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