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 <asm/uaccess.h>
  15 #include <asm/system.h>
  16 #include <linux/bitops.h>
  17 #include <linux/config.h>
  18 #include <linux/module.h>
  19 #include <linux/types.h>
  20 #include <linux/kernel.h>
  21 #include <linux/sched.h>
  22 #include <linux/string.h>
  23 #include <linux/mm.h>
  24 #include <linux/socket.h>
  25 #include <linux/sockios.h>
  26 #include <linux/in.h>
  27 #include <linux/errno.h>
  28 #include <linux/interrupt.h>
  29 #include <linux/netdevice.h>
  30 #include <linux/skbuff.h>
  31 #include <linux/rtnetlink.h>
  32 #include <linux/init.h>
  33 #include <linux/rcupdate.h>
  34 #include <linux/list.h>
  35 #include <net/sock.h>
  36 #include <net/pkt_sched.h>
  37
  38 /* Main transmission queue. */
  39
  40 /* Main qdisc structure lock.
  41
  42    However, modifications
  43    to data, participating in scheduling must be additionally
  44    protected with dev->queue_lock spinlock.
  45
  46    The idea is the following:
  47    - enqueue, dequeue are serialized via top level device
  48      spinlock dev->queue_lock.
  49    - tree walking is protected by read_lock_bh(qdisc_tree_lock)
  50      and this lock is used only in process context.
  51    - updates to tree are made under rtnl semaphore or
  52      from softirq context (__qdisc_destroy rcu-callback)
  53      hence this lock needs local bh disabling.
  54
  55    qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
  56  */
  57 DEFINE_RWLOCK(qdisc_tree_lock);
  58
  59 void qdisc_lock_tree(struct net_device *dev)
  60 {
  61         write_lock_bh(&qdisc_tree_lock);
  62         spin_lock_bh(&dev->queue_lock);
  63 }
  64
  65 void qdisc_unlock_tree(struct net_device *dev)
  66 {
  67         spin_unlock_bh(&dev->queue_lock);
  68         write_unlock_bh(&qdisc_tree_lock);
  69 }
  70
  71 /*
  72    dev->queue_lock serializes queue accesses for this device
  73    AND dev->qdisc pointer itself.
  74
  75    dev->xmit_lock serializes accesses to device driver.
  76
  77    dev->queue_lock and dev->xmit_lock are mutually exclusive,
  78    if one is grabbed, another must be free.
  79  */
  80
  81
  82 /* Kick device.
  83    Note, that this procedure can be called by a watchdog timer, so that
  84    we do not check dev->tbusy flag here.
  85
  86    Returns:  0  - queue is empty.
  87             >0  - queue is not empty, but throttled.
  88             <0  - queue is not empty. Device is throttled, if dev->tbusy != 0.
  89
  90    NOTE: Called under dev->queue_lock with locally disabled BH.
  91 */
  92
  93 int qdisc_restart(struct net_device *dev)
  94 {
  95         struct Qdisc *q = dev->qdisc;
  96         struct sk_buff *skb;
  97
  98         /* Dequeue packet */
  99         if ((skb = q->dequeue(q)) != NULL) {
 100                 unsigned nolock = (dev->features & NETIF_F_LLTX);
 101                 /*
 102                  * When the driver has LLTX set it does its own locking
 103                  * in start_xmit. No need to add additional overhead by
 104                  * locking again. These checks are worth it because
 105                  * even uncongested locks can be quite expensive.
 106                  * The driver can do trylock like here too, in case
 107                  * of lock congestion it should return -1 and the packet
 108                  * will be requeued.
 109                  */
 110                 if (!nolock) {
 111                         if (!spin_trylock(&dev->xmit_lock)) {
 112                         collision:
 113                                 /* So, someone grabbed the driver. */
 114
 115                                 /* It may be transient configuration error,
 116                                    when hard_start_xmit() recurses. We detect
 117                                    it by checking xmit owner and drop the
 118                                    packet when deadloop is detected.
 119                                 */
 120                                 if (dev->xmit_lock_owner == smp_processor_id()) {
 121                                         kfree_skb(skb);
 122                                         if (net_ratelimit())
 123                                                 printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
 124                                         return -1;
 125                                 }
 126                                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
 127                                 goto requeue;
 128                         }
 129                         /* Remember that the driver is grabbed by us. */
 130                         dev->xmit_lock_owner = smp_processor_id();
 131                 }
 132
 133                 {
 134                         /* And release queue */
 135                         spin_unlock(&dev->queue_lock);
 136
 137                         if (!netif_queue_stopped(dev)) {
 138                                 int ret;
 139                                 if (netdev_nit)
 140                                         dev_queue_xmit_nit(skb, dev);
 141
 142                                 ret = dev->hard_start_xmit(skb, dev);
 143                                 if (ret == NETDEV_TX_OK) {
 144                                         if (!nolock) {
 145                                                 dev->xmit_lock_owner = -1;
 146                                                 spin_unlock(&dev->xmit_lock);
 147                                         }
 148                                         spin_lock(&dev->queue_lock);
 149                                         return -1;
 150                                 }
 151                                 if (ret == NETDEV_TX_LOCKED && nolock) {
 152                                         spin_lock(&dev->queue_lock);
 153                                         goto collision;
 154                                 }
 155                         }
 156
 157                         /* NETDEV_TX_BUSY - we need to requeue */
 158                         /* Release the driver */
 159                         if (!nolock) {
 160                                 dev->xmit_lock_owner = -1;
 161                                 spin_unlock(&dev->xmit_lock);
 162                         }
 163                         spin_lock(&dev->queue_lock);
 164                         q = dev->qdisc;
 165                 }
 166
 167                 /* Device kicked us out :(
 168                    This is possible in three cases:
 169
 170                    0. driver is locked
 171                    1. fastroute is enabled
 172                    2. device cannot determine busy state
 173                       before start of transmission (f.e. dialout)
 174                    3. device is buggy (ppp)
 175                  */
 176
 177 requeue:
 178                 q->ops->requeue(skb, q);
 179                 netif_schedule(dev);
 180                 return 1;
 181         }
 182         return q->q.qlen;
 183 }
 184
 185 static void dev_watchdog(unsigned long arg)
 186 {
 187         struct net_device *dev = (struct net_device *)arg;
 188
 189         spin_lock(&dev->xmit_lock);
 190         if (dev->qdisc != &noop_qdisc) {
 191                 if (netif_device_present(dev) &&
 192                     netif_running(dev) &&
 193                     netif_carrier_ok(dev)) {
 194                         if (netif_queue_stopped(dev) &&
 195                             (jiffies - dev->trans_start) > dev->watchdog_timeo) {
 196                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n", dev->name);
 197                                 dev->tx_timeout(dev);
 198                         }
 199                         if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
 200                                 dev_hold(dev);
 201                 }
 202         }
 203         spin_unlock(&dev->xmit_lock);
 204
 205         dev_put(dev);
 206 }
 207
 208 static void dev_watchdog_init(struct net_device *dev)
 209 {
 210         init_timer(&dev->watchdog_timer);
 211         dev->watchdog_timer.data = (unsigned long)dev;
 212         dev->watchdog_timer.function = dev_watchdog;
 213 }
 214
 215 void __netdev_watchdog_up(struct net_device *dev)
 216 {
 217         if (dev->tx_timeout) {
 218                 if (dev->watchdog_timeo <= 0)
 219                         dev->watchdog_timeo = 5*HZ;
 220                 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
 221                         dev_hold(dev);
 222         }
 223 }
 224
 225 static void dev_watchdog_up(struct net_device *dev)
 226 {
 227         spin_lock_bh(&dev->xmit_lock);
 228         __netdev_watchdog_up(dev);
 229         spin_unlock_bh(&dev->xmit_lock);
 230 }
 231
 232 static void dev_watchdog_down(struct net_device *dev)
 233 {
 234         spin_lock_bh(&dev->xmit_lock);
 235         if (del_timer(&dev->watchdog_timer))
 236                 __dev_put(dev);
 237         spin_unlock_bh(&dev->xmit_lock);
 238 }
 239
 240 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 241    under all circumstances. It is difficult to invent anything faster or
 242    cheaper.
 243  */
 244
 245 static int
 246 noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 247 {
 248         kfree_skb(skb);
 249         return NET_XMIT_CN;
 250 }
 251
 252 static struct sk_buff *
 253 noop_dequeue(struct Qdisc * qdisc)
 254 {
 255         return NULL;
 256 }
 257
 258 static int
 259 noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 260 {
 261         if (net_ratelimit())
 262                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n", skb->dev->name);
 263         kfree_skb(skb);
 264         return NET_XMIT_CN;
 265 }
 266
 267 struct Qdisc_ops noop_qdisc_ops = {
 268         .next           =       NULL,
 269         .cl_ops         =       NULL,
 270         .id             =       "noop",
 271         .priv_size      =       0,
 272         .enqueue        =       noop_enqueue,
 273         .dequeue        =       noop_dequeue,
 274         .requeue        =       noop_requeue,
 275         .owner          =       THIS_MODULE,
 276 };
 277
 278 struct Qdisc noop_qdisc = {
 279         .enqueue        =       noop_enqueue,
 280         .dequeue        =       noop_dequeue,
 281         .flags          =       TCQ_F_BUILTIN,
 282         .ops            =       &noop_qdisc_ops,
 283         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 284 };
 285
 286 static struct Qdisc_ops noqueue_qdisc_ops = {
 287         .next           =       NULL,
 288         .cl_ops         =       NULL,
 289         .id             =       "noqueue",
 290         .priv_size      =       0,
 291         .enqueue        =       noop_enqueue,
 292         .dequeue        =       noop_dequeue,
 293         .requeue        =       noop_requeue,
 294         .owner          =       THIS_MODULE,
 295 };
 296
 297 static struct Qdisc noqueue_qdisc = {
 298         .enqueue        =       NULL,
 299         .dequeue        =       noop_dequeue,
 300         .flags          =       TCQ_F_BUILTIN,
 301         .ops            =       &noqueue_qdisc_ops,
 302         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 303 };
 304
 305
 306 static const u8 prio2band[TC_PRIO_MAX+1] =
 307         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 308
 309 /* 3-band FIFO queue: old style, but should be a bit faster than
 310    generic prio+fifo combination.
 311  */
 312
 313 static int
 314 pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 315 {
 316         struct sk_buff_head *list = qdisc_priv(qdisc);
 317
 318         list += prio2band[skb->priority&TC_PRIO_MAX];
 319
 320         if (list->qlen < qdisc->dev->tx_queue_len) {
 321                 __skb_queue_tail(list, skb);
 322                 qdisc->q.qlen++;
 323                 qdisc->bstats.bytes += skb->len;
 324                 qdisc->bstats.packets++;
 325                 return 0;
 326         }
 327         qdisc->qstats.drops++;
 328         kfree_skb(skb);
 329         return NET_XMIT_DROP;
 330 }
 331
 332 static struct sk_buff *
 333 pfifo_fast_dequeue(struct Qdisc* qdisc)
 334 {
 335         int prio;
 336         struct sk_buff_head *list = qdisc_priv(qdisc);
 337         struct sk_buff *skb;
 338
 339         for (prio = 0; prio < 3; prio++, list++) {
 340                 skb = __skb_dequeue(list);
 341                 if (skb) {
 342                         qdisc->q.qlen--;
 343                         return skb;
 344                 }
 345         }
 346         return NULL;
 347 }
 348
 349 static int
 350 pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 351 {
 352         struct sk_buff_head *list = qdisc_priv(qdisc);
 353
 354         list += prio2band[skb->priority&TC_PRIO_MAX];
 355
 356         __skb_queue_head(list, skb);
 357         qdisc->q.qlen++;
 358         qdisc->qstats.requeues++;
 359         return 0;
 360 }
 361
 362 static void
 363 pfifo_fast_reset(struct Qdisc* qdisc)
 364 {
 365         int prio;
 366         struct sk_buff_head *list = qdisc_priv(qdisc);
 367
 368         for (prio=0; prio < 3; prio++)
 369                 skb_queue_purge(list+prio);
 370         qdisc->q.qlen = 0;
 371 }
 372
 373 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 374 {
 375         unsigned char    *b = skb->tail;
 376         struct tc_prio_qopt opt;
 377
 378         opt.bands = 3;
 379         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 380         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 381         return skb->len;
 382
 383 rtattr_failure:
 384         skb_trim(skb, b - skb->data);
 385         return -1;
 386 }
 387
 388 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 389 {
 390         int i;
 391         struct sk_buff_head *list = qdisc_priv(qdisc);
 392
 393         for (i=0; i<3; i++)
 394                 skb_queue_head_init(list+i);
 395
 396         return 0;
 397 }
 398
 399 static struct Qdisc_ops pfifo_fast_ops = {
 400         .next           =       NULL,
 401         .cl_ops         =       NULL,
 402         .id             =       "pfifo_fast",
 403         .priv_size      =       3 * sizeof(struct sk_buff_head),
 404         .enqueue        =       pfifo_fast_enqueue,
 405         .dequeue        =       pfifo_fast_dequeue,
 406         .requeue        =       pfifo_fast_requeue,
 407         .init           =       pfifo_fast_init,
 408         .reset          =       pfifo_fast_reset,
 409         .dump           =       pfifo_fast_dump,
 410         .owner          =       THIS_MODULE,
 411 };
 412
 413 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
 414 {
 415         void *p;
 416         struct Qdisc *sch;
 417         int size;
 418
 419         /* ensure that the Qdisc and the private data are 32-byte aligned */
 420         size = ((sizeof(*sch) + QDISC_ALIGN_CONST) & ~QDISC_ALIGN_CONST);
 421         size += ops->priv_size + QDISC_ALIGN_CONST;
 422
 423         p = kmalloc(size, GFP_KERNEL);
 424         if (!p)
 425                 return NULL;
 426         memset(p, 0, size);
 427
 428         sch = (struct Qdisc *)(((unsigned long)p + QDISC_ALIGN_CONST)
 429                                & ~QDISC_ALIGN_CONST);
 430         sch->padded = (char *)sch - (char *)p;
 431
 432         INIT_LIST_HEAD(&sch->list);
 433         skb_queue_head_init(&sch->q);
 434         sch->ops = ops;
 435         sch->enqueue = ops->enqueue;
 436         sch->dequeue = ops->dequeue;
 437         sch->dev = dev;
 438         dev_hold(dev);
 439         sch->stats_lock = &dev->queue_lock;
 440         atomic_set(&sch->refcnt, 1);
 441         if (!ops->init || ops->init(sch, NULL) == 0)
 442                 return sch;
 443
 444         dev_put(dev);
 445         kfree(p);
 446         return NULL;
 447 }
 448
 449 /* Under dev->queue_lock and BH! */
 450
 451 void qdisc_reset(struct Qdisc *qdisc)
 452 {
 453         struct Qdisc_ops *ops = qdisc->ops;
 454
 455         if (ops->reset)
 456                 ops->reset(qdisc);
 457 }
 458
 459 /* this is the rcu callback function to clean up a qdisc when there
 460  * are no further references to it */
 461
 462 static void __qdisc_destroy(struct rcu_head *head)
 463 {
 464         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 465         struct Qdisc_ops  *ops = qdisc->ops;
 466
 467 #ifdef CONFIG_NET_ESTIMATOR
 468         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 469 #endif
 470         write_lock(&qdisc_tree_lock);
 471         if (ops->reset)
 472                 ops->reset(qdisc);
 473         if (ops->destroy)
 474                 ops->destroy(qdisc);
 475         write_unlock(&qdisc_tree_lock);
 476         module_put(ops->owner);
 477
 478         dev_put(qdisc->dev);
 479         kfree((char *) qdisc - qdisc->padded);
 480 }
 481
 482 /* Under dev->queue_lock and BH! */
 483
 484 void qdisc_destroy(struct Qdisc *qdisc)
 485 {
 486         struct list_head cql = LIST_HEAD_INIT(cql);
 487         struct Qdisc *cq, *q, *n;
 488
 489         if (qdisc->flags & TCQ_F_BUILTIN ||
 490                 !atomic_dec_and_test(&qdisc->refcnt))
 491                 return;
 492
 493         if (!list_empty(&qdisc->list)) {
 494                 if (qdisc->ops->cl_ops == NULL)
 495                         list_del(&qdisc->list);
 496                 else
 497                         list_move(&qdisc->list, &cql);
 498         }
 499
 500         /* unlink inner qdiscs from dev->qdisc_list immediately */
 501         list_for_each_entry(cq, &cql, list)
 502                 list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
 503                         if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
 504                                 if (q->ops->cl_ops == NULL)
 505                                         list_del_init(&q->list);
 506                                 else
 507                                         list_move_tail(&q->list, &cql);
 508                         }
 509         list_for_each_entry_safe(cq, n, &cql, list)
 510                 list_del_init(&cq->list);
 511
 512         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 513 }
 514
 515 void dev_activate(struct net_device *dev)
 516 {
 517         /* No queueing discipline is attached to device;
 518            create default one i.e. pfifo_fast for devices,
 519            which need queueing and noqueue_qdisc for
 520            virtual interfaces
 521          */
 522
 523         if (dev->qdisc_sleeping == &noop_qdisc) {
 524                 struct Qdisc *qdisc;
 525                 if (dev->tx_queue_len) {
 526                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
 527                         if (qdisc == NULL) {
 528                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 529                                 return;
 530                         }
 531                         write_lock_bh(&qdisc_tree_lock);
 532                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 533                         write_unlock_bh(&qdisc_tree_lock);
 534                 } else {
 535                         qdisc =  &noqueue_qdisc;
 536                 }
 537                 write_lock_bh(&qdisc_tree_lock);
 538                 dev->qdisc_sleeping = qdisc;
 539                 write_unlock_bh(&qdisc_tree_lock);
 540         }
 541
 542         spin_lock_bh(&dev->queue_lock);
 543         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 544         if (dev->qdisc != &noqueue_qdisc) {
 545                 dev->trans_start = jiffies;
 546                 dev_watchdog_up(dev);
 547         }
 548         spin_unlock_bh(&dev->queue_lock);
 549 }
 550
 551 void dev_deactivate(struct net_device *dev)
 552 {
 553         struct Qdisc *qdisc;
 554
 555         spin_lock_bh(&dev->queue_lock);
 556         qdisc = dev->qdisc;
 557         dev->qdisc = &noop_qdisc;
 558
 559         qdisc_reset(qdisc);
 560
 561         spin_unlock_bh(&dev->queue_lock);
 562
 563         dev_watchdog_down(dev);
 564
 565         while (test_bit(__LINK_STATE_SCHED, &dev->state))
 566                 yield();
 567
 568         spin_unlock_wait(&dev->xmit_lock);
 569 }
 570
 571 void dev_init_scheduler(struct net_device *dev)
 572 {
 573         qdisc_lock_tree(dev);
 574         dev->qdisc = &noop_qdisc;
 575         dev->qdisc_sleeping = &noop_qdisc;
 576         INIT_LIST_HEAD(&dev->qdisc_list);
 577         qdisc_unlock_tree(dev);
 578
 579         dev_watchdog_init(dev);
 580 }
 581
 582 void dev_shutdown(struct net_device *dev)
 583 {
 584         struct Qdisc *qdisc;
 585
 586         qdisc_lock_tree(dev);
 587         qdisc = dev->qdisc_sleeping;
 588         dev->qdisc = &noop_qdisc;
 589         dev->qdisc_sleeping = &noop_qdisc;
 590         qdisc_destroy(qdisc);
 591 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 592         if ((qdisc = dev->qdisc_ingress) != NULL) {
 593                 dev->qdisc_ingress = NULL;
 594                 qdisc_destroy(qdisc);
 595         }
 596 #endif
 597         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 598         qdisc_unlock_tree(dev);
 599 }
 600
 601 EXPORT_SYMBOL(__netdev_watchdog_up);
 602 EXPORT_SYMBOL(noop_qdisc);
 603 EXPORT_SYMBOL(noop_qdisc_ops);
 604 EXPORT_SYMBOL(qdisc_create_dflt);
 605 EXPORT_SYMBOL(qdisc_destroy);
 606 EXPORT_SYMBOL(qdisc_reset);
 607 EXPORT_SYMBOL(qdisc_restart);
 608 EXPORT_SYMBOL(qdisc_lock_tree);
 609 EXPORT_SYMBOL(qdisc_unlock_tree);