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