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 <net/pkt_sched.h>
  28
  29 /* Main transmission queue. */
  30
  31 /* Modifications to data participating in scheduling must be protected with
  32  * dev->queue_lock spinlock.
  33  *
  34  * The idea is the following:
  35  * - enqueue, dequeue are serialized via top level device
  36  *   spinlock dev->queue_lock.
  37  * - ingress filtering is serialized via top level device
  38  *   spinlock dev->ingress_lock.
  39  * - updates to tree and tree walking are only done under the rtnl mutex.
  40  */
  41
  42 void qdisc_lock_tree(struct net_device *dev)
  43 {
  44         spin_lock_bh(&dev->queue_lock);
  45         spin_lock(&dev->ingress_lock);
  46 }
  47
  48 void qdisc_unlock_tree(struct net_device *dev)
  49 {
  50         spin_unlock(&dev->ingress_lock);
  51         spin_unlock_bh(&dev->queue_lock);
  52 }
  53
  54 static inline int qdisc_qlen(struct Qdisc *q)
  55 {
  56         return q->q.qlen;
  57 }
  58
  59 static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
  60                                   struct Qdisc *q)
  61 {
  62         if (unlikely(skb->next))
  63                 dev->gso_skb = skb;
  64         else
  65                 q->ops->requeue(skb, q);
  66
  67         netif_schedule(dev);
  68         return 0;
  69 }
  70
  71 static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
  72                                               struct Qdisc *q)
  73 {
  74         struct sk_buff *skb;
  75
  76         if ((skb = dev->gso_skb))
  77                 dev->gso_skb = NULL;
  78         else
  79                 skb = q->dequeue(q);
  80
  81         return skb;
  82 }
  83
  84 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  85                                            struct net_device *dev,
  86                                            struct Qdisc *q)
  87 {
  88         int ret;
  89
  90         if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
  91                 /*
  92                  * Same CPU holding the lock. It may be a transient
  93                  * configuration error, when hard_start_xmit() recurses. We
  94                  * detect it by checking xmit owner and drop the packet when
  95                  * deadloop is detected. Return OK to try the next skb.
  96                  */
  97                 kfree_skb(skb);
  98                 if (net_ratelimit())
  99                         printk(KERN_WARNING "Dead loop on netdevice %s, "
 100                                "fix it urgently!\n", dev->name);
 101                 ret = qdisc_qlen(q);
 102         } else {
 103                 /*
 104                  * Another cpu is holding lock, requeue & delay xmits for
 105                  * some time.
 106                  */
 107                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
 108                 ret = dev_requeue_skb(skb, dev, q);
 109         }
 110
 111         return ret;
 112 }
 113
 114 /*
 115  * NOTE: Called under dev->queue_lock with locally disabled BH.
 116  *
 117  * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
 118  * device at a time. dev->queue_lock serializes queue accesses for
 119  * this device AND dev->qdisc pointer itself.
 120  *
 121  *  netif_tx_lock serializes accesses to device driver.
 122  *
 123  *  dev->queue_lock and netif_tx_lock are mutually exclusive,
 124  *  if one is grabbed, another must be free.
 125  *
 126  * Note, that this procedure can be called by a watchdog timer
 127  *
 128  * Returns to the caller:
 129  *                              0  - queue is empty or throttled.
 130  *                              >0 - queue is not empty.
 131  *
 132  */
 133 static inline int qdisc_restart(struct net_device *dev)
 134 {
 135         struct Qdisc *q = dev->qdisc;
 136         struct sk_buff *skb;
 137         int ret = NETDEV_TX_BUSY;
 138
 139         /* Dequeue packet */
 140         if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
 141                 return 0;
 142
 143
 144         /* And release queue */
 145         spin_unlock(&dev->queue_lock);
 146
 147         HARD_TX_LOCK(dev, smp_processor_id());
 148         if (!netif_subqueue_stopped(dev, skb))
 149                 ret = dev_hard_start_xmit(skb, dev);
 150         HARD_TX_UNLOCK(dev);
 151
 152         spin_lock(&dev->queue_lock);
 153         q = dev->qdisc;
 154
 155         switch (ret) {
 156         case NETDEV_TX_OK:
 157                 /* Driver sent out skb successfully */
 158                 ret = qdisc_qlen(q);
 159                 break;
 160
 161         case NETDEV_TX_LOCKED:
 162                 /* Driver try lock failed */
 163                 ret = handle_dev_cpu_collision(skb, dev, q);
 164                 break;
 165
 166         default:
 167                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 168                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 169                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 170                                dev->name, ret, q->q.qlen);
 171
 172                 ret = dev_requeue_skb(skb, dev, q);
 173                 break;
 174         }
 175
 176         return ret;
 177 }
 178
 179 void __qdisc_run(struct net_device *dev)
 180 {
 181         do {
 182                 if (!qdisc_restart(dev))
 183                         break;
 184         } while (!netif_queue_stopped(dev));
 185
 186         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 187 }
 188
 189 static void dev_watchdog(unsigned long arg)
 190 {
 191         struct net_device *dev = (struct net_device *)arg;
 192
 193         netif_tx_lock(dev);
 194         if (dev->qdisc != &noop_qdisc) {
 195                 if (netif_device_present(dev) &&
 196                     netif_running(dev) &&
 197                     netif_carrier_ok(dev)) {
 198                         if (netif_queue_stopped(dev) &&
 199                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 200
 201                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 202                                        dev->name);
 203                                 dev->tx_timeout(dev);
 204                         }
 205                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 206                                 dev_hold(dev);
 207                 }
 208         }
 209         netif_tx_unlock(dev);
 210
 211         dev_put(dev);
 212 }
 213
 214 void __netdev_watchdog_up(struct net_device *dev)
 215 {
 216         if (dev->tx_timeout) {
 217                 if (dev->watchdog_timeo <= 0)
 218                         dev->watchdog_timeo = 5*HZ;
 219                 if (!mod_timer(&dev->watchdog_timer,
 220                                round_jiffies(jiffies + dev->watchdog_timeo)))
 221                         dev_hold(dev);
 222         }
 223 }
 224
 225 static void dev_watchdog_up(struct net_device *dev)
 226 {
 227         __netdev_watchdog_up(dev);
 228 }
 229
 230 static void dev_watchdog_down(struct net_device *dev)
 231 {
 232         netif_tx_lock_bh(dev);
 233         if (del_timer(&dev->watchdog_timer))
 234                 dev_put(dev);
 235         netif_tx_unlock_bh(dev);
 236 }
 237
 238 /**
 239  *      netif_carrier_on - set carrier
 240  *      @dev: network device
 241  *
 242  * Device has detected that carrier.
 243  */
 244 void netif_carrier_on(struct net_device *dev)
 245 {
 246         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 247                 linkwatch_fire_event(dev);
 248                 if (netif_running(dev))
 249                         __netdev_watchdog_up(dev);
 250         }
 251 }
 252
 253 /**
 254  *      netif_carrier_off - clear carrier
 255  *      @dev: network device
 256  *
 257  * Device has detected loss of carrier.
 258  */
 259 void netif_carrier_off(struct net_device *dev)
 260 {
 261         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 262                 linkwatch_fire_event(dev);
 263 }
 264
 265 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 266    under all circumstances. It is difficult to invent anything faster or
 267    cheaper.
 268  */
 269
 270 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 271 {
 272         kfree_skb(skb);
 273         return NET_XMIT_CN;
 274 }
 275
 276 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 277 {
 278         return NULL;
 279 }
 280
 281 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 282 {
 283         if (net_ratelimit())
 284                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 285                        skb->dev->name);
 286         kfree_skb(skb);
 287         return NET_XMIT_CN;
 288 }
 289
 290 struct Qdisc_ops noop_qdisc_ops = {
 291         .id             =       "noop",
 292         .priv_size      =       0,
 293         .enqueue        =       noop_enqueue,
 294         .dequeue        =       noop_dequeue,
 295         .requeue        =       noop_requeue,
 296         .owner          =       THIS_MODULE,
 297 };
 298
 299 struct Qdisc noop_qdisc = {
 300         .enqueue        =       noop_enqueue,
 301         .dequeue        =       noop_dequeue,
 302         .flags          =       TCQ_F_BUILTIN,
 303         .ops            =       &noop_qdisc_ops,
 304         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 305 };
 306
 307 static struct Qdisc_ops noqueue_qdisc_ops = {
 308         .id             =       "noqueue",
 309         .priv_size      =       0,
 310         .enqueue        =       noop_enqueue,
 311         .dequeue        =       noop_dequeue,
 312         .requeue        =       noop_requeue,
 313         .owner          =       THIS_MODULE,
 314 };
 315
 316 static struct Qdisc noqueue_qdisc = {
 317         .enqueue        =       NULL,
 318         .dequeue        =       noop_dequeue,
 319         .flags          =       TCQ_F_BUILTIN,
 320         .ops            =       &noqueue_qdisc_ops,
 321         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 322 };
 323
 324
 325 static const u8 prio2band[TC_PRIO_MAX+1] =
 326         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 327
 328 /* 3-band FIFO queue: old style, but should be a bit faster than
 329    generic prio+fifo combination.
 330  */
 331
 332 #define PFIFO_FAST_BANDS 3
 333
 334 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 335                                              struct Qdisc *qdisc)
 336 {
 337         struct sk_buff_head *list = qdisc_priv(qdisc);
 338         return list + prio2band[skb->priority & TC_PRIO_MAX];
 339 }
 340
 341 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 342 {
 343         struct sk_buff_head *list = prio2list(skb, qdisc);
 344
 345         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 346                 qdisc->q.qlen++;
 347                 return __qdisc_enqueue_tail(skb, qdisc, list);
 348         }
 349
 350         return qdisc_drop(skb, qdisc);
 351 }
 352
 353 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 354 {
 355         int prio;
 356         struct sk_buff_head *list = qdisc_priv(qdisc);
 357
 358         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 359                 if (!skb_queue_empty(list + prio)) {
 360                         qdisc->q.qlen--;
 361                         return __qdisc_dequeue_head(qdisc, list + prio);
 362                 }
 363         }
 364
 365         return NULL;
 366 }
 367
 368 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 369 {
 370         qdisc->q.qlen++;
 371         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 372 }
 373
 374 static void pfifo_fast_reset(struct Qdisc* qdisc)
 375 {
 376         int prio;
 377         struct sk_buff_head *list = qdisc_priv(qdisc);
 378
 379         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 380                 __qdisc_reset_queue(qdisc, list + prio);
 381
 382         qdisc->qstats.backlog = 0;
 383         qdisc->q.qlen = 0;
 384 }
 385
 386 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 387 {
 388         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 389
 390         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 391         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 392         return skb->len;
 393
 394 rtattr_failure:
 395         return -1;
 396 }
 397
 398 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 399 {
 400         int prio;
 401         struct sk_buff_head *list = qdisc_priv(qdisc);
 402
 403         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 404                 skb_queue_head_init(list + prio);
 405
 406         return 0;
 407 }
 408
 409 static struct Qdisc_ops pfifo_fast_ops = {
 410         .id             =       "pfifo_fast",
 411         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 412         .enqueue        =       pfifo_fast_enqueue,
 413         .dequeue        =       pfifo_fast_dequeue,
 414         .requeue        =       pfifo_fast_requeue,
 415         .init           =       pfifo_fast_init,
 416         .reset          =       pfifo_fast_reset,
 417         .dump           =       pfifo_fast_dump,
 418         .owner          =       THIS_MODULE,
 419 };
 420
 421 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 422 {
 423         void *p;
 424         struct Qdisc *sch;
 425         unsigned int size;
 426         int err = -ENOBUFS;
 427
 428         /* ensure that the Qdisc and the private data are 32-byte aligned */
 429         size = QDISC_ALIGN(sizeof(*sch));
 430         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 431
 432         p = kzalloc(size, GFP_KERNEL);
 433         if (!p)
 434                 goto errout;
 435         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 436         sch->padded = (char *) sch - (char *) p;
 437
 438         INIT_LIST_HEAD(&sch->list);
 439         skb_queue_head_init(&sch->q);
 440         sch->ops = ops;
 441         sch->enqueue = ops->enqueue;
 442         sch->dequeue = ops->dequeue;
 443         sch->dev = dev;
 444         dev_hold(dev);
 445         atomic_set(&sch->refcnt, 1);
 446
 447         return sch;
 448 errout:
 449         return ERR_PTR(-err);
 450 }
 451
 452 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 453                                  unsigned int parentid)
 454 {
 455         struct Qdisc *sch;
 456
 457         sch = qdisc_alloc(dev, ops);
 458         if (IS_ERR(sch))
 459                 goto errout;
 460         sch->stats_lock = &dev->queue_lock;
 461         sch->parent = parentid;
 462
 463         if (!ops->init || ops->init(sch, NULL) == 0)
 464                 return sch;
 465
 466         qdisc_destroy(sch);
 467 errout:
 468         return NULL;
 469 }
 470
 471 /* Under dev->queue_lock and BH! */
 472
 473 void qdisc_reset(struct Qdisc *qdisc)
 474 {
 475         struct Qdisc_ops *ops = qdisc->ops;
 476
 477         if (ops->reset)
 478                 ops->reset(qdisc);
 479 }
 480
 481 /* this is the rcu callback function to clean up a qdisc when there
 482  * are no further references to it */
 483
 484 static void __qdisc_destroy(struct rcu_head *head)
 485 {
 486         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 487         kfree((char *) qdisc - qdisc->padded);
 488 }
 489
 490 /* Under dev->queue_lock and BH! */
 491
 492 void qdisc_destroy(struct Qdisc *qdisc)
 493 {
 494         struct Qdisc_ops  *ops = qdisc->ops;
 495
 496         if (qdisc->flags & TCQ_F_BUILTIN ||
 497             !atomic_dec_and_test(&qdisc->refcnt))
 498                 return;
 499
 500         list_del(&qdisc->list);
 501         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 502         if (ops->reset)
 503                 ops->reset(qdisc);
 504         if (ops->destroy)
 505                 ops->destroy(qdisc);
 506
 507         module_put(ops->owner);
 508         dev_put(qdisc->dev);
 509         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 510 }
 511
 512 void dev_activate(struct net_device *dev)
 513 {
 514         /* No queueing discipline is attached to device;
 515            create default one i.e. pfifo_fast for devices,
 516            which need queueing and noqueue_qdisc for
 517            virtual interfaces
 518          */
 519
 520         if (dev->qdisc_sleeping == &noop_qdisc) {
 521                 struct Qdisc *qdisc;
 522                 if (dev->tx_queue_len) {
 523                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 524                                                   TC_H_ROOT);
 525                         if (qdisc == NULL) {
 526                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 527                                 return;
 528                         }
 529                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 530                 } else {
 531                         qdisc =  &noqueue_qdisc;
 532                 }
 533                 dev->qdisc_sleeping = qdisc;
 534         }
 535
 536         if (!netif_carrier_ok(dev))
 537                 /* Delay activation until next carrier-on event */
 538                 return;
 539
 540         spin_lock_bh(&dev->queue_lock);
 541         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 542         if (dev->qdisc != &noqueue_qdisc) {
 543                 dev->trans_start = jiffies;
 544                 dev_watchdog_up(dev);
 545         }
 546         spin_unlock_bh(&dev->queue_lock);
 547 }
 548
 549 void dev_deactivate(struct net_device *dev)
 550 {
 551         struct Qdisc *qdisc;
 552         struct sk_buff *skb;
 553         int running;
 554
 555         spin_lock_bh(&dev->queue_lock);
 556         qdisc = dev->qdisc;
 557         dev->qdisc = &noop_qdisc;
 558
 559         qdisc_reset(qdisc);
 560
 561         skb = dev->gso_skb;
 562         dev->gso_skb = NULL;
 563         spin_unlock_bh(&dev->queue_lock);
 564
 565         kfree_skb(skb);
 566
 567         dev_watchdog_down(dev);
 568
 569         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 570         synchronize_rcu();
 571
 572         /* Wait for outstanding qdisc_run calls. */
 573         do {
 574                 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 575                         yield();
 576
 577                 /*
 578                  * Double-check inside queue lock to ensure that all effects
 579                  * of the queue run are visible when we return.
 580                  */
 581                 spin_lock_bh(&dev->queue_lock);
 582                 running = test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 583                 spin_unlock_bh(&dev->queue_lock);
 584
 585                 /*
 586                  * The running flag should never be set at this point because
 587                  * we've already set dev->qdisc to noop_qdisc *inside* the same
 588                  * pair of spin locks.  That is, if any qdisc_run starts after
 589                  * our initial test it should see the noop_qdisc and then
 590                  * clear the RUNNING bit before dropping the queue lock.  So
 591                  * if it is set here then we've found a bug.
 592                  */
 593         } while (WARN_ON_ONCE(running));
 594 }
 595
 596 void dev_init_scheduler(struct net_device *dev)
 597 {
 598         qdisc_lock_tree(dev);
 599         dev->qdisc = &noop_qdisc;
 600         dev->qdisc_sleeping = &noop_qdisc;
 601         INIT_LIST_HEAD(&dev->qdisc_list);
 602         qdisc_unlock_tree(dev);
 603
 604         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 605 }
 606
 607 void dev_shutdown(struct net_device *dev)
 608 {
 609         struct Qdisc *qdisc;
 610
 611         qdisc_lock_tree(dev);
 612         qdisc = dev->qdisc_sleeping;
 613         dev->qdisc = &noop_qdisc;
 614         dev->qdisc_sleeping = &noop_qdisc;
 615         qdisc_destroy(qdisc);
 616 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 617         if ((qdisc = dev->qdisc_ingress) != NULL) {
 618                 dev->qdisc_ingress = NULL;
 619                 qdisc_destroy(qdisc);
 620         }
 621 #endif
 622         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 623         qdisc_unlock_tree(dev);
 624 }
 625
 626 EXPORT_SYMBOL(netif_carrier_on);
 627 EXPORT_SYMBOL(netif_carrier_off);
 628 EXPORT_SYMBOL(noop_qdisc);
 629 EXPORT_SYMBOL(qdisc_create_dflt);
 630 EXPORT_SYMBOL(qdisc_destroy);
 631 EXPORT_SYMBOL(qdisc_reset);
 632 EXPORT_SYMBOL(qdisc_lock_tree);
 633 EXPORT_SYMBOL(qdisc_unlock_tree);