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