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