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         __acquires(dev->queue_lock)
  44         __acquires(dev->ingress_lock)
  45 {
  46         spin_lock_bh(&dev->queue_lock);
  47         spin_lock(&dev->ingress_lock);
  48 }
  49 EXPORT_SYMBOL(qdisc_lock_tree);
  50
  51 void qdisc_unlock_tree(struct net_device *dev)
  52         __releases(dev->ingress_lock)
  53         __releases(dev->queue_lock)
  54 {
  55         spin_unlock(&dev->ingress_lock);
  56         spin_unlock_bh(&dev->queue_lock);
  57 }
  58 EXPORT_SYMBOL(qdisc_unlock_tree);
  59
  60 static inline int qdisc_qlen(struct Qdisc *q)
  61 {
  62         return q->q.qlen;
  63 }
  64
  65 static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
  66                                   struct Qdisc *q)
  67 {
  68         if (unlikely(skb->next))
  69                 dev->gso_skb = skb;
  70         else
  71                 q->ops->requeue(skb, q);
  72
  73         netif_schedule(dev);
  74         return 0;
  75 }
  76
  77 static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
  78                                               struct Qdisc *q)
  79 {
  80         struct sk_buff *skb;
  81
  82         if ((skb = dev->gso_skb))
  83                 dev->gso_skb = NULL;
  84         else
  85                 skb = q->dequeue(q);
  86
  87         return skb;
  88 }
  89
  90 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  91                                            struct net_device *dev,
  92                                            struct Qdisc *q)
  93 {
  94         int ret;
  95
  96         if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
  97                 /*
  98                  * Same CPU holding the lock. It may be a transient
  99                  * configuration error, when hard_start_xmit() recurses. We
 100                  * detect it by checking xmit owner and drop the packet when
 101                  * deadloop is detected. Return OK to try the next skb.
 102                  */
 103                 kfree_skb(skb);
 104                 if (net_ratelimit())
 105                         printk(KERN_WARNING "Dead loop on netdevice %s, "
 106                                "fix it urgently!\n", dev->name);
 107                 ret = qdisc_qlen(q);
 108         } else {
 109                 /*
 110                  * Another cpu is holding lock, requeue & delay xmits for
 111                  * some time.
 112                  */
 113                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
 114                 ret = dev_requeue_skb(skb, dev, q);
 115         }
 116
 117         return ret;
 118 }
 119
 120 /*
 121  * NOTE: Called under dev->queue_lock with locally disabled BH.
 122  *
 123  * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
 124  * device at a time. dev->queue_lock serializes queue accesses for
 125  * this device AND dev->qdisc pointer itself.
 126  *
 127  *  netif_tx_lock serializes accesses to device driver.
 128  *
 129  *  dev->queue_lock and netif_tx_lock are mutually exclusive,
 130  *  if one is grabbed, another must be free.
 131  *
 132  * Note, that this procedure can be called by a watchdog timer
 133  *
 134  * Returns to the caller:
 135  *                              0  - queue is empty or throttled.
 136  *                              >0 - queue is not empty.
 137  *
 138  */
 139 static inline int qdisc_restart(struct net_device *dev)
 140 {
 141         struct Qdisc *q = dev->qdisc;
 142         struct sk_buff *skb;
 143         int ret = NETDEV_TX_BUSY;
 144
 145         /* Dequeue packet */
 146         if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
 147                 return 0;
 148
 149
 150         /* And release queue */
 151         spin_unlock(&dev->queue_lock);
 152
 153         HARD_TX_LOCK(dev, smp_processor_id());
 154         if (!netif_subqueue_stopped(dev, skb))
 155                 ret = dev_hard_start_xmit(skb, dev);
 156         HARD_TX_UNLOCK(dev);
 157
 158         spin_lock(&dev->queue_lock);
 159         q = dev->qdisc;
 160
 161         switch (ret) {
 162         case NETDEV_TX_OK:
 163                 /* Driver sent out skb successfully */
 164                 ret = qdisc_qlen(q);
 165                 break;
 166
 167         case NETDEV_TX_LOCKED:
 168                 /* Driver try lock failed */
 169                 ret = handle_dev_cpu_collision(skb, dev, q);
 170                 break;
 171
 172         default:
 173                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 174                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 175                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 176                                dev->name, ret, q->q.qlen);
 177
 178                 ret = dev_requeue_skb(skb, dev, q);
 179                 break;
 180         }
 181
 182         return ret;
 183 }
 184
 185 void __qdisc_run(struct net_device *dev)
 186 {
 187         unsigned long start_time = jiffies;
 188
 189         while (qdisc_restart(dev)) {
 190                 if (netif_queue_stopped(dev))
 191                         break;
 192
 193                 /*
 194                  * Postpone processing if
 195                  * 1. another process needs the CPU;
 196                  * 2. we've been doing it for too long.
 197                  */
 198                 if (need_resched() || jiffies != start_time) {
 199                         netif_schedule(dev);
 200                         break;
 201                 }
 202         }
 203
 204         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 205 }
 206
 207 static void dev_watchdog(unsigned long arg)
 208 {
 209         struct net_device *dev = (struct net_device *)arg;
 210
 211         netif_tx_lock(dev);
 212         if (dev->qdisc != &noop_qdisc) {
 213                 if (netif_device_present(dev) &&
 214                     netif_running(dev) &&
 215                     netif_carrier_ok(dev)) {
 216                         if (netif_queue_stopped(dev) &&
 217                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 218
 219                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 220                                        dev->name);
 221                                 dev->tx_timeout(dev);
 222                                 WARN_ON_ONCE(1);
 223                         }
 224                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 225                                 dev_hold(dev);
 226                 }
 227         }
 228         netif_tx_unlock(dev);
 229
 230         dev_put(dev);
 231 }
 232
 233 void __netdev_watchdog_up(struct net_device *dev)
 234 {
 235         if (dev->tx_timeout) {
 236                 if (dev->watchdog_timeo <= 0)
 237                         dev->watchdog_timeo = 5*HZ;
 238                 if (!mod_timer(&dev->watchdog_timer,
 239                                round_jiffies(jiffies + dev->watchdog_timeo)))
 240                         dev_hold(dev);
 241         }
 242 }
 243
 244 static void dev_watchdog_up(struct net_device *dev)
 245 {
 246         __netdev_watchdog_up(dev);
 247 }
 248
 249 static void dev_watchdog_down(struct net_device *dev)
 250 {
 251         netif_tx_lock_bh(dev);
 252         if (del_timer(&dev->watchdog_timer))
 253                 dev_put(dev);
 254         netif_tx_unlock_bh(dev);
 255 }
 256
 257 /**
 258  *      netif_carrier_on - set carrier
 259  *      @dev: network device
 260  *
 261  * Device has detected that carrier.
 262  */
 263 void netif_carrier_on(struct net_device *dev)
 264 {
 265         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 266                 linkwatch_fire_event(dev);
 267                 if (netif_running(dev))
 268                         __netdev_watchdog_up(dev);
 269         }
 270 }
 271 EXPORT_SYMBOL(netif_carrier_on);
 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 EXPORT_SYMBOL(netif_carrier_off);
 285
 286 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 287    under all circumstances. It is difficult to invent anything faster or
 288    cheaper.
 289  */
 290
 291 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 292 {
 293         kfree_skb(skb);
 294         return NET_XMIT_CN;
 295 }
 296
 297 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 298 {
 299         return NULL;
 300 }
 301
 302 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 303 {
 304         if (net_ratelimit())
 305                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 306                        skb->dev->name);
 307         kfree_skb(skb);
 308         return NET_XMIT_CN;
 309 }
 310
 311 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 312         .id             =       "noop",
 313         .priv_size      =       0,
 314         .enqueue        =       noop_enqueue,
 315         .dequeue        =       noop_dequeue,
 316         .requeue        =       noop_requeue,
 317         .owner          =       THIS_MODULE,
 318 };
 319
 320 struct Qdisc noop_qdisc = {
 321         .enqueue        =       noop_enqueue,
 322         .dequeue        =       noop_dequeue,
 323         .flags          =       TCQ_F_BUILTIN,
 324         .ops            =       &noop_qdisc_ops,
 325         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 326 };
 327 EXPORT_SYMBOL(noop_qdisc);
 328
 329 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 330         .id             =       "noqueue",
 331         .priv_size      =       0,
 332         .enqueue        =       noop_enqueue,
 333         .dequeue        =       noop_dequeue,
 334         .requeue        =       noop_requeue,
 335         .owner          =       THIS_MODULE,
 336 };
 337
 338 static struct Qdisc noqueue_qdisc = {
 339         .enqueue        =       NULL,
 340         .dequeue        =       noop_dequeue,
 341         .flags          =       TCQ_F_BUILTIN,
 342         .ops            =       &noqueue_qdisc_ops,
 343         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 344 };
 345
 346
 347 static const u8 prio2band[TC_PRIO_MAX+1] =
 348         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 349
 350 /* 3-band FIFO queue: old style, but should be a bit faster than
 351    generic prio+fifo combination.
 352  */
 353
 354 #define PFIFO_FAST_BANDS 3
 355
 356 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 357                                              struct Qdisc *qdisc)
 358 {
 359         struct sk_buff_head *list = qdisc_priv(qdisc);
 360         return list + prio2band[skb->priority & TC_PRIO_MAX];
 361 }
 362
 363 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 364 {
 365         struct sk_buff_head *list = prio2list(skb, qdisc);
 366
 367         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 368                 qdisc->q.qlen++;
 369                 return __qdisc_enqueue_tail(skb, qdisc, list);
 370         }
 371
 372         return qdisc_drop(skb, qdisc);
 373 }
 374
 375 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 376 {
 377         int prio;
 378         struct sk_buff_head *list = qdisc_priv(qdisc);
 379
 380         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 381                 if (!skb_queue_empty(list + prio)) {
 382                         qdisc->q.qlen--;
 383                         return __qdisc_dequeue_head(qdisc, list + prio);
 384                 }
 385         }
 386
 387         return NULL;
 388 }
 389
 390 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 391 {
 392         qdisc->q.qlen++;
 393         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 394 }
 395
 396 static void pfifo_fast_reset(struct Qdisc* qdisc)
 397 {
 398         int prio;
 399         struct sk_buff_head *list = qdisc_priv(qdisc);
 400
 401         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 402                 __qdisc_reset_queue(qdisc, list + prio);
 403
 404         qdisc->qstats.backlog = 0;
 405         qdisc->q.qlen = 0;
 406 }
 407
 408 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 409 {
 410         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 411
 412         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 413         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 414         return skb->len;
 415
 416 nla_put_failure:
 417         return -1;
 418 }
 419
 420 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 421 {
 422         int prio;
 423         struct sk_buff_head *list = qdisc_priv(qdisc);
 424
 425         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 426                 skb_queue_head_init(list + prio);
 427
 428         return 0;
 429 }
 430
 431 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 432         .id             =       "pfifo_fast",
 433         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 434         .enqueue        =       pfifo_fast_enqueue,
 435         .dequeue        =       pfifo_fast_dequeue,
 436         .requeue        =       pfifo_fast_requeue,
 437         .init           =       pfifo_fast_init,
 438         .reset          =       pfifo_fast_reset,
 439         .dump           =       pfifo_fast_dump,
 440         .owner          =       THIS_MODULE,
 441 };
 442
 443 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 444 {
 445         void *p;
 446         struct Qdisc *sch;
 447         unsigned int size;
 448         int err = -ENOBUFS;
 449
 450         /* ensure that the Qdisc and the private data are 32-byte aligned */
 451         size = QDISC_ALIGN(sizeof(*sch));
 452         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 453
 454         p = kzalloc(size, GFP_KERNEL);
 455         if (!p)
 456                 goto errout;
 457         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 458         sch->padded = (char *) sch - (char *) p;
 459
 460         INIT_LIST_HEAD(&sch->list);
 461         skb_queue_head_init(&sch->q);
 462         sch->ops = ops;
 463         sch->enqueue = ops->enqueue;
 464         sch->dequeue = ops->dequeue;
 465         sch->dev = dev;
 466         dev_hold(dev);
 467         atomic_set(&sch->refcnt, 1);
 468
 469         return sch;
 470 errout:
 471         return ERR_PTR(err);
 472 }
 473
 474 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 475                                  unsigned int parentid)
 476 {
 477         struct Qdisc *sch;
 478
 479         sch = qdisc_alloc(dev, ops);
 480         if (IS_ERR(sch))
 481                 goto errout;
 482         sch->stats_lock = &dev->queue_lock;
 483         sch->parent = parentid;
 484
 485         if (!ops->init || ops->init(sch, NULL) == 0)
 486                 return sch;
 487
 488         qdisc_destroy(sch);
 489 errout:
 490         return NULL;
 491 }
 492 EXPORT_SYMBOL(qdisc_create_dflt);
 493
 494 /* Under dev->queue_lock and BH! */
 495
 496 void qdisc_reset(struct Qdisc *qdisc)
 497 {
 498         const struct Qdisc_ops *ops = qdisc->ops;
 499
 500         if (ops->reset)
 501                 ops->reset(qdisc);
 502 }
 503 EXPORT_SYMBOL(qdisc_reset);
 504
 505 /* this is the rcu callback function to clean up a qdisc when there
 506  * are no further references to it */
 507
 508 static void __qdisc_destroy(struct rcu_head *head)
 509 {
 510         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 511         kfree((char *) qdisc - qdisc->padded);
 512 }
 513
 514 /* Under dev->queue_lock and BH! */
 515
 516 void qdisc_destroy(struct Qdisc *qdisc)
 517 {
 518         const struct Qdisc_ops  *ops = qdisc->ops;
 519
 520         if (qdisc->flags & TCQ_F_BUILTIN ||
 521             !atomic_dec_and_test(&qdisc->refcnt))
 522                 return;
 523
 524         list_del(&qdisc->list);
 525         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 526         if (ops->reset)
 527                 ops->reset(qdisc);
 528         if (ops->destroy)
 529                 ops->destroy(qdisc);
 530
 531         module_put(ops->owner);
 532         dev_put(qdisc->dev);
 533         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 534 }
 535 EXPORT_SYMBOL(qdisc_destroy);
 536
 537 void dev_activate(struct net_device *dev)
 538 {
 539         /* No queueing discipline is attached to device;
 540            create default one i.e. pfifo_fast for devices,
 541            which need queueing and noqueue_qdisc for
 542            virtual interfaces
 543          */
 544
 545         if (dev->qdisc_sleeping == &noop_qdisc) {
 546                 struct Qdisc *qdisc;
 547                 if (dev->tx_queue_len) {
 548                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 549                                                   TC_H_ROOT);
 550                         if (qdisc == NULL) {
 551                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 552                                 return;
 553                         }
 554                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 555                 } else {
 556                         qdisc =  &noqueue_qdisc;
 557                 }
 558                 dev->qdisc_sleeping = qdisc;
 559         }
 560
 561         if (!netif_carrier_ok(dev))
 562                 /* Delay activation until next carrier-on event */
 563                 return;
 564
 565         spin_lock_bh(&dev->queue_lock);
 566         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 567         if (dev->qdisc != &noqueue_qdisc) {
 568                 dev->trans_start = jiffies;
 569                 dev_watchdog_up(dev);
 570         }
 571         spin_unlock_bh(&dev->queue_lock);
 572 }
 573
 574 void dev_deactivate(struct net_device *dev)
 575 {
 576         struct Qdisc *qdisc;
 577         struct sk_buff *skb;
 578         int running;
 579
 580         spin_lock_bh(&dev->queue_lock);
 581         qdisc = dev->qdisc;
 582         dev->qdisc = &noop_qdisc;
 583
 584         qdisc_reset(qdisc);
 585
 586         skb = dev->gso_skb;
 587         dev->gso_skb = NULL;
 588         spin_unlock_bh(&dev->queue_lock);
 589
 590         kfree_skb(skb);
 591
 592         dev_watchdog_down(dev);
 593
 594         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 595         synchronize_rcu();
 596
 597         /* Wait for outstanding qdisc_run calls. */
 598         do {
 599                 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 600                         yield();
 601
 602                 /*
 603                  * Double-check inside queue lock to ensure that all effects
 604                  * of the queue run are visible when we return.
 605                  */
 606                 spin_lock_bh(&dev->queue_lock);
 607                 running = test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 608                 spin_unlock_bh(&dev->queue_lock);
 609
 610                 /*
 611                  * The running flag should never be set at this point because
 612                  * we've already set dev->qdisc to noop_qdisc *inside* the same
 613                  * pair of spin locks.  That is, if any qdisc_run starts after
 614                  * our initial test it should see the noop_qdisc and then
 615                  * clear the RUNNING bit before dropping the queue lock.  So
 616                  * if it is set here then we've found a bug.
 617                  */
 618         } while (WARN_ON_ONCE(running));
 619 }
 620
 621 void dev_init_scheduler(struct net_device *dev)
 622 {
 623         qdisc_lock_tree(dev);
 624         dev->qdisc = &noop_qdisc;
 625         dev->qdisc_sleeping = &noop_qdisc;
 626         INIT_LIST_HEAD(&dev->qdisc_list);
 627         qdisc_unlock_tree(dev);
 628
 629         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 630 }
 631
 632 void dev_shutdown(struct net_device *dev)
 633 {
 634         struct Qdisc *qdisc;
 635
 636         qdisc_lock_tree(dev);
 637         qdisc = dev->qdisc_sleeping;
 638         dev->qdisc = &noop_qdisc;
 639         dev->qdisc_sleeping = &noop_qdisc;
 640         qdisc_destroy(qdisc);
 641 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 642         if ((qdisc = dev->qdisc_ingress) != NULL) {
 643                 dev->qdisc_ingress = NULL;
 644                 qdisc_destroy(qdisc);
 645         }
 646 #endif
 647         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 648         qdisc_unlock_tree(dev);
 649 }