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