1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * net/sched/sch_generic.c Generic packet scheduler routines. 4 * 5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> 6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601 7 * - Ingress support 8 */ 9 10 #include <linux/bitops.h> 11 #include <linux/module.h> 12 #include <linux/types.h> 13 #include <linux/kernel.h> 14 #include <linux/sched.h> 15 #include <linux/string.h> 16 #include <linux/errno.h> 17 #include <linux/netdevice.h> 18 #include <linux/skbuff.h> 19 #include <linux/rtnetlink.h> 20 #include <linux/init.h> 21 #include <linux/rcupdate.h> 22 #include <linux/list.h> 23 #include <linux/slab.h> 24 #include <linux/if_vlan.h> 25 #include <linux/skb_array.h> 26 #include <linux/if_macvlan.h> 27 #include <linux/bpf.h> 28 #include <net/sch_generic.h> 29 #include <net/pkt_sched.h> 30 #include <net/dst.h> 31 #include <net/hotdata.h> 32 #include <trace/events/qdisc.h> 33 #include <trace/events/net.h> 34 #include <net/xfrm.h> 35 36 /* Qdisc to use by default */ 37 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops; 38 EXPORT_SYMBOL(default_qdisc_ops); 39 40 static void qdisc_maybe_clear_missed(struct Qdisc *q, 41 const struct netdev_queue *txq) 42 { 43 clear_bit(__QDISC_STATE_MISSED, &q->state); 44 45 /* Make sure the below netif_xmit_frozen_or_stopped() 46 * checking happens after clearing STATE_MISSED. 47 */ 48 smp_mb__after_atomic(); 49 50 /* Checking netif_xmit_frozen_or_stopped() again to 51 * make sure STATE_MISSED is set if the STATE_MISSED 52 * set by netif_tx_wake_queue()'s rescheduling of 53 * net_tx_action() is cleared by the above clear_bit(). 54 */ 55 if (!netif_xmit_frozen_or_stopped(txq)) 56 set_bit(__QDISC_STATE_MISSED, &q->state); 57 else 58 set_bit(__QDISC_STATE_DRAINING, &q->state); 59 } 60 61 /* Main transmission queue. */ 62 63 /* Modifications to data participating in scheduling must be protected with 64 * qdisc_lock(qdisc) spinlock. 65 * 66 * The idea is the following: 67 * - enqueue, dequeue are serialized via qdisc root lock 68 * - ingress filtering is also serialized via qdisc root lock 69 * - updates to tree and tree walking are only done under the rtnl mutex. 70 */ 71 72 #define SKB_XOFF_MAGIC ((struct sk_buff *)1UL) 73 74 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q) 75 { 76 const struct netdev_queue *txq = q->dev_queue; 77 spinlock_t *lock = NULL; 78 struct sk_buff *skb; 79 80 if (q->flags & TCQ_F_NOLOCK) { 81 lock = qdisc_lock(q); 82 spin_lock(lock); 83 } 84 85 skb = skb_peek(&q->skb_bad_txq); 86 if (skb) { 87 /* check the reason of requeuing without tx lock first */ 88 txq = skb_get_tx_queue(txq->dev, skb); 89 if (!netif_xmit_frozen_or_stopped(txq)) { 90 skb = __skb_dequeue(&q->skb_bad_txq); 91 if (qdisc_is_percpu_stats(q)) { 92 qdisc_qstats_cpu_backlog_dec(q, skb); 93 qdisc_qstats_cpu_qlen_dec(q); 94 } else { 95 qdisc_qstats_backlog_dec(q, skb); 96 q->q.qlen--; 97 } 98 } else { 99 skb = SKB_XOFF_MAGIC; 100 qdisc_maybe_clear_missed(q, txq); 101 } 102 } 103 104 if (lock) 105 spin_unlock(lock); 106 107 return skb; 108 } 109 110 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q) 111 { 112 struct sk_buff *skb = skb_peek(&q->skb_bad_txq); 113 114 if (unlikely(skb)) 115 skb = __skb_dequeue_bad_txq(q); 116 117 return skb; 118 } 119 120 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q, 121 struct sk_buff *skb) 122 { 123 spinlock_t *lock = NULL; 124 125 if (q->flags & TCQ_F_NOLOCK) { 126 lock = qdisc_lock(q); 127 spin_lock(lock); 128 } 129 130 __skb_queue_tail(&q->skb_bad_txq, skb); 131 132 if (qdisc_is_percpu_stats(q)) { 133 qdisc_qstats_cpu_backlog_inc(q, skb); 134 qdisc_qstats_cpu_qlen_inc(q); 135 } else { 136 qdisc_qstats_backlog_inc(q, skb); 137 q->q.qlen++; 138 } 139 140 if (lock) 141 spin_unlock(lock); 142 } 143 144 static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q) 145 { 146 spinlock_t *lock = NULL; 147 148 if (q->flags & TCQ_F_NOLOCK) { 149 lock = qdisc_lock(q); 150 spin_lock(lock); 151 } 152 153 while (skb) { 154 struct sk_buff *next = skb->next; 155 156 __skb_queue_tail(&q->gso_skb, skb); 157 158 /* it's still part of the queue */ 159 if (qdisc_is_percpu_stats(q)) { 160 qdisc_qstats_cpu_requeues_inc(q); 161 qdisc_qstats_cpu_backlog_inc(q, skb); 162 qdisc_qstats_cpu_qlen_inc(q); 163 } else { 164 q->qstats.requeues++; 165 qdisc_qstats_backlog_inc(q, skb); 166 q->q.qlen++; 167 } 168 169 skb = next; 170 } 171 172 if (lock) { 173 spin_unlock(lock); 174 set_bit(__QDISC_STATE_MISSED, &q->state); 175 } else { 176 __netif_schedule(q); 177 } 178 } 179 180 static void try_bulk_dequeue_skb(struct Qdisc *q, 181 struct sk_buff *skb, 182 const struct netdev_queue *txq, 183 int *packets, int budget) 184 { 185 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len; 186 int cnt = 0; 187 188 while (bytelimit > 0) { 189 struct sk_buff *nskb = q->dequeue(q); 190 191 if (!nskb) 192 break; 193 194 bytelimit -= nskb->len; /* covers GSO len */ 195 skb->next = nskb; 196 skb = nskb; 197 if (++cnt >= budget) 198 break; 199 } 200 (*packets) += cnt; 201 skb_mark_not_on_list(skb); 202 } 203 204 /* This variant of try_bulk_dequeue_skb() makes sure 205 * all skbs in the chain are for the same txq 206 */ 207 static void try_bulk_dequeue_skb_slow(struct Qdisc *q, 208 struct sk_buff *skb, 209 int *packets) 210 { 211 int mapping = skb_get_queue_mapping(skb); 212 struct sk_buff *nskb; 213 int cnt = 0; 214 215 do { 216 nskb = q->dequeue(q); 217 if (!nskb) 218 break; 219 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) { 220 qdisc_enqueue_skb_bad_txq(q, nskb); 221 break; 222 } 223 skb->next = nskb; 224 skb = nskb; 225 } while (++cnt < 8); 226 (*packets) += cnt; 227 skb_mark_not_on_list(skb); 228 } 229 230 /* Note that dequeue_skb can possibly return a SKB list (via skb->next). 231 * A requeued skb (via q->gso_skb) can also be a SKB list. 232 */ 233 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, 234 int *packets, int budget) 235 { 236 const struct netdev_queue *txq = q->dev_queue; 237 struct sk_buff *skb = NULL; 238 239 *packets = 1; 240 if (unlikely(!skb_queue_empty(&q->gso_skb))) { 241 spinlock_t *lock = NULL; 242 243 if (q->flags & TCQ_F_NOLOCK) { 244 lock = qdisc_lock(q); 245 spin_lock(lock); 246 } 247 248 skb = skb_peek(&q->gso_skb); 249 250 /* skb may be null if another cpu pulls gso_skb off in between 251 * empty check and lock. 252 */ 253 if (!skb) { 254 if (lock) 255 spin_unlock(lock); 256 goto validate; 257 } 258 259 /* skb in gso_skb were already validated */ 260 *validate = false; 261 if (xfrm_offload(skb)) 262 *validate = true; 263 /* check the reason of requeuing without tx lock first */ 264 txq = skb_get_tx_queue(txq->dev, skb); 265 if (!netif_xmit_frozen_or_stopped(txq)) { 266 skb = __skb_dequeue(&q->gso_skb); 267 if (qdisc_is_percpu_stats(q)) { 268 qdisc_qstats_cpu_backlog_dec(q, skb); 269 qdisc_qstats_cpu_qlen_dec(q); 270 } else { 271 qdisc_qstats_backlog_dec(q, skb); 272 q->q.qlen--; 273 } 274 } else { 275 skb = NULL; 276 qdisc_maybe_clear_missed(q, txq); 277 } 278 if (lock) 279 spin_unlock(lock); 280 goto trace; 281 } 282 validate: 283 *validate = true; 284 285 if ((q->flags & TCQ_F_ONETXQUEUE) && 286 netif_xmit_frozen_or_stopped(txq)) { 287 qdisc_maybe_clear_missed(q, txq); 288 return skb; 289 } 290 291 skb = qdisc_dequeue_skb_bad_txq(q); 292 if (unlikely(skb)) { 293 if (skb == SKB_XOFF_MAGIC) 294 return NULL; 295 goto bulk; 296 } 297 skb = q->dequeue(q); 298 if (skb) { 299 bulk: 300 if (qdisc_may_bulk(q)) 301 try_bulk_dequeue_skb(q, skb, txq, packets, budget); 302 else 303 try_bulk_dequeue_skb_slow(q, skb, packets); 304 } 305 trace: 306 trace_qdisc_dequeue(q, txq, *packets, skb); 307 return skb; 308 } 309 310 /* 311 * Transmit possibly several skbs, and handle the return status as 312 * required. Owning qdisc running bit guarantees that only one CPU 313 * can execute this function. 314 * 315 * Returns to the caller: 316 * false - hardware queue frozen backoff 317 * true - feel free to send more pkts 318 */ 319 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q, 320 struct net_device *dev, struct netdev_queue *txq, 321 spinlock_t *root_lock, bool validate) 322 { 323 int ret = NETDEV_TX_BUSY; 324 bool again = false; 325 326 /* And release qdisc */ 327 if (root_lock) 328 spin_unlock(root_lock); 329 330 /* Note that we validate skb (GSO, checksum, ...) outside of locks */ 331 if (validate) 332 skb = validate_xmit_skb_list(skb, dev, &again); 333 334 #ifdef CONFIG_XFRM_OFFLOAD 335 if (unlikely(again)) { 336 if (root_lock) 337 spin_lock(root_lock); 338 339 dev_requeue_skb(skb, q); 340 return false; 341 } 342 #endif 343 344 if (likely(skb)) { 345 HARD_TX_LOCK(dev, txq, smp_processor_id()); 346 if (!netif_xmit_frozen_or_stopped(txq)) 347 skb = dev_hard_start_xmit(skb, dev, txq, &ret); 348 else 349 qdisc_maybe_clear_missed(q, txq); 350 351 HARD_TX_UNLOCK(dev, txq); 352 } else { 353 if (root_lock) 354 spin_lock(root_lock); 355 return true; 356 } 357 358 if (root_lock) 359 spin_lock(root_lock); 360 361 if (!dev_xmit_complete(ret)) { 362 /* Driver returned NETDEV_TX_BUSY - requeue skb */ 363 if (unlikely(ret != NETDEV_TX_BUSY)) 364 net_warn_ratelimited("BUG %s code %d qlen %d\n", 365 dev->name, ret, q->q.qlen); 366 367 dev_requeue_skb(skb, q); 368 return false; 369 } 370 371 return true; 372 } 373 374 /* 375 * NOTE: Called under qdisc_lock(q) with locally disabled BH. 376 * 377 * running seqcount guarantees only one CPU can process 378 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for 379 * this queue. 380 * 381 * netif_tx_lock serializes accesses to device driver. 382 * 383 * qdisc_lock(q) and netif_tx_lock are mutually exclusive, 384 * if one is grabbed, another must be free. 385 * 386 * Note, that this procedure can be called by a watchdog timer 387 * 388 * Returns to the caller: 389 * 0 - queue is empty or throttled. 390 * >0 - queue is not empty. 391 * 392 */ 393 static inline bool qdisc_restart(struct Qdisc *q, int *packets, int budget) 394 { 395 spinlock_t *root_lock = NULL; 396 struct netdev_queue *txq; 397 struct net_device *dev; 398 struct sk_buff *skb; 399 bool validate; 400 401 /* Dequeue packet */ 402 skb = dequeue_skb(q, &validate, packets, budget); 403 if (unlikely(!skb)) 404 return false; 405 406 if (!(q->flags & TCQ_F_NOLOCK)) 407 root_lock = qdisc_lock(q); 408 409 dev = qdisc_dev(q); 410 txq = skb_get_tx_queue(dev, skb); 411 412 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate); 413 } 414 415 void __qdisc_run(struct Qdisc *q) 416 { 417 int quota = READ_ONCE(net_hotdata.dev_tx_weight); 418 int packets; 419 420 while (qdisc_restart(q, &packets, quota)) { 421 quota -= packets; 422 if (quota <= 0) { 423 if (q->flags & TCQ_F_NOLOCK) 424 set_bit(__QDISC_STATE_MISSED, &q->state); 425 else 426 __netif_schedule(q); 427 428 break; 429 } 430 } 431 } 432 433 unsigned long dev_trans_start(struct net_device *dev) 434 { 435 unsigned long res = READ_ONCE(netdev_get_tx_queue(dev, 0)->trans_start); 436 unsigned long val; 437 unsigned int i; 438 439 for (i = 1; i < dev->num_tx_queues; i++) { 440 val = READ_ONCE(netdev_get_tx_queue(dev, i)->trans_start); 441 if (val && time_after(val, res)) 442 res = val; 443 } 444 445 return res; 446 } 447 EXPORT_SYMBOL(dev_trans_start); 448 449 static void netif_freeze_queues(struct net_device *dev) 450 { 451 unsigned int i; 452 int cpu; 453 454 cpu = smp_processor_id(); 455 for (i = 0; i < dev->num_tx_queues; i++) { 456 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 457 458 /* We are the only thread of execution doing a 459 * freeze, but we have to grab the _xmit_lock in 460 * order to synchronize with threads which are in 461 * the ->hard_start_xmit() handler and already 462 * checked the frozen bit. 463 */ 464 __netif_tx_lock(txq, cpu); 465 set_bit(__QUEUE_STATE_FROZEN, &txq->state); 466 __netif_tx_unlock(txq); 467 } 468 } 469 470 void netif_tx_lock(struct net_device *dev) 471 { 472 spin_lock(&dev->tx_global_lock); 473 netif_freeze_queues(dev); 474 } 475 EXPORT_SYMBOL(netif_tx_lock); 476 477 static void netif_unfreeze_queues(struct net_device *dev) 478 { 479 unsigned int i; 480 481 for (i = 0; i < dev->num_tx_queues; i++) { 482 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 483 484 /* No need to grab the _xmit_lock here. If the 485 * queue is not stopped for another reason, we 486 * force a schedule. 487 */ 488 clear_bit(__QUEUE_STATE_FROZEN, &txq->state); 489 netif_schedule_queue(txq); 490 } 491 } 492 493 void netif_tx_unlock(struct net_device *dev) 494 { 495 netif_unfreeze_queues(dev); 496 spin_unlock(&dev->tx_global_lock); 497 } 498 EXPORT_SYMBOL(netif_tx_unlock); 499 500 static void dev_watchdog(struct timer_list *t) 501 { 502 struct net_device *dev = timer_container_of(dev, t, watchdog_timer); 503 bool release = true; 504 505 spin_lock(&dev->tx_global_lock); 506 if (!qdisc_tx_is_noop(dev)) { 507 if (netif_device_present(dev) && 508 netif_running(dev) && 509 netif_carrier_ok(dev)) { 510 unsigned int timedout_ms = 0; 511 unsigned int i; 512 unsigned long trans_start; 513 unsigned long oldest_start = jiffies; 514 515 for (i = 0; i < dev->num_tx_queues; i++) { 516 struct netdev_queue *txq; 517 518 txq = netdev_get_tx_queue(dev, i); 519 if (!netif_xmit_stopped(txq)) 520 continue; 521 522 /* Paired with WRITE_ONCE() + smp_mb...() in 523 * netdev_tx_sent_queue() and netif_tx_stop_queue(). 524 */ 525 smp_mb(); 526 trans_start = READ_ONCE(txq->trans_start); 527 528 if (time_after(jiffies, trans_start + dev->watchdog_timeo)) { 529 timedout_ms = jiffies_to_msecs(jiffies - trans_start); 530 atomic_long_inc(&txq->trans_timeout); 531 break; 532 } 533 if (time_after(oldest_start, trans_start)) 534 oldest_start = trans_start; 535 } 536 537 if (unlikely(timedout_ms)) { 538 trace_net_dev_xmit_timeout(dev, i); 539 netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n", 540 raw_smp_processor_id(), 541 i, timedout_ms); 542 netif_freeze_queues(dev); 543 dev->netdev_ops->ndo_tx_timeout(dev, i); 544 netif_unfreeze_queues(dev); 545 } 546 if (!mod_timer(&dev->watchdog_timer, 547 round_jiffies(oldest_start + 548 dev->watchdog_timeo))) 549 release = false; 550 } 551 } 552 spin_unlock(&dev->tx_global_lock); 553 554 if (release) 555 netdev_put(dev, &dev->watchdog_dev_tracker); 556 } 557 558 void netdev_watchdog_up(struct net_device *dev) 559 { 560 if (!dev->netdev_ops->ndo_tx_timeout) 561 return; 562 if (dev->watchdog_timeo <= 0) 563 dev->watchdog_timeo = 5*HZ; 564 if (!mod_timer(&dev->watchdog_timer, 565 round_jiffies(jiffies + dev->watchdog_timeo))) 566 netdev_hold(dev, &dev->watchdog_dev_tracker, 567 GFP_ATOMIC); 568 } 569 EXPORT_SYMBOL_GPL(netdev_watchdog_up); 570 571 static void netdev_watchdog_down(struct net_device *dev) 572 { 573 netif_tx_lock_bh(dev); 574 if (timer_delete(&dev->watchdog_timer)) 575 netdev_put(dev, &dev->watchdog_dev_tracker); 576 netif_tx_unlock_bh(dev); 577 } 578 579 /** 580 * netif_carrier_on - set carrier 581 * @dev: network device 582 * 583 * Device has detected acquisition of carrier. 584 */ 585 void netif_carrier_on(struct net_device *dev) 586 { 587 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) { 588 if (dev->reg_state == NETREG_UNINITIALIZED) 589 return; 590 atomic_inc(&dev->carrier_up_count); 591 linkwatch_fire_event(dev); 592 if (netif_running(dev)) 593 netdev_watchdog_up(dev); 594 } 595 } 596 EXPORT_SYMBOL(netif_carrier_on); 597 598 /** 599 * netif_carrier_off - clear carrier 600 * @dev: network device 601 * 602 * Device has detected loss of carrier. 603 */ 604 void netif_carrier_off(struct net_device *dev) 605 { 606 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) { 607 if (dev->reg_state == NETREG_UNINITIALIZED) 608 return; 609 atomic_inc(&dev->carrier_down_count); 610 linkwatch_fire_event(dev); 611 } 612 } 613 EXPORT_SYMBOL(netif_carrier_off); 614 615 /** 616 * netif_carrier_event - report carrier state event 617 * @dev: network device 618 * 619 * Device has detected a carrier event but the carrier state wasn't changed. 620 * Use in drivers when querying carrier state asynchronously, to avoid missing 621 * events (link flaps) if link recovers before it's queried. 622 */ 623 void netif_carrier_event(struct net_device *dev) 624 { 625 if (dev->reg_state == NETREG_UNINITIALIZED) 626 return; 627 atomic_inc(&dev->carrier_up_count); 628 atomic_inc(&dev->carrier_down_count); 629 linkwatch_fire_event(dev); 630 } 631 EXPORT_SYMBOL_GPL(netif_carrier_event); 632 633 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces 634 under all circumstances. It is difficult to invent anything faster or 635 cheaper. 636 */ 637 638 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc, 639 struct sk_buff **to_free) 640 { 641 dev_core_stats_tx_dropped_inc(skb->dev); 642 __qdisc_drop(skb, to_free); 643 return NET_XMIT_CN; 644 } 645 646 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc) 647 { 648 return NULL; 649 } 650 651 struct Qdisc_ops noop_qdisc_ops __read_mostly = { 652 .id = "noop", 653 .priv_size = 0, 654 .enqueue = noop_enqueue, 655 .dequeue = noop_dequeue, 656 .peek = noop_dequeue, 657 .owner = THIS_MODULE, 658 }; 659 660 static struct netdev_queue noop_netdev_queue = { 661 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc), 662 RCU_POINTER_INITIALIZER(qdisc_sleeping, &noop_qdisc), 663 }; 664 665 struct Qdisc noop_qdisc = { 666 .enqueue = noop_enqueue, 667 .dequeue = noop_dequeue, 668 .flags = TCQ_F_BUILTIN, 669 .ops = &noop_qdisc_ops, 670 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock), 671 .dev_queue = &noop_netdev_queue, 672 .gso_skb = { 673 .next = (struct sk_buff *)&noop_qdisc.gso_skb, 674 .prev = (struct sk_buff *)&noop_qdisc.gso_skb, 675 .qlen = 0, 676 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock), 677 }, 678 .skb_bad_txq = { 679 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq, 680 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq, 681 .qlen = 0, 682 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock), 683 }, 684 }; 685 EXPORT_SYMBOL(noop_qdisc); 686 687 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt, 688 struct netlink_ext_ack *extack) 689 { 690 /* register_qdisc() assigns a default of noop_enqueue if unset, 691 * but __dev_queue_xmit() treats noqueue only as such 692 * if this is NULL - so clear it here. */ 693 qdisc->enqueue = NULL; 694 return 0; 695 } 696 697 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = { 698 .id = "noqueue", 699 .priv_size = 0, 700 .init = noqueue_init, 701 .enqueue = noop_enqueue, 702 .dequeue = noop_dequeue, 703 .peek = noop_dequeue, 704 .owner = THIS_MODULE, 705 }; 706 707 const u8 sch_default_prio2band[TC_PRIO_MAX + 1] = { 708 1, 2, 2, 2, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1 709 }; 710 EXPORT_SYMBOL(sch_default_prio2band); 711 712 /* 3-band FIFO queue: old style, but should be a bit faster than 713 generic prio+fifo combination. 714 */ 715 716 #define PFIFO_FAST_BANDS 3 717 718 /* 719 * Private data for a pfifo_fast scheduler containing: 720 * - rings for priority bands 721 */ 722 struct pfifo_fast_priv { 723 struct skb_array q[PFIFO_FAST_BANDS]; 724 }; 725 726 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv, 727 int band) 728 { 729 return &priv->q[band]; 730 } 731 732 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc, 733 struct sk_buff **to_free) 734 { 735 int band = sch_default_prio2band[skb->priority & TC_PRIO_MAX]; 736 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 737 struct skb_array *q = band2list(priv, band); 738 unsigned int pkt_len = qdisc_pkt_len(skb); 739 int err; 740 741 err = skb_array_produce(q, skb); 742 743 if (unlikely(err)) { 744 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_OVERLIMIT); 745 746 if (qdisc_is_percpu_stats(qdisc)) 747 return qdisc_drop_cpu(skb, qdisc, to_free); 748 else 749 return qdisc_drop(skb, qdisc, to_free); 750 } 751 752 qdisc_update_stats_at_enqueue(qdisc, pkt_len); 753 return NET_XMIT_SUCCESS; 754 } 755 756 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) 757 { 758 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 759 struct sk_buff *skb = NULL; 760 bool need_retry = true; 761 int band; 762 763 retry: 764 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) { 765 struct skb_array *q = band2list(priv, band); 766 767 if (__skb_array_empty(q)) 768 continue; 769 770 skb = __skb_array_consume(q); 771 } 772 if (likely(skb)) { 773 qdisc_update_stats_at_dequeue(qdisc, skb); 774 } else if (need_retry && 775 READ_ONCE(qdisc->state) & QDISC_STATE_NON_EMPTY) { 776 /* Delay clearing the STATE_MISSED here to reduce 777 * the overhead of the second spin_trylock() in 778 * qdisc_run_begin() and __netif_schedule() calling 779 * in qdisc_run_end(). 780 */ 781 clear_bit(__QDISC_STATE_MISSED, &qdisc->state); 782 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state); 783 784 /* Make sure dequeuing happens after clearing 785 * STATE_MISSED. 786 */ 787 smp_mb__after_atomic(); 788 789 need_retry = false; 790 791 goto retry; 792 } 793 794 return skb; 795 } 796 797 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc) 798 { 799 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 800 struct sk_buff *skb = NULL; 801 int band; 802 803 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) { 804 struct skb_array *q = band2list(priv, band); 805 806 skb = __skb_array_peek(q); 807 } 808 809 return skb; 810 } 811 812 static void pfifo_fast_reset(struct Qdisc *qdisc) 813 { 814 int i, band; 815 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 816 817 for (band = 0; band < PFIFO_FAST_BANDS; band++) { 818 struct skb_array *q = band2list(priv, band); 819 struct sk_buff *skb; 820 821 /* NULL ring is possible if destroy path is due to a failed 822 * skb_array_init() in pfifo_fast_init() case. 823 */ 824 if (!q->ring.queue) 825 continue; 826 827 while ((skb = __skb_array_consume(q)) != NULL) 828 kfree_skb(skb); 829 } 830 831 if (qdisc_is_percpu_stats(qdisc)) { 832 for_each_possible_cpu(i) { 833 struct gnet_stats_queue *q; 834 835 q = per_cpu_ptr(qdisc->cpu_qstats, i); 836 q->backlog = 0; 837 q->qlen = 0; 838 } 839 } 840 } 841 842 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb) 843 { 844 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS }; 845 846 memcpy(&opt.priomap, sch_default_prio2band, TC_PRIO_MAX + 1); 847 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 848 goto nla_put_failure; 849 return skb->len; 850 851 nla_put_failure: 852 return -1; 853 } 854 855 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt, 856 struct netlink_ext_ack *extack) 857 { 858 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len; 859 struct pfifo_fast_priv *priv = qdisc_priv(qdisc); 860 int prio; 861 862 /* guard against zero length rings */ 863 if (!qlen) 864 return -EINVAL; 865 866 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) { 867 struct skb_array *q = band2list(priv, prio); 868 int err; 869 870 err = skb_array_init(q, qlen, GFP_KERNEL); 871 if (err) 872 return -ENOMEM; 873 } 874 875 /* Can by-pass the queue discipline */ 876 qdisc->flags |= TCQ_F_CAN_BYPASS; 877 return 0; 878 } 879 880 static void pfifo_fast_destroy(struct Qdisc *sch) 881 { 882 struct pfifo_fast_priv *priv = qdisc_priv(sch); 883 int prio; 884 885 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) { 886 struct skb_array *q = band2list(priv, prio); 887 888 /* NULL ring is possible if destroy path is due to a failed 889 * skb_array_init() in pfifo_fast_init() case. 890 */ 891 if (!q->ring.queue) 892 continue; 893 /* Destroy ring but no need to kfree_skb because a call to 894 * pfifo_fast_reset() has already done that work. 895 */ 896 ptr_ring_cleanup(&q->ring, NULL); 897 } 898 } 899 900 static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch, 901 unsigned int new_len) 902 { 903 struct pfifo_fast_priv *priv = qdisc_priv(sch); 904 struct skb_array *bands[PFIFO_FAST_BANDS]; 905 int prio; 906 907 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) { 908 struct skb_array *q = band2list(priv, prio); 909 910 bands[prio] = q; 911 } 912 913 return skb_array_resize_multiple_bh(bands, PFIFO_FAST_BANDS, new_len, 914 GFP_KERNEL); 915 } 916 917 struct Qdisc_ops pfifo_fast_ops __read_mostly = { 918 .id = "pfifo_fast", 919 .priv_size = sizeof(struct pfifo_fast_priv), 920 .enqueue = pfifo_fast_enqueue, 921 .dequeue = pfifo_fast_dequeue, 922 .peek = pfifo_fast_peek, 923 .init = pfifo_fast_init, 924 .destroy = pfifo_fast_destroy, 925 .reset = pfifo_fast_reset, 926 .dump = pfifo_fast_dump, 927 .change_tx_queue_len = pfifo_fast_change_tx_queue_len, 928 .owner = THIS_MODULE, 929 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS, 930 }; 931 EXPORT_SYMBOL(pfifo_fast_ops); 932 933 static struct lock_class_key qdisc_tx_busylock; 934 935 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue, 936 const struct Qdisc_ops *ops, 937 struct netlink_ext_ack *extack) 938 { 939 struct Qdisc *sch; 940 unsigned int size = sizeof(*sch) + ops->priv_size; 941 int err = -ENOBUFS; 942 struct net_device *dev; 943 944 if (!dev_queue) { 945 NL_SET_ERR_MSG(extack, "No device queue given"); 946 err = -EINVAL; 947 goto errout; 948 } 949 950 dev = dev_queue->dev; 951 sch = kzalloc_node(size, GFP_KERNEL, netdev_queue_numa_node_read(dev_queue)); 952 953 if (!sch) 954 goto errout; 955 __skb_queue_head_init(&sch->gso_skb); 956 __skb_queue_head_init(&sch->skb_bad_txq); 957 gnet_stats_basic_sync_init(&sch->bstats); 958 lockdep_register_key(&sch->root_lock_key); 959 spin_lock_init(&sch->q.lock); 960 lockdep_set_class(&sch->q.lock, &sch->root_lock_key); 961 962 if (ops->static_flags & TCQ_F_CPUSTATS) { 963 sch->cpu_bstats = 964 netdev_alloc_pcpu_stats(struct gnet_stats_basic_sync); 965 if (!sch->cpu_bstats) 966 goto errout1; 967 968 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue); 969 if (!sch->cpu_qstats) { 970 free_percpu(sch->cpu_bstats); 971 goto errout1; 972 } 973 } 974 975 /* seqlock has the same scope of busylock, for NOLOCK qdisc */ 976 spin_lock_init(&sch->seqlock); 977 lockdep_set_class(&sch->seqlock, 978 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock); 979 980 sch->ops = ops; 981 sch->flags = ops->static_flags; 982 sch->enqueue = ops->enqueue; 983 sch->dequeue = ops->dequeue; 984 sch->dev_queue = dev_queue; 985 netdev_hold(dev, &sch->dev_tracker, GFP_KERNEL); 986 refcount_set(&sch->refcnt, 1); 987 988 return sch; 989 errout1: 990 lockdep_unregister_key(&sch->root_lock_key); 991 kfree(sch); 992 errout: 993 return ERR_PTR(err); 994 } 995 996 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue, 997 const struct Qdisc_ops *ops, 998 unsigned int parentid, 999 struct netlink_ext_ack *extack) 1000 { 1001 struct Qdisc *sch; 1002 1003 if (!bpf_try_module_get(ops, ops->owner)) { 1004 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter"); 1005 return NULL; 1006 } 1007 1008 sch = qdisc_alloc(dev_queue, ops, extack); 1009 if (IS_ERR(sch)) { 1010 bpf_module_put(ops, ops->owner); 1011 return NULL; 1012 } 1013 sch->parent = parentid; 1014 1015 if (!ops->init || ops->init(sch, NULL, extack) == 0) { 1016 trace_qdisc_create(ops, dev_queue->dev, parentid); 1017 return sch; 1018 } 1019 1020 qdisc_put(sch); 1021 return NULL; 1022 } 1023 EXPORT_SYMBOL(qdisc_create_dflt); 1024 1025 /* Under qdisc_lock(qdisc) and BH! */ 1026 1027 void qdisc_reset(struct Qdisc *qdisc) 1028 { 1029 const struct Qdisc_ops *ops = qdisc->ops; 1030 1031 trace_qdisc_reset(qdisc); 1032 1033 if (ops->reset) 1034 ops->reset(qdisc); 1035 1036 __skb_queue_purge(&qdisc->gso_skb); 1037 __skb_queue_purge(&qdisc->skb_bad_txq); 1038 1039 qdisc->q.qlen = 0; 1040 qdisc->qstats.backlog = 0; 1041 } 1042 EXPORT_SYMBOL(qdisc_reset); 1043 1044 void qdisc_free(struct Qdisc *qdisc) 1045 { 1046 if (qdisc_is_percpu_stats(qdisc)) { 1047 free_percpu(qdisc->cpu_bstats); 1048 free_percpu(qdisc->cpu_qstats); 1049 } 1050 1051 kfree(qdisc); 1052 } 1053 1054 static void qdisc_free_cb(struct rcu_head *head) 1055 { 1056 struct Qdisc *q = container_of(head, struct Qdisc, rcu); 1057 1058 qdisc_free(q); 1059 } 1060 1061 static void __qdisc_destroy(struct Qdisc *qdisc) 1062 { 1063 const struct Qdisc_ops *ops = qdisc->ops; 1064 struct net_device *dev = qdisc_dev(qdisc); 1065 1066 #ifdef CONFIG_NET_SCHED 1067 qdisc_hash_del(qdisc); 1068 1069 qdisc_put_stab(rtnl_dereference(qdisc->stab)); 1070 #endif 1071 gen_kill_estimator(&qdisc->rate_est); 1072 1073 qdisc_reset(qdisc); 1074 1075 1076 if (ops->destroy) 1077 ops->destroy(qdisc); 1078 1079 lockdep_unregister_key(&qdisc->root_lock_key); 1080 bpf_module_put(ops, ops->owner); 1081 netdev_put(dev, &qdisc->dev_tracker); 1082 1083 trace_qdisc_destroy(qdisc); 1084 1085 call_rcu(&qdisc->rcu, qdisc_free_cb); 1086 } 1087 1088 void qdisc_destroy(struct Qdisc *qdisc) 1089 { 1090 if (qdisc->flags & TCQ_F_BUILTIN) 1091 return; 1092 1093 __qdisc_destroy(qdisc); 1094 } 1095 1096 void qdisc_put(struct Qdisc *qdisc) 1097 { 1098 if (!qdisc) 1099 return; 1100 1101 if (qdisc->flags & TCQ_F_BUILTIN || 1102 !refcount_dec_and_test(&qdisc->refcnt)) 1103 return; 1104 1105 __qdisc_destroy(qdisc); 1106 } 1107 EXPORT_SYMBOL(qdisc_put); 1108 1109 /* Version of qdisc_put() that is called with rtnl mutex unlocked. 1110 * Intended to be used as optimization, this function only takes rtnl lock if 1111 * qdisc reference counter reached zero. 1112 */ 1113 1114 void qdisc_put_unlocked(struct Qdisc *qdisc) 1115 { 1116 if (qdisc->flags & TCQ_F_BUILTIN || 1117 !refcount_dec_and_rtnl_lock(&qdisc->refcnt)) 1118 return; 1119 1120 __qdisc_destroy(qdisc); 1121 rtnl_unlock(); 1122 } 1123 EXPORT_SYMBOL(qdisc_put_unlocked); 1124 1125 /* Attach toplevel qdisc to device queue. */ 1126 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue, 1127 struct Qdisc *qdisc) 1128 { 1129 struct Qdisc *oqdisc = rtnl_dereference(dev_queue->qdisc_sleeping); 1130 spinlock_t *root_lock; 1131 1132 root_lock = qdisc_lock(oqdisc); 1133 spin_lock_bh(root_lock); 1134 1135 /* ... and graft new one */ 1136 if (qdisc == NULL) 1137 qdisc = &noop_qdisc; 1138 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc); 1139 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc); 1140 1141 spin_unlock_bh(root_lock); 1142 1143 return oqdisc; 1144 } 1145 EXPORT_SYMBOL(dev_graft_qdisc); 1146 1147 static void shutdown_scheduler_queue(struct net_device *dev, 1148 struct netdev_queue *dev_queue, 1149 void *_qdisc_default) 1150 { 1151 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping); 1152 struct Qdisc *qdisc_default = _qdisc_default; 1153 1154 if (qdisc) { 1155 rcu_assign_pointer(dev_queue->qdisc, qdisc_default); 1156 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc_default); 1157 1158 qdisc_put(qdisc); 1159 } 1160 } 1161 1162 static void attach_one_default_qdisc(struct net_device *dev, 1163 struct netdev_queue *dev_queue, 1164 void *_unused) 1165 { 1166 struct Qdisc *qdisc; 1167 const struct Qdisc_ops *ops = default_qdisc_ops; 1168 1169 if (dev->priv_flags & IFF_NO_QUEUE) 1170 ops = &noqueue_qdisc_ops; 1171 else if(dev->type == ARPHRD_CAN) 1172 ops = &pfifo_fast_ops; 1173 1174 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL); 1175 if (!qdisc) 1176 return; 1177 1178 if (!netif_is_multiqueue(dev)) 1179 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 1180 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc); 1181 } 1182 1183 static void attach_default_qdiscs(struct net_device *dev) 1184 { 1185 struct netdev_queue *txq; 1186 struct Qdisc *qdisc; 1187 1188 txq = netdev_get_tx_queue(dev, 0); 1189 1190 if (!netif_is_multiqueue(dev) || 1191 dev->priv_flags & IFF_NO_QUEUE) { 1192 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL); 1193 qdisc = rtnl_dereference(txq->qdisc_sleeping); 1194 rcu_assign_pointer(dev->qdisc, qdisc); 1195 qdisc_refcount_inc(qdisc); 1196 } else { 1197 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL); 1198 if (qdisc) { 1199 rcu_assign_pointer(dev->qdisc, qdisc); 1200 qdisc->ops->attach(qdisc); 1201 } 1202 } 1203 qdisc = rtnl_dereference(dev->qdisc); 1204 1205 /* Detect default qdisc setup/init failed and fallback to "noqueue" */ 1206 if (qdisc == &noop_qdisc) { 1207 netdev_warn(dev, "default qdisc (%s) fail, fallback to %s\n", 1208 default_qdisc_ops->id, noqueue_qdisc_ops.id); 1209 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc); 1210 dev->priv_flags |= IFF_NO_QUEUE; 1211 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL); 1212 qdisc = rtnl_dereference(txq->qdisc_sleeping); 1213 rcu_assign_pointer(dev->qdisc, qdisc); 1214 qdisc_refcount_inc(qdisc); 1215 dev->priv_flags ^= IFF_NO_QUEUE; 1216 } 1217 1218 #ifdef CONFIG_NET_SCHED 1219 if (qdisc != &noop_qdisc) 1220 qdisc_hash_add(qdisc, false); 1221 #endif 1222 } 1223 1224 static void transition_one_qdisc(struct net_device *dev, 1225 struct netdev_queue *dev_queue, 1226 void *_need_watchdog) 1227 { 1228 struct Qdisc *new_qdisc = rtnl_dereference(dev_queue->qdisc_sleeping); 1229 int *need_watchdog_p = _need_watchdog; 1230 1231 if (!(new_qdisc->flags & TCQ_F_BUILTIN)) 1232 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state); 1233 1234 rcu_assign_pointer(dev_queue->qdisc, new_qdisc); 1235 if (need_watchdog_p) { 1236 WRITE_ONCE(dev_queue->trans_start, 0); 1237 *need_watchdog_p = 1; 1238 } 1239 } 1240 1241 void dev_activate(struct net_device *dev) 1242 { 1243 int need_watchdog; 1244 1245 /* No queueing discipline is attached to device; 1246 * create default one for devices, which need queueing 1247 * and noqueue_qdisc for virtual interfaces 1248 */ 1249 1250 if (rtnl_dereference(dev->qdisc) == &noop_qdisc) 1251 attach_default_qdiscs(dev); 1252 1253 if (!netif_carrier_ok(dev)) 1254 /* Delay activation until next carrier-on event */ 1255 return; 1256 1257 need_watchdog = 0; 1258 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog); 1259 if (dev_ingress_queue(dev)) 1260 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL); 1261 1262 if (need_watchdog) { 1263 netif_trans_update(dev); 1264 netdev_watchdog_up(dev); 1265 } 1266 } 1267 EXPORT_SYMBOL(dev_activate); 1268 1269 static void qdisc_deactivate(struct Qdisc *qdisc) 1270 { 1271 if (qdisc->flags & TCQ_F_BUILTIN) 1272 return; 1273 1274 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state); 1275 } 1276 1277 static void dev_deactivate_queue(struct net_device *dev, 1278 struct netdev_queue *dev_queue, 1279 void *_sync_needed) 1280 { 1281 bool *sync_needed = _sync_needed; 1282 struct Qdisc *qdisc; 1283 1284 qdisc = rtnl_dereference(dev_queue->qdisc); 1285 if (qdisc) { 1286 if (qdisc->enqueue) 1287 *sync_needed = true; 1288 qdisc_deactivate(qdisc); 1289 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc); 1290 } 1291 } 1292 1293 static void dev_reset_queue(struct net_device *dev, 1294 struct netdev_queue *dev_queue, 1295 void *_unused) 1296 { 1297 struct Qdisc *qdisc; 1298 bool nolock; 1299 1300 qdisc = rtnl_dereference(dev_queue->qdisc_sleeping); 1301 if (!qdisc) 1302 return; 1303 1304 nolock = qdisc->flags & TCQ_F_NOLOCK; 1305 1306 if (nolock) 1307 spin_lock_bh(&qdisc->seqlock); 1308 spin_lock_bh(qdisc_lock(qdisc)); 1309 1310 qdisc_reset(qdisc); 1311 1312 spin_unlock_bh(qdisc_lock(qdisc)); 1313 if (nolock) { 1314 clear_bit(__QDISC_STATE_MISSED, &qdisc->state); 1315 clear_bit(__QDISC_STATE_DRAINING, &qdisc->state); 1316 spin_unlock_bh(&qdisc->seqlock); 1317 } 1318 } 1319 1320 static bool some_qdisc_is_busy(struct net_device *dev) 1321 { 1322 unsigned int i; 1323 1324 for (i = 0; i < dev->num_tx_queues; i++) { 1325 struct netdev_queue *dev_queue; 1326 spinlock_t *root_lock; 1327 struct Qdisc *q; 1328 int val; 1329 1330 dev_queue = netdev_get_tx_queue(dev, i); 1331 q = rtnl_dereference(dev_queue->qdisc_sleeping); 1332 1333 root_lock = qdisc_lock(q); 1334 spin_lock_bh(root_lock); 1335 1336 val = (qdisc_is_running(q) || 1337 test_bit(__QDISC_STATE_SCHED, &q->state)); 1338 1339 spin_unlock_bh(root_lock); 1340 1341 if (val) 1342 return true; 1343 } 1344 return false; 1345 } 1346 1347 /** 1348 * dev_deactivate_many - deactivate transmissions on several devices 1349 * @head: list of devices to deactivate 1350 * 1351 * This function returns only when all outstanding transmissions 1352 * have completed, unless all devices are in dismantle phase. 1353 */ 1354 void dev_deactivate_many(struct list_head *head) 1355 { 1356 bool sync_needed = false; 1357 struct net_device *dev; 1358 1359 list_for_each_entry(dev, head, close_list) { 1360 netdev_for_each_tx_queue(dev, dev_deactivate_queue, 1361 &sync_needed); 1362 if (dev_ingress_queue(dev)) 1363 dev_deactivate_queue(dev, dev_ingress_queue(dev), 1364 &sync_needed); 1365 1366 netdev_watchdog_down(dev); 1367 } 1368 1369 /* Wait for outstanding qdisc enqueuing calls. */ 1370 if (sync_needed) 1371 synchronize_net(); 1372 1373 list_for_each_entry(dev, head, close_list) { 1374 netdev_for_each_tx_queue(dev, dev_reset_queue, NULL); 1375 1376 if (dev_ingress_queue(dev)) 1377 dev_reset_queue(dev, dev_ingress_queue(dev), NULL); 1378 } 1379 1380 /* Wait for outstanding qdisc_run calls. */ 1381 list_for_each_entry(dev, head, close_list) { 1382 while (some_qdisc_is_busy(dev)) { 1383 /* wait_event() would avoid this sleep-loop but would 1384 * require expensive checks in the fast paths of packet 1385 * processing which isn't worth it. 1386 */ 1387 schedule_timeout_uninterruptible(1); 1388 } 1389 } 1390 } 1391 1392 void dev_deactivate(struct net_device *dev) 1393 { 1394 LIST_HEAD(single); 1395 1396 list_add(&dev->close_list, &single); 1397 dev_deactivate_many(&single); 1398 list_del(&single); 1399 } 1400 EXPORT_SYMBOL(dev_deactivate); 1401 1402 static int qdisc_change_tx_queue_len(struct net_device *dev, 1403 struct netdev_queue *dev_queue) 1404 { 1405 struct Qdisc *qdisc = rtnl_dereference(dev_queue->qdisc_sleeping); 1406 const struct Qdisc_ops *ops = qdisc->ops; 1407 1408 if (ops->change_tx_queue_len) 1409 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len); 1410 return 0; 1411 } 1412 1413 void dev_qdisc_change_real_num_tx(struct net_device *dev, 1414 unsigned int new_real_tx) 1415 { 1416 struct Qdisc *qdisc = rtnl_dereference(dev->qdisc); 1417 1418 if (qdisc->ops->change_real_num_tx) 1419 qdisc->ops->change_real_num_tx(qdisc, new_real_tx); 1420 } 1421 1422 void mq_change_real_num_tx(struct Qdisc *sch, unsigned int new_real_tx) 1423 { 1424 #ifdef CONFIG_NET_SCHED 1425 struct net_device *dev = qdisc_dev(sch); 1426 struct Qdisc *qdisc; 1427 unsigned int i; 1428 1429 for (i = new_real_tx; i < dev->real_num_tx_queues; i++) { 1430 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping); 1431 /* Only update the default qdiscs we created, 1432 * qdiscs with handles are always hashed. 1433 */ 1434 if (qdisc != &noop_qdisc && !qdisc->handle) 1435 qdisc_hash_del(qdisc); 1436 } 1437 for (i = dev->real_num_tx_queues; i < new_real_tx; i++) { 1438 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc_sleeping); 1439 if (qdisc != &noop_qdisc && !qdisc->handle) 1440 qdisc_hash_add(qdisc, false); 1441 } 1442 #endif 1443 } 1444 EXPORT_SYMBOL(mq_change_real_num_tx); 1445 1446 int dev_qdisc_change_tx_queue_len(struct net_device *dev) 1447 { 1448 bool up = dev->flags & IFF_UP; 1449 unsigned int i; 1450 int ret = 0; 1451 1452 if (up) 1453 dev_deactivate(dev); 1454 1455 for (i = 0; i < dev->num_tx_queues; i++) { 1456 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]); 1457 1458 /* TODO: revert changes on a partial failure */ 1459 if (ret) 1460 break; 1461 } 1462 1463 if (up) 1464 dev_activate(dev); 1465 return ret; 1466 } 1467 1468 static void dev_init_scheduler_queue(struct net_device *dev, 1469 struct netdev_queue *dev_queue, 1470 void *_qdisc) 1471 { 1472 struct Qdisc *qdisc = _qdisc; 1473 1474 rcu_assign_pointer(dev_queue->qdisc, qdisc); 1475 rcu_assign_pointer(dev_queue->qdisc_sleeping, qdisc); 1476 } 1477 1478 void dev_init_scheduler(struct net_device *dev) 1479 { 1480 rcu_assign_pointer(dev->qdisc, &noop_qdisc); 1481 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc); 1482 if (dev_ingress_queue(dev)) 1483 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc); 1484 1485 timer_setup(&dev->watchdog_timer, dev_watchdog, 0); 1486 } 1487 1488 void dev_shutdown(struct net_device *dev) 1489 { 1490 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc); 1491 if (dev_ingress_queue(dev)) 1492 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc); 1493 qdisc_put(rtnl_dereference(dev->qdisc)); 1494 rcu_assign_pointer(dev->qdisc, &noop_qdisc); 1495 1496 WARN_ON(timer_pending(&dev->watchdog_timer)); 1497 } 1498 1499 /** 1500 * psched_ratecfg_precompute__() - Pre-compute values for reciprocal division 1501 * @rate: Rate to compute reciprocal division values of 1502 * @mult: Multiplier for reciprocal division 1503 * @shift: Shift for reciprocal division 1504 * 1505 * The multiplier and shift for reciprocal division by rate are stored 1506 * in mult and shift. 1507 * 1508 * The deal here is to replace a divide by a reciprocal one 1509 * in fast path (a reciprocal divide is a multiply and a shift) 1510 * 1511 * Normal formula would be : 1512 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps 1513 * 1514 * We compute mult/shift to use instead : 1515 * time_in_ns = (len * mult) >> shift; 1516 * 1517 * We try to get the highest possible mult value for accuracy, 1518 * but have to make sure no overflows will ever happen. 1519 * 1520 * reciprocal_value() is not used here it doesn't handle 64-bit values. 1521 */ 1522 static void psched_ratecfg_precompute__(u64 rate, u32 *mult, u8 *shift) 1523 { 1524 u64 factor = NSEC_PER_SEC; 1525 1526 *mult = 1; 1527 *shift = 0; 1528 1529 if (rate <= 0) 1530 return; 1531 1532 for (;;) { 1533 *mult = div64_u64(factor, rate); 1534 if (*mult & (1U << 31) || factor & (1ULL << 63)) 1535 break; 1536 factor <<= 1; 1537 (*shift)++; 1538 } 1539 } 1540 1541 void psched_ratecfg_precompute(struct psched_ratecfg *r, 1542 const struct tc_ratespec *conf, 1543 u64 rate64) 1544 { 1545 memset(r, 0, sizeof(*r)); 1546 r->overhead = conf->overhead; 1547 r->mpu = conf->mpu; 1548 r->rate_bytes_ps = max_t(u64, conf->rate, rate64); 1549 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK); 1550 psched_ratecfg_precompute__(r->rate_bytes_ps, &r->mult, &r->shift); 1551 } 1552 EXPORT_SYMBOL(psched_ratecfg_precompute); 1553 1554 void psched_ppscfg_precompute(struct psched_pktrate *r, u64 pktrate64) 1555 { 1556 r->rate_pkts_ps = pktrate64; 1557 psched_ratecfg_precompute__(r->rate_pkts_ps, &r->mult, &r->shift); 1558 } 1559 EXPORT_SYMBOL(psched_ppscfg_precompute); 1560 1561 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp, 1562 struct tcf_proto *tp_head) 1563 { 1564 /* Protected with chain0->filter_chain_lock. 1565 * Can't access chain directly because tp_head can be NULL. 1566 */ 1567 struct mini_Qdisc *miniq_old = 1568 rcu_dereference_protected(*miniqp->p_miniq, 1); 1569 struct mini_Qdisc *miniq; 1570 1571 if (!tp_head) { 1572 RCU_INIT_POINTER(*miniqp->p_miniq, NULL); 1573 } else { 1574 miniq = miniq_old != &miniqp->miniq1 ? 1575 &miniqp->miniq1 : &miniqp->miniq2; 1576 1577 /* We need to make sure that readers won't see the miniq 1578 * we are about to modify. So ensure that at least one RCU 1579 * grace period has elapsed since the miniq was made 1580 * inactive. 1581 */ 1582 if (IS_ENABLED(CONFIG_PREEMPT_RT)) 1583 cond_synchronize_rcu(miniq->rcu_state); 1584 else if (!poll_state_synchronize_rcu(miniq->rcu_state)) 1585 synchronize_rcu_expedited(); 1586 1587 miniq->filter_list = tp_head; 1588 rcu_assign_pointer(*miniqp->p_miniq, miniq); 1589 } 1590 1591 if (miniq_old) 1592 /* This is counterpart of the rcu sync above. We need to 1593 * block potential new user of miniq_old until all readers 1594 * are not seeing it. 1595 */ 1596 miniq_old->rcu_state = start_poll_synchronize_rcu(); 1597 } 1598 EXPORT_SYMBOL(mini_qdisc_pair_swap); 1599 1600 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp, 1601 struct tcf_block *block) 1602 { 1603 miniqp->miniq1.block = block; 1604 miniqp->miniq2.block = block; 1605 } 1606 EXPORT_SYMBOL(mini_qdisc_pair_block_init); 1607 1608 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc, 1609 struct mini_Qdisc __rcu **p_miniq) 1610 { 1611 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats; 1612 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats; 1613 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats; 1614 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats; 1615 miniqp->miniq1.rcu_state = get_state_synchronize_rcu(); 1616 miniqp->miniq2.rcu_state = miniqp->miniq1.rcu_state; 1617 miniqp->p_miniq = p_miniq; 1618 } 1619 EXPORT_SYMBOL(mini_qdisc_pair_init); 1620