xref: /openbmc/linux/drivers/md/dm-mpath.c (revision ca481398)
1 /*
2  * Copyright (C) 2003 Sistina Software Limited.
3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include <linux/device-mapper.h>
9 
10 #include "dm-rq.h"
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
14 
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
28 
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32 
33 /* Path properties */
34 struct pgpath {
35 	struct list_head list;
36 
37 	struct priority_group *pg;	/* Owning PG */
38 	unsigned fail_count;		/* Cumulative failure count */
39 
40 	struct dm_path path;
41 	struct delayed_work activate_path;
42 
43 	bool is_active:1;		/* Path status */
44 };
45 
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
47 
48 /*
49  * Paths are grouped into Priority Groups and numbered from 1 upwards.
50  * Each has a path selector which controls which path gets used.
51  */
52 struct priority_group {
53 	struct list_head list;
54 
55 	struct multipath *m;		/* Owning multipath instance */
56 	struct path_selector ps;
57 
58 	unsigned pg_num;		/* Reference number */
59 	unsigned nr_pgpaths;		/* Number of paths in PG */
60 	struct list_head pgpaths;
61 
62 	bool bypassed:1;		/* Temporarily bypass this PG? */
63 };
64 
65 /* Multipath context */
66 struct multipath {
67 	struct list_head list;
68 	struct dm_target *ti;
69 
70 	const char *hw_handler_name;
71 	char *hw_handler_params;
72 
73 	spinlock_t lock;
74 
75 	unsigned nr_priority_groups;
76 	struct list_head priority_groups;
77 
78 	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
79 
80 	struct pgpath *current_pgpath;
81 	struct priority_group *current_pg;
82 	struct priority_group *next_pg;	/* Switch to this PG if set */
83 
84 	unsigned long flags;		/* Multipath state flags */
85 
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 */
88 
89 	atomic_t nr_valid_paths;	/* Total number of usable paths */
90 	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
91 	atomic_t pg_init_count;		/* Number of times pg_init called */
92 
93 	enum dm_queue_mode queue_mode;
94 
95 	struct mutex work_mutex;
96 	struct work_struct trigger_event;
97 
98 	struct work_struct process_queued_bios;
99 	struct bio_list queued_bios;
100 };
101 
102 /*
103  * Context information attached to each io we process.
104  */
105 struct dm_mpath_io {
106 	struct pgpath *pgpath;
107 	size_t nr_bytes;
108 };
109 
110 typedef int (*action_fn) (struct pgpath *pgpath);
111 
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void trigger_event(struct work_struct *work);
114 static void activate_or_offline_path(struct pgpath *pgpath);
115 static void activate_path_work(struct work_struct *work);
116 static void process_queued_bios(struct work_struct *work);
117 
118 /*-----------------------------------------------
119  * Multipath state flags.
120  *-----------------------------------------------*/
121 
122 #define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3	/* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
129 
130 /*-----------------------------------------------
131  * Allocation routines
132  *-----------------------------------------------*/
133 
134 static struct pgpath *alloc_pgpath(void)
135 {
136 	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
137 
138 	if (pgpath) {
139 		pgpath->is_active = true;
140 		INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
141 	}
142 
143 	return pgpath;
144 }
145 
146 static void free_pgpath(struct pgpath *pgpath)
147 {
148 	kfree(pgpath);
149 }
150 
151 static struct priority_group *alloc_priority_group(void)
152 {
153 	struct priority_group *pg;
154 
155 	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
156 
157 	if (pg)
158 		INIT_LIST_HEAD(&pg->pgpaths);
159 
160 	return pg;
161 }
162 
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
164 {
165 	struct pgpath *pgpath, *tmp;
166 
167 	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168 		list_del(&pgpath->list);
169 		dm_put_device(ti, pgpath->path.dev);
170 		free_pgpath(pgpath);
171 	}
172 }
173 
174 static void free_priority_group(struct priority_group *pg,
175 				struct dm_target *ti)
176 {
177 	struct path_selector *ps = &pg->ps;
178 
179 	if (ps->type) {
180 		ps->type->destroy(ps);
181 		dm_put_path_selector(ps->type);
182 	}
183 
184 	free_pgpaths(&pg->pgpaths, ti);
185 	kfree(pg);
186 }
187 
188 static struct multipath *alloc_multipath(struct dm_target *ti)
189 {
190 	struct multipath *m;
191 
192 	m = kzalloc(sizeof(*m), GFP_KERNEL);
193 	if (m) {
194 		INIT_LIST_HEAD(&m->priority_groups);
195 		spin_lock_init(&m->lock);
196 		set_bit(MPATHF_QUEUE_IO, &m->flags);
197 		atomic_set(&m->nr_valid_paths, 0);
198 		atomic_set(&m->pg_init_in_progress, 0);
199 		atomic_set(&m->pg_init_count, 0);
200 		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
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 
207 		m->ti = ti;
208 		ti->private = m;
209 	}
210 
211 	return m;
212 }
213 
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
215 {
216 	if (m->queue_mode == DM_TYPE_NONE) {
217 		/*
218 		 * Default to request-based.
219 		 */
220 		if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221 			m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
222 		else
223 			m->queue_mode = DM_TYPE_REQUEST_BASED;
224 	} else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225 		INIT_WORK(&m->process_queued_bios, process_queued_bios);
226 		/*
227 		 * bio-based doesn't support any direct scsi_dh management;
228 		 * it just discovers if a scsi_dh is attached.
229 		 */
230 		set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
231 	}
232 
233 	dm_table_set_type(ti->table, m->queue_mode);
234 
235 	return 0;
236 }
237 
238 static void free_multipath(struct multipath *m)
239 {
240 	struct priority_group *pg, *tmp;
241 
242 	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
243 		list_del(&pg->list);
244 		free_priority_group(pg, m->ti);
245 	}
246 
247 	kfree(m->hw_handler_name);
248 	kfree(m->hw_handler_params);
249 	kfree(m);
250 }
251 
252 static struct dm_mpath_io *get_mpio(union map_info *info)
253 {
254 	return info->ptr;
255 }
256 
257 static size_t multipath_per_bio_data_size(void)
258 {
259 	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
260 }
261 
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
263 {
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 	void *bio_details = mpio + 1;
272 
273 	return bio_details;
274 }
275 
276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
277 					struct dm_bio_details **bio_details_p)
278 {
279 	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280 	struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
281 
282 	memset(mpio, 0, sizeof(*mpio));
283 	memset(bio_details, 0, sizeof(*bio_details));
284 	dm_bio_record(bio_details, bio);
285 
286 	if (mpio_p)
287 		*mpio_p = mpio;
288 	if (bio_details_p)
289 		*bio_details_p = bio_details;
290 }
291 
292 /*-----------------------------------------------
293  * Path selection
294  *-----------------------------------------------*/
295 
296 static int __pg_init_all_paths(struct multipath *m)
297 {
298 	struct pgpath *pgpath;
299 	unsigned long pg_init_delay = 0;
300 
301 	lockdep_assert_held(&m->lock);
302 
303 	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
304 		return 0;
305 
306 	atomic_inc(&m->pg_init_count);
307 	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
308 
309 	/* Check here to reset pg_init_required */
310 	if (!m->current_pg)
311 		return 0;
312 
313 	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
314 		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
315 						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
316 	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
317 		/* Skip failed paths */
318 		if (!pgpath->is_active)
319 			continue;
320 		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
321 				       pg_init_delay))
322 			atomic_inc(&m->pg_init_in_progress);
323 	}
324 	return atomic_read(&m->pg_init_in_progress);
325 }
326 
327 static int pg_init_all_paths(struct multipath *m)
328 {
329 	int ret;
330 	unsigned long flags;
331 
332 	spin_lock_irqsave(&m->lock, flags);
333 	ret = __pg_init_all_paths(m);
334 	spin_unlock_irqrestore(&m->lock, flags);
335 
336 	return ret;
337 }
338 
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
340 {
341 	m->current_pg = pg;
342 
343 	/* Must we initialise the PG first, and queue I/O till it's ready? */
344 	if (m->hw_handler_name) {
345 		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346 		set_bit(MPATHF_QUEUE_IO, &m->flags);
347 	} else {
348 		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349 		clear_bit(MPATHF_QUEUE_IO, &m->flags);
350 	}
351 
352 	atomic_set(&m->pg_init_count, 0);
353 }
354 
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356 					struct priority_group *pg,
357 					size_t nr_bytes)
358 {
359 	unsigned long flags;
360 	struct dm_path *path;
361 	struct pgpath *pgpath;
362 
363 	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
364 	if (!path)
365 		return ERR_PTR(-ENXIO);
366 
367 	pgpath = path_to_pgpath(path);
368 
369 	if (unlikely(lockless_dereference(m->current_pg) != pg)) {
370 		/* Only update current_pgpath if pg changed */
371 		spin_lock_irqsave(&m->lock, flags);
372 		m->current_pgpath = pgpath;
373 		__switch_pg(m, pg);
374 		spin_unlock_irqrestore(&m->lock, flags);
375 	}
376 
377 	return pgpath;
378 }
379 
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
381 {
382 	unsigned long flags;
383 	struct priority_group *pg;
384 	struct pgpath *pgpath;
385 	unsigned bypassed = 1;
386 
387 	if (!atomic_read(&m->nr_valid_paths)) {
388 		clear_bit(MPATHF_QUEUE_IO, &m->flags);
389 		goto failed;
390 	}
391 
392 	/* Were we instructed to switch PG? */
393 	if (lockless_dereference(m->next_pg)) {
394 		spin_lock_irqsave(&m->lock, flags);
395 		pg = m->next_pg;
396 		if (!pg) {
397 			spin_unlock_irqrestore(&m->lock, flags);
398 			goto check_current_pg;
399 		}
400 		m->next_pg = NULL;
401 		spin_unlock_irqrestore(&m->lock, flags);
402 		pgpath = choose_path_in_pg(m, pg, nr_bytes);
403 		if (!IS_ERR_OR_NULL(pgpath))
404 			return pgpath;
405 	}
406 
407 	/* Don't change PG until it has no remaining paths */
408 check_current_pg:
409 	pg = lockless_dereference(m->current_pg);
410 	if (pg) {
411 		pgpath = choose_path_in_pg(m, pg, nr_bytes);
412 		if (!IS_ERR_OR_NULL(pgpath))
413 			return pgpath;
414 	}
415 
416 	/*
417 	 * Loop through priority groups until we find a valid path.
418 	 * First time we skip PGs marked 'bypassed'.
419 	 * Second time we only try the ones we skipped, but set
420 	 * pg_init_delay_retry so we do not hammer controllers.
421 	 */
422 	do {
423 		list_for_each_entry(pg, &m->priority_groups, list) {
424 			if (pg->bypassed == !!bypassed)
425 				continue;
426 			pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 			if (!IS_ERR_OR_NULL(pgpath)) {
428 				if (!bypassed)
429 					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
430 				return pgpath;
431 			}
432 		}
433 	} while (bypassed--);
434 
435 failed:
436 	spin_lock_irqsave(&m->lock, flags);
437 	m->current_pgpath = NULL;
438 	m->current_pg = NULL;
439 	spin_unlock_irqrestore(&m->lock, flags);
440 
441 	return NULL;
442 }
443 
444 /*
445  * dm_report_EIO() is a macro instead of a function to make pr_debug()
446  * report the function name and line number of the function from which
447  * it has been invoked.
448  */
449 #define dm_report_EIO(m)						\
450 do {									\
451 	struct mapped_device *md = dm_table_get_md((m)->ti->table);	\
452 									\
453 	pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454 		 dm_device_name(md),					\
455 		 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
456 		 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags),	\
457 		 dm_noflush_suspending((m)->ti));			\
458 } while (0)
459 
460 /*
461  * Map cloned requests (request-based multipath)
462  */
463 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
464 				   union map_info *map_context,
465 				   struct request **__clone)
466 {
467 	struct multipath *m = ti->private;
468 	size_t nr_bytes = blk_rq_bytes(rq);
469 	struct pgpath *pgpath;
470 	struct block_device *bdev;
471 	struct dm_mpath_io *mpio = get_mpio(map_context);
472 	struct request_queue *q;
473 	struct request *clone;
474 
475 	/* Do we need to select a new pgpath? */
476 	pgpath = lockless_dereference(m->current_pgpath);
477 	if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
478 		pgpath = choose_pgpath(m, nr_bytes);
479 
480 	if (!pgpath) {
481 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
482 			return DM_MAPIO_DELAY_REQUEUE;
483 		dm_report_EIO(m);	/* Failed */
484 		return DM_MAPIO_KILL;
485 	} else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
486 		   test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
487 		if (pg_init_all_paths(m))
488 			return DM_MAPIO_DELAY_REQUEUE;
489 		return DM_MAPIO_REQUEUE;
490 	}
491 
492 	memset(mpio, 0, sizeof(*mpio));
493 	mpio->pgpath = pgpath;
494 	mpio->nr_bytes = nr_bytes;
495 
496 	bdev = pgpath->path.dev->bdev;
497 	q = bdev_get_queue(bdev);
498 	clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
499 	if (IS_ERR(clone)) {
500 		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501 		bool queue_dying = blk_queue_dying(q);
502 		DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
503 			    PTR_ERR(clone), queue_dying ? " (path offline)" : "");
504 		if (queue_dying) {
505 			atomic_inc(&m->pg_init_in_progress);
506 			activate_or_offline_path(pgpath);
507 		}
508 		return DM_MAPIO_DELAY_REQUEUE;
509 	}
510 	clone->bio = clone->biotail = NULL;
511 	clone->rq_disk = bdev->bd_disk;
512 	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
513 	*__clone = clone;
514 
515 	if (pgpath->pg->ps.type->start_io)
516 		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
517 					      &pgpath->path,
518 					      nr_bytes);
519 	return DM_MAPIO_REMAPPED;
520 }
521 
522 static void multipath_release_clone(struct request *clone)
523 {
524 	blk_put_request(clone);
525 }
526 
527 /*
528  * Map cloned bios (bio-based multipath)
529  */
530 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
531 {
532 	size_t nr_bytes = bio->bi_iter.bi_size;
533 	struct pgpath *pgpath;
534 	unsigned long flags;
535 	bool queue_io;
536 
537 	/* Do we need to select a new pgpath? */
538 	pgpath = lockless_dereference(m->current_pgpath);
539 	queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
540 	if (!pgpath || !queue_io)
541 		pgpath = choose_pgpath(m, nr_bytes);
542 
543 	if ((pgpath && queue_io) ||
544 	    (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
545 		/* Queue for the daemon to resubmit */
546 		spin_lock_irqsave(&m->lock, flags);
547 		bio_list_add(&m->queued_bios, bio);
548 		spin_unlock_irqrestore(&m->lock, flags);
549 		/* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
550 		if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
551 			pg_init_all_paths(m);
552 		else if (!queue_io)
553 			queue_work(kmultipathd, &m->process_queued_bios);
554 		return DM_MAPIO_SUBMITTED;
555 	}
556 
557 	if (!pgpath) {
558 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
559 			return DM_MAPIO_REQUEUE;
560 		dm_report_EIO(m);
561 		return DM_MAPIO_KILL;
562 	}
563 
564 	mpio->pgpath = pgpath;
565 	mpio->nr_bytes = nr_bytes;
566 
567 	bio->bi_status = 0;
568 	bio_set_dev(bio, pgpath->path.dev->bdev);
569 	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
570 
571 	if (pgpath->pg->ps.type->start_io)
572 		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
573 					      &pgpath->path,
574 					      nr_bytes);
575 	return DM_MAPIO_REMAPPED;
576 }
577 
578 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
579 {
580 	struct multipath *m = ti->private;
581 	struct dm_mpath_io *mpio = NULL;
582 
583 	multipath_init_per_bio_data(bio, &mpio, NULL);
584 
585 	return __multipath_map_bio(m, bio, mpio);
586 }
587 
588 static void process_queued_io_list(struct multipath *m)
589 {
590 	if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
591 		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
592 	else if (m->queue_mode == DM_TYPE_BIO_BASED)
593 		queue_work(kmultipathd, &m->process_queued_bios);
594 }
595 
596 static void process_queued_bios(struct work_struct *work)
597 {
598 	int r;
599 	unsigned long flags;
600 	struct bio *bio;
601 	struct bio_list bios;
602 	struct blk_plug plug;
603 	struct multipath *m =
604 		container_of(work, struct multipath, process_queued_bios);
605 
606 	bio_list_init(&bios);
607 
608 	spin_lock_irqsave(&m->lock, flags);
609 
610 	if (bio_list_empty(&m->queued_bios)) {
611 		spin_unlock_irqrestore(&m->lock, flags);
612 		return;
613 	}
614 
615 	bio_list_merge(&bios, &m->queued_bios);
616 	bio_list_init(&m->queued_bios);
617 
618 	spin_unlock_irqrestore(&m->lock, flags);
619 
620 	blk_start_plug(&plug);
621 	while ((bio = bio_list_pop(&bios))) {
622 		r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
623 		switch (r) {
624 		case DM_MAPIO_KILL:
625 			bio->bi_status = BLK_STS_IOERR;
626 			bio_endio(bio);
627 			break;
628 		case DM_MAPIO_REQUEUE:
629 			bio->bi_status = BLK_STS_DM_REQUEUE;
630 			bio_endio(bio);
631 			break;
632 		case DM_MAPIO_REMAPPED:
633 			generic_make_request(bio);
634 			break;
635 		case 0:
636 			break;
637 		default:
638 			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
639 		}
640 	}
641 	blk_finish_plug(&plug);
642 }
643 
644 static void assign_bit(bool value, long nr, unsigned long *addr)
645 {
646 	if (value)
647 		set_bit(nr, addr);
648 	else
649 		clear_bit(nr, addr);
650 }
651 
652 /*
653  * If we run out of usable paths, should we queue I/O or error it?
654  */
655 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
656 			    bool save_old_value)
657 {
658 	unsigned long flags;
659 
660 	spin_lock_irqsave(&m->lock, flags);
661 	assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
662 		   (!save_old_value && queue_if_no_path),
663 		   MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
664 	assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
665 		   MPATHF_QUEUE_IF_NO_PATH, &m->flags);
666 	spin_unlock_irqrestore(&m->lock, flags);
667 
668 	if (!queue_if_no_path) {
669 		dm_table_run_md_queue_async(m->ti->table);
670 		process_queued_io_list(m);
671 	}
672 
673 	return 0;
674 }
675 
676 /*
677  * An event is triggered whenever a path is taken out of use.
678  * Includes path failure and PG bypass.
679  */
680 static void trigger_event(struct work_struct *work)
681 {
682 	struct multipath *m =
683 		container_of(work, struct multipath, trigger_event);
684 
685 	dm_table_event(m->ti->table);
686 }
687 
688 /*-----------------------------------------------------------------
689  * Constructor/argument parsing:
690  * <#multipath feature args> [<arg>]*
691  * <#hw_handler args> [hw_handler [<arg>]*]
692  * <#priority groups>
693  * <initial priority group>
694  *     [<selector> <#selector args> [<arg>]*
695  *      <#paths> <#per-path selector args>
696  *         [<path> [<arg>]* ]+ ]+
697  *---------------------------------------------------------------*/
698 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
699 			       struct dm_target *ti)
700 {
701 	int r;
702 	struct path_selector_type *pst;
703 	unsigned ps_argc;
704 
705 	static const struct dm_arg _args[] = {
706 		{0, 1024, "invalid number of path selector args"},
707 	};
708 
709 	pst = dm_get_path_selector(dm_shift_arg(as));
710 	if (!pst) {
711 		ti->error = "unknown path selector type";
712 		return -EINVAL;
713 	}
714 
715 	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
716 	if (r) {
717 		dm_put_path_selector(pst);
718 		return -EINVAL;
719 	}
720 
721 	r = pst->create(&pg->ps, ps_argc, as->argv);
722 	if (r) {
723 		dm_put_path_selector(pst);
724 		ti->error = "path selector constructor failed";
725 		return r;
726 	}
727 
728 	pg->ps.type = pst;
729 	dm_consume_args(as, ps_argc);
730 
731 	return 0;
732 }
733 
734 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
735 			       struct dm_target *ti)
736 {
737 	int r;
738 	struct pgpath *p;
739 	struct multipath *m = ti->private;
740 	struct request_queue *q = NULL;
741 	const char *attached_handler_name;
742 
743 	/* we need at least a path arg */
744 	if (as->argc < 1) {
745 		ti->error = "no device given";
746 		return ERR_PTR(-EINVAL);
747 	}
748 
749 	p = alloc_pgpath();
750 	if (!p)
751 		return ERR_PTR(-ENOMEM);
752 
753 	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
754 			  &p->path.dev);
755 	if (r) {
756 		ti->error = "error getting device";
757 		goto bad;
758 	}
759 
760 	if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
761 		q = bdev_get_queue(p->path.dev->bdev);
762 
763 	if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
764 retain:
765 		attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
766 		if (attached_handler_name) {
767 			/*
768 			 * Clear any hw_handler_params associated with a
769 			 * handler that isn't already attached.
770 			 */
771 			if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
772 				kfree(m->hw_handler_params);
773 				m->hw_handler_params = NULL;
774 			}
775 
776 			/*
777 			 * Reset hw_handler_name to match the attached handler
778 			 *
779 			 * NB. This modifies the table line to show the actual
780 			 * handler instead of the original table passed in.
781 			 */
782 			kfree(m->hw_handler_name);
783 			m->hw_handler_name = attached_handler_name;
784 		}
785 	}
786 
787 	if (m->hw_handler_name) {
788 		r = scsi_dh_attach(q, m->hw_handler_name);
789 		if (r == -EBUSY) {
790 			char b[BDEVNAME_SIZE];
791 
792 			printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
793 				bdevname(p->path.dev->bdev, b));
794 			goto retain;
795 		}
796 		if (r < 0) {
797 			ti->error = "error attaching hardware handler";
798 			dm_put_device(ti, p->path.dev);
799 			goto bad;
800 		}
801 
802 		if (m->hw_handler_params) {
803 			r = scsi_dh_set_params(q, m->hw_handler_params);
804 			if (r < 0) {
805 				ti->error = "unable to set hardware "
806 							"handler parameters";
807 				dm_put_device(ti, p->path.dev);
808 				goto bad;
809 			}
810 		}
811 	}
812 
813 	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
814 	if (r) {
815 		dm_put_device(ti, p->path.dev);
816 		goto bad;
817 	}
818 
819 	return p;
820 
821  bad:
822 	free_pgpath(p);
823 	return ERR_PTR(r);
824 }
825 
826 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
827 						   struct multipath *m)
828 {
829 	static const struct dm_arg _args[] = {
830 		{1, 1024, "invalid number of paths"},
831 		{0, 1024, "invalid number of selector args"}
832 	};
833 
834 	int r;
835 	unsigned i, nr_selector_args, nr_args;
836 	struct priority_group *pg;
837 	struct dm_target *ti = m->ti;
838 
839 	if (as->argc < 2) {
840 		as->argc = 0;
841 		ti->error = "not enough priority group arguments";
842 		return ERR_PTR(-EINVAL);
843 	}
844 
845 	pg = alloc_priority_group();
846 	if (!pg) {
847 		ti->error = "couldn't allocate priority group";
848 		return ERR_PTR(-ENOMEM);
849 	}
850 	pg->m = m;
851 
852 	r = parse_path_selector(as, pg, ti);
853 	if (r)
854 		goto bad;
855 
856 	/*
857 	 * read the paths
858 	 */
859 	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
860 	if (r)
861 		goto bad;
862 
863 	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
864 	if (r)
865 		goto bad;
866 
867 	nr_args = 1 + nr_selector_args;
868 	for (i = 0; i < pg->nr_pgpaths; i++) {
869 		struct pgpath *pgpath;
870 		struct dm_arg_set path_args;
871 
872 		if (as->argc < nr_args) {
873 			ti->error = "not enough path parameters";
874 			r = -EINVAL;
875 			goto bad;
876 		}
877 
878 		path_args.argc = nr_args;
879 		path_args.argv = as->argv;
880 
881 		pgpath = parse_path(&path_args, &pg->ps, ti);
882 		if (IS_ERR(pgpath)) {
883 			r = PTR_ERR(pgpath);
884 			goto bad;
885 		}
886 
887 		pgpath->pg = pg;
888 		list_add_tail(&pgpath->list, &pg->pgpaths);
889 		dm_consume_args(as, nr_args);
890 	}
891 
892 	return pg;
893 
894  bad:
895 	free_priority_group(pg, ti);
896 	return ERR_PTR(r);
897 }
898 
899 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
900 {
901 	unsigned hw_argc;
902 	int ret;
903 	struct dm_target *ti = m->ti;
904 
905 	static const struct dm_arg _args[] = {
906 		{0, 1024, "invalid number of hardware handler args"},
907 	};
908 
909 	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
910 		return -EINVAL;
911 
912 	if (!hw_argc)
913 		return 0;
914 
915 	if (m->queue_mode == DM_TYPE_BIO_BASED) {
916 		dm_consume_args(as, hw_argc);
917 		DMERR("bio-based multipath doesn't allow hardware handler args");
918 		return 0;
919 	}
920 
921 	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
922 	if (!m->hw_handler_name)
923 		return -EINVAL;
924 
925 	if (hw_argc > 1) {
926 		char *p;
927 		int i, j, len = 4;
928 
929 		for (i = 0; i <= hw_argc - 2; i++)
930 			len += strlen(as->argv[i]) + 1;
931 		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
932 		if (!p) {
933 			ti->error = "memory allocation failed";
934 			ret = -ENOMEM;
935 			goto fail;
936 		}
937 		j = sprintf(p, "%d", hw_argc - 1);
938 		for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
939 			j = sprintf(p, "%s", as->argv[i]);
940 	}
941 	dm_consume_args(as, hw_argc - 1);
942 
943 	return 0;
944 fail:
945 	kfree(m->hw_handler_name);
946 	m->hw_handler_name = NULL;
947 	return ret;
948 }
949 
950 static int parse_features(struct dm_arg_set *as, struct multipath *m)
951 {
952 	int r;
953 	unsigned argc;
954 	struct dm_target *ti = m->ti;
955 	const char *arg_name;
956 
957 	static const struct dm_arg _args[] = {
958 		{0, 8, "invalid number of feature args"},
959 		{1, 50, "pg_init_retries must be between 1 and 50"},
960 		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
961 	};
962 
963 	r = dm_read_arg_group(_args, as, &argc, &ti->error);
964 	if (r)
965 		return -EINVAL;
966 
967 	if (!argc)
968 		return 0;
969 
970 	do {
971 		arg_name = dm_shift_arg(as);
972 		argc--;
973 
974 		if (!strcasecmp(arg_name, "queue_if_no_path")) {
975 			r = queue_if_no_path(m, true, false);
976 			continue;
977 		}
978 
979 		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
980 			set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
981 			continue;
982 		}
983 
984 		if (!strcasecmp(arg_name, "pg_init_retries") &&
985 		    (argc >= 1)) {
986 			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
987 			argc--;
988 			continue;
989 		}
990 
991 		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
992 		    (argc >= 1)) {
993 			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
994 			argc--;
995 			continue;
996 		}
997 
998 		if (!strcasecmp(arg_name, "queue_mode") &&
999 		    (argc >= 1)) {
1000 			const char *queue_mode_name = dm_shift_arg(as);
1001 
1002 			if (!strcasecmp(queue_mode_name, "bio"))
1003 				m->queue_mode = DM_TYPE_BIO_BASED;
1004 			else if (!strcasecmp(queue_mode_name, "rq"))
1005 				m->queue_mode = DM_TYPE_REQUEST_BASED;
1006 			else if (!strcasecmp(queue_mode_name, "mq"))
1007 				m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1008 			else {
1009 				ti->error = "Unknown 'queue_mode' requested";
1010 				r = -EINVAL;
1011 			}
1012 			argc--;
1013 			continue;
1014 		}
1015 
1016 		ti->error = "Unrecognised multipath feature request";
1017 		r = -EINVAL;
1018 	} while (argc && !r);
1019 
1020 	return r;
1021 }
1022 
1023 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1024 {
1025 	/* target arguments */
1026 	static const struct dm_arg _args[] = {
1027 		{0, 1024, "invalid number of priority groups"},
1028 		{0, 1024, "invalid initial priority group number"},
1029 	};
1030 
1031 	int r;
1032 	struct multipath *m;
1033 	struct dm_arg_set as;
1034 	unsigned pg_count = 0;
1035 	unsigned next_pg_num;
1036 
1037 	as.argc = argc;
1038 	as.argv = argv;
1039 
1040 	m = alloc_multipath(ti);
1041 	if (!m) {
1042 		ti->error = "can't allocate multipath";
1043 		return -EINVAL;
1044 	}
1045 
1046 	r = parse_features(&as, m);
1047 	if (r)
1048 		goto bad;
1049 
1050 	r = alloc_multipath_stage2(ti, m);
1051 	if (r)
1052 		goto bad;
1053 
1054 	r = parse_hw_handler(&as, m);
1055 	if (r)
1056 		goto bad;
1057 
1058 	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1059 	if (r)
1060 		goto bad;
1061 
1062 	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1063 	if (r)
1064 		goto bad;
1065 
1066 	if ((!m->nr_priority_groups && next_pg_num) ||
1067 	    (m->nr_priority_groups && !next_pg_num)) {
1068 		ti->error = "invalid initial priority group";
1069 		r = -EINVAL;
1070 		goto bad;
1071 	}
1072 
1073 	/* parse the priority groups */
1074 	while (as.argc) {
1075 		struct priority_group *pg;
1076 		unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1077 
1078 		pg = parse_priority_group(&as, m);
1079 		if (IS_ERR(pg)) {
1080 			r = PTR_ERR(pg);
1081 			goto bad;
1082 		}
1083 
1084 		nr_valid_paths += pg->nr_pgpaths;
1085 		atomic_set(&m->nr_valid_paths, nr_valid_paths);
1086 
1087 		list_add_tail(&pg->list, &m->priority_groups);
1088 		pg_count++;
1089 		pg->pg_num = pg_count;
1090 		if (!--next_pg_num)
1091 			m->next_pg = pg;
1092 	}
1093 
1094 	if (pg_count != m->nr_priority_groups) {
1095 		ti->error = "priority group count mismatch";
1096 		r = -EINVAL;
1097 		goto bad;
1098 	}
1099 
1100 	ti->num_flush_bios = 1;
1101 	ti->num_discard_bios = 1;
1102 	ti->num_write_same_bios = 1;
1103 	ti->num_write_zeroes_bios = 1;
1104 	if (m->queue_mode == DM_TYPE_BIO_BASED)
1105 		ti->per_io_data_size = multipath_per_bio_data_size();
1106 	else
1107 		ti->per_io_data_size = sizeof(struct dm_mpath_io);
1108 
1109 	return 0;
1110 
1111  bad:
1112 	free_multipath(m);
1113 	return r;
1114 }
1115 
1116 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1117 {
1118 	DEFINE_WAIT(wait);
1119 
1120 	while (1) {
1121 		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1122 
1123 		if (!atomic_read(&m->pg_init_in_progress))
1124 			break;
1125 
1126 		io_schedule();
1127 	}
1128 	finish_wait(&m->pg_init_wait, &wait);
1129 }
1130 
1131 static void flush_multipath_work(struct multipath *m)
1132 {
1133 	set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1134 	smp_mb__after_atomic();
1135 
1136 	flush_workqueue(kmpath_handlerd);
1137 	multipath_wait_for_pg_init_completion(m);
1138 	flush_workqueue(kmultipathd);
1139 	flush_work(&m->trigger_event);
1140 
1141 	clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1142 	smp_mb__after_atomic();
1143 }
1144 
1145 static void multipath_dtr(struct dm_target *ti)
1146 {
1147 	struct multipath *m = ti->private;
1148 
1149 	flush_multipath_work(m);
1150 	free_multipath(m);
1151 }
1152 
1153 /*
1154  * Take a path out of use.
1155  */
1156 static int fail_path(struct pgpath *pgpath)
1157 {
1158 	unsigned long flags;
1159 	struct multipath *m = pgpath->pg->m;
1160 
1161 	spin_lock_irqsave(&m->lock, flags);
1162 
1163 	if (!pgpath->is_active)
1164 		goto out;
1165 
1166 	DMWARN("Failing path %s.", pgpath->path.dev->name);
1167 
1168 	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1169 	pgpath->is_active = false;
1170 	pgpath->fail_count++;
1171 
1172 	atomic_dec(&m->nr_valid_paths);
1173 
1174 	if (pgpath == m->current_pgpath)
1175 		m->current_pgpath = NULL;
1176 
1177 	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1178 		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1179 
1180 	schedule_work(&m->trigger_event);
1181 
1182 out:
1183 	spin_unlock_irqrestore(&m->lock, flags);
1184 
1185 	return 0;
1186 }
1187 
1188 /*
1189  * Reinstate a previously-failed path
1190  */
1191 static int reinstate_path(struct pgpath *pgpath)
1192 {
1193 	int r = 0, run_queue = 0;
1194 	unsigned long flags;
1195 	struct multipath *m = pgpath->pg->m;
1196 	unsigned nr_valid_paths;
1197 
1198 	spin_lock_irqsave(&m->lock, flags);
1199 
1200 	if (pgpath->is_active)
1201 		goto out;
1202 
1203 	DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1204 
1205 	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1206 	if (r)
1207 		goto out;
1208 
1209 	pgpath->is_active = true;
1210 
1211 	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1212 	if (nr_valid_paths == 1) {
1213 		m->current_pgpath = NULL;
1214 		run_queue = 1;
1215 	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1216 		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1217 			atomic_inc(&m->pg_init_in_progress);
1218 	}
1219 
1220 	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1221 		       pgpath->path.dev->name, nr_valid_paths);
1222 
1223 	schedule_work(&m->trigger_event);
1224 
1225 out:
1226 	spin_unlock_irqrestore(&m->lock, flags);
1227 	if (run_queue) {
1228 		dm_table_run_md_queue_async(m->ti->table);
1229 		process_queued_io_list(m);
1230 	}
1231 
1232 	return r;
1233 }
1234 
1235 /*
1236  * Fail or reinstate all paths that match the provided struct dm_dev.
1237  */
1238 static int action_dev(struct multipath *m, struct dm_dev *dev,
1239 		      action_fn action)
1240 {
1241 	int r = -EINVAL;
1242 	struct pgpath *pgpath;
1243 	struct priority_group *pg;
1244 
1245 	list_for_each_entry(pg, &m->priority_groups, list) {
1246 		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1247 			if (pgpath->path.dev == dev)
1248 				r = action(pgpath);
1249 		}
1250 	}
1251 
1252 	return r;
1253 }
1254 
1255 /*
1256  * Temporarily try to avoid having to use the specified PG
1257  */
1258 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1259 		      bool bypassed)
1260 {
1261 	unsigned long flags;
1262 
1263 	spin_lock_irqsave(&m->lock, flags);
1264 
1265 	pg->bypassed = bypassed;
1266 	m->current_pgpath = NULL;
1267 	m->current_pg = NULL;
1268 
1269 	spin_unlock_irqrestore(&m->lock, flags);
1270 
1271 	schedule_work(&m->trigger_event);
1272 }
1273 
1274 /*
1275  * Switch to using the specified PG from the next I/O that gets mapped
1276  */
1277 static int switch_pg_num(struct multipath *m, const char *pgstr)
1278 {
1279 	struct priority_group *pg;
1280 	unsigned pgnum;
1281 	unsigned long flags;
1282 	char dummy;
1283 
1284 	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1285 	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1286 		DMWARN("invalid PG number supplied to switch_pg_num");
1287 		return -EINVAL;
1288 	}
1289 
1290 	spin_lock_irqsave(&m->lock, flags);
1291 	list_for_each_entry(pg, &m->priority_groups, list) {
1292 		pg->bypassed = false;
1293 		if (--pgnum)
1294 			continue;
1295 
1296 		m->current_pgpath = NULL;
1297 		m->current_pg = NULL;
1298 		m->next_pg = pg;
1299 	}
1300 	spin_unlock_irqrestore(&m->lock, flags);
1301 
1302 	schedule_work(&m->trigger_event);
1303 	return 0;
1304 }
1305 
1306 /*
1307  * Set/clear bypassed status of a PG.
1308  * PGs are numbered upwards from 1 in the order they were declared.
1309  */
1310 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1311 {
1312 	struct priority_group *pg;
1313 	unsigned pgnum;
1314 	char dummy;
1315 
1316 	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1317 	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1318 		DMWARN("invalid PG number supplied to bypass_pg");
1319 		return -EINVAL;
1320 	}
1321 
1322 	list_for_each_entry(pg, &m->priority_groups, list) {
1323 		if (!--pgnum)
1324 			break;
1325 	}
1326 
1327 	bypass_pg(m, pg, bypassed);
1328 	return 0;
1329 }
1330 
1331 /*
1332  * Should we retry pg_init immediately?
1333  */
1334 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1335 {
1336 	unsigned long flags;
1337 	bool limit_reached = false;
1338 
1339 	spin_lock_irqsave(&m->lock, flags);
1340 
1341 	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1342 	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1343 		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1344 	else
1345 		limit_reached = true;
1346 
1347 	spin_unlock_irqrestore(&m->lock, flags);
1348 
1349 	return limit_reached;
1350 }
1351 
1352 static void pg_init_done(void *data, int errors)
1353 {
1354 	struct pgpath *pgpath = data;
1355 	struct priority_group *pg = pgpath->pg;
1356 	struct multipath *m = pg->m;
1357 	unsigned long flags;
1358 	bool delay_retry = false;
1359 
1360 	/* device or driver problems */
1361 	switch (errors) {
1362 	case SCSI_DH_OK:
1363 		break;
1364 	case SCSI_DH_NOSYS:
1365 		if (!m->hw_handler_name) {
1366 			errors = 0;
1367 			break;
1368 		}
1369 		DMERR("Could not failover the device: Handler scsi_dh_%s "
1370 		      "Error %d.", m->hw_handler_name, errors);
1371 		/*
1372 		 * Fail path for now, so we do not ping pong
1373 		 */
1374 		fail_path(pgpath);
1375 		break;
1376 	case SCSI_DH_DEV_TEMP_BUSY:
1377 		/*
1378 		 * Probably doing something like FW upgrade on the
1379 		 * controller so try the other pg.
1380 		 */
1381 		bypass_pg(m, pg, true);
1382 		break;
1383 	case SCSI_DH_RETRY:
1384 		/* Wait before retrying. */
1385 		delay_retry = 1;
1386 		/* fall through */
1387 	case SCSI_DH_IMM_RETRY:
1388 	case SCSI_DH_RES_TEMP_UNAVAIL:
1389 		if (pg_init_limit_reached(m, pgpath))
1390 			fail_path(pgpath);
1391 		errors = 0;
1392 		break;
1393 	case SCSI_DH_DEV_OFFLINED:
1394 	default:
1395 		/*
1396 		 * We probably do not want to fail the path for a device
1397 		 * error, but this is what the old dm did. In future
1398 		 * patches we can do more advanced handling.
1399 		 */
1400 		fail_path(pgpath);
1401 	}
1402 
1403 	spin_lock_irqsave(&m->lock, flags);
1404 	if (errors) {
1405 		if (pgpath == m->current_pgpath) {
1406 			DMERR("Could not failover device. Error %d.", errors);
1407 			m->current_pgpath = NULL;
1408 			m->current_pg = NULL;
1409 		}
1410 	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1411 		pg->bypassed = false;
1412 
1413 	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1414 		/* Activations of other paths are still on going */
1415 		goto out;
1416 
1417 	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1418 		if (delay_retry)
1419 			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1420 		else
1421 			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1422 
1423 		if (__pg_init_all_paths(m))
1424 			goto out;
1425 	}
1426 	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1427 
1428 	process_queued_io_list(m);
1429 
1430 	/*
1431 	 * Wake up any thread waiting to suspend.
1432 	 */
1433 	wake_up(&m->pg_init_wait);
1434 
1435 out:
1436 	spin_unlock_irqrestore(&m->lock, flags);
1437 }
1438 
1439 static void activate_or_offline_path(struct pgpath *pgpath)
1440 {
1441 	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1442 
1443 	if (pgpath->is_active && !blk_queue_dying(q))
1444 		scsi_dh_activate(q, pg_init_done, pgpath);
1445 	else
1446 		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1447 }
1448 
1449 static void activate_path_work(struct work_struct *work)
1450 {
1451 	struct pgpath *pgpath =
1452 		container_of(work, struct pgpath, activate_path.work);
1453 
1454 	activate_or_offline_path(pgpath);
1455 }
1456 
1457 static int noretry_error(blk_status_t error)
1458 {
1459 	switch (error) {
1460 	case BLK_STS_NOTSUPP:
1461 	case BLK_STS_NOSPC:
1462 	case BLK_STS_TARGET:
1463 	case BLK_STS_NEXUS:
1464 	case BLK_STS_MEDIUM:
1465 		return 1;
1466 	}
1467 
1468 	/* Anything else could be a path failure, so should be retried */
1469 	return 0;
1470 }
1471 
1472 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1473 			    blk_status_t error, union map_info *map_context)
1474 {
1475 	struct dm_mpath_io *mpio = get_mpio(map_context);
1476 	struct pgpath *pgpath = mpio->pgpath;
1477 	int r = DM_ENDIO_DONE;
1478 
1479 	/*
1480 	 * We don't queue any clone request inside the multipath target
1481 	 * during end I/O handling, since those clone requests don't have
1482 	 * bio clones.  If we queue them inside the multipath target,
1483 	 * we need to make bio clones, that requires memory allocation.
1484 	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1485 	 *  don't have bio clones.)
1486 	 * Instead of queueing the clone request here, we queue the original
1487 	 * request into dm core, which will remake a clone request and
1488 	 * clone bios for it and resubmit it later.
1489 	 */
1490 	if (error && !noretry_error(error)) {
1491 		struct multipath *m = ti->private;
1492 
1493 		r = DM_ENDIO_REQUEUE;
1494 
1495 		if (pgpath)
1496 			fail_path(pgpath);
1497 
1498 		if (atomic_read(&m->nr_valid_paths) == 0 &&
1499 		    !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1500 			if (error == BLK_STS_IOERR)
1501 				dm_report_EIO(m);
1502 			/* complete with the original error */
1503 			r = DM_ENDIO_DONE;
1504 		}
1505 	}
1506 
1507 	if (pgpath) {
1508 		struct path_selector *ps = &pgpath->pg->ps;
1509 
1510 		if (ps->type->end_io)
1511 			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1512 	}
1513 
1514 	return r;
1515 }
1516 
1517 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1518 		blk_status_t *error)
1519 {
1520 	struct multipath *m = ti->private;
1521 	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1522 	struct pgpath *pgpath = mpio->pgpath;
1523 	unsigned long flags;
1524 	int r = DM_ENDIO_DONE;
1525 
1526 	if (!*error || noretry_error(*error))
1527 		goto done;
1528 
1529 	if (pgpath)
1530 		fail_path(pgpath);
1531 
1532 	if (atomic_read(&m->nr_valid_paths) == 0 &&
1533 	    !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1534 		dm_report_EIO(m);
1535 		*error = BLK_STS_IOERR;
1536 		goto done;
1537 	}
1538 
1539 	/* Queue for the daemon to resubmit */
1540 	dm_bio_restore(get_bio_details_from_bio(clone), clone);
1541 
1542 	spin_lock_irqsave(&m->lock, flags);
1543 	bio_list_add(&m->queued_bios, clone);
1544 	spin_unlock_irqrestore(&m->lock, flags);
1545 	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1546 		queue_work(kmultipathd, &m->process_queued_bios);
1547 
1548 	r = DM_ENDIO_INCOMPLETE;
1549 done:
1550 	if (pgpath) {
1551 		struct path_selector *ps = &pgpath->pg->ps;
1552 
1553 		if (ps->type->end_io)
1554 			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1555 	}
1556 
1557 	return r;
1558 }
1559 
1560 /*
1561  * Suspend can't complete until all the I/O is processed so if
1562  * the last path fails we must error any remaining I/O.
1563  * Note that if the freeze_bdev fails while suspending, the
1564  * queue_if_no_path state is lost - userspace should reset it.
1565  */
1566 static void multipath_presuspend(struct dm_target *ti)
1567 {
1568 	struct multipath *m = ti->private;
1569 
1570 	queue_if_no_path(m, false, true);
1571 }
1572 
1573 static void multipath_postsuspend(struct dm_target *ti)
1574 {
1575 	struct multipath *m = ti->private;
1576 
1577 	mutex_lock(&m->work_mutex);
1578 	flush_multipath_work(m);
1579 	mutex_unlock(&m->work_mutex);
1580 }
1581 
1582 /*
1583  * Restore the queue_if_no_path setting.
1584  */
1585 static void multipath_resume(struct dm_target *ti)
1586 {
1587 	struct multipath *m = ti->private;
1588 	unsigned long flags;
1589 
1590 	spin_lock_irqsave(&m->lock, flags);
1591 	assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1592 		   MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1593 	spin_unlock_irqrestore(&m->lock, flags);
1594 }
1595 
1596 /*
1597  * Info output has the following format:
1598  * num_multipath_feature_args [multipath_feature_args]*
1599  * num_handler_status_args [handler_status_args]*
1600  * num_groups init_group_number
1601  *            [A|D|E num_ps_status_args [ps_status_args]*
1602  *             num_paths num_selector_args
1603  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1604  *
1605  * Table output has the following format (identical to the constructor string):
1606  * num_feature_args [features_args]*
1607  * num_handler_args hw_handler [hw_handler_args]*
1608  * num_groups init_group_number
1609  *     [priority selector-name num_ps_args [ps_args]*
1610  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1611  */
1612 static void multipath_status(struct dm_target *ti, status_type_t type,
1613 			     unsigned status_flags, char *result, unsigned maxlen)
1614 {
1615 	int sz = 0;
1616 	unsigned long flags;
1617 	struct multipath *m = ti->private;
1618 	struct priority_group *pg;
1619 	struct pgpath *p;
1620 	unsigned pg_num;
1621 	char state;
1622 
1623 	spin_lock_irqsave(&m->lock, flags);
1624 
1625 	/* Features */
1626 	if (type == STATUSTYPE_INFO)
1627 		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1628 		       atomic_read(&m->pg_init_count));
1629 	else {
1630 		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1631 			      (m->pg_init_retries > 0) * 2 +
1632 			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1633 			      test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1634 			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1635 
1636 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1637 			DMEMIT("queue_if_no_path ");
1638 		if (m->pg_init_retries)
1639 			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1640 		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1641 			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1642 		if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1643 			DMEMIT("retain_attached_hw_handler ");
1644 		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1645 			switch(m->queue_mode) {
1646 			case DM_TYPE_BIO_BASED:
1647 				DMEMIT("queue_mode bio ");
1648 				break;
1649 			case DM_TYPE_MQ_REQUEST_BASED:
1650 				DMEMIT("queue_mode mq ");
1651 				break;
1652 			default:
1653 				WARN_ON_ONCE(true);
1654 				break;
1655 			}
1656 		}
1657 	}
1658 
1659 	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1660 		DMEMIT("0 ");
1661 	else
1662 		DMEMIT("1 %s ", m->hw_handler_name);
1663 
1664 	DMEMIT("%u ", m->nr_priority_groups);
1665 
1666 	if (m->next_pg)
1667 		pg_num = m->next_pg->pg_num;
1668 	else if (m->current_pg)
1669 		pg_num = m->current_pg->pg_num;
1670 	else
1671 		pg_num = (m->nr_priority_groups ? 1 : 0);
1672 
1673 	DMEMIT("%u ", pg_num);
1674 
1675 	switch (type) {
1676 	case STATUSTYPE_INFO:
1677 		list_for_each_entry(pg, &m->priority_groups, list) {
1678 			if (pg->bypassed)
1679 				state = 'D';	/* Disabled */
1680 			else if (pg == m->current_pg)
1681 				state = 'A';	/* Currently Active */
1682 			else
1683 				state = 'E';	/* Enabled */
1684 
1685 			DMEMIT("%c ", state);
1686 
1687 			if (pg->ps.type->status)
1688 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1689 							  result + sz,
1690 							  maxlen - sz);
1691 			else
1692 				DMEMIT("0 ");
1693 
1694 			DMEMIT("%u %u ", pg->nr_pgpaths,
1695 			       pg->ps.type->info_args);
1696 
1697 			list_for_each_entry(p, &pg->pgpaths, list) {
1698 				DMEMIT("%s %s %u ", p->path.dev->name,
1699 				       p->is_active ? "A" : "F",
1700 				       p->fail_count);
1701 				if (pg->ps.type->status)
1702 					sz += pg->ps.type->status(&pg->ps,
1703 					      &p->path, type, result + sz,
1704 					      maxlen - sz);
1705 			}
1706 		}
1707 		break;
1708 
1709 	case STATUSTYPE_TABLE:
1710 		list_for_each_entry(pg, &m->priority_groups, list) {
1711 			DMEMIT("%s ", pg->ps.type->name);
1712 
1713 			if (pg->ps.type->status)
1714 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1715 							  result + sz,
1716 							  maxlen - sz);
1717 			else
1718 				DMEMIT("0 ");
1719 
1720 			DMEMIT("%u %u ", pg->nr_pgpaths,
1721 			       pg->ps.type->table_args);
1722 
1723 			list_for_each_entry(p, &pg->pgpaths, list) {
1724 				DMEMIT("%s ", p->path.dev->name);
1725 				if (pg->ps.type->status)
1726 					sz += pg->ps.type->status(&pg->ps,
1727 					      &p->path, type, result + sz,
1728 					      maxlen - sz);
1729 			}
1730 		}
1731 		break;
1732 	}
1733 
1734 	spin_unlock_irqrestore(&m->lock, flags);
1735 }
1736 
1737 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1738 {
1739 	int r = -EINVAL;
1740 	struct dm_dev *dev;
1741 	struct multipath *m = ti->private;
1742 	action_fn action;
1743 
1744 	mutex_lock(&m->work_mutex);
1745 
1746 	if (dm_suspended(ti)) {
1747 		r = -EBUSY;
1748 		goto out;
1749 	}
1750 
1751 	if (argc == 1) {
1752 		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1753 			r = queue_if_no_path(m, true, false);
1754 			goto out;
1755 		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1756 			r = queue_if_no_path(m, false, false);
1757 			goto out;
1758 		}
1759 	}
1760 
1761 	if (argc != 2) {
1762 		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1763 		goto out;
1764 	}
1765 
1766 	if (!strcasecmp(argv[0], "disable_group")) {
1767 		r = bypass_pg_num(m, argv[1], true);
1768 		goto out;
1769 	} else if (!strcasecmp(argv[0], "enable_group")) {
1770 		r = bypass_pg_num(m, argv[1], false);
1771 		goto out;
1772 	} else if (!strcasecmp(argv[0], "switch_group")) {
1773 		r = switch_pg_num(m, argv[1]);
1774 		goto out;
1775 	} else if (!strcasecmp(argv[0], "reinstate_path"))
1776 		action = reinstate_path;
1777 	else if (!strcasecmp(argv[0], "fail_path"))
1778 		action = fail_path;
1779 	else {
1780 		DMWARN("Unrecognised multipath message received: %s", argv[0]);
1781 		goto out;
1782 	}
1783 
1784 	r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1785 	if (r) {
1786 		DMWARN("message: error getting device %s",
1787 		       argv[1]);
1788 		goto out;
1789 	}
1790 
1791 	r = action_dev(m, dev, action);
1792 
1793 	dm_put_device(ti, dev);
1794 
1795 out:
1796 	mutex_unlock(&m->work_mutex);
1797 	return r;
1798 }
1799 
1800 static int multipath_prepare_ioctl(struct dm_target *ti,
1801 		struct block_device **bdev, fmode_t *mode)
1802 {
1803 	struct multipath *m = ti->private;
1804 	struct pgpath *current_pgpath;
1805 	int r;
1806 
1807 	current_pgpath = lockless_dereference(m->current_pgpath);
1808 	if (!current_pgpath)
1809 		current_pgpath = choose_pgpath(m, 0);
1810 
1811 	if (current_pgpath) {
1812 		if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1813 			*bdev = current_pgpath->path.dev->bdev;
1814 			*mode = current_pgpath->path.dev->mode;
1815 			r = 0;
1816 		} else {
1817 			/* pg_init has not started or completed */
1818 			r = -ENOTCONN;
1819 		}
1820 	} else {
1821 		/* No path is available */
1822 		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1823 			r = -ENOTCONN;
1824 		else
1825 			r = -EIO;
1826 	}
1827 
1828 	if (r == -ENOTCONN) {
1829 		if (!lockless_dereference(m->current_pg)) {
1830 			/* Path status changed, redo selection */
1831 			(void) choose_pgpath(m, 0);
1832 		}
1833 		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1834 			pg_init_all_paths(m);
1835 		dm_table_run_md_queue_async(m->ti->table);
1836 		process_queued_io_list(m);
1837 	}
1838 
1839 	/*
1840 	 * Only pass ioctls through if the device sizes match exactly.
1841 	 */
1842 	if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1843 		return 1;
1844 	return r;
1845 }
1846 
1847 static int multipath_iterate_devices(struct dm_target *ti,
1848 				     iterate_devices_callout_fn fn, void *data)
1849 {
1850 	struct multipath *m = ti->private;
1851 	struct priority_group *pg;
1852 	struct pgpath *p;
1853 	int ret = 0;
1854 
1855 	list_for_each_entry(pg, &m->priority_groups, list) {
1856 		list_for_each_entry(p, &pg->pgpaths, list) {
1857 			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1858 			if (ret)
1859 				goto out;
1860 		}
1861 	}
1862 
1863 out:
1864 	return ret;
1865 }
1866 
1867 static int pgpath_busy(struct pgpath *pgpath)
1868 {
1869 	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1870 
1871 	return blk_lld_busy(q);
1872 }
1873 
1874 /*
1875  * We return "busy", only when we can map I/Os but underlying devices
1876  * are busy (so even if we map I/Os now, the I/Os will wait on
1877  * the underlying queue).
1878  * In other words, if we want to kill I/Os or queue them inside us
1879  * due to map unavailability, we don't return "busy".  Otherwise,
1880  * dm core won't give us the I/Os and we can't do what we want.
1881  */
1882 static int multipath_busy(struct dm_target *ti)
1883 {
1884 	bool busy = false, has_active = false;
1885 	struct multipath *m = ti->private;
1886 	struct priority_group *pg, *next_pg;
1887 	struct pgpath *pgpath;
1888 
1889 	/* pg_init in progress */
1890 	if (atomic_read(&m->pg_init_in_progress))
1891 		return true;
1892 
1893 	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1894 	if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1895 		return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1896 
1897 	/* Guess which priority_group will be used at next mapping time */
1898 	pg = lockless_dereference(m->current_pg);
1899 	next_pg = lockless_dereference(m->next_pg);
1900 	if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1901 		pg = next_pg;
1902 
1903 	if (!pg) {
1904 		/*
1905 		 * We don't know which pg will be used at next mapping time.
1906 		 * We don't call choose_pgpath() here to avoid to trigger
1907 		 * pg_init just by busy checking.
1908 		 * So we don't know whether underlying devices we will be using
1909 		 * at next mapping time are busy or not. Just try mapping.
1910 		 */
1911 		return busy;
1912 	}
1913 
1914 	/*
1915 	 * If there is one non-busy active path at least, the path selector
1916 	 * will be able to select it. So we consider such a pg as not busy.
1917 	 */
1918 	busy = true;
1919 	list_for_each_entry(pgpath, &pg->pgpaths, list) {
1920 		if (pgpath->is_active) {
1921 			has_active = true;
1922 			if (!pgpath_busy(pgpath)) {
1923 				busy = false;
1924 				break;
1925 			}
1926 		}
1927 	}
1928 
1929 	if (!has_active) {
1930 		/*
1931 		 * No active path in this pg, so this pg won't be used and
1932 		 * the current_pg will be changed at next mapping time.
1933 		 * We need to try mapping to determine it.
1934 		 */
1935 		busy = false;
1936 	}
1937 
1938 	return busy;
1939 }
1940 
1941 /*-----------------------------------------------------------------
1942  * Module setup
1943  *---------------------------------------------------------------*/
1944 static struct target_type multipath_target = {
1945 	.name = "multipath",
1946 	.version = {1, 12, 0},
1947 	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1948 	.module = THIS_MODULE,
1949 	.ctr = multipath_ctr,
1950 	.dtr = multipath_dtr,
1951 	.clone_and_map_rq = multipath_clone_and_map,
1952 	.release_clone_rq = multipath_release_clone,
1953 	.rq_end_io = multipath_end_io,
1954 	.map = multipath_map_bio,
1955 	.end_io = multipath_end_io_bio,
1956 	.presuspend = multipath_presuspend,
1957 	.postsuspend = multipath_postsuspend,
1958 	.resume = multipath_resume,
1959 	.status = multipath_status,
1960 	.message = multipath_message,
1961 	.prepare_ioctl = multipath_prepare_ioctl,
1962 	.iterate_devices = multipath_iterate_devices,
1963 	.busy = multipath_busy,
1964 };
1965 
1966 static int __init dm_multipath_init(void)
1967 {
1968 	int r;
1969 
1970 	r = dm_register_target(&multipath_target);
1971 	if (r < 0) {
1972 		DMERR("request-based register failed %d", r);
1973 		r = -EINVAL;
1974 		goto bad_register_target;
1975 	}
1976 
1977 	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1978 	if (!kmultipathd) {
1979 		DMERR("failed to create workqueue kmpathd");
1980 		r = -ENOMEM;
1981 		goto bad_alloc_kmultipathd;
1982 	}
1983 
1984 	/*
1985 	 * A separate workqueue is used to handle the device handlers
1986 	 * to avoid overloading existing workqueue. Overloading the
1987 	 * old workqueue would also create a bottleneck in the
1988 	 * path of the storage hardware device activation.
1989 	 */
1990 	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1991 						  WQ_MEM_RECLAIM);
1992 	if (!kmpath_handlerd) {
1993 		DMERR("failed to create workqueue kmpath_handlerd");
1994 		r = -ENOMEM;
1995 		goto bad_alloc_kmpath_handlerd;
1996 	}
1997 
1998 	return 0;
1999 
2000 bad_alloc_kmpath_handlerd:
2001 	destroy_workqueue(kmultipathd);
2002 bad_alloc_kmultipathd:
2003 	dm_unregister_target(&multipath_target);
2004 bad_register_target:
2005 	return r;
2006 }
2007 
2008 static void __exit dm_multipath_exit(void)
2009 {
2010 	destroy_workqueue(kmpath_handlerd);
2011 	destroy_workqueue(kmultipathd);
2012 
2013 	dm_unregister_target(&multipath_target);
2014 }
2015 
2016 module_init(dm_multipath_init);
2017 module_exit(dm_multipath_exit);
2018 
2019 MODULE_DESCRIPTION(DM_NAME " multipath target");
2020 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2021 MODULE_LICENSE("GPL");
2022