dm-mpath.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2003 Sistina Software Limited.
   4 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   5 *
   6 * This file is released under the GPL.
   7 */
   8
   9#include <linux/device-mapper.h>
  10
  11#include "dm-rq.h"
  12#include "dm-bio-record.h"
  13#include "dm-path-selector.h"
  14#include "dm-uevent.h"
  15
  16#include <linux/blkdev.h>
  17#include <linux/ctype.h>
  18#include <linux/init.h>
  19#include <linux/mempool.h>
  20#include <linux/module.h>
  21#include <linux/pagemap.h>
  22#include <linux/slab.h>
  23#include <linux/time.h>
  24#include <linux/timer.h>
  25#include <linux/workqueue.h>
  26#include <linux/delay.h>
  27#include <scsi/scsi_dh.h>
  28#include <linux/atomic.h>
  29#include <linux/blk-mq.h>
  30
  31static struct workqueue_struct *dm_mpath_wq;
  32
  33#define DM_MSG_PREFIX "multipath"
  34#define DM_PG_INIT_DELAY_MSECS 2000
  35#define DM_PG_INIT_DELAY_DEFAULT ((unsigned int) -1)
  36#define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
  37
  38static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
  39
  40/* Path properties */
  41struct pgpath {
  42	struct list_head list;
  43
  44	struct priority_group *pg;	/* Owning PG */
  45	unsigned int fail_count;		/* Cumulative failure count */
  46
  47	struct dm_path path;
  48	struct delayed_work activate_path;
  49
  50	bool is_active:1;		/* Path status */
  51};
  52
  53#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  54
  55/*
  56 * Paths are grouped into Priority Groups and numbered from 1 upwards.
  57 * Each has a path selector which controls which path gets used.
  58 */
  59struct priority_group {
  60	struct list_head list;
  61
  62	struct multipath *m;		/* Owning multipath instance */
  63	struct path_selector ps;
  64
  65	unsigned int pg_num;		/* Reference number */
  66	unsigned int nr_pgpaths;		/* Number of paths in PG */
  67	struct list_head pgpaths;
  68
  69	bool bypassed:1;		/* Temporarily bypass this PG? */
  70};
  71
  72/* Multipath context */
  73struct multipath {
  74	unsigned long flags;		/* Multipath state flags */
  75
  76	spinlock_t lock;
  77	enum dm_queue_mode queue_mode;
  78
  79	struct pgpath *current_pgpath;
  80	struct priority_group *current_pg;
  81	struct priority_group *next_pg;	/* Switch to this PG if set */
  82	struct priority_group *last_probed_pg;
  83
  84	atomic_t nr_valid_paths;	/* Total number of usable paths */
  85	unsigned int nr_priority_groups;
  86	struct list_head priority_groups;
  87
  88	const char *hw_handler_name;
  89	char *hw_handler_params;
  90	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
  91	wait_queue_head_t probe_wait;   /* Wait for probing paths */
  92	unsigned int pg_init_retries;	/* Number of times to retry pg_init */
  93	unsigned int pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
  94	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
  95	atomic_t pg_init_count;		/* Number of times pg_init called */
  96
  97	struct mutex work_mutex;
  98	struct work_struct trigger_event;
  99	struct dm_target *ti;
 100
 101	struct work_struct process_queued_bios;
 102	struct bio_list queued_bios;
 103
 104	struct timer_list nopath_timer;	/* Timeout for queue_if_no_path */
 105	bool is_suspending;
 106};
 107
 108/*
 109 * Context information attached to each io we process.
 110 */
 111struct dm_mpath_io {
 112	struct pgpath *pgpath;
 113	size_t nr_bytes;
 114	u64 start_time_ns;
 115};
 116
 117typedef int (*action_fn) (struct pgpath *pgpath);
 118
 119static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 120static void trigger_event(struct work_struct *work);
 121static void activate_or_offline_path(struct pgpath *pgpath);
 122static void activate_path_work(struct work_struct *work);
 123static void process_queued_bios(struct work_struct *work);
 124static void queue_if_no_path_timeout_work(struct timer_list *t);
 125
 126/*
 127 *-----------------------------------------------
 128 * Multipath state flags.
 129 *-----------------------------------------------
 130 */
 131#define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
 132#define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
 133#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
 134/* MPATHF_RETAIN_ATTACHED_HW_HANDLER no longer has any effect */
 135#define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
 136#define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
 137#define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
 138#define MPATHF_DELAY_PG_SWITCH 7		/* Delay switching pg if it still has paths */
 139#define MPATHF_NEED_PG_SWITCH 8			/* Need to switch pgs after the delay has ended */
 140
 141static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
 142{
 143	bool r = test_bit(MPATHF_bit, &m->flags);
 144
 145	if (r) {
 146		unsigned long flags;
 147
 148		spin_lock_irqsave(&m->lock, flags);
 149		r = test_bit(MPATHF_bit, &m->flags);
 150		spin_unlock_irqrestore(&m->lock, flags);
 151	}
 152
 153	return r;
 154}
 155
 156/*
 157 *-----------------------------------------------
 158 * Allocation routines
 159 *-----------------------------------------------
 160 */
 161static struct pgpath *alloc_pgpath(void)
 162{
 163	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 164
 165	if (!pgpath)
 166		return NULL;
 167
 168	pgpath->is_active = true;
 169
 170	return pgpath;
 171}
 172
 173static void free_pgpath(struct pgpath *pgpath)
 174{
 175	kfree(pgpath);
 176}
 177
 178static struct priority_group *alloc_priority_group(void)
 179{
 180	struct priority_group *pg;
 181
 182	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 183
 184	if (pg)
 185		INIT_LIST_HEAD(&pg->pgpaths);
 186
 187	return pg;
 188}
 189
 190static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 191{
 192	struct pgpath *pgpath, *tmp;
 193
 194	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 195		list_del(&pgpath->list);
 196		dm_put_device(ti, pgpath->path.dev);
 197		free_pgpath(pgpath);
 198	}
 199}
 200
 201static void free_priority_group(struct priority_group *pg,
 202				struct dm_target *ti)
 203{
 204	struct path_selector *ps = &pg->ps;
 205
 206	if (ps->type) {
 207		ps->type->destroy(ps);
 208		dm_put_path_selector(ps->type);
 209	}
 210
 211	free_pgpaths(&pg->pgpaths, ti);
 212	kfree(pg);
 213}
 214
 215static struct multipath *alloc_multipath(struct dm_target *ti)
 216{
 217	struct multipath *m;
 218
 219	m = kzalloc(sizeof(*m), GFP_KERNEL);
 220	if (m) {
 221		INIT_LIST_HEAD(&m->priority_groups);
 222		spin_lock_init(&m->lock);
 223		atomic_set(&m->nr_valid_paths, 0);
 224		INIT_WORK(&m->trigger_event, trigger_event);
 225		mutex_init(&m->work_mutex);
 226
 227		m->queue_mode = DM_TYPE_NONE;
 228		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 229
 230		m->ti = ti;
 231		ti->private = m;
 232
 233		timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
 234	}
 235
 236	return m;
 237}
 238
 239static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 240{
 241	if (m->queue_mode == DM_TYPE_NONE)
 242		m->queue_mode = DM_TYPE_REQUEST_BASED;
 243	else if (m->queue_mode == DM_TYPE_BIO_BASED)
 244		INIT_WORK(&m->process_queued_bios, process_queued_bios);
 245
 246	dm_table_set_type(ti->table, m->queue_mode);
 247
 248	/*
 249	 * Init fields that are only used when a scsi_dh is attached
 250	 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
 251	 */
 252	set_bit(MPATHF_QUEUE_IO, &m->flags);
 253	atomic_set(&m->pg_init_in_progress, 0);
 254	atomic_set(&m->pg_init_count, 0);
 255	init_waitqueue_head(&m->pg_init_wait);
 256	init_waitqueue_head(&m->probe_wait);
 257
 258	return 0;
 259}
 260
 261static void free_multipath(struct multipath *m)
 262{
 263	struct priority_group *pg, *tmp;
 264
 265	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 266		list_del(&pg->list);
 267		free_priority_group(pg, m->ti);
 268	}
 269
 270	kfree(m->hw_handler_name);
 271	kfree(m->hw_handler_params);
 272	mutex_destroy(&m->work_mutex);
 273	kfree(m);
 274}
 275
 276static struct dm_mpath_io *get_mpio(union map_info *info)
 277{
 278	return info->ptr;
 279}
 280
 281static size_t multipath_per_bio_data_size(void)
 282{
 283	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 284}
 285
 286static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 287{
 288	return dm_per_bio_data(bio, multipath_per_bio_data_size());
 289}
 290
 291static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
 292{
 293	/* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 294	void *bio_details = mpio + 1;
 295	return bio_details;
 296}
 297
 298static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 299{
 300	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 301	struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
 302
 303	mpio->nr_bytes = bio->bi_iter.bi_size;
 304	mpio->pgpath = NULL;
 305	mpio->start_time_ns = 0;
 306	*mpio_p = mpio;
 307
 308	dm_bio_record(bio_details, bio);
 309}
 310
 311/*
 312 *-----------------------------------------------
 313 * Path selection
 314 *-----------------------------------------------
 315 */
 316static int __pg_init_all_paths(struct multipath *m)
 317{
 318	struct pgpath *pgpath;
 319	unsigned long pg_init_delay = 0;
 320
 321	lockdep_assert_held(&m->lock);
 322
 323	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 324		return 0;
 325
 326	atomic_inc(&m->pg_init_count);
 327	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 328
 329	/* Check here to reset pg_init_required */
 330	if (!m->current_pg)
 331		return 0;
 332
 333	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 334		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 335						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 336	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 337		/* Skip failed paths */
 338		if (!pgpath->is_active)
 339			continue;
 340		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 341				       pg_init_delay))
 342			atomic_inc(&m->pg_init_in_progress);
 343	}
 344	return atomic_read(&m->pg_init_in_progress);
 345}
 346
 347static int pg_init_all_paths(struct multipath *m)
 348{
 349	int ret;
 350	unsigned long flags;
 351
 352	spin_lock_irqsave(&m->lock, flags);
 353	ret = __pg_init_all_paths(m);
 354	spin_unlock_irqrestore(&m->lock, flags);
 355
 356	return ret;
 357}
 358
 359static void __switch_pg(struct multipath *m, struct priority_group *pg)
 360{
 361	lockdep_assert_held(&m->lock);
 362
 363	m->current_pg = pg;
 364
 365	/* Must we initialise the PG first, and queue I/O till it's ready? */
 366	if (m->hw_handler_name) {
 367		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 368		set_bit(MPATHF_QUEUE_IO, &m->flags);
 369	} else {
 370		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 371		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 372	}
 373
 374	atomic_set(&m->pg_init_count, 0);
 375}
 376
 377static struct pgpath *choose_path_in_pg(struct multipath *m,
 378					struct priority_group *pg,
 379					size_t nr_bytes)
 380{
 381	unsigned long flags;
 382	struct dm_path *path;
 383	struct pgpath *pgpath;
 384
 385	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 386	if (!path)
 387		return ERR_PTR(-ENXIO);
 388
 389	pgpath = path_to_pgpath(path);
 390
 391	if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 392		/* Only update current_pgpath if pg changed */
 393		spin_lock_irqsave(&m->lock, flags);
 394		m->current_pgpath = pgpath;
 395		__switch_pg(m, pg);
 396		spin_unlock_irqrestore(&m->lock, flags);
 397	}
 398
 399	return pgpath;
 400}
 401
 402static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 403{
 404	unsigned long flags;
 405	struct priority_group *pg;
 406	struct pgpath *pgpath;
 407	unsigned int bypassed = 1;
 408
 409	if (!atomic_read(&m->nr_valid_paths)) {
 410		spin_lock_irqsave(&m->lock, flags);
 411		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 412		spin_unlock_irqrestore(&m->lock, flags);
 413		goto failed;
 414	}
 415
 416	/* Don't change PG until it has no remaining paths */
 417	pg = READ_ONCE(m->current_pg);
 418	if (pg) {
 419		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 420		if (!IS_ERR_OR_NULL(pgpath))
 421			return pgpath;
 422	}
 423
 424	/* Were we instructed to switch PG? */
 425	if (READ_ONCE(m->next_pg)) {
 426		spin_lock_irqsave(&m->lock, flags);
 427		pg = m->next_pg;
 428		if (!pg) {
 429			spin_unlock_irqrestore(&m->lock, flags);
 430			goto check_all_pgs;
 431		}
 432		m->next_pg = NULL;
 433		spin_unlock_irqrestore(&m->lock, flags);
 434		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 435		if (!IS_ERR_OR_NULL(pgpath))
 436			return pgpath;
 437	}
 438check_all_pgs:
 439	/*
 440	 * Loop through priority groups until we find a valid path.
 441	 * First time we skip PGs marked 'bypassed'.
 442	 * Second time we only try the ones we skipped, but set
 443	 * pg_init_delay_retry so we do not hammer controllers.
 444	 */
 445	do {
 446		list_for_each_entry(pg, &m->priority_groups, list) {
 447			if (pg->bypassed == !!bypassed)
 448				continue;
 449			pgpath = choose_path_in_pg(m, pg, nr_bytes);
 450			if (!IS_ERR_OR_NULL(pgpath)) {
 451				if (!bypassed) {
 452					spin_lock_irqsave(&m->lock, flags);
 453					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 454					spin_unlock_irqrestore(&m->lock, flags);
 455				}
 456				return pgpath;
 457			}
 458		}
 459	} while (bypassed--);
 460
 461failed:
 462	spin_lock_irqsave(&m->lock, flags);
 463	m->current_pgpath = NULL;
 464	m->current_pg = NULL;
 465	spin_unlock_irqrestore(&m->lock, flags);
 466
 467	return NULL;
 468}
 469
 470/*
 471 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
 472 * report the function name and line number of the function from which
 473 * it has been invoked.
 474 */
 475#define dm_report_EIO(m)						\
 476	DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
 477		      dm_table_device_name((m)->ti->table),		\
 478		      test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
 479		      test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
 480		      dm_noflush_suspending((m)->ti))
 481
 482/*
 483 * Check whether bios must be queued in the device-mapper core rather
 484 * than here in the target.
 485 */
 486static bool __must_push_back(struct multipath *m)
 487{
 488	return dm_noflush_suspending(m->ti);
 489}
 490
 491static bool must_push_back_rq(struct multipath *m)
 492{
 493	unsigned long flags;
 494	bool ret;
 495
 496	spin_lock_irqsave(&m->lock, flags);
 497	ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
 498	spin_unlock_irqrestore(&m->lock, flags);
 499
 500	return ret;
 501}
 502
 503/*
 504 * Map cloned requests (request-based multipath)
 505 */
 506static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 507				   union map_info *map_context,
 508				   struct request **__clone)
 509{
 510	struct multipath *m = ti->private;
 511	size_t nr_bytes = blk_rq_bytes(rq);
 512	struct pgpath *pgpath;
 513	struct block_device *bdev;
 514	struct dm_mpath_io *mpio = get_mpio(map_context);
 515	struct request_queue *q;
 516	struct request *clone;
 517
 518	/* Do we need to select a new pgpath? */
 519	pgpath = READ_ONCE(m->current_pgpath);
 520	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 521		pgpath = choose_pgpath(m, nr_bytes);
 522
 523	if (!pgpath) {
 524		if (must_push_back_rq(m))
 525			return DM_MAPIO_DELAY_REQUEUE;
 526		dm_report_EIO(m);	/* Failed */
 527		return DM_MAPIO_KILL;
 528	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 529		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 530		pg_init_all_paths(m);
 531		return DM_MAPIO_DELAY_REQUEUE;
 532	}
 533
 534	mpio->pgpath = pgpath;
 535	mpio->nr_bytes = nr_bytes;
 536
 537	bdev = pgpath->path.dev->bdev;
 538	q = bdev_get_queue(bdev);
 539	clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
 540			BLK_MQ_REQ_NOWAIT);
 541	if (IS_ERR(clone)) {
 542		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 543		if (blk_queue_dying(q)) {
 544			atomic_inc(&m->pg_init_in_progress);
 545			activate_or_offline_path(pgpath);
 546			return DM_MAPIO_DELAY_REQUEUE;
 547		}
 548
 549		/*
 550		 * blk-mq's SCHED_RESTART can cover this requeue, so we
 551		 * needn't deal with it by DELAY_REQUEUE. More importantly,
 552		 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
 553		 * get the queue busy feedback (via BLK_STS_RESOURCE),
 554		 * otherwise I/O merging can suffer.
 555		 */
 556		return DM_MAPIO_REQUEUE;
 557	}
 558	clone->bio = clone->biotail = NULL;
 559	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 560	*__clone = clone;
 561
 562	if (pgpath->pg->ps.type->start_io)
 563		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 564					      &pgpath->path,
 565					      nr_bytes);
 566	return DM_MAPIO_REMAPPED;
 567}
 568
 569static void multipath_release_clone(struct request *clone,
 570				    union map_info *map_context)
 571{
 572	if (unlikely(map_context)) {
 573		/*
 574		 * non-NULL map_context means caller is still map
 575		 * method; must undo multipath_clone_and_map()
 576		 */
 577		struct dm_mpath_io *mpio = get_mpio(map_context);
 578		struct pgpath *pgpath = mpio->pgpath;
 579
 580		if (pgpath && pgpath->pg->ps.type->end_io)
 581			pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
 582						    &pgpath->path,
 583						    mpio->nr_bytes,
 584						    clone->io_start_time_ns);
 585	}
 586
 587	blk_mq_free_request(clone);
 588}
 589
 590/*
 591 * Map cloned bios (bio-based multipath)
 592 */
 593
 594static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
 595{
 596	/* Queue for the daemon to resubmit */
 597	bio_list_add(&m->queued_bios, bio);
 598	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
 599		queue_work(kmultipathd, &m->process_queued_bios);
 600}
 601
 602static void multipath_queue_bio(struct multipath *m, struct bio *bio)
 603{
 604	unsigned long flags;
 605
 606	spin_lock_irqsave(&m->lock, flags);
 607	__multipath_queue_bio(m, bio);
 608	spin_unlock_irqrestore(&m->lock, flags);
 609}
 610
 611static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
 612{
 613	struct pgpath *pgpath;
 614
 615	/* Do we need to select a new pgpath? */
 616	pgpath = READ_ONCE(m->current_pgpath);
 617	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 618		pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
 619
 620	if (!pgpath) {
 621		spin_lock_irq(&m->lock);
 622		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 623			__multipath_queue_bio(m, bio);
 624			pgpath = ERR_PTR(-EAGAIN);
 625		}
 626		spin_unlock_irq(&m->lock);
 627
 628	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 629		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 630		multipath_queue_bio(m, bio);
 631		pg_init_all_paths(m);
 632		return ERR_PTR(-EAGAIN);
 633	}
 634
 635	return pgpath;
 636}
 637
 638static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 639			       struct dm_mpath_io *mpio)
 640{
 641	struct pgpath *pgpath = __map_bio(m, bio);
 642
 643	if (IS_ERR(pgpath))
 644		return DM_MAPIO_SUBMITTED;
 645
 646	if (!pgpath) {
 647		if (__must_push_back(m))
 648			return DM_MAPIO_REQUEUE;
 649		dm_report_EIO(m);
 650		return DM_MAPIO_KILL;
 651	}
 652
 653	mpio->pgpath = pgpath;
 654
 655	if (dm_ps_use_hr_timer(pgpath->pg->ps.type))
 656		mpio->start_time_ns = ktime_get_ns();
 657
 658	bio->bi_status = 0;
 659	bio_set_dev(bio, pgpath->path.dev->bdev);
 660	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 661
 662	if (pgpath->pg->ps.type->start_io)
 663		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 664					      &pgpath->path,
 665					      mpio->nr_bytes);
 666	return DM_MAPIO_REMAPPED;
 667}
 668
 669static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 670{
 671	struct multipath *m = ti->private;
 672	struct dm_mpath_io *mpio = NULL;
 673
 674	multipath_init_per_bio_data(bio, &mpio);
 675	return __multipath_map_bio(m, bio, mpio);
 676}
 677
 678static void process_queued_io_list(struct multipath *m)
 679{
 680	if (m->queue_mode == DM_TYPE_REQUEST_BASED)
 681		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 682	else if (m->queue_mode == DM_TYPE_BIO_BASED)
 683		queue_work(kmultipathd, &m->process_queued_bios);
 684}
 685
 686static void process_queued_bios(struct work_struct *work)
 687{
 688	int r;
 689	struct bio *bio;
 690	struct bio_list bios;
 691	struct blk_plug plug;
 692	struct multipath *m =
 693		container_of(work, struct multipath, process_queued_bios);
 694
 695	bio_list_init(&bios);
 696
 697	spin_lock_irq(&m->lock);
 698
 699	if (bio_list_empty(&m->queued_bios)) {
 700		spin_unlock_irq(&m->lock);
 701		return;
 702	}
 703
 704	bio_list_merge_init(&bios, &m->queued_bios);
 705
 706	spin_unlock_irq(&m->lock);
 707
 708	blk_start_plug(&plug);
 709	while ((bio = bio_list_pop(&bios))) {
 710		struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 711
 712		dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
 713		r = __multipath_map_bio(m, bio, mpio);
 714		switch (r) {
 715		case DM_MAPIO_KILL:
 716			bio->bi_status = BLK_STS_IOERR;
 717			bio_endio(bio);
 718			break;
 719		case DM_MAPIO_REQUEUE:
 720			bio->bi_status = BLK_STS_DM_REQUEUE;
 721			bio_endio(bio);
 722			break;
 723		case DM_MAPIO_REMAPPED:
 724			submit_bio_noacct(bio);
 725			break;
 726		case DM_MAPIO_SUBMITTED:
 727			break;
 728		default:
 729			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 730		}
 731	}
 732	blk_finish_plug(&plug);
 733}
 734
 735/*
 736 * If we run out of usable paths, should we queue I/O or error it?
 737 */
 738static int queue_if_no_path(struct multipath *m, bool f_queue_if_no_path,
 739			    bool save_old_value, const char *caller)
 740{
 741	unsigned long flags;
 742	bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
 743	const char *dm_dev_name = dm_table_device_name(m->ti->table);
 744
 745	DMDEBUG("%s: %s caller=%s f_queue_if_no_path=%d save_old_value=%d",
 746		dm_dev_name, __func__, caller, f_queue_if_no_path, save_old_value);
 747
 748	spin_lock_irqsave(&m->lock, flags);
 749
 750	queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 751	saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 752
 753	if (save_old_value) {
 754		if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
 755			DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
 756			      dm_dev_name);
 757		} else
 758			assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
 759	} else if (!f_queue_if_no_path && saved_queue_if_no_path_bit) {
 760		/* due to "fail_if_no_path" message, need to honor it. */
 761		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 762	}
 763	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, f_queue_if_no_path);
 764
 765	DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
 766		dm_dev_name, __func__,
 767		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
 768		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
 769		dm_noflush_suspending(m->ti));
 770
 771	spin_unlock_irqrestore(&m->lock, flags);
 772
 773	if (!f_queue_if_no_path) {
 774		dm_table_run_md_queue_async(m->ti->table);
 775		process_queued_io_list(m);
 776	}
 777
 778	return 0;
 779}
 780
 781/*
 782 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
 783 * process any queued I/O.
 784 */
 785static void queue_if_no_path_timeout_work(struct timer_list *t)
 786{
 787	struct multipath *m = timer_container_of(m, t, nopath_timer);
 788
 789	DMWARN("queue_if_no_path timeout on %s, failing queued IO",
 790	       dm_table_device_name(m->ti->table));
 791	queue_if_no_path(m, false, false, __func__);
 792}
 793
 794/*
 795 * Enable the queue_if_no_path timeout if necessary.
 796 * Called with m->lock held.
 797 */
 798static void enable_nopath_timeout(struct multipath *m)
 799{
 800	unsigned long queue_if_no_path_timeout =
 801		READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
 802
 803	lockdep_assert_held(&m->lock);
 804
 805	if (queue_if_no_path_timeout > 0 &&
 806	    atomic_read(&m->nr_valid_paths) == 0 &&
 807	    test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 808		mod_timer(&m->nopath_timer,
 809			  jiffies + queue_if_no_path_timeout);
 810	}
 811}
 812
 813static void disable_nopath_timeout(struct multipath *m)
 814{
 815	timer_delete_sync(&m->nopath_timer);
 816}
 817
 818/*
 819 * An event is triggered whenever a path is taken out of use.
 820 * Includes path failure and PG bypass.
 821 */
 822static void trigger_event(struct work_struct *work)
 823{
 824	struct multipath *m =
 825		container_of(work, struct multipath, trigger_event);
 826
 827	dm_table_event(m->ti->table);
 828}
 829
 830/*
 831 *---------------------------------------------------------------
 832 * Constructor/argument parsing:
 833 * <#multipath feature args> [<arg>]*
 834 * <#hw_handler args> [hw_handler [<arg>]*]
 835 * <#priority groups>
 836 * <initial priority group>
 837 *     [<selector> <#selector args> [<arg>]*
 838 *      <#paths> <#per-path selector args>
 839 *         [<path> [<arg>]* ]+ ]+
 840 *---------------------------------------------------------------
 841 */
 842static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 843			       struct dm_target *ti)
 844{
 845	int r;
 846	struct path_selector_type *pst;
 847	unsigned int ps_argc;
 848
 849	static const struct dm_arg _args[] = {
 850		{0, 1024, "invalid number of path selector args"},
 851	};
 852
 853	pst = dm_get_path_selector(dm_shift_arg(as));
 854	if (!pst) {
 855		ti->error = "unknown path selector type";
 856		return -EINVAL;
 857	}
 858
 859	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 860	if (r) {
 861		dm_put_path_selector(pst);
 862		return -EINVAL;
 863	}
 864
 865	r = pst->create(&pg->ps, ps_argc, as->argv);
 866	if (r) {
 867		dm_put_path_selector(pst);
 868		ti->error = "path selector constructor failed";
 869		return r;
 870	}
 871
 872	pg->ps.type = pst;
 873	dm_consume_args(as, ps_argc);
 874
 875	return 0;
 876}
 877
 878static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
 879			 const char **attached_handler_name, char **error)
 880{
 881	struct request_queue *q = bdev_get_queue(bdev);
 882	int r;
 883
 884	if (*attached_handler_name) {
 885		/*
 886		 * Clear any hw_handler_params associated with a
 887		 * handler that isn't already attached.
 888		 */
 889		if (m->hw_handler_name && strcmp(*attached_handler_name,
 890						 m->hw_handler_name)) {
 891			kfree(m->hw_handler_params);
 892			m->hw_handler_params = NULL;
 893		}
 894
 895		/*
 896		 * Reset hw_handler_name to match the attached handler
 897		 *
 898		 * NB. This modifies the table line to show the actual
 899		 * handler instead of the original table passed in.
 900		 */
 901		kfree(m->hw_handler_name);
 902		m->hw_handler_name = *attached_handler_name;
 903		*attached_handler_name = NULL;
 904	}
 905
 906	if (m->hw_handler_name) {
 907		r = scsi_dh_attach(q, m->hw_handler_name);
 908		if (r < 0) {
 909			*error = "error attaching hardware handler";
 910			return r;
 911		}
 912
 913		if (m->hw_handler_params) {
 914			r = scsi_dh_set_params(q, m->hw_handler_params);
 915			if (r < 0) {
 916				*error = "unable to set hardware handler parameters";
 917				return r;
 918			}
 919		}
 920	}
 921
 922	return 0;
 923}
 924
 925static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 926				 struct dm_target *ti)
 927{
 928	int r;
 929	struct pgpath *p;
 930	struct multipath *m = ti->private;
 931	struct request_queue *q;
 932	const char *attached_handler_name = NULL;
 933
 934	/* we need at least a path arg */
 935	if (as->argc < 1) {
 936		ti->error = "no device given";
 937		return ERR_PTR(-EINVAL);
 938	}
 939
 940	p = alloc_pgpath();
 941	if (!p)
 942		return ERR_PTR(-ENOMEM);
 943
 944	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 945			  &p->path.dev);
 946	if (r) {
 947		ti->error = "error getting device";
 948		goto bad;
 949	}
 950
 951	q = bdev_get_queue(p->path.dev->bdev);
 952	attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 953	if (IS_ERR(attached_handler_name)) {
 954		if (PTR_ERR(attached_handler_name) == -ENODEV) {
 955			if (m->hw_handler_name) {
 956				DMERR("hardware handlers are only allowed for SCSI devices");
 957				kfree(m->hw_handler_name);
 958				m->hw_handler_name = NULL;
 959			}
 960			attached_handler_name = NULL;
 961		} else {
 962			r = PTR_ERR(attached_handler_name);
 963			ti->error = "error allocating handler name";
 964			goto bad_put_device;
 965		}
 966	}
 967	if (attached_handler_name || m->hw_handler_name) {
 968		INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
 969		r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
 970		kfree(attached_handler_name);
 971		if (r)
 972			goto bad_put_device;
 973	}
 974
 975	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 976	if (r)
 977		goto bad_put_device;
 978
 979	return p;
 980
 981bad_put_device:
 982	dm_put_device(ti, p->path.dev);
 983bad:
 984	free_pgpath(p);
 985	return ERR_PTR(r);
 986}
 987
 988static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 989						   struct multipath *m)
 990{
 991	static const struct dm_arg _args[] = {
 992		{1, 1024, "invalid number of paths"},
 993		{0, 1024, "invalid number of selector args"}
 994	};
 995
 996	int r;
 997	unsigned int i, nr_selector_args, nr_args;
 998	struct priority_group *pg;
 999	struct dm_target *ti = m->ti;
1000
1001	if (as->argc < 2) {
1002		as->argc = 0;
1003		ti->error = "not enough priority group arguments";
1004		return ERR_PTR(-EINVAL);
1005	}
1006
1007	pg = alloc_priority_group();
1008	if (!pg) {
1009		ti->error = "couldn't allocate priority group";
1010		return ERR_PTR(-ENOMEM);
1011	}
1012	pg->m = m;
1013
1014	r = parse_path_selector(as, pg, ti);
1015	if (r)
1016		goto bad;
1017
1018	/*
1019	 * read the paths
1020	 */
1021	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1022	if (r)
1023		goto bad;
1024
1025	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1026	if (r)
1027		goto bad;
1028
1029	nr_args = 1 + nr_selector_args;
1030	for (i = 0; i < pg->nr_pgpaths; i++) {
1031		struct pgpath *pgpath;
1032		struct dm_arg_set path_args;
1033
1034		if (as->argc < nr_args) {
1035			ti->error = "not enough path parameters";
1036			r = -EINVAL;
1037			goto bad;
1038		}
1039
1040		path_args.argc = nr_args;
1041		path_args.argv = as->argv;
1042
1043		pgpath = parse_path(&path_args, &pg->ps, ti);
1044		if (IS_ERR(pgpath)) {
1045			r = PTR_ERR(pgpath);
1046			goto bad;
1047		}
1048
1049		pgpath->pg = pg;
1050		list_add_tail(&pgpath->list, &pg->pgpaths);
1051		dm_consume_args(as, nr_args);
1052	}
1053
1054	return pg;
1055
1056 bad:
1057	free_priority_group(pg, ti);
1058	return ERR_PTR(r);
1059}
1060
1061static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1062{
1063	unsigned int hw_argc;
1064	int ret;
1065	struct dm_target *ti = m->ti;
1066
1067	static const struct dm_arg _args[] = {
1068		{0, 1024, "invalid number of hardware handler args"},
1069	};
1070
1071	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1072		return -EINVAL;
1073
1074	if (!hw_argc)
1075		return 0;
1076
1077	if (m->queue_mode == DM_TYPE_BIO_BASED) {
1078		dm_consume_args(as, hw_argc);
1079		DMERR("bio-based multipath doesn't allow hardware handler args");
1080		return 0;
1081	}
1082
1083	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1084	if (!m->hw_handler_name)
1085		return -EINVAL;
1086
1087	if (hw_argc > 1) {
1088		char *p;
1089		int i, j, len = 4;
1090
1091		for (i = 0; i <= hw_argc - 2; i++)
1092			len += strlen(as->argv[i]) + 1;
1093		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1094		if (!p) {
1095			ti->error = "memory allocation failed";
1096			ret = -ENOMEM;
1097			goto fail;
1098		}
1099		j = sprintf(p, "%d", hw_argc - 1);
1100		for (i = 0, p += j + 1; i <= hw_argc - 2; i++, p += j + 1)
1101			j = sprintf(p, "%s", as->argv[i]);
1102	}
1103	dm_consume_args(as, hw_argc - 1);
1104
1105	return 0;
1106fail:
1107	kfree(m->hw_handler_name);
1108	m->hw_handler_name = NULL;
1109	return ret;
1110}
1111
1112static int parse_features(struct dm_arg_set *as, struct multipath *m)
1113{
1114	int r;
1115	unsigned int argc;
1116	struct dm_target *ti = m->ti;
1117	const char *arg_name;
1118
1119	static const struct dm_arg _args[] = {
1120		{0, 8, "invalid number of feature args"},
1121		{1, 50, "pg_init_retries must be between 1 and 50"},
1122		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1123	};
1124
1125	r = dm_read_arg_group(_args, as, &argc, &ti->error);
1126	if (r)
1127		return -EINVAL;
1128
1129	if (!argc)
1130		return 0;
1131
1132	do {
1133		arg_name = dm_shift_arg(as);
1134		argc--;
1135
1136		if (!strcasecmp(arg_name, "queue_if_no_path")) {
1137			r = queue_if_no_path(m, true, false, __func__);
1138			continue;
1139		}
1140
1141		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1142			/* no longer has any effect */
1143			continue;
1144		}
1145
1146		if (!strcasecmp(arg_name, "pg_init_retries") &&
1147		    (argc >= 1)) {
1148			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1149			argc--;
1150			continue;
1151		}
1152
1153		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1154		    (argc >= 1)) {
1155			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1156			argc--;
1157			continue;
1158		}
1159
1160		if (!strcasecmp(arg_name, "queue_mode") &&
1161		    (argc >= 1)) {
1162			const char *queue_mode_name = dm_shift_arg(as);
1163
1164			if (!strcasecmp(queue_mode_name, "bio"))
1165				m->queue_mode = DM_TYPE_BIO_BASED;
1166			else if (!strcasecmp(queue_mode_name, "rq") ||
1167				 !strcasecmp(queue_mode_name, "mq"))
1168				m->queue_mode = DM_TYPE_REQUEST_BASED;
1169			else {
1170				ti->error = "Unknown 'queue_mode' requested";
1171				r = -EINVAL;
1172			}
1173			argc--;
1174			continue;
1175		}
1176
1177		ti->error = "Unrecognised multipath feature request";
1178		r = -EINVAL;
1179	} while (argc && !r);
1180
1181	return r;
1182}
1183
1184static int multipath_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1185{
1186	/* target arguments */
1187	static const struct dm_arg _args[] = {
1188		{0, 1024, "invalid number of priority groups"},
1189		{0, 1024, "invalid initial priority group number"},
1190	};
1191
1192	int r;
1193	struct multipath *m;
1194	struct dm_arg_set as;
1195	unsigned int pg_count = 0;
1196	unsigned int next_pg_num;
1197
1198	as.argc = argc;
1199	as.argv = argv;
1200
1201	m = alloc_multipath(ti);
1202	if (!m) {
1203		ti->error = "can't allocate multipath";
1204		return -EINVAL;
1205	}
1206
1207	r = parse_features(&as, m);
1208	if (r)
1209		goto bad;
1210
1211	r = alloc_multipath_stage2(ti, m);
1212	if (r)
1213		goto bad;
1214
1215	r = parse_hw_handler(&as, m);
1216	if (r)
1217		goto bad;
1218
1219	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1220	if (r)
1221		goto bad;
1222
1223	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1224	if (r)
1225		goto bad;
1226
1227	if ((!m->nr_priority_groups && next_pg_num) ||
1228	    (m->nr_priority_groups && !next_pg_num)) {
1229		ti->error = "invalid initial priority group";
1230		r = -EINVAL;
1231		goto bad;
1232	}
1233
1234	/* parse the priority groups */
1235	while (as.argc) {
1236		struct priority_group *pg;
1237		unsigned int nr_valid_paths = atomic_read(&m->nr_valid_paths);
1238
1239		pg = parse_priority_group(&as, m);
1240		if (IS_ERR(pg)) {
1241			r = PTR_ERR(pg);
1242			goto bad;
1243		}
1244
1245		nr_valid_paths += pg->nr_pgpaths;
1246		atomic_set(&m->nr_valid_paths, nr_valid_paths);
1247
1248		list_add_tail(&pg->list, &m->priority_groups);
1249		pg_count++;
1250		pg->pg_num = pg_count;
1251		if (!--next_pg_num)
1252			m->next_pg = pg;
1253	}
1254
1255	if (pg_count != m->nr_priority_groups) {
1256		ti->error = "priority group count mismatch";
1257		r = -EINVAL;
1258		goto bad;
1259	}
1260
1261	spin_lock_irq(&m->lock);
1262	enable_nopath_timeout(m);
1263	spin_unlock_irq(&m->lock);
1264
1265	ti->num_flush_bios = 1;
1266	ti->num_discard_bios = 1;
1267	ti->num_write_zeroes_bios = 1;
1268	if (m->queue_mode == DM_TYPE_BIO_BASED)
1269		ti->per_io_data_size = multipath_per_bio_data_size();
1270	else
1271		ti->per_io_data_size = sizeof(struct dm_mpath_io);
1272
1273	return 0;
1274
1275 bad:
1276	free_multipath(m);
1277	return r;
1278}
1279
1280static void multipath_wait_for_pg_init_completion(struct multipath *m)
1281{
1282	DEFINE_WAIT(wait);
1283
1284	while (1) {
1285		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1286
1287		if (!atomic_read(&m->pg_init_in_progress))
1288			break;
1289
1290		io_schedule();
1291	}
1292	finish_wait(&m->pg_init_wait, &wait);
1293}
1294
1295static void flush_multipath_work(struct multipath *m)
1296{
1297	if (m->hw_handler_name) {
1298		if (!atomic_read(&m->pg_init_in_progress))
1299			goto skip;
1300
1301		spin_lock_irq(&m->lock);
1302		if (atomic_read(&m->pg_init_in_progress) &&
1303		    !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1304			spin_unlock_irq(&m->lock);
1305
1306			flush_workqueue(kmpath_handlerd);
1307			multipath_wait_for_pg_init_completion(m);
1308
1309			spin_lock_irq(&m->lock);
1310			clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1311		}
1312		spin_unlock_irq(&m->lock);
1313	}
1314skip:
1315	if (m->queue_mode == DM_TYPE_BIO_BASED)
1316		flush_work(&m->process_queued_bios);
1317	flush_work(&m->trigger_event);
1318}
1319
1320static void multipath_dtr(struct dm_target *ti)
1321{
1322	struct multipath *m = ti->private;
1323
1324	disable_nopath_timeout(m);
1325	flush_multipath_work(m);
1326	free_multipath(m);
1327}
1328
1329/*
1330 * Take a path out of use.
1331 */
1332static int fail_path(struct pgpath *pgpath)
1333{
1334	unsigned long flags;
1335	struct multipath *m = pgpath->pg->m;
1336
1337	spin_lock_irqsave(&m->lock, flags);
1338
1339	if (!pgpath->is_active)
1340		goto out;
1341
1342	DMWARN("%s: Failing path %s.",
1343	       dm_table_device_name(m->ti->table),
1344	       pgpath->path.dev->name);
1345
1346	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1347	pgpath->is_active = false;
1348	pgpath->fail_count++;
1349
1350	atomic_dec(&m->nr_valid_paths);
1351
1352	if (pgpath == m->current_pgpath)
1353		m->current_pgpath = NULL;
1354
1355	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1356		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1357
1358	queue_work(dm_mpath_wq, &m->trigger_event);
1359
1360	enable_nopath_timeout(m);
1361
1362out:
1363	spin_unlock_irqrestore(&m->lock, flags);
1364
1365	return 0;
1366}
1367
1368/*
1369 * Reinstate a previously-failed path
1370 */
1371static int reinstate_path(struct pgpath *pgpath)
1372{
1373	int r = 0, run_queue = 0;
1374	struct multipath *m = pgpath->pg->m;
1375	unsigned int nr_valid_paths;
1376
1377	spin_lock_irq(&m->lock);
1378
1379	if (pgpath->is_active)
1380		goto out;
1381
1382	DMWARN("%s: Reinstating path %s.",
1383	       dm_table_device_name(m->ti->table),
1384	       pgpath->path.dev->name);
1385
1386	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1387	if (r)
1388		goto out;
1389
1390	pgpath->is_active = true;
1391
1392	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1393	if (nr_valid_paths == 1) {
1394		m->current_pgpath = NULL;
1395		run_queue = 1;
1396	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1397		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1398			atomic_inc(&m->pg_init_in_progress);
1399	}
1400
1401	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1402		       pgpath->path.dev->name, nr_valid_paths);
1403
1404	schedule_work(&m->trigger_event);
1405
1406out:
1407	spin_unlock_irq(&m->lock);
1408	if (run_queue) {
1409		dm_table_run_md_queue_async(m->ti->table);
1410		process_queued_io_list(m);
1411	}
1412
1413	if (pgpath->is_active)
1414		disable_nopath_timeout(m);
1415
1416	return r;
1417}
1418
1419/*
1420 * Fail or reinstate all paths that match the provided struct dm_dev.
1421 */
1422static int action_dev(struct multipath *m, dev_t dev, action_fn action)
1423{
1424	int r = -EINVAL;
1425	struct pgpath *pgpath;
1426	struct priority_group *pg;
1427
1428	list_for_each_entry(pg, &m->priority_groups, list) {
1429		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1430			if (pgpath->path.dev->bdev->bd_dev == dev)
1431				r = action(pgpath);
1432		}
1433	}
1434
1435	return r;
1436}
1437
1438/*
1439 * Temporarily try to avoid having to use the specified PG
1440 */
1441static void bypass_pg(struct multipath *m, struct priority_group *pg,
1442		      bool bypassed, bool can_be_delayed)
1443{
1444	unsigned long flags;
1445
1446	spin_lock_irqsave(&m->lock, flags);
1447
1448	pg->bypassed = bypassed;
1449	if (can_be_delayed && test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags))
1450		set_bit(MPATHF_NEED_PG_SWITCH, &m->flags);
1451	else {
1452		m->current_pgpath = NULL;
1453		m->current_pg = NULL;
1454	}
1455
1456	spin_unlock_irqrestore(&m->lock, flags);
1457
1458	schedule_work(&m->trigger_event);
1459}
1460
1461/*
1462 * Switch to using the specified PG from the next I/O that gets mapped
1463 */
1464static int switch_pg_num(struct multipath *m, const char *pgstr)
1465{
1466	struct priority_group *pg;
1467	unsigned int pgnum;
1468	char dummy;
1469
1470	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1471	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1472		DMWARN("invalid PG number supplied to %s", __func__);
1473		return -EINVAL;
1474	}
1475
1476	spin_lock_irq(&m->lock);
1477	list_for_each_entry(pg, &m->priority_groups, list) {
1478		pg->bypassed = false;
1479		if (--pgnum)
1480			continue;
1481
1482		if (test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags))
1483			set_bit(MPATHF_NEED_PG_SWITCH, &m->flags);
1484		else {
1485			m->current_pgpath = NULL;
1486			m->current_pg = NULL;
1487		}
1488		m->next_pg = pg;
1489	}
1490	spin_unlock_irq(&m->lock);
1491
1492	schedule_work(&m->trigger_event);
1493	return 0;
1494}
1495
1496/*
1497 * Set/clear bypassed status of a PG.
1498 * PGs are numbered upwards from 1 in the order they were declared.
1499 */
1500static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1501{
1502	struct priority_group *pg;
1503	unsigned int pgnum;
1504	char dummy;
1505
1506	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1507	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1508		DMWARN("invalid PG number supplied to bypass_pg");
1509		return -EINVAL;
1510	}
1511
1512	list_for_each_entry(pg, &m->priority_groups, list) {
1513		if (!--pgnum)
1514			break;
1515	}
1516
1517	bypass_pg(m, pg, bypassed, true);
1518	return 0;
1519}
1520
1521/*
1522 * Should we retry pg_init immediately?
1523 */
1524static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1525{
1526	unsigned long flags;
1527	bool limit_reached = false;
1528
1529	spin_lock_irqsave(&m->lock, flags);
1530
1531	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1532	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1533		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1534	else
1535		limit_reached = true;
1536
1537	spin_unlock_irqrestore(&m->lock, flags);
1538
1539	return limit_reached;
1540}
1541
1542static void pg_init_done(void *data, int errors)
1543{
1544	struct pgpath *pgpath = data;
1545	struct priority_group *pg = pgpath->pg;
1546	struct multipath *m = pg->m;
1547	unsigned long flags;
1548	bool delay_retry = false;
1549
1550	/* device or driver problems */
1551	switch (errors) {
1552	case SCSI_DH_OK:
1553		break;
1554	case SCSI_DH_NOSYS:
1555		if (!m->hw_handler_name) {
1556			errors = 0;
1557			break;
1558		}
1559		DMERR("Could not failover the device: Handler scsi_dh_%s "
1560		      "Error %d.", m->hw_handler_name, errors);
1561		/*
1562		 * Fail path for now, so we do not ping pong
1563		 */
1564		fail_path(pgpath);
1565		break;
1566	case SCSI_DH_DEV_TEMP_BUSY:
1567		/*
1568		 * Probably doing something like FW upgrade on the
1569		 * controller so try the other pg.
1570		 */
1571		bypass_pg(m, pg, true, false);
1572		break;
1573	case SCSI_DH_RETRY:
1574		/* Wait before retrying. */
1575		delay_retry = true;
1576		fallthrough;
1577	case SCSI_DH_IMM_RETRY:
1578	case SCSI_DH_RES_TEMP_UNAVAIL:
1579		if (pg_init_limit_reached(m, pgpath))
1580			fail_path(pgpath);
1581		errors = 0;
1582		break;
1583	case SCSI_DH_DEV_OFFLINED:
1584	default:
1585		/*
1586		 * We probably do not want to fail the path for a device
1587		 * error, but this is what the old dm did. In future
1588		 * patches we can do more advanced handling.
1589		 */
1590		fail_path(pgpath);
1591	}
1592
1593	spin_lock_irqsave(&m->lock, flags);
1594	if (errors) {
1595		if (pgpath == m->current_pgpath) {
1596			DMERR("Could not failover device. Error %d.", errors);
1597			m->current_pgpath = NULL;
1598			m->current_pg = NULL;
1599		}
1600	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1601		pg->bypassed = false;
1602
1603	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1604		/* Activations of other paths are still on going */
1605		goto out;
1606
1607	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1608		if (delay_retry)
1609			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1610		else
1611			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1612
1613		if (__pg_init_all_paths(m))
1614			goto out;
1615	}
1616	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1617
1618	process_queued_io_list(m);
1619
1620	/*
1621	 * Wake up any thread waiting to suspend.
1622	 */
1623	wake_up(&m->pg_init_wait);
1624
1625out:
1626	spin_unlock_irqrestore(&m->lock, flags);
1627}
1628
1629static void activate_or_offline_path(struct pgpath *pgpath)
1630{
1631	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1632
1633	if (pgpath->is_active && !blk_queue_dying(q))
1634		scsi_dh_activate(q, pg_init_done, pgpath);
1635	else
1636		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1637}
1638
1639static void activate_path_work(struct work_struct *work)
1640{
1641	struct pgpath *pgpath =
1642		container_of(work, struct pgpath, activate_path.work);
1643
1644	activate_or_offline_path(pgpath);
1645}
1646
1647static int multipath_end_io(struct dm_target *ti, struct request *clone,
1648			    blk_status_t error, union map_info *map_context)
1649{
1650	struct dm_mpath_io *mpio = get_mpio(map_context);
1651	struct pgpath *pgpath = mpio->pgpath;
1652	int r = DM_ENDIO_DONE;
1653
1654	/*
1655	 * We don't queue any clone request inside the multipath target
1656	 * during end I/O handling, since those clone requests don't have
1657	 * bio clones.  If we queue them inside the multipath target,
1658	 * we need to make bio clones, that requires memory allocation.
1659	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1660	 *  don't have bio clones.)
1661	 * Instead of queueing the clone request here, we queue the original
1662	 * request into dm core, which will remake a clone request and
1663	 * clone bios for it and resubmit it later.
1664	 */
1665	if (error && blk_path_error(error)) {
1666		struct multipath *m = ti->private;
1667
1668		if (error == BLK_STS_RESOURCE)
1669			r = DM_ENDIO_DELAY_REQUEUE;
1670		else
1671			r = DM_ENDIO_REQUEUE;
1672
1673		if (pgpath)
1674			fail_path(pgpath);
1675
1676		if (!atomic_read(&m->nr_valid_paths) &&
1677		    !must_push_back_rq(m)) {
1678			if (error == BLK_STS_IOERR)
1679				dm_report_EIO(m);
1680			/* complete with the original error */
1681			r = DM_ENDIO_DONE;
1682		}
1683	}
1684
1685	if (pgpath) {
1686		struct path_selector *ps = &pgpath->pg->ps;
1687
1688		if (ps->type->end_io)
1689			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1690					 clone->io_start_time_ns);
1691	}
1692
1693	return r;
1694}
1695
1696static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1697				blk_status_t *error)
1698{
1699	struct multipath *m = ti->private;
1700	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1701	struct pgpath *pgpath = mpio->pgpath;
1702	unsigned long flags;
1703	int r = DM_ENDIO_DONE;
1704
1705	if (!*error || !blk_path_error(*error))
1706		goto done;
1707
1708	if (pgpath)
1709		fail_path(pgpath);
1710
1711	if (!atomic_read(&m->nr_valid_paths)) {
1712		spin_lock_irqsave(&m->lock, flags);
1713		if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1714			if (__must_push_back(m)) {
1715				r = DM_ENDIO_REQUEUE;
1716			} else {
1717				dm_report_EIO(m);
1718				*error = BLK_STS_IOERR;
1719			}
1720			spin_unlock_irqrestore(&m->lock, flags);
1721			goto done;
1722		}
1723		spin_unlock_irqrestore(&m->lock, flags);
1724	}
1725
1726	multipath_queue_bio(m, clone);
1727	r = DM_ENDIO_INCOMPLETE;
1728done:
1729	if (pgpath) {
1730		struct path_selector *ps = &pgpath->pg->ps;
1731
1732		if (ps->type->end_io)
1733			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1734					 (mpio->start_time_ns ?:
1735					  dm_start_time_ns_from_clone(clone)));
1736	}
1737
1738	return r;
1739}
1740
1741/*
1742 * Suspend with flush can't complete until all the I/O is processed
1743 * so if the last path fails we must error any remaining I/O.
1744 * - Note that if the freeze_bdev fails while suspending, the
1745 *   queue_if_no_path state is lost - userspace should reset it.
1746 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1747 */
1748static void multipath_presuspend(struct dm_target *ti)
1749{
1750	struct multipath *m = ti->private;
1751
1752	spin_lock_irq(&m->lock);
1753	m->is_suspending = true;
1754	spin_unlock_irq(&m->lock);
1755	/* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1756	if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1757		queue_if_no_path(m, false, true, __func__);
1758}
1759
1760static void multipath_postsuspend(struct dm_target *ti)
1761{
1762	struct multipath *m = ti->private;
1763
1764	mutex_lock(&m->work_mutex);
1765	flush_multipath_work(m);
1766	mutex_unlock(&m->work_mutex);
1767}
1768
1769/*
1770 * Restore the queue_if_no_path setting.
1771 */
1772static void multipath_resume(struct dm_target *ti)
1773{
1774	struct multipath *m = ti->private;
1775
1776	spin_lock_irq(&m->lock);
1777	m->is_suspending = false;
1778	if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1779		set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1780		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1781	}
1782
1783	DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1784		dm_table_device_name(m->ti->table), __func__,
1785		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1786		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1787
1788	spin_unlock_irq(&m->lock);
1789}
1790
1791/*
1792 * Info output has the following format:
1793 * num_multipath_feature_args [multipath_feature_args]*
1794 * num_handler_status_args [handler_status_args]*
1795 * num_groups init_group_number
1796 *            [A|D|E num_ps_status_args [ps_status_args]*
1797 *             num_paths num_selector_args
1798 *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1799 *
1800 * Table output has the following format (identical to the constructor string):
1801 * num_feature_args [features_args]*
1802 * num_handler_args hw_handler [hw_handler_args]*
1803 * num_groups init_group_number
1804 *     [priority selector-name num_ps_args [ps_args]*
1805 *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1806 */
1807static void multipath_status(struct dm_target *ti, status_type_t type,
1808			     unsigned int status_flags, char *result, unsigned int maxlen)
1809{
1810	int sz = 0, pg_counter, pgpath_counter;
1811	struct multipath *m = ti->private;
1812	struct priority_group *pg;
1813	struct pgpath *p;
1814	unsigned int pg_num;
1815	char state;
1816
1817	spin_lock_irq(&m->lock);
1818
1819	/* Features */
1820	if (type == STATUSTYPE_INFO)
1821		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1822		       atomic_read(&m->pg_init_count));
1823	else {
1824		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1825			      (m->pg_init_retries > 0) * 2 +
1826			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1827			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1828
1829		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1830			DMEMIT("queue_if_no_path ");
1831		if (m->pg_init_retries)
1832			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1833		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1834			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1835		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1836			switch (m->queue_mode) {
1837			case DM_TYPE_BIO_BASED:
1838				DMEMIT("queue_mode bio ");
1839				break;
1840			default:
1841				WARN_ON_ONCE(true);
1842				break;
1843			}
1844		}
1845	}
1846
1847	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1848		DMEMIT("0 ");
1849	else
1850		DMEMIT("1 %s ", m->hw_handler_name);
1851
1852	DMEMIT("%u ", m->nr_priority_groups);
1853
1854	if (m->current_pg)
1855		pg_num = m->current_pg->pg_num;
1856	else if (m->next_pg)
1857		pg_num = m->next_pg->pg_num;
1858	else
1859		pg_num = (m->nr_priority_groups ? 1 : 0);
1860
1861	DMEMIT("%u ", pg_num);
1862
1863	switch (type) {
1864	case STATUSTYPE_INFO:
1865		list_for_each_entry(pg, &m->priority_groups, list) {
1866			if (pg->bypassed)
1867				state = 'D';	/* Disabled */
1868			else if (pg == m->current_pg)
1869				state = 'A';	/* Currently Active */
1870			else
1871				state = 'E';	/* Enabled */
1872
1873			DMEMIT("%c ", state);
1874
1875			if (pg->ps.type->status)
1876				sz += pg->ps.type->status(&pg->ps, NULL, type,
1877							  result + sz,
1878							  maxlen - sz);
1879			else
1880				DMEMIT("0 ");
1881
1882			DMEMIT("%u %u ", pg->nr_pgpaths,
1883			       pg->ps.type->info_args);
1884
1885			list_for_each_entry(p, &pg->pgpaths, list) {
1886				DMEMIT("%s %s %u ", p->path.dev->name,
1887				       p->is_active ? "A" : "F",
1888				       p->fail_count);
1889				if (pg->ps.type->status)
1890					sz += pg->ps.type->status(&pg->ps,
1891					      &p->path, type, result + sz,
1892					      maxlen - sz);
1893			}
1894		}
1895		break;
1896
1897	case STATUSTYPE_TABLE:
1898		list_for_each_entry(pg, &m->priority_groups, list) {
1899			DMEMIT("%s ", pg->ps.type->name);
1900
1901			if (pg->ps.type->status)
1902				sz += pg->ps.type->status(&pg->ps, NULL, type,
1903							  result + sz,
1904							  maxlen - sz);
1905			else
1906				DMEMIT("0 ");
1907
1908			DMEMIT("%u %u ", pg->nr_pgpaths,
1909			       pg->ps.type->table_args);
1910
1911			list_for_each_entry(p, &pg->pgpaths, list) {
1912				DMEMIT("%s ", p->path.dev->name);
1913				if (pg->ps.type->status)
1914					sz += pg->ps.type->status(&pg->ps,
1915					      &p->path, type, result + sz,
1916					      maxlen - sz);
1917			}
1918		}
1919		break;
1920
1921	case STATUSTYPE_IMA:
1922		sz = 0; /*reset the result pointer*/
1923
1924		DMEMIT_TARGET_NAME_VERSION(ti->type);
1925		DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1926
1927		pg_counter = 0;
1928		list_for_each_entry(pg, &m->priority_groups, list) {
1929			if (pg->bypassed)
1930				state = 'D';	/* Disabled */
1931			else if (pg == m->current_pg)
1932				state = 'A';	/* Currently Active */
1933			else
1934				state = 'E';	/* Enabled */
1935			DMEMIT(",pg_state_%d=%c", pg_counter, state);
1936			DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1937			DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1938
1939			pgpath_counter = 0;
1940			list_for_each_entry(p, &pg->pgpaths, list) {
1941				DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1942				       pg_counter, pgpath_counter, p->path.dev->name,
1943				       pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1944				       pg_counter, pgpath_counter, p->fail_count);
1945				if (pg->ps.type->status) {
1946					DMEMIT(",path_selector_status_%d_%d=",
1947					       pg_counter, pgpath_counter);
1948					sz += pg->ps.type->status(&pg->ps, &p->path,
1949								  type, result + sz,
1950								  maxlen - sz);
1951				}
1952				pgpath_counter++;
1953			}
1954			pg_counter++;
1955		}
1956		DMEMIT(";");
1957		break;
1958	}
1959
1960	spin_unlock_irq(&m->lock);
1961}
1962
1963static int multipath_message(struct dm_target *ti, unsigned int argc, char **argv,
1964			     char *result, unsigned int maxlen)
1965{
1966	int r = -EINVAL;
1967	dev_t dev;
1968	struct multipath *m = ti->private;
1969	action_fn action;
1970
1971	mutex_lock(&m->work_mutex);
1972
1973	if (dm_suspended(ti)) {
1974		r = -EBUSY;
1975		goto out;
1976	}
1977
1978	if (argc == 1) {
1979		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1980			r = queue_if_no_path(m, true, false, __func__);
1981			spin_lock_irq(&m->lock);
1982			enable_nopath_timeout(m);
1983			spin_unlock_irq(&m->lock);
1984			goto out;
1985		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1986			r = queue_if_no_path(m, false, false, __func__);
1987			disable_nopath_timeout(m);
1988			goto out;
1989		}
1990	}
1991
1992	if (argc != 2) {
1993		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1994		goto out;
1995	}
1996
1997	if (!strcasecmp(argv[0], "disable_group")) {
1998		r = bypass_pg_num(m, argv[1], true);
1999		goto out;
2000	} else if (!strcasecmp(argv[0], "enable_group")) {
2001		r = bypass_pg_num(m, argv[1], false);
2002		goto out;
2003	} else if (!strcasecmp(argv[0], "switch_group")) {
2004		r = switch_pg_num(m, argv[1]);
2005		goto out;
2006	} else if (!strcasecmp(argv[0], "reinstate_path"))
2007		action = reinstate_path;
2008	else if (!strcasecmp(argv[0], "fail_path"))
2009		action = fail_path;
2010	else {
2011		DMWARN("Unrecognised multipath message received: %s", argv[0]);
2012		goto out;
2013	}
2014
2015	r = dm_devt_from_path(argv[1], &dev);
2016	if (r) {
2017		DMWARN("message: error getting device %s",
2018		       argv[1]);
2019		goto out;
2020	}
2021
2022	r = action_dev(m, dev, action);
2023
2024out:
2025	mutex_unlock(&m->work_mutex);
2026	return r;
2027}
2028
2029/*
2030 * Perform a minimal read from the given path to find out whether the
2031 * path still works.  If a path error occurs, fail it.
2032 */
2033static int probe_path(struct pgpath *pgpath)
2034{
2035	struct block_device *bdev = pgpath->path.dev->bdev;
2036	unsigned int read_size = bdev_logical_block_size(bdev);
2037	struct page *page;
2038	struct bio *bio;
2039	blk_status_t status;
2040	int r = 0;
2041
2042	if (WARN_ON_ONCE(read_size > PAGE_SIZE))
2043		return -EINVAL;
2044
2045	page = alloc_page(GFP_KERNEL);
2046	if (!page)
2047		return -ENOMEM;
2048
2049	/* Perform a minimal read: Sector 0, length read_size */
2050	bio = bio_alloc(bdev, 1, REQ_OP_READ, GFP_KERNEL);
2051	if (!bio) {
2052		r = -ENOMEM;
2053		goto out;
2054	}
2055
2056	bio->bi_iter.bi_sector = 0;
2057	__bio_add_page(bio, page, read_size, 0);
2058	submit_bio_wait(bio);
2059	status = bio->bi_status;
2060	bio_put(bio);
2061
2062	if (status && blk_path_error(status))
2063		fail_path(pgpath);
2064
2065out:
2066	__free_page(page);
2067	return r;
2068}
2069
2070/*
2071 * Probe all active paths in current_pg to find out whether they still work.
2072 * Fail all paths that do not work.
2073 *
2074 * Return -ENOTCONN if no valid path is left (even outside of current_pg). We
2075 * cannot probe paths in other pgs without switching current_pg, so if valid
2076 * paths are only in different pgs, they may or may not work. Additionally
2077 * we should not probe paths in a pathgroup that is in the process of
2078 * Initializing. Userspace can submit a request and we'll switch and wait
2079 * for the pathgroup to be initialized. If the request fails, it may need to
2080 * probe again.
2081 */
2082static int probe_active_paths(struct multipath *m)
2083{
2084	struct pgpath *pgpath;
2085	struct priority_group *pg = NULL;
2086	int r = 0;
2087
2088	spin_lock_irq(&m->lock);
2089	if (test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags)) {
2090		wait_event_lock_irq(m->probe_wait,
2091				    !test_bit(MPATHF_DELAY_PG_SWITCH, &m->flags),
2092				    m->lock);
2093		/*
2094		 * if we waited because a probe was already in progress,
2095		 * and it probed the current active pathgroup, don't
2096		 * reprobe. Just return the number of valid paths
2097		 */
2098		if (m->current_pg == m->last_probed_pg)
2099			goto skip_probe;
2100	}
2101	if (!m->current_pg || m->is_suspending ||
2102	    test_bit(MPATHF_QUEUE_IO, &m->flags))
2103		goto skip_probe;
2104	set_bit(MPATHF_DELAY_PG_SWITCH, &m->flags);
2105	pg = m->last_probed_pg = m->current_pg;
2106	spin_unlock_irq(&m->lock);
2107
2108	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2109		if (pg != READ_ONCE(m->current_pg) ||
2110		    READ_ONCE(m->is_suspending))
2111			goto out;
2112		if (!pgpath->is_active)
2113			continue;
2114
2115		r = probe_path(pgpath);
2116		if (r < 0)
2117			goto out;
2118	}
2119
2120out:
2121	spin_lock_irq(&m->lock);
2122	clear_bit(MPATHF_DELAY_PG_SWITCH, &m->flags);
2123	if (test_and_clear_bit(MPATHF_NEED_PG_SWITCH, &m->flags)) {
2124		m->current_pgpath = NULL;
2125		m->current_pg = NULL;
2126	}
2127skip_probe:
2128	if (r == 0 && !atomic_read(&m->nr_valid_paths))
2129		r = -ENOTCONN;
2130	spin_unlock_irq(&m->lock);
2131	if (pg)
2132		wake_up(&m->probe_wait);
2133	return r;
2134}
2135
2136static int multipath_prepare_ioctl(struct dm_target *ti,
2137				   struct block_device **bdev,
2138				   unsigned int cmd, unsigned long arg,
2139				   bool *forward)
2140{
2141	struct multipath *m = ti->private;
2142	struct pgpath *pgpath;
2143	int r;
2144
2145	if (_IOC_TYPE(cmd) == DM_IOCTL) {
2146		*forward = false;
2147		switch (cmd) {
2148		case DM_MPATH_PROBE_PATHS:
2149			return probe_active_paths(m);
2150		default:
2151			return -ENOTTY;
2152		}
2153	}
2154
2155	pgpath = READ_ONCE(m->current_pgpath);
2156	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2157		pgpath = choose_pgpath(m, 0);
2158
2159	if (pgpath) {
2160		if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2161			*bdev = pgpath->path.dev->bdev;
2162			r = 0;
2163		} else {
2164			/* pg_init has not started or completed */
2165			r = -ENOTCONN;
2166		}
2167	} else {
2168		/* No path is available */
2169		r = -EIO;
2170		spin_lock_irq(&m->lock);
2171		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2172			r = -ENOTCONN;
2173		spin_unlock_irq(&m->lock);
2174	}
2175
2176	if (r == -ENOTCONN) {
2177		if (!READ_ONCE(m->current_pg)) {
2178			/* Path status changed, redo selection */
2179			(void) choose_pgpath(m, 0);
2180		}
2181		spin_lock_irq(&m->lock);
2182		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2183			(void) __pg_init_all_paths(m);
2184		spin_unlock_irq(&m->lock);
2185		dm_table_run_md_queue_async(m->ti->table);
2186		process_queued_io_list(m);
2187	}
2188
2189	/*
2190	 * Only pass ioctls through if the device sizes match exactly.
2191	 */
2192	if (!r && ti->len != bdev_nr_sectors((*bdev)))
2193		return 1;
2194	return r;
2195}
2196
2197static int multipath_iterate_devices(struct dm_target *ti,
2198				     iterate_devices_callout_fn fn, void *data)
2199{
2200	struct multipath *m = ti->private;
2201	struct priority_group *pg;
2202	struct pgpath *p;
2203	int ret = 0;
2204
2205	list_for_each_entry(pg, &m->priority_groups, list) {
2206		list_for_each_entry(p, &pg->pgpaths, list) {
2207			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2208			if (ret)
2209				goto out;
2210		}
2211	}
2212
2213out:
2214	return ret;
2215}
2216
2217static int pgpath_busy(struct pgpath *pgpath)
2218{
2219	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2220
2221	return blk_lld_busy(q);
2222}
2223
2224/*
2225 * We return "busy", only when we can map I/Os but underlying devices
2226 * are busy (so even if we map I/Os now, the I/Os will wait on
2227 * the underlying queue).
2228 * In other words, if we want to kill I/Os or queue them inside us
2229 * due to map unavailability, we don't return "busy".  Otherwise,
2230 * dm core won't give us the I/Os and we can't do what we want.
2231 */
2232static int multipath_busy(struct dm_target *ti)
2233{
2234	bool busy = false, has_active = false;
2235	struct multipath *m = ti->private;
2236	struct priority_group *pg, *next_pg;
2237	struct pgpath *pgpath;
2238
2239	/* pg_init in progress */
2240	if (atomic_read(&m->pg_init_in_progress))
2241		return true;
2242
2243	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2244	if (!atomic_read(&m->nr_valid_paths)) {
2245		unsigned long flags;
2246
2247		spin_lock_irqsave(&m->lock, flags);
2248		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2249			spin_unlock_irqrestore(&m->lock, flags);
2250			return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2251		}
2252		spin_unlock_irqrestore(&m->lock, flags);
2253	}
2254
2255	/* Guess which priority_group will be used at next mapping time */
2256	pg = READ_ONCE(m->current_pg);
2257	next_pg = READ_ONCE(m->next_pg);
2258	if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2259		pg = next_pg;
2260
2261	if (!pg) {
2262		/*
2263		 * We don't know which pg will be used at next mapping time.
2264		 * We don't call choose_pgpath() here to avoid to trigger
2265		 * pg_init just by busy checking.
2266		 * So we don't know whether underlying devices we will be using
2267		 * at next mapping time are busy or not. Just try mapping.
2268		 */
2269		return busy;
2270	}
2271
2272	/*
2273	 * If there is one non-busy active path at least, the path selector
2274	 * will be able to select it. So we consider such a pg as not busy.
2275	 */
2276	busy = true;
2277	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2278		if (pgpath->is_active) {
2279			has_active = true;
2280			if (!pgpath_busy(pgpath)) {
2281				busy = false;
2282				break;
2283			}
2284		}
2285	}
2286
2287	if (!has_active) {
2288		/*
2289		 * No active path in this pg, so this pg won't be used and
2290		 * the current_pg will be changed at next mapping time.
2291		 * We need to try mapping to determine it.
2292		 */
2293		busy = false;
2294	}
2295
2296	return busy;
2297}
2298
2299/*
2300 *---------------------------------------------------------------
2301 * Module setup
2302 *---------------------------------------------------------------
2303 */
2304static struct target_type multipath_target = {
2305	.name = "multipath",
2306	.version = {1, 15, 0},
2307	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2308		    DM_TARGET_PASSES_INTEGRITY | DM_TARGET_ATOMIC_WRITES,
2309	.module = THIS_MODULE,
2310	.ctr = multipath_ctr,
2311	.dtr = multipath_dtr,
2312	.clone_and_map_rq = multipath_clone_and_map,
2313	.release_clone_rq = multipath_release_clone,
2314	.rq_end_io = multipath_end_io,
2315	.map = multipath_map_bio,
2316	.end_io = multipath_end_io_bio,
2317	.presuspend = multipath_presuspend,
2318	.postsuspend = multipath_postsuspend,
2319	.resume = multipath_resume,
2320	.status = multipath_status,
2321	.message = multipath_message,
2322	.prepare_ioctl = multipath_prepare_ioctl,
2323	.iterate_devices = multipath_iterate_devices,
2324	.busy = multipath_busy,
2325};
2326
2327static int __init dm_multipath_init(void)
2328{
2329	int r = -ENOMEM;
2330
2331	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2332	if (!kmultipathd) {
2333		DMERR("failed to create workqueue kmpathd");
2334		goto bad_alloc_kmultipathd;
2335	}
2336
2337	/*
2338	 * A separate workqueue is used to handle the device handlers
2339	 * to avoid overloading existing workqueue. Overloading the
2340	 * old workqueue would also create a bottleneck in the
2341	 * path of the storage hardware device activation.
2342	 */
2343	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2344						  WQ_MEM_RECLAIM);
2345	if (!kmpath_handlerd) {
2346		DMERR("failed to create workqueue kmpath_handlerd");
2347		goto bad_alloc_kmpath_handlerd;
2348	}
2349
2350	dm_mpath_wq = alloc_workqueue("dm_mpath_wq", 0, 0);
2351	if (!dm_mpath_wq) {
2352		DMERR("failed to create workqueue dm_mpath_wq");
2353		goto bad_alloc_dm_mpath_wq;
2354	}
2355
2356	r = dm_register_target(&multipath_target);
2357	if (r < 0)
2358		goto bad_register_target;
2359
2360	return 0;
2361
2362bad_register_target:
2363	destroy_workqueue(dm_mpath_wq);
2364bad_alloc_dm_mpath_wq:
2365	destroy_workqueue(kmpath_handlerd);
2366bad_alloc_kmpath_handlerd:
2367	destroy_workqueue(kmultipathd);
2368bad_alloc_kmultipathd:
2369	return r;
2370}
2371
2372static void __exit dm_multipath_exit(void)
2373{
2374	destroy_workqueue(dm_mpath_wq);
2375	destroy_workqueue(kmpath_handlerd);
2376	destroy_workqueue(kmultipathd);
2377
2378	dm_unregister_target(&multipath_target);
2379}
2380
2381module_init(dm_multipath_init);
2382module_exit(dm_multipath_exit);
2383
2384module_param_named(queue_if_no_path_timeout_secs, queue_if_no_path_timeout_secs, ulong, 0644);
2385MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2386
2387MODULE_DESCRIPTION(DM_NAME " multipath target");
2388MODULE_AUTHOR("Sistina Software <dm-devel@lists.linux.dev>");
2389MODULE_LICENSE("GPL");