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;
 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         ret = dev_hard_start_xmit(skb, dev);
 149         HARD_TX_UNLOCK(dev);
 150
 151         spin_lock(&dev->queue_lock);
 152         q = dev->qdisc;
 153
 154         switch (ret) {
 155         case NETDEV_TX_OK:
 156                 /* Driver sent out skb successfully */
 157                 ret = qdisc_qlen(q);
 158                 break;
 159
 160         case NETDEV_TX_LOCKED:
 161                 /* Driver try lock failed */
 162                 ret = handle_dev_cpu_collision(skb, dev, q);
 163                 break;
 164
 165         default:
 166                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 167                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 168                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 169                                dev->name, ret, q->q.qlen);
 170
 171                 ret = dev_requeue_skb(skb, dev, q);
 172                 break;
 173         }
 174
 175         return ret;
 176 }
 177
 178 void __qdisc_run(struct net_device *dev)
 179 {
 180         do {
 181                 if (!qdisc_restart(dev))
 182                         break;
 183         } while (!netif_queue_stopped(dev));
 184
 185         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 186 }
 187
 188 static void dev_watchdog(unsigned long arg)
 189 {
 190         struct net_device *dev = (struct net_device *)arg;
 191
 192         netif_tx_lock(dev);
 193         if (dev->qdisc != &noop_qdisc) {
 194                 if (netif_device_present(dev) &&
 195                     netif_running(dev) &&
 196                     netif_carrier_ok(dev)) {
 197                         if (netif_queue_stopped(dev) &&
 198                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 199
 200                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 201                                        dev->name);
 202                                 dev->tx_timeout(dev);
 203                         }
 204                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 205                                 dev_hold(dev);
 206                 }
 207         }
 208         netif_tx_unlock(dev);
 209
 210         dev_put(dev);
 211 }
 212
 213 static void dev_watchdog_init(struct net_device *dev)
 214 {
 215         init_timer(&dev->watchdog_timer);
 216         dev->watchdog_timer.data = (unsigned long)dev;
 217         dev->watchdog_timer.function = dev_watchdog;
 218 }
 219
 220 void __netdev_watchdog_up(struct net_device *dev)
 221 {
 222         if (dev->tx_timeout) {
 223                 if (dev->watchdog_timeo <= 0)
 224                         dev->watchdog_timeo = 5*HZ;
 225                 if (!mod_timer(&dev->watchdog_timer,
 226                                round_jiffies(jiffies + dev->watchdog_timeo)))
 227                         dev_hold(dev);
 228         }
 229 }
 230
 231 static void dev_watchdog_up(struct net_device *dev)
 232 {
 233         __netdev_watchdog_up(dev);
 234 }
 235
 236 static void dev_watchdog_down(struct net_device *dev)
 237 {
 238         netif_tx_lock_bh(dev);
 239         if (del_timer(&dev->watchdog_timer))
 240                 dev_put(dev);
 241         netif_tx_unlock_bh(dev);
 242 }
 243
 244 /**
 245  *      netif_carrier_on - set carrier
 246  *      @dev: network device
 247  *
 248  * Device has detected that carrier.
 249  */
 250 void netif_carrier_on(struct net_device *dev)
 251 {
 252         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
 253                 linkwatch_fire_event(dev);
 254         if (netif_running(dev))
 255                 __netdev_watchdog_up(dev);
 256 }
 257
 258 /**
 259  *      netif_carrier_off - clear carrier
 260  *      @dev: network device
 261  *
 262  * Device has detected loss of carrier.
 263  */
 264 void netif_carrier_off(struct net_device *dev)
 265 {
 266         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 267                 linkwatch_fire_event(dev);
 268 }
 269
 270 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 271    under all circumstances. It is difficult to invent anything faster or
 272    cheaper.
 273  */
 274
 275 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 276 {
 277         kfree_skb(skb);
 278         return NET_XMIT_CN;
 279 }
 280
 281 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 282 {
 283         return NULL;
 284 }
 285
 286 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 287 {
 288         if (net_ratelimit())
 289                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 290                        skb->dev->name);
 291         kfree_skb(skb);
 292         return NET_XMIT_CN;
 293 }
 294
 295 struct Qdisc_ops noop_qdisc_ops = {
 296         .id             =       "noop",
 297         .priv_size      =       0,
 298         .enqueue        =       noop_enqueue,
 299         .dequeue        =       noop_dequeue,
 300         .requeue        =       noop_requeue,
 301         .owner          =       THIS_MODULE,
 302 };
 303
 304 struct Qdisc noop_qdisc = {
 305         .enqueue        =       noop_enqueue,
 306         .dequeue        =       noop_dequeue,
 307         .flags          =       TCQ_F_BUILTIN,
 308         .ops            =       &noop_qdisc_ops,
 309         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 310 };
 311
 312 static struct Qdisc_ops noqueue_qdisc_ops = {
 313         .id             =       "noqueue",
 314         .priv_size      =       0,
 315         .enqueue        =       noop_enqueue,
 316         .dequeue        =       noop_dequeue,
 317         .requeue        =       noop_requeue,
 318         .owner          =       THIS_MODULE,
 319 };
 320
 321 static struct Qdisc noqueue_qdisc = {
 322         .enqueue        =       NULL,
 323         .dequeue        =       noop_dequeue,
 324         .flags          =       TCQ_F_BUILTIN,
 325         .ops            =       &noqueue_qdisc_ops,
 326         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 327 };
 328
 329
 330 static const u8 prio2band[TC_PRIO_MAX+1] =
 331         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 332
 333 /* 3-band FIFO queue: old style, but should be a bit faster than
 334    generic prio+fifo combination.
 335  */
 336
 337 #define PFIFO_FAST_BANDS 3
 338
 339 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 340                                              struct Qdisc *qdisc)
 341 {
 342         struct sk_buff_head *list = qdisc_priv(qdisc);
 343         return list + prio2band[skb->priority & TC_PRIO_MAX];
 344 }
 345
 346 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 347 {
 348         struct sk_buff_head *list = prio2list(skb, qdisc);
 349
 350         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 351                 qdisc->q.qlen++;
 352                 return __qdisc_enqueue_tail(skb, qdisc, list);
 353         }
 354
 355         return qdisc_drop(skb, qdisc);
 356 }
 357
 358 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 359 {
 360         int prio;
 361         struct sk_buff_head *list = qdisc_priv(qdisc);
 362
 363         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 364                 if (!skb_queue_empty(list + prio)) {
 365                         qdisc->q.qlen--;
 366                         return __qdisc_dequeue_head(qdisc, list + prio);
 367                 }
 368         }
 369
 370         return NULL;
 371 }
 372
 373 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 374 {
 375         qdisc->q.qlen++;
 376         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 377 }
 378
 379 static void pfifo_fast_reset(struct Qdisc* qdisc)
 380 {
 381         int prio;
 382         struct sk_buff_head *list = qdisc_priv(qdisc);
 383
 384         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 385                 __qdisc_reset_queue(qdisc, list + prio);
 386
 387         qdisc->qstats.backlog = 0;
 388         qdisc->q.qlen = 0;
 389 }
 390
 391 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 392 {
 393         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 394
 395         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 396         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 397         return skb->len;
 398
 399 rtattr_failure:
 400         return -1;
 401 }
 402
 403 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 404 {
 405         int prio;
 406         struct sk_buff_head *list = qdisc_priv(qdisc);
 407
 408         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 409                 skb_queue_head_init(list + prio);
 410
 411         return 0;
 412 }
 413
 414 static struct Qdisc_ops pfifo_fast_ops = {
 415         .id             =       "pfifo_fast",
 416         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 417         .enqueue        =       pfifo_fast_enqueue,
 418         .dequeue        =       pfifo_fast_dequeue,
 419         .requeue        =       pfifo_fast_requeue,
 420         .init           =       pfifo_fast_init,
 421         .reset          =       pfifo_fast_reset,
 422         .dump           =       pfifo_fast_dump,
 423         .owner          =       THIS_MODULE,
 424 };
 425
 426 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 427 {
 428         void *p;
 429         struct Qdisc *sch;
 430         unsigned int size;
 431         int err = -ENOBUFS;
 432
 433         /* ensure that the Qdisc and the private data are 32-byte aligned */
 434         size = QDISC_ALIGN(sizeof(*sch));
 435         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 436
 437         p = kzalloc(size, GFP_KERNEL);
 438         if (!p)
 439                 goto errout;
 440         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 441         sch->padded = (char *) sch - (char *) p;
 442
 443         INIT_LIST_HEAD(&sch->list);
 444         skb_queue_head_init(&sch->q);
 445         sch->ops = ops;
 446         sch->enqueue = ops->enqueue;
 447         sch->dequeue = ops->dequeue;
 448         sch->dev = dev;
 449         dev_hold(dev);
 450         atomic_set(&sch->refcnt, 1);
 451
 452         return sch;
 453 errout:
 454         return ERR_PTR(-err);
 455 }
 456
 457 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 458                                  unsigned int parentid)
 459 {
 460         struct Qdisc *sch;
 461
 462         sch = qdisc_alloc(dev, ops);
 463         if (IS_ERR(sch))
 464                 goto errout;
 465         sch->stats_lock = &dev->queue_lock;
 466         sch->parent = parentid;
 467
 468         if (!ops->init || ops->init(sch, NULL) == 0)
 469                 return sch;
 470
 471         qdisc_destroy(sch);
 472 errout:
 473         return NULL;
 474 }
 475
 476 /* Under dev->queue_lock and BH! */
 477
 478 void qdisc_reset(struct Qdisc *qdisc)
 479 {
 480         struct Qdisc_ops *ops = qdisc->ops;
 481
 482         if (ops->reset)
 483                 ops->reset(qdisc);
 484 }
 485
 486 /* this is the rcu callback function to clean up a qdisc when there
 487  * are no further references to it */
 488
 489 static void __qdisc_destroy(struct rcu_head *head)
 490 {
 491         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 492         kfree((char *) qdisc - qdisc->padded);
 493 }
 494
 495 /* Under dev->queue_lock and BH! */
 496
 497 void qdisc_destroy(struct Qdisc *qdisc)
 498 {
 499         struct Qdisc_ops  *ops = qdisc->ops;
 500
 501         if (qdisc->flags & TCQ_F_BUILTIN ||
 502             !atomic_dec_and_test(&qdisc->refcnt))
 503                 return;
 504
 505         list_del(&qdisc->list);
 506         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 507         if (ops->reset)
 508                 ops->reset(qdisc);
 509         if (ops->destroy)
 510                 ops->destroy(qdisc);
 511
 512         module_put(ops->owner);
 513         dev_put(qdisc->dev);
 514         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 515 }
 516
 517 void dev_activate(struct net_device *dev)
 518 {
 519         /* No queueing discipline is attached to device;
 520            create default one i.e. pfifo_fast for devices,
 521            which need queueing and noqueue_qdisc for
 522            virtual interfaces
 523          */
 524
 525         if (dev->qdisc_sleeping == &noop_qdisc) {
 526                 struct Qdisc *qdisc;
 527                 if (dev->tx_queue_len) {
 528                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 529                                                   TC_H_ROOT);
 530                         if (qdisc == NULL) {
 531                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 532                                 return;
 533                         }
 534                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 535                 } else {
 536                         qdisc =  &noqueue_qdisc;
 537                 }
 538                 dev->qdisc_sleeping = qdisc;
 539         }
 540
 541         if (!netif_carrier_ok(dev))
 542                 /* Delay activation until next carrier-on event */
 543                 return;
 544
 545         spin_lock_bh(&dev->queue_lock);
 546         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 547         if (dev->qdisc != &noqueue_qdisc) {
 548                 dev->trans_start = jiffies;
 549                 dev_watchdog_up(dev);
 550         }
 551         spin_unlock_bh(&dev->queue_lock);
 552 }
 553
 554 void dev_deactivate(struct net_device *dev)
 555 {
 556         struct Qdisc *qdisc;
 557         struct sk_buff *skb;
 558
 559         spin_lock_bh(&dev->queue_lock);
 560         qdisc = dev->qdisc;
 561         dev->qdisc = &noop_qdisc;
 562
 563         qdisc_reset(qdisc);
 564
 565         skb = dev->gso_skb;
 566         dev->gso_skb = NULL;
 567         spin_unlock_bh(&dev->queue_lock);
 568
 569         kfree_skb(skb);
 570
 571         dev_watchdog_down(dev);
 572
 573         /* Wait for outstanding dev_queue_xmit calls. */
 574         synchronize_rcu();
 575
 576         /* Wait for outstanding qdisc_run calls. */
 577         while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 578                 yield();
 579 }
 580
 581 void dev_init_scheduler(struct net_device *dev)
 582 {
 583         qdisc_lock_tree(dev);
 584         dev->qdisc = &noop_qdisc;
 585         dev->qdisc_sleeping = &noop_qdisc;
 586         INIT_LIST_HEAD(&dev->qdisc_list);
 587         qdisc_unlock_tree(dev);
 588
 589         dev_watchdog_init(dev);
 590 }
 591
 592 void dev_shutdown(struct net_device *dev)
 593 {
 594         struct Qdisc *qdisc;
 595
 596         qdisc_lock_tree(dev);
 597         qdisc = dev->qdisc_sleeping;
 598         dev->qdisc = &noop_qdisc;
 599         dev->qdisc_sleeping = &noop_qdisc;
 600         qdisc_destroy(qdisc);
 601 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 602         if ((qdisc = dev->qdisc_ingress) != NULL) {
 603                 dev->qdisc_ingress = NULL;
 604                 qdisc_destroy(qdisc);
 605         }
 606 #endif
 607         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 608         qdisc_unlock_tree(dev);
 609 }
 610
 611 EXPORT_SYMBOL(netif_carrier_on);
 612 EXPORT_SYMBOL(netif_carrier_off);
 613 EXPORT_SYMBOL(noop_qdisc);
 614 EXPORT_SYMBOL(qdisc_create_dflt);
 615 EXPORT_SYMBOL(qdisc_destroy);
 616 EXPORT_SYMBOL(qdisc_reset);
 617 EXPORT_SYMBOL(qdisc_lock_tree);
 618 EXPORT_SYMBOL(qdisc_unlock_tree);