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         unsigned lockless;
 138         int ret;
 139
 140         /* Dequeue packet */
 141         if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
 142                 return 0;
 143
 144         /*
 145          * When the driver has LLTX set, it does its own locking in
 146          * start_xmit. These checks are worth it because even uncongested
 147          * locks can be quite expensive. The driver can do a trylock, as
 148          * is being done here; in case of lock contention it should return
 149          * NETDEV_TX_LOCKED and the packet will be requeued.
 150          */
 151         lockless = (dev->features & NETIF_F_LLTX);
 152
 153         if (!lockless && !netif_tx_trylock(dev)) {
 154                 /* Another CPU grabbed the driver tx lock */
 155                 return handle_dev_cpu_collision(skb, dev, q);
 156         }
 157
 158         /* And release queue */
 159         spin_unlock(&dev->queue_lock);
 160
 161         ret = dev_hard_start_xmit(skb, dev);
 162
 163         if (!lockless)
 164                 netif_tx_unlock(dev);
 165
 166         spin_lock(&dev->queue_lock);
 167         q = dev->qdisc;
 168
 169         switch (ret) {
 170         case NETDEV_TX_OK:
 171                 /* Driver sent out skb successfully */
 172                 ret = qdisc_qlen(q);
 173                 break;
 174
 175         case NETDEV_TX_LOCKED:
 176                 /* Driver try lock failed */
 177                 ret = handle_dev_cpu_collision(skb, dev, q);
 178                 break;
 179
 180         default:
 181                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 182                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 183                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 184                                dev->name, ret, q->q.qlen);
 185
 186                 ret = dev_requeue_skb(skb, dev, q);
 187                 break;
 188         }
 189
 190         return ret;
 191 }
 192
 193 void __qdisc_run(struct net_device *dev)
 194 {
 195         do {
 196                 if (!qdisc_restart(dev))
 197                         break;
 198         } while (!netif_queue_stopped(dev));
 199
 200         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 201 }
 202
 203 static void dev_watchdog(unsigned long arg)
 204 {
 205         struct net_device *dev = (struct net_device *)arg;
 206
 207         netif_tx_lock(dev);
 208         if (dev->qdisc != &noop_qdisc) {
 209                 if (netif_device_present(dev) &&
 210                     netif_running(dev) &&
 211                     netif_carrier_ok(dev)) {
 212                         if (netif_queue_stopped(dev) &&
 213                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 214
 215                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 216                                        dev->name);
 217                                 dev->tx_timeout(dev);
 218                         }
 219                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 220                                 dev_hold(dev);
 221                 }
 222         }
 223         netif_tx_unlock(dev);
 224
 225         dev_put(dev);
 226 }
 227
 228 static void dev_watchdog_init(struct net_device *dev)
 229 {
 230         init_timer(&dev->watchdog_timer);
 231         dev->watchdog_timer.data = (unsigned long)dev;
 232         dev->watchdog_timer.function = dev_watchdog;
 233 }
 234
 235 void __netdev_watchdog_up(struct net_device *dev)
 236 {
 237         if (dev->tx_timeout) {
 238                 if (dev->watchdog_timeo <= 0)
 239                         dev->watchdog_timeo = 5*HZ;
 240                 if (!mod_timer(&dev->watchdog_timer,
 241                                round_jiffies(jiffies + dev->watchdog_timeo)))
 242                         dev_hold(dev);
 243         }
 244 }
 245
 246 static void dev_watchdog_up(struct net_device *dev)
 247 {
 248         __netdev_watchdog_up(dev);
 249 }
 250
 251 static void dev_watchdog_down(struct net_device *dev)
 252 {
 253         netif_tx_lock_bh(dev);
 254         if (del_timer(&dev->watchdog_timer))
 255                 dev_put(dev);
 256         netif_tx_unlock_bh(dev);
 257 }
 258
 259 /**
 260  *      netif_carrier_on - set carrier
 261  *      @dev: network device
 262  *
 263  * Device has detected that carrier.
 264  */
 265 void netif_carrier_on(struct net_device *dev)
 266 {
 267         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
 268                 linkwatch_fire_event(dev);
 269         if (netif_running(dev))
 270                 __netdev_watchdog_up(dev);
 271 }
 272
 273 /**
 274  *      netif_carrier_off - clear carrier
 275  *      @dev: network device
 276  *
 277  * Device has detected loss of carrier.
 278  */
 279 void netif_carrier_off(struct net_device *dev)
 280 {
 281         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 282                 linkwatch_fire_event(dev);
 283 }
 284
 285 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 286    under all circumstances. It is difficult to invent anything faster or
 287    cheaper.
 288  */
 289
 290 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 291 {
 292         kfree_skb(skb);
 293         return NET_XMIT_CN;
 294 }
 295
 296 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 297 {
 298         return NULL;
 299 }
 300
 301 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 302 {
 303         if (net_ratelimit())
 304                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 305                        skb->dev->name);
 306         kfree_skb(skb);
 307         return NET_XMIT_CN;
 308 }
 309
 310 struct Qdisc_ops noop_qdisc_ops = {
 311         .id             =       "noop",
 312         .priv_size      =       0,
 313         .enqueue        =       noop_enqueue,
 314         .dequeue        =       noop_dequeue,
 315         .requeue        =       noop_requeue,
 316         .owner          =       THIS_MODULE,
 317 };
 318
 319 struct Qdisc noop_qdisc = {
 320         .enqueue        =       noop_enqueue,
 321         .dequeue        =       noop_dequeue,
 322         .flags          =       TCQ_F_BUILTIN,
 323         .ops            =       &noop_qdisc_ops,
 324         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 325 };
 326
 327 static struct Qdisc_ops noqueue_qdisc_ops = {
 328         .id             =       "noqueue",
 329         .priv_size      =       0,
 330         .enqueue        =       noop_enqueue,
 331         .dequeue        =       noop_dequeue,
 332         .requeue        =       noop_requeue,
 333         .owner          =       THIS_MODULE,
 334 };
 335
 336 static struct Qdisc noqueue_qdisc = {
 337         .enqueue        =       NULL,
 338         .dequeue        =       noop_dequeue,
 339         .flags          =       TCQ_F_BUILTIN,
 340         .ops            =       &noqueue_qdisc_ops,
 341         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 342 };
 343
 344
 345 static const u8 prio2band[TC_PRIO_MAX+1] =
 346         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 347
 348 /* 3-band FIFO queue: old style, but should be a bit faster than
 349    generic prio+fifo combination.
 350  */
 351
 352 #define PFIFO_FAST_BANDS 3
 353
 354 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 355                                              struct Qdisc *qdisc)
 356 {
 357         struct sk_buff_head *list = qdisc_priv(qdisc);
 358         return list + prio2band[skb->priority & TC_PRIO_MAX];
 359 }
 360
 361 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 362 {
 363         struct sk_buff_head *list = prio2list(skb, qdisc);
 364
 365         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 366                 qdisc->q.qlen++;
 367                 return __qdisc_enqueue_tail(skb, qdisc, list);
 368         }
 369
 370         return qdisc_drop(skb, qdisc);
 371 }
 372
 373 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 374 {
 375         int prio;
 376         struct sk_buff_head *list = qdisc_priv(qdisc);
 377
 378         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 379                 if (!skb_queue_empty(list + prio)) {
 380                         qdisc->q.qlen--;
 381                         return __qdisc_dequeue_head(qdisc, list + prio);
 382                 }
 383         }
 384
 385         return NULL;
 386 }
 387
 388 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 389 {
 390         qdisc->q.qlen++;
 391         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 392 }
 393
 394 static void pfifo_fast_reset(struct Qdisc* qdisc)
 395 {
 396         int prio;
 397         struct sk_buff_head *list = qdisc_priv(qdisc);
 398
 399         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 400                 __qdisc_reset_queue(qdisc, list + prio);
 401
 402         qdisc->qstats.backlog = 0;
 403         qdisc->q.qlen = 0;
 404 }
 405
 406 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 407 {
 408         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 409
 410         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 411         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 412         return skb->len;
 413
 414 rtattr_failure:
 415         return -1;
 416 }
 417
 418 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 419 {
 420         int prio;
 421         struct sk_buff_head *list = qdisc_priv(qdisc);
 422
 423         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 424                 skb_queue_head_init(list + prio);
 425
 426         return 0;
 427 }
 428
 429 static struct Qdisc_ops pfifo_fast_ops = {
 430         .id             =       "pfifo_fast",
 431         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 432         .enqueue        =       pfifo_fast_enqueue,
 433         .dequeue        =       pfifo_fast_dequeue,
 434         .requeue        =       pfifo_fast_requeue,
 435         .init           =       pfifo_fast_init,
 436         .reset          =       pfifo_fast_reset,
 437         .dump           =       pfifo_fast_dump,
 438         .owner          =       THIS_MODULE,
 439 };
 440
 441 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 442 {
 443         void *p;
 444         struct Qdisc *sch;
 445         unsigned int size;
 446         int err = -ENOBUFS;
 447
 448         /* ensure that the Qdisc and the private data are 32-byte aligned */
 449         size = QDISC_ALIGN(sizeof(*sch));
 450         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 451
 452         p = kzalloc(size, GFP_KERNEL);
 453         if (!p)
 454                 goto errout;
 455         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 456         sch->padded = (char *) sch - (char *) p;
 457
 458         INIT_LIST_HEAD(&sch->list);
 459         skb_queue_head_init(&sch->q);
 460         sch->ops = ops;
 461         sch->enqueue = ops->enqueue;
 462         sch->dequeue = ops->dequeue;
 463         sch->dev = dev;
 464         dev_hold(dev);
 465         atomic_set(&sch->refcnt, 1);
 466
 467         return sch;
 468 errout:
 469         return ERR_PTR(-err);
 470 }
 471
 472 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 473                                  unsigned int parentid)
 474 {
 475         struct Qdisc *sch;
 476
 477         sch = qdisc_alloc(dev, ops);
 478         if (IS_ERR(sch))
 479                 goto errout;
 480         sch->stats_lock = &dev->queue_lock;
 481         sch->parent = parentid;
 482
 483         if (!ops->init || ops->init(sch, NULL) == 0)
 484                 return sch;
 485
 486         qdisc_destroy(sch);
 487 errout:
 488         return NULL;
 489 }
 490
 491 /* Under dev->queue_lock and BH! */
 492
 493 void qdisc_reset(struct Qdisc *qdisc)
 494 {
 495         struct Qdisc_ops *ops = qdisc->ops;
 496
 497         if (ops->reset)
 498                 ops->reset(qdisc);
 499 }
 500
 501 /* this is the rcu callback function to clean up a qdisc when there
 502  * are no further references to it */
 503
 504 static void __qdisc_destroy(struct rcu_head *head)
 505 {
 506         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 507         kfree((char *) qdisc - qdisc->padded);
 508 }
 509
 510 /* Under dev->queue_lock and BH! */
 511
 512 void qdisc_destroy(struct Qdisc *qdisc)
 513 {
 514         struct Qdisc_ops  *ops = qdisc->ops;
 515
 516         if (qdisc->flags & TCQ_F_BUILTIN ||
 517             !atomic_dec_and_test(&qdisc->refcnt))
 518                 return;
 519
 520         list_del(&qdisc->list);
 521         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 522         if (ops->reset)
 523                 ops->reset(qdisc);
 524         if (ops->destroy)
 525                 ops->destroy(qdisc);
 526
 527         module_put(ops->owner);
 528         dev_put(qdisc->dev);
 529         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 530 }
 531
 532 void dev_activate(struct net_device *dev)
 533 {
 534         /* No queueing discipline is attached to device;
 535            create default one i.e. pfifo_fast for devices,
 536            which need queueing and noqueue_qdisc for
 537            virtual interfaces
 538          */
 539
 540         if (dev->qdisc_sleeping == &noop_qdisc) {
 541                 struct Qdisc *qdisc;
 542                 if (dev->tx_queue_len) {
 543                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 544                                                   TC_H_ROOT);
 545                         if (qdisc == NULL) {
 546                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 547                                 return;
 548                         }
 549                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 550                 } else {
 551                         qdisc =  &noqueue_qdisc;
 552                 }
 553                 dev->qdisc_sleeping = qdisc;
 554         }
 555
 556         if (!netif_carrier_ok(dev))
 557                 /* Delay activation until next carrier-on event */
 558                 return;
 559
 560         spin_lock_bh(&dev->queue_lock);
 561         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 562         if (dev->qdisc != &noqueue_qdisc) {
 563                 dev->trans_start = jiffies;
 564                 dev_watchdog_up(dev);
 565         }
 566         spin_unlock_bh(&dev->queue_lock);
 567 }
 568
 569 void dev_deactivate(struct net_device *dev)
 570 {
 571         struct Qdisc *qdisc;
 572         struct sk_buff *skb;
 573
 574         spin_lock_bh(&dev->queue_lock);
 575         qdisc = dev->qdisc;
 576         dev->qdisc = &noop_qdisc;
 577
 578         qdisc_reset(qdisc);
 579
 580         skb = dev->gso_skb;
 581         dev->gso_skb = NULL;
 582         spin_unlock_bh(&dev->queue_lock);
 583
 584         kfree_skb(skb);
 585
 586         dev_watchdog_down(dev);
 587
 588         /* Wait for outstanding dev_queue_xmit calls. */
 589         synchronize_rcu();
 590
 591         /* Wait for outstanding qdisc_run calls. */
 592         while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 593                 yield();
 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         dev_watchdog_init(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);