Linux Audio

Check our new training course

Loading...
v5.4
   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 <net/sch_generic.h>
  28#include <net/pkt_sched.h>
  29#include <net/dst.h>
  30#include <trace/events/qdisc.h>
  31#include <trace/events/net.h>
  32#include <net/xfrm.h>
  33
  34/* Qdisc to use by default */
  35const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
  36EXPORT_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#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
  50
  51static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
  52{
  53	const struct netdev_queue *txq = q->dev_queue;
  54	spinlock_t *lock = NULL;
  55	struct sk_buff *skb;
  56
  57	if (q->flags & TCQ_F_NOLOCK) {
  58		lock = qdisc_lock(q);
  59		spin_lock(lock);
  60	}
  61
  62	skb = skb_peek(&q->skb_bad_txq);
  63	if (skb) {
  64		/* check the reason of requeuing without tx lock first */
  65		txq = skb_get_tx_queue(txq->dev, skb);
  66		if (!netif_xmit_frozen_or_stopped(txq)) {
  67			skb = __skb_dequeue(&q->skb_bad_txq);
  68			if (qdisc_is_percpu_stats(q)) {
  69				qdisc_qstats_cpu_backlog_dec(q, skb);
  70				qdisc_qstats_cpu_qlen_dec(q);
  71			} else {
  72				qdisc_qstats_backlog_dec(q, skb);
  73				q->q.qlen--;
  74			}
  75		} else {
  76			skb = SKB_XOFF_MAGIC;
  77		}
  78	}
  79
  80	if (lock)
  81		spin_unlock(lock);
  82
  83	return skb;
  84}
  85
  86static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
  87{
  88	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
  89
  90	if (unlikely(skb))
  91		skb = __skb_dequeue_bad_txq(q);
 
  92
  93	return skb;
  94}
  95
  96static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
  97					     struct sk_buff *skb)
  98{
  99	spinlock_t *lock = NULL;
 100
 101	if (q->flags & TCQ_F_NOLOCK) {
 102		lock = qdisc_lock(q);
 103		spin_lock(lock);
 104	}
 105
 106	__skb_queue_tail(&q->skb_bad_txq, skb);
 107
 108	if (qdisc_is_percpu_stats(q)) {
 109		qdisc_qstats_cpu_backlog_inc(q, skb);
 110		qdisc_qstats_cpu_qlen_inc(q);
 
 
 
 
 111	} else {
 112		qdisc_qstats_backlog_inc(q, skb);
 113		q->q.qlen++;
 114	}
 115
 116	if (lock)
 117		spin_unlock(lock);
 118}
 119
 120static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 121{
 122	spinlock_t *lock = NULL;
 123
 124	if (q->flags & TCQ_F_NOLOCK) {
 125		lock = qdisc_lock(q);
 126		spin_lock(lock);
 127	}
 128
 129	while (skb) {
 130		struct sk_buff *next = skb->next;
 131
 132		__skb_queue_tail(&q->gso_skb, skb);
 133
 134		/* it's still part of the queue */
 135		if (qdisc_is_percpu_stats(q)) {
 136			qdisc_qstats_cpu_requeues_inc(q);
 137			qdisc_qstats_cpu_backlog_inc(q, skb);
 138			qdisc_qstats_cpu_qlen_inc(q);
 139		} else {
 140			q->qstats.requeues++;
 141			qdisc_qstats_backlog_inc(q, skb);
 142			q->q.qlen++;
 143		}
 144
 145		skb = next;
 146	}
 147	if (lock)
 148		spin_unlock(lock);
 149	__netif_schedule(q);
 150}
 151
 152static void try_bulk_dequeue_skb(struct Qdisc *q,
 153				 struct sk_buff *skb,
 154				 const struct netdev_queue *txq,
 155				 int *packets)
 156{
 157	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
 158
 159	while (bytelimit > 0) {
 160		struct sk_buff *nskb = q->dequeue(q);
 161
 162		if (!nskb)
 163			break;
 164
 165		bytelimit -= nskb->len; /* covers GSO len */
 166		skb->next = nskb;
 167		skb = nskb;
 168		(*packets)++; /* GSO counts as one pkt */
 169	}
 170	skb_mark_not_on_list(skb);
 171}
 172
 173/* This variant of try_bulk_dequeue_skb() makes sure
 174 * all skbs in the chain are for the same txq
 175 */
 176static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
 177				      struct sk_buff *skb,
 178				      int *packets)
 179{
 180	int mapping = skb_get_queue_mapping(skb);
 181	struct sk_buff *nskb;
 182	int cnt = 0;
 183
 184	do {
 185		nskb = q->dequeue(q);
 186		if (!nskb)
 187			break;
 188		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
 189			qdisc_enqueue_skb_bad_txq(q, nskb);
 190			break;
 191		}
 192		skb->next = nskb;
 193		skb = nskb;
 194	} while (++cnt < 8);
 195	(*packets) += cnt;
 196	skb_mark_not_on_list(skb);
 197}
 198
 199/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
 200 * A requeued skb (via q->gso_skb) can also be a SKB list.
 201 */
 202static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
 203				   int *packets)
 204{
 205	const struct netdev_queue *txq = q->dev_queue;
 206	struct sk_buff *skb = NULL;
 207
 208	*packets = 1;
 209	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
 210		spinlock_t *lock = NULL;
 211
 212		if (q->flags & TCQ_F_NOLOCK) {
 213			lock = qdisc_lock(q);
 214			spin_lock(lock);
 215		}
 216
 217		skb = skb_peek(&q->gso_skb);
 218
 219		/* skb may be null if another cpu pulls gso_skb off in between
 220		 * empty check and lock.
 221		 */
 222		if (!skb) {
 223			if (lock)
 224				spin_unlock(lock);
 225			goto validate;
 226		}
 227
 228		/* skb in gso_skb were already validated */
 229		*validate = false;
 230		if (xfrm_offload(skb))
 231			*validate = true;
 232		/* check the reason of requeuing without tx lock first */
 233		txq = skb_get_tx_queue(txq->dev, skb);
 234		if (!netif_xmit_frozen_or_stopped(txq)) {
 235			skb = __skb_dequeue(&q->gso_skb);
 236			if (qdisc_is_percpu_stats(q)) {
 237				qdisc_qstats_cpu_backlog_dec(q, skb);
 238				qdisc_qstats_cpu_qlen_dec(q);
 239			} else {
 240				qdisc_qstats_backlog_dec(q, skb);
 241				q->q.qlen--;
 242			}
 243		} else {
 244			skb = NULL;
 245		}
 246		if (lock)
 247			spin_unlock(lock);
 248		goto trace;
 249	}
 250validate:
 251	*validate = true;
 252
 253	if ((q->flags & TCQ_F_ONETXQUEUE) &&
 254	    netif_xmit_frozen_or_stopped(txq))
 255		return skb;
 256
 257	skb = qdisc_dequeue_skb_bad_txq(q);
 258	if (unlikely(skb)) {
 259		if (skb == SKB_XOFF_MAGIC)
 260			return NULL;
 261		goto bulk;
 262	}
 263	skb = q->dequeue(q);
 264	if (skb) {
 265bulk:
 266		if (qdisc_may_bulk(q))
 267			try_bulk_dequeue_skb(q, skb, txq, packets);
 268		else
 269			try_bulk_dequeue_skb_slow(q, skb, packets);
 270	}
 271trace:
 272	trace_qdisc_dequeue(q, txq, *packets, skb);
 273	return skb;
 274}
 275
 276/*
 277 * Transmit possibly several skbs, and handle the return status as
 278 * required. Owning running seqcount bit guarantees that
 279 * only one CPU can execute this function.
 280 *
 281 * Returns to the caller:
 282 *				false  - hardware queue frozen backoff
 283 *				true   - feel free to send more pkts
 284 */
 285bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
 286		     struct net_device *dev, struct netdev_queue *txq,
 287		     spinlock_t *root_lock, bool validate)
 288{
 289	int ret = NETDEV_TX_BUSY;
 290	bool again = false;
 291
 292	/* And release qdisc */
 293	if (root_lock)
 294		spin_unlock(root_lock);
 295
 296	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
 297	if (validate)
 298		skb = validate_xmit_skb_list(skb, dev, &again);
 299
 300#ifdef CONFIG_XFRM_OFFLOAD
 301	if (unlikely(again)) {
 302		if (root_lock)
 303			spin_lock(root_lock);
 304
 305		dev_requeue_skb(skb, q);
 306		return false;
 307	}
 308#endif
 309
 310	if (likely(skb)) {
 311		HARD_TX_LOCK(dev, txq, smp_processor_id());
 312		if (!netif_xmit_frozen_or_stopped(txq))
 313			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
 314
 315		HARD_TX_UNLOCK(dev, txq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 316	} else {
 317		if (root_lock)
 318			spin_lock(root_lock);
 319		return true;
 320	}
 321
 322	if (root_lock)
 323		spin_lock(root_lock);
 324
 325	if (!dev_xmit_complete(ret)) {
 326		/* Driver returned NETDEV_TX_BUSY - requeue skb */
 327		if (unlikely(ret != NETDEV_TX_BUSY))
 328			net_warn_ratelimited("BUG %s code %d qlen %d\n",
 329					     dev->name, ret, q->q.qlen);
 330
 331		dev_requeue_skb(skb, q);
 332		return false;
 333	}
 334
 335	return true;
 
 
 
 336}
 337
 338/*
 339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
 340 *
 341 * running seqcount guarantees only one CPU can process
 342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
 343 * this queue.
 344 *
 345 *  netif_tx_lock serializes accesses to device driver.
 346 *
 347 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
 348 *  if one is grabbed, another must be free.
 349 *
 350 * Note, that this procedure can be called by a watchdog timer
 351 *
 352 * Returns to the caller:
 353 *				0  - queue is empty or throttled.
 354 *				>0 - queue is not empty.
 355 *
 356 */
 357static inline bool qdisc_restart(struct Qdisc *q, int *packets)
 358{
 359	spinlock_t *root_lock = NULL;
 360	struct netdev_queue *txq;
 361	struct net_device *dev;
 
 362	struct sk_buff *skb;
 363	bool validate;
 364
 365	/* Dequeue packet */
 366	skb = dequeue_skb(q, &validate, packets);
 367	if (unlikely(!skb))
 368		return false;
 369
 370	if (!(q->flags & TCQ_F_NOLOCK))
 371		root_lock = qdisc_lock(q);
 372
 373	dev = qdisc_dev(q);
 374	txq = skb_get_tx_queue(dev, skb);
 375
 376	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
 377}
 378
 379void __qdisc_run(struct Qdisc *q)
 380{
 381	int quota = dev_tx_weight;
 382	int packets;
 383
 384	while (qdisc_restart(q, &packets)) {
 385		/*
 386		 * Ordered by possible occurrence: Postpone processing if
 387		 * 1. we've exceeded packet quota
 388		 * 2. another process needs the CPU;
 389		 */
 390		quota -= packets;
 391		if (quota <= 0 || need_resched()) {
 392			__netif_schedule(q);
 393			break;
 394		}
 395	}
 
 
 396}
 397
 398unsigned long dev_trans_start(struct net_device *dev)
 399{
 400	unsigned long val, res;
 401	unsigned int i;
 402
 403	if (is_vlan_dev(dev))
 404		dev = vlan_dev_real_dev(dev);
 405	else if (netif_is_macvlan(dev))
 406		dev = macvlan_dev_real_dev(dev);
 407	res = netdev_get_tx_queue(dev, 0)->trans_start;
 408	for (i = 1; i < dev->num_tx_queues; i++) {
 409		val = netdev_get_tx_queue(dev, i)->trans_start;
 410		if (val && time_after(val, res))
 411			res = val;
 412	}
 413
 414	return res;
 415}
 416EXPORT_SYMBOL(dev_trans_start);
 417
 418static void dev_watchdog(struct timer_list *t)
 419{
 420	struct net_device *dev = from_timer(dev, t, watchdog_timer);
 421
 422	netif_tx_lock(dev);
 423	if (!qdisc_tx_is_noop(dev)) {
 424		if (netif_device_present(dev) &&
 425		    netif_running(dev) &&
 426		    netif_carrier_ok(dev)) {
 427			int some_queue_timedout = 0;
 428			unsigned int i;
 429			unsigned long trans_start;
 430
 431			for (i = 0; i < dev->num_tx_queues; i++) {
 432				struct netdev_queue *txq;
 433
 434				txq = netdev_get_tx_queue(dev, i);
 435				trans_start = txq->trans_start;
 
 
 
 436				if (netif_xmit_stopped(txq) &&
 437				    time_after(jiffies, (trans_start +
 438							 dev->watchdog_timeo))) {
 439					some_queue_timedout = 1;
 440					txq->trans_timeout++;
 441					break;
 442				}
 443			}
 444
 445			if (some_queue_timedout) {
 446				trace_net_dev_xmit_timeout(dev, i);
 447				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
 448				       dev->name, netdev_drivername(dev), i);
 449				dev->netdev_ops->ndo_tx_timeout(dev);
 450			}
 451			if (!mod_timer(&dev->watchdog_timer,
 452				       round_jiffies(jiffies +
 453						     dev->watchdog_timeo)))
 454				dev_hold(dev);
 455		}
 456	}
 457	netif_tx_unlock(dev);
 458
 459	dev_put(dev);
 460}
 461
 462void __netdev_watchdog_up(struct net_device *dev)
 463{
 464	if (dev->netdev_ops->ndo_tx_timeout) {
 465		if (dev->watchdog_timeo <= 0)
 466			dev->watchdog_timeo = 5*HZ;
 467		if (!mod_timer(&dev->watchdog_timer,
 468			       round_jiffies(jiffies + dev->watchdog_timeo)))
 469			dev_hold(dev);
 470	}
 471}
 472
 473static void dev_watchdog_up(struct net_device *dev)
 474{
 475	__netdev_watchdog_up(dev);
 476}
 477
 478static void dev_watchdog_down(struct net_device *dev)
 479{
 480	netif_tx_lock_bh(dev);
 481	if (del_timer(&dev->watchdog_timer))
 482		dev_put(dev);
 483	netif_tx_unlock_bh(dev);
 484}
 485
 486/**
 487 *	netif_carrier_on - set carrier
 488 *	@dev: network device
 489 *
 490 * Device has detected acquisition of carrier.
 491 */
 492void netif_carrier_on(struct net_device *dev)
 493{
 494	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 495		if (dev->reg_state == NETREG_UNINITIALIZED)
 496			return;
 497		atomic_inc(&dev->carrier_up_count);
 498		linkwatch_fire_event(dev);
 499		if (netif_running(dev))
 500			__netdev_watchdog_up(dev);
 501	}
 502}
 503EXPORT_SYMBOL(netif_carrier_on);
 504
 505/**
 506 *	netif_carrier_off - clear carrier
 507 *	@dev: network device
 508 *
 509 * Device has detected loss of carrier.
 510 */
 511void netif_carrier_off(struct net_device *dev)
 512{
 513	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
 514		if (dev->reg_state == NETREG_UNINITIALIZED)
 515			return;
 516		atomic_inc(&dev->carrier_down_count);
 517		linkwatch_fire_event(dev);
 518	}
 519}
 520EXPORT_SYMBOL(netif_carrier_off);
 521
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 522/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
 523   under all circumstances. It is difficult to invent anything faster or
 524   cheaper.
 525 */
 526
 527static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 528			struct sk_buff **to_free)
 529{
 530	__qdisc_drop(skb, to_free);
 531	return NET_XMIT_CN;
 532}
 533
 534static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
 535{
 536	return NULL;
 537}
 538
 539struct Qdisc_ops noop_qdisc_ops __read_mostly = {
 540	.id		=	"noop",
 541	.priv_size	=	0,
 542	.enqueue	=	noop_enqueue,
 543	.dequeue	=	noop_dequeue,
 544	.peek		=	noop_dequeue,
 545	.owner		=	THIS_MODULE,
 546};
 547
 548static struct netdev_queue noop_netdev_queue = {
 549	RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
 550	.qdisc_sleeping	=	&noop_qdisc,
 551};
 552
 553struct Qdisc noop_qdisc = {
 554	.enqueue	=	noop_enqueue,
 555	.dequeue	=	noop_dequeue,
 556	.flags		=	TCQ_F_BUILTIN,
 557	.ops		=	&noop_qdisc_ops,
 
 558	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
 559	.dev_queue	=	&noop_netdev_queue,
 560	.running	=	SEQCNT_ZERO(noop_qdisc.running),
 561	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 562	.gso_skb = {
 563		.next = (struct sk_buff *)&noop_qdisc.gso_skb,
 564		.prev = (struct sk_buff *)&noop_qdisc.gso_skb,
 565		.qlen = 0,
 566		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
 567	},
 568	.skb_bad_txq = {
 569		.next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 570		.prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
 571		.qlen = 0,
 572		.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
 573	},
 574};
 575EXPORT_SYMBOL(noop_qdisc);
 576
 577static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
 578			struct netlink_ext_ack *extack)
 579{
 580	/* register_qdisc() assigns a default of noop_enqueue if unset,
 581	 * but __dev_queue_xmit() treats noqueue only as such
 582	 * if this is NULL - so clear it here. */
 583	qdisc->enqueue = NULL;
 584	return 0;
 585}
 586
 587struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 588	.id		=	"noqueue",
 589	.priv_size	=	0,
 590	.init		=	noqueue_init,
 591	.enqueue	=	noop_enqueue,
 592	.dequeue	=	noop_dequeue,
 593	.peek		=	noop_dequeue,
 594	.owner		=	THIS_MODULE,
 595};
 596
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 597static const u8 prio2band[TC_PRIO_MAX + 1] = {
 598	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 599};
 600
 601/* 3-band FIFO queue: old style, but should be a bit faster than
 602   generic prio+fifo combination.
 603 */
 604
 605#define PFIFO_FAST_BANDS 3
 606
 607/*
 608 * Private data for a pfifo_fast scheduler containing:
 609 *	- rings for priority bands
 
 610 */
 611struct pfifo_fast_priv {
 612	struct skb_array q[PFIFO_FAST_BANDS];
 
 613};
 614
 615static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
 616					  int band)
 
 
 
 
 
 
 
 
 617{
 618	return &priv->q[band];
 619}
 620
 621static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
 622			      struct sk_buff **to_free)
 623{
 624	int band = prio2band[skb->priority & TC_PRIO_MAX];
 625	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 626	struct skb_array *q = band2list(priv, band);
 627	unsigned int pkt_len = qdisc_pkt_len(skb);
 628	int err;
 629
 630	err = skb_array_produce(q, skb);
 631
 632	if (unlikely(err)) {
 633		if (qdisc_is_percpu_stats(qdisc))
 634			return qdisc_drop_cpu(skb, qdisc, to_free);
 635		else
 636			return qdisc_drop(skb, qdisc, to_free);
 637	}
 638
 639	qdisc_update_stats_at_enqueue(qdisc, pkt_len);
 640	return NET_XMIT_SUCCESS;
 641}
 642
 643static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
 644{
 645	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 646	struct sk_buff *skb = NULL;
 647	int band;
 648
 649	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 650		struct skb_array *q = band2list(priv, band);
 651
 652		if (__skb_array_empty(q))
 653			continue;
 
 
 
 
 
 654
 655		skb = __skb_array_consume(q);
 656	}
 657	if (likely(skb)) {
 658		qdisc_update_stats_at_dequeue(qdisc, skb);
 659	} else {
 660		qdisc->empty = true;
 661	}
 662
 663	return skb;
 664}
 665
 666static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
 667{
 668	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 669	struct sk_buff *skb = NULL;
 670	int band;
 671
 672	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
 673		struct skb_array *q = band2list(priv, band);
 674
 675		skb = __skb_array_peek(q);
 676	}
 677
 678	return skb;
 679}
 680
 681static void pfifo_fast_reset(struct Qdisc *qdisc)
 682{
 683	int i, band;
 684	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 685
 686	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
 687		struct skb_array *q = band2list(priv, band);
 688		struct sk_buff *skb;
 689
 690		/* NULL ring is possible if destroy path is due to a failed
 691		 * skb_array_init() in pfifo_fast_init() case.
 692		 */
 693		if (!q->ring.queue)
 694			continue;
 695
 696		while ((skb = __skb_array_consume(q)) != NULL)
 697			kfree_skb(skb);
 698	}
 699
 700	if (qdisc_is_percpu_stats(qdisc)) {
 701		for_each_possible_cpu(i) {
 702			struct gnet_stats_queue *q;
 703
 704			q = per_cpu_ptr(qdisc->cpu_qstats, i);
 705			q->backlog = 0;
 706			q->qlen = 0;
 707		}
 708	}
 709}
 710
 711static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 712{
 713	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 714
 715	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
 716	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 717		goto nla_put_failure;
 718	return skb->len;
 719
 720nla_put_failure:
 721	return -1;
 722}
 723
 724static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
 725			   struct netlink_ext_ack *extack)
 726{
 727	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
 728	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
 729	int prio;
 
 730
 731	/* guard against zero length rings */
 732	if (!qlen)
 733		return -EINVAL;
 734
 735	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 736		struct skb_array *q = band2list(priv, prio);
 737		int err;
 738
 739		err = skb_array_init(q, qlen, GFP_KERNEL);
 740		if (err)
 741			return -ENOMEM;
 742	}
 743
 744	/* Can by-pass the queue discipline */
 745	qdisc->flags |= TCQ_F_CAN_BYPASS;
 746	return 0;
 747}
 748
 749static void pfifo_fast_destroy(struct Qdisc *sch)
 750{
 751	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 752	int prio;
 753
 754	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 755		struct skb_array *q = band2list(priv, prio);
 756
 757		/* NULL ring is possible if destroy path is due to a failed
 758		 * skb_array_init() in pfifo_fast_init() case.
 759		 */
 760		if (!q->ring.queue)
 761			continue;
 762		/* Destroy ring but no need to kfree_skb because a call to
 763		 * pfifo_fast_reset() has already done that work.
 764		 */
 765		ptr_ring_cleanup(&q->ring, NULL);
 766	}
 767}
 768
 769static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
 770					  unsigned int new_len)
 771{
 772	struct pfifo_fast_priv *priv = qdisc_priv(sch);
 773	struct skb_array *bands[PFIFO_FAST_BANDS];
 774	int prio;
 775
 776	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
 777		struct skb_array *q = band2list(priv, prio);
 778
 779		bands[prio] = q;
 780	}
 781
 782	return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
 783					 GFP_KERNEL);
 784}
 785
 786struct Qdisc_ops pfifo_fast_ops __read_mostly = {
 787	.id		=	"pfifo_fast",
 788	.priv_size	=	sizeof(struct pfifo_fast_priv),
 789	.enqueue	=	pfifo_fast_enqueue,
 790	.dequeue	=	pfifo_fast_dequeue,
 791	.peek		=	pfifo_fast_peek,
 792	.init		=	pfifo_fast_init,
 793	.destroy	=	pfifo_fast_destroy,
 794	.reset		=	pfifo_fast_reset,
 795	.dump		=	pfifo_fast_dump,
 796	.change_tx_queue_len =  pfifo_fast_change_tx_queue_len,
 797	.owner		=	THIS_MODULE,
 798	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
 799};
 800EXPORT_SYMBOL(pfifo_fast_ops);
 801
 802struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
 803			  const struct Qdisc_ops *ops,
 804			  struct netlink_ext_ack *extack)
 805{
 806	void *p;
 807	struct Qdisc *sch;
 808	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
 809	int err = -ENOBUFS;
 810	struct net_device *dev;
 811
 812	if (!dev_queue) {
 813		NL_SET_ERR_MSG(extack, "No device queue given");
 814		err = -EINVAL;
 815		goto errout;
 816	}
 817
 818	dev = dev_queue->dev;
 819	p = kzalloc_node(size, GFP_KERNEL,
 820			 netdev_queue_numa_node_read(dev_queue));
 821
 822	if (!p)
 823		goto errout;
 824	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 825	/* if we got non aligned memory, ask more and do alignment ourself */
 826	if (sch != p) {
 827		kfree(p);
 828		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
 829				 netdev_queue_numa_node_read(dev_queue));
 830		if (!p)
 831			goto errout;
 832		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
 833		sch->padded = (char *) sch - (char *) p;
 834	}
 835	__skb_queue_head_init(&sch->gso_skb);
 836	__skb_queue_head_init(&sch->skb_bad_txq);
 837	qdisc_skb_head_init(&sch->q);
 838	spin_lock_init(&sch->q.lock);
 839
 840	if (ops->static_flags & TCQ_F_CPUSTATS) {
 841		sch->cpu_bstats =
 842			netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
 843		if (!sch->cpu_bstats)
 844			goto errout1;
 845
 846		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
 847		if (!sch->cpu_qstats) {
 848			free_percpu(sch->cpu_bstats);
 849			goto errout1;
 850		}
 851	}
 852
 853	spin_lock_init(&sch->busylock);
 854	/* seqlock has the same scope of busylock, for NOLOCK qdisc */
 855	spin_lock_init(&sch->seqlock);
 856	seqcount_init(&sch->running);
 857
 858	sch->ops = ops;
 859	sch->flags = ops->static_flags;
 860	sch->enqueue = ops->enqueue;
 861	sch->dequeue = ops->dequeue;
 862	sch->dev_queue = dev_queue;
 863	sch->empty = true;
 864	dev_hold(dev);
 865	refcount_set(&sch->refcnt, 1);
 866
 867	if (sch != &noop_qdisc) {
 868		lockdep_set_class(&sch->busylock, &dev->qdisc_tx_busylock_key);
 869		lockdep_set_class(&sch->seqlock, &dev->qdisc_tx_busylock_key);
 870		lockdep_set_class(&sch->running, &dev->qdisc_running_key);
 871	}
 872
 873	return sch;
 874errout1:
 875	kfree(p);
 876errout:
 877	return ERR_PTR(err);
 878}
 879
 880struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
 881				const struct Qdisc_ops *ops,
 882				unsigned int parentid,
 883				struct netlink_ext_ack *extack)
 884{
 885	struct Qdisc *sch;
 886
 887	if (!try_module_get(ops->owner)) {
 888		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
 889		return NULL;
 890	}
 891
 892	sch = qdisc_alloc(dev_queue, ops, extack);
 893	if (IS_ERR(sch)) {
 894		module_put(ops->owner);
 895		return NULL;
 896	}
 897	sch->parent = parentid;
 898
 899	if (!ops->init || ops->init(sch, NULL, extack) == 0)
 900		return sch;
 901
 902	qdisc_put(sch);
 
 903	return NULL;
 904}
 905EXPORT_SYMBOL(qdisc_create_dflt);
 906
 907/* Under qdisc_lock(qdisc) and BH! */
 908
 909void qdisc_reset(struct Qdisc *qdisc)
 910{
 911	const struct Qdisc_ops *ops = qdisc->ops;
 912	struct sk_buff *skb, *tmp;
 913
 914	if (ops->reset)
 915		ops->reset(qdisc);
 916
 917	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 918		__skb_unlink(skb, &qdisc->gso_skb);
 919		kfree_skb_list(skb);
 920	}
 921
 922	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 923		__skb_unlink(skb, &qdisc->skb_bad_txq);
 924		kfree_skb_list(skb);
 925	}
 926
 927	qdisc->q.qlen = 0;
 928	qdisc->qstats.backlog = 0;
 929}
 930EXPORT_SYMBOL(qdisc_reset);
 931
 932void qdisc_free(struct Qdisc *qdisc)
 933{
 934	if (qdisc_is_percpu_stats(qdisc)) {
 935		free_percpu(qdisc->cpu_bstats);
 936		free_percpu(qdisc->cpu_qstats);
 937	}
 938
 939	kfree((char *) qdisc - qdisc->padded);
 940}
 941
 942static void qdisc_free_cb(struct rcu_head *head)
 943{
 944	struct Qdisc *q = container_of(head, struct Qdisc, rcu);
 945
 946	qdisc_free(q);
 947}
 948
 949static void qdisc_destroy(struct Qdisc *qdisc)
 950{
 951	const struct Qdisc_ops  *ops = qdisc->ops;
 952	struct sk_buff *skb, *tmp;
 
 
 
 953
 954#ifdef CONFIG_NET_SCHED
 955	qdisc_hash_del(qdisc);
 956
 957	qdisc_put_stab(rtnl_dereference(qdisc->stab));
 958#endif
 959	gen_kill_estimator(&qdisc->rate_est);
 960	if (ops->reset)
 961		ops->reset(qdisc);
 962	if (ops->destroy)
 963		ops->destroy(qdisc);
 964
 965	module_put(ops->owner);
 966	dev_put(qdisc_dev(qdisc));
 967
 968	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
 969		__skb_unlink(skb, &qdisc->gso_skb);
 970		kfree_skb_list(skb);
 971	}
 972
 973	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
 974		__skb_unlink(skb, &qdisc->skb_bad_txq);
 975		kfree_skb_list(skb);
 976	}
 977
 978	call_rcu(&qdisc->rcu, qdisc_free_cb);
 979}
 980
 981void qdisc_put(struct Qdisc *qdisc)
 982{
 983	if (!qdisc)
 984		return;
 985
 986	if (qdisc->flags & TCQ_F_BUILTIN ||
 987	    !refcount_dec_and_test(&qdisc->refcnt))
 988		return;
 989
 990	qdisc_destroy(qdisc);
 991}
 992EXPORT_SYMBOL(qdisc_put);
 993
 994/* Version of qdisc_put() that is called with rtnl mutex unlocked.
 995 * Intended to be used as optimization, this function only takes rtnl lock if
 996 * qdisc reference counter reached zero.
 997 */
 998
 999void qdisc_put_unlocked(struct Qdisc *qdisc)
1000{
1001	if (qdisc->flags & TCQ_F_BUILTIN ||
1002	    !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1003		return;
1004
1005	qdisc_destroy(qdisc);
1006	rtnl_unlock();
1007}
1008EXPORT_SYMBOL(qdisc_put_unlocked);
1009
1010/* Attach toplevel qdisc to device queue. */
1011struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1012			      struct Qdisc *qdisc)
1013{
1014	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1015	spinlock_t *root_lock;
1016
1017	root_lock = qdisc_lock(oqdisc);
1018	spin_lock_bh(root_lock);
1019
 
 
 
 
1020	/* ... and graft new one */
1021	if (qdisc == NULL)
1022		qdisc = &noop_qdisc;
1023	dev_queue->qdisc_sleeping = qdisc;
1024	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1025
1026	spin_unlock_bh(root_lock);
1027
1028	return oqdisc;
1029}
1030EXPORT_SYMBOL(dev_graft_qdisc);
1031
1032static void attach_one_default_qdisc(struct net_device *dev,
1033				     struct netdev_queue *dev_queue,
1034				     void *_unused)
1035{
1036	struct Qdisc *qdisc;
1037	const struct Qdisc_ops *ops = default_qdisc_ops;
1038
1039	if (dev->priv_flags & IFF_NO_QUEUE)
1040		ops = &noqueue_qdisc_ops;
1041	else if(dev->type == ARPHRD_CAN)
1042		ops = &pfifo_fast_ops;
1043
1044	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1045	if (!qdisc) {
1046		netdev_info(dev, "activation failed\n");
1047		return;
1048	}
1049	if (!netif_is_multiqueue(dev))
1050		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1051	dev_queue->qdisc_sleeping = qdisc;
1052}
1053
1054static void attach_default_qdiscs(struct net_device *dev)
1055{
1056	struct netdev_queue *txq;
1057	struct Qdisc *qdisc;
1058
1059	txq = netdev_get_tx_queue(dev, 0);
1060
1061	if (!netif_is_multiqueue(dev) ||
1062	    dev->priv_flags & IFF_NO_QUEUE) {
1063		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1064		dev->qdisc = txq->qdisc_sleeping;
1065		qdisc_refcount_inc(dev->qdisc);
1066	} else {
1067		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1068		if (qdisc) {
1069			dev->qdisc = qdisc;
1070			qdisc->ops->attach(qdisc);
 
1071		}
1072	}
1073#ifdef CONFIG_NET_SCHED
1074	if (dev->qdisc != &noop_qdisc)
1075		qdisc_hash_add(dev->qdisc, false);
1076#endif
1077}
1078
1079static void transition_one_qdisc(struct net_device *dev,
1080				 struct netdev_queue *dev_queue,
1081				 void *_need_watchdog)
1082{
1083	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1084	int *need_watchdog_p = _need_watchdog;
1085
1086	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1087		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1088
1089	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1090	if (need_watchdog_p) {
1091		dev_queue->trans_start = 0;
1092		*need_watchdog_p = 1;
1093	}
1094}
1095
1096void dev_activate(struct net_device *dev)
1097{
1098	int need_watchdog;
1099
1100	/* No queueing discipline is attached to device;
1101	 * create default one for devices, which need queueing
1102	 * and noqueue_qdisc for virtual interfaces
 
1103	 */
1104
1105	if (dev->qdisc == &noop_qdisc)
1106		attach_default_qdiscs(dev);
1107
1108	if (!netif_carrier_ok(dev))
1109		/* Delay activation until next carrier-on event */
1110		return;
1111
1112	need_watchdog = 0;
1113	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1114	if (dev_ingress_queue(dev))
1115		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1116
1117	if (need_watchdog) {
1118		netif_trans_update(dev);
1119		dev_watchdog_up(dev);
1120	}
1121}
1122EXPORT_SYMBOL(dev_activate);
1123
1124static void dev_deactivate_queue(struct net_device *dev,
1125				 struct netdev_queue *dev_queue,
1126				 void *_qdisc_default)
1127{
1128	struct Qdisc *qdisc_default = _qdisc_default;
1129	struct Qdisc *qdisc;
1130
1131	qdisc = rtnl_dereference(dev_queue->qdisc);
1132	if (qdisc) {
1133		bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1134
1135		if (nolock)
1136			spin_lock_bh(&qdisc->seqlock);
1137		spin_lock_bh(qdisc_lock(qdisc));
1138
1139		if (!(qdisc->flags & TCQ_F_BUILTIN))
1140			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1141
1142		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1143		qdisc_reset(qdisc);
1144
1145		spin_unlock_bh(qdisc_lock(qdisc));
1146		if (nolock)
1147			spin_unlock_bh(&qdisc->seqlock);
1148	}
1149}
1150
1151static bool some_qdisc_is_busy(struct net_device *dev)
1152{
1153	unsigned int i;
1154
1155	for (i = 0; i < dev->num_tx_queues; i++) {
1156		struct netdev_queue *dev_queue;
1157		spinlock_t *root_lock;
1158		struct Qdisc *q;
1159		int val;
1160
1161		dev_queue = netdev_get_tx_queue(dev, i);
1162		q = dev_queue->qdisc_sleeping;
1163
1164		root_lock = qdisc_lock(q);
 
1165		spin_lock_bh(root_lock);
1166
1167		val = (qdisc_is_running(q) ||
1168		       test_bit(__QDISC_STATE_SCHED, &q->state));
1169
1170		spin_unlock_bh(root_lock);
1171
1172		if (val)
1173			return true;
1174	}
1175	return false;
1176}
1177
1178static void dev_qdisc_reset(struct net_device *dev,
1179			    struct netdev_queue *dev_queue,
1180			    void *none)
1181{
1182	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1183
1184	if (qdisc)
1185		qdisc_reset(qdisc);
1186}
1187
1188/**
1189 * 	dev_deactivate_many - deactivate transmissions on several devices
1190 * 	@head: list of devices to deactivate
1191 *
1192 *	This function returns only when all outstanding transmissions
1193 *	have completed, unless all devices are in dismantle phase.
1194 */
1195void dev_deactivate_many(struct list_head *head)
1196{
1197	struct net_device *dev;
 
1198
1199	list_for_each_entry(dev, head, close_list) {
1200		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1201					 &noop_qdisc);
1202		if (dev_ingress_queue(dev))
1203			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1204					     &noop_qdisc);
1205
1206		dev_watchdog_down(dev);
 
1207	}
1208
1209	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
1210	 * This is avoided if all devices are in dismantle phase :
1211	 * Caller will call synchronize_net() for us
1212	 */
1213	synchronize_net();
 
1214
1215	/* Wait for outstanding qdisc_run calls. */
1216	list_for_each_entry(dev, head, close_list) {
1217		while (some_qdisc_is_busy(dev))
1218			yield();
1219		/* The new qdisc is assigned at this point so we can safely
1220		 * unwind stale skb lists and qdisc statistics
1221		 */
1222		netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1223		if (dev_ingress_queue(dev))
1224			dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1225	}
1226}
1227
1228void dev_deactivate(struct net_device *dev)
1229{
1230	LIST_HEAD(single);
1231
1232	list_add(&dev->close_list, &single);
1233	dev_deactivate_many(&single);
1234	list_del(&single);
1235}
1236EXPORT_SYMBOL(dev_deactivate);
1237
1238static int qdisc_change_tx_queue_len(struct net_device *dev,
1239				     struct netdev_queue *dev_queue)
1240{
1241	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1242	const struct Qdisc_ops *ops = qdisc->ops;
1243
1244	if (ops->change_tx_queue_len)
1245		return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1246	return 0;
1247}
1248
1249int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1250{
1251	bool up = dev->flags & IFF_UP;
1252	unsigned int i;
1253	int ret = 0;
1254
1255	if (up)
1256		dev_deactivate(dev);
1257
1258	for (i = 0; i < dev->num_tx_queues; i++) {
1259		ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1260
1261		/* TODO: revert changes on a partial failure */
1262		if (ret)
1263			break;
1264	}
1265
1266	if (up)
1267		dev_activate(dev);
1268	return ret;
1269}
1270
1271static void dev_init_scheduler_queue(struct net_device *dev,
1272				     struct netdev_queue *dev_queue,
1273				     void *_qdisc)
1274{
1275	struct Qdisc *qdisc = _qdisc;
1276
1277	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1278	dev_queue->qdisc_sleeping = qdisc;
1279}
1280
1281void dev_init_scheduler(struct net_device *dev)
1282{
1283	dev->qdisc = &noop_qdisc;
1284	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1285	if (dev_ingress_queue(dev))
1286		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1287
1288	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1289}
1290
1291static void shutdown_scheduler_queue(struct net_device *dev,
1292				     struct netdev_queue *dev_queue,
1293				     void *_qdisc_default)
1294{
1295	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1296	struct Qdisc *qdisc_default = _qdisc_default;
1297
1298	if (qdisc) {
1299		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1300		dev_queue->qdisc_sleeping = qdisc_default;
1301
1302		qdisc_put(qdisc);
1303	}
1304}
1305
1306void dev_shutdown(struct net_device *dev)
1307{
1308	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1309	if (dev_ingress_queue(dev))
1310		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311	qdisc_put(dev->qdisc);
1312	dev->qdisc = &noop_qdisc;
1313
1314	WARN_ON(timer_pending(&dev->watchdog_timer));
1315}
1316
1317void psched_ratecfg_precompute(struct psched_ratecfg *r,
1318			       const struct tc_ratespec *conf,
1319			       u64 rate64)
1320{
1321	memset(r, 0, sizeof(*r));
1322	r->overhead = conf->overhead;
1323	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1324	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1325	r->mult = 1;
1326	/*
1327	 * The deal here is to replace a divide by a reciprocal one
1328	 * in fast path (a reciprocal divide is a multiply and a shift)
1329	 *
1330	 * Normal formula would be :
1331	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1332	 *
1333	 * We compute mult/shift to use instead :
1334	 *  time_in_ns = (len * mult) >> shift;
1335	 *
1336	 * We try to get the highest possible mult value for accuracy,
1337	 * but have to make sure no overflows will ever happen.
1338	 */
1339	if (r->rate_bytes_ps > 0) {
1340		u64 factor = NSEC_PER_SEC;
1341
1342		for (;;) {
1343			r->mult = div64_u64(factor, r->rate_bytes_ps);
1344			if (r->mult & (1U << 31) || factor & (1ULL << 63))
1345				break;
1346			factor <<= 1;
1347			r->shift++;
1348		}
1349	}
1350}
1351EXPORT_SYMBOL(psched_ratecfg_precompute);
1352
1353static void mini_qdisc_rcu_func(struct rcu_head *head)
1354{
1355}
1356
1357void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1358			  struct tcf_proto *tp_head)
1359{
1360	/* Protected with chain0->filter_chain_lock.
1361	 * Can't access chain directly because tp_head can be NULL.
1362	 */
1363	struct mini_Qdisc *miniq_old =
1364		rcu_dereference_protected(*miniqp->p_miniq, 1);
1365	struct mini_Qdisc *miniq;
1366
1367	if (!tp_head) {
1368		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1369		/* Wait for flying RCU callback before it is freed. */
1370		rcu_barrier();
1371		return;
1372	}
1373
1374	miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1375		&miniqp->miniq1 : &miniqp->miniq2;
1376
1377	/* We need to make sure that readers won't see the miniq
1378	 * we are about to modify. So wait until previous call_rcu callback
1379	 * is done.
1380	 */
1381	rcu_barrier();
1382	miniq->filter_list = tp_head;
1383	rcu_assign_pointer(*miniqp->p_miniq, miniq);
1384
1385	if (miniq_old)
1386		/* This is counterpart of the rcu barriers above. We need to
1387		 * block potential new user of miniq_old until all readers
1388		 * are not seeing it.
1389		 */
1390		call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1391}
1392EXPORT_SYMBOL(mini_qdisc_pair_swap);
1393
1394void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1395			  struct mini_Qdisc __rcu **p_miniq)
1396{
1397	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1398	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1399	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1400	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1401	miniqp->p_miniq = p_miniq;
1402}
1403EXPORT_SYMBOL(mini_qdisc_pair_init);
v3.5.6
 
  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 <net/pkt_sched.h>
 29#include <net/dst.h>
 
 
 
 
 
 
 
 30
 31/* Main transmission queue. */
 32
 33/* Modifications to data participating in scheduling must be protected with
 34 * qdisc_lock(qdisc) spinlock.
 35 *
 36 * The idea is the following:
 37 * - enqueue, dequeue are serialized via qdisc root lock
 38 * - ingress filtering is also serialized via qdisc root lock
 39 * - updates to tree and tree walking are only done under the rtnl mutex.
 40 */
 41
 42static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 43{
 44	skb_dst_force(skb);
 45	q->gso_skb = skb;
 46	q->qstats.requeues++;
 47	q->q.qlen++;	/* it's still part of the queue */
 48	__netif_schedule(q);
 49
 50	return 0;
 51}
 52
 53static inline struct sk_buff *dequeue_skb(struct Qdisc *q)
 
 54{
 55	struct sk_buff *skb = q->gso_skb;
 56
 57	if (unlikely(skb)) {
 58		struct net_device *dev = qdisc_dev(q);
 59		struct netdev_queue *txq;
 
 
 
 60
 61		/* check the reason of requeuing without tx lock first */
 62		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 63		if (!netif_xmit_frozen_or_stopped(txq)) {
 64			q->gso_skb = NULL;
 65			q->q.qlen--;
 66		} else
 67			skb = NULL;
 68	} else {
 69		skb = q->dequeue(q);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 70	}
 71
 72	return skb;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 73}
 74
 75static inline int handle_dev_cpu_collision(struct sk_buff *skb,
 76					   struct netdev_queue *dev_queue,
 77					   struct Qdisc *q)
 
 
 78{
 79	int ret;
 
 80
 81	if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
 82		/*
 83		 * Same CPU holding the lock. It may be a transient
 84		 * configuration error, when hard_start_xmit() recurses. We
 85		 * detect it by checking xmit owner and drop the packet when
 86		 * deadloop is detected. Return OK to try the next skb.
 
 
 
 
 
 
 
 87		 */
 88		kfree_skb(skb);
 89		net_warn_ratelimited("Dead loop on netdevice %s, fix it urgently!\n",
 90				     dev_queue->dev->name);
 91		ret = qdisc_qlen(q);
 92	} else {
 93		/*
 94		 * Another cpu is holding lock, requeue & delay xmits for
 95		 * some time.
 96		 */
 97		__this_cpu_inc(softnet_data.cpu_collision);
 98		ret = dev_requeue_skb(skb, q);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 99	}
 
 
100
101	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102}
103
104/*
105 * Transmit one skb, and handle the return status as required. Holding the
106 * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this
107 * function.
108 *
109 * Returns to the caller:
110 *				0  - queue is empty or throttled.
111 *				>0 - queue is not empty.
112 */
113int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
114		    struct net_device *dev, struct netdev_queue *txq,
115		    spinlock_t *root_lock)
116{
117	int ret = NETDEV_TX_BUSY;
 
118
119	/* And release qdisc */
120	spin_unlock(root_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
121
122	HARD_TX_LOCK(dev, txq, smp_processor_id());
123	if (!netif_xmit_frozen_or_stopped(txq))
124		ret = dev_hard_start_xmit(skb, dev, txq);
125
126	HARD_TX_UNLOCK(dev, txq);
127
128	spin_lock(root_lock);
129
130	if (dev_xmit_complete(ret)) {
131		/* Driver sent out skb successfully or skb was consumed */
132		ret = qdisc_qlen(q);
133	} else if (ret == NETDEV_TX_LOCKED) {
134		/* Driver try lock failed */
135		ret = handle_dev_cpu_collision(skb, txq, q);
136	} else {
 
 
 
 
 
 
 
 
 
137		/* Driver returned NETDEV_TX_BUSY - requeue skb */
138		if (unlikely(ret != NETDEV_TX_BUSY))
139			net_warn_ratelimited("BUG %s code %d qlen %d\n",
140					     dev->name, ret, q->q.qlen);
141
142		ret = dev_requeue_skb(skb, q);
 
143	}
144
145	if (ret && netif_xmit_frozen_or_stopped(txq))
146		ret = 0;
147
148	return ret;
149}
150
151/*
152 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
153 *
154 * __QDISC_STATE_RUNNING guarantees only one CPU can process
155 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
156 * this queue.
157 *
158 *  netif_tx_lock serializes accesses to device driver.
159 *
160 *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
161 *  if one is grabbed, another must be free.
162 *
163 * Note, that this procedure can be called by a watchdog timer
164 *
165 * Returns to the caller:
166 *				0  - queue is empty or throttled.
167 *				>0 - queue is not empty.
168 *
169 */
170static inline int qdisc_restart(struct Qdisc *q)
171{
 
172	struct netdev_queue *txq;
173	struct net_device *dev;
174	spinlock_t *root_lock;
175	struct sk_buff *skb;
 
176
177	/* Dequeue packet */
178	skb = dequeue_skb(q);
179	if (unlikely(!skb))
180		return 0;
181	WARN_ON_ONCE(skb_dst_is_noref(skb));
182	root_lock = qdisc_lock(q);
 
 
183	dev = qdisc_dev(q);
184	txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
185
186	return sch_direct_xmit(skb, q, dev, txq, root_lock);
187}
188
189void __qdisc_run(struct Qdisc *q)
190{
191	int quota = weight_p;
 
192
193	while (qdisc_restart(q)) {
194		/*
195		 * Ordered by possible occurrence: Postpone processing if
196		 * 1. we've exceeded packet quota
197		 * 2. another process needs the CPU;
198		 */
199		if (--quota <= 0 || need_resched()) {
 
200			__netif_schedule(q);
201			break;
202		}
203	}
204
205	qdisc_run_end(q);
206}
207
208unsigned long dev_trans_start(struct net_device *dev)
209{
210	unsigned long val, res = dev->trans_start;
211	unsigned int i;
212
213	for (i = 0; i < dev->num_tx_queues; i++) {
 
 
 
 
 
214		val = netdev_get_tx_queue(dev, i)->trans_start;
215		if (val && time_after(val, res))
216			res = val;
217	}
218	dev->trans_start = res;
219	return res;
220}
221EXPORT_SYMBOL(dev_trans_start);
222
223static void dev_watchdog(unsigned long arg)
224{
225	struct net_device *dev = (struct net_device *)arg;
226
227	netif_tx_lock(dev);
228	if (!qdisc_tx_is_noop(dev)) {
229		if (netif_device_present(dev) &&
230		    netif_running(dev) &&
231		    netif_carrier_ok(dev)) {
232			int some_queue_timedout = 0;
233			unsigned int i;
234			unsigned long trans_start;
235
236			for (i = 0; i < dev->num_tx_queues; i++) {
237				struct netdev_queue *txq;
238
239				txq = netdev_get_tx_queue(dev, i);
240				/*
241				 * old device drivers set dev->trans_start
242				 */
243				trans_start = txq->trans_start ? : dev->trans_start;
244				if (netif_xmit_stopped(txq) &&
245				    time_after(jiffies, (trans_start +
246							 dev->watchdog_timeo))) {
247					some_queue_timedout = 1;
248					txq->trans_timeout++;
249					break;
250				}
251			}
252
253			if (some_queue_timedout) {
 
254				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
255				       dev->name, netdev_drivername(dev), i);
256				dev->netdev_ops->ndo_tx_timeout(dev);
257			}
258			if (!mod_timer(&dev->watchdog_timer,
259				       round_jiffies(jiffies +
260						     dev->watchdog_timeo)))
261				dev_hold(dev);
262		}
263	}
264	netif_tx_unlock(dev);
265
266	dev_put(dev);
267}
268
269void __netdev_watchdog_up(struct net_device *dev)
270{
271	if (dev->netdev_ops->ndo_tx_timeout) {
272		if (dev->watchdog_timeo <= 0)
273			dev->watchdog_timeo = 5*HZ;
274		if (!mod_timer(&dev->watchdog_timer,
275			       round_jiffies(jiffies + dev->watchdog_timeo)))
276			dev_hold(dev);
277	}
278}
279
280static void dev_watchdog_up(struct net_device *dev)
281{
282	__netdev_watchdog_up(dev);
283}
284
285static void dev_watchdog_down(struct net_device *dev)
286{
287	netif_tx_lock_bh(dev);
288	if (del_timer(&dev->watchdog_timer))
289		dev_put(dev);
290	netif_tx_unlock_bh(dev);
291}
292
293/**
294 *	netif_carrier_on - set carrier
295 *	@dev: network device
296 *
297 * Device has detected that carrier.
298 */
299void netif_carrier_on(struct net_device *dev)
300{
301	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
302		if (dev->reg_state == NETREG_UNINITIALIZED)
303			return;
 
304		linkwatch_fire_event(dev);
305		if (netif_running(dev))
306			__netdev_watchdog_up(dev);
307	}
308}
309EXPORT_SYMBOL(netif_carrier_on);
310
311/**
312 *	netif_carrier_off - clear carrier
313 *	@dev: network device
314 *
315 * Device has detected loss of carrier.
316 */
317void netif_carrier_off(struct net_device *dev)
318{
319	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
320		if (dev->reg_state == NETREG_UNINITIALIZED)
321			return;
 
322		linkwatch_fire_event(dev);
323	}
324}
325EXPORT_SYMBOL(netif_carrier_off);
326
327/**
328 * 	netif_notify_peers - notify network peers about existence of @dev
329 * 	@dev: network device
330 *
331 * Generate traffic such that interested network peers are aware of
332 * @dev, such as by generating a gratuitous ARP. This may be used when
333 * a device wants to inform the rest of the network about some sort of
334 * reconfiguration such as a failover event or virtual machine
335 * migration.
336 */
337void netif_notify_peers(struct net_device *dev)
338{
339	rtnl_lock();
340	call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
341	rtnl_unlock();
342}
343EXPORT_SYMBOL(netif_notify_peers);
344
345/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
346   under all circumstances. It is difficult to invent anything faster or
347   cheaper.
348 */
349
350static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 
351{
352	kfree_skb(skb);
353	return NET_XMIT_CN;
354}
355
356static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
357{
358	return NULL;
359}
360
361struct Qdisc_ops noop_qdisc_ops __read_mostly = {
362	.id		=	"noop",
363	.priv_size	=	0,
364	.enqueue	=	noop_enqueue,
365	.dequeue	=	noop_dequeue,
366	.peek		=	noop_dequeue,
367	.owner		=	THIS_MODULE,
368};
369
370static struct netdev_queue noop_netdev_queue = {
371	.qdisc		=	&noop_qdisc,
372	.qdisc_sleeping	=	&noop_qdisc,
373};
374
375struct Qdisc noop_qdisc = {
376	.enqueue	=	noop_enqueue,
377	.dequeue	=	noop_dequeue,
378	.flags		=	TCQ_F_BUILTIN,
379	.ops		=	&noop_qdisc_ops,
380	.list		=	LIST_HEAD_INIT(noop_qdisc.list),
381	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
382	.dev_queue	=	&noop_netdev_queue,
 
383	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
 
 
 
 
 
 
 
 
 
 
 
 
384};
385EXPORT_SYMBOL(noop_qdisc);
386
387static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
 
 
 
 
 
 
 
 
 
 
388	.id		=	"noqueue",
389	.priv_size	=	0,
 
390	.enqueue	=	noop_enqueue,
391	.dequeue	=	noop_dequeue,
392	.peek		=	noop_dequeue,
393	.owner		=	THIS_MODULE,
394};
395
396static struct Qdisc noqueue_qdisc;
397static struct netdev_queue noqueue_netdev_queue = {
398	.qdisc		=	&noqueue_qdisc,
399	.qdisc_sleeping	=	&noqueue_qdisc,
400};
401
402static struct Qdisc noqueue_qdisc = {
403	.enqueue	=	NULL,
404	.dequeue	=	noop_dequeue,
405	.flags		=	TCQ_F_BUILTIN,
406	.ops		=	&noqueue_qdisc_ops,
407	.list		=	LIST_HEAD_INIT(noqueue_qdisc.list),
408	.q.lock		=	__SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock),
409	.dev_queue	=	&noqueue_netdev_queue,
410	.busylock	=	__SPIN_LOCK_UNLOCKED(noqueue_qdisc.busylock),
411};
412
413
414static const u8 prio2band[TC_PRIO_MAX + 1] = {
415	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
416};
417
418/* 3-band FIFO queue: old style, but should be a bit faster than
419   generic prio+fifo combination.
420 */
421
422#define PFIFO_FAST_BANDS 3
423
424/*
425 * Private data for a pfifo_fast scheduler containing:
426 * 	- queues for the three band
427 * 	- bitmap indicating which of the bands contain skbs
428 */
429struct pfifo_fast_priv {
430	u32 bitmap;
431	struct sk_buff_head q[PFIFO_FAST_BANDS];
432};
433
434/*
435 * Convert a bitmap to the first band number where an skb is queued, where:
436 * 	bitmap=0 means there are no skbs on any band.
437 * 	bitmap=1 means there is an skb on band 0.
438 *	bitmap=7 means there are skbs on all 3 bands, etc.
439 */
440static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0};
441
442static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv,
443					     int band)
444{
445	return priv->q + band;
446}
447
448static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc)
 
449{
450	if (skb_queue_len(&qdisc->q) < qdisc_dev(qdisc)->tx_queue_len) {
451		int band = prio2band[skb->priority & TC_PRIO_MAX];
452		struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
453		struct sk_buff_head *list = band2list(priv, band);
 
454
455		priv->bitmap |= (1 << band);
456		qdisc->q.qlen++;
457		return __qdisc_enqueue_tail(skb, qdisc, list);
 
 
 
 
458	}
459
460	return qdisc_drop(skb, qdisc);
 
461}
462
463static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
464{
465	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
466	int band = bitmap2band[priv->bitmap];
 
 
 
 
467
468	if (likely(band >= 0)) {
469		struct sk_buff_head *list = band2list(priv, band);
470		struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list);
471
472		qdisc->q.qlen--;
473		if (skb_queue_empty(list))
474			priv->bitmap &= ~(1 << band);
475
476		return skb;
 
 
 
 
 
477	}
478
479	return NULL;
480}
481
482static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
483{
484	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
485	int band = bitmap2band[priv->bitmap];
 
486
487	if (band >= 0) {
488		struct sk_buff_head *list = band2list(priv, band);
489
490		return skb_peek(list);
491	}
492
493	return NULL;
494}
495
496static void pfifo_fast_reset(struct Qdisc *qdisc)
497{
498	int prio;
499	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
500
501	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
502		__qdisc_reset_queue(qdisc, band2list(priv, prio));
 
503
504	priv->bitmap = 0;
505	qdisc->qstats.backlog = 0;
506	qdisc->q.qlen = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
507}
508
509static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
510{
511	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
512
513	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
514	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
515		goto nla_put_failure;
516	return skb->len;
517
518nla_put_failure:
519	return -1;
520}
521
522static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt)
 
523{
 
 
524	int prio;
525	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
526
527	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
528		skb_queue_head_init(band2list(priv, prio));
 
 
 
 
 
 
 
 
 
 
529
530	/* Can by-pass the queue discipline */
531	qdisc->flags |= TCQ_F_CAN_BYPASS;
532	return 0;
533}
534
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
535struct Qdisc_ops pfifo_fast_ops __read_mostly = {
536	.id		=	"pfifo_fast",
537	.priv_size	=	sizeof(struct pfifo_fast_priv),
538	.enqueue	=	pfifo_fast_enqueue,
539	.dequeue	=	pfifo_fast_dequeue,
540	.peek		=	pfifo_fast_peek,
541	.init		=	pfifo_fast_init,
 
542	.reset		=	pfifo_fast_reset,
543	.dump		=	pfifo_fast_dump,
 
544	.owner		=	THIS_MODULE,
 
545};
546EXPORT_SYMBOL(pfifo_fast_ops);
547
548struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
549			  struct Qdisc_ops *ops)
 
550{
551	void *p;
552	struct Qdisc *sch;
553	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
554	int err = -ENOBUFS;
 
555
 
 
 
 
 
 
 
556	p = kzalloc_node(size, GFP_KERNEL,
557			 netdev_queue_numa_node_read(dev_queue));
558
559	if (!p)
560		goto errout;
561	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
562	/* if we got non aligned memory, ask more and do alignment ourself */
563	if (sch != p) {
564		kfree(p);
565		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
566				 netdev_queue_numa_node_read(dev_queue));
567		if (!p)
568			goto errout;
569		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
570		sch->padded = (char *) sch - (char *) p;
571	}
572	INIT_LIST_HEAD(&sch->list);
573	skb_queue_head_init(&sch->q);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
574	spin_lock_init(&sch->busylock);
 
 
 
 
575	sch->ops = ops;
 
576	sch->enqueue = ops->enqueue;
577	sch->dequeue = ops->dequeue;
578	sch->dev_queue = dev_queue;
579	dev_hold(qdisc_dev(sch));
580	atomic_set(&sch->refcnt, 1);
 
 
 
 
 
 
 
581
582	return sch;
 
 
583errout:
584	return ERR_PTR(err);
585}
586
587struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
588				struct Qdisc_ops *ops, unsigned int parentid)
 
 
589{
590	struct Qdisc *sch;
591
592	sch = qdisc_alloc(dev_queue, ops);
593	if (IS_ERR(sch))
594		goto errout;
 
 
 
 
 
 
 
595	sch->parent = parentid;
596
597	if (!ops->init || ops->init(sch, NULL) == 0)
598		return sch;
599
600	qdisc_destroy(sch);
601errout:
602	return NULL;
603}
604EXPORT_SYMBOL(qdisc_create_dflt);
605
606/* Under qdisc_lock(qdisc) and BH! */
607
608void qdisc_reset(struct Qdisc *qdisc)
609{
610	const struct Qdisc_ops *ops = qdisc->ops;
 
611
612	if (ops->reset)
613		ops->reset(qdisc);
614
615	if (qdisc->gso_skb) {
616		kfree_skb(qdisc->gso_skb);
617		qdisc->gso_skb = NULL;
618		qdisc->q.qlen = 0;
 
 
 
 
619	}
 
 
 
620}
621EXPORT_SYMBOL(qdisc_reset);
622
623static void qdisc_rcu_free(struct rcu_head *head)
624{
625	struct Qdisc *qdisc = container_of(head, struct Qdisc, rcu_head);
 
 
 
626
627	kfree((char *) qdisc - qdisc->padded);
628}
629
630void qdisc_destroy(struct Qdisc *qdisc)
 
 
 
 
 
 
 
631{
632	const struct Qdisc_ops  *ops = qdisc->ops;
633
634	if (qdisc->flags & TCQ_F_BUILTIN ||
635	    !atomic_dec_and_test(&qdisc->refcnt))
636		return;
637
638#ifdef CONFIG_NET_SCHED
639	qdisc_list_del(qdisc);
640
641	qdisc_put_stab(rtnl_dereference(qdisc->stab));
642#endif
643	gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
644	if (ops->reset)
645		ops->reset(qdisc);
646	if (ops->destroy)
647		ops->destroy(qdisc);
648
649	module_put(ops->owner);
650	dev_put(qdisc_dev(qdisc));
651
652	kfree_skb(qdisc->gso_skb);
653	/*
654	 * gen_estimator est_timer() might access qdisc->q.lock,
655	 * wait a RCU grace period before freeing qdisc.
656	 */
657	call_rcu(&qdisc->rcu_head, qdisc_rcu_free);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
658}
659EXPORT_SYMBOL(qdisc_destroy);
660
661/* Attach toplevel qdisc to device queue. */
662struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
663			      struct Qdisc *qdisc)
664{
665	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
666	spinlock_t *root_lock;
667
668	root_lock = qdisc_lock(oqdisc);
669	spin_lock_bh(root_lock);
670
671	/* Prune old scheduler */
672	if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1)
673		qdisc_reset(oqdisc);
674
675	/* ... and graft new one */
676	if (qdisc == NULL)
677		qdisc = &noop_qdisc;
678	dev_queue->qdisc_sleeping = qdisc;
679	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
680
681	spin_unlock_bh(root_lock);
682
683	return oqdisc;
684}
685EXPORT_SYMBOL(dev_graft_qdisc);
686
687static void attach_one_default_qdisc(struct net_device *dev,
688				     struct netdev_queue *dev_queue,
689				     void *_unused)
690{
691	struct Qdisc *qdisc = &noqueue_qdisc;
 
692
693	if (dev->tx_queue_len) {
694		qdisc = qdisc_create_dflt(dev_queue,
695					  &pfifo_fast_ops, TC_H_ROOT);
696		if (!qdisc) {
697			netdev_info(dev, "activation failed\n");
698			return;
699		}
 
 
700	}
 
 
701	dev_queue->qdisc_sleeping = qdisc;
702}
703
704static void attach_default_qdiscs(struct net_device *dev)
705{
706	struct netdev_queue *txq;
707	struct Qdisc *qdisc;
708
709	txq = netdev_get_tx_queue(dev, 0);
710
711	if (!netif_is_multiqueue(dev) || dev->tx_queue_len == 0) {
 
712		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
713		dev->qdisc = txq->qdisc_sleeping;
714		atomic_inc(&dev->qdisc->refcnt);
715	} else {
716		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
717		if (qdisc) {
 
718			qdisc->ops->attach(qdisc);
719			dev->qdisc = qdisc;
720		}
721	}
 
 
 
 
722}
723
724static void transition_one_qdisc(struct net_device *dev,
725				 struct netdev_queue *dev_queue,
726				 void *_need_watchdog)
727{
728	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
729	int *need_watchdog_p = _need_watchdog;
730
731	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
732		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
733
734	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
735	if (need_watchdog_p && new_qdisc != &noqueue_qdisc) {
736		dev_queue->trans_start = 0;
737		*need_watchdog_p = 1;
738	}
739}
740
741void dev_activate(struct net_device *dev)
742{
743	int need_watchdog;
744
745	/* No queueing discipline is attached to device;
746	   create default one i.e. pfifo_fast for devices,
747	   which need queueing and noqueue_qdisc for
748	   virtual interfaces
749	 */
750
751	if (dev->qdisc == &noop_qdisc)
752		attach_default_qdiscs(dev);
753
754	if (!netif_carrier_ok(dev))
755		/* Delay activation until next carrier-on event */
756		return;
757
758	need_watchdog = 0;
759	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
760	if (dev_ingress_queue(dev))
761		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
762
763	if (need_watchdog) {
764		dev->trans_start = jiffies;
765		dev_watchdog_up(dev);
766	}
767}
768EXPORT_SYMBOL(dev_activate);
769
770static void dev_deactivate_queue(struct net_device *dev,
771				 struct netdev_queue *dev_queue,
772				 void *_qdisc_default)
773{
774	struct Qdisc *qdisc_default = _qdisc_default;
775	struct Qdisc *qdisc;
776
777	qdisc = dev_queue->qdisc;
778	if (qdisc) {
 
 
 
 
779		spin_lock_bh(qdisc_lock(qdisc));
780
781		if (!(qdisc->flags & TCQ_F_BUILTIN))
782			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
783
784		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
785		qdisc_reset(qdisc);
786
787		spin_unlock_bh(qdisc_lock(qdisc));
 
 
788	}
789}
790
791static bool some_qdisc_is_busy(struct net_device *dev)
792{
793	unsigned int i;
794
795	for (i = 0; i < dev->num_tx_queues; i++) {
796		struct netdev_queue *dev_queue;
797		spinlock_t *root_lock;
798		struct Qdisc *q;
799		int val;
800
801		dev_queue = netdev_get_tx_queue(dev, i);
802		q = dev_queue->qdisc_sleeping;
 
803		root_lock = qdisc_lock(q);
804
805		spin_lock_bh(root_lock);
806
807		val = (qdisc_is_running(q) ||
808		       test_bit(__QDISC_STATE_SCHED, &q->state));
809
810		spin_unlock_bh(root_lock);
811
812		if (val)
813			return true;
814	}
815	return false;
816}
817
 
 
 
 
 
 
 
 
 
 
818/**
819 * 	dev_deactivate_many - deactivate transmissions on several devices
820 * 	@head: list of devices to deactivate
821 *
822 *	This function returns only when all outstanding transmissions
823 *	have completed, unless all devices are in dismantle phase.
824 */
825void dev_deactivate_many(struct list_head *head)
826{
827	struct net_device *dev;
828	bool sync_needed = false;
829
830	list_for_each_entry(dev, head, unreg_list) {
831		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
832					 &noop_qdisc);
833		if (dev_ingress_queue(dev))
834			dev_deactivate_queue(dev, dev_ingress_queue(dev),
835					     &noop_qdisc);
836
837		dev_watchdog_down(dev);
838		sync_needed |= !dev->dismantle;
839	}
840
841	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
842	 * This is avoided if all devices are in dismantle phase :
843	 * Caller will call synchronize_net() for us
844	 */
845	if (sync_needed)
846		synchronize_net();
847
848	/* Wait for outstanding qdisc_run calls. */
849	list_for_each_entry(dev, head, unreg_list)
850		while (some_qdisc_is_busy(dev))
851			yield();
 
 
 
 
 
 
 
852}
853
854void dev_deactivate(struct net_device *dev)
855{
856	LIST_HEAD(single);
857
858	list_add(&dev->unreg_list, &single);
859	dev_deactivate_many(&single);
860	list_del(&single);
861}
862EXPORT_SYMBOL(dev_deactivate);
863
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
864static void dev_init_scheduler_queue(struct net_device *dev,
865				     struct netdev_queue *dev_queue,
866				     void *_qdisc)
867{
868	struct Qdisc *qdisc = _qdisc;
869
870	dev_queue->qdisc = qdisc;
871	dev_queue->qdisc_sleeping = qdisc;
872}
873
874void dev_init_scheduler(struct net_device *dev)
875{
876	dev->qdisc = &noop_qdisc;
877	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
878	if (dev_ingress_queue(dev))
879		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
880
881	setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
882}
883
884static void shutdown_scheduler_queue(struct net_device *dev,
885				     struct netdev_queue *dev_queue,
886				     void *_qdisc_default)
887{
888	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
889	struct Qdisc *qdisc_default = _qdisc_default;
890
891	if (qdisc) {
892		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
893		dev_queue->qdisc_sleeping = qdisc_default;
894
895		qdisc_destroy(qdisc);
896	}
897}
898
899void dev_shutdown(struct net_device *dev)
900{
901	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
902	if (dev_ingress_queue(dev))
903		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
904	qdisc_destroy(dev->qdisc);
905	dev->qdisc = &noop_qdisc;
906
907	WARN_ON(timer_pending(&dev->watchdog_timer));
908}