dm-mpath.c 49.0 KB
Newer Older
L
Linus Torvalds 已提交
1 2 3 4 5 6 7
/*
 * Copyright (C) 2003 Sistina Software Limited.
 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
 *
 * This file is released under the GPL.
 */

8 9
#include <linux/device-mapper.h>

10
#include "dm-rq.h"
11
#include "dm-bio-record.h"
L
Linus Torvalds 已提交
12
#include "dm-path-selector.h"
M
Mike Anderson 已提交
13
#include "dm-uevent.h"
L
Linus Torvalds 已提交
14

15
#include <linux/blkdev.h>
L
Linus Torvalds 已提交
16 17 18 19 20 21 22 23
#include <linux/ctype.h>
#include <linux/init.h>
#include <linux/mempool.h>
#include <linux/module.h>
#include <linux/pagemap.h>
#include <linux/slab.h>
#include <linux/time.h>
#include <linux/workqueue.h>
24
#include <linux/delay.h>
25
#include <scsi/scsi_dh.h>
A
Arun Sharma 已提交
26
#include <linux/atomic.h>
27
#include <linux/blk-mq.h>
L
Linus Torvalds 已提交
28

29
#define DM_MSG_PREFIX "multipath"
30 31
#define DM_PG_INIT_DELAY_MSECS 2000
#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
L
Linus Torvalds 已提交
32 33 34 35 36 37 38 39

/* Path properties */
struct pgpath {
	struct list_head list;

	struct priority_group *pg;	/* Owning PG */
	unsigned fail_count;		/* Cumulative failure count */

40
	struct dm_path path;
41
	struct delayed_work activate_path;
42 43

	bool is_active:1;		/* Path status */
L
Linus Torvalds 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
};

#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)

/*
 * Paths are grouped into Priority Groups and numbered from 1 upwards.
 * Each has a path selector which controls which path gets used.
 */
struct priority_group {
	struct list_head list;

	struct multipath *m;		/* Owning multipath instance */
	struct path_selector ps;

	unsigned pg_num;		/* Reference number */
	unsigned nr_pgpaths;		/* Number of paths in PG */
	struct list_head pgpaths;
61 62

	bool bypassed:1;		/* Temporarily bypass this PG? */
L
Linus Torvalds 已提交
63 64 65 66 67 68 69
};

/* Multipath context */
struct multipath {
	struct list_head list;
	struct dm_target *ti;

70
	const char *hw_handler_name;
71
	char *hw_handler_params;
72

73 74
	spinlock_t lock;

L
Linus Torvalds 已提交
75 76
	unsigned nr_priority_groups;
	struct list_head priority_groups;
77 78 79

	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */

L
Linus Torvalds 已提交
80 81 82 83
	struct pgpath *current_pgpath;
	struct priority_group *current_pg;
	struct priority_group *next_pg;	/* Switch to this PG if set */

84
	unsigned long flags;		/* Multipath state flags */
85

D
Dave Wysochanski 已提交
86
	unsigned pg_init_retries;	/* Number of times to retry pg_init */
87
	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
L
Linus Torvalds 已提交
88

89 90 91 92
	atomic_t nr_valid_paths;	/* Total number of usable paths */
	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
	atomic_t pg_init_count;		/* Number of times pg_init called */

93
	enum dm_queue_mode queue_mode;
94

95
	struct mutex work_mutex;
96
	struct work_struct trigger_event;
97 98 99

	struct work_struct process_queued_bios;
	struct bio_list queued_bios;
L
Linus Torvalds 已提交
100 101 102
};

/*
103
 * Context information attached to each io we process.
L
Linus Torvalds 已提交
104
 */
A
Alasdair G Kergon 已提交
105
struct dm_mpath_io {
L
Linus Torvalds 已提交
106
	struct pgpath *pgpath;
107
	size_t nr_bytes;
L
Linus Torvalds 已提交
108 109 110 111
};

typedef int (*action_fn) (struct pgpath *pgpath);

112
static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
D
David Howells 已提交
113
static void trigger_event(struct work_struct *work);
114 115
static void activate_or_offline_path(struct pgpath *pgpath);
static void activate_path_work(struct work_struct *work);
116
static void process_queued_bios(struct work_struct *work);
L
Linus Torvalds 已提交
117

118 119 120 121 122 123 124 125 126 127 128
/*-----------------------------------------------
 * Multipath state flags.
 *-----------------------------------------------*/

#define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
#define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
#define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3	/* If there's already a hw_handler present, don't change it. */
#define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
#define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
#define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
L
Linus Torvalds 已提交
129 130 131 132 133 134 135

/*-----------------------------------------------
 * Allocation routines
 *-----------------------------------------------*/

static struct pgpath *alloc_pgpath(void)
{
M
Micha³ Miros³aw 已提交
136
	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
L
Linus Torvalds 已提交
137

138
	if (pgpath) {
139
		pgpath->is_active = true;
140
		INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
141
	}
L
Linus Torvalds 已提交
142 143 144 145

	return pgpath;
}

A
Alasdair G Kergon 已提交
146
static void free_pgpath(struct pgpath *pgpath)
L
Linus Torvalds 已提交
147 148 149 150 151 152 153 154
{
	kfree(pgpath);
}

static struct priority_group *alloc_priority_group(void)
{
	struct priority_group *pg;

M
Micha³ Miros³aw 已提交
155
	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
L
Linus Torvalds 已提交
156

M
Micha³ Miros³aw 已提交
157 158
	if (pg)
		INIT_LIST_HEAD(&pg->pgpaths);
L
Linus Torvalds 已提交
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187

	return pg;
}

static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
{
	struct pgpath *pgpath, *tmp;

	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
		list_del(&pgpath->list);
		dm_put_device(ti, pgpath->path.dev);
		free_pgpath(pgpath);
	}
}

static void free_priority_group(struct priority_group *pg,
				struct dm_target *ti)
{
	struct path_selector *ps = &pg->ps;

	if (ps->type) {
		ps->type->destroy(ps);
		dm_put_path_selector(ps->type);
	}

	free_pgpaths(&pg->pgpaths, ti);
	kfree(pg);
}

188
static struct multipath *alloc_multipath(struct dm_target *ti)
L
Linus Torvalds 已提交
189 190 191
{
	struct multipath *m;

M
Micha³ Miros³aw 已提交
192
	m = kzalloc(sizeof(*m), GFP_KERNEL);
L
Linus Torvalds 已提交
193 194 195
	if (m) {
		INIT_LIST_HEAD(&m->priority_groups);
		spin_lock_init(&m->lock);
196
		set_bit(MPATHF_QUEUE_IO, &m->flags);
197 198 199
		atomic_set(&m->nr_valid_paths, 0);
		atomic_set(&m->pg_init_in_progress, 0);
		atomic_set(&m->pg_init_count, 0);
200
		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
D
David Howells 已提交
201
		INIT_WORK(&m->trigger_event, trigger_event);
202
		init_waitqueue_head(&m->pg_init_wait);
203
		mutex_init(&m->work_mutex);
204

205
		m->queue_mode = DM_TYPE_NONE;
206

M
Micha³ Miros³aw 已提交
207 208
		m->ti = ti;
		ti->private = m;
L
Linus Torvalds 已提交
209 210 211 212 213
	}

	return m;
}

214 215 216 217 218 219 220 221 222 223
static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
{
	if (m->queue_mode == DM_TYPE_NONE) {
		/*
		 * Default to request-based.
		 */
		if (dm_use_blk_mq(dm_table_get_md(ti->table)))
			m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
		else
			m->queue_mode = DM_TYPE_REQUEST_BASED;
224
	} else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225 226 227 228 229 230 231 232 233 234 235 236 237
		INIT_WORK(&m->process_queued_bios, process_queued_bios);
		/*
		 * bio-based doesn't support any direct scsi_dh management;
		 * it just discovers if a scsi_dh is attached.
		 */
		set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
	}

	dm_table_set_type(ti->table, m->queue_mode);

	return 0;
}

L
Linus Torvalds 已提交
238 239 240 241 242 243 244 245 246
static void free_multipath(struct multipath *m)
{
	struct priority_group *pg, *tmp;

	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
		list_del(&pg->list);
		free_priority_group(pg, m->ti);
	}

247
	kfree(m->hw_handler_name);
248
	kfree(m->hw_handler_params);
L
Linus Torvalds 已提交
249 250 251
	kfree(m);
}

252 253 254 255 256
static struct dm_mpath_io *get_mpio(union map_info *info)
{
	return info->ptr;
}

257 258 259 260 261
static size_t multipath_per_bio_data_size(void)
{
	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
}

262 263
static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
{
264
	return dm_per_bio_data(bio, multipath_per_bio_data_size());
265 266
}

267
static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
268
{
269
	/* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270
	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271 272 273 274 275
	void *bio_details = mpio + 1;

	return bio_details;
}

276
static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
277 278 279
{
	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
	struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
280 281

	memset(mpio, 0, sizeof(*mpio));
282 283
	memset(bio_details, 0, sizeof(*bio_details));
	dm_bio_record(bio_details, bio);
284

285 286
	if (mpio_p)
		*mpio_p = mpio;
287 288
}

L
Linus Torvalds 已提交
289 290 291 292
/*-----------------------------------------------
 * Path selection
 *-----------------------------------------------*/

293
static int __pg_init_all_paths(struct multipath *m)
K
Kiyoshi Ueda 已提交
294 295
{
	struct pgpath *pgpath;
296
	unsigned long pg_init_delay = 0;
K
Kiyoshi Ueda 已提交
297

298 299
	lockdep_assert_held(&m->lock);

300
	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
301
		return 0;
302

303
	atomic_inc(&m->pg_init_count);
304
	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
305 306 307 308 309

	/* Check here to reset pg_init_required */
	if (!m->current_pg)
		return 0;

310
	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
311 312
		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
K
Kiyoshi Ueda 已提交
313 314 315 316
	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
		/* Skip failed paths */
		if (!pgpath->is_active)
			continue;
317 318
		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
				       pg_init_delay))
319
			atomic_inc(&m->pg_init_in_progress);
K
Kiyoshi Ueda 已提交
320
	}
321
	return atomic_read(&m->pg_init_in_progress);
K
Kiyoshi Ueda 已提交
322 323
}

324
static int pg_init_all_paths(struct multipath *m)
L
Linus Torvalds 已提交
325
{
326
	int ret;
327 328 329
	unsigned long flags;

	spin_lock_irqsave(&m->lock, flags);
330
	ret = __pg_init_all_paths(m);
331
	spin_unlock_irqrestore(&m->lock, flags);
332 333

	return ret;
334 335 336 337 338
}

static void __switch_pg(struct multipath *m, struct priority_group *pg)
{
	m->current_pg = pg;
L
Linus Torvalds 已提交
339 340

	/* Must we initialise the PG first, and queue I/O till it's ready? */
341
	if (m->hw_handler_name) {
342 343
		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
		set_bit(MPATHF_QUEUE_IO, &m->flags);
L
Linus Torvalds 已提交
344
	} else {
345 346
		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
		clear_bit(MPATHF_QUEUE_IO, &m->flags);
L
Linus Torvalds 已提交
347
	}
D
Dave Wysochanski 已提交
348

349
	atomic_set(&m->pg_init_count, 0);
L
Linus Torvalds 已提交
350 351
}

352 353 354
static struct pgpath *choose_path_in_pg(struct multipath *m,
					struct priority_group *pg,
					size_t nr_bytes)
L
Linus Torvalds 已提交
355
{
356
	unsigned long flags;
357
	struct dm_path *path;
358
	struct pgpath *pgpath;
L
Linus Torvalds 已提交
359

360
	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
L
Linus Torvalds 已提交
361
	if (!path)
362
		return ERR_PTR(-ENXIO);
L
Linus Torvalds 已提交
363

364
	pgpath = path_to_pgpath(path);
L
Linus Torvalds 已提交
365

366
	if (unlikely(READ_ONCE(m->current_pg) != pg)) {
367 368 369 370 371 372
		/* Only update current_pgpath if pg changed */
		spin_lock_irqsave(&m->lock, flags);
		m->current_pgpath = pgpath;
		__switch_pg(m, pg);
		spin_unlock_irqrestore(&m->lock, flags);
	}
L
Linus Torvalds 已提交
373

374
	return pgpath;
L
Linus Torvalds 已提交
375 376
}

377
static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
L
Linus Torvalds 已提交
378
{
379
	unsigned long flags;
L
Linus Torvalds 已提交
380
	struct priority_group *pg;
381
	struct pgpath *pgpath;
382
	unsigned bypassed = 1;
L
Linus Torvalds 已提交
383

384
	if (!atomic_read(&m->nr_valid_paths)) {
385
		clear_bit(MPATHF_QUEUE_IO, &m->flags);
L
Linus Torvalds 已提交
386
		goto failed;
387
	}
L
Linus Torvalds 已提交
388 389

	/* Were we instructed to switch PG? */
390
	if (READ_ONCE(m->next_pg)) {
391
		spin_lock_irqsave(&m->lock, flags);
L
Linus Torvalds 已提交
392
		pg = m->next_pg;
393 394 395 396
		if (!pg) {
			spin_unlock_irqrestore(&m->lock, flags);
			goto check_current_pg;
		}
L
Linus Torvalds 已提交
397
		m->next_pg = NULL;
398 399 400 401
		spin_unlock_irqrestore(&m->lock, flags);
		pgpath = choose_path_in_pg(m, pg, nr_bytes);
		if (!IS_ERR_OR_NULL(pgpath))
			return pgpath;
L
Linus Torvalds 已提交
402 403 404
	}

	/* Don't change PG until it has no remaining paths */
405
check_current_pg:
406
	pg = READ_ONCE(m->current_pg);
407 408 409 410 411
	if (pg) {
		pgpath = choose_path_in_pg(m, pg, nr_bytes);
		if (!IS_ERR_OR_NULL(pgpath))
			return pgpath;
	}
L
Linus Torvalds 已提交
412 413 414 415

	/*
	 * Loop through priority groups until we find a valid path.
	 * First time we skip PGs marked 'bypassed'.
416 417
	 * Second time we only try the ones we skipped, but set
	 * pg_init_delay_retry so we do not hammer controllers.
L
Linus Torvalds 已提交
418 419 420
	 */
	do {
		list_for_each_entry(pg, &m->priority_groups, list) {
421
			if (pg->bypassed == !!bypassed)
L
Linus Torvalds 已提交
422
				continue;
423 424
			pgpath = choose_path_in_pg(m, pg, nr_bytes);
			if (!IS_ERR_OR_NULL(pgpath)) {
425
				if (!bypassed)
426
					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
427
				return pgpath;
428
			}
L
Linus Torvalds 已提交
429 430 431 432
		}
	} while (bypassed--);

failed:
433
	spin_lock_irqsave(&m->lock, flags);
L
Linus Torvalds 已提交
434 435
	m->current_pgpath = NULL;
	m->current_pg = NULL;
436 437 438
	spin_unlock_irqrestore(&m->lock, flags);

	return NULL;
L
Linus Torvalds 已提交
439 440
}

441
/*
442 443 444
 * dm_report_EIO() is a macro instead of a function to make pr_debug()
 * report the function name and line number of the function from which
 * it has been invoked.
445
 */
446
#define dm_report_EIO(m)						\
447
do {									\
448 449 450 451 452 453 454
	struct mapped_device *md = dm_table_get_md((m)->ti->table);	\
									\
	pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
		 dm_device_name(md),					\
		 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
		 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags),	\
		 dm_noflush_suspending((m)->ti));			\
455
} while (0)
456

457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
/*
 * Check whether bios must be queued in the device-mapper core rather
 * than here in the target.
 *
 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
 * the same value then we are not between multipath_presuspend()
 * and multipath_resume() calls and we have no need to check
 * for the DMF_NOFLUSH_SUSPENDING flag.
 */
static bool __must_push_back(struct multipath *m, unsigned long flags)
{
	return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
		 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
		dm_noflush_suspending(m->ti));
}

/*
 * Following functions use READ_ONCE to get atomic access to
 * all m->flags to avoid taking spinlock
 */
static bool must_push_back_rq(struct multipath *m)
{
	unsigned long flags = READ_ONCE(m->flags);
	return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
}

static bool must_push_back_bio(struct multipath *m)
{
	unsigned long flags = READ_ONCE(m->flags);
	return __must_push_back(m, flags);
}

H
Hannes Reinecke 已提交
489
/*
490
 * Map cloned requests (request-based multipath)
H
Hannes Reinecke 已提交
491
 */
492 493 494
static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
				   union map_info *map_context,
				   struct request **__clone)
L
Linus Torvalds 已提交
495
{
496
	struct multipath *m = ti->private;
497
	size_t nr_bytes = blk_rq_bytes(rq);
L
Linus Torvalds 已提交
498
	struct pgpath *pgpath;
499
	struct block_device *bdev;
500
	struct dm_mpath_io *mpio = get_mpio(map_context);
501
	struct request_queue *q;
502
	struct request *clone;
L
Linus Torvalds 已提交
503 504

	/* Do we need to select a new pgpath? */
505
	pgpath = READ_ONCE(m->current_pgpath);
506 507
	if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
		pgpath = choose_pgpath(m, nr_bytes);
L
Linus Torvalds 已提交
508

509
	if (!pgpath) {
510
		if (must_push_back_rq(m))
511
			return DM_MAPIO_DELAY_REQUEUE;
512
		dm_report_EIO(m);	/* Failed */
513
		return DM_MAPIO_KILL;
514 515
	} else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
		   test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
516 517
		if (pg_init_all_paths(m))
			return DM_MAPIO_DELAY_REQUEUE;
518
		return DM_MAPIO_REQUEUE;
519
	}
520

521
	memset(mpio, 0, sizeof(*mpio));
522 523 524
	mpio->pgpath = pgpath;
	mpio->nr_bytes = nr_bytes;

525
	bdev = pgpath->path.dev->bdev;
526 527
	q = bdev_get_queue(bdev);
	clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
528 529
	if (IS_ERR(clone)) {
		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
530 531 532 533 534
		bool queue_dying = blk_queue_dying(q);
		if (queue_dying) {
			atomic_inc(&m->pg_init_in_progress);
			activate_or_offline_path(pgpath);
		}
535
		return DM_MAPIO_DELAY_REQUEUE;
536
	}
537 538 539 540
	clone->bio = clone->biotail = NULL;
	clone->rq_disk = bdev->bd_disk;
	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
	*__clone = clone;
541

542 543 544 545
	if (pgpath->pg->ps.type->start_io)
		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
					      &pgpath->path,
					      nr_bytes);
546
	return DM_MAPIO_REMAPPED;
L
Linus Torvalds 已提交
547 548
}

549 550
static void multipath_release_clone(struct request *clone)
{
551
	blk_put_request(clone);
552 553
}

554 555 556 557 558 559 560 561 562 563 564
/*
 * Map cloned bios (bio-based multipath)
 */
static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
{
	size_t nr_bytes = bio->bi_iter.bi_size;
	struct pgpath *pgpath;
	unsigned long flags;
	bool queue_io;

	/* Do we need to select a new pgpath? */
565
	pgpath = READ_ONCE(m->current_pgpath);
566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
	queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
	if (!pgpath || !queue_io)
		pgpath = choose_pgpath(m, nr_bytes);

	if ((pgpath && queue_io) ||
	    (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
		/* Queue for the daemon to resubmit */
		spin_lock_irqsave(&m->lock, flags);
		bio_list_add(&m->queued_bios, bio);
		spin_unlock_irqrestore(&m->lock, flags);
		/* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
		if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
			pg_init_all_paths(m);
		else if (!queue_io)
			queue_work(kmultipathd, &m->process_queued_bios);
		return DM_MAPIO_SUBMITTED;
	}

	if (!pgpath) {
585
		if (must_push_back_bio(m))
586
			return DM_MAPIO_REQUEUE;
587
		dm_report_EIO(m);
588
		return DM_MAPIO_KILL;
589 590 591 592 593
	}

	mpio->pgpath = pgpath;
	mpio->nr_bytes = nr_bytes;

594
	bio->bi_status = 0;
595
	bio_set_dev(bio, pgpath->path.dev->bdev);
J
Jens Axboe 已提交
596
	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
597 598 599 600 601 602 603 604 605 606 607

	if (pgpath->pg->ps.type->start_io)
		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
					      &pgpath->path,
					      nr_bytes);
	return DM_MAPIO_REMAPPED;
}

static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
{
	struct multipath *m = ti->private;
608 609
	struct dm_mpath_io *mpio = NULL;

610
	multipath_init_per_bio_data(bio, &mpio);
611 612 613
	return __multipath_map_bio(m, bio, mpio);
}

614
static void process_queued_io_list(struct multipath *m)
615
{
616 617 618
	if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
	else if (m->queue_mode == DM_TYPE_BIO_BASED)
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
		queue_work(kmultipathd, &m->process_queued_bios);
}

static void process_queued_bios(struct work_struct *work)
{
	int r;
	unsigned long flags;
	struct bio *bio;
	struct bio_list bios;
	struct blk_plug plug;
	struct multipath *m =
		container_of(work, struct multipath, process_queued_bios);

	bio_list_init(&bios);

	spin_lock_irqsave(&m->lock, flags);

	if (bio_list_empty(&m->queued_bios)) {
		spin_unlock_irqrestore(&m->lock, flags);
		return;
	}

	bio_list_merge(&bios, &m->queued_bios);
	bio_list_init(&m->queued_bios);

	spin_unlock_irqrestore(&m->lock, flags);

	blk_start_plug(&plug);
	while ((bio = bio_list_pop(&bios))) {
		r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
649 650
		switch (r) {
		case DM_MAPIO_KILL:
651 652
			bio->bi_status = BLK_STS_IOERR;
			bio_endio(bio);
653
			break;
654
		case DM_MAPIO_REQUEUE:
655
			bio->bi_status = BLK_STS_DM_REQUEUE;
656
			bio_endio(bio);
657 658
			break;
		case DM_MAPIO_REMAPPED:
659
			generic_make_request(bio);
660
			break;
661 662 663 664
		case 0:
			break;
		default:
			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
665
		}
666 667 668 669
	}
	blk_finish_plug(&plug);
}

L
Linus Torvalds 已提交
670 671 672
/*
 * If we run out of usable paths, should we queue I/O or error it?
 */
673 674
static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
			    bool save_old_value)
L
Linus Torvalds 已提交
675 676 677 678
{
	unsigned long flags;

	spin_lock_irqsave(&m->lock, flags);
L
Lukas Wunner 已提交
679 680 681
	assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
		   (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
		   (!save_old_value && queue_if_no_path));
682
	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
L
Linus Torvalds 已提交
683 684
	spin_unlock_irqrestore(&m->lock, flags);

685
	if (!queue_if_no_path) {
686
		dm_table_run_md_queue_async(m->ti->table);
687
		process_queued_io_list(m);
688
	}
689

L
Linus Torvalds 已提交
690 691 692 693 694 695 696
	return 0;
}

/*
 * An event is triggered whenever a path is taken out of use.
 * Includes path failure and PG bypass.
 */
D
David Howells 已提交
697
static void trigger_event(struct work_struct *work)
L
Linus Torvalds 已提交
698
{
D
David Howells 已提交
699 700
	struct multipath *m =
		container_of(work, struct multipath, trigger_event);
L
Linus Torvalds 已提交
701 702 703 704 705 706 707 708 709 710 711 712 713 714

	dm_table_event(m->ti->table);
}

/*-----------------------------------------------------------------
 * Constructor/argument parsing:
 * <#multipath feature args> [<arg>]*
 * <#hw_handler args> [hw_handler [<arg>]*]
 * <#priority groups>
 * <initial priority group>
 *     [<selector> <#selector args> [<arg>]*
 *      <#paths> <#per-path selector args>
 *         [<path> [<arg>]* ]+ ]+
 *---------------------------------------------------------------*/
715
static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
L
Linus Torvalds 已提交
716 717 718 719 720 721
			       struct dm_target *ti)
{
	int r;
	struct path_selector_type *pst;
	unsigned ps_argc;

E
Eric Biggers 已提交
722
	static const struct dm_arg _args[] = {
723
		{0, 1024, "invalid number of path selector args"},
L
Linus Torvalds 已提交
724 725
	};

726
	pst = dm_get_path_selector(dm_shift_arg(as));
L
Linus Torvalds 已提交
727
	if (!pst) {
728
		ti->error = "unknown path selector type";
L
Linus Torvalds 已提交
729 730 731
		return -EINVAL;
	}

732
	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
733 734
	if (r) {
		dm_put_path_selector(pst);
L
Linus Torvalds 已提交
735
		return -EINVAL;
736
	}
L
Linus Torvalds 已提交
737 738 739 740

	r = pst->create(&pg->ps, ps_argc, as->argv);
	if (r) {
		dm_put_path_selector(pst);
741
		ti->error = "path selector constructor failed";
L
Linus Torvalds 已提交
742 743 744 745
		return r;
	}

	pg->ps.type = pst;
746
	dm_consume_args(as, ps_argc);
L
Linus Torvalds 已提交
747 748 749 750

	return 0;
}

751
static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
L
Linus Torvalds 已提交
752 753 754 755
			       struct dm_target *ti)
{
	int r;
	struct pgpath *p;
756
	struct multipath *m = ti->private;
757 758
	struct request_queue *q = NULL;
	const char *attached_handler_name;
L
Linus Torvalds 已提交
759 760 761

	/* we need at least a path arg */
	if (as->argc < 1) {
762
		ti->error = "no device given";
763
		return ERR_PTR(-EINVAL);
L
Linus Torvalds 已提交
764 765 766 767
	}

	p = alloc_pgpath();
	if (!p)
768
		return ERR_PTR(-ENOMEM);
L
Linus Torvalds 已提交
769

770
	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
771
			  &p->path.dev);
L
Linus Torvalds 已提交
772
	if (r) {
773
		ti->error = "error getting device";
L
Linus Torvalds 已提交
774 775 776
		goto bad;
	}

777
	if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
778 779
		q = bdev_get_queue(p->path.dev->bdev);

780
	if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
781
retain:
782 783
		attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
		if (attached_handler_name) {
784 785 786 787 788 789 790 791 792
			/*
			 * Clear any hw_handler_params associated with a
			 * handler that isn't already attached.
			 */
			if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
				kfree(m->hw_handler_params);
				m->hw_handler_params = NULL;
			}

793 794 795 796 797 798 799 800 801 802
			/*
			 * Reset hw_handler_name to match the attached handler
			 *
			 * NB. This modifies the table line to show the actual
			 * handler instead of the original table passed in.
			 */
			kfree(m->hw_handler_name);
			m->hw_handler_name = attached_handler_name;
		}
	}
803

804
	if (m->hw_handler_name) {
805 806
		r = scsi_dh_attach(q, m->hw_handler_name);
		if (r == -EBUSY) {
807
			char b[BDEVNAME_SIZE];
808

809 810 811 812
			printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
				bdevname(p->path.dev->bdev, b));
			goto retain;
		}
813
		if (r < 0) {
814
			ti->error = "error attaching hardware handler";
815 816 817
			dm_put_device(ti, p->path.dev);
			goto bad;
		}
818 819 820 821 822 823 824 825 826 827

		if (m->hw_handler_params) {
			r = scsi_dh_set_params(q, m->hw_handler_params);
			if (r < 0) {
				ti->error = "unable to set hardware "
							"handler parameters";
				dm_put_device(ti, p->path.dev);
				goto bad;
			}
		}
828 829
	}

L
Linus Torvalds 已提交
830 831 832 833 834 835 836 837 838 839
	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
	if (r) {
		dm_put_device(ti, p->path.dev);
		goto bad;
	}

	return p;

 bad:
	free_pgpath(p);
840
	return ERR_PTR(r);
L
Linus Torvalds 已提交
841 842
}

843
static struct priority_group *parse_priority_group(struct dm_arg_set *as,
M
Micha³ Miros³aw 已提交
844
						   struct multipath *m)
L
Linus Torvalds 已提交
845
{
E
Eric Biggers 已提交
846
	static const struct dm_arg _args[] = {
847 848
		{1, 1024, "invalid number of paths"},
		{0, 1024, "invalid number of selector args"}
L
Linus Torvalds 已提交
849 850 851
	};

	int r;
852
	unsigned i, nr_selector_args, nr_args;
L
Linus Torvalds 已提交
853
	struct priority_group *pg;
M
Micha³ Miros³aw 已提交
854
	struct dm_target *ti = m->ti;
L
Linus Torvalds 已提交
855 856 857

	if (as->argc < 2) {
		as->argc = 0;
858 859
		ti->error = "not enough priority group arguments";
		return ERR_PTR(-EINVAL);
L
Linus Torvalds 已提交
860 861 862 863
	}

	pg = alloc_priority_group();
	if (!pg) {
864
		ti->error = "couldn't allocate priority group";
865
		return ERR_PTR(-ENOMEM);
L
Linus Torvalds 已提交
866 867 868 869 870 871 872 873 874 875
	}
	pg->m = m;

	r = parse_path_selector(as, pg, ti);
	if (r)
		goto bad;

	/*
	 * read the paths
	 */
876
	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
L
Linus Torvalds 已提交
877 878 879
	if (r)
		goto bad;

880
	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
L
Linus Torvalds 已提交
881 882 883
	if (r)
		goto bad;

884
	nr_args = 1 + nr_selector_args;
L
Linus Torvalds 已提交
885 886
	for (i = 0; i < pg->nr_pgpaths; i++) {
		struct pgpath *pgpath;
887
		struct dm_arg_set path_args;
L
Linus Torvalds 已提交
888

889
		if (as->argc < nr_args) {
890
			ti->error = "not enough path parameters";
891
			r = -EINVAL;
L
Linus Torvalds 已提交
892
			goto bad;
893
		}
L
Linus Torvalds 已提交
894

895
		path_args.argc = nr_args;
L
Linus Torvalds 已提交
896 897 898
		path_args.argv = as->argv;

		pgpath = parse_path(&path_args, &pg->ps, ti);
899 900
		if (IS_ERR(pgpath)) {
			r = PTR_ERR(pgpath);
L
Linus Torvalds 已提交
901
			goto bad;
902
		}
L
Linus Torvalds 已提交
903 904 905

		pgpath->pg = pg;
		list_add_tail(&pgpath->list, &pg->pgpaths);
906
		dm_consume_args(as, nr_args);
L
Linus Torvalds 已提交
907 908 909 910 911 912
	}

	return pg;

 bad:
	free_priority_group(pg, ti);
913
	return ERR_PTR(r);
L
Linus Torvalds 已提交
914 915
}

916
static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
L
Linus Torvalds 已提交
917 918
{
	unsigned hw_argc;
919
	int ret;
M
Micha³ Miros³aw 已提交
920
	struct dm_target *ti = m->ti;
L
Linus Torvalds 已提交
921

E
Eric Biggers 已提交
922
	static const struct dm_arg _args[] = {
923
		{0, 1024, "invalid number of hardware handler args"},
L
Linus Torvalds 已提交
924 925
	};

926
	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
L
Linus Torvalds 已提交
927 928 929 930 931
		return -EINVAL;

	if (!hw_argc)
		return 0;

932
	if (m->queue_mode == DM_TYPE_BIO_BASED) {
933 934 935 936 937
		dm_consume_args(as, hw_argc);
		DMERR("bio-based multipath doesn't allow hardware handler args");
		return 0;
	}

938
	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
939 940
	if (!m->hw_handler_name)
		return -EINVAL;
941

942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957
	if (hw_argc > 1) {
		char *p;
		int i, j, len = 4;

		for (i = 0; i <= hw_argc - 2; i++)
			len += strlen(as->argv[i]) + 1;
		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
		if (!p) {
			ti->error = "memory allocation failed";
			ret = -ENOMEM;
			goto fail;
		}
		j = sprintf(p, "%d", hw_argc - 1);
		for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
			j = sprintf(p, "%s", as->argv[i]);
	}
958
	dm_consume_args(as, hw_argc - 1);
L
Linus Torvalds 已提交
959 960

	return 0;
961 962 963 964
fail:
	kfree(m->hw_handler_name);
	m->hw_handler_name = NULL;
	return ret;
L
Linus Torvalds 已提交
965 966
}

967
static int parse_features(struct dm_arg_set *as, struct multipath *m)
L
Linus Torvalds 已提交
968 969 970
{
	int r;
	unsigned argc;
M
Micha³ Miros³aw 已提交
971
	struct dm_target *ti = m->ti;
972
	const char *arg_name;
L
Linus Torvalds 已提交
973

E
Eric Biggers 已提交
974
	static const struct dm_arg _args[] = {
975
		{0, 8, "invalid number of feature args"},
D
Dave Wysochanski 已提交
976
		{1, 50, "pg_init_retries must be between 1 and 50"},
977
		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
L
Linus Torvalds 已提交
978 979
	};

980
	r = dm_read_arg_group(_args, as, &argc, &ti->error);
L
Linus Torvalds 已提交
981 982 983 984 985 986
	if (r)
		return -EINVAL;

	if (!argc)
		return 0;

D
Dave Wysochanski 已提交
987
	do {
988
		arg_name = dm_shift_arg(as);
D
Dave Wysochanski 已提交
989 990
		argc--;

991
		if (!strcasecmp(arg_name, "queue_if_no_path")) {
992
			r = queue_if_no_path(m, true, false);
D
Dave Wysochanski 已提交
993 994 995
			continue;
		}

996
		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
997
			set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
998 999 1000
			continue;
		}

1001
		if (!strcasecmp(arg_name, "pg_init_retries") &&
D
Dave Wysochanski 已提交
1002
		    (argc >= 1)) {
1003
			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
D
Dave Wysochanski 已提交
1004 1005 1006 1007
			argc--;
			continue;
		}

1008
		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1009
		    (argc >= 1)) {
1010
			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1011 1012 1013 1014
			argc--;
			continue;
		}

1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032
		if (!strcasecmp(arg_name, "queue_mode") &&
		    (argc >= 1)) {
			const char *queue_mode_name = dm_shift_arg(as);

			if (!strcasecmp(queue_mode_name, "bio"))
				m->queue_mode = DM_TYPE_BIO_BASED;
			else if (!strcasecmp(queue_mode_name, "rq"))
				m->queue_mode = DM_TYPE_REQUEST_BASED;
			else if (!strcasecmp(queue_mode_name, "mq"))
				m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
			else {
				ti->error = "Unknown 'queue_mode' requested";
				r = -EINVAL;
			}
			argc--;
			continue;
		}

L
Linus Torvalds 已提交
1033
		ti->error = "Unrecognised multipath feature request";
D
Dave Wysochanski 已提交
1034 1035 1036 1037
		r = -EINVAL;
	} while (argc && !r);

	return r;
L
Linus Torvalds 已提交
1038 1039
}

1040
static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
L
Linus Torvalds 已提交
1041
{
1042
	/* target arguments */
E
Eric Biggers 已提交
1043
	static const struct dm_arg _args[] = {
1044 1045
		{0, 1024, "invalid number of priority groups"},
		{0, 1024, "invalid initial priority group number"},
L
Linus Torvalds 已提交
1046 1047 1048 1049
	};

	int r;
	struct multipath *m;
1050
	struct dm_arg_set as;
L
Linus Torvalds 已提交
1051 1052 1053 1054 1055 1056
	unsigned pg_count = 0;
	unsigned next_pg_num;

	as.argc = argc;
	as.argv = argv;

1057
	m = alloc_multipath(ti);
L
Linus Torvalds 已提交
1058
	if (!m) {
1059
		ti->error = "can't allocate multipath";
L
Linus Torvalds 已提交
1060 1061 1062
		return -EINVAL;
	}

M
Micha³ Miros³aw 已提交
1063
	r = parse_features(&as, m);
L
Linus Torvalds 已提交
1064 1065 1066
	if (r)
		goto bad;

1067 1068 1069 1070
	r = alloc_multipath_stage2(ti, m);
	if (r)
		goto bad;

M
Micha³ Miros³aw 已提交
1071
	r = parse_hw_handler(&as, m);
L
Linus Torvalds 已提交
1072 1073 1074
	if (r)
		goto bad;

1075
	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
L
Linus Torvalds 已提交
1076 1077 1078
	if (r)
		goto bad;

1079
	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
L
Linus Torvalds 已提交
1080 1081 1082
	if (r)
		goto bad;

1083 1084 1085 1086 1087 1088 1089
	if ((!m->nr_priority_groups && next_pg_num) ||
	    (m->nr_priority_groups && !next_pg_num)) {
		ti->error = "invalid initial priority group";
		r = -EINVAL;
		goto bad;
	}

L
Linus Torvalds 已提交
1090 1091 1092
	/* parse the priority groups */
	while (as.argc) {
		struct priority_group *pg;
1093
		unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
L
Linus Torvalds 已提交
1094

M
Micha³ Miros³aw 已提交
1095
		pg = parse_priority_group(&as, m);
1096 1097
		if (IS_ERR(pg)) {
			r = PTR_ERR(pg);
L
Linus Torvalds 已提交
1098 1099 1100
			goto bad;
		}

1101 1102 1103
		nr_valid_paths += pg->nr_pgpaths;
		atomic_set(&m->nr_valid_paths, nr_valid_paths);

L
Linus Torvalds 已提交
1104 1105 1106 1107 1108 1109 1110 1111
		list_add_tail(&pg->list, &m->priority_groups);
		pg_count++;
		pg->pg_num = pg_count;
		if (!--next_pg_num)
			m->next_pg = pg;
	}

	if (pg_count != m->nr_priority_groups) {
1112
		ti->error = "priority group count mismatch";
L
Linus Torvalds 已提交
1113 1114 1115 1116
		r = -EINVAL;
		goto bad;
	}

1117 1118
	ti->num_flush_bios = 1;
	ti->num_discard_bios = 1;
1119
	ti->num_write_same_bios = 1;
1120
	ti->num_write_zeroes_bios = 1;
1121
	if (m->queue_mode == DM_TYPE_BIO_BASED)
1122
		ti->per_io_data_size = multipath_per_bio_data_size();
1123
	else
1124
		ti->per_io_data_size = sizeof(struct dm_mpath_io);
M
Mikulas Patocka 已提交
1125

L
Linus Torvalds 已提交
1126 1127 1128 1129 1130 1131 1132
	return 0;

 bad:
	free_multipath(m);
	return r;
}

1133 1134
static void multipath_wait_for_pg_init_completion(struct multipath *m)
{
1135
	DEFINE_WAIT(wait);
1136 1137

	while (1) {
1138
		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1139

1140
		if (!atomic_read(&m->pg_init_in_progress))
1141 1142 1143 1144
			break;

		io_schedule();
	}
1145
	finish_wait(&m->pg_init_wait, &wait);
1146 1147 1148
}

static void flush_multipath_work(struct multipath *m)
L
Linus Torvalds 已提交
1149
{
1150 1151
	set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
	smp_mb__after_atomic();
1152

1153
	flush_workqueue(kmpath_handlerd);
1154
	multipath_wait_for_pg_init_completion(m);
1155
	flush_workqueue(kmultipathd);
1156
	flush_work(&m->trigger_event);
1157

1158 1159
	clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
	smp_mb__after_atomic();
1160 1161 1162 1163 1164 1165
}

static void multipath_dtr(struct dm_target *ti)
{
	struct multipath *m = ti->private;

1166
	flush_multipath_work(m);
L
Linus Torvalds 已提交
1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179
	free_multipath(m);
}

/*
 * Take a path out of use.
 */
static int fail_path(struct pgpath *pgpath)
{
	unsigned long flags;
	struct multipath *m = pgpath->pg->m;

	spin_lock_irqsave(&m->lock, flags);

1180
	if (!pgpath->is_active)
L
Linus Torvalds 已提交
1181 1182
		goto out;

1183
	DMWARN("Failing path %s.", pgpath->path.dev->name);
L
Linus Torvalds 已提交
1184 1185

	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1186
	pgpath->is_active = false;
L
Linus Torvalds 已提交
1187 1188
	pgpath->fail_count++;

1189
	atomic_dec(&m->nr_valid_paths);
L
Linus Torvalds 已提交
1190 1191 1192 1193

	if (pgpath == m->current_pgpath)
		m->current_pgpath = NULL;

M
Mike Anderson 已提交
1194
	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1195
		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
M
Mike Anderson 已提交
1196

1197
	schedule_work(&m->trigger_event);
L
Linus Torvalds 已提交
1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209

out:
	spin_unlock_irqrestore(&m->lock, flags);

	return 0;
}

/*
 * Reinstate a previously-failed path
 */
static int reinstate_path(struct pgpath *pgpath)
{
1210
	int r = 0, run_queue = 0;
L
Linus Torvalds 已提交
1211 1212
	unsigned long flags;
	struct multipath *m = pgpath->pg->m;
1213
	unsigned nr_valid_paths;
L
Linus Torvalds 已提交
1214 1215 1216

	spin_lock_irqsave(&m->lock, flags);

1217
	if (pgpath->is_active)
L
Linus Torvalds 已提交
1218 1219
		goto out;

1220
	DMWARN("Reinstating path %s.", pgpath->path.dev->name);
L
Linus Torvalds 已提交
1221 1222 1223 1224 1225

	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
	if (r)
		goto out;

1226
	pgpath->is_active = true;
L
Linus Torvalds 已提交
1227

1228 1229
	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
	if (nr_valid_paths == 1) {
1230
		m->current_pgpath = NULL;
1231
		run_queue = 1;
1232
	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1233
		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1234
			atomic_inc(&m->pg_init_in_progress);
1235
	}
L
Linus Torvalds 已提交
1236

M
Mike Anderson 已提交
1237
	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1238
		       pgpath->path.dev->name, nr_valid_paths);
M
Mike Anderson 已提交
1239

1240
	schedule_work(&m->trigger_event);
L
Linus Torvalds 已提交
1241 1242 1243

out:
	spin_unlock_irqrestore(&m->lock, flags);
1244
	if (run_queue) {
1245
		dm_table_run_md_queue_async(m->ti->table);
1246
		process_queued_io_list(m);
1247
	}
L
Linus Torvalds 已提交
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257

	return r;
}

/*
 * Fail or reinstate all paths that match the provided struct dm_dev.
 */
static int action_dev(struct multipath *m, struct dm_dev *dev,
		      action_fn action)
{
1258
	int r = -EINVAL;
L
Linus Torvalds 已提交
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275
	struct pgpath *pgpath;
	struct priority_group *pg;

	list_for_each_entry(pg, &m->priority_groups, list) {
		list_for_each_entry(pgpath, &pg->pgpaths, list) {
			if (pgpath->path.dev == dev)
				r = action(pgpath);
		}
	}

	return r;
}

/*
 * Temporarily try to avoid having to use the specified PG
 */
static void bypass_pg(struct multipath *m, struct priority_group *pg,
1276
		      bool bypassed)
L
Linus Torvalds 已提交
1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287
{
	unsigned long flags;

	spin_lock_irqsave(&m->lock, flags);

	pg->bypassed = bypassed;
	m->current_pgpath = NULL;
	m->current_pg = NULL;

	spin_unlock_irqrestore(&m->lock, flags);

1288
	schedule_work(&m->trigger_event);
L
Linus Torvalds 已提交
1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
}

/*
 * Switch to using the specified PG from the next I/O that gets mapped
 */
static int switch_pg_num(struct multipath *m, const char *pgstr)
{
	struct priority_group *pg;
	unsigned pgnum;
	unsigned long flags;
1299
	char dummy;
L
Linus Torvalds 已提交
1300

1301
	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1302
	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
L
Linus Torvalds 已提交
1303 1304 1305 1306 1307 1308
		DMWARN("invalid PG number supplied to switch_pg_num");
		return -EINVAL;
	}

	spin_lock_irqsave(&m->lock, flags);
	list_for_each_entry(pg, &m->priority_groups, list) {
1309
		pg->bypassed = false;
L
Linus Torvalds 已提交
1310 1311 1312 1313 1314 1315 1316 1317 1318
		if (--pgnum)
			continue;

		m->current_pgpath = NULL;
		m->current_pg = NULL;
		m->next_pg = pg;
	}
	spin_unlock_irqrestore(&m->lock, flags);

1319
	schedule_work(&m->trigger_event);
L
Linus Torvalds 已提交
1320 1321 1322 1323 1324 1325 1326
	return 0;
}

/*
 * Set/clear bypassed status of a PG.
 * PGs are numbered upwards from 1 in the order they were declared.
 */
1327
static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
L
Linus Torvalds 已提交
1328 1329 1330
{
	struct priority_group *pg;
	unsigned pgnum;
1331
	char dummy;
L
Linus Torvalds 已提交
1332

1333
	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1334
	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
L
Linus Torvalds 已提交
1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347
		DMWARN("invalid PG number supplied to bypass_pg");
		return -EINVAL;
	}

	list_for_each_entry(pg, &m->priority_groups, list) {
		if (!--pgnum)
			break;
	}

	bypass_pg(m, pg, bypassed);
	return 0;
}

D
Dave Wysochanski 已提交
1348 1349 1350
/*
 * Should we retry pg_init immediately?
 */
1351
static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
D
Dave Wysochanski 已提交
1352 1353
{
	unsigned long flags;
1354
	bool limit_reached = false;
D
Dave Wysochanski 已提交
1355 1356 1357

	spin_lock_irqsave(&m->lock, flags);

1358 1359
	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1360
		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
D
Dave Wysochanski 已提交
1361
	else
1362
		limit_reached = true;
D
Dave Wysochanski 已提交
1363 1364 1365 1366 1367 1368

	spin_unlock_irqrestore(&m->lock, flags);

	return limit_reached;
}

1369
static void pg_init_done(void *data, int errors)
1370
{
1371
	struct pgpath *pgpath = data;
1372 1373 1374
	struct priority_group *pg = pgpath->pg;
	struct multipath *m = pg->m;
	unsigned long flags;
1375
	bool delay_retry = false;
1376 1377 1378 1379 1380 1381 1382 1383 1384 1385

	/* device or driver problems */
	switch (errors) {
	case SCSI_DH_OK:
		break;
	case SCSI_DH_NOSYS:
		if (!m->hw_handler_name) {
			errors = 0;
			break;
		}
1386 1387
		DMERR("Could not failover the device: Handler scsi_dh_%s "
		      "Error %d.", m->hw_handler_name, errors);
1388 1389 1390 1391 1392 1393 1394 1395 1396 1397
		/*
		 * Fail path for now, so we do not ping pong
		 */
		fail_path(pgpath);
		break;
	case SCSI_DH_DEV_TEMP_BUSY:
		/*
		 * Probably doing something like FW upgrade on the
		 * controller so try the other pg.
		 */
1398
		bypass_pg(m, pg, true);
1399 1400
		break;
	case SCSI_DH_RETRY:
1401 1402
		/* Wait before retrying. */
		delay_retry = 1;
1403
		/* fall through */
1404 1405 1406 1407 1408 1409
	case SCSI_DH_IMM_RETRY:
	case SCSI_DH_RES_TEMP_UNAVAIL:
		if (pg_init_limit_reached(m, pgpath))
			fail_path(pgpath);
		errors = 0;
		break;
1410
	case SCSI_DH_DEV_OFFLINED:
1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421
	default:
		/*
		 * We probably do not want to fail the path for a device
		 * error, but this is what the old dm did. In future
		 * patches we can do more advanced handling.
		 */
		fail_path(pgpath);
	}

	spin_lock_irqsave(&m->lock, flags);
	if (errors) {
1422 1423 1424 1425 1426
		if (pgpath == m->current_pgpath) {
			DMERR("Could not failover device. Error %d.", errors);
			m->current_pgpath = NULL;
			m->current_pg = NULL;
		}
1427
	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1428
		pg->bypassed = false;
1429

1430
	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1431 1432 1433
		/* Activations of other paths are still on going */
		goto out;

1434 1435 1436 1437 1438 1439
	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
		if (delay_retry)
			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
		else
			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);

1440 1441 1442
		if (__pg_init_all_paths(m))
			goto out;
	}
1443
	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1444

1445
	process_queued_io_list(m);
1446

1447 1448 1449 1450 1451
	/*
	 * Wake up any thread waiting to suspend.
	 */
	wake_up(&m->pg_init_wait);

1452
out:
1453 1454 1455
	spin_unlock_irqrestore(&m->lock, flags);
}

1456
static void activate_or_offline_path(struct pgpath *pgpath)
1457
{
1458
	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1459

1460 1461
	if (pgpath->is_active && !blk_queue_dying(q))
		scsi_dh_activate(q, pg_init_done, pgpath);
1462 1463
	else
		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1464 1465
}

1466 1467 1468 1469 1470 1471 1472 1473
static void activate_path_work(struct work_struct *work)
{
	struct pgpath *pgpath =
		container_of(work, struct pgpath, activate_path.work);

	activate_or_offline_path(pgpath);
}

1474
static int noretry_error(blk_status_t error)
1475 1476
{
	switch (error) {
1477 1478 1479 1480 1481
	case BLK_STS_NOTSUPP:
	case BLK_STS_NOSPC:
	case BLK_STS_TARGET:
	case BLK_STS_NEXUS:
	case BLK_STS_MEDIUM:
1482 1483 1484 1485 1486 1487 1488
		return 1;
	}

	/* Anything else could be a path failure, so should be retried */
	return 0;
}

1489
static int multipath_end_io(struct dm_target *ti, struct request *clone,
1490
			    blk_status_t error, union map_info *map_context)
L
Linus Torvalds 已提交
1491
{
1492 1493
	struct dm_mpath_io *mpio = get_mpio(map_context);
	struct pgpath *pgpath = mpio->pgpath;
1494
	int r = DM_ENDIO_DONE;
1495

1496 1497 1498 1499 1500
	/*
	 * We don't queue any clone request inside the multipath target
	 * during end I/O handling, since those clone requests don't have
	 * bio clones.  If we queue them inside the multipath target,
	 * we need to make bio clones, that requires memory allocation.
1501
	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1502 1503 1504 1505 1506
	 *  don't have bio clones.)
	 * Instead of queueing the clone request here, we queue the original
	 * request into dm core, which will remake a clone request and
	 * clone bios for it and resubmit it later.
	 */
1507 1508
	if (error && !noretry_error(error)) {
		struct multipath *m = ti->private;
L
Linus Torvalds 已提交
1509

1510
		r = DM_ENDIO_REQUEUE;
L
Linus Torvalds 已提交
1511

1512 1513
		if (pgpath)
			fail_path(pgpath);
L
Linus Torvalds 已提交
1514

1515
		if (atomic_read(&m->nr_valid_paths) == 0 &&
1516
		    !must_push_back_rq(m)) {
1517
			if (error == BLK_STS_IOERR)
1518
				dm_report_EIO(m);
1519 1520 1521
			/* complete with the original error */
			r = DM_ENDIO_DONE;
		}
1522
	}
1523

L
Linus Torvalds 已提交
1524
	if (pgpath) {
1525 1526
		struct path_selector *ps = &pgpath->pg->ps;

L
Linus Torvalds 已提交
1527
		if (ps->type->end_io)
1528
			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
L
Linus Torvalds 已提交
1529 1530
	}

1531
	return r;
L
Linus Torvalds 已提交
1532 1533
}

1534 1535
static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
		blk_status_t *error)
1536
{
1537 1538 1539
	struct multipath *m = ti->private;
	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
	struct pgpath *pgpath = mpio->pgpath;
1540
	unsigned long flags;
1541
	int r = DM_ENDIO_DONE;
1542

1543
	if (!*error || noretry_error(*error))
1544
		goto done;
1545

1546 1547
	if (pgpath)
		fail_path(pgpath);
1548

1549
	if (atomic_read(&m->nr_valid_paths) == 0 &&
1550
	    !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1551 1552 1553 1554 1555 1556
		if (must_push_back_bio(m)) {
			r = DM_ENDIO_REQUEUE;
		} else {
			dm_report_EIO(m);
			*error = BLK_STS_IOERR;
		}
1557
		goto done;
1558
	}
1559 1560

	/* Queue for the daemon to resubmit */
1561
	dm_bio_restore(get_bio_details_from_bio(clone), clone);
1562 1563 1564 1565 1566 1567 1568

	spin_lock_irqsave(&m->lock, flags);
	bio_list_add(&m->queued_bios, clone);
	spin_unlock_irqrestore(&m->lock, flags);
	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
		queue_work(kmultipathd, &m->process_queued_bios);

1569
	r = DM_ENDIO_INCOMPLETE;
1570
done:
1571
	if (pgpath) {
1572 1573
		struct path_selector *ps = &pgpath->pg->ps;

1574 1575 1576 1577
		if (ps->type->end_io)
			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
	}

1578
	return r;
1579 1580
}

L
Linus Torvalds 已提交
1581 1582
/*
 * Suspend can't complete until all the I/O is processed so if
1583 1584 1585
 * the last path fails we must error any remaining I/O.
 * Note that if the freeze_bdev fails while suspending, the
 * queue_if_no_path state is lost - userspace should reset it.
L
Linus Torvalds 已提交
1586 1587 1588
 */
static void multipath_presuspend(struct dm_target *ti)
{
1589
	struct multipath *m = ti->private;
L
Linus Torvalds 已提交
1590

1591
	queue_if_no_path(m, false, true);
L
Linus Torvalds 已提交
1592 1593
}

1594 1595
static void multipath_postsuspend(struct dm_target *ti)
{
1596 1597 1598
	struct multipath *m = ti->private;

	mutex_lock(&m->work_mutex);
1599
	flush_multipath_work(m);
1600
	mutex_unlock(&m->work_mutex);
1601 1602
}

1603 1604 1605
/*
 * Restore the queue_if_no_path setting.
 */
L
Linus Torvalds 已提交
1606 1607
static void multipath_resume(struct dm_target *ti)
{
1608
	struct multipath *m = ti->private;
1609
	unsigned long flags;
L
Linus Torvalds 已提交
1610

1611
	spin_lock_irqsave(&m->lock, flags);
L
Lukas Wunner 已提交
1612 1613
	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
		   test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1614
	spin_unlock_irqrestore(&m->lock, flags);
L
Linus Torvalds 已提交
1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
}

/*
 * Info output has the following format:
 * num_multipath_feature_args [multipath_feature_args]*
 * num_handler_status_args [handler_status_args]*
 * num_groups init_group_number
 *            [A|D|E num_ps_status_args [ps_status_args]*
 *             num_paths num_selector_args
 *             [path_dev A|F fail_count [selector_args]* ]+ ]+
 *
 * Table output has the following format (identical to the constructor string):
 * num_feature_args [features_args]*
 * num_handler_args hw_handler [hw_handler_args]*
 * num_groups init_group_number
 *     [priority selector-name num_ps_args [ps_args]*
 *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
 */
1633 1634
static void multipath_status(struct dm_target *ti, status_type_t type,
			     unsigned status_flags, char *result, unsigned maxlen)
L
Linus Torvalds 已提交
1635 1636 1637
{
	int sz = 0;
	unsigned long flags;
1638
	struct multipath *m = ti->private;
L
Linus Torvalds 已提交
1639 1640 1641 1642 1643 1644 1645 1646 1647
	struct priority_group *pg;
	struct pgpath *p;
	unsigned pg_num;
	char state;

	spin_lock_irqsave(&m->lock, flags);

	/* Features */
	if (type == STATUSTYPE_INFO)
1648 1649
		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
		       atomic_read(&m->pg_init_count));
D
Dave Wysochanski 已提交
1650
	else {
1651
		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1652
			      (m->pg_init_retries > 0) * 2 +
1653
			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1654 1655 1656
			      test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);

1657
		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
D
Dave Wysochanski 已提交
1658 1659 1660
			DMEMIT("queue_if_no_path ");
		if (m->pg_init_retries)
			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1661 1662
		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1663
		if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1664
			DMEMIT("retain_attached_hw_handler ");
1665 1666 1667 1668 1669 1670 1671 1672
		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
			switch(m->queue_mode) {
			case DM_TYPE_BIO_BASED:
				DMEMIT("queue_mode bio ");
				break;
			case DM_TYPE_MQ_REQUEST_BASED:
				DMEMIT("queue_mode mq ");
				break;
1673 1674 1675
			default:
				WARN_ON_ONCE(true);
				break;
1676 1677
			}
		}
D
Dave Wysochanski 已提交
1678
	}
L
Linus Torvalds 已提交
1679

1680
	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
L
Linus Torvalds 已提交
1681 1682
		DMEMIT("0 ");
	else
1683
		DMEMIT("1 %s ", m->hw_handler_name);
L
Linus Torvalds 已提交
1684 1685 1686 1687 1688 1689 1690 1691

	DMEMIT("%u ", m->nr_priority_groups);

	if (m->next_pg)
		pg_num = m->next_pg->pg_num;
	else if (m->current_pg)
		pg_num = m->current_pg->pg_num;
	else
1692
		pg_num = (m->nr_priority_groups ? 1 : 0);
L
Linus Torvalds 已提交
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719

	DMEMIT("%u ", pg_num);

	switch (type) {
	case STATUSTYPE_INFO:
		list_for_each_entry(pg, &m->priority_groups, list) {
			if (pg->bypassed)
				state = 'D';	/* Disabled */
			else if (pg == m->current_pg)
				state = 'A';	/* Currently Active */
			else
				state = 'E';	/* Enabled */

			DMEMIT("%c ", state);

			if (pg->ps.type->status)
				sz += pg->ps.type->status(&pg->ps, NULL, type,
							  result + sz,
							  maxlen - sz);
			else
				DMEMIT("0 ");

			DMEMIT("%u %u ", pg->nr_pgpaths,
			       pg->ps.type->info_args);

			list_for_each_entry(p, &pg->pgpaths, list) {
				DMEMIT("%s %s %u ", p->path.dev->name,
1720
				       p->is_active ? "A" : "F",
L
Linus Torvalds 已提交
1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759
				       p->fail_count);
				if (pg->ps.type->status)
					sz += pg->ps.type->status(&pg->ps,
					      &p->path, type, result + sz,
					      maxlen - sz);
			}
		}
		break;

	case STATUSTYPE_TABLE:
		list_for_each_entry(pg, &m->priority_groups, list) {
			DMEMIT("%s ", pg->ps.type->name);

			if (pg->ps.type->status)
				sz += pg->ps.type->status(&pg->ps, NULL, type,
							  result + sz,
							  maxlen - sz);
			else
				DMEMIT("0 ");

			DMEMIT("%u %u ", pg->nr_pgpaths,
			       pg->ps.type->table_args);

			list_for_each_entry(p, &pg->pgpaths, list) {
				DMEMIT("%s ", p->path.dev->name);
				if (pg->ps.type->status)
					sz += pg->ps.type->status(&pg->ps,
					      &p->path, type, result + sz,
					      maxlen - sz);
			}
		}
		break;
	}

	spin_unlock_irqrestore(&m->lock, flags);
}

static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
{
1760
	int r = -EINVAL;
L
Linus Torvalds 已提交
1761
	struct dm_dev *dev;
1762
	struct multipath *m = ti->private;
L
Linus Torvalds 已提交
1763 1764
	action_fn action;

1765 1766
	mutex_lock(&m->work_mutex);

1767 1768 1769 1770 1771
	if (dm_suspended(ti)) {
		r = -EBUSY;
		goto out;
	}

L
Linus Torvalds 已提交
1772
	if (argc == 1) {
1773
		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1774
			r = queue_if_no_path(m, true, false);
1775
			goto out;
1776
		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1777
			r = queue_if_no_path(m, false, false);
1778 1779
			goto out;
		}
L
Linus Torvalds 已提交
1780 1781
	}

1782
	if (argc != 2) {
1783
		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1784 1785
		goto out;
	}
L
Linus Torvalds 已提交
1786

1787
	if (!strcasecmp(argv[0], "disable_group")) {
1788
		r = bypass_pg_num(m, argv[1], true);
1789
		goto out;
1790
	} else if (!strcasecmp(argv[0], "enable_group")) {
1791
		r = bypass_pg_num(m, argv[1], false);
1792
		goto out;
1793
	} else if (!strcasecmp(argv[0], "switch_group")) {
1794 1795
		r = switch_pg_num(m, argv[1]);
		goto out;
1796
	} else if (!strcasecmp(argv[0], "reinstate_path"))
L
Linus Torvalds 已提交
1797
		action = reinstate_path;
1798
	else if (!strcasecmp(argv[0], "fail_path"))
L
Linus Torvalds 已提交
1799
		action = fail_path;
1800
	else {
1801
		DMWARN("Unrecognised multipath message received: %s", argv[0]);
1802 1803
		goto out;
	}
L
Linus Torvalds 已提交
1804

1805
	r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
L
Linus Torvalds 已提交
1806
	if (r) {
1807
		DMWARN("message: error getting device %s",
L
Linus Torvalds 已提交
1808
		       argv[1]);
1809
		goto out;
L
Linus Torvalds 已提交
1810 1811 1812 1813 1814 1815
	}

	r = action_dev(m, dev, action);

	dm_put_device(ti, dev);

1816 1817
out:
	mutex_unlock(&m->work_mutex);
L
Linus Torvalds 已提交
1818 1819 1820
	return r;
}

C
Christoph Hellwig 已提交
1821 1822
static int multipath_prepare_ioctl(struct dm_target *ti,
		struct block_device **bdev, fmode_t *mode)
M
Milan Broz 已提交
1823
{
1824
	struct multipath *m = ti->private;
1825
	struct pgpath *current_pgpath;
1826 1827
	int r;

1828
	current_pgpath = READ_ONCE(m->current_pgpath);
1829 1830
	if (!current_pgpath)
		current_pgpath = choose_pgpath(m, 0);
M
Milan Broz 已提交
1831

1832
	if (current_pgpath) {
1833
		if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1834 1835
			*bdev = current_pgpath->path.dev->bdev;
			*mode = current_pgpath->path.dev->mode;
1836 1837 1838 1839 1840 1841 1842
			r = 0;
		} else {
			/* pg_init has not started or completed */
			r = -ENOTCONN;
		}
	} else {
		/* No path is available */
1843
		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1844 1845 1846
			r = -ENOTCONN;
		else
			r = -EIO;
1847
	}
M
Milan Broz 已提交
1848

1849
	if (r == -ENOTCONN) {
1850
		if (!READ_ONCE(m->current_pg)) {
1851
			/* Path status changed, redo selection */
1852
			(void) choose_pgpath(m, 0);
1853
		}
1854
		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1855
			pg_init_all_paths(m);
1856
		dm_table_run_md_queue_async(m->ti->table);
1857
		process_queued_io_list(m);
1858
	}
1859

C
Christoph Hellwig 已提交
1860 1861 1862 1863 1864 1865
	/*
	 * Only pass ioctls through if the device sizes match exactly.
	 */
	if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
		return 1;
	return r;
M
Milan Broz 已提交
1866 1867
}

1868 1869 1870 1871 1872 1873 1874 1875 1876 1877
static int multipath_iterate_devices(struct dm_target *ti,
				     iterate_devices_callout_fn fn, void *data)
{
	struct multipath *m = ti->private;
	struct priority_group *pg;
	struct pgpath *p;
	int ret = 0;

	list_for_each_entry(pg, &m->priority_groups, list) {
		list_for_each_entry(p, &pg->pgpaths, list) {
1878
			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1879 1880 1881 1882 1883 1884 1885 1886 1887
			if (ret)
				goto out;
		}
	}

out:
	return ret;
}

1888
static int pgpath_busy(struct pgpath *pgpath)
1889 1890 1891
{
	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);

1892
	return blk_lld_busy(q);
1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904
}

/*
 * We return "busy", only when we can map I/Os but underlying devices
 * are busy (so even if we map I/Os now, the I/Os will wait on
 * the underlying queue).
 * In other words, if we want to kill I/Os or queue them inside us
 * due to map unavailability, we don't return "busy".  Otherwise,
 * dm core won't give us the I/Os and we can't do what we want.
 */
static int multipath_busy(struct dm_target *ti)
{
1905
	bool busy = false, has_active = false;
1906
	struct multipath *m = ti->private;
1907
	struct priority_group *pg, *next_pg;
1908 1909
	struct pgpath *pgpath;

1910 1911
	/* pg_init in progress */
	if (atomic_read(&m->pg_init_in_progress))
1912 1913
		return true;

1914 1915 1916 1917
	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
	if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
		return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);

1918
	/* Guess which priority_group will be used at next mapping time */
1919 1920 1921
	pg = READ_ONCE(m->current_pg);
	next_pg = READ_ONCE(m->next_pg);
	if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1922 1923 1924
		pg = next_pg;

	if (!pg) {
1925 1926
		/*
		 * We don't know which pg will be used at next mapping time.
1927
		 * We don't call choose_pgpath() here to avoid to trigger
1928 1929 1930 1931
		 * pg_init just by busy checking.
		 * So we don't know whether underlying devices we will be using
		 * at next mapping time are busy or not. Just try mapping.
		 */
1932 1933
		return busy;
	}
1934 1935 1936 1937 1938

	/*
	 * If there is one non-busy active path at least, the path selector
	 * will be able to select it. So we consider such a pg as not busy.
	 */
1939
	busy = true;
1940
	list_for_each_entry(pgpath, &pg->pgpaths, list) {
1941
		if (pgpath->is_active) {
1942
			has_active = true;
1943
			if (!pgpath_busy(pgpath)) {
1944
				busy = false;
1945 1946 1947
				break;
			}
		}
1948
	}
1949

1950
	if (!has_active) {
1951 1952 1953 1954 1955
		/*
		 * No active path in this pg, so this pg won't be used and
		 * the current_pg will be changed at next mapping time.
		 * We need to try mapping to determine it.
		 */
1956
		busy = false;
1957
	}
1958 1959 1960 1961

	return busy;
}

L
Linus Torvalds 已提交
1962 1963 1964 1965 1966
/*-----------------------------------------------------------------
 * Module setup
 *---------------------------------------------------------------*/
static struct target_type multipath_target = {
	.name = "multipath",
1967
	.version = {1, 12, 0},
M
Mike Snitzer 已提交
1968
	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
L
Linus Torvalds 已提交
1969 1970 1971
	.module = THIS_MODULE,
	.ctr = multipath_ctr,
	.dtr = multipath_dtr,
1972 1973
	.clone_and_map_rq = multipath_clone_and_map,
	.release_clone_rq = multipath_release_clone,
1974
	.rq_end_io = multipath_end_io,
1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986
	.map = multipath_map_bio,
	.end_io = multipath_end_io_bio,
	.presuspend = multipath_presuspend,
	.postsuspend = multipath_postsuspend,
	.resume = multipath_resume,
	.status = multipath_status,
	.message = multipath_message,
	.prepare_ioctl = multipath_prepare_ioctl,
	.iterate_devices = multipath_iterate_devices,
	.busy = multipath_busy,
};

L
Linus Torvalds 已提交
1987 1988 1989 1990
static int __init dm_multipath_init(void)
{
	int r;

1991
	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1992
	if (!kmultipathd) {
1993
		DMERR("failed to create workqueue kmpathd");
1994 1995
		r = -ENOMEM;
		goto bad_alloc_kmultipathd;
1996 1997
	}

1998 1999 2000 2001 2002 2003
	/*
	 * A separate workqueue is used to handle the device handlers
	 * to avoid overloading existing workqueue. Overloading the
	 * old workqueue would also create a bottleneck in the
	 * path of the storage hardware device activation.
	 */
2004 2005
	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
						  WQ_MEM_RECLAIM);
2006 2007
	if (!kmpath_handlerd) {
		DMERR("failed to create workqueue kmpath_handlerd");
2008 2009
		r = -ENOMEM;
		goto bad_alloc_kmpath_handlerd;
2010 2011
	}

2012 2013 2014 2015 2016 2017 2018
	r = dm_register_target(&multipath_target);
	if (r < 0) {
		DMERR("request-based register failed %d", r);
		r = -EINVAL;
		goto bad_register_target;
	}

2019 2020
	return 0;

2021 2022
bad_register_target:
	destroy_workqueue(kmpath_handlerd);
2023 2024 2025
bad_alloc_kmpath_handlerd:
	destroy_workqueue(kmultipathd);
bad_alloc_kmultipathd:
L
Linus Torvalds 已提交
2026 2027 2028 2029 2030
	return r;
}

static void __exit dm_multipath_exit(void)
{
2031
	destroy_workqueue(kmpath_handlerd);
2032 2033
	destroy_workqueue(kmultipathd);

2034
	dm_unregister_target(&multipath_target);
L
Linus Torvalds 已提交
2035 2036 2037 2038 2039 2040 2041 2042
}

module_init(dm_multipath_init);
module_exit(dm_multipath_exit);

MODULE_DESCRIPTION(DM_NAME " multipath target");
MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
MODULE_LICENSE("GPL");