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 <net/sch_generic.h>
  29 #include <net/pkt_sched.h>
  30 #include <net/dst.h>
  31
  32 /* Main transmission queue. */
  33
  34 /* Modifications to data participating in scheduling must be protected with
  35  * qdisc_lock(qdisc) spinlock.
  36  *
  37  * The idea is the following:
  38  * - enqueue, dequeue are serialized via qdisc root lock
  39  * - ingress filtering is also serialized via qdisc root lock
  40  * - updates to tree and tree walking are only done under the rtnl mutex.
  41  */
  42
  43 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
  44 {
  45         skb_dst_force(skb);
  46         q->gso_skb = skb;
  47         q->qstats.requeues++;
  48         q->q.qlen++;    /* it's still part of the queue */
  49         __netif_schedule(q);
  50
  51         return 0;
  52 }
  53
  54 static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
  55 {
  56         struct sk_buff *skb = q->gso_skb;
  57         const struct netdev_queue *txq = q->dev_queue;
  58
  59         if (unlikely(skb)) {
  60                 /* check the reason of requeuing without tx lock first */
  61                 txq = netdev_get_tx_queue(txq->dev, skb_get_queue_mapping(skb));
  62                 if (!netif_xmit_frozen_or_stopped(txq)) {
  63                         q->gso_skb = NULL;
  64                         q->q.qlen--;
  65                 } else
  66                         skb = NULL;
  67         } else {
  68                 if (!(q->flags & TCQ_F_ONETXQUEUE) || !netif_xmit_frozen_or_stopped(txq))
  69                         skb = q->dequeue(q);
  70         }
  71
  72         return skb;
  73 }
  74
  75 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  76                                            struct netdev_queue *dev_queue,
  77                                            struct Qdisc *q)
  78 {
  79         int ret;
  80
  81         if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
  82                 /*
  83                  * Same CPU holding the lock. It may be a transient
  84                  * configuration error, when hard_start_xmit() recurses. We
  85                  * detect it by checking xmit owner and drop the packet when
  86                  * deadloop is detected. Return OK to try the next skb.
  87                  */
  88                 kfree_skb(skb);
  89                 net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
  90                                      dev_queue->dev->name);
  91                 ret = qdisc_qlen(q);
  92         } else {
  93                 /*
  94                  * Another cpu is holding lock, requeue & delay xmits for
  95                  * some time.
  96                  */
  97                 __this_cpu_inc(softnet_data.cpu_collision);
  98                 ret = dev_requeue_skb(skb, q);
  99         }
 100
 101         return ret;
 102 }
 103
 104 /*
 105  * Transmit one skb, and handle the return status as required. Holding the
 106  * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
 107  * function.
 108  *
 109  * Returns to the caller:
 110  *                              0  - queue is empty or throttled.
 111  *                              >0 - queue is not empty.
 112  */
 113 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 114                     struct net_device *dev, struct netdev_queue *txq,
 115                     spinlock_t *root_lock)
 116 {
 117         int ret = NETDEV_TX_BUSY;
 118
 119         /* And release qdisc */
 120         spin_unlock(root_lock);
 121
 122         HARD_TX_LOCK(dev, txq, smp_processor_id());
 123         if (!netif_xmit_frozen_or_stopped(txq))
 124                 ret = dev_hard_start_xmit(skb, dev, txq);
 125
 126         HARD_TX_UNLOCK(dev, txq);
 127
 128         spin_lock(root_lock);
 129
 130         if (dev_xmit_complete(ret)) {
 131                 /* Driver sent out skb successfully or skb was consumed */
 132                 ret = qdisc_qlen(q);
 133         } else if (ret == NETDEV_TX_LOCKED) {
 134                 /* Driver try lock failed */
 135                 ret = handle_dev_cpu_collision(skb, txq, q);
 136         } else {
 137                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 138                 if (unlikely(ret != NETDEV_TX_BUSY))
 139                         net_warn_ratelimited("BUG %s code %d qlen %d\n",
 140                                              dev->name, ret, q->q.qlen);
 141
 142                 ret = dev_requeue_skb(skb, q);
 143         }
 144
 145         if (ret && netif_xmit_frozen_or_stopped(txq))
 146                 ret = 0;
 147
 148         return ret;
 149 }
 150
 151 /*
 152  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 153  *
 154  * __QDISC_STATE_RUNNING guarantees only one CPU can process
 155  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 156  * this queue.
 157  *
 158  *  netif_tx_lock serializes accesses to device driver.
 159  *
 160  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 161  *  if one is grabbed, another must be free.
 162  *
 163  * Note, that this procedure can be called by a watchdog timer
 164  *
 165  * Returns to the caller:
 166  *                              0  - queue is empty or throttled.
 167  *                              >0 - queue is not empty.
 168  *
 169  */
 170 static inline int qdisc_restart(struct Qdisc *q)
 171 {
 172         struct netdev_queue *txq;
 173         struct net_device *dev;
 174         spinlock_t *root_lock;
 175         struct sk_buff *skb;
 176
 177         /* Dequeue packet */
 178         skb = dequeue_skb(q);
 179         if (unlikely(!skb))
 180                 return 0;
 181         WARN_ON_ONCE(skb_dst_is_noref(skb));
 182         root_lock = qdisc_lock(q);
 183         dev = qdisc_dev(q);
 184         txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 185
 186         return sch_direct_xmit(skb, q, dev, txq, root_lock);
 187 }
 188
 189 void __qdisc_run(struct Qdisc *q)
 190 {
 191         int quota = weight_p;
 192
 193         while (qdisc_restart(q)) {
 194                 /*
 195                  * Ordered by possible occurrence: Postpone processing if
 196                  * 1. we've exceeded packet quota
 197                  * 2. another process needs the CPU;
 198                  */
 199                 if (--quota <= 0 || need_resched()) {
 200                         __netif_schedule(q);
 201                         break;
 202                 }
 203         }
 204
 205         qdisc_run_end(q);
 206 }
 207
 208 unsigned long dev_trans_start(struct net_device *dev)
 209 {
 210         unsigned long val, res = dev->trans_start;
 211         unsigned int i;
 212
 213         for (i = 0; i < dev->num_tx_queues; i++) {
 214                 val = netdev_get_tx_queue(dev, i)->trans_start;
 215                 if (val && time_after(val, res))
 216                         res = val;
 217         }
 218         dev->trans_start = res;
 219         return res;
 220 }
 221 EXPORT_SYMBOL(dev_trans_start);
 222
 223 static void dev_watchdog(unsigned long arg)
 224 {
 225         struct net_device *dev = (struct net_device *)arg;
 226
 227         netif_tx_lock(dev);
 228         if (!qdisc_tx_is_noop(dev)) {
 229                 if (netif_device_present(dev) &&
 230                     netif_running(dev) &&
 231                     netif_carrier_ok(dev)) {
 232                         int some_queue_timedout = 0;
 233                         unsigned int i;
 234                         unsigned long trans_start;
 235
 236                         for (i = 0; i < dev->num_tx_queues; i++) {
 237                                 struct netdev_queue *txq;
 238
 239                                 txq = netdev_get_tx_queue(dev, i);
 240                                 /*
 241                                  * old device drivers set dev->trans_start
 242                                  */
 243                                 trans_start = txq->trans_start ? : dev->trans_start;
 244                                 if (netif_xmit_stopped(txq) &&
 245                                     time_after(jiffies, (trans_start +
 246                                                          dev->watchdog_timeo))) {
 247                                         some_queue_timedout = 1;
 248                                         txq->trans_timeout++;
 249                                         break;
 250                                 }
 251                         }
 252
 253                         if (some_queue_timedout) {
 254                                 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 255                                        dev->name, netdev_drivername(dev), i);
 256                                 dev->netdev_ops->ndo_tx_timeout(dev);
 257                         }
 258                         if (!mod_timer(&dev->watchdog_timer,
 259                                        round_jiffies(jiffies +
 260                                                      dev->watchdog_timeo)))
 261                                 dev_hold(dev);
 262                 }
 263         }
 264         netif_tx_unlock(dev);
 265
 266         dev_put(dev);
 267 }
 268
 269 void __netdev_watchdog_up(struct net_device *dev)
 270 {
 271         if (dev->netdev_ops->ndo_tx_timeout) {
 272                 if (dev->watchdog_timeo <= 0)
 273                         dev->watchdog_timeo = 5*HZ;
 274                 if (!mod_timer(&dev->watchdog_timer,
 275                                round_jiffies(jiffies + dev->watchdog_timeo)))
 276                         dev_hold(dev);
 277         }
 278 }
 279
 280 static void dev_watchdog_up(struct net_device *dev)
 281 {
 282         __netdev_watchdog_up(dev);
 283 }
 284
 285 static void dev_watchdog_down(struct net_device *dev)
 286 {
 287         netif_tx_lock_bh(dev);
 288         if (del_timer(&dev->watchdog_timer))
 289                 dev_put(dev);
 290         netif_tx_unlock_bh(dev);
 291 }
 292
 293 /**
 294  *      netif_carrier_on - set carrier
 295  *      @dev: network device
 296  *
 297  * Device has detected that carrier.
 298  */
 299 void netif_carrier_on(struct net_device *dev)
 300 {
 301         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 302                 if (dev->reg_state == NETREG_UNINITIALIZED)
 303                         return;
 304                 linkwatch_fire_event(dev);
 305                 if (netif_running(dev))
 306                         __netdev_watchdog_up(dev);
 307         }
 308 }
 309 EXPORT_SYMBOL(netif_carrier_on);
 310
 311 /**
 312  *      netif_carrier_off - clear carrier
 313  *      @dev: network device
 314  *
 315  * Device has detected loss of carrier.
 316  */
 317 void netif_carrier_off(struct net_device *dev)
 318 {
 319         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 320                 if (dev->reg_state == NETREG_UNINITIALIZED)
 321                         return;
 322                 linkwatch_fire_event(dev);
 323         }
 324 }
 325 EXPORT_SYMBOL(netif_carrier_off);
 326
 327 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 328    under all circumstances. It is difficult to invent anything faster or
 329    cheaper.
 330  */
 331
 332 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 333 {
 334         kfree_skb(skb);
 335         return NET_XMIT_CN;
 336 }
 337
 338 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 339 {
 340         return NULL;
 341 }
 342
 343 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 344         .id             =       "noop",
 345         .priv_size      =       0,
 346         .enqueue        =       noop_enqueue,
 347         .dequeue        =       noop_dequeue,
 348         .peek           =       noop_dequeue,
 349         .owner          =       THIS_MODULE,
 350 };
 351
 352 static struct netdev_queue noop_netdev_queue = {
 353         .qdisc          =       &noop_qdisc,
 354         .qdisc_sleeping =       &noop_qdisc,
 355 };
 356
 357 struct Qdisc noop_qdisc = {
 358         .enqueue        =       noop_enqueue,
 359         .dequeue        =       noop_dequeue,
 360         .flags          =       TCQ_F_BUILTIN,
 361         .ops            =       &noop_qdisc_ops,
 362         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 363         .q.lock         =       __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 364         .dev_queue      =       &noop_netdev_queue,
 365         .busylock       =       __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 366 };
 367 EXPORT_SYMBOL(noop_qdisc);
 368
 369 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 370         .id             =       "noqueue",
 371         .priv_size      =       0,
 372         .enqueue        =       noop_enqueue,
 373         .dequeue        =       noop_dequeue,
 374         .peek           =       noop_dequeue,
 375         .owner          =       THIS_MODULE,
 376 };
 377
 378 static struct Qdisc noqueue_qdisc;
 379 static struct netdev_queue noqueue_netdev_queue = {
 380         .qdisc          =       &noqueue_qdisc,
 381         .qdisc_sleeping =       &noqueue_qdisc,
 382 };
 383
 384 static struct Qdisc noqueue_qdisc = {
 385         .enqueue        =       NULL,
 386         .dequeue        =       noop_dequeue,
 387         .flags          =       TCQ_F_BUILTIN,
 388         .ops            =       &noqueue_qdisc_ops,
 389         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 390         .q.lock         =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
 391         .dev_queue      =       &noqueue_netdev_queue,
 392         .busylock       =       __SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
 393 };
 394
 395
 396 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 397         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 398 };
 399
 400 /* 3-band FIFO queue: old style, but should be a bit faster than
 401    generic prio+fifo combination.
 402  */
 403
 404 #define PFIFO_FAST_BANDS 3
 405
 406 /*
 407  * Private data for a pfifo_fast scheduler containing:
 408  *      - queues for the three band
 409  *      - bitmap indicating which of the bands contain skbs
 410  */
 411 struct pfifo_fast_priv {
 412         u32 bitmap;
 413         struct sk_buff_head q[PFIFO_FAST_BANDS];
 414 };
 415
 416 /*
 417  * Convert a bitmap to the first band number where an skb is queued, where:
 418  *      bitmap=0 means there are no skbs on any band.
 419  *      bitmap=1 means there is an skb on band 0.
 420  *      bitmap=7 means there are skbs on all 3 bands, etc.
 421  */
 422 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
 423
 424 static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
 425                                              int band)
 426 {
 427         return priv->q + band;
 428 }
 429
 430 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 431 {
 432         if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
 433                 int band = prio2band[skb->priority & TC_PRIO_MAX];
 434                 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 435                 struct sk_buff_head *list = band2list(priv, band);
 436
 437                 priv->bitmap |= (1 << band);
 438                 qdisc->q.qlen++;
 439                 return __qdisc_enqueue_tail(skb, qdisc, list);
 440         }
 441
 442         return qdisc_drop(skb, qdisc);
 443 }
 444
 445 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 446 {
 447         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 448         int band = bitmap2band[priv->bitmap];
 449
 450         if (likely(band >= 0)) {
 451                 struct sk_buff_head *list = band2list(priv, band);
 452                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 453
 454                 qdisc->q.qlen--;
 455                 if (skb_queue_empty(list))
 456                         priv->bitmap &= ~(1 << band);
 457
 458                 return skb;
 459         }
 460
 461         return NULL;
 462 }
 463
 464 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 465 {
 466         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 467         int band = bitmap2band[priv->bitmap];
 468
 469         if (band >= 0) {
 470                 struct sk_buff_head *list = band2list(priv, band);
 471
 472                 return skb_peek(list);
 473         }
 474
 475         return NULL;
 476 }
 477
 478 static void pfifo_fast_reset(struct Qdisc *qdisc)
 479 {
 480         int prio;
 481         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 482
 483         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 484                 __qdisc_reset_queue(qdisc, band2list(priv, prio));
 485
 486         priv->bitmap = 0;
 487         qdisc->qstats.backlog = 0;
 488         qdisc->q.qlen = 0;
 489 }
 490
 491 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 492 {
 493         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 494
 495         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 496         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 497                 goto nla_put_failure;
 498         return skb->len;
 499
 500 nla_put_failure:
 501         return -1;
 502 }
 503
 504 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 505 {
 506         int prio;
 507         struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 508
 509         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 510                 skb_queue_head_init(band2list(priv, prio));
 511
 512         /* Can by-pass the queue discipline */
 513         qdisc->flags |= TCQ_F_CAN_BYPASS;
 514         return 0;
 515 }
 516
 517 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 518         .id             =       "pfifo_fast",
 519         .priv_size      =       sizeof(struct pfifo_fast_priv),
 520         .enqueue        =       pfifo_fast_enqueue,
 521         .dequeue        =       pfifo_fast_dequeue,
 522         .peek           =       pfifo_fast_peek,
 523         .init           =       pfifo_fast_init,
 524         .reset          =       pfifo_fast_reset,
 525         .dump           =       pfifo_fast_dump,
 526         .owner          =       THIS_MODULE,
 527 };
 528 EXPORT_SYMBOL(pfifo_fast_ops);
 529
 530 static struct lock_class_key qdisc_tx_busylock;
 531
 532 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 533                           struct Qdisc_ops *ops)
 534 {
 535         void *p;
 536         struct Qdisc *sch;
 537         unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 538         int err = -ENOBUFS;
 539         struct net_device *dev = dev_queue->dev;
 540
 541         p = kzalloc_node(size, GFP_KERNEL,
 542                          netdev_queue_numa_node_read(dev_queue));
 543
 544         if (!p)
 545                 goto errout;
 546         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 547         /* if we got non aligned memory, ask more and do alignment ourself */
 548         if (sch != p) {
 549                 kfree(p);
 550                 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 551                                  netdev_queue_numa_node_read(dev_queue));
 552                 if (!p)
 553                         goto errout;
 554                 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 555                 sch->padded = (char *) sch - (char *) p;
 556         }
 557         INIT_LIST_HEAD(&sch->list);
 558         skb_queue_head_init(&sch->q);
 559
 560         spin_lock_init(&sch->busylock);
 561         lockdep_set_class(&sch->busylock,
 562                           dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
 563
 564         sch->ops = ops;
 565         sch->enqueue = ops->enqueue;
 566         sch->dequeue = ops->dequeue;
 567         sch->dev_queue = dev_queue;
 568         dev_hold(dev);
 569         atomic_set(&sch->refcnt, 1);
 570
 571         return sch;
 572 errout:
 573         return ERR_PTR(err);
 574 }
 575
 576 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 577                                 struct Qdisc_ops *ops, unsigned int parentid)
 578 {
 579         struct Qdisc *sch;
 580
 581         sch = qdisc_alloc(dev_queue, ops);
 582         if (IS_ERR(sch))
 583                 goto errout;
 584         sch->parent = parentid;
 585
 586         if (!ops->init || ops->init(sch, NULL) == 0)
 587                 return sch;
 588
 589         qdisc_destroy(sch);
 590 errout:
 591         return NULL;
 592 }
 593 EXPORT_SYMBOL(qdisc_create_dflt);
 594
 595 /* Under qdisc_lock(qdisc) and BH! */
 596
 597 void qdisc_reset(struct Qdisc *qdisc)
 598 {
 599         const struct Qdisc_ops *ops = qdisc->ops;
 600
 601         if (ops->reset)
 602                 ops->reset(qdisc);
 603
 604         if (qdisc->gso_skb) {
 605                 kfree_skb(qdisc->gso_skb);
 606                 qdisc->gso_skb = NULL;
 607                 qdisc->q.qlen = 0;
 608         }
 609 }
 610 EXPORT_SYMBOL(qdisc_reset);
 611
 612 static void qdisc_rcu_free(struct rcu_head *head)
 613 {
 614         struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 615
 616         kfree((char *) qdisc - qdisc->padded);
 617 }
 618
 619 void qdisc_destroy(struct Qdisc *qdisc)
 620 {
 621         const struct Qdisc_ops  *ops = qdisc->ops;
 622
 623         if (qdisc->flags & TCQ_F_BUILTIN ||
 624             !atomic_dec_and_test(&qdisc->refcnt))
 625                 return;
 626
 627 #ifdef CONFIG_NET_SCHED
 628         qdisc_list_del(qdisc);
 629
 630         qdisc_put_stab(rtnl_dereference(qdisc->stab));
 631 #endif
 632         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 633         if (ops->reset)
 634                 ops->reset(qdisc);
 635         if (ops->destroy)
 636                 ops->destroy(qdisc);
 637
 638         module_put(ops->owner);
 639         dev_put(qdisc_dev(qdisc));
 640
 641         kfree_skb(qdisc->gso_skb);
 642         /*
 643          * gen_estimator est_timer() might access qdisc->q.lock,
 644          * wait a RCU grace period before freeing qdisc.
 645          */
 646         call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 647 }
 648 EXPORT_SYMBOL(qdisc_destroy);
 649
 650 /* Attach toplevel qdisc to device queue. */
 651 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
 652                               struct Qdisc *qdisc)
 653 {
 654         struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
 655         spinlock_t *root_lock;
 656
 657         root_lock = qdisc_lock(oqdisc);
 658         spin_lock_bh(root_lock);
 659
 660         /* Prune old scheduler */
 661         if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
 662                 qdisc_reset(oqdisc);
 663
 664         /* ... and graft new one */
 665         if (qdisc == NULL)
 666                 qdisc = &noop_qdisc;
 667         dev_queue->qdisc_sleeping = qdisc;
 668         rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 669
 670         spin_unlock_bh(root_lock);
 671
 672         return oqdisc;
 673 }
 674 EXPORT_SYMBOL(dev_graft_qdisc);
 675
 676 static void attach_one_default_qdisc(struct net_device *dev,
 677                                      struct netdev_queue *dev_queue,
 678                                      void *_unused)
 679 {
 680         struct Qdisc *qdisc = &noqueue_qdisc;
 681
 682         if (dev->tx_queue_len) {
 683                 qdisc = qdisc_create_dflt(dev_queue,
 684                                           &pfifo_fast_ops, TC_H_ROOT);
 685                 if (!qdisc) {
 686                         netdev_info(dev, "activation failed\n");
 687                         return;
 688                 }
 689                 if (!netif_is_multiqueue(dev))
 690                         qdisc->flags |= TCQ_F_ONETXQUEUE;
 691         }
 692         dev_queue->qdisc_sleeping = qdisc;
 693 }
 694
 695 static void attach_default_qdiscs(struct net_device *dev)
 696 {
 697         struct netdev_queue *txq;
 698         struct Qdisc *qdisc;
 699
 700         txq = netdev_get_tx_queue(dev, 0);
 701
 702         if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 703                 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 704                 dev->qdisc = txq->qdisc_sleeping;
 705                 atomic_inc(&dev->qdisc->refcnt);
 706         } else {
 707                 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
 708                 if (qdisc) {
 709                         qdisc->ops->attach(qdisc);
 710                         dev->qdisc = qdisc;
 711                 }
 712         }
 713 }
 714
 715 static void transition_one_qdisc(struct net_device *dev,
 716                                  struct netdev_queue *dev_queue,
 717                                  void *_need_watchdog)
 718 {
 719         struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
 720         int *need_watchdog_p = _need_watchdog;
 721
 722         if (!(new_qdisc->flags & TCQ_F_BUILTIN))
 723                 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
 724
 725         rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
 726         if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
 727                 dev_queue->trans_start = 0;
 728                 *need_watchdog_p = 1;
 729         }
 730 }
 731
 732 void dev_activate(struct net_device *dev)
 733 {
 734         int need_watchdog;
 735
 736         /* No queueing discipline is attached to device;
 737            create default one i.e. pfifo_fast for devices,
 738            which need queueing and noqueue_qdisc for
 739            virtual interfaces
 740          */
 741
 742         if (dev->qdisc == &noop_qdisc)
 743                 attach_default_qdiscs(dev);
 744
 745         if (!netif_carrier_ok(dev))
 746                 /* Delay activation until next carrier-on event */
 747                 return;
 748
 749         need_watchdog = 0;
 750         netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
 751         if (dev_ingress_queue(dev))
 752                 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
 753
 754         if (need_watchdog) {
 755                 dev->trans_start = jiffies;
 756                 dev_watchdog_up(dev);
 757         }
 758 }
 759 EXPORT_SYMBOL(dev_activate);
 760
 761 static void dev_deactivate_queue(struct net_device *dev,
 762                                  struct netdev_queue *dev_queue,
 763                                  void *_qdisc_default)
 764 {
 765         struct Qdisc *qdisc_default = _qdisc_default;
 766         struct Qdisc *qdisc;
 767
 768         qdisc = dev_queue->qdisc;
 769         if (qdisc) {
 770                 spin_lock_bh(qdisc_lock(qdisc));
 771
 772                 if (!(qdisc->flags & TCQ_F_BUILTIN))
 773                         set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
 774
 775                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 776                 qdisc_reset(qdisc);
 777
 778                 spin_unlock_bh(qdisc_lock(qdisc));
 779         }
 780 }
 781
 782 static bool some_qdisc_is_busy(struct net_device *dev)
 783 {
 784         unsigned int i;
 785
 786         for (i = 0; i < dev->num_tx_queues; i++) {
 787                 struct netdev_queue *dev_queue;
 788                 spinlock_t *root_lock;
 789                 struct Qdisc *q;
 790                 int val;
 791
 792                 dev_queue = netdev_get_tx_queue(dev, i);
 793                 q = dev_queue->qdisc_sleeping;
 794                 root_lock = qdisc_lock(q);
 795
 796                 spin_lock_bh(root_lock);
 797
 798                 val = (qdisc_is_running(q) ||
 799                        test_bit(__QDISC_STATE_SCHED, &q->state));
 800
 801                 spin_unlock_bh(root_lock);
 802
 803                 if (val)
 804                         return true;
 805         }
 806         return false;
 807 }
 808
 809 /**
 810  *      dev_deactivate_many - deactivate transmissions on several devices
 811  *      @head: list of devices to deactivate
 812  *
 813  *      This function returns only when all outstanding transmissions
 814  *      have completed, unless all devices are in dismantle phase.
 815  */
 816 void dev_deactivate_many(struct list_head *head)
 817 {
 818         struct net_device *dev;
 819         bool sync_needed = false;
 820
 821         list_for_each_entry(dev, head, unreg_list) {
 822                 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
 823                                          &noop_qdisc);
 824                 if (dev_ingress_queue(dev))
 825                         dev_deactivate_queue(dev, dev_ingress_queue(dev),
 826                                              &noop_qdisc);
 827
 828                 dev_watchdog_down(dev);
 829                 sync_needed |= !dev->dismantle;
 830         }
 831
 832         /* Wait for outstanding qdisc-less dev_queue_xmit calls.
 833          * This is avoided if all devices are in dismantle phase :
 834          * Caller will call synchronize_net() for us
 835          */
 836         if (sync_needed)
 837                 synchronize_net();
 838
 839         /* Wait for outstanding qdisc_run calls. */
 840         list_for_each_entry(dev, head, unreg_list)
 841                 while (some_qdisc_is_busy(dev))
 842                         yield();
 843 }
 844
 845 void dev_deactivate(struct net_device *dev)
 846 {
 847         LIST_HEAD(single);
 848
 849         list_add(&dev->unreg_list, &single);
 850         dev_deactivate_many(&single);
 851         list_del(&single);
 852 }
 853 EXPORT_SYMBOL(dev_deactivate);
 854
 855 static void dev_init_scheduler_queue(struct net_device *dev,
 856                                      struct netdev_queue *dev_queue,
 857                                      void *_qdisc)
 858 {
 859         struct Qdisc *qdisc = _qdisc;
 860
 861         dev_queue->qdisc = qdisc;
 862         dev_queue->qdisc_sleeping = qdisc;
 863 }
 864
 865 void dev_init_scheduler(struct net_device *dev)
 866 {
 867         dev->qdisc = &noop_qdisc;
 868         netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
 869         if (dev_ingress_queue(dev))
 870                 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 871
 872         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 873 }
 874
 875 static void shutdown_scheduler_queue(struct net_device *dev,
 876                                      struct netdev_queue *dev_queue,
 877                                      void *_qdisc_default)
 878 {
 879         struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
 880         struct Qdisc *qdisc_default = _qdisc_default;
 881
 882         if (qdisc) {
 883                 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
 884                 dev_queue->qdisc_sleeping = qdisc_default;
 885
 886                 qdisc_destroy(qdisc);
 887         }
 888 }
 889
 890 void dev_shutdown(struct net_device *dev)
 891 {
 892         netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
 893         if (dev_ingress_queue(dev))
 894                 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
 895         qdisc_destroy(dev->qdisc);
 896         dev->qdisc = &noop_qdisc;
 897
 898         WARN_ON(timer_pending(&dev->watchdog_timer));
 899 }
 900
 901 void psched_ratecfg_precompute(struct psched_ratecfg *r, u32 rate)
 902 {
 903         u64 factor;
 904         u64 mult;
 905         int shift;
 906
 907         r->rate_bps = (u64)rate << 3;
 908         r->shift = 0;
 909         r->mult = 1;
 910         /*
 911          * Calibrate mult, shift so that token counting is accurate
 912          * for smallest packet size (64 bytes).  Token (time in ns) is
 913          * computed as (bytes * 8) * NSEC_PER_SEC / rate_bps.  It will
 914          * work as long as the smallest packet transfer time can be
 915          * accurately represented in nanosec.
 916          */
 917         if (r->rate_bps > 0) {
 918                 /*
 919                  * Higher shift gives better accuracy.  Find the largest
 920                  * shift such that mult fits in 32 bits.
 921                  */
 922                 for (shift = 0; shift < 16; shift++) {
 923                         r->shift = shift;
 924                         factor = 8LLU * NSEC_PER_SEC * (1 << r->shift);
 925                         mult = div64_u64(factor, r->rate_bps);
 926                         if (mult > UINT_MAX)
 927                                 break;
 928                 }
 929
 930                 r->shift = shift - 1;
 931                 factor = 8LLU * NSEC_PER_SEC * (1 << r->shift);
 932                 r->mult = div64_u64(factor, r->rate_bps);
 933         }
 934 }
 935 EXPORT_SYMBOL(psched_ratecfg_precompute);