net/sched/sch_generic.c

   1 /*
   2  * net/sched/sch_generic.c      Generic packet scheduler routines.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
  11  *              - Ingress support
  12  */
  13
  14 #include <linux/bitops.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/string.h>
  20 #include <linux/errno.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/rtnetlink.h>
  24 #include <linux/init.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/list.h>
  27 #include <linux/slab.h>
  28 #include <linux/if_vlan.h>
  29 #include <net/sch_generic.h>
  30 #include <net/pkt_sched.h>
  31 #include <net/dst.h>
  32 #include <trace/events/qdisc.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 int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  50 {
  51         q->gso_skb = skb;
  52         q->qstats.requeues++;
  53         qdisc_qstats_backlog_inc(q, skb);
  54         q->q.qlen++;    /* it's still part of the queue */
  55         __netif_schedule(q);
  56
  57         return 0;
  58 }
  59
  60 static void try_bulk_dequeue_skb(struct Qdisc *q,
  61                                  struct sk_buff *skb,
  62                                  const struct netdev_queue *txq,
  63                                  int *packets)
  64 {
  65         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  66
  67         while (bytelimit > 0) {
  68                 struct sk_buff *nskb = q->dequeue(q);
  69
  70                 if (!nskb)
  71                         break;
  72
  73                 bytelimit -= nskb->len; /* covers GSO len */
  74                 skb->next = nskb;
  75                 skb = nskb;
  76                 (*packets)++; /* GSO counts as one pkt */
  77         }
  78         skb->next = NULL;
  79 }
  80
  81 /* This variant of try_bulk_dequeue_skb() makes sure
  82  * all skbs in the chain are for the same txq
  83  */
  84 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
  85                                       struct sk_buff *skb,
  86                                       int *packets)
  87 {
  88         int mapping = skb_get_queue_mapping(skb);
  89         struct sk_buff *nskb;
  90         int cnt = 0;
  91
  92         do {
  93                 nskb = q->dequeue(q);
  94                 if (!nskb)
  95                         break;
  96                 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
  97                         q->skb_bad_txq = nskb;
  98                         qdisc_qstats_backlog_inc(q, nskb);
  99                         q->q.qlen++;
 100                         break;
 101                 }
 102                 skb->next = nskb;
 103                 skb = nskb;
 104         } while (++cnt < 8);
 105         (*packets) += cnt;
 106         skb->next = NULL;
 107 }
 108
 109 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 110  * A requeued skb (via q->gso_skb) can also be a SKB list.
 111  */
 112 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 113                                    int *packets)
 114 {
 115         struct sk_buff *skb = q->gso_skb;
 116         const struct netdev_queue *txq = q->dev_queue;
 117
 118         *packets = 1;
 119         if (unlikely(skb)) {
 120                 /* skb in gso_skb were already validated */
 121                 *validate = false;
 122                 /* check the reason of requeuing without tx lock first */
 123                 txq = skb_get_tx_queue(txq->dev, skb);
 124                 if (!netif_xmit_frozen_or_stopped(txq)) {
 125                         q->gso_skb = NULL;
 126                         qdisc_qstats_backlog_dec(q, skb);
 127                         q->q.qlen--;
 128                 } else
 129                         skb = NULL;
 130                 goto trace;
 131         }
 132         *validate = true;
 133         skb = q->skb_bad_txq;
 134         if (unlikely(skb)) {
 135                 /* check the reason of requeuing without tx lock first */
 136                 txq = skb_get_tx_queue(txq->dev, skb);
 137                 if (!netif_xmit_frozen_or_stopped(txq)) {
 138                         q->skb_bad_txq = NULL;
 139                         qdisc_qstats_backlog_dec(q, skb);
 140                         q->q.qlen--;
 141                         goto bulk;
 142                 }
 143                 skb = NULL;
 144                 goto trace;
 145         }
 146         if (!(q->flags & TCQ_F_ONETXQUEUE) ||
 147             !netif_xmit_frozen_or_stopped(txq))
 148                 skb = q->dequeue(q);
 149         if (skb) {
 150 bulk:
 151                 if (qdisc_may_bulk(q))
 152                         try_bulk_dequeue_skb(q, skb, txq, packets);
 153                 else
 154                         try_bulk_dequeue_skb_slow(q, skb, packets);
 155         }
 156 trace:
 157         trace_qdisc_dequeue(q, txq, *packets, skb);
 158         return skb;
 159 }
 160
 161 /*
 162  * Transmit possibly several skbs, and handle the return status as
 163  * required. Owning running seqcount bit guarantees that
 164  * only one CPU can execute this function.
 165  *
 166  * Returns to the caller:
 167  *                              0  - queue is empty or throttled.
 168  *                              >0 - queue is not empty.
 169  */
 170 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 171                     struct net_device *dev, struct netdev_queue *txq,
 172                     spinlock_t *root_lock, bool validate)
 173 {
 174         int ret = NETDEV_TX_BUSY;
 175
 176         /* And release qdisc */
 177         spin_unlock(root_lock);
 178
 179         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 180         if (validate)
 181                 skb = validate_xmit_skb_list(skb, dev);
 182
 183         if (likely(skb)) {
 184                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 185                 if (!netif_xmit_frozen_or_stopped(txq))
 186                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 187
 188                 HARD_TX_UNLOCK(dev, txq);
 189         } else {
 190                 spin_lock(root_lock);
 191                 return qdisc_qlen(q);
 192         }
 193         spin_lock(root_lock);
 194
 195         if (dev_xmit_complete(ret)) {
 196                 /* Driver sent out skb successfully or skb was consumed */
 197                 ret = qdisc_qlen(q);
 198         } else {
 199                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 200                 if (unlikely(ret != NETDEV_TX_BUSY))
 201                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 202                                              dev->name, ret, q->q.qlen);
 203
 204                 ret = dev_requeue_skb(skb, q);
 205         }
 206
 207         if (ret && netif_xmit_frozen_or_stopped(txq))
 208                 ret = 0;
 209
 210         return ret;
 211 }
 212
 213 /*
 214  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 215  *
 216  * running seqcount guarantees only one CPU can process
 217  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 218  * this queue.
 219  *
 220  *  netif_tx_lock serializes accesses to device driver.
 221  *
 222  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 223  *  if one is grabbed, another must be free.
 224  *
 225  * Note, that this procedure can be called by a watchdog timer
 226  *
 227  * Returns to the caller:
 228  *                              0  - queue is empty or throttled.
 229  *                              >0 - queue is not empty.
 230  *
 231  */
 232 static inline int qdisc_restart(struct Qdisc *q, int *packets)
 233 {
 234         struct netdev_queue *txq;
 235         struct net_device *dev;
 236         spinlock_t *root_lock;
 237         struct sk_buff *skb;
 238         bool validate;
 239
 240         /* Dequeue packet */
 241         skb = dequeue_skb(q, &validate, packets);
 242         if (unlikely(!skb))
 243                 return 0;
 244
 245         root_lock = qdisc_lock(q);
 246         dev = qdisc_dev(q);
 247         txq = skb_get_tx_queue(dev, skb);
 248
 249         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 250 }
 251
 252 void __qdisc_run(struct Qdisc *q)
 253 {
 254         int quota = dev_tx_weight;
 255         int packets;
 256
 257         while (qdisc_restart(q, &packets)) {
 258                 /*
 259                  * Ordered by possible occurrence: Postpone processing if
 260                  * 1. we've exceeded packet quota
 261                  * 2. another process needs the CPU;
 262                  */
 263                 quota -= packets;
 264                 if (quota <= 0 || need_resched()) {
 265                         __netif_schedule(q);
 266                         break;
 267                 }
 268         }
 269
 270         qdisc_run_end(q);
 271 }
 272
 273 unsigned long dev_trans_start(struct net_device *dev)
 274 {
 275         unsigned long val, res;
 276         unsigned int i;
 277
 278         if (is_vlan_dev(dev))
 279                 dev = vlan_dev_real_dev(dev);
 280         res = netdev_get_tx_queue(dev, 0)->trans_start;
 281         for (i = 1; i < dev->num_tx_queues; i++) {
 282                 val = netdev_get_tx_queue(dev, i)->trans_start;
 283                 if (val && time_after(val, res))
 284                         res = val;
 285         }
 286
 287         return res;
 288 }
 289 EXPORT_SYMBOL(dev_trans_start);
 290
 291 static void dev_watchdog(unsigned long arg)
 292 {
 293         struct net_device *dev = (struct net_device *)arg;
 294
 295         netif_tx_lock(dev);
 296         if (!qdisc_tx_is_noop(dev)) {
 297                 if (netif_device_present(dev) &&
 298                     netif_running(dev) &&
 299                     netif_carrier_ok(dev)) {
 300                         int some_queue_timedout = 0;
 301                         unsigned int i;
 302                         unsigned long trans_start;
 303
 304                         for (i = 0; i < dev->num_tx_queues; i++) {
 305                                 struct netdev_queue *txq;
 306
 307                                 txq = netdev_get_tx_queue(dev, i);
 308                                 trans_start = txq->trans_start;
 309                                 if (netif_xmit_stopped(txq) &&
 310                                     time_after(jiffies, (trans_start +
 311                                                          dev->watchdog_timeo))) {
 312                                         some_queue_timedout = 1;
 313                                         txq->trans_timeout++;
 314                                         break;
 315                                 }
 316                         }
 317
 318                         if (some_queue_timedout) {
 319                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 320                                        dev->name, netdev_drivername(dev), i);
 321                                 dev->netdev_ops->ndo_tx_timeout(dev);
 322                         }
 323                         if (!mod_timer(&dev->watchdog_timer,
 324                                        round_jiffies(jiffies +
 325                                                      dev->watchdog_timeo)))
 326                                 dev_hold(dev);
 327                 }
 328         }
 329         netif_tx_unlock(dev);
 330
 331         dev_put(dev);
 332 }
 333
 334 void __netdev_watchdog_up(struct net_device *dev)
 335 {
 336         if (dev->netdev_ops->ndo_tx_timeout) {
 337                 if (dev->watchdog_timeo <= 0)
 338                         dev->watchdog_timeo = 5*HZ;
 339                 if (!mod_timer(&dev->watchdog_timer,
 340                                round_jiffies(jiffies + dev->watchdog_timeo)))
 341                         dev_hold(dev);
 342         }
 343 }
 344 EXPORT_SYMBOL_GPL(__netdev_watchdog_up);
 345
 346 static void dev_watchdog_up(struct net_device *dev)
 347 {
 348         __netdev_watchdog_up(dev);
 349 }
 350
 351 static void dev_watchdog_down(struct net_device *dev)
 352 {
 353         netif_tx_lock_bh(dev);
 354         if (del_timer(&dev->watchdog_timer))
 355                 dev_put(dev);
 356         netif_tx_unlock_bh(dev);
 357 }
 358
 359 /**
 360  *      netif_carrier_on - set carrier
 361  *      @dev: network device
 362  *
 363  * Device has detected that carrier.
 364  */
 365 void netif_carrier_on(struct net_device *dev)
 366 {
 367         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 368                 if (dev->reg_state == NETREG_UNINITIALIZED)
 369                         return;
 370                 atomic_inc(&dev->carrier_changes);
 371                 linkwatch_fire_event(dev);
 372                 if (netif_running(dev))
 373                         __netdev_watchdog_up(dev);
 374         }
 375 }
 376 EXPORT_SYMBOL(netif_carrier_on);
 377
 378 /**
 379  *      netif_carrier_off - clear carrier
 380  *      @dev: network device
 381  *
 382  * Device has detected loss of carrier.
 383  */
 384 void netif_carrier_off(struct net_device *dev)
 385 {
 386         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 387                 if (dev->reg_state == NETREG_UNINITIALIZED)
 388                         return;
 389                 atomic_inc(&dev->carrier_changes);
 390                 linkwatch_fire_event(dev);
 391         }
 392 }
 393 EXPORT_SYMBOL(netif_carrier_off);
 394
 395 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 396    under all circumstances. It is difficult to invent anything faster or
 397    cheaper.
 398  */
 399
 400 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 401                         struct sk_buff **to_free)
 402 {
 403         __qdisc_drop(skb, to_free);
 404         return NET_XMIT_CN;
 405 }
 406
 407 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 408 {
 409         return NULL;
 410 }
 411
 412 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 413         .id             =       "noop",
 414         .priv_size      =       0,
 415         .enqueue        =       noop_enqueue,
 416         .dequeue        =       noop_dequeue,
 417         .peek           =       noop_dequeue,
 418         .owner          =       THIS_MODULE,
 419 };
 420
 421 static struct netdev_queue noop_netdev_queue = {
 422         .qdisc          =       &noop_qdisc,
 423         .qdisc_sleeping =       &noop_qdisc,
 424 };
 425
 426 struct Qdisc noop_qdisc = {
 427         .enqueue        =       noop_enqueue,
 428         .dequeue        =       noop_dequeue,
 429         .flags          =       TCQ_F_BUILTIN,
 430         .ops            =       &noop_qdisc_ops,
 431         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 432         .dev_queue      =       &noop_netdev_queue,
 433         .running        =       SEQCNT_ZERO(noop_qdisc.running),
 434         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 435 };
 436 EXPORT_SYMBOL(noop_qdisc);
 437
 438 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt)
 439 {
 440         /* register_qdisc() assigns a default of noop_enqueue if unset,
 441          * but __dev_queue_xmit() treats noqueue only as such
 442          * if this is NULL - so clear it here. */
 443         qdisc->enqueue = NULL;
 444         return 0;
 445 }
 446
 447 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 448         .id             =       "noqueue",
 449         .priv_size      =       0,
 450         .init           =       noqueue_init,
 451         .enqueue        =       noop_enqueue,
 452         .dequeue        =       noop_dequeue,
 453         .peek           =       noop_dequeue,
 454         .owner          =       THIS_MODULE,
 455 };
 456
 457 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 458         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 459 };
 460
 461 /* 3-band FIFO queue: old style, but should be a bit faster than
 462    generic prio+fifo combination.
 463  */
 464
 465 #define PFIFO_FAST_BANDS 3
 466
 467 /*
 468  * Private data for a pfifo_fast scheduler containing:
 469  *      - queues for the three band
 470  *      - bitmap indicating which of the bands contain skbs
 471  */
 472 struct pfifo_fast_priv {
 473         u32 bitmap;
 474         struct qdisc_skb_head q[PFIFO_FAST_BANDS];
 475 };
 476
 477 /*
 478  * Convert a bitmap to the first band number where an skb is queued, where:
 479  *      bitmap=0 means there are no skbs on any band.
 480  *      bitmap=1 means there is an skb on band 0.
 481  *      bitmap=7 means there are skbs on all 3 bands, etc.
 482  */
 483 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 484
 485 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv,
 486                                              int band)
 487 {
 488         return priv->q + band;
 489 }
 490
 491 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 492                               struct sk_buff **to_free)
 493 {
 494         if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) {
 495                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 496                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 497                 struct qdisc_skb_head *list = band2list(priv, band);
 498
 499                 priv->bitmap |= (1 << band);
 500                 qdisc->q.qlen++;
 501                 return __qdisc_enqueue_tail(skb, qdisc, list);
 502         }
 503
 504         return qdisc_drop(skb, qdisc, to_free);
 505 }
 506
 507 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 508 {
 509         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 510         int band = bitmap2band[priv->bitmap];
 511
 512         if (likely(band >= 0)) {
 513                 struct qdisc_skb_head *qh = band2list(priv, band);
 514                 struct sk_buff *skb = __qdisc_dequeue_head(qh);
 515
 516                 if (likely(skb != NULL)) {
 517                         qdisc_qstats_backlog_dec(qdisc, skb);
 518                         qdisc_bstats_update(qdisc, skb);
 519                 }
 520
 521                 qdisc->q.qlen--;
 522                 if (qh->qlen == 0)
 523                         priv->bitmap &= ~(1 << band);
 524
 525                 return skb;
 526         }
 527
 528         return NULL;
 529 }
 530
 531 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 532 {
 533         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 534         int band = bitmap2band[priv->bitmap];
 535
 536         if (band >= 0) {
 537                 struct qdisc_skb_head *qh = band2list(priv, band);
 538
 539                 return qh->head;
 540         }
 541
 542         return NULL;
 543 }
 544
 545 static void pfifo_fast_reset(struct Qdisc *qdisc)
 546 {
 547         int prio;
 548         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 549
 550         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 551                 __qdisc_reset_queue(band2list(priv, prio));
 552
 553         priv->bitmap = 0;
 554         qdisc->qstats.backlog = 0;
 555         qdisc->q.qlen = 0;
 556 }
 557
 558 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 559 {
 560         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 561
 562         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 563         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 564                 goto nla_put_failure;
 565         return skb->len;
 566
 567 nla_put_failure:
 568         return -1;
 569 }
 570
 571 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 572 {
 573         int prio;
 574         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 575
 576         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 577                 qdisc_skb_head_init(band2list(priv, prio));
 578
 579         /* Can by-pass the queue discipline */
 580         qdisc->flags |= TCQ_F_CAN_BYPASS;
 581         return 0;
 582 }
 583
 584 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 585         .id             =       "pfifo_fast",
 586         .priv_size      =       sizeof(struct pfifo_fast_priv),
 587         .enqueue        =       pfifo_fast_enqueue,
 588         .dequeue        =       pfifo_fast_dequeue,
 589         .peek           =       pfifo_fast_peek,
 590         .init           =       pfifo_fast_init,
 591         .reset          =       pfifo_fast_reset,
 592         .dump           =       pfifo_fast_dump,
 593         .owner          =       THIS_MODULE,
 594 };
 595 EXPORT_SYMBOL(pfifo_fast_ops);
 596
 597 static struct lock_class_key qdisc_tx_busylock;
 598 static struct lock_class_key qdisc_running_key;
 599
 600 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 601                           const struct Qdisc_ops *ops)
 602 {
 603         void *p;
 604         struct Qdisc *sch;
 605         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 606         int err = -ENOBUFS;
 607         struct net_device *dev = dev_queue->dev;
 608
 609         p = kzalloc_node(size, GFP_KERNEL,
 610                          netdev_queue_numa_node_read(dev_queue));
 611
 612         if (!p)
 613                 goto errout;
 614         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 615         /* if we got non aligned memory, ask more and do alignment ourself */
 616         if (sch != p) {
 617                 kfree(p);
 618                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 619                                  netdev_queue_numa_node_read(dev_queue));
 620                 if (!p)
 621                         goto errout;
 622                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 623                 sch->padded = (char *) sch - (char *) p;
 624         }
 625         qdisc_skb_head_init(&sch->q);
 626         spin_lock_init(&sch->q.lock);
 627
 628         spin_lock_init(&sch->busylock);
 629         lockdep_set_class(&sch->busylock,
 630                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 631
 632         seqcount_init(&sch->running);
 633         lockdep_set_class(&sch->running,
 634                           dev->qdisc_running_key ?: &qdisc_running_key);
 635
 636         sch->ops = ops;
 637         sch->enqueue = ops->enqueue;
 638         sch->dequeue = ops->dequeue;
 639         sch->dev_queue = dev_queue;
 640         dev_hold(dev);
 641         refcount_set(&sch->refcnt, 1);
 642
 643         return sch;
 644 errout:
 645         return ERR_PTR(err);
 646 }
 647
 648 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 649                                 const struct Qdisc_ops *ops,
 650                                 unsigned int parentid)
 651 {
 652         struct Qdisc *sch;
 653
 654         if (!try_module_get(ops->owner))
 655                 return NULL;
 656
 657         sch = qdisc_alloc(dev_queue, ops);
 658         if (IS_ERR(sch)) {
 659                 module_put(ops->owner);
 660                 return NULL;
 661         }
 662         sch->parent = parentid;
 663
 664         if (!ops->init || ops->init(sch, NULL) == 0)
 665                 return sch;
 666
 667         qdisc_destroy(sch);
 668         return NULL;
 669 }
 670 EXPORT_SYMBOL(qdisc_create_dflt);
 671
 672 /* Under qdisc_lock(qdisc) and BH! */
 673
 674 void qdisc_reset(struct Qdisc *qdisc)
 675 {
 676         const struct Qdisc_ops *ops = qdisc->ops;
 677
 678         if (ops->reset)
 679                 ops->reset(qdisc);
 680
 681         kfree_skb(qdisc->skb_bad_txq);
 682         qdisc->skb_bad_txq = NULL;
 683
 684         if (qdisc->gso_skb) {
 685                 kfree_skb_list(qdisc->gso_skb);
 686                 qdisc->gso_skb = NULL;
 687         }
 688         qdisc->q.qlen = 0;
 689         qdisc->qstats.backlog = 0;
 690 }
 691 EXPORT_SYMBOL(qdisc_reset);
 692
 693 static void qdisc_rcu_free(struct rcu_head *head)
 694 {
 695         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 696
 697         if (qdisc_is_percpu_stats(qdisc)) {
 698                 free_percpu(qdisc->cpu_bstats);
 699                 free_percpu(qdisc->cpu_qstats);
 700         }
 701
 702         kfree((char *) qdisc - qdisc->padded);
 703 }
 704
 705 void qdisc_destroy(struct Qdisc *qdisc)
 706 {
 707         const struct Qdisc_ops *ops;
 708
 709         if (!qdisc)
 710                 return;
 711         ops = qdisc->ops;
 712
 713         if (qdisc->flags & TCQ_F_BUILTIN ||
 714             !refcount_dec_and_test(&qdisc->refcnt))
 715                 return;
 716
 717 #ifdef CONFIG_NET_SCHED
 718         qdisc_hash_del(qdisc);
 719
 720         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 721 #endif
 722         gen_kill_estimator(&qdisc->rate_est);
 723         if (ops->reset)
 724                 ops->reset(qdisc);
 725         if (ops->destroy)
 726                 ops->destroy(qdisc);
 727
 728         module_put(ops->owner);
 729         dev_put(qdisc_dev(qdisc));
 730
 731         kfree_skb_list(qdisc->gso_skb);
 732         kfree_skb(qdisc->skb_bad_txq);
 733         /*
 734          * gen_estimator est_timer() might access qdisc->q.lock,
 735          * wait a RCU grace period before freeing qdisc.
 736          */
 737         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 738 }
 739 EXPORT_SYMBOL(qdisc_destroy);
 740
 741 /* Attach toplevel qdisc to device queue. */
 742 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 743                               struct Qdisc *qdisc)
 744 {
 745         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 746         spinlock_t *root_lock;
 747
 748         root_lock = qdisc_lock(oqdisc);
 749         spin_lock_bh(root_lock);
 750
 751         /* ... and graft new one */
 752         if (qdisc == NULL)
 753                 qdisc = &noop_qdisc;
 754         dev_queue->qdisc_sleeping = qdisc;
 755         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 756
 757         spin_unlock_bh(root_lock);
 758
 759         return oqdisc;
 760 }
 761 EXPORT_SYMBOL(dev_graft_qdisc);
 762
 763 static void attach_one_default_qdisc(struct net_device *dev,
 764                                      struct netdev_queue *dev_queue,
 765                                      void *_unused)
 766 {
 767         struct Qdisc *qdisc;
 768         const struct Qdisc_ops *ops = default_qdisc_ops;
 769
 770         if (dev->priv_flags & IFF_NO_QUEUE)
 771                 ops = &noqueue_qdisc_ops;
 772
 773         qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT);
 774         if (!qdisc) {
 775                 netdev_info(dev, "activation failed\n");
 776                 return;
 777         }
 778         if (!netif_is_multiqueue(dev))
 779                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 780         dev_queue->qdisc_sleeping = qdisc;
 781 }
 782
 783 static void attach_default_qdiscs(struct net_device *dev)
 784 {
 785         struct netdev_queue *txq;
 786         struct Qdisc *qdisc;
 787
 788         txq = netdev_get_tx_queue(dev, 0);
 789
 790         if (!netif_is_multiqueue(dev) ||
 791             dev->priv_flags & IFF_NO_QUEUE) {
 792                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 793                 dev->qdisc = txq->qdisc_sleeping;
 794                 qdisc_refcount_inc(dev->qdisc);
 795         } else {
 796                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 797                 if (qdisc) {
 798                         dev->qdisc = qdisc;
 799                         qdisc->ops->attach(qdisc);
 800                 }
 801         }
 802 #ifdef CONFIG_NET_SCHED
 803         if (dev->qdisc != &noop_qdisc)
 804                 qdisc_hash_add(dev->qdisc, false);
 805 #endif
 806 }
 807
 808 static void transition_one_qdisc(struct net_device *dev,
 809                                  struct netdev_queue *dev_queue,
 810                                  void *_need_watchdog)
 811 {
 812         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 813         int *need_watchdog_p = _need_watchdog;
 814
 815         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 816                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 817
 818         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 819         if (need_watchdog_p) {
 820                 dev_queue->trans_start = 0;
 821                 *need_watchdog_p = 1;
 822         }
 823 }
 824
 825 void dev_activate(struct net_device *dev)
 826 {
 827         int need_watchdog;
 828
 829         /* No queueing discipline is attached to device;
 830          * create default one for devices, which need queueing
 831          * and noqueue_qdisc for virtual interfaces
 832          */
 833
 834         if (dev->qdisc == &noop_qdisc)
 835                 attach_default_qdiscs(dev);
 836
 837         if (!netif_carrier_ok(dev))
 838                 /* Delay activation until next carrier-on event */
 839                 return;
 840
 841         need_watchdog = 0;
 842         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 843         if (dev_ingress_queue(dev))
 844                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 845
 846         if (need_watchdog) {
 847                 netif_trans_update(dev);
 848                 dev_watchdog_up(dev);
 849         }
 850 }
 851 EXPORT_SYMBOL(dev_activate);
 852
 853 static void dev_deactivate_queue(struct net_device *dev,
 854                                  struct netdev_queue *dev_queue,
 855                                  void *_qdisc_default)
 856 {
 857         struct Qdisc *qdisc_default = _qdisc_default;
 858         struct Qdisc *qdisc;
 859
 860         qdisc = rtnl_dereference(dev_queue->qdisc);
 861         if (qdisc) {
 862                 spin_lock_bh(qdisc_lock(qdisc));
 863
 864                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 865                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 866
 867                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 868                 qdisc_reset(qdisc);
 869
 870                 spin_unlock_bh(qdisc_lock(qdisc));
 871         }
 872 }
 873
 874 static bool some_qdisc_is_busy(struct net_device *dev)
 875 {
 876         unsigned int i;
 877
 878         for (i = 0; i < dev->num_tx_queues; i++) {
 879                 struct netdev_queue *dev_queue;
 880                 spinlock_t *root_lock;
 881                 struct Qdisc *q;
 882                 int val;
 883
 884                 dev_queue = netdev_get_tx_queue(dev, i);
 885                 q = dev_queue->qdisc_sleeping;
 886                 root_lock = qdisc_lock(q);
 887
 888                 spin_lock_bh(root_lock);
 889
 890                 val = (qdisc_is_running(q) ||
 891                        test_bit(__QDISC_STATE_SCHED, &q->state));
 892
 893                 spin_unlock_bh(root_lock);
 894
 895                 if (val)
 896                         return true;
 897         }
 898         return false;
 899 }
 900
 901 static void dev_qdisc_reset(struct net_device *dev,
 902                             struct netdev_queue *dev_queue,
 903                             void *none)
 904 {
 905         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 906
 907         if (qdisc)
 908                 qdisc_reset(qdisc);
 909 }
 910
 911 /**
 912  *      dev_deactivate_many - deactivate transmissions on several devices
 913  *      @head: list of devices to deactivate
 914  *
 915  *      This function returns only when all outstanding transmissions
 916  *      have completed, unless all devices are in dismantle phase.
 917  */
 918 void dev_deactivate_many(struct list_head *head)
 919 {
 920         struct net_device *dev;
 921
 922         list_for_each_entry(dev, head, close_list) {
 923                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 924                                          &noop_qdisc);
 925                 if (dev_ingress_queue(dev))
 926                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 927                                              &noop_qdisc);
 928
 929                 dev_watchdog_down(dev);
 930         }
 931
 932         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 933          * This is avoided if all devices are in dismantle phase :
 934          * Caller will call synchronize_net() for us
 935          */
 936         synchronize_net();
 937
 938         /* Wait for outstanding qdisc_run calls. */
 939         list_for_each_entry(dev, head, close_list) {
 940                 while (some_qdisc_is_busy(dev))
 941                         yield();
 942                 /* The new qdisc is assigned at this point so we can safely
 943                  * unwind stale skb lists and qdisc statistics
 944                  */
 945                 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
 946                 if (dev_ingress_queue(dev))
 947                         dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
 948         }
 949 }
 950
 951 void dev_deactivate(struct net_device *dev)
 952 {
 953         LIST_HEAD(single);
 954
 955         list_add(&dev->close_list, &single);
 956         dev_deactivate_many(&single);
 957         list_del(&single);
 958 }
 959 EXPORT_SYMBOL(dev_deactivate);
 960
 961 static void dev_init_scheduler_queue(struct net_device *dev,
 962                                      struct netdev_queue *dev_queue,
 963                                      void *_qdisc)
 964 {
 965         struct Qdisc *qdisc = _qdisc;
 966
 967         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 968         dev_queue->qdisc_sleeping = qdisc;
 969 }
 970
 971 void dev_init_scheduler(struct net_device *dev)
 972 {
 973         dev->qdisc = &noop_qdisc;
 974         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 975         if (dev_ingress_queue(dev))
 976                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 977
 978         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 979 }
 980
 981 static void shutdown_scheduler_queue(struct net_device *dev,
 982                                      struct netdev_queue *dev_queue,
 983                                      void *_qdisc_default)
 984 {
 985         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 986         struct Qdisc *qdisc_default = _qdisc_default;
 987
 988         if (qdisc) {
 989                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 990                 dev_queue->qdisc_sleeping = qdisc_default;
 991
 992                 qdisc_destroy(qdisc);
 993         }
 994 }
 995
 996 void dev_shutdown(struct net_device *dev)
 997 {
 998         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 999         if (dev_ingress_queue(dev))
1000                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1001         qdisc_destroy(dev->qdisc);
1002         dev->qdisc = &noop_qdisc;
1003
1004         WARN_ON(timer_pending(&dev->watchdog_timer));
1005 }
1006
1007 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1008                                const struct tc_ratespec *conf,
1009                                u64 rate64)
1010 {
1011         memset(r, 0, sizeof(*r));
1012         r->overhead = conf->overhead;
1013         r->mpu = conf->mpu;
1014         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1015         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1016         r->mult = 1;
1017         /*
1018          * The deal here is to replace a divide by a reciprocal one
1019          * in fast path (a reciprocal divide is a multiply and a shift)
1020          *
1021          * Normal formula would be :
1022          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1023          *
1024          * We compute mult/shift to use instead :
1025          *  time_in_ns = (len * mult) >> shift;
1026          *
1027          * We try to get the highest possible mult value for accuracy,
1028          * but have to make sure no overflows will ever happen.
1029          */
1030         if (r->rate_bytes_ps > 0) {
1031                 u64 factor = NSEC_PER_SEC;
1032
1033                 for (;;) {
1034                         r->mult = div64_u64(factor, r->rate_bytes_ps);
1035                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
1036                                 break;
1037                         factor <<= 1;
1038                         r->shift++;
1039                 }
1040         }
1041 }
1042 EXPORT_SYMBOL(psched_ratecfg_precompute);