_tx);
1447
1448 int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1449 {
1450         bool up = dev->flags & IFF_UP;
1451         unsigned int i;
1452         int ret = 0;
1453
1454         if (up)
1455                 dev_deactivate(dev);
1456
1457         for (i = 0; i < dev->num_tx_queues; i++) {
1458                 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1459
1460                 /* TODO: revert changes on a partial failure */
1461                 if (ret)
1462                         break;
1463         }
1464
1465         if (up)
1466                 dev_activate(dev);
1467         return ret;
1468 }
1469
1470 static void dev_init_scheduler_queue(struct net_device *dev,
1471                                      struct netdev_queue *dev_queue,
1472                                      void *_qdisc)
1473 {
1474         struct Qdisc *qdisc = _qdisc;
1475
1476         rcu_assign_pointer(dev_queue->qdisc, qdisc);
1477         rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc);
1478 }
1479
1480 void dev_init_scheduler(struct net_device *dev)
1481 {
1482         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1483         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1484         if (dev_ingress_queue(dev))
1485                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1486
1487         timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1488 }
1489
1490 void dev_shutdown(struct net_device *dev)
1491 {
1492         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1493         if (dev_ingress_queue(dev))
1494                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1495         qdisc_put(rtnl_dereference(dev->qdisc));
1496         rcu_assign_pointer(dev->qdisc, &noop_qdisc);
1497
1498         WARN_ON(timer_pending(&dev->watchdog_timer));
1499 }
1500
1501 /**
1502  * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division
1503  * @rate:   Rate to compute reciprocal division values of
1504  * @mult:   Multiplier for reciprocal division
1505  * @shift:  Shift for reciprocal division
1506  *
1507  * The multiplier and shift for reciprocal division by rate are stored
1508  * in mult and shift.
1509  *
1510  * The deal here is to replace a divide by a reciprocal one
1511  * in fast path (a reciprocal divide is a multiply and a shift)
1512  *
1513  * Normal formula would be :
1514  *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1515  *
1516  * We compute mult/shift to use instead :
1517  *  time_in_ns = (len * mult) >> shift;
1518  *
1519  * We try to get the highest possible mult value for accuracy,
1520  * but have to make sure no overflows will ever happen.
1521  *
1522  * reciprocal_value() is not used here it doesn't handle 64-bit values.
1523  */
1524 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift)
1525 {
1526         u64 factor = NSEC_PER_SEC;
1527
1528         *mult = 1;
1529         *shift = 0;
1530
1531         if (rate <= 0)
1532                 return;
1533
1534         for (;;) {
1535                 *mult = div64_u64(factor, rate);
1536                 if (*mult & (1U << 31) || factor & (1ULL << 63))
1537                         break;
1538                 factor <<= 1;
1539                 (*shift)++;
1540         }
1541 }
1542
1543 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1544                                const struct tc_ratespec *conf,
1545                                u64 rate64)
1546 {
1547         memset(r, 0, sizeof(*r));
1548         r->overhead = conf->overhead;
1549         r->mpu = conf->mpu;
1550         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1551         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1552         psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift);
1553 }
1554 EXPORT_SYMBOL(psched_ratecfg_precompute);
1555
1556 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64)
1557 {
1558         r->rate_pkts_ps = pktrate64;
1559         psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift);
1560 }
1561 EXPORT_SYMBOL(psched_ppscfg_precompute);
1562
1563 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1564                           struct tcf_proto *tp_head)
1565 {
1566         /* Protected with chain0->filter_chain_lock.
1567          * Can't access chain directly because tp_head can be NULL.
1568          */
1569         struct mini_Qdisc *miniq_old =
1570                 rcu_dereference_protected(*miniqp->p_miniq, 1);
1571         struct mini_Qdisc *miniq;
1572
1573         if (!tp_head) {
1574                 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1575         } else {
1576                 miniq = miniq_old != &miniqp->miniq1 ?
1577                         &miniqp->miniq1 : &miniqp->miniq2;
1578
1579                 /* We need to make sure that readers won't see the miniq
1580                  * we are about to modify. So ensure that at least one RCU
1581                  * grace period has elapsed since the miniq was made
1582                  * inactive.
1583                  */
1584                 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1585                         cond_synchronize_rcu(miniq->rcu_state);
1586                 else if (!poll_state_synchronize_rcu(miniq->rcu_state))
1587                         synchronize_rcu_expedited();
1588
1589                 miniq->filter_list = tp_head;
1590                 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1591         }
1592
1593         if (miniq_old)
1594                 /* This is counterpart of the rcu sync above. We need to
1595                  * block potential new user of miniq_old until all readers
1596                  * are not seeing it.
1597                  */
1598                 miniq_old->rcu_state = start_poll_synchronize_rcu();
1599 }
1600 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1601
1602 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
1603                                 struct tcf_block *block)
1604 {
1605         miniqp->miniq1.block = block;
1606         miniqp->miniq2.block = block;
1607 }
1608 EXPORT_SYMBOL(mini_qdisc_pair_block_init);
1609
1610 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1611                           struct mini_Qdisc __rcu **p_miniq)
1612 {
1613         miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1614         miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1615         miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1616         miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1617         miniqp->miniq1.rcu_state = get_state_synchronize_rcu();
1618         miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state;
1619         miniqp->p_miniq = p_miniq;
1620 }
1621 EXPORT_SYMBOL(mini_qdisc_pair_init);