net/sched/sch_generic.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/sch_generic.c      Generic packet scheduler routines.
   4  *
   5  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   6  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
   7  *              - Ingress support
   8  */
   9
  10 #include <linux/bitops.h>
  11 #include <linux/module.h>
  12 #include <linux/types.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/string.h>
  16 #include <linux/errno.h>
  17 #include <linux/netdevice.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/rtnetlink.h>
  20 #include <linux/init.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/list.h>
  23 #include <linux/slab.h>
  24 #include <linux/if_vlan.h>
  25 #include <linux/skb_array.h>
  26 #include <linux/if_macvlan.h>
  27 #include <net/sch_generic.h>
  28 #include <net/pkt_sched.h>
  29 #include <net/dst.h>
  30 #include <trace/events/qdisc.h>
  31 #include <trace/events/net.h>
  32 #include <net/xfrm.h>
  33
  34 /* Qdisc to use by default */
  35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  36 EXPORT_SYMBOL(default_qdisc_ops);
  37
  38 static void qdisc_maybe_clear_missed(struct Qdisc *q,
  39                                      const struct netdev_queue *txq)
  40 {
  41         clear_bit(__QDISC_STATE_MISSED, &q->state);
  42
  43         /* Make sure the below netif_xmit_frozen_or_stopped()
  44          * checking happens after clearing STATE_MISSED.
  45          */
  46         smp_mb__after_atomic();
  47
  48         /* Checking netif_xmit_frozen_or_stopped() again to
  49          * make sure STATE_MISSED is set if the STATE_MISSED
  50          * set by netif_tx_wake_queue()'s rescheduling of
  51          * net_tx_action() is cleared by the above clear_bit().
  52          */
  53         if (!netif_xmit_frozen_or_stopped(txq))
  54                 set_bit(__QDISC_STATE_MISSED, &q->state);
  55         else
  56                 set_bit(__QDISC_STATE_DRAINING, &q->state);
  57 }
  58
  59 /* Main transmission queue. */
  60
  61 /* Modifications to data participating in scheduling must be protected with
  62  * qdisc_lock(qdisc) spinlock.
  63  *
  64  * The idea is the following:
  65  * - enqueue, dequeue are serialized via qdisc root lock
  66  * - ingress filtering is also serialized via qdisc root lock
  67  * - updates to tree and tree walking are only done under the rtnl mutex.
  68  */
  69
  70 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  71
  72 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  73 {
  74         const struct netdev_queue *txq = q->dev_queue;
  75         spinlock_t *lock = NULL;
  76         struct sk_buff *skb;
  77
  78         if (q->flags & TCQ_F_NOLOCK) {
  79                 lock = qdisc_lock(q);
  80                 spin_lock(lock);
  81         }
  82
  83         skb = skb_peek(&q->skb_bad_txq);
  84         if (skb) {
  85                 /* check the reason of requeuing without tx lock first */
  86                 txq = skb_get_tx_queue(txq->dev, skb);
  87                 if (!netif_xmit_frozen_or_stopped(txq)) {
  88                         skb = __skb_dequeue(&q->skb_bad_txq);
  89                         if (qdisc_is_percpu_stats(q)) {
  90                                 qdisc_qstats_cpu_backlog_dec(q, skb);
  91                                 qdisc_qstats_cpu_qlen_dec(q);
  92                         } else {
  93                                 qdisc_qstats_backlog_dec(q, skb);
  94                                 q->q.qlen--;
  95                         }
  96                 } else {
  97                         skb = SKB_XOFF_MAGIC;
  98                         qdisc_maybe_clear_missed(q, txq);
  99                 }
 100         }
 101
 102         if (lock)
 103                 spin_unlock(lock);
 104
 105         return skb;
 106 }
 107
 108 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
 109 {
 110         struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
 111
 112         if (unlikely(skb))
 113                 skb = __skb_dequeue_bad_txq(q);
 114
 115         return skb;
 116 }
 117
 118 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
 119                                              struct sk_buff *skb)
 120 {
 121         spinlock_t *lock = NULL;
 122
 123         if (q->flags & TCQ_F_NOLOCK) {
 124                 lock = qdisc_lock(q);
 125                 spin_lock(lock);
 126         }
 127
 128         __skb_queue_tail(&q->skb_bad_txq, skb);
 129
 130         if (qdisc_is_percpu_stats(q)) {
 131                 qdisc_qstats_cpu_backlog_inc(q, skb);
 132                 qdisc_qstats_cpu_qlen_inc(q);
 133         } else {
 134                 qdisc_qstats_backlog_inc(q, skb);
 135                 q->q.qlen++;
 136         }
 137
 138         if (lock)
 139                 spin_unlock(lock);
 140 }
 141
 142 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 143 {
 144         spinlock_t *lock = NULL;
 145
 146         if (q->flags & TCQ_F_NOLOCK) {
 147                 lock = qdisc_lock(q);
 148                 spin_lock(lock);
 149         }
 150
 151         while (skb) {
 152                 struct sk_buff *next = skb->next;
 153
 154                 __skb_queue_tail(&q->gso_skb, skb);
 155
 156                 /* it's still part of the queue */
 157                 if (qdisc_is_percpu_stats(q)) {
 158                         qdisc_qstats_cpu_requeues_inc(q);
 159                         qdisc_qstats_cpu_backlog_inc(q, skb);
 160                         qdisc_qstats_cpu_qlen_inc(q);
 161                 } else {
 162                         q->qstats.requeues++;
 163                         qdisc_qstats_backlog_inc(q, skb);
 164                         q->q.qlen++;
 165                 }
 166
 167                 skb = next;
 168         }
 169
 170         if (lock) {
 171                 spin_unlock(lock);
 172                 set_bit(__QDISC_STATE_MISSED, &q->state);
 173         } else {
 174                 __netif_schedule(q);
 175         }
 176 }
 177
 178 static void try_bulk_dequeue_skb(struct Qdisc *q,
 179                                  struct sk_buff *skb,
 180                                  const struct netdev_queue *txq,
 181                                  int *packets)
 182 {
 183         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 184
 185         while (bytelimit > 0) {
 186                 struct sk_buff *nskb = q->dequeue(q);
 187
 188                 if (!nskb)
 189                         break;
 190
 191                 bytelimit -= nskb->len; /* covers GSO len */
 192                 skb->next = nskb;
 193                 skb = nskb;
 194                 (*packets)++; /* GSO counts as one pkt */
 195         }
 196         skb_mark_not_on_list(skb);
 197 }
 198
 199 /* This variant of try_bulk_dequeue_skb() makes sure
 200  * all skbs in the chain are for the same txq
 201  */
 202 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 203                                       struct sk_buff *skb,
 204                                       int *packets)
 205 {
 206         int mapping = skb_get_queue_mapping(skb);
 207         struct sk_buff *nskb;
 208         int cnt = 0;
 209
 210         do {
 211                 nskb = q->dequeue(q);
 212                 if (!nskb)
 213                         break;
 214                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 215                         qdisc_enqueue_skb_bad_txq(q, nskb);
 216                         break;
 217                 }
 218                 skb->next = nskb;
 219                 skb = nskb;
 220         } while (++cnt < 8);
 221         (*packets) += cnt;
 222         skb_mark_not_on_list(skb);
 223 }
 224
 225 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 226  * A requeued skb (via q->gso_skb) can also be a SKB list.
 227  */
 228 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 229                                    int *packets)
 230 {
 231         const struct netdev_queue *txq = q->dev_queue;
 232         struct sk_buff *skb = NULL;
 233
 234         *packets = 1;
 235         if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 236                 spinlock_t *lock = NULL;
 237
 238                 if (q->flags & TCQ_F_NOLOCK) {
 239                         lock = qdisc_lock(q);
 240                         spin_lock(lock);
 241                 }
 242
 243                 skb = skb_peek(&q->gso_skb);
 244
 245                 /* skb may be null if another cpu pulls gso_skb off in between
 246                  * empty check and lock.
 247                  */
 248                 if (!skb) {
 249                         if (lock)
 250                                 spin_unlock(lock);
 251                         goto validate;
 252                 }
 253
 254                 /* skb in gso_skb were already validated */
 255                 *validate = false;
 256                 if (xfrm_offload(skb))
 257                         *validate = true;
 258                 /* check the reason of requeuing without tx lock first */
 259                 txq = skb_get_tx_queue(txq->dev, skb);
 260                 if (!netif_xmit_frozen_or_stopped(txq)) {
 261                         skb = __skb_dequeue(&q->gso_skb);
 262                         if (qdisc_is_percpu_stats(q)) {
 263                                 qdisc_qstats_cpu_backlog_dec(q, skb);
 264                                 qdisc_qstats_cpu_qlen_dec(q);
 265                         } else {
 266                                 qdisc_qstats_backlog_dec(q, skb);
 267                                 q->q.qlen--;
 268                         }
 269                 } else {
 270                         skb = NULL;
 271                         qdisc_maybe_clear_missed(q, txq);
 272                 }
 273                 if (lock)
 274                         spin_unlock(lock);
 275                 goto trace;
 276         }
 277 validate:
 278         *validate = true;
 279
 280         if ((q->flags & TCQ_F_ONETXQUEUE) &&
 281             netif_xmit_frozen_or_stopped(txq)) {
 282                 qdisc_maybe_clear_missed(q, txq);
 283                 return skb;
 284         }
 285
 286         skb = qdisc_dequeue_skb_bad_txq(q);
 287         if (unlikely(skb)) {
 288                 if (skb == SKB_XOFF_MAGIC)
 289                         return NULL;
 290                 goto bulk;
 291         }
 292         skb = q->dequeue(q);
 293         if (skb) {
 294 bulk:
 295                 if (qdisc_may_bulk(q))
 296                         try_bulk_dequeue_skb(q, skb, txq, packets);
 297                 else
 298                         try_bulk_dequeue_skb_slow(q, skb, packets);
 299         }
 300 trace:
 301         trace_qdisc_dequeue(q, txq, *packets, skb);
 302         return skb;
 303 }
 304
 305 /*
 306  * Transmit possibly several skbs, and handle the return status as
 307  * required. Owning qdisc running bit guarantees that only one CPU
 308  * can execute this function.
 309  *
 310  * Returns to the caller:
 311  *                              false  - hardware queue frozen backoff
 312  *                              true   - feel free to send more pkts
 313  */
 314 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 315                      struct net_device *dev, struct netdev_queue *txq,
 316                      spinlock_t *root_lock, bool validate)
 317 {
 318         int ret = NETDEV_TX_BUSY;
 319         bool again = false;
 320
 321         /* And release qdisc */
 322         if (root_lock)
 323                 spin_unlock(root_lock);
 324
 325         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 326         if (validate)
 327                 skb = validate_xmit_skb_list(skb, dev, &again);
 328
 329 #ifdef CONFIG_XFRM_OFFLOAD
 330         if (unlikely(again)) {
 331                 if (root_lock)
 332                         spin_lock(root_lock);
 333
 334                 dev_requeue_skb(skb, q);
 335                 return false;
 336         }
 337 #endif
 338
 339         if (likely(skb)) {
 340                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 341                 if (!netif_xmit_frozen_or_stopped(txq))
 342                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 343                 else
 344                         qdisc_maybe_clear_missed(q, txq);
 345
 346                 HARD_TX_UNLOCK(dev, txq);
 347         } else {
 348                 if (root_lock)
 349                         spin_lock(root_lock);
 350                 return true;
 351         }
 352
 353         if (root_lock)
 354                 spin_lock(root_lock);
 355
 356         if (!dev_xmit_complete(ret)) {
 357                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 358                 if (unlikely(ret != NETDEV_TX_BUSY))
 359                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 360                                              dev->name, ret, q->q.qlen);
 361
 362                 dev_requeue_skb(skb, q);
 363                 return false;
 364         }
 365
 366         return true;
 367 }
 368
 369 /*
 370  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 371  *
 372  * running seqcount guarantees only one CPU can process
 373  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 374  * this queue.
 375  *
 376  *  netif_tx_lock serializes accesses to device driver.
 377  *
 378  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 379  *  if one is grabbed, another must be free.
 380  *
 381  * Note, that this procedure can be called by a watchdog timer
 382  *
 383  * Returns to the caller:
 384  *                              0  - queue is empty or throttled.
 385  *                              >0 - queue is not empty.
 386  *
 387  */
 388 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 389 {
 390         spinlock_t *root_lock = NULL;
 391         struct netdev_queue *txq;
 392         struct net_device *dev;
 393         struct sk_buff *skb;
 394         bool validate;
 395
 396         /* Dequeue packet */
 397         skb = dequeue_skb(q, &validate, packets);
 398         if (unlikely(!skb))
 399                 return false;
 400
 401         if (!(q->flags & TCQ_F_NOLOCK))
 402                 root_lock = qdisc_lock(q);
 403
 404         dev = qdisc_dev(q);
 405         txq = skb_get_tx_queue(dev, skb);
 406
 407         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 408 }
 409
 410 void __qdisc_run(struct Qdisc *q)
 411 {
 412         int quota = READ_ONCE(dev_tx_weight);
 413         int packets;
 414
 415         while (qdisc_restart(q, &packets)) {
 416                 quota -= packets;
 417                 if (quota <= 0) {
 418                         if (q->flags & TCQ_F_NOLOCK)
 419                                 set_bit(__QDISC_STATE_MISSED, &q->state);
 420                         else
 421                                 __netif_schedule(q);
 422
 423                         break;
 424                 }
 425         }
 426 }
 427
 428 unsigned long dev_trans_start(struct net_device *dev)
 429 {
 430         unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start);
 431         unsigned long val;
 432         unsigned int i;
 433
 434         for (i = 1; i < dev->num_tx_queues; i++) {
 435                 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start);
 436                 if (val && time_after(val, res))
 437                         res = val;
 438         }
 439
 440         return res;
 441 }
 442 EXPORT_SYMBOL(dev_trans_start);
 443
 444 static void netif_freeze_queues(struct net_device *dev)
 445 {
 446         unsigned int i;
 447         int cpu;
 448
 449         cpu = smp_processor_id();
 450         for (i = 0; i < dev->num_tx_queues; i++) {
 451                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 452
 453                 /* We are the only thread of execution doing a
 454                  * freeze, but we have to grab the _xmit_lock in
 455                  * order to synchronize with threads which are in
 456                  * the ->hard_start_xmit() handler and already
 457                  * checked the frozen bit.
 458                  */
 459                 __netif_tx_lock(txq, cpu);
 460                 set_bit(__QUEUE_STATE_FROZEN, &txq->state);
 461                 __netif_tx_unlock(txq);
 462         }
 463 }
 464
 465 void netif_tx_lock(struct net_device *dev)
 466 {
 467         spin_lock(&dev->tx_global_lock);
 468         netif_freeze_queues(dev);
 469 }
 470 EXPORT_SYMBOL(netif_tx_lock);
 471
 472 static void netif_unfreeze_queues(struct net_device *dev)
 473 {
 474         unsigned int i;
 475
 476         for (i = 0; i < dev->num_tx_queues; i++) {
 477                 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
 478
 479                 /* No need to grab the _xmit_lock here.  If the
 480                  * queue is not stopped for another reason, we
 481                  * force a schedule.
 482                  */
 483                 clear_bit(__QUEUE_STATE_FROZEN, &txq->state);
 484                 netif_schedule_queue(txq);
 485         }
 486 }
 487
 488 void netif_tx_unlock(struct net_device *dev)
 489 {
 490         netif_unfreeze_queues(dev);
 491         spin_unlock(&dev->tx_global_lock);
 492 }
 493 EXPORT_SYMBOL(netif_tx_unlock);
 494
 495 static void dev_watchdog(struct timer_list *t)
 496 {
 497         struct net_device *dev = from_timer(dev, t, watchdog_timer);
 498         bool release = true;
 499
 500         spin_lock(&dev->tx_global_lock);
 501         if (!qdisc_tx_is_noop(dev)) {
 502                 if (netif_device_present(dev) &&
 503                     netif_running(dev) &&
 504                     netif_carrier_ok(dev)) {
 505                         unsigned int timedout_ms = 0;
 506                         unsigned int i;
 507                         unsigned long trans_start;
 508
 509                         for (i = 0; i < dev->num_tx_queues; i++) {
 510                                 struct netdev_queue *txq;
 511
 512                                 txq = netdev_get_tx_queue(dev, i);
 513                                 trans_start = READ_ONCE(txq->trans_start);
 514                                 if (netif_xmit_stopped(txq) &&
 515                                     time_after(jiffies, (trans_start +
 516                                                          dev->watchdog_timeo))) {
 517                                         timedout_ms = jiffies_to_msecs(jiffies - trans_start);
 518                                         atomic_long_inc(&txq->trans_timeout);
 519                                         break;
 520                                 }
 521                         }
 522
 523                         if (unlikely(timedout_ms)) {
 524                                 trace_net_dev_xmit_timeout(dev, i);
 525                                 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
 526                                             raw_smp_processor_id(),
 527                                             i, timedout_ms);
 528                                 netif_freeze_queues(dev);
 529                                 dev->netdev_ops->ndo_tx_timeout(dev, i);
 530                                 netif_unfreeze_queues(dev);
 531                         }
 532                         if (!mod_timer(&dev->watchdog_timer,
 533                                        round_jiffies(jiffies +
 534                                                      dev->watchdog_timeo)))
 535                                 release = false;
 536                 }
 537         }
 538         spin_unlock(&dev->tx_global_lock);
 539
 540         if (release)
 541                 netdev_put(dev, &dev->watchdog_dev_tracker);
 542 }
 543
 544 void __netdev_watchdog_up(struct net_device *dev)
 545 {
 546         if (dev->netdev_ops->ndo_tx_timeout) {
 547                 if (dev->watchdog_timeo <= 0)
 548                         dev->watchdog_timeo = 5*HZ;
 549                 if (!mod_timer(&dev->watchdog_timer,
 550                                round_jiffies(jiffies + dev->watchdog_timeo)))
 551                         netdev_hold(dev, &dev->watchdog_dev_tracker,
 552                                     GFP_ATOMIC);
 553         }
 554 }
 555 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 556
 557 static void dev_watchdog_up(struct net_device *dev)
 558 {
 559         __netdev_watchdog_up(dev);
 560 }
 561
 562 static void dev_watchdog_down(struct net_device *dev)
 563 {
 564         netif_tx_lock_bh(dev);
 565         if (del_timer(&dev->watchdog_timer))
 566                 netdev_put(dev, &dev->watchdog_dev_tracker);
 567         netif_tx_unlock_bh(dev);
 568 }
 569
 570 /**
 571  *      netif_carrier_on - set carrier
 572  *      @dev: network device
 573  *
 574  * Device has detected acquisition of carrier.
 575  */
 576 void netif_carrier_on(struct net_device *dev)
 577 {
 578         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 579                 if (dev->reg_state == NETREG_UNINITIALIZED)
 580                         return;
 581                 atomic_inc(&dev->carrier_up_count);
 582                 linkwatch_fire_event(dev);
 583                 if (netif_running(dev))
 584                         __netdev_watchdog_up(dev);
 585         }
 586 }
 587 EXPORT_SYMBOL(netif_carrier_on);
 588
 589 /**
 590  *      netif_carrier_off - clear carrier
 591  *      @dev: network device
 592  *
 593  * Device has detected loss of carrier.
 594  */
 595 void netif_carrier_off(struct net_device *dev)
 596 {
 597         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 598                 if (dev->reg_state == NETREG_UNINITIALIZED)
 599                         return;
 600                 atomic_inc(&dev->carrier_down_count);
 601                 linkwatch_fire_event(dev);
 602         }
 603 }
 604 EXPORT_SYMBOL(netif_carrier_off);
 605
 606 /**
 607  *      netif_carrier_event - report carrier state event
 608  *      @dev: network device
 609  *
 610  * Device has detected a carrier event but the carrier state wasn't changed.
 611  * Use in drivers when querying carrier state asynchronously, to avoid missing
 612  * events (link flaps) if link recovers before it's queried.
 613  */
 614 void netif_carrier_event(struct net_device *dev)
 615 {
 616         if (dev->reg_state == NETREG_UNINITIALIZED)
 617                 return;
 618         atomic_inc(&dev->carrier_up_count);
 619         atomic_inc(&dev->carrier_down_count);
 620         linkwatch_fire_event(dev);
 621 }
 622 EXPORT_SYMBOL_GPL(netif_carrier_event);
 623
 624 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 625    under all circumstances. It is difficult to invent anything faster or
 626    cheaper.
 627  */
 628
 629 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 630                         struct sk_buff **to_free)
 631 {
 632         __qdisc_drop(skb, to_free);
 633         return NET_XMIT_CN;
 634 }
 635
 636 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 637 {
 638         return NULL;
 639 }
 640
 641 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 642         .id             =       "noop",
 643         .priv_size      =       0,
 644         .enqueue        =       noop_enqueue,
 645         .dequeue        =       noop_dequeue,
 646         .peek           =       noop_dequeue,
 647         .owner          =       THIS_MODULE,
 648 };
 649
 650 static struct netdev_queue noop_netdev_queue = {
 651         RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 652         RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc),
 653 };
 654
 655 struct Qdisc noop_qdisc = {
 656         .enqueue        =       noop_enqueue,
 657         .dequeue        =       noop_dequeue,
 658         .flags          =       TCQ_F_BUILTIN,
 659         .ops            =       &noop_qdisc_ops,
 660         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 661         .dev_queue      =       &noop_netdev_queue,
 662         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 663         .gso_skb = {
 664                 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
 665                 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 666                 .qlen = 0,
 667                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 668         },
 669         .skb_bad_txq = {
 670                 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 671                 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 672                 .qlen = 0,
 673                 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 674         },
 675 };
 676 EXPORT_SYMBOL(noop_qdisc);
 677
 678 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 679                         struct netlink_ext_ack *extack)
 680 {
 681         /* register_qdisc() assigns a default of noop_enqueue if unset,
 682          * but __dev_queue_xmit() treats noqueue only as such
 683          * if this is NULL - so clear it here. */
 684         qdisc->enqueue = NULL;
 685         return 0;
 686 }
 687
 688 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 689         .id             =       "noqueue",
 690         .priv_size      =       0,
 691         .init           =       noqueue_init,
 692         .enqueue        =       noop_enqueue,
 693         .dequeue        =       noop_dequeue,
 694         .peek           =       noop_dequeue,
 695         .owner          =       THIS_MODULE,
 696 };
 697
 698 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = {
 699         1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1
 700 };
 701 EXPORT_SYMBOL(sch_default_prio2band);
 702
 703 /* 3-band FIFO queue: old style, but should be a bit faster than
 704    generic prio+fifo combination.
 705  */
 706
 707 #define PFIFO_FAST_BANDS 3
 708
 709 /*
 710  * Private data for a pfifo_fast scheduler containing:
 711  *      - rings for priority bands
 712  */
 713 struct pfifo_fast_priv {
 714         struct skb_array q[PFIFO_FAST_BANDS];
 715 };
 716
 717 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 718                                           int band)
 719 {
 720         return &priv->q[band];
 721 }
 722
 723 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 724                               struct sk_buff **to_free)
 725 {
 726         int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX];
 727         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 728         struct skb_array *q = band2list(priv, band);
 729         unsigned int pkt_len = qdisc_pkt_len(skb);
 730         int err;
 731
 732         err = skb_array_produce(q, skb);
 733
 734         if (unlikely(err)) {
 735                 if (qdisc_is_percpu_stats(qdisc))
 736                         return qdisc_drop_cpu(skb, qdisc, to_free);
 737                 else
 738                         return qdisc_drop(skb, qdisc, to_free);
 739         }
 740
 741         qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 742         return NET_XMIT_SUCCESS;
 743 }
 744
 745 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 746 {
 747         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 748         struct sk_buff *skb = NULL;
 749         bool need_retry = true;
 750         int band;
 751
 752 retry:
 753         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 754                 struct skb_array *q = band2list(priv, band);
 755
 756                 if (__skb_array_empty(q))
 757                         continue;
 758
 759                 skb = __skb_array_consume(q);
 760         }
 761         if (likely(skb)) {
 762                 qdisc_update_stats_at_dequeue(qdisc, skb);
 763         } else if (need_retry &&
 764                    READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) {
 765                 /* Delay clearing the STATE_MISSED here to reduce
 766                  * the overhead of the second spin_trylock() in
 767                  * qdisc_run_begin() and __netif_schedule() calling
 768                  * in qdisc_run_end().
 769                  */
 770                 clear_bit(__QDISC_STATE_MISSED, &qdisc->state);
 771                 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state);
 772
 773                 /* Make sure dequeuing happens after clearing
 774                  * STATE_MISSED.
 775                  */
 776                 smp_mb__after_atomic();
 777
 778                 need_retry = false;
 779
 780                 goto retry;
 781         }
 782
 783         return skb;
 784 }
 785
 786 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 787 {
 788         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 789         struct sk_buff *skb = NULL;
 790         int band;
 791
 792         for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 793                 struct skb_array *q = band2list(priv, band);
 794
 795                 skb = __skb_array_peek(q);
 796         }
 797
 798         return skb;
 799 }
 800
 801 static void pfifo_fast_reset(struct Qdisc *qdisc)
 802 {
 803         int i, band;
 804         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 805
 806         for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 807                 struct skb_array *q = band2list(priv, band);
 808                 struct sk_buff *skb;
 809
 810                 /* NULL ring is possible if destroy path is due to a failed
 811                  * skb_array_init() in pfifo_fast_init() case.
 812                  */
 813                 if (!q->ring.queue)
 814                         continue;
 815
 816                 while ((skb = __skb_array_consume(q)) != NULL)
 817                         kfree_skb(skb);
 818         }
 819
 820         if (qdisc_is_percpu_stats(qdisc)) {
 821                 for_each_possible_cpu(i) {
 822                         struct gnet_stats_queue *q;
 823
 824                         q = per_cpu_ptr(qdisc->cpu_qstats, i);
 825                         q->backlog = 0;
 826                         q->qlen = 0;
 827                 }
 828         }
 829 }
 830
 831 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 832 {
 833         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 834
 835         memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1);
 836         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 837                 goto nla_put_failure;
 838         return skb->len;
 839
 840 nla_put_failure:
 841         return -1;
 842 }
 843
 844 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 845                            struct netlink_ext_ack *extack)
 846 {
 847         unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 848         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 849         int prio;
 850
 851         /* guard against zero length rings */
 852         if (!qlen)
 853                 return -EINVAL;
 854
 855         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 856                 struct skb_array *q = band2list(priv, prio);
 857                 int err;
 858
 859                 err = skb_array_init(q, qlen, GFP_KERNEL);
 860                 if (err)
 861                         return -ENOMEM;
 862         }
 863
 864         /* Can by-pass the queue discipline */
 865         qdisc->flags |= TCQ_F_CAN_BYPASS;
 866         return 0;
 867 }
 868
 869 static void pfifo_fast_destroy(struct Qdisc *sch)
 870 {
 871         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 872         int prio;
 873
 874         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 875                 struct skb_array *q = band2list(priv, prio);
 876
 877                 /* NULL ring is possible if destroy path is due to a failed
 878                  * skb_array_init() in pfifo_fast_init() case.
 879                  */
 880                 if (!q->ring.queue)
 881                         continue;
 882                 /* Destroy ring but no need to kfree_skb because a call to
 883                  * pfifo_fast_reset() has already done that work.
 884                  */
 885                 ptr_ring_cleanup(&q->ring, NULL);
 886         }
 887 }
 888
 889 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 890                                           unsigned int new_len)
 891 {
 892         struct pfifo_fast_priv *priv = qdisc_priv(sch);
 893         struct skb_array *bands[PFIFO_FAST_BANDS];
 894         int prio;
 895
 896         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 897                 struct skb_array *q = band2list(priv, prio);
 898
 899                 bands[prio] = q;
 900         }
 901
 902         return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 903                                          GFP_KERNEL);
 904 }
 905
 906 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 907         .id             =       "pfifo_fast",
 908         .priv_size      =       sizeof(struct pfifo_fast_priv),
 909         .enqueue        =       pfifo_fast_enqueue,
 910         .dequeue        =       pfifo_fast_dequeue,
 911         .peek           =       pfifo_fast_peek,
 912         .init           =       pfifo_fast_init,
 913         .destroy        =       pfifo_fast_destroy,
 914         .reset          =       pfifo_fast_reset,
 915         .dump           =       pfifo_fast_dump,
 916         .change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 917         .owner          =       THIS_MODULE,
 918         .static_flags   =       TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 919 };
 920 EXPORT_SYMBOL(pfifo_fast_ops);
 921
 922 static struct lock_class_key qdisc_tx_busylock;
 923
 924 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 925                           const struct Qdisc_ops *ops,
 926                           struct netlink_ext_ack *extack)
 927 {
 928         struct Qdisc *sch;
 929         unsigned int size = sizeof(*sch) + ops->priv_size;
 930         int err = -ENOBUFS;
 931         struct net_device *dev;
 932
 933         if (!dev_queue) {
 934                 NL_SET_ERR_MSG(extack, "No device queue given");
 935                 err = -EINVAL;
 936                 goto errout;
 937         }
 938
 939         dev = dev_queue->dev;
 940         sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue));
 941
 942         if (!sch)
 943                 goto errout;
 944         __skb_queue_head_init(&sch->gso_skb);
 945         __skb_queue_head_init(&sch->skb_bad_txq);
 946         gnet_stats_basic_sync_init(&sch->bstats);
 947         spin_lock_init(&sch->q.lock);
 948
 949         if (ops->static_flags & TCQ_F_CPUSTATS) {
 950                 sch->cpu_bstats =
 951                         netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync);
 952                 if (!sch->cpu_bstats)
 953                         goto errout1;
 954
 955                 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 956                 if (!sch->cpu_qstats) {
 957                         free_percpu(sch->cpu_bstats);
 958                         goto errout1;
 959                 }
 960         }
 961
 962         spin_lock_init(&sch->busylock);
 963         lockdep_set_class(&sch->busylock,
 964                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 965
 966         /* seqlock has the same scope of busylock, for NOLOCK qdisc */
 967         spin_lock_init(&sch->seqlock);
 968         lockdep_set_class(&sch->seqlock,
 969                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 970
 971         sch->ops = ops;
 972         sch->flags = ops->static_flags;
 973         sch->enqueue = ops->enqueue;
 974         sch->dequeue = ops->dequeue;
 975         sch->dev_queue = dev_queue;
 976         sch->owner = -1;
 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);