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
  33 /* Qdisc to use by default */
  34 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  35 EXPORT_SYMBOL(default_qdisc_ops);
  36
  37 /* Main transmission queue. */
  38
  39 /* Modifications to data participating in scheduling must be protected with
  40  * qdisc_lock(qdisc) spinlock.
  41  *
  42  * The idea is the following:
  43  * - enqueue, dequeue are serialized via qdisc root lock
  44  * - ingress filtering is also serialized via qdisc root lock
  45  * - updates to tree and tree walking are only done under the rtnl mutex.
  46  */
  47
  48 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  49 {
  50         skb_dst_force(skb);
  51         q->gso_skb = skb;
  52         q->qstats.requeues++;
  53         q->q.qlen++;    /* it's still part of the queue */
  54         __netif_schedule(q);
  55
  56         return 0;
  57 }
  58
  59 static void try_bulk_dequeue_skb(struct Qdisc *q,
  60                                  struct sk_buff *skb,
  61                                  const struct netdev_queue *txq)
  62 {
  63         int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
  64
  65         while (bytelimit > 0) {
  66                 struct sk_buff *nskb = q->dequeue(q);
  67
  68                 if (!nskb)
  69                         break;
  70
  71                 bytelimit -= nskb->len; /* covers GSO len */
  72                 skb->next = nskb;
  73                 skb = nskb;
  74         }
  75         skb->next = NULL;
  76 }
  77
  78 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
  79  * A requeued skb (via q->gso_skb) can also be a SKB list.
  80  */
  81 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate)
  82 {
  83         struct sk_buff *skb = q->gso_skb;
  84         const struct netdev_queue *txq = q->dev_queue;
  85
  86         *validate = true;
  87         if (unlikely(skb)) {
  88                 /* check the reason of requeuing without tx lock first */
  89                 txq = skb_get_tx_queue(txq->dev, skb);
  90                 if (!netif_xmit_frozen_or_stopped(txq)) {
  91                         q->gso_skb = NULL;
  92                         q->q.qlen--;
  93                 } else
  94                         skb = NULL;
  95                 /* skb in gso_skb were already validated */
  96                 *validate = false;
  97         } else {
  98                 if (!(q->flags & TCQ_F_ONETXQUEUE) ||
  99                     !netif_xmit_frozen_or_stopped(txq)) {
 100                         skb = q->dequeue(q);
 101                         if (skb && qdisc_may_bulk(q))
 102                                 try_bulk_dequeue_skb(q, skb, txq);
 103                 }
 104         }
 105         return skb;
 106 }
 107
 108 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
 109                                            struct netdev_queue *dev_queue,
 110                                            struct Qdisc *q)
 111 {
 112         int ret;
 113
 114         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
 115                 /*
 116                  * Same CPU holding the lock. It may be a transient
 117                  * configuration error, when hard_start_xmit() recurses. We
 118                  * detect it by checking xmit owner and drop the packet when
 119                  * deadloop is detected. Return OK to try the next skb.
 120                  */
 121                 kfree_skb_list(skb);
 122                 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
 123                                      dev_queue->dev->name);
 124                 ret = qdisc_qlen(q);
 125         } else {
 126                 /*
 127                  * Another cpu is holding lock, requeue & delay xmits for
 128                  * some time.
 129                  */
 130                 __this_cpu_inc(softnet_data.cpu_collision);
 131                 ret = dev_requeue_skb(skb, q);
 132         }
 133
 134         return ret;
 135 }
 136
 137 /*
 138  * Transmit possibly several skbs, and handle the return status as
 139  * required. Holding the __QDISC___STATE_RUNNING bit guarantees that
 140  * only one CPU can execute this function.
 141  *
 142  * Returns to the caller:
 143  *                              0  - queue is empty or throttled.
 144  *                              >0 - queue is not empty.
 145  */
 146 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 147                     struct net_device *dev, struct netdev_queue *txq,
 148                     spinlock_t *root_lock, bool validate)
 149 {
 150         int ret = NETDEV_TX_BUSY;
 151
 152         /* And release qdisc */
 153         spin_unlock(root_lock);
 154
 155         /* Note that we validate skb (GSO, checksum, ...) outside of locks */
 156         if (validate)
 157                 skb = validate_xmit_skb_list(skb, dev);
 158
 159         if (skb) {
 160                 HARD_TX_LOCK(dev, txq, smp_processor_id());
 161                 if (!netif_xmit_frozen_or_stopped(txq))
 162                         skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 163
 164                 HARD_TX_UNLOCK(dev, txq);
 165         }
 166         spin_lock(root_lock);
 167
 168         if (dev_xmit_complete(ret)) {
 169                 /* Driver sent out skb successfully or skb was consumed */
 170                 ret = qdisc_qlen(q);
 171         } else if (ret == NETDEV_TX_LOCKED) {
 172                 /* Driver try lock failed */
 173                 ret = handle_dev_cpu_collision(skb, txq, q);
 174         } else {
 175                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 176                 if (unlikely(ret != NETDEV_TX_BUSY))
 177                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 178                                              dev->name, ret, q->q.qlen);
 179
 180                 ret = dev_requeue_skb(skb, q);
 181         }
 182
 183         if (ret && netif_xmit_frozen_or_stopped(txq))
 184                 ret = 0;
 185
 186         return ret;
 187 }
 188
 189 /*
 190  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 191  *
 192  * __QDISC___STATE_RUNNING guarantees only one CPU can process
 193  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 194  * this queue.
 195  *
 196  *  netif_tx_lock serializes accesses to device driver.
 197  *
 198  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 199  *  if one is grabbed, another must be free.
 200  *
 201  * Note, that this procedure can be called by a watchdog timer
 202  *
 203  * Returns to the caller:
 204  *                              0  - queue is empty or throttled.
 205  *                              >0 - queue is not empty.
 206  *
 207  */
 208 static inline int qdisc_restart(struct Qdisc *q)
 209 {
 210         struct netdev_queue *txq;
 211         struct net_device *dev;
 212         spinlock_t *root_lock;
 213         struct sk_buff *skb;
 214         bool validate;
 215
 216         /* Dequeue packet */
 217         skb = dequeue_skb(q, &validate);
 218         if (unlikely(!skb))
 219                 return 0;
 220
 221         WARN_ON_ONCE(skb_dst_is_noref(skb));
 222
 223         root_lock = qdisc_lock(q);
 224         dev = qdisc_dev(q);
 225         txq = skb_get_tx_queue(dev, skb);
 226
 227         return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 228 }
 229
 230 void __qdisc_run(struct Qdisc *q)
 231 {
 232         int quota = weight_p;
 233
 234         while (qdisc_restart(q)) {
 235                 /*
 236                  * Ordered by possible occurrence: Postpone processing if
 237                  * 1. we've exceeded packet quota
 238                  * 2. another process needs the CPU;
 239                  */
 240                 if (--quota <= 0 || need_resched()) {
 241                         __netif_schedule(q);
 242                         break;
 243                 }
 244         }
 245
 246         qdisc_run_end(q);
 247 }
 248
 249 unsigned long dev_trans_start(struct net_device *dev)
 250 {
 251         unsigned long val, res;
 252         unsigned int i;
 253
 254         if (is_vlan_dev(dev))
 255                 dev = vlan_dev_real_dev(dev);
 256         res = dev->trans_start;
 257         for (i = 0; i < dev->num_tx_queues; i++) {
 258                 val = netdev_get_tx_queue(dev, i)->trans_start;
 259                 if (val && time_after(val, res))
 260                         res = val;
 261         }
 262         dev->trans_start = res;
 263
 264         return res;
 265 }
 266 EXPORT_SYMBOL(dev_trans_start);
 267
 268 static void dev_watchdog(unsigned long arg)
 269 {
 270         struct net_device *dev = (struct net_device *)arg;
 271
 272         netif_tx_lock(dev);
 273         if (!qdisc_tx_is_noop(dev)) {
 274                 if (netif_device_present(dev) &&
 275                     netif_running(dev) &&
 276                     netif_carrier_ok(dev)) {
 277                         int some_queue_timedout = 0;
 278                         unsigned int i;
 279                         unsigned long trans_start;
 280
 281                         for (i = 0; i < dev->num_tx_queues; i++) {
 282                                 struct netdev_queue *txq;
 283
 284                                 txq = netdev_get_tx_queue(dev, i);
 285                                 /*
 286                                  * old device drivers set dev->trans_start
 287                                  */
 288                                 trans_start = txq->trans_start ? : dev->trans_start;
 289                                 if (netif_xmit_stopped(txq) &&
 290                                     time_after(jiffies, (trans_start +
 291                                                          dev->watchdog_timeo))) {
 292                                         some_queue_timedout = 1;
 293                                         txq->trans_timeout++;
 294                                         break;
 295                                 }
 296                         }
 297
 298                         if (some_queue_timedout) {
 299                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 300                                        dev->name, netdev_drivername(dev), i);
 301                                 dev->netdev_ops->ndo_tx_timeout(dev);
 302                         }
 303                         if (!mod_timer(&dev->watchdog_timer,
 304                                        round_jiffies(jiffies +
 305                                                      dev->watchdog_timeo)))
 306                                 dev_hold(dev);
 307                 }
 308         }
 309         netif_tx_unlock(dev);
 310
 311         dev_put(dev);
 312 }
 313
 314 void __netdev_watchdog_up(struct net_device *dev)
 315 {
 316         if (dev->netdev_ops->ndo_tx_timeout) {
 317                 if (dev->watchdog_timeo <= 0)
 318                         dev->watchdog_timeo = 5*HZ;
 319                 if (!mod_timer(&dev->watchdog_timer,
 320                                round_jiffies(jiffies + dev->watchdog_timeo)))
 321                         dev_hold(dev);
 322         }
 323 }
 324
 325 static void dev_watchdog_up(struct net_device *dev)
 326 {
 327         __netdev_watchdog_up(dev);
 328 }
 329
 330 static void dev_watchdog_down(struct net_device *dev)
 331 {
 332         netif_tx_lock_bh(dev);
 333         if (del_timer(&dev->watchdog_timer))
 334                 dev_put(dev);
 335         netif_tx_unlock_bh(dev);
 336 }
 337
 338 /**
 339  *      netif_carrier_on - set carrier
 340  *      @dev: network device
 341  *
 342  * Device has detected that carrier.
 343  */
 344 void netif_carrier_on(struct net_device *dev)
 345 {
 346         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 347                 if (dev->reg_state == NETREG_UNINITIALIZED)
 348                         return;
 349                 atomic_inc(&dev->carrier_changes);
 350                 linkwatch_fire_event(dev);
 351                 if (netif_running(dev))
 352                         __netdev_watchdog_up(dev);
 353         }
 354 }
 355 EXPORT_SYMBOL(netif_carrier_on);
 356
 357 /**
 358  *      netif_carrier_off - clear carrier
 359  *      @dev: network device
 360  *
 361  * Device has detected loss of carrier.
 362  */
 363 void netif_carrier_off(struct net_device *dev)
 364 {
 365         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 366                 if (dev->reg_state == NETREG_UNINITIALIZED)
 367                         return;
 368                 atomic_inc(&dev->carrier_changes);
 369                 linkwatch_fire_event(dev);
 370         }
 371 }
 372 EXPORT_SYMBOL(netif_carrier_off);
 373
 374 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 375    under all circumstances. It is difficult to invent anything faster or
 376    cheaper.
 377  */
 378
 379 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 380 {
 381         kfree_skb(skb);
 382         return NET_XMIT_CN;
 383 }
 384
 385 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 386 {
 387         return NULL;
 388 }
 389
 390 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 391         .id             =       "noop",
 392         .priv_size      =       0,
 393         .enqueue        =       noop_enqueue,
 394         .dequeue        =       noop_dequeue,
 395         .peek           =       noop_dequeue,
 396         .owner          =       THIS_MODULE,
 397 };
 398
 399 static struct netdev_queue noop_netdev_queue = {
 400         .qdisc          =       &noop_qdisc,
 401         .qdisc_sleeping =       &noop_qdisc,
 402 };
 403
 404 struct Qdisc noop_qdisc = {
 405         .enqueue        =       noop_enqueue,
 406         .dequeue        =       noop_dequeue,
 407         .flags          =       TCQ_F_BUILTIN,
 408         .ops            =       &noop_qdisc_ops,
 409         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 410         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 411         .dev_queue      =       &noop_netdev_queue,
 412         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 413 };
 414 EXPORT_SYMBOL(noop_qdisc);
 415
 416 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 417         .id             =       "noqueue",
 418         .priv_size      =       0,
 419         .enqueue        =       noop_enqueue,
 420         .dequeue        =       noop_dequeue,
 421         .peek           =       noop_dequeue,
 422         .owner          =       THIS_MODULE,
 423 };
 424
 425 static struct Qdisc noqueue_qdisc;
 426 static struct netdev_queue noqueue_netdev_queue = {
 427         .qdisc          =       &noqueue_qdisc,
 428         .qdisc_sleeping =       &noqueue_qdisc,
 429 };
 430
 431 static struct Qdisc noqueue_qdisc = {
 432         .enqueue        =       NULL,
 433         .dequeue        =       noop_dequeue,
 434         .flags          =       TCQ_F_BUILTIN,
 435         .ops            =       &noqueue_qdisc_ops,
 436         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 437         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 438         .dev_queue      =       &noqueue_netdev_queue,
 439         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 440 };
 441
 442
 443 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 444         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 445 };
 446
 447 /* 3-band FIFO queue: old style, but should be a bit faster than
 448    generic prio+fifo combination.
 449  */
 450
 451 #define PFIFO_FAST_BANDS 3
 452
 453 /*
 454  * Private data for a pfifo_fast scheduler containing:
 455  *      - queues for the three band
 456  *      - bitmap indicating which of the bands contain skbs
 457  */
 458 struct pfifo_fast_priv {
 459         u32 bitmap;
 460         struct sk_buff_head q[PFIFO_FAST_BANDS];
 461 };
 462
 463 /*
 464  * Convert a bitmap to the first band number where an skb is queued, where:
 465  *      bitmap=0 means there are no skbs on any band.
 466  *      bitmap=1 means there is an skb on band 0.
 467  *      bitmap=7 means there are skbs on all 3 bands, etc.
 468  */
 469 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 470
 471 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 472                                              int band)
 473 {
 474         return priv->q + band;
 475 }
 476
 477 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 478 {
 479         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 480                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 481                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 482                 struct sk_buff_head *list = band2list(priv, band);
 483
 484                 priv->bitmap |= (1 << band);
 485                 qdisc->q.qlen++;
 486                 return __qdisc_enqueue_tail(skb, qdisc, list);
 487         }
 488
 489         return qdisc_drop(skb, qdisc);
 490 }
 491
 492 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 493 {
 494         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 495         int band = bitmap2band[priv->bitmap];
 496
 497         if (likely(band >= 0)) {
 498                 struct sk_buff_head *list = band2list(priv, band);
 499                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 500
 501                 qdisc->q.qlen--;
 502                 if (skb_queue_empty(list))
 503                         priv->bitmap &= ~(1 << band);
 504
 505                 return skb;
 506         }
 507
 508         return NULL;
 509 }
 510
 511 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 512 {
 513         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 514         int band = bitmap2band[priv->bitmap];
 515
 516         if (band >= 0) {
 517                 struct sk_buff_head *list = band2list(priv, band);
 518
 519                 return skb_peek(list);
 520         }
 521
 522         return NULL;
 523 }
 524
 525 static void pfifo_fast_reset(struct Qdisc *qdisc)
 526 {
 527         int prio;
 528         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 529
 530         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 531                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 532
 533         priv->bitmap = 0;
 534         qdisc->qstats.backlog = 0;
 535         qdisc->q.qlen = 0;
 536 }
 537
 538 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 539 {
 540         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 541
 542         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 543         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 544                 goto nla_put_failure;
 545         return skb->len;
 546
 547 nla_put_failure:
 548         return -1;
 549 }
 550
 551 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 552 {
 553         int prio;
 554         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 555
 556         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 557                 __skb_queue_head_init(band2list(priv, prio));
 558
 559         /* Can by-pass the queue discipline */
 560         qdisc->flags |= TCQ_F_CAN_BYPASS;
 561         return 0;
 562 }
 563
 564 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 565         .id             =       "pfifo_fast",
 566         .priv_size      =       sizeof(struct pfifo_fast_priv),
 567         .enqueue        =       pfifo_fast_enqueue,
 568         .dequeue        =       pfifo_fast_dequeue,
 569         .peek           =       pfifo_fast_peek,
 570         .init           =       pfifo_fast_init,
 571         .reset          =       pfifo_fast_reset,
 572         .dump           =       pfifo_fast_dump,
 573         .owner          =       THIS_MODULE,
 574 };
 575
 576 static struct lock_class_key qdisc_tx_busylock;
 577
 578 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 579                           const struct Qdisc_ops *ops)
 580 {
 581         void *p;
 582         struct Qdisc *sch;
 583         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 584         int err = -ENOBUFS;
 585         struct net_device *dev = dev_queue->dev;
 586
 587         p = kzalloc_node(size, GFP_KERNEL,
 588                          netdev_queue_numa_node_read(dev_queue));
 589
 590         if (!p)
 591                 goto errout;
 592         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 593         /* if we got non aligned memory, ask more and do alignment ourself */
 594         if (sch != p) {
 595                 kfree(p);
 596                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 597                                  netdev_queue_numa_node_read(dev_queue));
 598                 if (!p)
 599                         goto errout;
 600                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 601                 sch->padded = (char *) sch - (char *) p;
 602         }
 603         INIT_LIST_HEAD(&sch->list);
 604         skb_queue_head_init(&sch->q);
 605
 606         spin_lock_init(&sch->busylock);
 607         lockdep_set_class(&sch->busylock,
 608                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 609
 610         sch->ops = ops;
 611         sch->enqueue = ops->enqueue;
 612         sch->dequeue = ops->dequeue;
 613         sch->dev_queue = dev_queue;
 614         dev_hold(dev);
 615         atomic_set(&sch->refcnt, 1);
 616
 617         return sch;
 618 errout:
 619         return ERR_PTR(err);
 620 }
 621
 622 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 623                                 const struct Qdisc_ops *ops,
 624                                 unsigned int parentid)
 625 {
 626         struct Qdisc *sch;
 627
 628         if (!try_module_get(ops->owner))
 629                 goto errout;
 630
 631         sch = qdisc_alloc(dev_queue, ops);
 632         if (IS_ERR(sch))
 633                 goto errout;
 634         sch->parent = parentid;
 635
 636         if (!ops->init || ops->init(sch, NULL) == 0)
 637                 return sch;
 638
 639         qdisc_destroy(sch);
 640 errout:
 641         return NULL;
 642 }
 643 EXPORT_SYMBOL(qdisc_create_dflt);
 644
 645 /* Under qdisc_lock(qdisc) and BH! */
 646
 647 void qdisc_reset(struct Qdisc *qdisc)
 648 {
 649         const struct Qdisc_ops *ops = qdisc->ops;
 650
 651         if (ops->reset)
 652                 ops->reset(qdisc);
 653
 654         if (qdisc->gso_skb) {
 655                 kfree_skb_list(qdisc->gso_skb);
 656                 qdisc->gso_skb = NULL;
 657                 qdisc->q.qlen = 0;
 658         }
 659 }
 660 EXPORT_SYMBOL(qdisc_reset);
 661
 662 static void qdisc_rcu_free(struct rcu_head *head)
 663 {
 664         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 665
 666         if (qdisc_is_percpu_stats(qdisc))
 667                 free_percpu(qdisc->cpu_bstats);
 668
 669         kfree((char *) qdisc - qdisc->padded);
 670 }
 671
 672 void qdisc_destroy(struct Qdisc *qdisc)
 673 {
 674         const struct Qdisc_ops  *ops = qdisc->ops;
 675
 676         if (qdisc->flags & TCQ_F_BUILTIN ||
 677             !atomic_dec_and_test(&qdisc->refcnt))
 678                 return;
 679
 680 #ifdef CONFIG_NET_SCHED
 681         qdisc_list_del(qdisc);
 682
 683         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 684 #endif
 685         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 686         if (ops->reset)
 687                 ops->reset(qdisc);
 688         if (ops->destroy)
 689                 ops->destroy(qdisc);
 690
 691         module_put(ops->owner);
 692         dev_put(qdisc_dev(qdisc));
 693
 694         kfree_skb_list(qdisc->gso_skb);
 695         /*
 696          * gen_estimator est_timer() might access qdisc->q.lock,
 697          * wait a RCU grace period before freeing qdisc.
 698          */
 699         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 700 }
 701 EXPORT_SYMBOL(qdisc_destroy);
 702
 703 /* Attach toplevel qdisc to device queue. */
 704 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 705                               struct Qdisc *qdisc)
 706 {
 707         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 708         spinlock_t *root_lock;
 709
 710         root_lock = qdisc_lock(oqdisc);
 711         spin_lock_bh(root_lock);
 712
 713         /* Prune old scheduler */
 714         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 715                 qdisc_reset(oqdisc);
 716
 717         /* ... and graft new one */
 718         if (qdisc == NULL)
 719                 qdisc = &noop_qdisc;
 720         dev_queue->qdisc_sleeping = qdisc;
 721         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 722
 723         spin_unlock_bh(root_lock);
 724
 725         return oqdisc;
 726 }
 727 EXPORT_SYMBOL(dev_graft_qdisc);
 728
 729 static void attach_one_default_qdisc(struct net_device *dev,
 730                                      struct netdev_queue *dev_queue,
 731                                      void *_unused)
 732 {
 733         struct Qdisc *qdisc = &noqueue_qdisc;
 734
 735         if (dev->tx_queue_len) {
 736                 qdisc = qdisc_create_dflt(dev_queue,
 737                                           default_qdisc_ops, TC_H_ROOT);
 738                 if (!qdisc) {
 739                         netdev_info(dev, "activation failed\n");
 740                         return;
 741                 }
 742                 if (!netif_is_multiqueue(dev))
 743                         qdisc->flags |= TCQ_F_ONETXQUEUE;
 744         }
 745         dev_queue->qdisc_sleeping = qdisc;
 746 }
 747
 748 static void attach_default_qdiscs(struct net_device *dev)
 749 {
 750         struct netdev_queue *txq;
 751         struct Qdisc *qdisc;
 752
 753         txq = netdev_get_tx_queue(dev, 0);
 754
 755         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 756                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 757                 dev->qdisc = txq->qdisc_sleeping;
 758                 atomic_inc(&dev->qdisc->refcnt);
 759         } else {
 760                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 761                 if (qdisc) {
 762                         dev->qdisc = qdisc;
 763                         qdisc->ops->attach(qdisc);
 764                 }
 765         }
 766 }
 767
 768 static void transition_one_qdisc(struct net_device *dev,
 769                                  struct netdev_queue *dev_queue,
 770                                  void *_need_watchdog)
 771 {
 772         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 773         int *need_watchdog_p = _need_watchdog;
 774
 775         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 776                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 777
 778         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 779         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 780                 dev_queue->trans_start = 0;
 781                 *need_watchdog_p = 1;
 782         }
 783 }
 784
 785 void dev_activate(struct net_device *dev)
 786 {
 787         int need_watchdog;
 788
 789         /* No queueing discipline is attached to device;
 790          * create default one for devices, which need queueing
 791          * and noqueue_qdisc for virtual interfaces
 792          */
 793
 794         if (dev->qdisc == &noop_qdisc)
 795                 attach_default_qdiscs(dev);
 796
 797         if (!netif_carrier_ok(dev))
 798                 /* Delay activation until next carrier-on event */
 799                 return;
 800
 801         need_watchdog = 0;
 802         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 803         if (dev_ingress_queue(dev))
 804                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 805
 806         if (need_watchdog) {
 807                 dev->trans_start = jiffies;
 808                 dev_watchdog_up(dev);
 809         }
 810 }
 811 EXPORT_SYMBOL(dev_activate);
 812
 813 static void dev_deactivate_queue(struct net_device *dev,
 814                                  struct netdev_queue *dev_queue,
 815                                  void *_qdisc_default)
 816 {
 817         struct Qdisc *qdisc_default = _qdisc_default;
 818         struct Qdisc *qdisc;
 819
 820         qdisc = rtnl_dereference(dev_queue->qdisc);
 821         if (qdisc) {
 822                 spin_lock_bh(qdisc_lock(qdisc));
 823
 824                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 825                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 826
 827                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 828                 qdisc_reset(qdisc);
 829
 830                 spin_unlock_bh(qdisc_lock(qdisc));
 831         }
 832 }
 833
 834 static bool some_qdisc_is_busy(struct net_device *dev)
 835 {
 836         unsigned int i;
 837
 838         for (i = 0; i < dev->num_tx_queues; i++) {
 839                 struct netdev_queue *dev_queue;
 840                 spinlock_t *root_lock;
 841                 struct Qdisc *q;
 842                 int val;
 843
 844                 dev_queue = netdev_get_tx_queue(dev, i);
 845                 q = dev_queue->qdisc_sleeping;
 846                 root_lock = qdisc_lock(q);
 847
 848                 spin_lock_bh(root_lock);
 849
 850                 val = (qdisc_is_running(q) ||
 851                        test_bit(__QDISC_STATE_SCHED, &q->state));
 852
 853                 spin_unlock_bh(root_lock);
 854
 855                 if (val)
 856                         return true;
 857         }
 858         return false;
 859 }
 860
 861 /**
 862  *      dev_deactivate_many - deactivate transmissions on several devices
 863  *      @head: list of devices to deactivate
 864  *
 865  *      This function returns only when all outstanding transmissions
 866  *      have completed, unless all devices are in dismantle phase.
 867  */
 868 void dev_deactivate_many(struct list_head *head)
 869 {
 870         struct net_device *dev;
 871         bool sync_needed = false;
 872
 873         list_for_each_entry(dev, head, close_list) {
 874                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 875                                          &noop_qdisc);
 876                 if (dev_ingress_queue(dev))
 877                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 878                                              &noop_qdisc);
 879
 880                 dev_watchdog_down(dev);
 881                 sync_needed |= !dev->dismantle;
 882         }
 883
 884         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 885          * This is avoided if all devices are in dismantle phase :
 886          * Caller will call synchronize_net() for us
 887          */
 888         if (sync_needed)
 889                 synchronize_net();
 890
 891         /* Wait for outstanding qdisc_run calls. */
 892         list_for_each_entry(dev, head, close_list)
 893                 while (some_qdisc_is_busy(dev))
 894                         yield();
 895 }
 896
 897 void dev_deactivate(struct net_device *dev)
 898 {
 899         LIST_HEAD(single);
 900
 901         list_add(&dev->close_list, &single);
 902         dev_deactivate_many(&single);
 903         list_del(&single);
 904 }
 905 EXPORT_SYMBOL(dev_deactivate);
 906
 907 static void dev_init_scheduler_queue(struct net_device *dev,
 908                                      struct netdev_queue *dev_queue,
 909                                      void *_qdisc)
 910 {
 911         struct Qdisc *qdisc = _qdisc;
 912
 913         rcu_assign_pointer(dev_queue->qdisc, qdisc);
 914         dev_queue->qdisc_sleeping = qdisc;
 915 }
 916
 917 void dev_init_scheduler(struct net_device *dev)
 918 {
 919         dev->qdisc = &noop_qdisc;
 920         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 921         if (dev_ingress_queue(dev))
 922                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 923
 924         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 925 }
 926
 927 static void shutdown_scheduler_queue(struct net_device *dev,
 928                                      struct netdev_queue *dev_queue,
 929                                      void *_qdisc_default)
 930 {
 931         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 932         struct Qdisc *qdisc_default = _qdisc_default;
 933
 934         if (qdisc) {
 935                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 936                 dev_queue->qdisc_sleeping = qdisc_default;
 937
 938                 qdisc_destroy(qdisc);
 939         }
 940 }
 941
 942 void dev_shutdown(struct net_device *dev)
 943 {
 944         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 945         if (dev_ingress_queue(dev))
 946                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 947         qdisc_destroy(dev->qdisc);
 948         dev->qdisc = &noop_qdisc;
 949
 950         WARN_ON(timer_pending(&dev->watchdog_timer));
 951 }
 952
 953 void psched_ratecfg_precompute(struct psched_ratecfg *r,
 954                                const struct tc_ratespec *conf,
 955                                u64 rate64)
 956 {
 957         memset(r, 0, sizeof(*r));
 958         r->overhead = conf->overhead;
 959         r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
 960         r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
 961         r->mult = 1;
 962         /*
 963          * The deal here is to replace a divide by a reciprocal one
 964          * in fast path (a reciprocal divide is a multiply and a shift)
 965          *
 966          * Normal formula would be :
 967          *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
 968          *
 969          * We compute mult/shift to use instead :
 970          *  time_in_ns = (len * mult) >> shift;
 971          *
 972          * We try to get the highest possible mult value for accuracy,
 973          * but have to make sure no overflows will ever happen.
 974          */
 975         if (r->rate_bytes_ps > 0) {
 976                 u64 factor = NSEC_PER_SEC;
 977
 978                 for (;;) {
 979                         r->mult = div64_u64(factor, r->rate_bytes_ps);
 980                         if (r->mult & (1U << 31) || factor & (1ULL << 63))
 981                                 break;
 982                         factor <<= 1;
 983                         r->shift++;
 984                 }
 985         }
 986 }
 987 EXPORT_SYMBOL(psched_ratecfg_precompute);