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