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         BUG_ON((int) q->q.qlen < 0);
 183         return q->q.qlen;
 184 }
 185
 186 static void dev_watchdog(unsigned long arg)
 187 {
 188         struct net_device *dev = (struct net_device *)arg;
 189
 190         spin_lock(&dev->xmit_lock);
 191         if (dev->qdisc != &noop_qdisc) {
 192                 if (netif_device_present(dev) &&
 193                     netif_running(dev) &&
 194                     netif_carrier_ok(dev)) {
 195                         if (netif_queue_stopped(dev) &&
 196                             (jiffies - dev->trans_start) > dev->watchdog_timeo) {
 197                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n", dev->name);
 198                                 dev->tx_timeout(dev);
 199                         }
 200                         if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
 201                                 dev_hold(dev);
 202                 }
 203         }
 204         spin_unlock(&dev->xmit_lock);
 205
 206         dev_put(dev);
 207 }
 208
 209 static void dev_watchdog_init(struct net_device *dev)
 210 {
 211         init_timer(&dev->watchdog_timer);
 212         dev->watchdog_timer.data = (unsigned long)dev;
 213         dev->watchdog_timer.function = dev_watchdog;
 214 }
 215
 216 void __netdev_watchdog_up(struct net_device *dev)
 217 {
 218         if (dev->tx_timeout) {
 219                 if (dev->watchdog_timeo <= 0)
 220                         dev->watchdog_timeo = 5*HZ;
 221                 if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
 222                         dev_hold(dev);
 223         }
 224 }
 225
 226 static void dev_watchdog_up(struct net_device *dev)
 227 {
 228         spin_lock_bh(&dev->xmit_lock);
 229         __netdev_watchdog_up(dev);
 230         spin_unlock_bh(&dev->xmit_lock);
 231 }
 232
 233 static void dev_watchdog_down(struct net_device *dev)
 234 {
 235         spin_lock_bh(&dev->xmit_lock);
 236         if (del_timer(&dev->watchdog_timer))
 237                 __dev_put(dev);
 238         spin_unlock_bh(&dev->xmit_lock);
 239 }
 240
 241 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 242    under all circumstances. It is difficult to invent anything faster or
 243    cheaper.
 244  */
 245
 246 static int
 247 noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 248 {
 249         kfree_skb(skb);
 250         return NET_XMIT_CN;
 251 }
 252
 253 static struct sk_buff *
 254 noop_dequeue(struct Qdisc * qdisc)
 255 {
 256         return NULL;
 257 }
 258
 259 static int
 260 noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 261 {
 262         if (net_ratelimit())
 263                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n", skb->dev->name);
 264         kfree_skb(skb);
 265         return NET_XMIT_CN;
 266 }
 267
 268 struct Qdisc_ops noop_qdisc_ops = {
 269         .next           =       NULL,
 270         .cl_ops         =       NULL,
 271         .id             =       "noop",
 272         .priv_size      =       0,
 273         .enqueue        =       noop_enqueue,
 274         .dequeue        =       noop_dequeue,
 275         .requeue        =       noop_requeue,
 276         .owner          =       THIS_MODULE,
 277 };
 278
 279 struct Qdisc noop_qdisc = {
 280         .enqueue        =       noop_enqueue,
 281         .dequeue        =       noop_dequeue,
 282         .flags          =       TCQ_F_BUILTIN,
 283         .ops            =       &noop_qdisc_ops,
 284         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 285 };
 286
 287 static struct Qdisc_ops noqueue_qdisc_ops = {
 288         .next           =       NULL,
 289         .cl_ops         =       NULL,
 290         .id             =       "noqueue",
 291         .priv_size      =       0,
 292         .enqueue        =       noop_enqueue,
 293         .dequeue        =       noop_dequeue,
 294         .requeue        =       noop_requeue,
 295         .owner          =       THIS_MODULE,
 296 };
 297
 298 static struct Qdisc noqueue_qdisc = {
 299         .enqueue        =       NULL,
 300         .dequeue        =       noop_dequeue,
 301         .flags          =       TCQ_F_BUILTIN,
 302         .ops            =       &noqueue_qdisc_ops,
 303         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 304 };
 305
 306
 307 static const u8 prio2band[TC_PRIO_MAX+1] =
 308         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 309
 310 /* 3-band FIFO queue: old style, but should be a bit faster than
 311    generic prio+fifo combination.
 312  */
 313
 314 #define PFIFO_FAST_BANDS 3
 315
 316 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 317                                              struct Qdisc *qdisc)
 318 {
 319         struct sk_buff_head *list = qdisc_priv(qdisc);
 320         return list + prio2band[skb->priority & TC_PRIO_MAX];
 321 }
 322
 323 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 324 {
 325         struct sk_buff_head *list = prio2list(skb, qdisc);
 326
 327         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 328                 qdisc->q.qlen++;
 329                 return __qdisc_enqueue_tail(skb, qdisc, list);
 330         }
 331
 332         return qdisc_drop(skb, qdisc);
 333 }
 334
 335 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 336 {
 337         int prio;
 338         struct sk_buff_head *list = qdisc_priv(qdisc);
 339
 340         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++, list++) {
 341                 struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
 342                 if (skb) {
 343                         qdisc->q.qlen--;
 344                         return skb;
 345                 }
 346         }
 347
 348         return NULL;
 349 }
 350
 351 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 352 {
 353         qdisc->q.qlen++;
 354         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 355 }
 356
 357 static void pfifo_fast_reset(struct Qdisc* qdisc)
 358 {
 359         int prio;
 360         struct sk_buff_head *list = qdisc_priv(qdisc);
 361
 362         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 363                 __qdisc_reset_queue(qdisc, list + prio);
 364
 365         qdisc->qstats.backlog = 0;
 366         qdisc->q.qlen = 0;
 367 }
 368
 369 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 370 {
 371         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 372
 373         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 374         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 375         return skb->len;
 376
 377 rtattr_failure:
 378         return -1;
 379 }
 380
 381 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 382 {
 383         int prio;
 384         struct sk_buff_head *list = qdisc_priv(qdisc);
 385
 386         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 387                 skb_queue_head_init(list + prio);
 388
 389         return 0;
 390 }
 391
 392 static struct Qdisc_ops pfifo_fast_ops = {
 393         .id             =       "pfifo_fast",
 394         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 395         .enqueue        =       pfifo_fast_enqueue,
 396         .dequeue        =       pfifo_fast_dequeue,
 397         .requeue        =       pfifo_fast_requeue,
 398         .init           =       pfifo_fast_init,
 399         .reset          =       pfifo_fast_reset,
 400         .dump           =       pfifo_fast_dump,
 401         .owner          =       THIS_MODULE,
 402 };
 403
 404 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
 405 {
 406         void *p;
 407         struct Qdisc *sch;
 408         int size;
 409
 410         /* ensure that the Qdisc and the private data are 32-byte aligned */
 411         size = ((sizeof(*sch) + QDISC_ALIGN_CONST) & ~QDISC_ALIGN_CONST);
 412         size += ops->priv_size + QDISC_ALIGN_CONST;
 413
 414         p = kmalloc(size, GFP_KERNEL);
 415         if (!p)
 416                 return NULL;
 417         memset(p, 0, size);
 418
 419         sch = (struct Qdisc *)(((unsigned long)p + QDISC_ALIGN_CONST)
 420                                & ~QDISC_ALIGN_CONST);
 421         sch->padded = (char *)sch - (char *)p;
 422
 423         INIT_LIST_HEAD(&sch->list);
 424         skb_queue_head_init(&sch->q);
 425         sch->ops = ops;
 426         sch->enqueue = ops->enqueue;
 427         sch->dequeue = ops->dequeue;
 428         sch->dev = dev;
 429         dev_hold(dev);
 430         sch->stats_lock = &dev->queue_lock;
 431         atomic_set(&sch->refcnt, 1);
 432         if (!ops->init || ops->init(sch, NULL) == 0)
 433                 return sch;
 434
 435         dev_put(dev);
 436         kfree(p);
 437         return NULL;
 438 }
 439
 440 /* Under dev->queue_lock and BH! */
 441
 442 void qdisc_reset(struct Qdisc *qdisc)
 443 {
 444         struct Qdisc_ops *ops = qdisc->ops;
 445
 446         if (ops->reset)
 447                 ops->reset(qdisc);
 448 }
 449
 450 /* this is the rcu callback function to clean up a qdisc when there
 451  * are no further references to it */
 452
 453 static void __qdisc_destroy(struct rcu_head *head)
 454 {
 455         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 456         struct Qdisc_ops  *ops = qdisc->ops;
 457
 458 #ifdef CONFIG_NET_ESTIMATOR
 459         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 460 #endif
 461         write_lock(&qdisc_tree_lock);
 462         if (ops->reset)
 463                 ops->reset(qdisc);
 464         if (ops->destroy)
 465                 ops->destroy(qdisc);
 466         write_unlock(&qdisc_tree_lock);
 467         module_put(ops->owner);
 468
 469         dev_put(qdisc->dev);
 470         kfree((char *) qdisc - qdisc->padded);
 471 }
 472
 473 /* Under dev->queue_lock and BH! */
 474
 475 void qdisc_destroy(struct Qdisc *qdisc)
 476 {
 477         struct list_head cql = LIST_HEAD_INIT(cql);
 478         struct Qdisc *cq, *q, *n;
 479
 480         if (qdisc->flags & TCQ_F_BUILTIN ||
 481                 !atomic_dec_and_test(&qdisc->refcnt))
 482                 return;
 483
 484         if (!list_empty(&qdisc->list)) {
 485                 if (qdisc->ops->cl_ops == NULL)
 486                         list_del(&qdisc->list);
 487                 else
 488                         list_move(&qdisc->list, &cql);
 489         }
 490
 491         /* unlink inner qdiscs from dev->qdisc_list immediately */
 492         list_for_each_entry(cq, &cql, list)
 493                 list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
 494                         if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
 495                                 if (q->ops->cl_ops == NULL)
 496                                         list_del_init(&q->list);
 497                                 else
 498                                         list_move_tail(&q->list, &cql);
 499                         }
 500         list_for_each_entry_safe(cq, n, &cql, list)
 501                 list_del_init(&cq->list);
 502
 503         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 504 }
 505
 506 void dev_activate(struct net_device *dev)
 507 {
 508         /* No queueing discipline is attached to device;
 509            create default one i.e. pfifo_fast for devices,
 510            which need queueing and noqueue_qdisc for
 511            virtual interfaces
 512          */
 513
 514         if (dev->qdisc_sleeping == &noop_qdisc) {
 515                 struct Qdisc *qdisc;
 516                 if (dev->tx_queue_len) {
 517                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
 518                         if (qdisc == NULL) {
 519                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 520                                 return;
 521                         }
 522                         write_lock_bh(&qdisc_tree_lock);
 523                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 524                         write_unlock_bh(&qdisc_tree_lock);
 525                 } else {
 526                         qdisc =  &noqueue_qdisc;
 527                 }
 528                 write_lock_bh(&qdisc_tree_lock);
 529                 dev->qdisc_sleeping = qdisc;
 530                 write_unlock_bh(&qdisc_tree_lock);
 531         }
 532
 533         if (!netif_carrier_ok(dev))
 534                 /* Delay activation until next carrier-on event */
 535                 return;
 536
 537         spin_lock_bh(&dev->queue_lock);
 538         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 539         if (dev->qdisc != &noqueue_qdisc) {
 540                 dev->trans_start = jiffies;
 541                 dev_watchdog_up(dev);
 542         }
 543         spin_unlock_bh(&dev->queue_lock);
 544 }
 545
 546 void dev_deactivate(struct net_device *dev)
 547 {
 548         struct Qdisc *qdisc;
 549
 550         spin_lock_bh(&dev->queue_lock);
 551         qdisc = dev->qdisc;
 552         dev->qdisc = &noop_qdisc;
 553
 554         qdisc_reset(qdisc);
 555
 556         spin_unlock_bh(&dev->queue_lock);
 557
 558         dev_watchdog_down(dev);
 559
 560         while (test_bit(__LINK_STATE_SCHED, &dev->state))
 561                 yield();
 562
 563         spin_unlock_wait(&dev->xmit_lock);
 564 }
 565
 566 void dev_init_scheduler(struct net_device *dev)
 567 {
 568         qdisc_lock_tree(dev);
 569         dev->qdisc = &noop_qdisc;
 570         dev->qdisc_sleeping = &noop_qdisc;
 571         INIT_LIST_HEAD(&dev->qdisc_list);
 572         qdisc_unlock_tree(dev);
 573
 574         dev_watchdog_init(dev);
 575 }
 576
 577 void dev_shutdown(struct net_device *dev)
 578 {
 579         struct Qdisc *qdisc;
 580
 581         qdisc_lock_tree(dev);
 582         qdisc = dev->qdisc_sleeping;
 583         dev->qdisc = &noop_qdisc;
 584         dev->qdisc_sleeping = &noop_qdisc;
 585         qdisc_destroy(qdisc);
 586 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 587         if ((qdisc = dev->qdisc_ingress) != NULL) {
 588                 dev->qdisc_ingress = NULL;
 589                 qdisc_destroy(qdisc);
 590         }
 591 #endif
 592         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 593         qdisc_unlock_tree(dev);
 594 }
 595
 596 EXPORT_SYMBOL(__netdev_watchdog_up);
 597 EXPORT_SYMBOL(noop_qdisc);
 598 EXPORT_SYMBOL(noop_qdisc_ops);
 599 EXPORT_SYMBOL(qdisc_create_dflt);
 600 EXPORT_SYMBOL(qdisc_destroy);
 601 EXPORT_SYMBOL(qdisc_reset);
 602 EXPORT_SYMBOL(qdisc_restart);
 603 EXPORT_SYMBOL(qdisc_lock_tree);
 604 EXPORT_SYMBOL(qdisc_unlock_tree);