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         do {
 188                 if (!qdisc_restart(dev))
 189                         break;
 190         } while (!netif_queue_stopped(dev));
 191
 192         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 193 }
 194
 195 static void dev_watchdog(unsigned long arg)
 196 {
 197         struct net_device *dev = (struct net_device *)arg;
 198
 199         netif_tx_lock(dev);
 200         if (dev->qdisc != &noop_qdisc) {
 201                 if (netif_device_present(dev) &&
 202                     netif_running(dev) &&
 203                     netif_carrier_ok(dev)) {
 204                         if (netif_queue_stopped(dev) &&
 205                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 206
 207                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 208                                        dev->name);
 209                                 dev->tx_timeout(dev);
 210                         }
 211                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 212                                 dev_hold(dev);
 213                 }
 214         }
 215         netif_tx_unlock(dev);
 216
 217         dev_put(dev);
 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 EXPORT_SYMBOL(netif_carrier_on);
 259
 260 /**
 261  *      netif_carrier_off - clear carrier
 262  *      @dev: network device
 263  *
 264  * Device has detected loss of carrier.
 265  */
 266 void netif_carrier_off(struct net_device *dev)
 267 {
 268         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 269                 linkwatch_fire_event(dev);
 270 }
 271 EXPORT_SYMBOL(netif_carrier_off);
 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 __read_mostly = {
 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 EXPORT_SYMBOL(noop_qdisc);
 315
 316 static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 317         .id             =       "noqueue",
 318         .priv_size      =       0,
 319         .enqueue        =       noop_enqueue,
 320         .dequeue        =       noop_dequeue,
 321         .requeue        =       noop_requeue,
 322         .owner          =       THIS_MODULE,
 323 };
 324
 325 static struct Qdisc noqueue_qdisc = {
 326         .enqueue        =       NULL,
 327         .dequeue        =       noop_dequeue,
 328         .flags          =       TCQ_F_BUILTIN,
 329         .ops            =       &noqueue_qdisc_ops,
 330         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 331 };
 332
 333
 334 static const u8 prio2band[TC_PRIO_MAX+1] =
 335         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 336
 337 /* 3-band FIFO queue: old style, but should be a bit faster than
 338    generic prio+fifo combination.
 339  */
 340
 341 #define PFIFO_FAST_BANDS 3
 342
 343 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 344                                              struct Qdisc *qdisc)
 345 {
 346         struct sk_buff_head *list = qdisc_priv(qdisc);
 347         return list + prio2band[skb->priority & TC_PRIO_MAX];
 348 }
 349
 350 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 351 {
 352         struct sk_buff_head *list = prio2list(skb, qdisc);
 353
 354         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 355                 qdisc->q.qlen++;
 356                 return __qdisc_enqueue_tail(skb, qdisc, list);
 357         }
 358
 359         return qdisc_drop(skb, qdisc);
 360 }
 361
 362 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 363 {
 364         int prio;
 365         struct sk_buff_head *list = qdisc_priv(qdisc);
 366
 367         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 368                 if (!skb_queue_empty(list + prio)) {
 369                         qdisc->q.qlen--;
 370                         return __qdisc_dequeue_head(qdisc, list + prio);
 371                 }
 372         }
 373
 374         return NULL;
 375 }
 376
 377 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 378 {
 379         qdisc->q.qlen++;
 380         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 381 }
 382
 383 static void pfifo_fast_reset(struct Qdisc* qdisc)
 384 {
 385         int prio;
 386         struct sk_buff_head *list = qdisc_priv(qdisc);
 387
 388         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 389                 __qdisc_reset_queue(qdisc, list + prio);
 390
 391         qdisc->qstats.backlog = 0;
 392         qdisc->q.qlen = 0;
 393 }
 394
 395 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 396 {
 397         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 398
 399         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 400         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 401         return skb->len;
 402
 403 nla_put_failure:
 404         return -1;
 405 }
 406
 407 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 408 {
 409         int prio;
 410         struct sk_buff_head *list = qdisc_priv(qdisc);
 411
 412         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 413                 skb_queue_head_init(list + prio);
 414
 415         return 0;
 416 }
 417
 418 static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 419         .id             =       "pfifo_fast",
 420         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 421         .enqueue        =       pfifo_fast_enqueue,
 422         .dequeue        =       pfifo_fast_dequeue,
 423         .requeue        =       pfifo_fast_requeue,
 424         .init           =       pfifo_fast_init,
 425         .reset          =       pfifo_fast_reset,
 426         .dump           =       pfifo_fast_dump,
 427         .owner          =       THIS_MODULE,
 428 };
 429
 430 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 431 {
 432         void *p;
 433         struct Qdisc *sch;
 434         unsigned int size;
 435         int err = -ENOBUFS;
 436
 437         /* ensure that the Qdisc and the private data are 32-byte aligned */
 438         size = QDISC_ALIGN(sizeof(*sch));
 439         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 440
 441         p = kzalloc(size, GFP_KERNEL);
 442         if (!p)
 443                 goto errout;
 444         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 445         sch->padded = (char *) sch - (char *) p;
 446
 447         INIT_LIST_HEAD(&sch->list);
 448         skb_queue_head_init(&sch->q);
 449         sch->ops = ops;
 450         sch->enqueue = ops->enqueue;
 451         sch->dequeue = ops->dequeue;
 452         sch->dev = dev;
 453         dev_hold(dev);
 454         atomic_set(&sch->refcnt, 1);
 455
 456         return sch;
 457 errout:
 458         return ERR_PTR(-err);
 459 }
 460
 461 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 462                                  unsigned int parentid)
 463 {
 464         struct Qdisc *sch;
 465
 466         sch = qdisc_alloc(dev, ops);
 467         if (IS_ERR(sch))
 468                 goto errout;
 469         sch->stats_lock = &dev->queue_lock;
 470         sch->parent = parentid;
 471
 472         if (!ops->init || ops->init(sch, NULL) == 0)
 473                 return sch;
 474
 475         qdisc_destroy(sch);
 476 errout:
 477         return NULL;
 478 }
 479 EXPORT_SYMBOL(qdisc_create_dflt);
 480
 481 /* Under dev->queue_lock and BH! */
 482
 483 void qdisc_reset(struct Qdisc *qdisc)
 484 {
 485         const struct Qdisc_ops *ops = qdisc->ops;
 486
 487         if (ops->reset)
 488                 ops->reset(qdisc);
 489 }
 490 EXPORT_SYMBOL(qdisc_reset);
 491
 492 /* this is the rcu callback function to clean up a qdisc when there
 493  * are no further references to it */
 494
 495 static void __qdisc_destroy(struct rcu_head *head)
 496 {
 497         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 498         kfree((char *) qdisc - qdisc->padded);
 499 }
 500
 501 /* Under dev->queue_lock and BH! */
 502
 503 void qdisc_destroy(struct Qdisc *qdisc)
 504 {
 505         const struct Qdisc_ops  *ops = qdisc->ops;
 506
 507         if (qdisc->flags & TCQ_F_BUILTIN ||
 508             !atomic_dec_and_test(&qdisc->refcnt))
 509                 return;
 510
 511         list_del(&qdisc->list);
 512         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 513         if (ops->reset)
 514                 ops->reset(qdisc);
 515         if (ops->destroy)
 516                 ops->destroy(qdisc);
 517
 518         module_put(ops->owner);
 519         dev_put(qdisc->dev);
 520         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 521 }
 522 EXPORT_SYMBOL(qdisc_destroy);
 523
 524 void dev_activate(struct net_device *dev)
 525 {
 526         /* No queueing discipline is attached to device;
 527            create default one i.e. pfifo_fast for devices,
 528            which need queueing and noqueue_qdisc for
 529            virtual interfaces
 530          */
 531
 532         if (dev->qdisc_sleeping == &noop_qdisc) {
 533                 struct Qdisc *qdisc;
 534                 if (dev->tx_queue_len) {
 535                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 536                                                   TC_H_ROOT);
 537                         if (qdisc == NULL) {
 538                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 539                                 return;
 540                         }
 541                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 542                 } else {
 543                         qdisc =  &noqueue_qdisc;
 544                 }
 545                 dev->qdisc_sleeping = qdisc;
 546         }
 547
 548         if (!netif_carrier_ok(dev))
 549                 /* Delay activation until next carrier-on event */
 550                 return;
 551
 552         spin_lock_bh(&dev->queue_lock);
 553         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 554         if (dev->qdisc != &noqueue_qdisc) {
 555                 dev->trans_start = jiffies;
 556                 dev_watchdog_up(dev);
 557         }
 558         spin_unlock_bh(&dev->queue_lock);
 559 }
 560
 561 void dev_deactivate(struct net_device *dev)
 562 {
 563         struct Qdisc *qdisc;
 564         struct sk_buff *skb;
 565         int running;
 566
 567         spin_lock_bh(&dev->queue_lock);
 568         qdisc = dev->qdisc;
 569         dev->qdisc = &noop_qdisc;
 570
 571         qdisc_reset(qdisc);
 572
 573         skb = dev->gso_skb;
 574         dev->gso_skb = NULL;
 575         spin_unlock_bh(&dev->queue_lock);
 576
 577         kfree_skb(skb);
 578
 579         dev_watchdog_down(dev);
 580
 581         /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
 582         synchronize_rcu();
 583
 584         /* Wait for outstanding qdisc_run calls. */
 585         do {
 586                 while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 587                         yield();
 588
 589                 /*
 590                  * Double-check inside queue lock to ensure that all effects
 591                  * of the queue run are visible when we return.
 592                  */
 593                 spin_lock_bh(&dev->queue_lock);
 594                 running = test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 595                 spin_unlock_bh(&dev->queue_lock);
 596
 597                 /*
 598                  * The running flag should never be set at this point because
 599                  * we've already set dev->qdisc to noop_qdisc *inside* the same
 600                  * pair of spin locks.  That is, if any qdisc_run starts after
 601                  * our initial test it should see the noop_qdisc and then
 602                  * clear the RUNNING bit before dropping the queue lock.  So
 603                  * if it is set here then we've found a bug.
 604                  */
 605         } while (WARN_ON_ONCE(running));
 606 }
 607
 608 void dev_init_scheduler(struct net_device *dev)
 609 {
 610         qdisc_lock_tree(dev);
 611         dev->qdisc = &noop_qdisc;
 612         dev->qdisc_sleeping = &noop_qdisc;
 613         INIT_LIST_HEAD(&dev->qdisc_list);
 614         qdisc_unlock_tree(dev);
 615
 616         setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 617 }
 618
 619 void dev_shutdown(struct net_device *dev)
 620 {
 621         struct Qdisc *qdisc;
 622
 623         qdisc_lock_tree(dev);
 624         qdisc = dev->qdisc_sleeping;
 625         dev->qdisc = &noop_qdisc;
 626         dev->qdisc_sleeping = &noop_qdisc;
 627         qdisc_destroy(qdisc);
 628 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 629         if ((qdisc = dev->qdisc_ingress) != NULL) {
 630                 dev->qdisc_ingress = NULL;
 631                 qdisc_destroy(qdisc);
 632         }
 633 #endif
 634         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 635         qdisc_unlock_tree(dev);
 636 }