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 void netif_carrier_on(struct net_device *dev)
 260 {
 261         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
 262                 linkwatch_fire_event(dev);
 263         if (netif_running(dev))
 264                 __netdev_watchdog_up(dev);
 265 }
 266
 267 void netif_carrier_off(struct net_device *dev)
 268 {
 269         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 270                 linkwatch_fire_event(dev);
 271 }
 272
 273 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 274    under all circumstances. It is difficult to invent anything faster or
 275    cheaper.
 276  */
 277
 278 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 279 {
 280         kfree_skb(skb);
 281         return NET_XMIT_CN;
 282 }
 283
 284 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 285 {
 286         return NULL;
 287 }
 288
 289 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 290 {
 291         if (net_ratelimit())
 292                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 293                        skb->dev->name);
 294         kfree_skb(skb);
 295         return NET_XMIT_CN;
 296 }
 297
 298 struct Qdisc_ops noop_qdisc_ops = {
 299         .id             =       "noop",
 300         .priv_size      =       0,
 301         .enqueue        =       noop_enqueue,
 302         .dequeue        =       noop_dequeue,
 303         .requeue        =       noop_requeue,
 304         .owner          =       THIS_MODULE,
 305 };
 306
 307 struct Qdisc noop_qdisc = {
 308         .enqueue        =       noop_enqueue,
 309         .dequeue        =       noop_dequeue,
 310         .flags          =       TCQ_F_BUILTIN,
 311         .ops            =       &noop_qdisc_ops,
 312         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 313 };
 314
 315 static struct Qdisc_ops noqueue_qdisc_ops = {
 316         .id             =       "noqueue",
 317         .priv_size      =       0,
 318         .enqueue        =       noop_enqueue,
 319         .dequeue        =       noop_dequeue,
 320         .requeue        =       noop_requeue,
 321         .owner          =       THIS_MODULE,
 322 };
 323
 324 static struct Qdisc noqueue_qdisc = {
 325         .enqueue        =       NULL,
 326         .dequeue        =       noop_dequeue,
 327         .flags          =       TCQ_F_BUILTIN,
 328         .ops            =       &noqueue_qdisc_ops,
 329         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 330 };
 331
 332
 333 static const u8 prio2band[TC_PRIO_MAX+1] =
 334         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 335
 336 /* 3-band FIFO queue: old style, but should be a bit faster than
 337    generic prio+fifo combination.
 338  */
 339
 340 #define PFIFO_FAST_BANDS 3
 341
 342 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 343                                              struct Qdisc *qdisc)
 344 {
 345         struct sk_buff_head *list = qdisc_priv(qdisc);
 346         return list + prio2band[skb->priority & TC_PRIO_MAX];
 347 }
 348
 349 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 350 {
 351         struct sk_buff_head *list = prio2list(skb, qdisc);
 352
 353         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 354                 qdisc->q.qlen++;
 355                 return __qdisc_enqueue_tail(skb, qdisc, list);
 356         }
 357
 358         return qdisc_drop(skb, qdisc);
 359 }
 360
 361 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 362 {
 363         int prio;
 364         struct sk_buff_head *list = qdisc_priv(qdisc);
 365
 366         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 367                 if (!skb_queue_empty(list + prio)) {
 368                         qdisc->q.qlen--;
 369                         return __qdisc_dequeue_head(qdisc, list + prio);
 370                 }
 371         }
 372
 373         return NULL;
 374 }
 375
 376 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 377 {
 378         qdisc->q.qlen++;
 379         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 380 }
 381
 382 static void pfifo_fast_reset(struct Qdisc* qdisc)
 383 {
 384         int prio;
 385         struct sk_buff_head *list = qdisc_priv(qdisc);
 386
 387         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 388                 __qdisc_reset_queue(qdisc, list + prio);
 389
 390         qdisc->qstats.backlog = 0;
 391         qdisc->q.qlen = 0;
 392 }
 393
 394 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 395 {
 396         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 397
 398         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 399         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 400         return skb->len;
 401
 402 rtattr_failure:
 403         return -1;
 404 }
 405
 406 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 407 {
 408         int prio;
 409         struct sk_buff_head *list = qdisc_priv(qdisc);
 410
 411         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 412                 skb_queue_head_init(list + prio);
 413
 414         return 0;
 415 }
 416
 417 static struct Qdisc_ops pfifo_fast_ops = {
 418         .id             =       "pfifo_fast",
 419         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 420         .enqueue        =       pfifo_fast_enqueue,
 421         .dequeue        =       pfifo_fast_dequeue,
 422         .requeue        =       pfifo_fast_requeue,
 423         .init           =       pfifo_fast_init,
 424         .reset          =       pfifo_fast_reset,
 425         .dump           =       pfifo_fast_dump,
 426         .owner          =       THIS_MODULE,
 427 };
 428
 429 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 430 {
 431         void *p;
 432         struct Qdisc *sch;
 433         unsigned int size;
 434         int err = -ENOBUFS;
 435
 436         /* ensure that the Qdisc and the private data are 32-byte aligned */
 437         size = QDISC_ALIGN(sizeof(*sch));
 438         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 439
 440         p = kzalloc(size, GFP_KERNEL);
 441         if (!p)
 442                 goto errout;
 443         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 444         sch->padded = (char *) sch - (char *) p;
 445
 446         INIT_LIST_HEAD(&sch->list);
 447         skb_queue_head_init(&sch->q);
 448         sch->ops = ops;
 449         sch->enqueue = ops->enqueue;
 450         sch->dequeue = ops->dequeue;
 451         sch->dev = dev;
 452         dev_hold(dev);
 453         atomic_set(&sch->refcnt, 1);
 454
 455         return sch;
 456 errout:
 457         return ERR_PTR(-err);
 458 }
 459
 460 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 461                                  unsigned int parentid)
 462 {
 463         struct Qdisc *sch;
 464
 465         sch = qdisc_alloc(dev, ops);
 466         if (IS_ERR(sch))
 467                 goto errout;
 468         sch->stats_lock = &dev->queue_lock;
 469         sch->parent = parentid;
 470
 471         if (!ops->init || ops->init(sch, NULL) == 0)
 472                 return sch;
 473
 474         qdisc_destroy(sch);
 475 errout:
 476         return NULL;
 477 }
 478
 479 /* Under dev->queue_lock and BH! */
 480
 481 void qdisc_reset(struct Qdisc *qdisc)
 482 {
 483         struct Qdisc_ops *ops = qdisc->ops;
 484
 485         if (ops->reset)
 486                 ops->reset(qdisc);
 487 }
 488
 489 /* this is the rcu callback function to clean up a qdisc when there
 490  * are no further references to it */
 491
 492 static void __qdisc_destroy(struct rcu_head *head)
 493 {
 494         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 495         kfree((char *) qdisc - qdisc->padded);
 496 }
 497
 498 /* Under dev->queue_lock and BH! */
 499
 500 void qdisc_destroy(struct Qdisc *qdisc)
 501 {
 502         struct Qdisc_ops  *ops = qdisc->ops;
 503
 504         if (qdisc->flags & TCQ_F_BUILTIN ||
 505             !atomic_dec_and_test(&qdisc->refcnt))
 506                 return;
 507
 508         list_del(&qdisc->list);
 509         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 510         if (ops->reset)
 511                 ops->reset(qdisc);
 512         if (ops->destroy)
 513                 ops->destroy(qdisc);
 514
 515         module_put(ops->owner);
 516         dev_put(qdisc->dev);
 517         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 518 }
 519
 520 void dev_activate(struct net_device *dev)
 521 {
 522         /* No queueing discipline is attached to device;
 523            create default one i.e. pfifo_fast for devices,
 524            which need queueing and noqueue_qdisc for
 525            virtual interfaces
 526          */
 527
 528         if (dev->qdisc_sleeping == &noop_qdisc) {
 529                 struct Qdisc *qdisc;
 530                 if (dev->tx_queue_len) {
 531                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 532                                                   TC_H_ROOT);
 533                         if (qdisc == NULL) {
 534                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 535                                 return;
 536                         }
 537                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 538                 } else {
 539                         qdisc =  &noqueue_qdisc;
 540                 }
 541                 dev->qdisc_sleeping = qdisc;
 542         }
 543
 544         if (!netif_carrier_ok(dev))
 545                 /* Delay activation until next carrier-on event */
 546                 return;
 547
 548         spin_lock_bh(&dev->queue_lock);
 549         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 550         if (dev->qdisc != &noqueue_qdisc) {
 551                 dev->trans_start = jiffies;
 552                 dev_watchdog_up(dev);
 553         }
 554         spin_unlock_bh(&dev->queue_lock);
 555 }
 556
 557 void dev_deactivate(struct net_device *dev)
 558 {
 559         struct Qdisc *qdisc;
 560         struct sk_buff *skb;
 561
 562         spin_lock_bh(&dev->queue_lock);
 563         qdisc = dev->qdisc;
 564         dev->qdisc = &noop_qdisc;
 565
 566         qdisc_reset(qdisc);
 567
 568         skb = dev->gso_skb;
 569         dev->gso_skb = NULL;
 570         spin_unlock_bh(&dev->queue_lock);
 571
 572         kfree_skb(skb);
 573
 574         dev_watchdog_down(dev);
 575
 576         /* Wait for outstanding dev_queue_xmit calls. */
 577         synchronize_rcu();
 578
 579         /* Wait for outstanding qdisc_run calls. */
 580         while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 581                 yield();
 582 }
 583
 584 void dev_init_scheduler(struct net_device *dev)
 585 {
 586         qdisc_lock_tree(dev);
 587         dev->qdisc = &noop_qdisc;
 588         dev->qdisc_sleeping = &noop_qdisc;
 589         INIT_LIST_HEAD(&dev->qdisc_list);
 590         qdisc_unlock_tree(dev);
 591
 592         dev_watchdog_init(dev);
 593 }
 594
 595 void dev_shutdown(struct net_device *dev)
 596 {
 597         struct Qdisc *qdisc;
 598
 599         qdisc_lock_tree(dev);
 600         qdisc = dev->qdisc_sleeping;
 601         dev->qdisc = &noop_qdisc;
 602         dev->qdisc_sleeping = &noop_qdisc;
 603         qdisc_destroy(qdisc);
 604 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 605         if ((qdisc = dev->qdisc_ingress) != NULL) {
 606                 dev->qdisc_ingress = NULL;
 607                 qdisc_destroy(qdisc);
 608         }
 609 #endif
 610         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 611         qdisc_unlock_tree(dev);
 612 }
 613
 614 EXPORT_SYMBOL(netif_carrier_on);
 615 EXPORT_SYMBOL(netif_carrier_off);
 616 EXPORT_SYMBOL(noop_qdisc);
 617 EXPORT_SYMBOL(qdisc_create_dflt);
 618 EXPORT_SYMBOL(qdisc_destroy);
 619 EXPORT_SYMBOL(qdisc_reset);
 620 EXPORT_SYMBOL(qdisc_lock_tree);
 621 EXPORT_SYMBOL(qdisc_unlock_tree);