1/*
   2 * Copyright (C) 2003 Sistina Software Limited.
   3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4 *
   5 * This file is released under the GPL.
   6 */
   7
   8#include <linux/device-mapper.h>
   9
  10#include "dm-rq.h"
  11#include "dm-bio-record.h"
  12#include "dm-path-selector.h"
  13#include "dm-uevent.h"
  14
  15#include <linux/blkdev.h>
  16#include <linux/ctype.h>
  17#include <linux/init.h>
  18#include <linux/mempool.h>
  19#include <linux/module.h>
  20#include <linux/pagemap.h>
  21#include <linux/slab.h>
  22#include <linux/time.h>
  23#include <linux/timer.h>
  24#include <linux/workqueue.h>
  25#include <linux/delay.h>
  26#include <scsi/scsi_dh.h>
  27#include <linux/atomic.h>
  28#include <linux/blk-mq.h>
  29
  30#define DM_MSG_PREFIX "multipath"
  31#define DM_PG_INIT_DELAY_MSECS 2000
  32#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  33#define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
  34
  35static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
  36
  37/* Path properties */
  38struct pgpath {
  39	struct list_head list;
  40
  41	struct priority_group *pg;	/* Owning PG */
  42	unsigned fail_count;		/* Cumulative failure count */
  43
  44	struct dm_path path;
  45	struct delayed_work activate_path;
  46
  47	bool is_active:1;		/* Path status */
  48};
  49
  50#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  51
  52/*
  53 * Paths are grouped into Priority Groups and numbered from 1 upwards.
  54 * Each has a path selector which controls which path gets used.
  55 */
  56struct priority_group {
  57	struct list_head list;
  58
  59	struct multipath *m;		/* Owning multipath instance */
  60	struct path_selector ps;
  61
  62	unsigned pg_num;		/* Reference number */
  63	unsigned nr_pgpaths;		/* Number of paths in PG */
  64	struct list_head pgpaths;
  65
  66	bool bypassed:1;		/* Temporarily bypass this PG? */
  67};
  68
  69/* Multipath context */
  70struct multipath {
  71	unsigned long flags;		/* Multipath state flags */
  72
  73	spinlock_t lock;
  74	enum dm_queue_mode queue_mode;
  75
  76	struct pgpath *current_pgpath;
  77	struct priority_group *current_pg;
  78	struct priority_group *next_pg;	/* Switch to this PG if set */
  79
  80	atomic_t nr_valid_paths;	/* Total number of usable paths */
  81	unsigned nr_priority_groups;
  82	struct list_head priority_groups;
  83
  84	const char *hw_handler_name;
  85	char *hw_handler_params;
  86	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
  87	unsigned pg_init_retries;	/* Number of times to retry pg_init */
  88	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
  89	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
  90	atomic_t pg_init_count;		/* Number of times pg_init called */
  91
  92	struct mutex work_mutex;
  93	struct work_struct trigger_event;
  94	struct dm_target *ti;
  95
  96	struct work_struct process_queued_bios;
  97	struct bio_list queued_bios;
  98
  99	struct timer_list nopath_timer;	/* Timeout for queue_if_no_path */
 100};
 101
 102/*
 103 * Context information attached to each io we process.
 104 */
 105struct dm_mpath_io {
 106	struct pgpath *pgpath;
 107	size_t nr_bytes;
 108};
 109
 110typedef int (*action_fn) (struct pgpath *pgpath);
 111
 112static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 113static void trigger_event(struct work_struct *work);
 114static void activate_or_offline_path(struct pgpath *pgpath);
 115static void activate_path_work(struct work_struct *work);
 116static void process_queued_bios(struct work_struct *work);
 117static void queue_if_no_path_timeout_work(struct timer_list *t);
 118
 119/*-----------------------------------------------
 120 * Multipath state flags.
 121 *-----------------------------------------------*/
 122
 123#define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
 124#define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
 125#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
 126#define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3	/* If there's already a hw_handler present, don't change it. */
 127#define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
 128#define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
 129#define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
 130
 131static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
 132{
 133	bool r = test_bit(MPATHF_bit, &m->flags);
 134
 135	if (r) {
 136		unsigned long flags;
 137		spin_lock_irqsave(&m->lock, flags);
 138		r = test_bit(MPATHF_bit, &m->flags);
 139		spin_unlock_irqrestore(&m->lock, flags);
 140	}
 141
 142	return r;
 143}
 144
 145/*-----------------------------------------------
 146 * Allocation routines
 147 *-----------------------------------------------*/
 148
 149static struct pgpath *alloc_pgpath(void)
 150{
 151	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 152
 153	if (!pgpath)
 154		return NULL;
 155
 156	pgpath->is_active = true;
 157
 158	return pgpath;
 159}
 160
 161static void free_pgpath(struct pgpath *pgpath)
 162{
 163	kfree(pgpath);
 164}
 165
 166static struct priority_group *alloc_priority_group(void)
 167{
 168	struct priority_group *pg;
 169
 170	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 171
 172	if (pg)
 173		INIT_LIST_HEAD(&pg->pgpaths);
 174
 175	return pg;
 176}
 177
 178static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 179{
 180	struct pgpath *pgpath, *tmp;
 181
 182	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 183		list_del(&pgpath->list);
 184		dm_put_device(ti, pgpath->path.dev);
 185		free_pgpath(pgpath);
 186	}
 187}
 188
 189static void free_priority_group(struct priority_group *pg,
 190				struct dm_target *ti)
 191{
 192	struct path_selector *ps = &pg->ps;
 193
 194	if (ps->type) {
 195		ps->type->destroy(ps);
 196		dm_put_path_selector(ps->type);
 197	}
 198
 199	free_pgpaths(&pg->pgpaths, ti);
 200	kfree(pg);
 201}
 202
 203static struct multipath *alloc_multipath(struct dm_target *ti)
 204{
 205	struct multipath *m;
 206
 207	m = kzalloc(sizeof(*m), GFP_KERNEL);
 208	if (m) {
 209		INIT_LIST_HEAD(&m->priority_groups);
 210		spin_lock_init(&m->lock);
 211		atomic_set(&m->nr_valid_paths, 0);
 212		INIT_WORK(&m->trigger_event, trigger_event);
 213		mutex_init(&m->work_mutex);
 214
 215		m->queue_mode = DM_TYPE_NONE;
 216
 217		m->ti = ti;
 218		ti->private = m;
 219
 220		timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
 221	}
 222
 223	return m;
 224}
 225
 226static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 227{
 228	if (m->queue_mode == DM_TYPE_NONE) {
 229		m->queue_mode = DM_TYPE_REQUEST_BASED;
 230	} else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 231		INIT_WORK(&m->process_queued_bios, process_queued_bios);
 232		/*
 233		 * bio-based doesn't support any direct scsi_dh management;
 234		 * it just discovers if a scsi_dh is attached.
 235		 */
 236		set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 237	}
 238
 239	dm_table_set_type(ti->table, m->queue_mode);
 240
 241	/*
 242	 * Init fields that are only used when a scsi_dh is attached
 243	 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
 244	 */
 245	set_bit(MPATHF_QUEUE_IO, &m->flags);
 246	atomic_set(&m->pg_init_in_progress, 0);
 247	atomic_set(&m->pg_init_count, 0);
 248	m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 249	init_waitqueue_head(&m->pg_init_wait);
 250
 251	return 0;
 252}
 253
 254static void free_multipath(struct multipath *m)
 255{
 256	struct priority_group *pg, *tmp;
 257
 258	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 259		list_del(&pg->list);
 260		free_priority_group(pg, m->ti);
 261	}
 262
 263	kfree(m->hw_handler_name);
 264	kfree(m->hw_handler_params);
 265	mutex_destroy(&m->work_mutex);
 266	kfree(m);
 267}
 268
 269static struct dm_mpath_io *get_mpio(union map_info *info)
 270{
 271	return info->ptr;
 272}
 273
 274static size_t multipath_per_bio_data_size(void)
 275{
 276	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 277}
 278
 279static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 280{
 281	return dm_per_bio_data(bio, multipath_per_bio_data_size());
 282}
 283
 284static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
 285{
 286	/* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 287	void *bio_details = mpio + 1;
 288	return bio_details;
 289}
 290
 291static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 292{
 293	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 294	struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
 295
 296	mpio->nr_bytes = bio->bi_iter.bi_size;
 297	mpio->pgpath = NULL;
 298	*mpio_p = mpio;
 299
 300	dm_bio_record(bio_details, bio);
 301}
 302
 303/*-----------------------------------------------
 304 * Path selection
 305 *-----------------------------------------------*/
 306
 307static int __pg_init_all_paths(struct multipath *m)
 308{
 309	struct pgpath *pgpath;
 310	unsigned long pg_init_delay = 0;
 311
 312	lockdep_assert_held(&m->lock);
 313
 314	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 315		return 0;
 316
 317	atomic_inc(&m->pg_init_count);
 318	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 319
 320	/* Check here to reset pg_init_required */
 321	if (!m->current_pg)
 322		return 0;
 323
 324	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 325		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 326						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 327	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 328		/* Skip failed paths */
 329		if (!pgpath->is_active)
 330			continue;
 331		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 332				       pg_init_delay))
 333			atomic_inc(&m->pg_init_in_progress);
 334	}
 335	return atomic_read(&m->pg_init_in_progress);
 336}
 337
 338static int pg_init_all_paths(struct multipath *m)
 339{
 340	int ret;
 341	unsigned long flags;
 342
 343	spin_lock_irqsave(&m->lock, flags);
 344	ret = __pg_init_all_paths(m);
 345	spin_unlock_irqrestore(&m->lock, flags);
 346
 347	return ret;
 348}
 349
 350static void __switch_pg(struct multipath *m, struct priority_group *pg)
 351{
 352	lockdep_assert_held(&m->lock);
 353
 354	m->current_pg = pg;
 355
 356	/* Must we initialise the PG first, and queue I/O till it's ready? */
 357	if (m->hw_handler_name) {
 358		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 359		set_bit(MPATHF_QUEUE_IO, &m->flags);
 360	} else {
 361		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 362		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 363	}
 364
 365	atomic_set(&m->pg_init_count, 0);
 366}
 367
 368static struct pgpath *choose_path_in_pg(struct multipath *m,
 369					struct priority_group *pg,
 370					size_t nr_bytes)
 371{
 372	unsigned long flags;
 373	struct dm_path *path;
 374	struct pgpath *pgpath;
 375
 376	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 377	if (!path)
 378		return ERR_PTR(-ENXIO);
 379
 380	pgpath = path_to_pgpath(path);
 381
 382	if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 383		/* Only update current_pgpath if pg changed */
 384		spin_lock_irqsave(&m->lock, flags);
 385		m->current_pgpath = pgpath;
 386		__switch_pg(m, pg);
 387		spin_unlock_irqrestore(&m->lock, flags);
 388	}
 389
 390	return pgpath;
 391}
 392
 393static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 394{
 395	unsigned long flags;
 396	struct priority_group *pg;
 397	struct pgpath *pgpath;
 398	unsigned bypassed = 1;
 399
 400	if (!atomic_read(&m->nr_valid_paths)) {
 401		spin_lock_irqsave(&m->lock, flags);
 402		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 403		spin_unlock_irqrestore(&m->lock, flags);
 404		goto failed;
 405	}
 406
 407	/* Were we instructed to switch PG? */
 408	if (READ_ONCE(m->next_pg)) {
 409		spin_lock_irqsave(&m->lock, flags);
 410		pg = m->next_pg;
 411		if (!pg) {
 412			spin_unlock_irqrestore(&m->lock, flags);
 413			goto check_current_pg;
 414		}
 415		m->next_pg = NULL;
 416		spin_unlock_irqrestore(&m->lock, flags);
 417		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 418		if (!IS_ERR_OR_NULL(pgpath))
 419			return pgpath;
 420	}
 421
 422	/* Don't change PG until it has no remaining paths */
 423check_current_pg:
 424	pg = READ_ONCE(m->current_pg);
 425	if (pg) {
 426		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 427		if (!IS_ERR_OR_NULL(pgpath))
 428			return pgpath;
 429	}
 430
 431	/*
 432	 * Loop through priority groups until we find a valid path.
 433	 * First time we skip PGs marked 'bypassed'.
 434	 * Second time we only try the ones we skipped, but set
 435	 * pg_init_delay_retry so we do not hammer controllers.
 436	 */
 437	do {
 438		list_for_each_entry(pg, &m->priority_groups, list) {
 439			if (pg->bypassed == !!bypassed)
 440				continue;
 441			pgpath = choose_path_in_pg(m, pg, nr_bytes);
 442			if (!IS_ERR_OR_NULL(pgpath)) {
 443				if (!bypassed) {
 444					spin_lock_irqsave(&m->lock, flags);
 445					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 446					spin_unlock_irqrestore(&m->lock, flags);
 447				}
 448				return pgpath;
 449			}
 450		}
 451	} while (bypassed--);
 452
 453failed:
 454	spin_lock_irqsave(&m->lock, flags);
 455	m->current_pgpath = NULL;
 456	m->current_pg = NULL;
 457	spin_unlock_irqrestore(&m->lock, flags);
 458
 459	return NULL;
 460}
 461
 462/*
 463 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
 464 * report the function name and line number of the function from which
 465 * it has been invoked.
 466 */
 467#define dm_report_EIO(m)						\
 468do {									\
 469	struct mapped_device *md = dm_table_get_md((m)->ti->table);	\
 470									\
 471	DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
 472		      dm_device_name(md),				\
 473		      test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
 474		      test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
 475		      dm_noflush_suspending((m)->ti));			\
 476} while (0)
 477
 478/*
 479 * Check whether bios must be queued in the device-mapper core rather
 480 * than here in the target.
 481 */
 482static bool __must_push_back(struct multipath *m)
 483{
 484	return dm_noflush_suspending(m->ti);
 485}
 486
 487static bool must_push_back_rq(struct multipath *m)
 488{
 489	unsigned long flags;
 490	bool ret;
 491
 492	spin_lock_irqsave(&m->lock, flags);
 493	ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
 494	spin_unlock_irqrestore(&m->lock, flags);
 495
 496	return ret;
 497}
 498
 499/*
 500 * Map cloned requests (request-based multipath)
 501 */
 502static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 503				   union map_info *map_context,
 504				   struct request **__clone)
 505{
 506	struct multipath *m = ti->private;
 507	size_t nr_bytes = blk_rq_bytes(rq);
 508	struct pgpath *pgpath;
 509	struct block_device *bdev;
 510	struct dm_mpath_io *mpio = get_mpio(map_context);
 511	struct request_queue *q;
 512	struct request *clone;
 513
 514	/* Do we need to select a new pgpath? */
 515	pgpath = READ_ONCE(m->current_pgpath);
 516	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 517		pgpath = choose_pgpath(m, nr_bytes);
 518
 519	if (!pgpath) {
 520		if (must_push_back_rq(m))
 521			return DM_MAPIO_DELAY_REQUEUE;
 522		dm_report_EIO(m);	/* Failed */
 523		return DM_MAPIO_KILL;
 524	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 525		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 526		pg_init_all_paths(m);
 527		return DM_MAPIO_DELAY_REQUEUE;
 528	}
 529
 530	mpio->pgpath = pgpath;
 531	mpio->nr_bytes = nr_bytes;
 532
 533	bdev = pgpath->path.dev->bdev;
 534	q = bdev_get_queue(bdev);
 535	clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
 536			BLK_MQ_REQ_NOWAIT);
 537	if (IS_ERR(clone)) {
 538		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 539		if (blk_queue_dying(q)) {
 540			atomic_inc(&m->pg_init_in_progress);
 541			activate_or_offline_path(pgpath);
 542			return DM_MAPIO_DELAY_REQUEUE;
 543		}
 544
 545		/*
 546		 * blk-mq's SCHED_RESTART can cover this requeue, so we
 547		 * needn't deal with it by DELAY_REQUEUE. More importantly,
 548		 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
 549		 * get the queue busy feedback (via BLK_STS_RESOURCE),
 550		 * otherwise I/O merging can suffer.
 551		 */
 552		return DM_MAPIO_REQUEUE;
 553	}
 554	clone->bio = clone->biotail = NULL;
 555	clone->rq_disk = bdev->bd_disk;
 556	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 557	*__clone = clone;
 558
 559	if (pgpath->pg->ps.type->start_io)
 560		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 561					      &pgpath->path,
 562					      nr_bytes);
 563	return DM_MAPIO_REMAPPED;
 564}
 565
 566static void multipath_release_clone(struct request *clone,
 567				    union map_info *map_context)
 568{
 569	if (unlikely(map_context)) {
 570		/*
 571		 * non-NULL map_context means caller is still map
 572		 * method; must undo multipath_clone_and_map()
 573		 */
 574		struct dm_mpath_io *mpio = get_mpio(map_context);
 575		struct pgpath *pgpath = mpio->pgpath;
 576
 577		if (pgpath && pgpath->pg->ps.type->end_io)
 578			pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
 579						    &pgpath->path,
 580						    mpio->nr_bytes,
 581						    clone->io_start_time_ns);
 582	}
 583
 584	blk_put_request(clone);
 585}
 586
 587/*
 588 * Map cloned bios (bio-based multipath)
 589 */
 590
 591static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
 592{
 593	/* Queue for the daemon to resubmit */
 594	bio_list_add(&m->queued_bios, bio);
 595	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
 596		queue_work(kmultipathd, &m->process_queued_bios);
 597}
 598
 599static void multipath_queue_bio(struct multipath *m, struct bio *bio)
 600{
 601	unsigned long flags;
 602
 603	spin_lock_irqsave(&m->lock, flags);
 604	__multipath_queue_bio(m, bio);
 605	spin_unlock_irqrestore(&m->lock, flags);
 606}
 607
 608static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
 609{
 610	struct pgpath *pgpath;
 611	unsigned long flags;
 612
 613	/* Do we need to select a new pgpath? */
 614	pgpath = READ_ONCE(m->current_pgpath);
 615	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 616		pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
 617
 618	if (!pgpath) {
 619		spin_lock_irqsave(&m->lock, flags);
 620		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 621			__multipath_queue_bio(m, bio);
 622			pgpath = ERR_PTR(-EAGAIN);
 623		}
 624		spin_unlock_irqrestore(&m->lock, flags);
 625
 626	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 627		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 628		multipath_queue_bio(m, bio);
 629		pg_init_all_paths(m);
 630		return ERR_PTR(-EAGAIN);
 631	}
 632
 633	return pgpath;
 634}
 635
 636static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 637			       struct dm_mpath_io *mpio)
 638{
 639	struct pgpath *pgpath = __map_bio(m, bio);
 640
 641	if (IS_ERR(pgpath))
 642		return DM_MAPIO_SUBMITTED;
 643
 644	if (!pgpath) {
 645		if (__must_push_back(m))
 646			return DM_MAPIO_REQUEUE;
 647		dm_report_EIO(m);
 648		return DM_MAPIO_KILL;
 649	}
 650
 651	mpio->pgpath = pgpath;
 652
 653	bio->bi_status = 0;
 654	bio_set_dev(bio, pgpath->path.dev->bdev);
 655	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 656
 657	if (pgpath->pg->ps.type->start_io)
 658		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 659					      &pgpath->path,
 660					      mpio->nr_bytes);
 661	return DM_MAPIO_REMAPPED;
 662}
 663
 664static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 665{
 666	struct multipath *m = ti->private;
 667	struct dm_mpath_io *mpio = NULL;
 668
 669	multipath_init_per_bio_data(bio, &mpio);
 670	return __multipath_map_bio(m, bio, mpio);
 671}
 672
 673static void process_queued_io_list(struct multipath *m)
 674{
 675	if (m->queue_mode == DM_TYPE_REQUEST_BASED)
 676		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 677	else if (m->queue_mode == DM_TYPE_BIO_BASED)
 678		queue_work(kmultipathd, &m->process_queued_bios);
 679}
 680
 681static void process_queued_bios(struct work_struct *work)
 682{
 683	int r;
 684	unsigned long flags;
 685	struct bio *bio;
 686	struct bio_list bios;
 687	struct blk_plug plug;
 688	struct multipath *m =
 689		container_of(work, struct multipath, process_queued_bios);
 690
 691	bio_list_init(&bios);
 692
 693	spin_lock_irqsave(&m->lock, flags);
 694
 695	if (bio_list_empty(&m->queued_bios)) {
 696		spin_unlock_irqrestore(&m->lock, flags);
 697		return;
 698	}
 699
 700	bio_list_merge(&bios, &m->queued_bios);
 701	bio_list_init(&m->queued_bios);
 702
 703	spin_unlock_irqrestore(&m->lock, flags);
 704
 705	blk_start_plug(&plug);
 706	while ((bio = bio_list_pop(&bios))) {
 707		struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 708		dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
 709		r = __multipath_map_bio(m, bio, mpio);
 710		switch (r) {
 711		case DM_MAPIO_KILL:
 712			bio->bi_status = BLK_STS_IOERR;
 713			bio_endio(bio);
 714			break;
 715		case DM_MAPIO_REQUEUE:
 716			bio->bi_status = BLK_STS_DM_REQUEUE;
 717			bio_endio(bio);
 718			break;
 719		case DM_MAPIO_REMAPPED:
 720			submit_bio_noacct(bio);
 721			break;
 722		case DM_MAPIO_SUBMITTED:
 723			break;
 724		default:
 725			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 726		}
 727	}
 728	blk_finish_plug(&plug);
 729}
 730
 731/*
 732 * If we run out of usable paths, should we queue I/O or error it?
 733 */
 734static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 735			    bool save_old_value, const char *caller)
 736{
 737	unsigned long flags;
 738	bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
 739	const char *dm_dev_name = dm_device_name(dm_table_get_md(m->ti->table));
 740
 741	DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
 742		dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
 743
 744	spin_lock_irqsave(&m->lock, flags);
 745
 746	queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 747	saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 748
 749	if (save_old_value) {
 750		if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
 751			DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
 752			      dm_dev_name);
 753		} else
 754			assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
 755	} else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
 756		/* due to "fail_if_no_path" message, need to honor it. */
 757		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 758	}
 759	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
 760
 761	DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
 762		dm_dev_name, __func__,
 763		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
 764		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
 765		dm_noflush_suspending(m->ti));
 766
 767	spin_unlock_irqrestore(&m->lock, flags);
 768
 769	if (!queue_if_no_path) {
 770		dm_table_run_md_queue_async(m->ti->table);
 771		process_queued_io_list(m);
 772	}
 773
 774	return 0;
 775}
 776
 777/*
 778 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
 779 * process any queued I/O.
 780 */
 781static void queue_if_no_path_timeout_work(struct timer_list *t)
 782{
 783	struct multipath *m = from_timer(m, t, nopath_timer);
 784	struct mapped_device *md = dm_table_get_md(m->ti->table);
 785
 786	DMWARN("queue_if_no_path timeout on %s, failing queued IO", dm_device_name(md));
 
 787	queue_if_no_path(m, false, false, __func__);
 788}
 789
 790/*
 791 * Enable the queue_if_no_path timeout if necessary.
 792 * Called with m->lock held.
 793 */
 794static void enable_nopath_timeout(struct multipath *m)
 795{
 796	unsigned long queue_if_no_path_timeout =
 797		READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
 798
 799	lockdep_assert_held(&m->lock);
 800
 801	if (queue_if_no_path_timeout > 0 &&
 802	    atomic_read(&m->nr_valid_paths) == 0 &&
 803	    test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 804		mod_timer(&m->nopath_timer,
 805			  jiffies + queue_if_no_path_timeout);
 806	}
 807}
 808
 809static void disable_nopath_timeout(struct multipath *m)
 810{
 811	del_timer_sync(&m->nopath_timer);
 812}
 813
 814/*
 815 * An event is triggered whenever a path is taken out of use.
 816 * Includes path failure and PG bypass.
 817 */
 818static void trigger_event(struct work_struct *work)
 819{
 820	struct multipath *m =
 821		container_of(work, struct multipath, trigger_event);
 822
 823	dm_table_event(m->ti->table);
 824}
 825
 826/*-----------------------------------------------------------------
 827 * Constructor/argument parsing:
 828 * <#multipath feature args> [<arg>]*
 829 * <#hw_handler args> [hw_handler [<arg>]*]
 830 * <#priority groups>
 831 * <initial priority group>
 832 *     [<selector> <#selector args> [<arg>]*
 833 *      <#paths> <#per-path selector args>
 834 *         [<path> [<arg>]* ]+ ]+
 835 *---------------------------------------------------------------*/
 836static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 837			       struct dm_target *ti)
 838{
 839	int r;
 840	struct path_selector_type *pst;
 841	unsigned ps_argc;
 842
 843	static const struct dm_arg _args[] = {
 844		{0, 1024, "invalid number of path selector args"},
 845	};
 846
 847	pst = dm_get_path_selector(dm_shift_arg(as));
 848	if (!pst) {
 849		ti->error = "unknown path selector type";
 850		return -EINVAL;
 851	}
 852
 853	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 854	if (r) {
 855		dm_put_path_selector(pst);
 856		return -EINVAL;
 857	}
 858
 859	r = pst->create(&pg->ps, ps_argc, as->argv);
 860	if (r) {
 861		dm_put_path_selector(pst);
 862		ti->error = "path selector constructor failed";
 863		return r;
 864	}
 865
 866	pg->ps.type = pst;
 867	dm_consume_args(as, ps_argc);
 868
 869	return 0;
 870}
 871
 872static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
 873			 const char **attached_handler_name, char **error)
 874{
 875	struct request_queue *q = bdev_get_queue(bdev);
 876	int r;
 877
 878	if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
 879retain:
 880		if (*attached_handler_name) {
 881			/*
 882			 * Clear any hw_handler_params associated with a
 883			 * handler that isn't already attached.
 884			 */
 885			if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
 886				kfree(m->hw_handler_params);
 887				m->hw_handler_params = NULL;
 888			}
 889
 890			/*
 891			 * Reset hw_handler_name to match the attached handler
 892			 *
 893			 * NB. This modifies the table line to show the actual
 894			 * handler instead of the original table passed in.
 895			 */
 896			kfree(m->hw_handler_name);
 897			m->hw_handler_name = *attached_handler_name;
 898			*attached_handler_name = NULL;
 899		}
 900	}
 901
 902	if (m->hw_handler_name) {
 903		r = scsi_dh_attach(q, m->hw_handler_name);
 904		if (r == -EBUSY) {
 905			char b[BDEVNAME_SIZE];
 906
 907			printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 908			       bdevname(bdev, b));
 909			goto retain;
 910		}
 911		if (r < 0) {
 912			*error = "error attaching hardware handler";
 913			return r;
 914		}
 915
 916		if (m->hw_handler_params) {
 917			r = scsi_dh_set_params(q, m->hw_handler_params);
 918			if (r < 0) {
 919				*error = "unable to set hardware handler parameters";
 920				return r;
 921			}
 922		}
 923	}
 924
 925	return 0;
 926}
 927
 928static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 929				 struct dm_target *ti)
 930{
 931	int r;
 932	struct pgpath *p;
 933	struct multipath *m = ti->private;
 934	struct request_queue *q;
 935	const char *attached_handler_name = NULL;
 936
 937	/* we need at least a path arg */
 938	if (as->argc < 1) {
 939		ti->error = "no device given";
 940		return ERR_PTR(-EINVAL);
 941	}
 942
 943	p = alloc_pgpath();
 944	if (!p)
 945		return ERR_PTR(-ENOMEM);
 946
 947	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 948			  &p->path.dev);
 949	if (r) {
 950		ti->error = "error getting device";
 951		goto bad;
 952	}
 953
 954	q = bdev_get_queue(p->path.dev->bdev);
 955	attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 956	if (attached_handler_name || m->hw_handler_name) {
 957		INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
 958		r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
 959		kfree(attached_handler_name);
 960		if (r) {
 961			dm_put_device(ti, p->path.dev);
 962			goto bad;
 963		}
 964	}
 965
 966	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 967	if (r) {
 968		dm_put_device(ti, p->path.dev);
 969		goto bad;
 970	}
 971
 972	return p;
 973 bad:
 974	free_pgpath(p);
 975	return ERR_PTR(r);
 976}
 977
 978static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 979						   struct multipath *m)
 980{
 981	static const struct dm_arg _args[] = {
 982		{1, 1024, "invalid number of paths"},
 983		{0, 1024, "invalid number of selector args"}
 984	};
 985
 986	int r;
 987	unsigned i, nr_selector_args, nr_args;
 988	struct priority_group *pg;
 989	struct dm_target *ti = m->ti;
 990
 991	if (as->argc < 2) {
 992		as->argc = 0;
 993		ti->error = "not enough priority group arguments";
 994		return ERR_PTR(-EINVAL);
 995	}
 996
 997	pg = alloc_priority_group();
 998	if (!pg) {
 999		ti->error = "couldn't allocate priority group";
1000		return ERR_PTR(-ENOMEM);
1001	}
1002	pg->m = m;
1003
1004	r = parse_path_selector(as, pg, ti);
1005	if (r)
1006		goto bad;
1007
1008	/*
1009	 * read the paths
1010	 */
1011	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1012	if (r)
1013		goto bad;
1014
1015	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1016	if (r)
1017		goto bad;
1018
1019	nr_args = 1 + nr_selector_args;
1020	for (i = 0; i < pg->nr_pgpaths; i++) {
1021		struct pgpath *pgpath;
1022		struct dm_arg_set path_args;
1023
1024		if (as->argc < nr_args) {
1025			ti->error = "not enough path parameters";
1026			r = -EINVAL;
1027			goto bad;
1028		}
1029
1030		path_args.argc = nr_args;
1031		path_args.argv = as->argv;
1032
1033		pgpath = parse_path(&path_args, &pg->ps, ti);
1034		if (IS_ERR(pgpath)) {
1035			r = PTR_ERR(pgpath);
1036			goto bad;
1037		}
1038
1039		pgpath->pg = pg;
1040		list_add_tail(&pgpath->list, &pg->pgpaths);
1041		dm_consume_args(as, nr_args);
1042	}
1043
1044	return pg;
1045
1046 bad:
1047	free_priority_group(pg, ti);
1048	return ERR_PTR(r);
1049}
1050
1051static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1052{
1053	unsigned hw_argc;
1054	int ret;
1055	struct dm_target *ti = m->ti;
1056
1057	static const struct dm_arg _args[] = {
1058		{0, 1024, "invalid number of hardware handler args"},
1059	};
1060
1061	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1062		return -EINVAL;
1063
1064	if (!hw_argc)
1065		return 0;
1066
1067	if (m->queue_mode == DM_TYPE_BIO_BASED) {
1068		dm_consume_args(as, hw_argc);
1069		DMERR("bio-based multipath doesn't allow hardware handler args");
1070		return 0;
1071	}
1072
1073	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1074	if (!m->hw_handler_name)
1075		return -EINVAL;
1076
1077	if (hw_argc > 1) {
1078		char *p;
1079		int i, j, len = 4;
1080
1081		for (i = 0; i <= hw_argc - 2; i++)
1082			len += strlen(as->argv[i]) + 1;
1083		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1084		if (!p) {
1085			ti->error = "memory allocation failed";
1086			ret = -ENOMEM;
1087			goto fail;
1088		}
1089		j = sprintf(p, "%d", hw_argc - 1);
1090		for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1091			j = sprintf(p, "%s", as->argv[i]);
1092	}
1093	dm_consume_args(as, hw_argc - 1);
1094
1095	return 0;
1096fail:
1097	kfree(m->hw_handler_name);
1098	m->hw_handler_name = NULL;
1099	return ret;
1100}
1101
1102static int parse_features(struct dm_arg_set *as, struct multipath *m)
1103{
1104	int r;
1105	unsigned argc;
1106	struct dm_target *ti = m->ti;
1107	const char *arg_name;
1108
1109	static const struct dm_arg _args[] = {
1110		{0, 8, "invalid number of feature args"},
1111		{1, 50, "pg_init_retries must be between 1 and 50"},
1112		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1113	};
1114
1115	r = dm_read_arg_group(_args, as, &argc, &ti->error);
1116	if (r)
1117		return -EINVAL;
1118
1119	if (!argc)
1120		return 0;
1121
1122	do {
1123		arg_name = dm_shift_arg(as);
1124		argc--;
1125
1126		if (!strcasecmp(arg_name, "queue_if_no_path")) {
1127			r = queue_if_no_path(m, true, false, __func__);
1128			continue;
1129		}
1130
1131		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1132			set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1133			continue;
1134		}
1135
1136		if (!strcasecmp(arg_name, "pg_init_retries") &&
1137		    (argc >= 1)) {
1138			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1139			argc--;
1140			continue;
1141		}
1142
1143		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1144		    (argc >= 1)) {
1145			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1146			argc--;
1147			continue;
1148		}
1149
1150		if (!strcasecmp(arg_name, "queue_mode") &&
1151		    (argc >= 1)) {
1152			const char *queue_mode_name = dm_shift_arg(as);
1153
1154			if (!strcasecmp(queue_mode_name, "bio"))
1155				m->queue_mode = DM_TYPE_BIO_BASED;
1156			else if (!strcasecmp(queue_mode_name, "rq") ||
1157				 !strcasecmp(queue_mode_name, "mq"))
1158				m->queue_mode = DM_TYPE_REQUEST_BASED;
1159			else {
1160				ti->error = "Unknown 'queue_mode' requested";
1161				r = -EINVAL;
1162			}
1163			argc--;
1164			continue;
1165		}
1166
1167		ti->error = "Unrecognised multipath feature request";
1168		r = -EINVAL;
1169	} while (argc && !r);
1170
1171	return r;
1172}
1173
1174static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1175{
1176	/* target arguments */
1177	static const struct dm_arg _args[] = {
1178		{0, 1024, "invalid number of priority groups"},
1179		{0, 1024, "invalid initial priority group number"},
1180	};
1181
1182	int r;
1183	struct multipath *m;
1184	struct dm_arg_set as;
1185	unsigned pg_count = 0;
1186	unsigned next_pg_num;
1187	unsigned long flags;
1188
1189	as.argc = argc;
1190	as.argv = argv;
1191
1192	m = alloc_multipath(ti);
1193	if (!m) {
1194		ti->error = "can't allocate multipath";
1195		return -EINVAL;
1196	}
1197
1198	r = parse_features(&as, m);
1199	if (r)
1200		goto bad;
1201
1202	r = alloc_multipath_stage2(ti, m);
1203	if (r)
1204		goto bad;
1205
1206	r = parse_hw_handler(&as, m);
1207	if (r)
1208		goto bad;
1209
1210	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1211	if (r)
1212		goto bad;
1213
1214	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1215	if (r)
1216		goto bad;
1217
1218	if ((!m->nr_priority_groups && next_pg_num) ||
1219	    (m->nr_priority_groups && !next_pg_num)) {
1220		ti->error = "invalid initial priority group";
1221		r = -EINVAL;
1222		goto bad;
1223	}
1224
1225	/* parse the priority groups */
1226	while (as.argc) {
1227		struct priority_group *pg;
1228		unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1229
1230		pg = parse_priority_group(&as, m);
1231		if (IS_ERR(pg)) {
1232			r = PTR_ERR(pg);
1233			goto bad;
1234		}
1235
1236		nr_valid_paths += pg->nr_pgpaths;
1237		atomic_set(&m->nr_valid_paths, nr_valid_paths);
1238
1239		list_add_tail(&pg->list, &m->priority_groups);
1240		pg_count++;
1241		pg->pg_num = pg_count;
1242		if (!--next_pg_num)
1243			m->next_pg = pg;
1244	}
1245
1246	if (pg_count != m->nr_priority_groups) {
1247		ti->error = "priority group count mismatch";
1248		r = -EINVAL;
1249		goto bad;
1250	}
1251
1252	spin_lock_irqsave(&m->lock, flags);
1253	enable_nopath_timeout(m);
1254	spin_unlock_irqrestore(&m->lock, flags);
1255
1256	ti->num_flush_bios = 1;
1257	ti->num_discard_bios = 1;
1258	ti->num_write_same_bios = 1;
1259	ti->num_write_zeroes_bios = 1;
1260	if (m->queue_mode == DM_TYPE_BIO_BASED)
1261		ti->per_io_data_size = multipath_per_bio_data_size();
1262	else
1263		ti->per_io_data_size = sizeof(struct dm_mpath_io);
1264
1265	return 0;
1266
1267 bad:
1268	free_multipath(m);
1269	return r;
1270}
1271
1272static void multipath_wait_for_pg_init_completion(struct multipath *m)
1273{
1274	DEFINE_WAIT(wait);
1275
1276	while (1) {
1277		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1278
1279		if (!atomic_read(&m->pg_init_in_progress))
1280			break;
1281
1282		io_schedule();
1283	}
1284	finish_wait(&m->pg_init_wait, &wait);
1285}
1286
1287static void flush_multipath_work(struct multipath *m)
1288{
1289	if (m->hw_handler_name) {
1290		unsigned long flags;
1291
1292		if (!atomic_read(&m->pg_init_in_progress))
1293			goto skip;
1294
1295		spin_lock_irqsave(&m->lock, flags);
1296		if (atomic_read(&m->pg_init_in_progress) &&
1297		    !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1298			spin_unlock_irqrestore(&m->lock, flags);
1299
1300			flush_workqueue(kmpath_handlerd);
1301			multipath_wait_for_pg_init_completion(m);
1302
1303			spin_lock_irqsave(&m->lock, flags);
1304			clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1305		}
1306		spin_unlock_irqrestore(&m->lock, flags);
1307	}
1308skip:
1309	if (m->queue_mode == DM_TYPE_BIO_BASED)
1310		flush_work(&m->process_queued_bios);
1311	flush_work(&m->trigger_event);
1312}
1313
1314static void multipath_dtr(struct dm_target *ti)
1315{
1316	struct multipath *m = ti->private;
1317
1318	disable_nopath_timeout(m);
1319	flush_multipath_work(m);
1320	free_multipath(m);
1321}
1322
1323/*
1324 * Take a path out of use.
1325 */
1326static int fail_path(struct pgpath *pgpath)
1327{
1328	unsigned long flags;
1329	struct multipath *m = pgpath->pg->m;
1330
1331	spin_lock_irqsave(&m->lock, flags);
1332
1333	if (!pgpath->is_active)
1334		goto out;
1335
1336	DMWARN("%s: Failing path %s.",
1337	       dm_device_name(dm_table_get_md(m->ti->table)),
1338	       pgpath->path.dev->name);
1339
1340	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1341	pgpath->is_active = false;
1342	pgpath->fail_count++;
1343
1344	atomic_dec(&m->nr_valid_paths);
1345
1346	if (pgpath == m->current_pgpath)
1347		m->current_pgpath = NULL;
1348
1349	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1350		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1351
1352	schedule_work(&m->trigger_event);
1353
1354	enable_nopath_timeout(m);
1355
1356out:
1357	spin_unlock_irqrestore(&m->lock, flags);
1358
1359	return 0;
1360}
1361
1362/*
1363 * Reinstate a previously-failed path
1364 */
1365static int reinstate_path(struct pgpath *pgpath)
1366{
1367	int r = 0, run_queue = 0;
1368	unsigned long flags;
1369	struct multipath *m = pgpath->pg->m;
1370	unsigned nr_valid_paths;
1371
1372	spin_lock_irqsave(&m->lock, flags);
1373
1374	if (pgpath->is_active)
1375		goto out;
1376
1377	DMWARN("%s: Reinstating path %s.",
1378	       dm_device_name(dm_table_get_md(m->ti->table)),
1379	       pgpath->path.dev->name);
1380
1381	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1382	if (r)
1383		goto out;
1384
1385	pgpath->is_active = true;
1386
1387	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1388	if (nr_valid_paths == 1) {
1389		m->current_pgpath = NULL;
1390		run_queue = 1;
1391	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1392		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1393			atomic_inc(&m->pg_init_in_progress);
1394	}
1395
1396	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1397		       pgpath->path.dev->name, nr_valid_paths);
1398
1399	schedule_work(&m->trigger_event);
1400
1401out:
1402	spin_unlock_irqrestore(&m->lock, flags);
1403	if (run_queue) {
1404		dm_table_run_md_queue_async(m->ti->table);
1405		process_queued_io_list(m);
1406	}
1407
1408	if (pgpath->is_active)
1409		disable_nopath_timeout(m);
1410
1411	return r;
1412}
1413
1414/*
1415 * Fail or reinstate all paths that match the provided struct dm_dev.
1416 */
1417static int action_dev(struct multipath *m, struct dm_dev *dev,
1418		      action_fn action)
1419{
1420	int r = -EINVAL;
1421	struct pgpath *pgpath;
1422	struct priority_group *pg;
1423
1424	list_for_each_entry(pg, &m->priority_groups, list) {
1425		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1426			if (pgpath->path.dev == dev)
1427				r = action(pgpath);
1428		}
1429	}
1430
1431	return r;
1432}
1433
1434/*
1435 * Temporarily try to avoid having to use the specified PG
1436 */
1437static void bypass_pg(struct multipath *m, struct priority_group *pg,
1438		      bool bypassed)
1439{
1440	unsigned long flags;
1441
1442	spin_lock_irqsave(&m->lock, flags);
1443
1444	pg->bypassed = bypassed;
1445	m->current_pgpath = NULL;
1446	m->current_pg = NULL;
1447
1448	spin_unlock_irqrestore(&m->lock, flags);
1449
1450	schedule_work(&m->trigger_event);
1451}
1452
1453/*
1454 * Switch to using the specified PG from the next I/O that gets mapped
1455 */
1456static int switch_pg_num(struct multipath *m, const char *pgstr)
1457{
1458	struct priority_group *pg;
1459	unsigned pgnum;
1460	unsigned long flags;
1461	char dummy;
1462
1463	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1464	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1465		DMWARN("invalid PG number supplied to switch_pg_num");
1466		return -EINVAL;
1467	}
1468
1469	spin_lock_irqsave(&m->lock, flags);
1470	list_for_each_entry(pg, &m->priority_groups, list) {
1471		pg->bypassed = false;
1472		if (--pgnum)
1473			continue;
1474
1475		m->current_pgpath = NULL;
1476		m->current_pg = NULL;
1477		m->next_pg = pg;
1478	}
1479	spin_unlock_irqrestore(&m->lock, flags);
1480
1481	schedule_work(&m->trigger_event);
1482	return 0;
1483}
1484
1485/*
1486 * Set/clear bypassed status of a PG.
1487 * PGs are numbered upwards from 1 in the order they were declared.
1488 */
1489static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1490{
1491	struct priority_group *pg;
1492	unsigned pgnum;
1493	char dummy;
1494
1495	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1496	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1497		DMWARN("invalid PG number supplied to bypass_pg");
1498		return -EINVAL;
1499	}
1500
1501	list_for_each_entry(pg, &m->priority_groups, list) {
1502		if (!--pgnum)
1503			break;
1504	}
1505
1506	bypass_pg(m, pg, bypassed);
1507	return 0;
1508}
1509
1510/*
1511 * Should we retry pg_init immediately?
1512 */
1513static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1514{
1515	unsigned long flags;
1516	bool limit_reached = false;
1517
1518	spin_lock_irqsave(&m->lock, flags);
1519
1520	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1521	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1522		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1523	else
1524		limit_reached = true;
1525
1526	spin_unlock_irqrestore(&m->lock, flags);
1527
1528	return limit_reached;
1529}
1530
1531static void pg_init_done(void *data, int errors)
1532{
1533	struct pgpath *pgpath = data;
1534	struct priority_group *pg = pgpath->pg;
1535	struct multipath *m = pg->m;
1536	unsigned long flags;
1537	bool delay_retry = false;
1538
1539	/* device or driver problems */
1540	switch (errors) {
1541	case SCSI_DH_OK:
1542		break;
1543	case SCSI_DH_NOSYS:
1544		if (!m->hw_handler_name) {
1545			errors = 0;
1546			break;
1547		}
1548		DMERR("Could not failover the device: Handler scsi_dh_%s "
1549		      "Error %d.", m->hw_handler_name, errors);
1550		/*
1551		 * Fail path for now, so we do not ping pong
1552		 */
1553		fail_path(pgpath);
1554		break;
1555	case SCSI_DH_DEV_TEMP_BUSY:
1556		/*
1557		 * Probably doing something like FW upgrade on the
1558		 * controller so try the other pg.
1559		 */
1560		bypass_pg(m, pg, true);
1561		break;
1562	case SCSI_DH_RETRY:
1563		/* Wait before retrying. */
1564		delay_retry = true;
1565		fallthrough;
1566	case SCSI_DH_IMM_RETRY:
1567	case SCSI_DH_RES_TEMP_UNAVAIL:
1568		if (pg_init_limit_reached(m, pgpath))
1569			fail_path(pgpath);
1570		errors = 0;
1571		break;
1572	case SCSI_DH_DEV_OFFLINED:
1573	default:
1574		/*
1575		 * We probably do not want to fail the path for a device
1576		 * error, but this is what the old dm did. In future
1577		 * patches we can do more advanced handling.
1578		 */
1579		fail_path(pgpath);
1580	}
1581
1582	spin_lock_irqsave(&m->lock, flags);
1583	if (errors) {
1584		if (pgpath == m->current_pgpath) {
1585			DMERR("Could not failover device. Error %d.", errors);
1586			m->current_pgpath = NULL;
1587			m->current_pg = NULL;
1588		}
1589	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1590		pg->bypassed = false;
1591
1592	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1593		/* Activations of other paths are still on going */
1594		goto out;
1595
1596	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1597		if (delay_retry)
1598			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1599		else
1600			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1601
1602		if (__pg_init_all_paths(m))
1603			goto out;
1604	}
1605	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1606
1607	process_queued_io_list(m);
1608
1609	/*
1610	 * Wake up any thread waiting to suspend.
1611	 */
1612	wake_up(&m->pg_init_wait);
1613
1614out:
1615	spin_unlock_irqrestore(&m->lock, flags);
1616}
1617
1618static void activate_or_offline_path(struct pgpath *pgpath)
1619{
1620	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1621
1622	if (pgpath->is_active && !blk_queue_dying(q))
1623		scsi_dh_activate(q, pg_init_done, pgpath);
1624	else
1625		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1626}
1627
1628static void activate_path_work(struct work_struct *work)
1629{
1630	struct pgpath *pgpath =
1631		container_of(work, struct pgpath, activate_path.work);
1632
1633	activate_or_offline_path(pgpath);
1634}
1635
1636static int multipath_end_io(struct dm_target *ti, struct request *clone,
1637			    blk_status_t error, union map_info *map_context)
1638{
1639	struct dm_mpath_io *mpio = get_mpio(map_context);
1640	struct pgpath *pgpath = mpio->pgpath;
1641	int r = DM_ENDIO_DONE;
1642
1643	/*
1644	 * We don't queue any clone request inside the multipath target
1645	 * during end I/O handling, since those clone requests don't have
1646	 * bio clones.  If we queue them inside the multipath target,
1647	 * we need to make bio clones, that requires memory allocation.
1648	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1649	 *  don't have bio clones.)
1650	 * Instead of queueing the clone request here, we queue the original
1651	 * request into dm core, which will remake a clone request and
1652	 * clone bios for it and resubmit it later.
1653	 */
1654	if (error && blk_path_error(error)) {
1655		struct multipath *m = ti->private;
1656
1657		if (error == BLK_STS_RESOURCE)
1658			r = DM_ENDIO_DELAY_REQUEUE;
1659		else
1660			r = DM_ENDIO_REQUEUE;
1661
1662		if (pgpath)
1663			fail_path(pgpath);
1664
1665		if (!atomic_read(&m->nr_valid_paths) &&
1666		    !must_push_back_rq(m)) {
1667			if (error == BLK_STS_IOERR)
1668				dm_report_EIO(m);
1669			/* complete with the original error */
1670			r = DM_ENDIO_DONE;
1671		}
1672	}
1673
1674	if (pgpath) {
1675		struct path_selector *ps = &pgpath->pg->ps;
1676
1677		if (ps->type->end_io)
1678			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1679					 clone->io_start_time_ns);
1680	}
1681
1682	return r;
1683}
1684
1685static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1686				blk_status_t *error)
1687{
1688	struct multipath *m = ti->private;
1689	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1690	struct pgpath *pgpath = mpio->pgpath;
1691	unsigned long flags;
1692	int r = DM_ENDIO_DONE;
1693
1694	if (!*error || !blk_path_error(*error))
1695		goto done;
1696
1697	if (pgpath)
1698		fail_path(pgpath);
1699
1700	if (!atomic_read(&m->nr_valid_paths)) {
1701		spin_lock_irqsave(&m->lock, flags);
1702		if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1703			if (__must_push_back(m)) {
1704				r = DM_ENDIO_REQUEUE;
1705			} else {
1706				dm_report_EIO(m);
1707				*error = BLK_STS_IOERR;
1708			}
1709			spin_unlock_irqrestore(&m->lock, flags);
1710			goto done;
1711		}
1712		spin_unlock_irqrestore(&m->lock, flags);
1713	}
1714
1715	multipath_queue_bio(m, clone);
1716	r = DM_ENDIO_INCOMPLETE;
1717done:
1718	if (pgpath) {
1719		struct path_selector *ps = &pgpath->pg->ps;
1720
1721		if (ps->type->end_io)
1722			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1723					 dm_start_time_ns_from_clone(clone));
1724	}
1725
1726	return r;
1727}
1728
1729/*
1730 * Suspend with flush can't complete until all the I/O is processed
1731 * so if the last path fails we must error any remaining I/O.
1732 * - Note that if the freeze_bdev fails while suspending, the
1733 *   queue_if_no_path state is lost - userspace should reset it.
1734 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1735 */
1736static void multipath_presuspend(struct dm_target *ti)
1737{
1738	struct multipath *m = ti->private;
1739
1740	/* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1741	if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1742		queue_if_no_path(m, false, true, __func__);
1743}
1744
1745static void multipath_postsuspend(struct dm_target *ti)
1746{
1747	struct multipath *m = ti->private;
1748
1749	mutex_lock(&m->work_mutex);
1750	flush_multipath_work(m);
1751	mutex_unlock(&m->work_mutex);
1752}
1753
1754/*
1755 * Restore the queue_if_no_path setting.
1756 */
1757static void multipath_resume(struct dm_target *ti)
1758{
1759	struct multipath *m = ti->private;
1760	unsigned long flags;
1761
1762	spin_lock_irqsave(&m->lock, flags);
1763	if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1764		set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1765		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1766	}
1767
1768	DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1769		dm_device_name(dm_table_get_md(m->ti->table)), __func__,
1770		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1771		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1772
1773	spin_unlock_irqrestore(&m->lock, flags);
1774}
1775
1776/*
1777 * Info output has the following format:
1778 * num_multipath_feature_args [multipath_feature_args]*
1779 * num_handler_status_args [handler_status_args]*
1780 * num_groups init_group_number
1781 *            [A|D|E num_ps_status_args [ps_status_args]*
1782 *             num_paths num_selector_args
1783 *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1784 *
1785 * Table output has the following format (identical to the constructor string):
1786 * num_feature_args [features_args]*
1787 * num_handler_args hw_handler [hw_handler_args]*
1788 * num_groups init_group_number
1789 *     [priority selector-name num_ps_args [ps_args]*
1790 *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1791 */
1792static void multipath_status(struct dm_target *ti, status_type_t type,
1793			     unsigned status_flags, char *result, unsigned maxlen)
1794{
1795	int sz = 0;
1796	unsigned long flags;
1797	struct multipath *m = ti->private;
1798	struct priority_group *pg;
1799	struct pgpath *p;
1800	unsigned pg_num;
1801	char state;
1802
1803	spin_lock_irqsave(&m->lock, flags);
1804
1805	/* Features */
1806	if (type == STATUSTYPE_INFO)
1807		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1808		       atomic_read(&m->pg_init_count));
1809	else {
1810		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1811			      (m->pg_init_retries > 0) * 2 +
1812			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1813			      test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1814			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1815
1816		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1817			DMEMIT("queue_if_no_path ");
1818		if (m->pg_init_retries)
1819			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1820		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1821			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1822		if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1823			DMEMIT("retain_attached_hw_handler ");
1824		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1825			switch(m->queue_mode) {
1826			case DM_TYPE_BIO_BASED:
1827				DMEMIT("queue_mode bio ");
1828				break;
1829			default:
1830				WARN_ON_ONCE(true);
1831				break;
1832			}
1833		}
1834	}
1835
1836	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1837		DMEMIT("0 ");
1838	else
1839		DMEMIT("1 %s ", m->hw_handler_name);
1840
1841	DMEMIT("%u ", m->nr_priority_groups);
1842
1843	if (m->next_pg)
1844		pg_num = m->next_pg->pg_num;
1845	else if (m->current_pg)
1846		pg_num = m->current_pg->pg_num;
1847	else
1848		pg_num = (m->nr_priority_groups ? 1 : 0);
1849
1850	DMEMIT("%u ", pg_num);
1851
1852	switch (type) {
1853	case STATUSTYPE_INFO:
1854		list_for_each_entry(pg, &m->priority_groups, list) {
1855			if (pg->bypassed)
1856				state = 'D';	/* Disabled */
1857			else if (pg == m->current_pg)
1858				state = 'A';	/* Currently Active */
1859			else
1860				state = 'E';	/* Enabled */
1861
1862			DMEMIT("%c ", state);
1863
1864			if (pg->ps.type->status)
1865				sz += pg->ps.type->status(&pg->ps, NULL, type,
1866							  result + sz,
1867							  maxlen - sz);
1868			else
1869				DMEMIT("0 ");
1870
1871			DMEMIT("%u %u ", pg->nr_pgpaths,
1872			       pg->ps.type->info_args);
1873
1874			list_for_each_entry(p, &pg->pgpaths, list) {
1875				DMEMIT("%s %s %u ", p->path.dev->name,
1876				       p->is_active ? "A" : "F",
1877				       p->fail_count);
1878				if (pg->ps.type->status)
1879					sz += pg->ps.type->status(&pg->ps,
1880					      &p->path, type, result + sz,
1881					      maxlen - sz);
1882			}
1883		}
1884		break;
1885
1886	case STATUSTYPE_TABLE:
1887		list_for_each_entry(pg, &m->priority_groups, list) {
1888			DMEMIT("%s ", pg->ps.type->name);
1889
1890			if (pg->ps.type->status)
1891				sz += pg->ps.type->status(&pg->ps, NULL, type,
1892							  result + sz,
1893							  maxlen - sz);
1894			else
1895				DMEMIT("0 ");
1896
1897			DMEMIT("%u %u ", pg->nr_pgpaths,
1898			       pg->ps.type->table_args);
1899
1900			list_for_each_entry(p, &pg->pgpaths, list) {
1901				DMEMIT("%s ", p->path.dev->name);
1902				if (pg->ps.type->status)
1903					sz += pg->ps.type->status(&pg->ps,
1904					      &p->path, type, result + sz,
1905					      maxlen - sz);
1906			}
1907		}
1908		break;
1909	}
1910
1911	spin_unlock_irqrestore(&m->lock, flags);
1912}
1913
1914static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1915			     char *result, unsigned maxlen)
1916{
1917	int r = -EINVAL;
1918	struct dm_dev *dev;
1919	struct multipath *m = ti->private;
1920	action_fn action;
1921	unsigned long flags;
1922
1923	mutex_lock(&m->work_mutex);
1924
1925	if (dm_suspended(ti)) {
1926		r = -EBUSY;
1927		goto out;
1928	}
1929
1930	if (argc == 1) {
1931		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1932			r = queue_if_no_path(m, true, false, __func__);
1933			spin_lock_irqsave(&m->lock, flags);
1934			enable_nopath_timeout(m);
1935			spin_unlock_irqrestore(&m->lock, flags);
1936			goto out;
1937		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1938			r = queue_if_no_path(m, false, false, __func__);
1939			disable_nopath_timeout(m);
1940			goto out;
1941		}
1942	}
1943
1944	if (argc != 2) {
1945		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1946		goto out;
1947	}
1948
1949	if (!strcasecmp(argv[0], "disable_group")) {
1950		r = bypass_pg_num(m, argv[1], true);
1951		goto out;
1952	} else if (!strcasecmp(argv[0], "enable_group")) {
1953		r = bypass_pg_num(m, argv[1], false);
1954		goto out;
1955	} else if (!strcasecmp(argv[0], "switch_group")) {
1956		r = switch_pg_num(m, argv[1]);
1957		goto out;
1958	} else if (!strcasecmp(argv[0], "reinstate_path"))
1959		action = reinstate_path;
1960	else if (!strcasecmp(argv[0], "fail_path"))
1961		action = fail_path;
1962	else {
1963		DMWARN("Unrecognised multipath message received: %s", argv[0]);
1964		goto out;
1965	}
1966
1967	r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1968	if (r) {
1969		DMWARN("message: error getting device %s",
1970		       argv[1]);
1971		goto out;
1972	}
1973
1974	r = action_dev(m, dev, action);
1975
1976	dm_put_device(ti, dev);
1977
1978out:
1979	mutex_unlock(&m->work_mutex);
1980	return r;
1981}
1982
1983static int multipath_prepare_ioctl(struct dm_target *ti,
1984				   struct block_device **bdev)
1985{
1986	struct multipath *m = ti->private;
1987	struct pgpath *pgpath;
1988	unsigned long flags;
1989	int r;
1990
1991	pgpath = READ_ONCE(m->current_pgpath);
1992	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
1993		pgpath = choose_pgpath(m, 0);
1994
1995	if (pgpath) {
1996		if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
1997			*bdev = pgpath->path.dev->bdev;
1998			r = 0;
1999		} else {
2000			/* pg_init has not started or completed */
2001			r = -ENOTCONN;
2002		}
2003	} else {
2004		/* No path is available */
2005		r = -EIO;
2006		spin_lock_irqsave(&m->lock, flags);
2007		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2008			r = -ENOTCONN;
2009		spin_unlock_irqrestore(&m->lock, flags);
2010	}
2011
2012	if (r == -ENOTCONN) {
2013		if (!READ_ONCE(m->current_pg)) {
2014			/* Path status changed, redo selection */
2015			(void) choose_pgpath(m, 0);
2016		}
2017		spin_lock_irqsave(&m->lock, flags);
2018		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2019			(void) __pg_init_all_paths(m);
2020		spin_unlock_irqrestore(&m->lock, flags);
2021		dm_table_run_md_queue_async(m->ti->table);
2022		process_queued_io_list(m);
2023	}
2024
2025	/*
2026	 * Only pass ioctls through if the device sizes match exactly.
2027	 */
2028	if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
2029		return 1;
2030	return r;
2031}
2032
2033static int multipath_iterate_devices(struct dm_target *ti,
2034				     iterate_devices_callout_fn fn, void *data)
2035{
2036	struct multipath *m = ti->private;
2037	struct priority_group *pg;
2038	struct pgpath *p;
2039	int ret = 0;
2040
2041	list_for_each_entry(pg, &m->priority_groups, list) {
2042		list_for_each_entry(p, &pg->pgpaths, list) {
2043			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2044			if (ret)
2045				goto out;
2046		}
2047	}
2048
2049out:
2050	return ret;
2051}
2052
2053static int pgpath_busy(struct pgpath *pgpath)
2054{
2055	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2056
2057	return blk_lld_busy(q);
2058}
2059
2060/*
2061 * We return "busy", only when we can map I/Os but underlying devices
2062 * are busy (so even if we map I/Os now, the I/Os will wait on
2063 * the underlying queue).
2064 * In other words, if we want to kill I/Os or queue them inside us
2065 * due to map unavailability, we don't return "busy".  Otherwise,
2066 * dm core won't give us the I/Os and we can't do what we want.
2067 */
2068static int multipath_busy(struct dm_target *ti)
2069{
2070	bool busy = false, has_active = false;
2071	struct multipath *m = ti->private;
2072	struct priority_group *pg, *next_pg;
2073	struct pgpath *pgpath;
2074
2075	/* pg_init in progress */
2076	if (atomic_read(&m->pg_init_in_progress))
2077		return true;
2078
2079	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2080	if (!atomic_read(&m->nr_valid_paths)) {
2081		unsigned long flags;
2082		spin_lock_irqsave(&m->lock, flags);
2083		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2084			spin_unlock_irqrestore(&m->lock, flags);
2085			return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2086		}
2087		spin_unlock_irqrestore(&m->lock, flags);
2088	}
2089
2090	/* Guess which priority_group will be used at next mapping time */
2091	pg = READ_ONCE(m->current_pg);
2092	next_pg = READ_ONCE(m->next_pg);
2093	if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2094		pg = next_pg;
2095
2096	if (!pg) {
2097		/*
2098		 * We don't know which pg will be used at next mapping time.
2099		 * We don't call choose_pgpath() here to avoid to trigger
2100		 * pg_init just by busy checking.
2101		 * So we don't know whether underlying devices we will be using
2102		 * at next mapping time are busy or not. Just try mapping.
2103		 */
2104		return busy;
2105	}
2106
2107	/*
2108	 * If there is one non-busy active path at least, the path selector
2109	 * will be able to select it. So we consider such a pg as not busy.
2110	 */
2111	busy = true;
2112	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2113		if (pgpath->is_active) {
2114			has_active = true;
2115			if (!pgpath_busy(pgpath)) {
2116				busy = false;
2117				break;
2118			}
2119		}
2120	}
2121
2122	if (!has_active) {
2123		/*
2124		 * No active path in this pg, so this pg won't be used and
2125		 * the current_pg will be changed at next mapping time.
2126		 * We need to try mapping to determine it.
2127		 */
2128		busy = false;
2129	}
2130
2131	return busy;
2132}
2133
2134/*-----------------------------------------------------------------
2135 * Module setup
2136 *---------------------------------------------------------------*/
2137static struct target_type multipath_target = {
2138	.name = "multipath",
2139	.version = {1, 14, 0},
2140	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2141		    DM_TARGET_PASSES_INTEGRITY,
2142	.module = THIS_MODULE,
2143	.ctr = multipath_ctr,
2144	.dtr = multipath_dtr,
2145	.clone_and_map_rq = multipath_clone_and_map,
2146	.release_clone_rq = multipath_release_clone,
2147	.rq_end_io = multipath_end_io,
2148	.map = multipath_map_bio,
2149	.end_io = multipath_end_io_bio,
2150	.presuspend = multipath_presuspend,
2151	.postsuspend = multipath_postsuspend,
2152	.resume = multipath_resume,
2153	.status = multipath_status,
2154	.message = multipath_message,
2155	.prepare_ioctl = multipath_prepare_ioctl,
2156	.iterate_devices = multipath_iterate_devices,
2157	.busy = multipath_busy,
2158};
2159
2160static int __init dm_multipath_init(void)
2161{
2162	int r;
2163
2164	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2165	if (!kmultipathd) {
2166		DMERR("failed to create workqueue kmpathd");
2167		r = -ENOMEM;
2168		goto bad_alloc_kmultipathd;
2169	}
2170
2171	/*
2172	 * A separate workqueue is used to handle the device handlers
2173	 * to avoid overloading existing workqueue. Overloading the
2174	 * old workqueue would also create a bottleneck in the
2175	 * path of the storage hardware device activation.
2176	 */
2177	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2178						  WQ_MEM_RECLAIM);
2179	if (!kmpath_handlerd) {
2180		DMERR("failed to create workqueue kmpath_handlerd");
2181		r = -ENOMEM;
2182		goto bad_alloc_kmpath_handlerd;
2183	}
2184
2185	r = dm_register_target(&multipath_target);
2186	if (r < 0) {
2187		DMERR("request-based register failed %d", r);
2188		r = -EINVAL;
2189		goto bad_register_target;
2190	}
2191
2192	return 0;
2193
2194bad_register_target:
2195	destroy_workqueue(kmpath_handlerd);
2196bad_alloc_kmpath_handlerd:
2197	destroy_workqueue(kmultipathd);
2198bad_alloc_kmultipathd:
2199	return r;
2200}
2201
2202static void __exit dm_multipath_exit(void)
2203{
2204	destroy_workqueue(kmpath_handlerd);
2205	destroy_workqueue(kmultipathd);
2206
2207	dm_unregister_target(&multipath_target);
2208}
2209
2210module_init(dm_multipath_init);
2211module_exit(dm_multipath_exit);
2212
2213module_param_named(queue_if_no_path_timeout_secs,
2214		   queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2215MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2216
2217MODULE_DESCRIPTION(DM_NAME " multipath target");
2218MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2219MODULE_LICENSE("GPL");