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 <linux/slab.h> 28 #include <linux/if_vlan.h> 29 #include <net/sch_generic.h> 30 #include <net/pkt_sched.h> 31 #include <net/dst.h> 32 #include <trace/events/qdisc.h> 33 34 /* Qdisc to use by default */ 35 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops; 36 EXPORT_SYMBOL(default_qdisc_ops); 37 38 /* Main transmission queue. */ 39 40 /* Modifications to data participating in scheduling must be protected with 41 * qdisc_lock(qdisc) spinlock. 42 * 43 * The idea is the following: 44 * - enqueue, dequeue are serialized via qdisc root lock 45 * - ingress filtering is also serialized via qdisc root lock 46 * - updates to tree and tree walking are only done under the rtnl mutex. 47 */ 48 49 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q) 50 { 51 q->gso_skb = skb; 52 q->qstats.requeues++; 53 qdisc_qstats_backlog_inc(q, skb); 54 q->q.qlen++; /* it's still part of the queue */ 55 __netif_schedule(q); 56 57 return 0; 58 } 59 60 static void try_bulk_dequeue_skb(struct Qdisc *q, 61 struct sk_buff *skb, 62 const struct netdev_queue *txq, 63 int *packets) 64 { 65 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len; 66 67 while (bytelimit > 0) { 68 struct sk_buff *nskb = q->dequeue(q); 69 70 if (!nskb) 71 break; 72 73 bytelimit -= nskb->len; /* covers GSO len */ 74 skb->next = nskb; 75 skb = nskb; 76 (*packets)++; /* GSO counts as one pkt */ 77 } 78 skb->next = NULL; 79 } 80 81 /* This variant of try_bulk_dequeue_skb() makes sure 82 * all skbs in the chain are for the same txq 83 */ 84 static void try_bulk_dequeue_skb_slow(struct Qdisc *q, 85 struct sk_buff *skb, 86 int *packets) 87 { 88 int mapping = skb_get_queue_mapping(skb); 89 struct sk_buff *nskb; 90 int cnt = 0; 91 92 do { 93 nskb = q->dequeue(q); 94 if (!nskb) 95 break; 96 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) { 97 q->skb_bad_txq = nskb; 98 qdisc_qstats_backlog_inc(q, nskb); 99 q->q.qlen++; 100 break; 101 } 102 skb->next = nskb; 103 skb = nskb; 104 } while (++cnt < 8); 105 (*packets) += cnt; 106 skb->next = NULL; 107 } 108 109 /* Note that dequeue_skb can possibly return a SKB list (via skb->next). 110 * A requeued skb (via q->gso_skb) can also be a SKB list. 111 */ 112 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, 113 int *packets) 114 { 115 struct sk_buff *skb = q->gso_skb; 116 const struct netdev_queue *txq = q->dev_queue; 117 118 *packets = 1; 119 if (unlikely(skb)) { 120 /* skb in gso_skb were already validated */ 121 *validate = false; 122 /* check the reason of requeuing without tx lock first */ 123 txq = skb_get_tx_queue(txq->dev, skb); 124 if (!netif_xmit_frozen_or_stopped(txq)) { 125 q->gso_skb = NULL; 126 qdisc_qstats_backlog_dec(q, skb); 127 q->q.qlen--; 128 } else 129 skb = NULL; 130 goto trace; 131 } 132 *validate = true; 133 skb = q->skb_bad_txq; 134 if (unlikely(skb)) { 135 /* check the reason of requeuing without tx lock first */ 136 txq = skb_get_tx_queue(txq->dev, skb); 137 if (!netif_xmit_frozen_or_stopped(txq)) { 138 q->skb_bad_txq = NULL; 139 qdisc_qstats_backlog_dec(q, skb); 140 q->q.qlen--; 141 goto bulk; 142 } 143 skb = NULL; 144 goto trace; 145 } 146 if (!(q->flags & TCQ_F_ONETXQUEUE) || 147 !netif_xmit_frozen_or_stopped(txq)) 148 skb = q->dequeue(q); 149 if (skb) { 150 bulk: 151 if (qdisc_may_bulk(q)) 152 try_bulk_dequeue_skb(q, skb, txq, packets); 153 else 154 try_bulk_dequeue_skb_slow(q, skb, packets); 155 } 156 trace: 157 trace_qdisc_dequeue(q, txq, *packets, skb); 158 return skb; 159 } 160 161 /* 162 * Transmit possibly several skbs, and handle the return status as 163 * required. Owning running seqcount bit guarantees that 164 * only one CPU can execute this function. 165 * 166 * Returns to the caller: 167 * 0 - queue is empty or throttled. 168 * >0 - queue is not empty. 169 */ 170 int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q, 171 struct net_device *dev, struct netdev_queue *txq, 172 spinlock_t *root_lock, bool validate) 173 { 174 int ret = NETDEV_TX_BUSY; 175 176 /* And release qdisc */ 177 if (root_lock) 178 spin_unlock(root_lock); 179 180 /* Note that we validate skb (GSO, checksum, ...) outside of locks */ 181 if (validate) 182 skb = validate_xmit_skb_list(skb, dev); 183 184 if (likely(skb)) { 185 HARD_TX_LOCK(dev, txq, smp_processor_id()); 186 if (!netif_xmit_frozen_or_stopped(txq)) 187 skb = dev_hard_start_xmit(skb, dev, txq, &ret); 188 189 HARD_TX_UNLOCK(dev, txq); 190 } else { 191 if (root_lock) 192 spin_lock(root_lock); 193 return qdisc_qlen(q); 194 } 195 196 if (root_lock) 197 spin_lock(root_lock); 198 199 if (dev_xmit_complete(ret)) { 200 /* Driver sent out skb successfully or skb was consumed */ 201 ret = qdisc_qlen(q); 202 } else { 203 /* Driver returned NETDEV_TX_BUSY - requeue skb */ 204 if (unlikely(ret != NETDEV_TX_BUSY)) 205 net_warn_ratelimited("BUG %s code %d qlen %d\n", 206 dev->name, ret, q->q.qlen); 207 208 ret = dev_requeue_skb(skb, q); 209 } 210 211 if (ret && netif_xmit_frozen_or_stopped(txq)) 212 ret = 0; 213 214 return ret; 215 } 216 217 /* 218 * NOTE: Called under qdisc_lock(q) with locally disabled BH. 219 * 220 * running seqcount guarantees only one CPU can process 221 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for 222 * this queue. 223 * 224 * netif_tx_lock serializes accesses to device driver. 225 * 226 * qdisc_lock(q) and netif_tx_lock are mutually exclusive, 227 * if one is grabbed, another must be free. 228 * 229 * Note, that this procedure can be called by a watchdog timer 230 * 231 * Returns to the caller: 232 * 0 - queue is empty or throttled. 233 * >0 - queue is not empty. 234 * 235 */ 236 static inline int qdisc_restart(struct Qdisc *q, int *packets) 237 { 238 spinlock_t *root_lock = NULL; 239 struct netdev_queue *txq; 240 struct net_device *dev; 241 struct sk_buff *skb; 242 bool validate; 243 244 /* Dequeue packet */ 245 skb = dequeue_skb(q, &validate, packets); 246 if (unlikely(!skb)) 247 return 0; 248 249 if (!(q->flags & TCQ_F_NOLOCK)) 250 root_lock = qdisc_lock(q); 251 252 dev = qdisc_dev(q); 253 txq = skb_get_tx_queue(dev, skb); 254 255 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate); 256 } 257 258 void __qdisc_run(struct Qdisc *q) 259 { 260 int quota = dev_tx_weight; 261 int packets; 262 263 while (qdisc_restart(q, &packets)) { 264 /* 265 * Ordered by possible occurrence: Postpone processing if 266 * 1. we've exceeded packet quota 267 * 2. another process needs the CPU; 268 */ 269 quota -= packets; 270 if (quota <= 0 || need_resched()) { 271 __netif_schedule(q); 272 break; 273 } 274 } 275 } 276 277 unsigned long dev_trans_start(struct net_device *dev) 278 { 279 unsigned long val, res; 280 unsigned int i; 281 282 if (is_vlan_dev(dev)) 283 dev = vlan_dev_real_dev(dev); 284 res = netdev_get_tx_queue(dev, 0)->trans_start; 285 for (i = 1; i < dev->num_tx_queues; i++) { 286 val = netdev_get_tx_queue(dev, i)->trans_start; 287 if (val && time_after(val, res)) 288 res = val; 289 } 290 291 return res; 292 } 293 EXPORT_SYMBOL(dev_trans_start); 294 295 static void dev_watchdog(struct timer_list *t) 296 { 297 struct net_device *dev = from_timer(dev, t, watchdog_timer); 298 299 netif_tx_lock(dev); 300 if (!qdisc_tx_is_noop(dev)) { 301 if (netif_device_present(dev) && 302 netif_running(dev) && 303 netif_carrier_ok(dev)) { 304 int some_queue_timedout = 0; 305 unsigned int i; 306 unsigned long trans_start; 307 308 for (i = 0; i < dev->num_tx_queues; i++) { 309 struct netdev_queue *txq; 310 311 txq = netdev_get_tx_queue(dev, i); 312 trans_start = txq->trans_start; 313 if (netif_xmit_stopped(txq) && 314 time_after(jiffies, (trans_start + 315 dev->watchdog_timeo))) { 316 some_queue_timedout = 1; 317 txq->trans_timeout++; 318 break; 319 } 320 } 321 322 if (some_queue_timedout) { 323 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n", 324 dev->name, netdev_drivername(dev), i); 325 dev->netdev_ops->ndo_tx_timeout(dev); 326 } 327 if (!mod_timer(&dev->watchdog_timer, 328 round_jiffies(jiffies + 329 dev->watchdog_timeo))) 330 dev_hold(dev); 331 } 332 } 333 netif_tx_unlock(dev); 334 335 dev_put(dev); 336 } 337 338 void __netdev_watchdog_up(struct net_device *dev) 339 { 340 if (dev->netdev_ops->ndo_tx_timeout) { 341 if (dev->watchdog_timeo <= 0) 342 dev->watchdog_timeo = 5*HZ; 343 if (!mod_timer(&dev->watchdog_timer, 344 round_jiffies(jiffies + dev->watchdog_timeo))) 345 dev_hold(dev); 346 } 347 } 348 349 static void dev_watchdog_up(struct net_device *dev) 350 { 351 __netdev_watchdog_up(dev); 352 } 353 354 static void dev_watchdog_down(struct net_device *dev) 355 { 356 netif_tx_lock_bh(dev); 357 if (del_timer(&dev->watchdog_timer)) 358 dev_put(dev); 359 netif_tx_unlock_bh(dev); 360 } 361 362 /** 363 * netif_carrier_on - set carrier 364 * @dev: network device 365 * 366 * Device has detected that carrier. 367 */ 368 void netif_carrier_on(struct net_device *dev) 369 { 370 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) { 371 if (dev->reg_state == NETREG_UNINITIALIZED) 372 return; 373 atomic_inc(&dev->carrier_changes); 374 linkwatch_fire_event(dev); 375 if (netif_running(dev)) 376 __netdev_watchdog_up(dev); 377 } 378 } 379 EXPORT_SYMBOL(netif_carrier_on); 380 381 /** 382 * netif_carrier_off - clear carrier 383 * @dev: network device 384 * 385 * Device has detected loss of carrier. 386 */ 387 void netif_carrier_off(struct net_device *dev) 388 { 389 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) { 390 if (dev->reg_state == NETREG_UNINITIALIZED) 391 return; 392 atomic_inc(&dev->carrier_changes); 393 linkwatch_fire_event(dev); 394 } 395 } 396 EXPORT_SYMBOL(netif_carrier_off); 397 398 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces 399 under all circumstances. It is difficult to invent anything faster or 400 cheaper. 401 */ 402 403 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc, 404 struct sk_buff **to_free) 405 { 406 __qdisc_drop(skb, to_free); 407 return NET_XMIT_CN; 408 } 409 410 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc) 411 { 412 return NULL; 413 } 414 415 struct Qdisc_ops noop_qdisc_ops __read_mostly = { 416 .id = "noop", 417 .priv_size = 0, 418 .enqueue = noop_enqueue, 419 .dequeue = noop_dequeue, 420 .peek = noop_dequeue, 421 .owner = THIS_MODULE, 422 }; 423 424 static struct netdev_queue noop_netdev_queue = { 425 .qdisc = &noop_qdisc, 426 .qdisc_sleeping = &noop_qdisc, 427 }; 428 429 struct Qdisc noop_qdisc = { 430 .enqueue = noop_enqueue, 431 .dequeue = noop_dequeue, 432 .flags = TCQ_F_BUILTIN, 433 .ops = &noop_qdisc_ops, 434 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock), 435 .dev_queue = &noop_netdev_queue, 436 .running = SEQCNT_ZERO(noop_qdisc.running), 437 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock), 438 }; 439 EXPORT_SYMBOL(noop_qdisc); 440 441 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt) 442 { 443 /* register_qdisc() assigns a default of noop_enqueue if unset, 444 * but __dev_queue_xmit() treats noqueue only as such 445 * if this is NULL - so clear it here. */ 446 qdisc->enqueue = NULL; 447 return 0; 448 } 449 450 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = { 451 .id = "noqueue", 452 .priv_size = 0, 453 .init = noqueue_init, 454 .enqueue = noop_enqueue, 455 .dequeue = noop_dequeue, 456 .peek = noop_dequeue, 457 .owner = THIS_MODULE, 458 }; 459 460 static const u8 prio2band[TC_PRIO_MAX + 1] = { 461 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 462 }; 463 464 /* 3-band FIFO queue: old style, but should be a bit faster than 465 generic prio+fifo combination. 466 */ 467 468 #define PFIFO_FAST_BANDS 3 469 470 /* 471 * Private data for a pfifo_fast scheduler containing: 472 * - queues for the three band 473 * - bitmap indicating which of the bands contain skbs 474 */ 475 struct pfifo_fast_priv { 476 u32 bitmap; 477 struct qdisc_skb_head q[PFIFO_FAST_BANDS]; 478 }; 479 480 /* 481 * Convert a bitmap to the first band number where an skb is queued, where: 482 * bitmap=0 means there are no skbs on any band. 483 * bitmap=1 means there is an skb on band 0. 484 * bitmap=7 means there are skbs on all 3 bands, etc. 485 */ 486 static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0}; 487 488 static inline struct qdisc_skb_head *band2list(struct pfifo_fast_priv *priv, 489 int band) 490 { 491 return priv->q + band; 492 } 493 494 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc, 495 struct sk_buff **to_free) 496 { 497 if (qdisc->q.qlen < qdisc_dev(qdisc)->tx_queue_len) { 498 int band = prio2band[skb->priority & TC_PRIO_MAX]; 499 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 500 struct qdisc_skb_head *list = band2list(priv, band); 501 502 priv->bitmap |= (1 << band); 503 qdisc->q.qlen++; 504 return __qdisc_enqueue_tail(skb, qdisc, list); 505 } 506 507 return qdisc_drop(skb, qdisc, to_free); 508 } 509 510 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) 511 { 512 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 513 int band = bitmap2band[priv->bitmap]; 514 515 if (likely(band >= 0)) { 516 struct qdisc_skb_head *qh = band2list(priv, band); 517 struct sk_buff *skb = __qdisc_dequeue_head(qh); 518 519 if (likely(skb != NULL)) { 520 qdisc_qstats_backlog_dec(qdisc, skb); 521 qdisc_bstats_update(qdisc, skb); 522 } 523 524 qdisc->q.qlen--; 525 if (qh->qlen == 0) 526 priv->bitmap &= ~(1 << band); 527 528 return skb; 529 } 530 531 return NULL; 532 } 533 534 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc) 535 { 536 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 537 int band = bitmap2band[priv->bitmap]; 538 539 if (band >= 0) { 540 struct qdisc_skb_head *qh = band2list(priv, band); 541 542 return qh->head; 543 } 544 545 return NULL; 546 } 547 548 static void pfifo_fast_reset(struct Qdisc *qdisc) 549 { 550 int prio; 551 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 552 553 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) 554 __qdisc_reset_queue(band2list(priv, prio)); 555 556 priv->bitmap = 0; 557 qdisc->qstats.backlog = 0; 558 qdisc->q.qlen = 0; 559 } 560 561 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb) 562 { 563 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS }; 564 565 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1); 566 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 567 goto nla_put_failure; 568 return skb->len; 569 570 nla_put_failure: 571 return -1; 572 } 573 574 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt) 575 { 576 int prio; 577 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 578 579 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) 580 qdisc_skb_head_init(band2list(priv, prio)); 581 582 /* Can by-pass the queue discipline */ 583 qdisc->flags |= TCQ_F_CAN_BYPASS; 584 return 0; 585 } 586 587 struct Qdisc_ops pfifo_fast_ops __read_mostly = { 588 .id = "pfifo_fast", 589 .priv_size = sizeof(struct pfifo_fast_priv), 590 .enqueue = pfifo_fast_enqueue, 591 .dequeue = pfifo_fast_dequeue, 592 .peek = pfifo_fast_peek, 593 .init = pfifo_fast_init, 594 .reset = pfifo_fast_reset, 595 .dump = pfifo_fast_dump, 596 .owner = THIS_MODULE, 597 }; 598 EXPORT_SYMBOL(pfifo_fast_ops); 599 600 static struct lock_class_key qdisc_tx_busylock; 601 static struct lock_class_key qdisc_running_key; 602 603 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue, 604 const struct Qdisc_ops *ops) 605 { 606 void *p; 607 struct Qdisc *sch; 608 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size; 609 int err = -ENOBUFS; 610 struct net_device *dev; 611 612 if (!dev_queue) { 613 err = -EINVAL; 614 goto errout; 615 } 616 617 dev = dev_queue->dev; 618 p = kzalloc_node(size, GFP_KERNEL, 619 netdev_queue_numa_node_read(dev_queue)); 620 621 if (!p) 622 goto errout; 623 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p); 624 /* if we got non aligned memory, ask more and do alignment ourself */ 625 if (sch != p) { 626 kfree(p); 627 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL, 628 netdev_queue_numa_node_read(dev_queue)); 629 if (!p) 630 goto errout; 631 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p); 632 sch->padded = (char *) sch - (char *) p; 633 } 634 qdisc_skb_head_init(&sch->q); 635 spin_lock_init(&sch->q.lock); 636 637 spin_lock_init(&sch->busylock); 638 lockdep_set_class(&sch->busylock, 639 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock); 640 641 seqcount_init(&sch->running); 642 lockdep_set_class(&sch->running, 643 dev->qdisc_running_key ?: &qdisc_running_key); 644 645 sch->ops = ops; 646 sch->enqueue = ops->enqueue; 647 sch->dequeue = ops->dequeue; 648 sch->dev_queue = dev_queue; 649 dev_hold(dev); 650 refcount_set(&sch->refcnt, 1); 651 652 return sch; 653 errout: 654 return ERR_PTR(err); 655 } 656 657 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue, 658 const struct Qdisc_ops *ops, 659 unsigned int parentid) 660 { 661 struct Qdisc *sch; 662 663 if (!try_module_get(ops->owner)) 664 return NULL; 665 666 sch = qdisc_alloc(dev_queue, ops); 667 if (IS_ERR(sch)) { 668 module_put(ops->owner); 669 return NULL; 670 } 671 sch->parent = parentid; 672 673 if (!ops->init || ops->init(sch, NULL) == 0) 674 return sch; 675 676 qdisc_destroy(sch); 677 return NULL; 678 } 679 EXPORT_SYMBOL(qdisc_create_dflt); 680 681 /* Under qdisc_lock(qdisc) and BH! */ 682 683 void qdisc_reset(struct Qdisc *qdisc) 684 { 685 const struct Qdisc_ops *ops = qdisc->ops; 686 687 if (ops->reset) 688 ops->reset(qdisc); 689 690 kfree_skb(qdisc->skb_bad_txq); 691 qdisc->skb_bad_txq = NULL; 692 693 if (qdisc->gso_skb) { 694 kfree_skb_list(qdisc->gso_skb); 695 qdisc->gso_skb = NULL; 696 } 697 qdisc->q.qlen = 0; 698 qdisc->qstats.backlog = 0; 699 } 700 EXPORT_SYMBOL(qdisc_reset); 701 702 static void qdisc_free(struct Qdisc *qdisc) 703 { 704 if (qdisc_is_percpu_stats(qdisc)) { 705 free_percpu(qdisc->cpu_bstats); 706 free_percpu(qdisc->cpu_qstats); 707 } 708 709 kfree((char *) qdisc - qdisc->padded); 710 } 711 712 void qdisc_destroy(struct Qdisc *qdisc) 713 { 714 const struct Qdisc_ops *ops = qdisc->ops; 715 716 if (qdisc->flags & TCQ_F_BUILTIN || 717 !refcount_dec_and_test(&qdisc->refcnt)) 718 return; 719 720 #ifdef CONFIG_NET_SCHED 721 qdisc_hash_del(qdisc); 722 723 qdisc_put_stab(rtnl_dereference(qdisc->stab)); 724 #endif 725 gen_kill_estimator(&qdisc->rate_est); 726 if (ops->reset) 727 ops->reset(qdisc); 728 if (ops->destroy) 729 ops->destroy(qdisc); 730 731 module_put(ops->owner); 732 dev_put(qdisc_dev(qdisc)); 733 734 kfree_skb_list(qdisc->gso_skb); 735 kfree_skb(qdisc->skb_bad_txq); 736 qdisc_free(qdisc); 737 } 738 EXPORT_SYMBOL(qdisc_destroy); 739 740 /* Attach toplevel qdisc to device queue. */ 741 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue, 742 struct Qdisc *qdisc) 743 { 744 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping; 745 spinlock_t *root_lock; 746 747 root_lock = qdisc_lock(oqdisc); 748 spin_lock_bh(root_lock); 749 750 /* Prune old scheduler */ 751 if (oqdisc && refcount_read(&oqdisc->refcnt) <= 1) 752 qdisc_reset(oqdisc); 753 754 /* ... and graft new one */ 755 if (qdisc == NULL) 756 qdisc = &noop_qdisc; 757 dev_queue->qdisc_sleeping = qdisc; 758 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc); 759 760 spin_unlock_bh(root_lock); 761 762 return oqdisc; 763 } 764 EXPORT_SYMBOL(dev_graft_qdisc); 765 766 static void attach_one_default_qdisc(struct net_device *dev, 767 struct netdev_queue *dev_queue, 768 void *_unused) 769 { 770 struct Qdisc *qdisc; 771 const struct Qdisc_ops *ops = default_qdisc_ops; 772 773 if (dev->priv_flags & IFF_NO_QUEUE) 774 ops = &noqueue_qdisc_ops; 775 776 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT); 777 if (!qdisc) { 778 netdev_info(dev, "activation failed\n"); 779 return; 780 } 781 if (!netif_is_multiqueue(dev)) 782 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 783 dev_queue->qdisc_sleeping = qdisc; 784 } 785 786 static void attach_default_qdiscs(struct net_device *dev) 787 { 788 struct netdev_queue *txq; 789 struct Qdisc *qdisc; 790 791 txq = netdev_get_tx_queue(dev, 0); 792 793 if (!netif_is_multiqueue(dev) || 794 dev->priv_flags & IFF_NO_QUEUE) { 795 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL); 796 dev->qdisc = txq->qdisc_sleeping; 797 qdisc_refcount_inc(dev->qdisc); 798 } else { 799 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT); 800 if (qdisc) { 801 dev->qdisc = qdisc; 802 qdisc->ops->attach(qdisc); 803 } 804 } 805 #ifdef CONFIG_NET_SCHED 806 if (dev->qdisc != &noop_qdisc) 807 qdisc_hash_add(dev->qdisc, false); 808 #endif 809 } 810 811 static void transition_one_qdisc(struct net_device *dev, 812 struct netdev_queue *dev_queue, 813 void *_need_watchdog) 814 { 815 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping; 816 int *need_watchdog_p = _need_watchdog; 817 818 if (!(new_qdisc->flags & TCQ_F_BUILTIN)) 819 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state); 820 821 rcu_assign_pointer(dev_queue->qdisc, new_qdisc); 822 if (need_watchdog_p) { 823 dev_queue->trans_start = 0; 824 *need_watchdog_p = 1; 825 } 826 } 827 828 void dev_activate(struct net_device *dev) 829 { 830 int need_watchdog; 831 832 /* No queueing discipline is attached to device; 833 * create default one for devices, which need queueing 834 * and noqueue_qdisc for virtual interfaces 835 */ 836 837 if (dev->qdisc == &noop_qdisc) 838 attach_default_qdiscs(dev); 839 840 if (!netif_carrier_ok(dev)) 841 /* Delay activation until next carrier-on event */ 842 return; 843 844 need_watchdog = 0; 845 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog); 846 if (dev_ingress_queue(dev)) 847 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL); 848 849 if (need_watchdog) { 850 netif_trans_update(dev); 851 dev_watchdog_up(dev); 852 } 853 } 854 EXPORT_SYMBOL(dev_activate); 855 856 static void dev_deactivate_queue(struct net_device *dev, 857 struct netdev_queue *dev_queue, 858 void *_qdisc_default) 859 { 860 struct Qdisc *qdisc_default = _qdisc_default; 861 struct Qdisc *qdisc; 862 863 qdisc = rtnl_dereference(dev_queue->qdisc); 864 if (qdisc) { 865 spin_lock_bh(qdisc_lock(qdisc)); 866 867 if (!(qdisc->flags & TCQ_F_BUILTIN)) 868 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state); 869 870 rcu_assign_pointer(dev_queue->qdisc, qdisc_default); 871 qdisc_reset(qdisc); 872 873 spin_unlock_bh(qdisc_lock(qdisc)); 874 } 875 } 876 877 static bool some_qdisc_is_busy(struct net_device *dev) 878 { 879 unsigned int i; 880 881 for (i = 0; i < dev->num_tx_queues; i++) { 882 struct netdev_queue *dev_queue; 883 spinlock_t *root_lock; 884 struct Qdisc *q; 885 int val; 886 887 dev_queue = netdev_get_tx_queue(dev, i); 888 q = dev_queue->qdisc_sleeping; 889 890 if (q->flags & TCQ_F_NOLOCK) { 891 val = test_bit(__QDISC_STATE_SCHED, &q->state); 892 } else { 893 root_lock = qdisc_lock(q); 894 spin_lock_bh(root_lock); 895 896 val = (qdisc_is_running(q) || 897 test_bit(__QDISC_STATE_SCHED, &q->state)); 898 899 spin_unlock_bh(root_lock); 900 } 901 902 if (val) 903 return true; 904 } 905 return false; 906 } 907 908 /** 909 * dev_deactivate_many - deactivate transmissions on several devices 910 * @head: list of devices to deactivate 911 * 912 * This function returns only when all outstanding transmissions 913 * have completed, unless all devices are in dismantle phase. 914 */ 915 void dev_deactivate_many(struct list_head *head) 916 { 917 struct net_device *dev; 918 bool sync_needed = false; 919 920 list_for_each_entry(dev, head, close_list) { 921 netdev_for_each_tx_queue(dev, dev_deactivate_queue, 922 &noop_qdisc); 923 if (dev_ingress_queue(dev)) 924 dev_deactivate_queue(dev, dev_ingress_queue(dev), 925 &noop_qdisc); 926 927 dev_watchdog_down(dev); 928 sync_needed |= !dev->dismantle; 929 } 930 931 /* Wait for outstanding qdisc-less dev_queue_xmit calls. 932 * This is avoided if all devices are in dismantle phase : 933 * Caller will call synchronize_net() for us 934 */ 935 if (sync_needed) 936 synchronize_net(); 937 938 /* Wait for outstanding qdisc_run calls. */ 939 list_for_each_entry(dev, head, close_list) 940 while (some_qdisc_is_busy(dev)) 941 yield(); 942 } 943 944 void dev_deactivate(struct net_device *dev) 945 { 946 LIST_HEAD(single); 947 948 list_add(&dev->close_list, &single); 949 dev_deactivate_many(&single); 950 list_del(&single); 951 } 952 EXPORT_SYMBOL(dev_deactivate); 953 954 static void dev_init_scheduler_queue(struct net_device *dev, 955 struct netdev_queue *dev_queue, 956 void *_qdisc) 957 { 958 struct Qdisc *qdisc = _qdisc; 959 960 rcu_assign_pointer(dev_queue->qdisc, qdisc); 961 dev_queue->qdisc_sleeping = qdisc; 962 } 963 964 void dev_init_scheduler(struct net_device *dev) 965 { 966 dev->qdisc = &noop_qdisc; 967 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc); 968 if (dev_ingress_queue(dev)) 969 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc); 970 971 timer_setup(&dev->watchdog_timer, dev_watchdog, 0); 972 } 973 974 static void shutdown_scheduler_queue(struct net_device *dev, 975 struct netdev_queue *dev_queue, 976 void *_qdisc_default) 977 { 978 struct Qdisc *qdisc = dev_queue->qdisc_sleeping; 979 struct Qdisc *qdisc_default = _qdisc_default; 980 981 if (qdisc) { 982 rcu_assign_pointer(dev_queue->qdisc, qdisc_default); 983 dev_queue->qdisc_sleeping = qdisc_default; 984 985 qdisc_destroy(qdisc); 986 } 987 } 988 989 void dev_shutdown(struct net_device *dev) 990 { 991 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc); 992 if (dev_ingress_queue(dev)) 993 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc); 994 qdisc_destroy(dev->qdisc); 995 dev->qdisc = &noop_qdisc; 996 997 WARN_ON(timer_pending(&dev->watchdog_timer)); 998 } 999 1000 void psched_ratecfg_precompute(struct psched_ratecfg *r, 1001 const struct tc_ratespec *conf, 1002 u64 rate64) 1003 { 1004 memset(r, 0, sizeof(*r)); 1005 r->overhead = conf->overhead; 1006 r->rate_bytes_ps = max_t(u64, conf->rate, rate64); 1007 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK); 1008 r->mult = 1; 1009 /* 1010 * The deal here is to replace a divide by a reciprocal one 1011 * in fast path (a reciprocal divide is a multiply and a shift) 1012 * 1013 * Normal formula would be : 1014 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps 1015 * 1016 * We compute mult/shift to use instead : 1017 * time_in_ns = (len * mult) >> shift; 1018 * 1019 * We try to get the highest possible mult value for accuracy, 1020 * but have to make sure no overflows will ever happen. 1021 */ 1022 if (r->rate_bytes_ps > 0) { 1023 u64 factor = NSEC_PER_SEC; 1024 1025 for (;;) { 1026 r->mult = div64_u64(factor, r->rate_bytes_ps); 1027 if (r->mult & (1U << 31) || factor & (1ULL << 63)) 1028 break; 1029 factor <<= 1; 1030 r->shift++; 1031 } 1032 } 1033 } 1034 EXPORT_SYMBOL(psched_ratecfg_precompute); 1035 1036 static void mini_qdisc_rcu_func(struct rcu_head *head) 1037 { 1038 } 1039 1040 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp, 1041 struct tcf_proto *tp_head) 1042 { 1043 struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq); 1044 struct mini_Qdisc *miniq; 1045 1046 if (!tp_head) { 1047 RCU_INIT_POINTER(*miniqp->p_miniq, NULL); 1048 return; 1049 } 1050 1051 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ? 1052 &miniqp->miniq1 : &miniqp->miniq2; 1053 1054 /* We need to make sure that readers won't see the miniq 1055 * we are about to modify. So wait until previous call_rcu_bh callback 1056 * is done. 1057 */ 1058 rcu_barrier_bh(); 1059 miniq->filter_list = tp_head; 1060 rcu_assign_pointer(*miniqp->p_miniq, miniq); 1061 1062 if (miniq_old) 1063 /* This is counterpart of the rcu barrier above. We need to 1064 * block potential new user of miniq_old until all readers 1065 * are not seeing it. 1066 */ 1067 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func); 1068 } 1069 EXPORT_SYMBOL(mini_qdisc_pair_swap); 1070 1071 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc, 1072 struct mini_Qdisc __rcu **p_miniq) 1073 { 1074 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats; 1075 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats; 1076 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats; 1077 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats; 1078 miniqp->p_miniq = p_miniq; 1079 } 1080 EXPORT_SYMBOL(mini_qdisc_pair_init); 1081