Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 *  Block device elevator/IO-scheduler.
   3 *
   4 *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5 *
   6 * 30042000 Jens Axboe <axboe@kernel.dk> :
   7 *
   8 * Split the elevator a bit so that it is possible to choose a different
   9 * one or even write a new "plug in". There are three pieces:
  10 * - elevator_fn, inserts a new request in the queue list
  11 * - elevator_merge_fn, decides whether a new buffer can be merged with
  12 *   an existing request
  13 * - elevator_dequeue_fn, called when a request is taken off the active list
  14 *
  15 * 20082000 Dave Jones <davej@suse.de> :
  16 * Removed tests for max-bomb-segments, which was breaking elvtune
  17 *  when run without -bN
  18 *
  19 * Jens:
  20 * - Rework again to work with bio instead of buffer_heads
  21 * - loose bi_dev comparisons, partition handling is right now
  22 * - completely modularize elevator setup and teardown
  23 *
  24 */
  25#include <linux/kernel.h>
  26#include <linux/fs.h>
  27#include <linux/blkdev.h>
  28#include <linux/elevator.h>
  29#include <linux/bio.h>
  30#include <linux/module.h>
  31#include <linux/slab.h>
  32#include <linux/init.h>
  33#include <linux/compiler.h>
  34#include <linux/blktrace_api.h>
  35#include <linux/hash.h>
  36#include <linux/uaccess.h>
  37#include <linux/pm_runtime.h>
  38#include <linux/blk-cgroup.h>
  39
  40#include <trace/events/block.h>
  41
  42#include "blk.h"
 
 
 
  43
  44static DEFINE_SPINLOCK(elv_list_lock);
  45static LIST_HEAD(elv_list);
  46
  47/*
  48 * Merge hash stuff.
  49 */
  50#define rq_hash_key(rq)		(blk_rq_pos(rq) + blk_rq_sectors(rq))
  51
  52/*
  53 * Query io scheduler to see if the current process issuing bio may be
  54 * merged with rq.
  55 */
  56static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  57{
  58	struct request_queue *q = rq->q;
  59	struct elevator_queue *e = q->elevator;
  60
  61	if (e->type->ops.elevator_allow_merge_fn)
  62		return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
  63
  64	return 1;
  65}
  66
  67/*
  68 * can we safely merge with this request?
  69 */
  70bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
  71{
  72	if (!blk_rq_merge_ok(rq, bio))
  73		return 0;
  74
  75	if (!elv_iosched_allow_merge(rq, bio))
  76		return 0;
  77
  78	return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  79}
  80EXPORT_SYMBOL(elv_rq_merge_ok);
  81
  82static struct elevator_type *elevator_find(const char *name)
 
 
 
 
 
 
 
 
 
  83{
  84	struct elevator_type *e;
  85
  86	list_for_each_entry(e, &elv_list, list) {
  87		if (!strcmp(e->elevator_name, name))
  88			return e;
  89	}
  90
  91	return NULL;
  92}
  93
  94static void elevator_put(struct elevator_type *e)
  95{
  96	module_put(e->elevator_owner);
  97}
  98
  99static struct elevator_type *elevator_get(const char *name, bool try_loading)
 
 100{
 101	struct elevator_type *e;
 102
 103	spin_lock(&elv_list_lock);
 104
 105	e = elevator_find(name);
 106	if (!e && try_loading) {
 107		spin_unlock(&elv_list_lock);
 108		request_module("%s-iosched", name);
 109		spin_lock(&elv_list_lock);
 110		e = elevator_find(name);
 111	}
 112
 113	if (e && !try_module_get(e->elevator_owner))
 114		e = NULL;
 115
 116	spin_unlock(&elv_list_lock);
 117
 118	return e;
 119}
 120
 121static char chosen_elevator[ELV_NAME_MAX];
 122
 123static int __init elevator_setup(char *str)
 124{
 125	/*
 126	 * Be backwards-compatible with previous kernels, so users
 127	 * won't get the wrong elevator.
 128	 */
 129	strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 130	return 1;
 131}
 132
 133__setup("elevator=", elevator_setup);
 134
 135/* called during boot to load the elevator chosen by the elevator param */
 136void __init load_default_elevator_module(void)
 137{
 138	struct elevator_type *e;
 139
 140	if (!chosen_elevator[0])
 141		return;
 142
 143	spin_lock(&elv_list_lock);
 144	e = elevator_find(chosen_elevator);
 145	spin_unlock(&elv_list_lock);
 146
 147	if (!e)
 148		request_module("%s-iosched", chosen_elevator);
 149}
 150
 151static struct kobj_type elv_ktype;
 152
 153struct elevator_queue *elevator_alloc(struct request_queue *q,
 154				  struct elevator_type *e)
 155{
 156	struct elevator_queue *eq;
 157
 158	eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
 159	if (unlikely(!eq))
 160		return NULL;
 161
 162	eq->type = e;
 163	kobject_init(&eq->kobj, &elv_ktype);
 164	mutex_init(&eq->sysfs_lock);
 165	hash_init(eq->hash);
 166
 167	return eq;
 168}
 169EXPORT_SYMBOL(elevator_alloc);
 170
 171static void elevator_release(struct kobject *kobj)
 172{
 173	struct elevator_queue *e;
 174
 175	e = container_of(kobj, struct elevator_queue, kobj);
 176	elevator_put(e->type);
 177	kfree(e);
 178}
 179
 180int elevator_init(struct request_queue *q, char *name)
 181{
 182	struct elevator_type *e = NULL;
 183	int err;
 184
 185	/*
 186	 * q->sysfs_lock must be held to provide mutual exclusion between
 187	 * elevator_switch() and here.
 188	 */
 189	lockdep_assert_held(&q->sysfs_lock);
 190
 191	if (unlikely(q->elevator))
 192		return 0;
 193
 194	INIT_LIST_HEAD(&q->queue_head);
 195	q->last_merge = NULL;
 196	q->end_sector = 0;
 197	q->boundary_rq = NULL;
 198
 199	if (name) {
 200		e = elevator_get(name, true);
 201		if (!e)
 202			return -EINVAL;
 203	}
 204
 205	/*
 206	 * Use the default elevator specified by config boot param or
 207	 * config option.  Don't try to load modules as we could be running
 208	 * off async and request_module() isn't allowed from async.
 209	 */
 210	if (!e && *chosen_elevator) {
 211		e = elevator_get(chosen_elevator, false);
 212		if (!e)
 213			printk(KERN_ERR "I/O scheduler %s not found\n",
 214							chosen_elevator);
 215	}
 216
 217	if (!e) {
 218		e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
 219		if (!e) {
 220			printk(KERN_ERR
 221				"Default I/O scheduler not found. " \
 222				"Using noop.\n");
 223			e = elevator_get("noop", false);
 224		}
 225	}
 226
 227	err = e->ops.elevator_init_fn(q, e);
 228	if (err)
 229		elevator_put(e);
 230	return err;
 231}
 232EXPORT_SYMBOL(elevator_init);
 233
 234void elevator_exit(struct elevator_queue *e)
 235{
 236	mutex_lock(&e->sysfs_lock);
 237	if (e->type->ops.elevator_exit_fn)
 238		e->type->ops.elevator_exit_fn(e);
 239	mutex_unlock(&e->sysfs_lock);
 240
 241	kobject_put(&e->kobj);
 242}
 243EXPORT_SYMBOL(elevator_exit);
 244
 245static inline void __elv_rqhash_del(struct request *rq)
 246{
 247	hash_del(&rq->hash);
 248	rq->cmd_flags &= ~REQ_HASHED;
 249}
 250
 251static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 252{
 253	if (ELV_ON_HASH(rq))
 254		__elv_rqhash_del(rq);
 255}
 
 256
 257static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 258{
 259	struct elevator_queue *e = q->elevator;
 260
 261	BUG_ON(ELV_ON_HASH(rq));
 262	hash_add(e->hash, &rq->hash, rq_hash_key(rq));
 263	rq->cmd_flags |= REQ_HASHED;
 264}
 
 265
 266static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 267{
 268	__elv_rqhash_del(rq);
 269	elv_rqhash_add(q, rq);
 270}
 271
 272static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 273{
 274	struct elevator_queue *e = q->elevator;
 275	struct hlist_node *next;
 276	struct request *rq;
 277
 278	hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
 279		BUG_ON(!ELV_ON_HASH(rq));
 280
 281		if (unlikely(!rq_mergeable(rq))) {
 282			__elv_rqhash_del(rq);
 283			continue;
 284		}
 285
 286		if (rq_hash_key(rq) == offset)
 287			return rq;
 288	}
 289
 290	return NULL;
 291}
 292
 293/*
 294 * RB-tree support functions for inserting/lookup/removal of requests
 295 * in a sorted RB tree.
 296 */
 297void elv_rb_add(struct rb_root *root, struct request *rq)
 298{
 299	struct rb_node **p = &root->rb_node;
 300	struct rb_node *parent = NULL;
 301	struct request *__rq;
 302
 303	while (*p) {
 304		parent = *p;
 305		__rq = rb_entry(parent, struct request, rb_node);
 306
 307		if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 308			p = &(*p)->rb_left;
 309		else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 310			p = &(*p)->rb_right;
 311	}
 312
 313	rb_link_node(&rq->rb_node, parent, p);
 314	rb_insert_color(&rq->rb_node, root);
 315}
 316EXPORT_SYMBOL(elv_rb_add);
 317
 318void elv_rb_del(struct rb_root *root, struct request *rq)
 319{
 320	BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 321	rb_erase(&rq->rb_node, root);
 322	RB_CLEAR_NODE(&rq->rb_node);
 323}
 324EXPORT_SYMBOL(elv_rb_del);
 325
 326struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 327{
 328	struct rb_node *n = root->rb_node;
 329	struct request *rq;
 330
 331	while (n) {
 332		rq = rb_entry(n, struct request, rb_node);
 333
 334		if (sector < blk_rq_pos(rq))
 335			n = n->rb_left;
 336		else if (sector > blk_rq_pos(rq))
 337			n = n->rb_right;
 338		else
 339			return rq;
 340	}
 341
 342	return NULL;
 343}
 344EXPORT_SYMBOL(elv_rb_find);
 345
 346/*
 347 * Insert rq into dispatch queue of q.  Queue lock must be held on
 348 * entry.  rq is sort instead into the dispatch queue. To be used by
 349 * specific elevators.
 350 */
 351void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 352{
 353	sector_t boundary;
 354	struct list_head *entry;
 355	int stop_flags;
 356
 357	if (q->last_merge == rq)
 358		q->last_merge = NULL;
 359
 360	elv_rqhash_del(q, rq);
 361
 362	q->nr_sorted--;
 363
 364	boundary = q->end_sector;
 365	stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
 366	list_for_each_prev(entry, &q->queue_head) {
 367		struct request *pos = list_entry_rq(entry);
 368
 369		if ((rq->cmd_flags & REQ_DISCARD) !=
 370		    (pos->cmd_flags & REQ_DISCARD))
 371			break;
 372		if (rq_data_dir(rq) != rq_data_dir(pos))
 373			break;
 374		if (pos->cmd_flags & stop_flags)
 375			break;
 376		if (blk_rq_pos(rq) >= boundary) {
 377			if (blk_rq_pos(pos) < boundary)
 378				continue;
 379		} else {
 380			if (blk_rq_pos(pos) >= boundary)
 381				break;
 382		}
 383		if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 384			break;
 385	}
 386
 387	list_add(&rq->queuelist, entry);
 388}
 389EXPORT_SYMBOL(elv_dispatch_sort);
 390
 391/*
 392 * Insert rq into dispatch queue of q.  Queue lock must be held on
 393 * entry.  rq is added to the back of the dispatch queue. To be used by
 394 * specific elevators.
 395 */
 396void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 397{
 398	if (q->last_merge == rq)
 399		q->last_merge = NULL;
 400
 401	elv_rqhash_del(q, rq);
 402
 403	q->nr_sorted--;
 404
 405	q->end_sector = rq_end_sector(rq);
 406	q->boundary_rq = rq;
 407	list_add_tail(&rq->queuelist, &q->queue_head);
 408}
 409EXPORT_SYMBOL(elv_dispatch_add_tail);
 410
 411int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 412{
 413	struct elevator_queue *e = q->elevator;
 414	struct request *__rq;
 415	int ret;
 416
 417	/*
 418	 * Levels of merges:
 419	 * 	nomerges:  No merges at all attempted
 420	 * 	noxmerges: Only simple one-hit cache try
 421	 * 	merges:	   All merge tries attempted
 422	 */
 423	if (blk_queue_nomerges(q) || !bio_mergeable(bio))
 424		return ELEVATOR_NO_MERGE;
 425
 426	/*
 427	 * First try one-hit cache.
 428	 */
 429	if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
 430		ret = blk_try_merge(q->last_merge, bio);
 
 431		if (ret != ELEVATOR_NO_MERGE) {
 432			*req = q->last_merge;
 433			return ret;
 434		}
 435	}
 436
 437	if (blk_queue_noxmerges(q))
 438		return ELEVATOR_NO_MERGE;
 439
 440	/*
 441	 * See if our hash lookup can find a potential backmerge.
 442	 */
 443	__rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
 444	if (__rq && elv_rq_merge_ok(__rq, bio)) {
 445		*req = __rq;
 446		return ELEVATOR_BACK_MERGE;
 447	}
 448
 449	if (e->type->ops.elevator_merge_fn)
 450		return e->type->ops.elevator_merge_fn(q, req, bio);
 451
 452	return ELEVATOR_NO_MERGE;
 453}
 454
 455/*
 456 * Attempt to do an insertion back merge. Only check for the case where
 457 * we can append 'rq' to an existing request, so we can throw 'rq' away
 458 * afterwards.
 459 *
 460 * Returns true if we merged, false otherwise
 461 */
 462static bool elv_attempt_insert_merge(struct request_queue *q,
 463				     struct request *rq)
 464{
 465	struct request *__rq;
 466	bool ret;
 467
 468	if (blk_queue_nomerges(q))
 469		return false;
 470
 471	/*
 472	 * First try one-hit cache.
 473	 */
 474	if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 475		return true;
 476
 477	if (blk_queue_noxmerges(q))
 478		return false;
 479
 480	ret = false;
 481	/*
 482	 * See if our hash lookup can find a potential backmerge.
 483	 */
 484	while (1) {
 485		__rq = elv_rqhash_find(q, blk_rq_pos(rq));
 486		if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
 487			break;
 488
 489		/* The merged request could be merged with others, try again */
 490		ret = true;
 491		rq = __rq;
 492	}
 493
 494	return ret;
 495}
 496
 497void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 
 498{
 499	struct elevator_queue *e = q->elevator;
 500
 501	if (e->type->ops.elevator_merged_fn)
 502		e->type->ops.elevator_merged_fn(q, rq, type);
 503
 504	if (type == ELEVATOR_BACK_MERGE)
 505		elv_rqhash_reposition(q, rq);
 506
 507	q->last_merge = rq;
 508}
 509
 510void elv_merge_requests(struct request_queue *q, struct request *rq,
 511			     struct request *next)
 512{
 513	struct elevator_queue *e = q->elevator;
 514	const int next_sorted = next->cmd_flags & REQ_SORTED;
 515
 516	if (next_sorted && e->type->ops.elevator_merge_req_fn)
 517		e->type->ops.elevator_merge_req_fn(q, rq, next);
 518
 519	elv_rqhash_reposition(q, rq);
 520
 521	if (next_sorted) {
 522		elv_rqhash_del(q, next);
 523		q->nr_sorted--;
 524	}
 525
 526	q->last_merge = rq;
 527}
 528
 529void elv_bio_merged(struct request_queue *q, struct request *rq,
 530			struct bio *bio)
 531{
 532	struct elevator_queue *e = q->elevator;
 533
 534	if (e->type->ops.elevator_bio_merged_fn)
 535		e->type->ops.elevator_bio_merged_fn(q, rq, bio);
 536}
 537
 538#ifdef CONFIG_PM
 539static void blk_pm_requeue_request(struct request *rq)
 540{
 541	if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
 542		rq->q->nr_pending--;
 543}
 544
 545static void blk_pm_add_request(struct request_queue *q, struct request *rq)
 546{
 547	if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
 548	    (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
 549		pm_request_resume(q->dev);
 550}
 551#else
 552static inline void blk_pm_requeue_request(struct request *rq) {}
 553static inline void blk_pm_add_request(struct request_queue *q,
 554				      struct request *rq)
 555{
 556}
 557#endif
 558
 559void elv_requeue_request(struct request_queue *q, struct request *rq)
 560{
 561	/*
 562	 * it already went through dequeue, we need to decrement the
 563	 * in_flight count again
 564	 */
 565	if (blk_account_rq(rq)) {
 566		q->in_flight[rq_is_sync(rq)]--;
 567		if (rq->cmd_flags & REQ_SORTED)
 568			elv_deactivate_rq(q, rq);
 569	}
 570
 571	rq->cmd_flags &= ~REQ_STARTED;
 572
 573	blk_pm_requeue_request(rq);
 574
 575	__elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 576}
 577
 578void elv_drain_elevator(struct request_queue *q)
 579{
 580	static int printed;
 581
 582	lockdep_assert_held(q->queue_lock);
 583
 584	while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
 585		;
 586	if (q->nr_sorted && printed++ < 10) {
 587		printk(KERN_ERR "%s: forced dispatching is broken "
 588		       "(nr_sorted=%u), please report this\n",
 589		       q->elevator->type->elevator_name, q->nr_sorted);
 590	}
 591}
 592
 593void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 594{
 595	trace_block_rq_insert(q, rq);
 596
 597	blk_pm_add_request(q, rq);
 598
 599	rq->q = q;
 600
 601	if (rq->cmd_flags & REQ_SOFTBARRIER) {
 602		/* barriers are scheduling boundary, update end_sector */
 603		if (rq->cmd_type == REQ_TYPE_FS) {
 604			q->end_sector = rq_end_sector(rq);
 605			q->boundary_rq = rq;
 606		}
 607	} else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 608		    (where == ELEVATOR_INSERT_SORT ||
 609		     where == ELEVATOR_INSERT_SORT_MERGE))
 610		where = ELEVATOR_INSERT_BACK;
 611
 612	switch (where) {
 613	case ELEVATOR_INSERT_REQUEUE:
 614	case ELEVATOR_INSERT_FRONT:
 615		rq->cmd_flags |= REQ_SOFTBARRIER;
 616		list_add(&rq->queuelist, &q->queue_head);
 617		break;
 618
 619	case ELEVATOR_INSERT_BACK:
 620		rq->cmd_flags |= REQ_SOFTBARRIER;
 621		elv_drain_elevator(q);
 622		list_add_tail(&rq->queuelist, &q->queue_head);
 623		/*
 624		 * We kick the queue here for the following reasons.
 625		 * - The elevator might have returned NULL previously
 626		 *   to delay requests and returned them now.  As the
 627		 *   queue wasn't empty before this request, ll_rw_blk
 628		 *   won't run the queue on return, resulting in hang.
 629		 * - Usually, back inserted requests won't be merged
 630		 *   with anything.  There's no point in delaying queue
 631		 *   processing.
 632		 */
 633		__blk_run_queue(q);
 634		break;
 635
 636	case ELEVATOR_INSERT_SORT_MERGE:
 637		/*
 638		 * If we succeed in merging this request with one in the
 639		 * queue already, we are done - rq has now been freed,
 640		 * so no need to do anything further.
 641		 */
 642		if (elv_attempt_insert_merge(q, rq))
 643			break;
 644	case ELEVATOR_INSERT_SORT:
 645		BUG_ON(rq->cmd_type != REQ_TYPE_FS);
 646		rq->cmd_flags |= REQ_SORTED;
 647		q->nr_sorted++;
 648		if (rq_mergeable(rq)) {
 649			elv_rqhash_add(q, rq);
 650			if (!q->last_merge)
 651				q->last_merge = rq;
 652		}
 653
 654		/*
 655		 * Some ioscheds (cfq) run q->request_fn directly, so
 656		 * rq cannot be accessed after calling
 657		 * elevator_add_req_fn.
 658		 */
 659		q->elevator->type->ops.elevator_add_req_fn(q, rq);
 660		break;
 661
 662	case ELEVATOR_INSERT_FLUSH:
 663		rq->cmd_flags |= REQ_SOFTBARRIER;
 664		blk_insert_flush(rq);
 665		break;
 666	default:
 667		printk(KERN_ERR "%s: bad insertion point %d\n",
 668		       __func__, where);
 669		BUG();
 670	}
 671}
 672EXPORT_SYMBOL(__elv_add_request);
 673
 674void elv_add_request(struct request_queue *q, struct request *rq, int where)
 675{
 676	unsigned long flags;
 677
 678	spin_lock_irqsave(q->queue_lock, flags);
 679	__elv_add_request(q, rq, where);
 680	spin_unlock_irqrestore(q->queue_lock, flags);
 681}
 682EXPORT_SYMBOL(elv_add_request);
 683
 684struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 685{
 686	struct elevator_queue *e = q->elevator;
 687
 688	if (e->type->ops.elevator_latter_req_fn)
 689		return e->type->ops.elevator_latter_req_fn(q, rq);
 690	return NULL;
 691}
 692
 693struct request *elv_former_request(struct request_queue *q, struct request *rq)
 694{
 695	struct elevator_queue *e = q->elevator;
 696
 697	if (e->type->ops.elevator_former_req_fn)
 698		return e->type->ops.elevator_former_req_fn(q, rq);
 699	return NULL;
 700}
 701
 702int elv_set_request(struct request_queue *q, struct request *rq,
 703		    struct bio *bio, gfp_t gfp_mask)
 704{
 705	struct elevator_queue *e = q->elevator;
 706
 707	if (e->type->ops.elevator_set_req_fn)
 708		return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
 709	return 0;
 710}
 711
 712void elv_put_request(struct request_queue *q, struct request *rq)
 713{
 714	struct elevator_queue *e = q->elevator;
 715
 716	if (e->type->ops.elevator_put_req_fn)
 717		e->type->ops.elevator_put_req_fn(rq);
 718}
 719
 720int elv_may_queue(struct request_queue *q, int rw)
 721{
 722	struct elevator_queue *e = q->elevator;
 723
 724	if (e->type->ops.elevator_may_queue_fn)
 725		return e->type->ops.elevator_may_queue_fn(q, rw);
 726
 727	return ELV_MQUEUE_MAY;
 728}
 729
 730void elv_completed_request(struct request_queue *q, struct request *rq)
 731{
 732	struct elevator_queue *e = q->elevator;
 733
 734	/*
 735	 * request is released from the driver, io must be done
 736	 */
 737	if (blk_account_rq(rq)) {
 738		q->in_flight[rq_is_sync(rq)]--;
 739		if ((rq->cmd_flags & REQ_SORTED) &&
 740		    e->type->ops.elevator_completed_req_fn)
 741			e->type->ops.elevator_completed_req_fn(q, rq);
 742	}
 743}
 744
 745#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 746
 747static ssize_t
 748elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 749{
 750	struct elv_fs_entry *entry = to_elv(attr);
 751	struct elevator_queue *e;
 752	ssize_t error;
 753
 754	if (!entry->show)
 755		return -EIO;
 756
 757	e = container_of(kobj, struct elevator_queue, kobj);
 758	mutex_lock(&e->sysfs_lock);
 759	error = e->type ? entry->show(e, page) : -ENOENT;
 760	mutex_unlock(&e->sysfs_lock);
 761	return error;
 762}
 763
 764static ssize_t
 765elv_attr_store(struct kobject *kobj, struct attribute *attr,
 766	       const char *page, size_t length)
 767{
 768	struct elv_fs_entry *entry = to_elv(attr);
 769	struct elevator_queue *e;
 770	ssize_t error;
 771
 772	if (!entry->store)
 773		return -EIO;
 774
 775	e = container_of(kobj, struct elevator_queue, kobj);
 776	mutex_lock(&e->sysfs_lock);
 777	error = e->type ? entry->store(e, page, length) : -ENOENT;
 778	mutex_unlock(&e->sysfs_lock);
 779	return error;
 780}
 781
 782static const struct sysfs_ops elv_sysfs_ops = {
 783	.show	= elv_attr_show,
 784	.store	= elv_attr_store,
 785};
 786
 787static struct kobj_type elv_ktype = {
 788	.sysfs_ops	= &elv_sysfs_ops,
 789	.release	= elevator_release,
 790};
 791
 792int elv_register_queue(struct request_queue *q)
 
 
 
 
 
 793{
 794	struct elevator_queue *e = q->elevator;
 795	int error;
 796
 797	error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 798	if (!error) {
 799		struct elv_fs_entry *attr = e->type->elevator_attrs;
 800		if (attr) {
 801			while (attr->attr.name) {
 802				if (sysfs_create_file(&e->kobj, &attr->attr))
 803					break;
 804				attr++;
 805			}
 806		}
 807		kobject_uevent(&e->kobj, KOBJ_ADD);
 
 
 808		e->registered = 1;
 809		if (e->type->ops.elevator_registered_fn)
 810			e->type->ops.elevator_registered_fn(q);
 811	}
 812	return error;
 813}
 814EXPORT_SYMBOL(elv_register_queue);
 815
 
 
 
 
 
 816void elv_unregister_queue(struct request_queue *q)
 817{
 818	if (q) {
 819		struct elevator_queue *e = q->elevator;
 820
 821		kobject_uevent(&e->kobj, KOBJ_REMOVE);
 822		kobject_del(&e->kobj);
 
 823		e->registered = 0;
 
 
 824	}
 825}
 826EXPORT_SYMBOL(elv_unregister_queue);
 827
 828int elv_register(struct elevator_type *e)
 829{
 830	char *def = "";
 831
 832	/* create icq_cache if requested */
 833	if (e->icq_size) {
 834		if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 835		    WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 836			return -EINVAL;
 837
 838		snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 839			 "%s_io_cq", e->elevator_name);
 840		e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 841						 e->icq_align, 0, NULL);
 842		if (!e->icq_cache)
 843			return -ENOMEM;
 844	}
 845
 846	/* register, don't allow duplicate names */
 847	spin_lock(&elv_list_lock);
 848	if (elevator_find(e->elevator_name)) {
 849		spin_unlock(&elv_list_lock);
 850		if (e->icq_cache)
 851			kmem_cache_destroy(e->icq_cache);
 852		return -EBUSY;
 853	}
 854	list_add_tail(&e->list, &elv_list);
 855	spin_unlock(&elv_list_lock);
 856
 857	/* print pretty message */
 858	if (!strcmp(e->elevator_name, chosen_elevator) ||
 859			(!*chosen_elevator &&
 860			 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 861				def = " (default)";
 862
 863	printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 864								def);
 865	return 0;
 866}
 867EXPORT_SYMBOL_GPL(elv_register);
 868
 869void elv_unregister(struct elevator_type *e)
 870{
 871	/* unregister */
 872	spin_lock(&elv_list_lock);
 873	list_del_init(&e->list);
 874	spin_unlock(&elv_list_lock);
 875
 876	/*
 877	 * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 878	 * sure all RCU operations are complete before proceeding.
 879	 */
 880	if (e->icq_cache) {
 881		rcu_barrier();
 882		kmem_cache_destroy(e->icq_cache);
 883		e->icq_cache = NULL;
 884	}
 885}
 886EXPORT_SYMBOL_GPL(elv_unregister);
 887
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 888/*
 889 * switch to new_e io scheduler. be careful not to introduce deadlocks -
 890 * we don't free the old io scheduler, before we have allocated what we
 891 * need for the new one. this way we have a chance of going back to the old
 892 * one, if the new one fails init for some reason.
 893 */
 894static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 895{
 896	struct elevator_queue *old = q->elevator;
 897	bool registered = old->registered;
 898	int err;
 899
 900	/*
 901	 * Turn on BYPASS and drain all requests w/ elevator private data.
 902	 * Block layer doesn't call into a quiesced elevator - all requests
 903	 * are directly put on the dispatch list without elevator data
 904	 * using INSERT_BACK.  All requests have SOFTBARRIER set and no
 905	 * merge happens either.
 906	 */
 907	blk_queue_bypass_start(q);
 908
 909	/* unregister and clear all auxiliary data of the old elevator */
 910	if (registered)
 911		elv_unregister_queue(q);
 912
 913	spin_lock_irq(q->queue_lock);
 914	ioc_clear_queue(q);
 915	spin_unlock_irq(q->queue_lock);
 916
 917	/* allocate, init and register new elevator */
 918	err = new_e->ops.elevator_init_fn(q, new_e);
 919	if (err)
 920		goto fail_init;
 921
 922	if (registered) {
 923		err = elv_register_queue(q);
 924		if (err)
 925			goto fail_register;
 926	}
 927
 928	/* done, kill the old one and finish */
 929	elevator_exit(old);
 930	blk_queue_bypass_end(q);
 931
 932	blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 
 933
 934	return 0;
 935
 936fail_register:
 937	elevator_exit(q->elevator);
 938fail_init:
 939	/* switch failed, restore and re-register old elevator */
 940	q->elevator = old;
 941	elv_register_queue(q);
 942	blk_queue_bypass_end(q);
 943
 944	return err;
 945}
 946
 947/*
 948 * Switch this queue to the given IO scheduler.
 949 */
 950static int __elevator_change(struct request_queue *q, const char *name)
 951{
 952	char elevator_name[ELV_NAME_MAX];
 953	struct elevator_type *e;
 954
 955	if (!q->elevator)
 956		return -ENXIO;
 
 
 
 
 
 
 
 
 
 
 957
 958	strlcpy(elevator_name, name, sizeof(elevator_name));
 959	e = elevator_get(strstrip(elevator_name), true);
 960	if (!e) {
 961		printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 962		return -EINVAL;
 963	}
 964
 965	if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
 
 966		elevator_put(e);
 967		return 0;
 968	}
 969
 970	return elevator_switch(q, e);
 971}
 972
 973int elevator_change(struct request_queue *q, const char *name)
 974{
 975	int ret;
 976
 977	/* Protect q->elevator from elevator_init() */
 978	mutex_lock(&q->sysfs_lock);
 979	ret = __elevator_change(q, name);
 980	mutex_unlock(&q->sysfs_lock);
 981
 982	return ret;
 983}
 984EXPORT_SYMBOL(elevator_change);
 985
 986ssize_t elv_iosched_store(struct request_queue *q, const char *name,
 987			  size_t count)
 988{
 989	int ret;
 990
 991	if (!q->elevator)
 992		return count;
 993
 994	ret = __elevator_change(q, name);
 995	if (!ret)
 996		return count;
 997
 998	printk(KERN_ERR "elevator: switch to %s failed\n", name);
 999	return ret;
1000}
1001
1002ssize_t elv_iosched_show(struct request_queue *q, char *name)
1003{
1004	struct elevator_queue *e = q->elevator;
1005	struct elevator_type *elv;
1006	struct elevator_type *__e;
1007	int len = 0;
1008
1009	if (!q->elevator || !blk_queue_stackable(q))
1010		return sprintf(name, "none\n");
1011
1012	elv = e->type;
 
 
 
1013
1014	spin_lock(&elv_list_lock);
1015	list_for_each_entry(__e, &elv_list, list) {
1016		if (!strcmp(elv->elevator_name, __e->elevator_name))
1017			len += sprintf(name+len, "[%s] ", elv->elevator_name);
1018		else
 
 
 
 
1019			len += sprintf(name+len, "%s ", __e->elevator_name);
1020	}
1021	spin_unlock(&elv_list_lock);
1022
 
 
 
1023	len += sprintf(len+name, "\n");
1024	return len;
1025}
1026
1027struct request *elv_rb_former_request(struct request_queue *q,
1028				      struct request *rq)
1029{
1030	struct rb_node *rbprev = rb_prev(&rq->rb_node);
1031
1032	if (rbprev)
1033		return rb_entry_rq(rbprev);
1034
1035	return NULL;
1036}
1037EXPORT_SYMBOL(elv_rb_former_request);
1038
1039struct request *elv_rb_latter_request(struct request_queue *q,
1040				      struct request *rq)
1041{
1042	struct rb_node *rbnext = rb_next(&rq->rb_node);
1043
1044	if (rbnext)
1045		return rb_entry_rq(rbnext);
1046
1047	return NULL;
1048}
1049EXPORT_SYMBOL(elv_rb_latter_request);
v5.9
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *  Block device elevator/IO-scheduler.
  4 *
  5 *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
  6 *
  7 * 30042000 Jens Axboe <axboe@kernel.dk> :
  8 *
  9 * Split the elevator a bit so that it is possible to choose a different
 10 * one or even write a new "plug in". There are three pieces:
 11 * - elevator_fn, inserts a new request in the queue list
 12 * - elevator_merge_fn, decides whether a new buffer can be merged with
 13 *   an existing request
 14 * - elevator_dequeue_fn, called when a request is taken off the active list
 15 *
 16 * 20082000 Dave Jones <davej@suse.de> :
 17 * Removed tests for max-bomb-segments, which was breaking elvtune
 18 *  when run without -bN
 19 *
 20 * Jens:
 21 * - Rework again to work with bio instead of buffer_heads
 22 * - loose bi_dev comparisons, partition handling is right now
 23 * - completely modularize elevator setup and teardown
 24 *
 25 */
 26#include <linux/kernel.h>
 27#include <linux/fs.h>
 28#include <linux/blkdev.h>
 29#include <linux/elevator.h>
 30#include <linux/bio.h>
 31#include <linux/module.h>
 32#include <linux/slab.h>
 33#include <linux/init.h>
 34#include <linux/compiler.h>
 35#include <linux/blktrace_api.h>
 36#include <linux/hash.h>
 37#include <linux/uaccess.h>
 38#include <linux/pm_runtime.h>
 39#include <linux/blk-cgroup.h>
 40
 41#include <trace/events/block.h>
 42
 43#include "blk.h"
 44#include "blk-mq-sched.h"
 45#include "blk-pm.h"
 46#include "blk-wbt.h"
 47
 48static DEFINE_SPINLOCK(elv_list_lock);
 49static LIST_HEAD(elv_list);
 50
 51/*
 52 * Merge hash stuff.
 53 */
 54#define rq_hash_key(rq)		(blk_rq_pos(rq) + blk_rq_sectors(rq))
 55
 56/*
 57 * Query io scheduler to see if the current process issuing bio may be
 58 * merged with rq.
 59 */
 60static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
 61{
 62	struct request_queue *q = rq->q;
 63	struct elevator_queue *e = q->elevator;
 64
 65	if (e->type->ops.allow_merge)
 66		return e->type->ops.allow_merge(q, rq, bio);
 67
 68	return 1;
 69}
 70
 71/*
 72 * can we safely merge with this request?
 73 */
 74bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
 75{
 76	if (!blk_rq_merge_ok(rq, bio))
 77		return false;
 78
 79	if (!elv_iosched_allow_bio_merge(rq, bio))
 80		return false;
 81
 82	return true;
 83}
 84EXPORT_SYMBOL(elv_bio_merge_ok);
 85
 86static inline bool elv_support_features(unsigned int elv_features,
 87					unsigned int required_features)
 88{
 89	return (required_features & elv_features) == required_features;
 90}
 91
 92/**
 93 * elevator_match - Test an elevator name and features
 94 * @e: Scheduler to test
 95 * @name: Elevator name to test
 96 * @required_features: Features that the elevator must provide
 97 *
 98 * Return true if the elevator @e name matches @name and if @e provides all
 99 * the features specified by @required_features.
100 */
101static bool elevator_match(const struct elevator_type *e, const char *name,
102			   unsigned int required_features)
103{
104	if (!elv_support_features(e->elevator_features, required_features))
105		return false;
106	if (!strcmp(e->elevator_name, name))
107		return true;
108	if (e->elevator_alias && !strcmp(e->elevator_alias, name))
109		return true;
110
111	return false;
112}
 
113
114/**
115 * elevator_find - Find an elevator
116 * @name: Name of the elevator to find
117 * @required_features: Features that the elevator must provide
118 *
119 * Return the first registered scheduler with name @name and supporting the
120 * features @required_features and NULL otherwise.
121 */
122static struct elevator_type *elevator_find(const char *name,
123					   unsigned int required_features)
124{
125	struct elevator_type *e;
126
127	list_for_each_entry(e, &elv_list, list) {
128		if (elevator_match(e, name, required_features))
129			return e;
130	}
131
132	return NULL;
133}
134
135static void elevator_put(struct elevator_type *e)
136{
137	module_put(e->elevator_owner);
138}
139
140static struct elevator_type *elevator_get(struct request_queue *q,
141					  const char *name, bool try_loading)
142{
143	struct elevator_type *e;
144
145	spin_lock(&elv_list_lock);
146
147	e = elevator_find(name, q->required_elevator_features);
148	if (!e && try_loading) {
149		spin_unlock(&elv_list_lock);
150		request_module("%s-iosched", name);
151		spin_lock(&elv_list_lock);
152		e = elevator_find(name, q->required_elevator_features);
153	}
154
155	if (e && !try_module_get(e->elevator_owner))
156		e = NULL;
157
158	spin_unlock(&elv_list_lock);
 
159	return e;
160}
161
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
162static struct kobj_type elv_ktype;
163
164struct elevator_queue *elevator_alloc(struct request_queue *q,
165				  struct elevator_type *e)
166{
167	struct elevator_queue *eq;
168
169	eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
170	if (unlikely(!eq))
171		return NULL;
172
173	eq->type = e;
174	kobject_init(&eq->kobj, &elv_ktype);
175	mutex_init(&eq->sysfs_lock);
176	hash_init(eq->hash);
177
178	return eq;
179}
180EXPORT_SYMBOL(elevator_alloc);
181
182static void elevator_release(struct kobject *kobj)
183{
184	struct elevator_queue *e;
185
186	e = container_of(kobj, struct elevator_queue, kobj);
187	elevator_put(e->type);
188	kfree(e);
189}
190
191void __elevator_exit(struct request_queue *q, struct elevator_queue *e)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
192{
193	mutex_lock(&e->sysfs_lock);
194	if (e->type->ops.exit_sched)
195		blk_mq_exit_sched(q, e);
196	mutex_unlock(&e->sysfs_lock);
197
198	kobject_put(&e->kobj);
199}
 
200
201static inline void __elv_rqhash_del(struct request *rq)
202{
203	hash_del(&rq->hash);
204	rq->rq_flags &= ~RQF_HASHED;
205}
206
207void elv_rqhash_del(struct request_queue *q, struct request *rq)
208{
209	if (ELV_ON_HASH(rq))
210		__elv_rqhash_del(rq);
211}
212EXPORT_SYMBOL_GPL(elv_rqhash_del);
213
214void elv_rqhash_add(struct request_queue *q, struct request *rq)
215{
216	struct elevator_queue *e = q->elevator;
217
218	BUG_ON(ELV_ON_HASH(rq));
219	hash_add(e->hash, &rq->hash, rq_hash_key(rq));
220	rq->rq_flags |= RQF_HASHED;
221}
222EXPORT_SYMBOL_GPL(elv_rqhash_add);
223
224void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
225{
226	__elv_rqhash_del(rq);
227	elv_rqhash_add(q, rq);
228}
229
230struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
231{
232	struct elevator_queue *e = q->elevator;
233	struct hlist_node *next;
234	struct request *rq;
235
236	hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
237		BUG_ON(!ELV_ON_HASH(rq));
238
239		if (unlikely(!rq_mergeable(rq))) {
240			__elv_rqhash_del(rq);
241			continue;
242		}
243
244		if (rq_hash_key(rq) == offset)
245			return rq;
246	}
247
248	return NULL;
249}
250
251/*
252 * RB-tree support functions for inserting/lookup/removal of requests
253 * in a sorted RB tree.
254 */
255void elv_rb_add(struct rb_root *root, struct request *rq)
256{
257	struct rb_node **p = &root->rb_node;
258	struct rb_node *parent = NULL;
259	struct request *__rq;
260
261	while (*p) {
262		parent = *p;
263		__rq = rb_entry(parent, struct request, rb_node);
264
265		if (blk_rq_pos(rq) < blk_rq_pos(__rq))
266			p = &(*p)->rb_left;
267		else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
268			p = &(*p)->rb_right;
269	}
270
271	rb_link_node(&rq->rb_node, parent, p);
272	rb_insert_color(&rq->rb_node, root);
273}
274EXPORT_SYMBOL(elv_rb_add);
275
276void elv_rb_del(struct rb_root *root, struct request *rq)
277{
278	BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
279	rb_erase(&rq->rb_node, root);
280	RB_CLEAR_NODE(&rq->rb_node);
281}
282EXPORT_SYMBOL(elv_rb_del);
283
284struct request *elv_rb_find(struct rb_root *root, sector_t sector)
285{
286	struct rb_node *n = root->rb_node;
287	struct request *rq;
288
289	while (n) {
290		rq = rb_entry(n, struct request, rb_node);
291
292		if (sector < blk_rq_pos(rq))
293			n = n->rb_left;
294		else if (sector > blk_rq_pos(rq))
295			n = n->rb_right;
296		else
297			return rq;
298	}
299
300	return NULL;
301}
302EXPORT_SYMBOL(elv_rb_find);
303
304enum elv_merge elv_merge(struct request_queue *q, struct request **req,
305		struct bio *bio)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
306{
307	struct elevator_queue *e = q->elevator;
308	struct request *__rq;
 
309
310	/*
311	 * Levels of merges:
312	 * 	nomerges:  No merges at all attempted
313	 * 	noxmerges: Only simple one-hit cache try
314	 * 	merges:	   All merge tries attempted
315	 */
316	if (blk_queue_nomerges(q) || !bio_mergeable(bio))
317		return ELEVATOR_NO_MERGE;
318
319	/*
320	 * First try one-hit cache.
321	 */
322	if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
323		enum elv_merge ret = blk_try_merge(q->last_merge, bio);
324
325		if (ret != ELEVATOR_NO_MERGE) {
326			*req = q->last_merge;
327			return ret;
328		}
329	}
330
331	if (blk_queue_noxmerges(q))
332		return ELEVATOR_NO_MERGE;
333
334	/*
335	 * See if our hash lookup can find a potential backmerge.
336	 */
337	__rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
338	if (__rq && elv_bio_merge_ok(__rq, bio)) {
339		*req = __rq;
340		return ELEVATOR_BACK_MERGE;
341	}
342
343	if (e->type->ops.request_merge)
344		return e->type->ops.request_merge(q, req, bio);
345
346	return ELEVATOR_NO_MERGE;
347}
348
349/*
350 * Attempt to do an insertion back merge. Only check for the case where
351 * we can append 'rq' to an existing request, so we can throw 'rq' away
352 * afterwards.
353 *
354 * Returns true if we merged, false otherwise
355 */
356bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
 
357{
358	struct request *__rq;
359	bool ret;
360
361	if (blk_queue_nomerges(q))
362		return false;
363
364	/*
365	 * First try one-hit cache.
366	 */
367	if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
368		return true;
369
370	if (blk_queue_noxmerges(q))
371		return false;
372
373	ret = false;
374	/*
375	 * See if our hash lookup can find a potential backmerge.
376	 */
377	while (1) {
378		__rq = elv_rqhash_find(q, blk_rq_pos(rq));
379		if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
380			break;
381
382		/* The merged request could be merged with others, try again */
383		ret = true;
384		rq = __rq;
385	}
386
387	return ret;
388}
389
390void elv_merged_request(struct request_queue *q, struct request *rq,
391		enum elv_merge type)
392{
393	struct elevator_queue *e = q->elevator;
394
395	if (e->type->ops.request_merged)
396		e->type->ops.request_merged(q, rq, type);
397
398	if (type == ELEVATOR_BACK_MERGE)
399		elv_rqhash_reposition(q, rq);
400
401	q->last_merge = rq;
402}
403
404void elv_merge_requests(struct request_queue *q, struct request *rq,
405			     struct request *next)
406{
407	struct elevator_queue *e = q->elevator;
 
408
409	if (e->type->ops.requests_merged)
410		e->type->ops.requests_merged(q, rq, next);
411
412	elv_rqhash_reposition(q, rq);
 
 
 
 
 
 
413	q->last_merge = rq;
414}
415
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
416struct request *elv_latter_request(struct request_queue *q, struct request *rq)
417{
418	struct elevator_queue *e = q->elevator;
419
420	if (e->type->ops.next_request)
421		return e->type->ops.next_request(q, rq);
 
 
422
 
 
 
 
 
 
423	return NULL;
424}
425
426struct request *elv_former_request(struct request_queue *q, struct request *rq)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
427{
428	struct elevator_queue *e = q->elevator;
429
430	if (e->type->ops.former_request)
431		return e->type->ops.former_request(q, rq);
432
433	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
434}
435
436#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
437
438static ssize_t
439elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
440{
441	struct elv_fs_entry *entry = to_elv(attr);
442	struct elevator_queue *e;
443	ssize_t error;
444
445	if (!entry->show)
446		return -EIO;
447
448	e = container_of(kobj, struct elevator_queue, kobj);
449	mutex_lock(&e->sysfs_lock);
450	error = e->type ? entry->show(e, page) : -ENOENT;
451	mutex_unlock(&e->sysfs_lock);
452	return error;
453}
454
455static ssize_t
456elv_attr_store(struct kobject *kobj, struct attribute *attr,
457	       const char *page, size_t length)
458{
459	struct elv_fs_entry *entry = to_elv(attr);
460	struct elevator_queue *e;
461	ssize_t error;
462
463	if (!entry->store)
464		return -EIO;
465
466	e = container_of(kobj, struct elevator_queue, kobj);
467	mutex_lock(&e->sysfs_lock);
468	error = e->type ? entry->store(e, page, length) : -ENOENT;
469	mutex_unlock(&e->sysfs_lock);
470	return error;
471}
472
473static const struct sysfs_ops elv_sysfs_ops = {
474	.show	= elv_attr_show,
475	.store	= elv_attr_store,
476};
477
478static struct kobj_type elv_ktype = {
479	.sysfs_ops	= &elv_sysfs_ops,
480	.release	= elevator_release,
481};
482
483/*
484 * elv_register_queue is called from either blk_register_queue or
485 * elevator_switch, elevator switch is prevented from being happen
486 * in the two paths, so it is safe to not hold q->sysfs_lock.
487 */
488int elv_register_queue(struct request_queue *q, bool uevent)
489{
490	struct elevator_queue *e = q->elevator;
491	int error;
492
493	error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
494	if (!error) {
495		struct elv_fs_entry *attr = e->type->elevator_attrs;
496		if (attr) {
497			while (attr->attr.name) {
498				if (sysfs_create_file(&e->kobj, &attr->attr))
499					break;
500				attr++;
501			}
502		}
503		if (uevent)
504			kobject_uevent(&e->kobj, KOBJ_ADD);
505
506		e->registered = 1;
 
 
507	}
508	return error;
509}
 
510
511/*
512 * elv_unregister_queue is called from either blk_unregister_queue or
513 * elevator_switch, elevator switch is prevented from being happen
514 * in the two paths, so it is safe to not hold q->sysfs_lock.
515 */
516void elv_unregister_queue(struct request_queue *q)
517{
518	if (q) {
519		struct elevator_queue *e = q->elevator;
520
521		kobject_uevent(&e->kobj, KOBJ_REMOVE);
522		kobject_del(&e->kobj);
523
524		e->registered = 0;
525		/* Re-enable throttling in case elevator disabled it */
526		wbt_enable_default(q);
527	}
528}
 
529
530int elv_register(struct elevator_type *e)
531{
 
 
532	/* create icq_cache if requested */
533	if (e->icq_size) {
534		if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
535		    WARN_ON(e->icq_align < __alignof__(struct io_cq)))
536			return -EINVAL;
537
538		snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
539			 "%s_io_cq", e->elevator_name);
540		e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
541						 e->icq_align, 0, NULL);
542		if (!e->icq_cache)
543			return -ENOMEM;
544	}
545
546	/* register, don't allow duplicate names */
547	spin_lock(&elv_list_lock);
548	if (elevator_find(e->elevator_name, 0)) {
549		spin_unlock(&elv_list_lock);
550		kmem_cache_destroy(e->icq_cache);
 
551		return -EBUSY;
552	}
553	list_add_tail(&e->list, &elv_list);
554	spin_unlock(&elv_list_lock);
555
556	printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name);
 
 
 
 
557
 
 
558	return 0;
559}
560EXPORT_SYMBOL_GPL(elv_register);
561
562void elv_unregister(struct elevator_type *e)
563{
564	/* unregister */
565	spin_lock(&elv_list_lock);
566	list_del_init(&e->list);
567	spin_unlock(&elv_list_lock);
568
569	/*
570	 * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
571	 * sure all RCU operations are complete before proceeding.
572	 */
573	if (e->icq_cache) {
574		rcu_barrier();
575		kmem_cache_destroy(e->icq_cache);
576		e->icq_cache = NULL;
577	}
578}
579EXPORT_SYMBOL_GPL(elv_unregister);
580
581int elevator_switch_mq(struct request_queue *q,
582			      struct elevator_type *new_e)
583{
584	int ret;
585
586	lockdep_assert_held(&q->sysfs_lock);
587
588	if (q->elevator) {
589		if (q->elevator->registered)
590			elv_unregister_queue(q);
591
592		ioc_clear_queue(q);
593		elevator_exit(q, q->elevator);
594	}
595
596	ret = blk_mq_init_sched(q, new_e);
597	if (ret)
598		goto out;
599
600	if (new_e) {
601		ret = elv_register_queue(q, true);
602		if (ret) {
603			elevator_exit(q, q->elevator);
604			goto out;
605		}
606	}
607
608	if (new_e)
609		blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
610	else
611		blk_add_trace_msg(q, "elv switch: none");
612
613out:
614	return ret;
615}
616
617static inline bool elv_support_iosched(struct request_queue *q)
618{
619	if (!q->mq_ops ||
620	    (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED)))
621		return false;
622	return true;
623}
624
625/*
626 * For single queue devices, default to using mq-deadline. If we have multiple
627 * queues or mq-deadline is not available, default to "none".
628 */
629static struct elevator_type *elevator_get_default(struct request_queue *q)
630{
631	if (q->nr_hw_queues != 1)
632		return NULL;
633
634	return elevator_get(q, "mq-deadline", false);
635}
636
637/*
638 * Get the first elevator providing the features required by the request queue.
639 * Default to "none" if no matching elevator is found.
640 */
641static struct elevator_type *elevator_get_by_features(struct request_queue *q)
642{
643	struct elevator_type *e, *found = NULL;
644
645	spin_lock(&elv_list_lock);
646
647	list_for_each_entry(e, &elv_list, list) {
648		if (elv_support_features(e->elevator_features,
649					 q->required_elevator_features)) {
650			found = e;
651			break;
652		}
653	}
654
655	if (found && !try_module_get(found->elevator_owner))
656		found = NULL;
657
658	spin_unlock(&elv_list_lock);
659	return found;
660}
661
662/*
663 * For a device queue that has no required features, use the default elevator
664 * settings. Otherwise, use the first elevator available matching the required
665 * features. If no suitable elevator is find or if the chosen elevator
666 * initialization fails, fall back to the "none" elevator (no elevator).
667 */
668void elevator_init_mq(struct request_queue *q)
669{
670	struct elevator_type *e;
671	int err;
672
673	if (!elv_support_iosched(q))
674		return;
675
676	WARN_ON_ONCE(test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags));
677
678	if (unlikely(q->elevator))
679		return;
680
681	if (!q->required_elevator_features)
682		e = elevator_get_default(q);
683	else
684		e = elevator_get_by_features(q);
685	if (!e)
686		return;
687
688	blk_mq_freeze_queue(q);
689	blk_mq_quiesce_queue(q);
690
691	err = blk_mq_init_sched(q, e);
692
693	blk_mq_unquiesce_queue(q);
694	blk_mq_unfreeze_queue(q);
695
696	if (err) {
697		pr_warn("\"%s\" elevator initialization failed, "
698			"falling back to \"none\"\n", e->elevator_name);
699		elevator_put(e);
700	}
701}
702
703
704/*
705 * switch to new_e io scheduler. be careful not to introduce deadlocks -
706 * we don't free the old io scheduler, before we have allocated what we
707 * need for the new one. this way we have a chance of going back to the old
708 * one, if the new one fails init for some reason.
709 */
710static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
711{
 
 
712	int err;
713
714	lockdep_assert_held(&q->sysfs_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
715
716	blk_mq_freeze_queue(q);
717	blk_mq_quiesce_queue(q);
718
719	err = elevator_switch_mq(q, new_e);
720
721	blk_mq_unquiesce_queue(q);
722	blk_mq_unfreeze_queue(q);
 
 
 
 
 
723
724	return err;
725}
726
727/*
728 * Switch this queue to the given IO scheduler.
729 */
730static int __elevator_change(struct request_queue *q, const char *name)
731{
732	char elevator_name[ELV_NAME_MAX];
733	struct elevator_type *e;
734
735	/* Make sure queue is not in the middle of being removed */
736	if (!blk_queue_registered(q))
737		return -ENOENT;
738
739	/*
740	 * Special case for mq, turn off scheduling
741	 */
742	if (!strncmp(name, "none", 4)) {
743		if (!q->elevator)
744			return 0;
745		return elevator_switch(q, NULL);
746	}
747
748	strlcpy(elevator_name, name, sizeof(elevator_name));
749	e = elevator_get(q, strstrip(elevator_name), true);
750	if (!e)
 
751		return -EINVAL;
 
752
753	if (q->elevator &&
754	    elevator_match(q->elevator->type, elevator_name, 0)) {
755		elevator_put(e);
756		return 0;
757	}
758
759	return elevator_switch(q, e);
760}
761
 
 
 
 
 
 
 
 
 
 
 
 
 
762ssize_t elv_iosched_store(struct request_queue *q, const char *name,
763			  size_t count)
764{
765	int ret;
766
767	if (!queue_is_mq(q) || !elv_support_iosched(q))
768		return count;
769
770	ret = __elevator_change(q, name);
771	if (!ret)
772		return count;
773
 
774	return ret;
775}
776
777ssize_t elv_iosched_show(struct request_queue *q, char *name)
778{
779	struct elevator_queue *e = q->elevator;
780	struct elevator_type *elv = NULL;
781	struct elevator_type *__e;
782	int len = 0;
783
784	if (!queue_is_mq(q))
785		return sprintf(name, "none\n");
786
787	if (!q->elevator)
788		len += sprintf(name+len, "[none] ");
789	else
790		elv = e->type;
791
792	spin_lock(&elv_list_lock);
793	list_for_each_entry(__e, &elv_list, list) {
794		if (elv && elevator_match(elv, __e->elevator_name, 0)) {
795			len += sprintf(name+len, "[%s] ", elv->elevator_name);
796			continue;
797		}
798		if (elv_support_iosched(q) &&
799		    elevator_match(__e, __e->elevator_name,
800				   q->required_elevator_features))
801			len += sprintf(name+len, "%s ", __e->elevator_name);
802	}
803	spin_unlock(&elv_list_lock);
804
805	if (q->elevator)
806		len += sprintf(name+len, "none");
807
808	len += sprintf(len+name, "\n");
809	return len;
810}
811
812struct request *elv_rb_former_request(struct request_queue *q,
813				      struct request *rq)
814{
815	struct rb_node *rbprev = rb_prev(&rq->rb_node);
816
817	if (rbprev)
818		return rb_entry_rq(rbprev);
819
820	return NULL;
821}
822EXPORT_SYMBOL(elv_rb_former_request);
823
824struct request *elv_rb_latter_request(struct request_queue *q,
825				      struct request *rq)
826{
827	struct rb_node *rbnext = rb_next(&rq->rb_node);
828
829	if (rbnext)
830		return rb_entry_rq(rbnext);
831
832	return NULL;
833}
834EXPORT_SYMBOL(elv_rb_latter_request);
835
836static int __init elevator_setup(char *str)
837{
838	pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
839		"Please use sysfs to set IO scheduler for individual devices.\n");
840	return 1;
841}
842
843__setup("elevator=", elevator_setup);