net/sched/sch_generic.c

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