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 <asm/uaccess.h>
  15 #include <asm/system.h>
  16 #include <linux/bitops.h>
  17 #include <linux/module.h>
  18 #include <linux/types.h>
  19 #include <linux/kernel.h>
  20 #include <linux/sched.h>
  21 #include <linux/string.h>
  22 #include <linux/mm.h>
  23 #include <linux/socket.h>
  24 #include <linux/sockios.h>
  25 #include <linux/in.h>
  26 #include <linux/errno.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/netdevice.h>
  29 #include <linux/skbuff.h>
  30 #include <linux/rtnetlink.h>
  31 #include <linux/init.h>
  32 #include <linux/rcupdate.h>
  33 #include <linux/list.h>
  34 #include <net/sock.h>
  35 #include <net/pkt_sched.h>
  36
  37 /* Main transmission queue. */
  38
  39 /* Modifications to data participating in scheduling must be protected with
  40  * dev->queue_lock spinlock.
  41  *
  42  * The idea is the following:
  43  * - enqueue, dequeue are serialized via top level device
  44  *   spinlock dev->queue_lock.
  45  * - ingress filtering is serialized via top level device
  46  *   spinlock dev->ingress_lock.
  47  * - updates to tree and tree walking are only done under the rtnl mutex.
  48  */
  49
  50 void qdisc_lock_tree(struct net_device *dev)
  51 {
  52         spin_lock_bh(&dev->queue_lock);
  53         spin_lock(&dev->ingress_lock);
  54 }
  55
  56 void qdisc_unlock_tree(struct net_device *dev)
  57 {
  58         spin_unlock(&dev->ingress_lock);
  59         spin_unlock_bh(&dev->queue_lock);
  60 }
  61
  62 static inline int qdisc_qlen(struct Qdisc *q)
  63 {
  64         BUG_ON((int) q->q.qlen < 0);
  65         return q->q.qlen;
  66 }
  67
  68 static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
  69                                   struct Qdisc *q)
  70 {
  71         if (unlikely(skb->next))
  72                 dev->gso_skb = skb;
  73         else
  74                 q->ops->requeue(skb, q);
  75
  76         netif_schedule(dev);
  77         return 0;
  78 }
  79
  80 static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
  81                                               struct Qdisc *q)
  82 {
  83         struct sk_buff *skb;
  84
  85         if ((skb = dev->gso_skb))
  86                 dev->gso_skb = NULL;
  87         else
  88                 skb = q->dequeue(q);
  89
  90         return skb;
  91 }
  92
  93 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  94                                            struct net_device *dev,
  95                                            struct Qdisc *q)
  96 {
  97         int ret;
  98
  99         if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
 100                 /*
 101                  * Same CPU holding the lock. It may be a transient
 102                  * configuration error, when hard_start_xmit() recurses. We
 103                  * detect it by checking xmit owner and drop the packet when
 104                  * deadloop is detected. Return OK to try the next skb.
 105                  */
 106                 kfree_skb(skb);
 107                 if (net_ratelimit())
 108                         printk(KERN_WARNING "Dead loop on netdevice %s, "
 109                                "fix it urgently!\n", dev->name);
 110                 ret = qdisc_qlen(q);
 111         } else {
 112                 /*
 113                  * Another cpu is holding lock, requeue & delay xmits for
 114                  * some time.
 115                  */
 116                 __get_cpu_var(netdev_rx_stat).cpu_collision++;
 117                 ret = dev_requeue_skb(skb, dev, q);
 118         }
 119
 120         return ret;
 121 }
 122
 123 /*
 124  * NOTE: Called under dev->queue_lock with locally disabled BH.
 125  *
 126  * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
 127  * device at a time. dev->queue_lock serializes queue accesses for
 128  * this device AND dev->qdisc pointer itself.
 129  *
 130  *  netif_tx_lock serializes accesses to device driver.
 131  *
 132  *  dev->queue_lock and netif_tx_lock are mutually exclusive,
 133  *  if one is grabbed, another must be free.
 134  *
 135  * Note, that this procedure can be called by a watchdog timer
 136  *
 137  * Returns to the caller:
 138  *                              0  - queue is empty or throttled.
 139  *                              >0 - queue is not empty.
 140  *
 141  */
 142 static inline int qdisc_restart(struct net_device *dev)
 143 {
 144         struct Qdisc *q = dev->qdisc;
 145         struct sk_buff *skb;
 146         unsigned lockless;
 147         int ret;
 148
 149         /* Dequeue packet */
 150         if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
 151                 return 0;
 152
 153         /*
 154          * When the driver has LLTX set, it does its own locking in
 155          * start_xmit. These checks are worth it because even uncongested
 156          * locks can be quite expensive. The driver can do a trylock, as
 157          * is being done here; in case of lock contention it should return
 158          * NETDEV_TX_LOCKED and the packet will be requeued.
 159          */
 160         lockless = (dev->features & NETIF_F_LLTX);
 161
 162         if (!lockless && !netif_tx_trylock(dev)) {
 163                 /* Another CPU grabbed the driver tx lock */
 164                 return handle_dev_cpu_collision(skb, dev, q);
 165         }
 166
 167         /* And release queue */
 168         spin_unlock(&dev->queue_lock);
 169
 170         ret = NETDEV_TX_BUSY;
 171         if (!netif_queue_stopped(dev))
 172                 ret = dev_hard_start_xmit(skb, dev);
 173
 174         if (!lockless)
 175                 netif_tx_unlock(dev);
 176
 177         spin_lock(&dev->queue_lock);
 178         q = dev->qdisc;
 179
 180         switch (ret) {
 181         case NETDEV_TX_OK:
 182                 /* Driver sent out skb successfully */
 183                 ret = qdisc_qlen(q);
 184                 break;
 185
 186         case NETDEV_TX_LOCKED:
 187                 /* Driver try lock failed */
 188                 ret = handle_dev_cpu_collision(skb, dev, q);
 189                 break;
 190
 191         default:
 192                 /* Driver returned NETDEV_TX_BUSY - requeue skb */
 193                 if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
 194                         printk(KERN_WARNING "BUG %s code %d qlen %d\n",
 195                                dev->name, ret, q->q.qlen);
 196
 197                 ret = dev_requeue_skb(skb, dev, q);
 198                 break;
 199         }
 200
 201         return ret;
 202 }
 203
 204 void __qdisc_run(struct net_device *dev)
 205 {
 206         do {
 207                 if (!qdisc_restart(dev))
 208                         break;
 209         } while (!netif_queue_stopped(dev));
 210
 211         clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 212 }
 213
 214 static void dev_watchdog(unsigned long arg)
 215 {
 216         struct net_device *dev = (struct net_device *)arg;
 217
 218         netif_tx_lock(dev);
 219         if (dev->qdisc != &noop_qdisc) {
 220                 if (netif_device_present(dev) &&
 221                     netif_running(dev) &&
 222                     netif_carrier_ok(dev)) {
 223                         if (netif_queue_stopped(dev) &&
 224                             time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
 225
 226                                 printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
 227                                        dev->name);
 228                                 dev->tx_timeout(dev);
 229                         }
 230                         if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
 231                                 dev_hold(dev);
 232                 }
 233         }
 234         netif_tx_unlock(dev);
 235
 236         dev_put(dev);
 237 }
 238
 239 static void dev_watchdog_init(struct net_device *dev)
 240 {
 241         init_timer(&dev->watchdog_timer);
 242         dev->watchdog_timer.data = (unsigned long)dev;
 243         dev->watchdog_timer.function = dev_watchdog;
 244 }
 245
 246 void __netdev_watchdog_up(struct net_device *dev)
 247 {
 248         if (dev->tx_timeout) {
 249                 if (dev->watchdog_timeo <= 0)
 250                         dev->watchdog_timeo = 5*HZ;
 251                 if (!mod_timer(&dev->watchdog_timer,
 252                                round_jiffies(jiffies + dev->watchdog_timeo)))
 253                         dev_hold(dev);
 254         }
 255 }
 256
 257 static void dev_watchdog_up(struct net_device *dev)
 258 {
 259         __netdev_watchdog_up(dev);
 260 }
 261
 262 static void dev_watchdog_down(struct net_device *dev)
 263 {
 264         netif_tx_lock_bh(dev);
 265         if (del_timer(&dev->watchdog_timer))
 266                 dev_put(dev);
 267         netif_tx_unlock_bh(dev);
 268 }
 269
 270 void netif_carrier_on(struct net_device *dev)
 271 {
 272         if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
 273                 linkwatch_fire_event(dev);
 274         if (netif_running(dev))
 275                 __netdev_watchdog_up(dev);
 276 }
 277
 278 void netif_carrier_off(struct net_device *dev)
 279 {
 280         if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
 281                 linkwatch_fire_event(dev);
 282 }
 283
 284 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 285    under all circumstances. It is difficult to invent anything faster or
 286    cheaper.
 287  */
 288
 289 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 290 {
 291         kfree_skb(skb);
 292         return NET_XMIT_CN;
 293 }
 294
 295 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 296 {
 297         return NULL;
 298 }
 299
 300 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 301 {
 302         if (net_ratelimit())
 303                 printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
 304                        skb->dev->name);
 305         kfree_skb(skb);
 306         return NET_XMIT_CN;
 307 }
 308
 309 struct Qdisc_ops noop_qdisc_ops = {
 310         .id             =       "noop",
 311         .priv_size      =       0,
 312         .enqueue        =       noop_enqueue,
 313         .dequeue        =       noop_dequeue,
 314         .requeue        =       noop_requeue,
 315         .owner          =       THIS_MODULE,
 316 };
 317
 318 struct Qdisc noop_qdisc = {
 319         .enqueue        =       noop_enqueue,
 320         .dequeue        =       noop_dequeue,
 321         .flags          =       TCQ_F_BUILTIN,
 322         .ops            =       &noop_qdisc_ops,
 323         .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 324 };
 325
 326 static struct Qdisc_ops noqueue_qdisc_ops = {
 327         .id             =       "noqueue",
 328         .priv_size      =       0,
 329         .enqueue        =       noop_enqueue,
 330         .dequeue        =       noop_dequeue,
 331         .requeue        =       noop_requeue,
 332         .owner          =       THIS_MODULE,
 333 };
 334
 335 static struct Qdisc noqueue_qdisc = {
 336         .enqueue        =       NULL,
 337         .dequeue        =       noop_dequeue,
 338         .flags          =       TCQ_F_BUILTIN,
 339         .ops            =       &noqueue_qdisc_ops,
 340         .list           =       LIST_HEAD_INIT(noqueue_qdisc.list),
 341 };
 342
 343
 344 static const u8 prio2band[TC_PRIO_MAX+1] =
 345         { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
 346
 347 /* 3-band FIFO queue: old style, but should be a bit faster than
 348    generic prio+fifo combination.
 349  */
 350
 351 #define PFIFO_FAST_BANDS 3
 352
 353 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
 354                                              struct Qdisc *qdisc)
 355 {
 356         struct sk_buff_head *list = qdisc_priv(qdisc);
 357         return list + prio2band[skb->priority & TC_PRIO_MAX];
 358 }
 359
 360 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 361 {
 362         struct sk_buff_head *list = prio2list(skb, qdisc);
 363
 364         if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
 365                 qdisc->q.qlen++;
 366                 return __qdisc_enqueue_tail(skb, qdisc, list);
 367         }
 368
 369         return qdisc_drop(skb, qdisc);
 370 }
 371
 372 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 373 {
 374         int prio;
 375         struct sk_buff_head *list = qdisc_priv(qdisc);
 376
 377         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 378                 if (!skb_queue_empty(list + prio)) {
 379                         qdisc->q.qlen--;
 380                         return __qdisc_dequeue_head(qdisc, list + prio);
 381                 }
 382         }
 383
 384         return NULL;
 385 }
 386
 387 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 388 {
 389         qdisc->q.qlen++;
 390         return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 391 }
 392
 393 static void pfifo_fast_reset(struct Qdisc* qdisc)
 394 {
 395         int prio;
 396         struct sk_buff_head *list = qdisc_priv(qdisc);
 397
 398         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 399                 __qdisc_reset_queue(qdisc, list + prio);
 400
 401         qdisc->qstats.backlog = 0;
 402         qdisc->q.qlen = 0;
 403 }
 404
 405 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 406 {
 407         struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 408
 409         memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
 410         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 411         return skb->len;
 412
 413 rtattr_failure:
 414         return -1;
 415 }
 416
 417 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 418 {
 419         int prio;
 420         struct sk_buff_head *list = qdisc_priv(qdisc);
 421
 422         for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
 423                 skb_queue_head_init(list + prio);
 424
 425         return 0;
 426 }
 427
 428 static struct Qdisc_ops pfifo_fast_ops = {
 429         .id             =       "pfifo_fast",
 430         .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
 431         .enqueue        =       pfifo_fast_enqueue,
 432         .dequeue        =       pfifo_fast_dequeue,
 433         .requeue        =       pfifo_fast_requeue,
 434         .init           =       pfifo_fast_init,
 435         .reset          =       pfifo_fast_reset,
 436         .dump           =       pfifo_fast_dump,
 437         .owner          =       THIS_MODULE,
 438 };
 439
 440 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 441 {
 442         void *p;
 443         struct Qdisc *sch;
 444         unsigned int size;
 445         int err = -ENOBUFS;
 446
 447         /* ensure that the Qdisc and the private data are 32-byte aligned */
 448         size = QDISC_ALIGN(sizeof(*sch));
 449         size += ops->priv_size + (QDISC_ALIGNTO - 1);
 450
 451         p = kzalloc(size, GFP_KERNEL);
 452         if (!p)
 453                 goto errout;
 454         sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 455         sch->padded = (char *) sch - (char *) p;
 456
 457         INIT_LIST_HEAD(&sch->list);
 458         skb_queue_head_init(&sch->q);
 459         sch->ops = ops;
 460         sch->enqueue = ops->enqueue;
 461         sch->dequeue = ops->dequeue;
 462         sch->dev = dev;
 463         dev_hold(dev);
 464         atomic_set(&sch->refcnt, 1);
 465
 466         return sch;
 467 errout:
 468         return ERR_PTR(-err);
 469 }
 470
 471 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
 472                                  unsigned int parentid)
 473 {
 474         struct Qdisc *sch;
 475
 476         sch = qdisc_alloc(dev, ops);
 477         if (IS_ERR(sch))
 478                 goto errout;
 479         sch->stats_lock = &dev->queue_lock;
 480         sch->parent = parentid;
 481
 482         if (!ops->init || ops->init(sch, NULL) == 0)
 483                 return sch;
 484
 485         qdisc_destroy(sch);
 486 errout:
 487         return NULL;
 488 }
 489
 490 /* Under dev->queue_lock and BH! */
 491
 492 void qdisc_reset(struct Qdisc *qdisc)
 493 {
 494         struct Qdisc_ops *ops = qdisc->ops;
 495
 496         if (ops->reset)
 497                 ops->reset(qdisc);
 498 }
 499
 500 /* this is the rcu callback function to clean up a qdisc when there
 501  * are no further references to it */
 502
 503 static void __qdisc_destroy(struct rcu_head *head)
 504 {
 505         struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
 506         kfree((char *) qdisc - qdisc->padded);
 507 }
 508
 509 /* Under dev->queue_lock and BH! */
 510
 511 void qdisc_destroy(struct Qdisc *qdisc)
 512 {
 513         struct Qdisc_ops  *ops = qdisc->ops;
 514
 515         if (qdisc->flags & TCQ_F_BUILTIN ||
 516             !atomic_dec_and_test(&qdisc->refcnt))
 517                 return;
 518
 519         list_del(&qdisc->list);
 520 #ifdef CONFIG_NET_ESTIMATOR
 521         gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
 522 #endif
 523         if (ops->reset)
 524                 ops->reset(qdisc);
 525         if (ops->destroy)
 526                 ops->destroy(qdisc);
 527
 528         module_put(ops->owner);
 529         dev_put(qdisc->dev);
 530         call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 531 }
 532
 533 void dev_activate(struct net_device *dev)
 534 {
 535         /* No queueing discipline is attached to device;
 536            create default one i.e. pfifo_fast for devices,
 537            which need queueing and noqueue_qdisc for
 538            virtual interfaces
 539          */
 540
 541         if (dev->qdisc_sleeping == &noop_qdisc) {
 542                 struct Qdisc *qdisc;
 543                 if (dev->tx_queue_len) {
 544                         qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
 545                                                   TC_H_ROOT);
 546                         if (qdisc == NULL) {
 547                                 printk(KERN_INFO "%s: activation failed\n", dev->name);
 548                                 return;
 549                         }
 550                         list_add_tail(&qdisc->list, &dev->qdisc_list);
 551                 } else {
 552                         qdisc =  &noqueue_qdisc;
 553                 }
 554                 dev->qdisc_sleeping = qdisc;
 555         }
 556
 557         if (!netif_carrier_ok(dev))
 558                 /* Delay activation until next carrier-on event */
 559                 return;
 560
 561         spin_lock_bh(&dev->queue_lock);
 562         rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
 563         if (dev->qdisc != &noqueue_qdisc) {
 564                 dev->trans_start = jiffies;
 565                 dev_watchdog_up(dev);
 566         }
 567         spin_unlock_bh(&dev->queue_lock);
 568 }
 569
 570 void dev_deactivate(struct net_device *dev)
 571 {
 572         struct Qdisc *qdisc;
 573         struct sk_buff *skb;
 574
 575         spin_lock_bh(&dev->queue_lock);
 576         qdisc = dev->qdisc;
 577         dev->qdisc = &noop_qdisc;
 578
 579         qdisc_reset(qdisc);
 580
 581         skb = dev->gso_skb;
 582         dev->gso_skb = NULL;
 583         spin_unlock_bh(&dev->queue_lock);
 584
 585         kfree_skb(skb);
 586
 587         dev_watchdog_down(dev);
 588
 589         /* Wait for outstanding dev_queue_xmit calls. */
 590         synchronize_rcu();
 591
 592         /* Wait for outstanding qdisc_run calls. */
 593         while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
 594                 yield();
 595 }
 596
 597 void dev_init_scheduler(struct net_device *dev)
 598 {
 599         qdisc_lock_tree(dev);
 600         dev->qdisc = &noop_qdisc;
 601         dev->qdisc_sleeping = &noop_qdisc;
 602         INIT_LIST_HEAD(&dev->qdisc_list);
 603         qdisc_unlock_tree(dev);
 604
 605         dev_watchdog_init(dev);
 606 }
 607
 608 void dev_shutdown(struct net_device *dev)
 609 {
 610         struct Qdisc *qdisc;
 611
 612         qdisc_lock_tree(dev);
 613         qdisc = dev->qdisc_sleeping;
 614         dev->qdisc = &noop_qdisc;
 615         dev->qdisc_sleeping = &noop_qdisc;
 616         qdisc_destroy(qdisc);
 617 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
 618         if ((qdisc = dev->qdisc_ingress) != NULL) {
 619                 dev->qdisc_ingress = NULL;
 620                 qdisc_destroy(qdisc);
 621         }
 622 #endif
 623         BUG_TRAP(!timer_pending(&dev->watchdog_timer));
 624         qdisc_unlock_tree(dev);
 625 }
 626
 627 EXPORT_SYMBOL(netif_carrier_on);
 628 EXPORT_SYMBOL(netif_carrier_off);
 629 EXPORT_SYMBOL(noop_qdisc);
 630 EXPORT_SYMBOL(qdisc_create_dflt);
 631 EXPORT_SYMBOL(qdisc_destroy);
 632 EXPORT_SYMBOL(qdisc_reset);
 633 EXPORT_SYMBOL(qdisc_lock_tree);
 634 EXPORT_SYMBOL(qdisc_unlock_tree);