2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
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>
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
35 struct list_head list;
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
41 struct delayed_work activate_path;
43 bool is_active:1; /* Path status */
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
52 struct priority_group {
53 struct list_head list;
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
62 bool bypassed:1; /* Temporarily bypass this PG? */
65 /* Multipath context */
67 struct list_head list;
70 const char *hw_handler_name;
71 char *hw_handler_params;
75 unsigned nr_priority_groups;
76 struct list_head priority_groups;
78 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned long flags; /* Multipath state flags */
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 */
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 */
93 enum dm_queue_mode queue_mode;
95 struct mutex work_mutex;
96 struct work_struct trigger_event;
98 struct work_struct process_queued_bios;
99 struct bio_list queued_bios;
103 * Context information attached to each io we process.
106 struct pgpath *pgpath;
110 typedef int (*action_fn) (struct pgpath *pgpath);
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);
118 /*-----------------------------------------------
119 * Multipath state flags.
120 *-----------------------------------------------*/
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? */
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
134 static struct pgpath *alloc_pgpath(void)
136 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
139 pgpath->is_active = true;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
146 static void free_pgpath(struct pgpath *pgpath)
151 static struct priority_group *alloc_priority_group(void)
153 struct priority_group *pg;
155 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
158 INIT_LIST_HEAD(&pg->pgpaths);
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
165 struct pgpath *pgpath, *tmp;
167 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168 list_del(&pgpath->list);
169 dm_put_device(ti, pgpath->path.dev);
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
177 struct path_selector *ps = &pg->ps;
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
184 free_pgpaths(&pg->pgpaths, ti);
188 static struct multipath *alloc_multipath(struct dm_target *ti)
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
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);
205 m->queue_mode = DM_TYPE_NONE;
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
216 if (m->queue_mode == DM_TYPE_NONE) {
218 * Default to request-based.
220 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
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);
227 * bio-based doesn't support any direct scsi_dh management;
228 * it just discovers if a scsi_dh is attached.
230 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
233 dm_table_set_type(ti->table, m->queue_mode);
238 static void free_multipath(struct multipath *m)
240 struct priority_group *pg, *tmp;
242 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
244 free_priority_group(pg, m->ti);
247 kfree(m->hw_handler_name);
248 kfree(m->hw_handler_params);
252 static struct dm_mpath_io *get_mpio(union map_info *info)
257 static size_t multipath_per_bio_data_size(void)
259 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
264 return dm_per_bio_data(bio, multipath_per_bio_data_size());
267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
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;
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)
279 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
282 memset(mpio, 0, sizeof(*mpio));
283 memset(bio_details, 0, sizeof(*bio_details));
284 dm_bio_record(bio_details, bio);
289 *bio_details_p = bio_details;
292 /*-----------------------------------------------
294 *-----------------------------------------------*/
296 static int __pg_init_all_paths(struct multipath *m)
298 struct pgpath *pgpath;
299 unsigned long pg_init_delay = 0;
301 lockdep_assert_held(&m->lock);
303 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
306 atomic_inc(&m->pg_init_count);
307 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
309 /* Check here to reset pg_init_required */
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)
320 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
322 atomic_inc(&m->pg_init_in_progress);
324 return atomic_read(&m->pg_init_in_progress);
327 static int pg_init_all_paths(struct multipath *m)
332 spin_lock_irqsave(&m->lock, flags);
333 ret = __pg_init_all_paths(m);
334 spin_unlock_irqrestore(&m->lock, flags);
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
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);
348 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349 clear_bit(MPATHF_QUEUE_IO, &m->flags);
352 atomic_set(&m->pg_init_count, 0);
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356 struct priority_group *pg,
360 struct dm_path *path;
361 struct pgpath *pgpath;
363 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
365 return ERR_PTR(-ENXIO);
367 pgpath = path_to_pgpath(path);
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;
374 spin_unlock_irqrestore(&m->lock, flags);
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
383 struct priority_group *pg;
384 struct pgpath *pgpath;
385 unsigned bypassed = 1;
387 if (!atomic_read(&m->nr_valid_paths)) {
388 clear_bit(MPATHF_QUEUE_IO, &m->flags);
392 /* Were we instructed to switch PG? */
393 if (lockless_dereference(m->next_pg)) {
394 spin_lock_irqsave(&m->lock, flags);
397 spin_unlock_irqrestore(&m->lock, flags);
398 goto check_current_pg;
401 spin_unlock_irqrestore(&m->lock, flags);
402 pgpath = choose_path_in_pg(m, pg, nr_bytes);
403 if (!IS_ERR_OR_NULL(pgpath))
407 /* Don't change PG until it has no remaining paths */
409 pg = lockless_dereference(m->current_pg);
411 pgpath = choose_path_in_pg(m, pg, nr_bytes);
412 if (!IS_ERR_OR_NULL(pgpath))
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.
423 list_for_each_entry(pg, &m->priority_groups, list) {
424 if (pg->bypassed == !!bypassed)
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath)) {
429 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
433 } while (bypassed--);
436 spin_lock_irqsave(&m->lock, flags);
437 m->current_pgpath = NULL;
438 m->current_pg = NULL;
439 spin_unlock_irqrestore(&m->lock, flags);
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.
449 #define dm_report_EIO(m) \
451 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
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)); \
461 * Map cloned requests (request-based multipath)
463 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
464 union map_info *map_context,
465 struct request **__clone)
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;
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);
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;
492 memset(mpio, 0, sizeof(*mpio));
493 mpio->pgpath = pgpath;
494 mpio->nr_bytes = nr_bytes;
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);
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)" : "");
505 atomic_inc(&m->pg_init_in_progress);
506 activate_or_offline_path(pgpath);
507 return DM_MAPIO_REQUEUE;
509 return DM_MAPIO_DELAY_REQUEUE;
511 clone->bio = clone->biotail = NULL;
512 clone->rq_disk = bdev->bd_disk;
513 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
516 if (pgpath->pg->ps.type->start_io)
517 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
520 return DM_MAPIO_REMAPPED;
523 static void multipath_release_clone(struct request *clone)
525 blk_put_request(clone);
529 * Map cloned bios (bio-based multipath)
531 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
533 size_t nr_bytes = bio->bi_iter.bi_size;
534 struct pgpath *pgpath;
538 /* Do we need to select a new pgpath? */
539 pgpath = lockless_dereference(m->current_pgpath);
540 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
541 if (!pgpath || !queue_io)
542 pgpath = choose_pgpath(m, nr_bytes);
544 if ((pgpath && queue_io) ||
545 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
546 /* Queue for the daemon to resubmit */
547 spin_lock_irqsave(&m->lock, flags);
548 bio_list_add(&m->queued_bios, bio);
549 spin_unlock_irqrestore(&m->lock, flags);
550 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
551 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
552 pg_init_all_paths(m);
554 queue_work(kmultipathd, &m->process_queued_bios);
555 return DM_MAPIO_SUBMITTED;
559 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
560 return DM_MAPIO_REQUEUE;
565 mpio->pgpath = pgpath;
566 mpio->nr_bytes = nr_bytes;
569 bio->bi_bdev = pgpath->path.dev->bdev;
570 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
572 if (pgpath->pg->ps.type->start_io)
573 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
576 return DM_MAPIO_REMAPPED;
579 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
581 struct multipath *m = ti->private;
582 struct dm_mpath_io *mpio = NULL;
584 multipath_init_per_bio_data(bio, &mpio, NULL);
586 return __multipath_map_bio(m, bio, mpio);
589 static void process_queued_io_list(struct multipath *m)
591 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
592 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
593 else if (m->queue_mode == DM_TYPE_BIO_BASED)
594 queue_work(kmultipathd, &m->process_queued_bios);
597 static void process_queued_bios(struct work_struct *work)
602 struct bio_list bios;
603 struct blk_plug plug;
604 struct multipath *m =
605 container_of(work, struct multipath, process_queued_bios);
607 bio_list_init(&bios);
609 spin_lock_irqsave(&m->lock, flags);
611 if (bio_list_empty(&m->queued_bios)) {
612 spin_unlock_irqrestore(&m->lock, flags);
616 bio_list_merge(&bios, &m->queued_bios);
617 bio_list_init(&m->queued_bios);
619 spin_unlock_irqrestore(&m->lock, flags);
621 blk_start_plug(&plug);
622 while ((bio = bio_list_pop(&bios))) {
623 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
624 if (r < 0 || r == DM_MAPIO_REQUEUE) {
627 } else if (r == DM_MAPIO_REMAPPED)
628 generic_make_request(bio);
630 blk_finish_plug(&plug);
633 static void assign_bit(bool value, long nr, unsigned long *addr)
642 * If we run out of usable paths, should we queue I/O or error it?
644 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
649 spin_lock_irqsave(&m->lock, flags);
650 assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
651 (!save_old_value && queue_if_no_path),
652 MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
653 assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
654 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
655 spin_unlock_irqrestore(&m->lock, flags);
657 if (!queue_if_no_path) {
658 dm_table_run_md_queue_async(m->ti->table);
659 process_queued_io_list(m);
666 * An event is triggered whenever a path is taken out of use.
667 * Includes path failure and PG bypass.
669 static void trigger_event(struct work_struct *work)
671 struct multipath *m =
672 container_of(work, struct multipath, trigger_event);
674 dm_table_event(m->ti->table);
677 /*-----------------------------------------------------------------
678 * Constructor/argument parsing:
679 * <#multipath feature args> [<arg>]*
680 * <#hw_handler args> [hw_handler [<arg>]*]
682 * <initial priority group>
683 * [<selector> <#selector args> [<arg>]*
684 * <#paths> <#per-path selector args>
685 * [<path> [<arg>]* ]+ ]+
686 *---------------------------------------------------------------*/
687 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
688 struct dm_target *ti)
691 struct path_selector_type *pst;
694 static struct dm_arg _args[] = {
695 {0, 1024, "invalid number of path selector args"},
698 pst = dm_get_path_selector(dm_shift_arg(as));
700 ti->error = "unknown path selector type";
704 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
706 dm_put_path_selector(pst);
710 r = pst->create(&pg->ps, ps_argc, as->argv);
712 dm_put_path_selector(pst);
713 ti->error = "path selector constructor failed";
718 dm_consume_args(as, ps_argc);
723 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
724 struct dm_target *ti)
728 struct multipath *m = ti->private;
729 struct request_queue *q = NULL;
730 const char *attached_handler_name;
732 /* we need at least a path arg */
734 ti->error = "no device given";
735 return ERR_PTR(-EINVAL);
740 return ERR_PTR(-ENOMEM);
742 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
745 ti->error = "error getting device";
749 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
750 q = bdev_get_queue(p->path.dev->bdev);
752 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
754 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
755 if (attached_handler_name) {
757 * Clear any hw_handler_params associated with a
758 * handler that isn't already attached.
760 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
761 kfree(m->hw_handler_params);
762 m->hw_handler_params = NULL;
766 * Reset hw_handler_name to match the attached handler
768 * NB. This modifies the table line to show the actual
769 * handler instead of the original table passed in.
771 kfree(m->hw_handler_name);
772 m->hw_handler_name = attached_handler_name;
776 if (m->hw_handler_name) {
777 r = scsi_dh_attach(q, m->hw_handler_name);
779 char b[BDEVNAME_SIZE];
781 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
782 bdevname(p->path.dev->bdev, b));
786 ti->error = "error attaching hardware handler";
787 dm_put_device(ti, p->path.dev);
791 if (m->hw_handler_params) {
792 r = scsi_dh_set_params(q, m->hw_handler_params);
794 ti->error = "unable to set hardware "
795 "handler parameters";
796 dm_put_device(ti, p->path.dev);
802 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
804 dm_put_device(ti, p->path.dev);
815 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
818 static struct dm_arg _args[] = {
819 {1, 1024, "invalid number of paths"},
820 {0, 1024, "invalid number of selector args"}
824 unsigned i, nr_selector_args, nr_args;
825 struct priority_group *pg;
826 struct dm_target *ti = m->ti;
830 ti->error = "not enough priority group arguments";
831 return ERR_PTR(-EINVAL);
834 pg = alloc_priority_group();
836 ti->error = "couldn't allocate priority group";
837 return ERR_PTR(-ENOMEM);
841 r = parse_path_selector(as, pg, ti);
848 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
852 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
856 nr_args = 1 + nr_selector_args;
857 for (i = 0; i < pg->nr_pgpaths; i++) {
858 struct pgpath *pgpath;
859 struct dm_arg_set path_args;
861 if (as->argc < nr_args) {
862 ti->error = "not enough path parameters";
867 path_args.argc = nr_args;
868 path_args.argv = as->argv;
870 pgpath = parse_path(&path_args, &pg->ps, ti);
871 if (IS_ERR(pgpath)) {
877 list_add_tail(&pgpath->list, &pg->pgpaths);
878 dm_consume_args(as, nr_args);
884 free_priority_group(pg, ti);
888 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
892 struct dm_target *ti = m->ti;
894 static struct dm_arg _args[] = {
895 {0, 1024, "invalid number of hardware handler args"},
898 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
904 if (m->queue_mode == DM_TYPE_BIO_BASED) {
905 dm_consume_args(as, hw_argc);
906 DMERR("bio-based multipath doesn't allow hardware handler args");
910 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
911 if (!m->hw_handler_name)
918 for (i = 0; i <= hw_argc - 2; i++)
919 len += strlen(as->argv[i]) + 1;
920 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
922 ti->error = "memory allocation failed";
926 j = sprintf(p, "%d", hw_argc - 1);
927 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
928 j = sprintf(p, "%s", as->argv[i]);
930 dm_consume_args(as, hw_argc - 1);
934 kfree(m->hw_handler_name);
935 m->hw_handler_name = NULL;
939 static int parse_features(struct dm_arg_set *as, struct multipath *m)
943 struct dm_target *ti = m->ti;
944 const char *arg_name;
946 static struct dm_arg _args[] = {
947 {0, 8, "invalid number of feature args"},
948 {1, 50, "pg_init_retries must be between 1 and 50"},
949 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
952 r = dm_read_arg_group(_args, as, &argc, &ti->error);
960 arg_name = dm_shift_arg(as);
963 if (!strcasecmp(arg_name, "queue_if_no_path")) {
964 r = queue_if_no_path(m, true, false);
968 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
969 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
973 if (!strcasecmp(arg_name, "pg_init_retries") &&
975 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
980 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
982 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
987 if (!strcasecmp(arg_name, "queue_mode") &&
989 const char *queue_mode_name = dm_shift_arg(as);
991 if (!strcasecmp(queue_mode_name, "bio"))
992 m->queue_mode = DM_TYPE_BIO_BASED;
993 else if (!strcasecmp(queue_mode_name, "rq"))
994 m->queue_mode = DM_TYPE_REQUEST_BASED;
995 else if (!strcasecmp(queue_mode_name, "mq"))
996 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
998 ti->error = "Unknown 'queue_mode' requested";
1005 ti->error = "Unrecognised multipath feature request";
1007 } while (argc && !r);
1012 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1014 /* target arguments */
1015 static struct dm_arg _args[] = {
1016 {0, 1024, "invalid number of priority groups"},
1017 {0, 1024, "invalid initial priority group number"},
1021 struct multipath *m;
1022 struct dm_arg_set as;
1023 unsigned pg_count = 0;
1024 unsigned next_pg_num;
1029 m = alloc_multipath(ti);
1031 ti->error = "can't allocate multipath";
1035 r = parse_features(&as, m);
1039 r = alloc_multipath_stage2(ti, m);
1043 r = parse_hw_handler(&as, m);
1047 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1051 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1055 if ((!m->nr_priority_groups && next_pg_num) ||
1056 (m->nr_priority_groups && !next_pg_num)) {
1057 ti->error = "invalid initial priority group";
1062 /* parse the priority groups */
1064 struct priority_group *pg;
1065 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1067 pg = parse_priority_group(&as, m);
1073 nr_valid_paths += pg->nr_pgpaths;
1074 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1076 list_add_tail(&pg->list, &m->priority_groups);
1078 pg->pg_num = pg_count;
1083 if (pg_count != m->nr_priority_groups) {
1084 ti->error = "priority group count mismatch";
1089 ti->num_flush_bios = 1;
1090 ti->num_discard_bios = 1;
1091 ti->num_write_same_bios = 1;
1092 ti->num_write_zeroes_bios = 1;
1093 if (m->queue_mode == DM_TYPE_BIO_BASED)
1094 ti->per_io_data_size = multipath_per_bio_data_size();
1096 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1105 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1110 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1112 if (!atomic_read(&m->pg_init_in_progress))
1117 finish_wait(&m->pg_init_wait, &wait);
1120 static void flush_multipath_work(struct multipath *m)
1122 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1123 smp_mb__after_atomic();
1125 flush_workqueue(kmpath_handlerd);
1126 multipath_wait_for_pg_init_completion(m);
1127 flush_workqueue(kmultipathd);
1128 flush_work(&m->trigger_event);
1130 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1131 smp_mb__after_atomic();
1134 static void multipath_dtr(struct dm_target *ti)
1136 struct multipath *m = ti->private;
1138 flush_multipath_work(m);
1143 * Take a path out of use.
1145 static int fail_path(struct pgpath *pgpath)
1147 unsigned long flags;
1148 struct multipath *m = pgpath->pg->m;
1150 spin_lock_irqsave(&m->lock, flags);
1152 if (!pgpath->is_active)
1155 DMWARN("Failing path %s.", pgpath->path.dev->name);
1157 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1158 pgpath->is_active = false;
1159 pgpath->fail_count++;
1161 atomic_dec(&m->nr_valid_paths);
1163 if (pgpath == m->current_pgpath)
1164 m->current_pgpath = NULL;
1166 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1167 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1169 schedule_work(&m->trigger_event);
1172 spin_unlock_irqrestore(&m->lock, flags);
1178 * Reinstate a previously-failed path
1180 static int reinstate_path(struct pgpath *pgpath)
1182 int r = 0, run_queue = 0;
1183 unsigned long flags;
1184 struct multipath *m = pgpath->pg->m;
1185 unsigned nr_valid_paths;
1187 spin_lock_irqsave(&m->lock, flags);
1189 if (pgpath->is_active)
1192 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1194 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1198 pgpath->is_active = true;
1200 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1201 if (nr_valid_paths == 1) {
1202 m->current_pgpath = NULL;
1204 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1205 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1206 atomic_inc(&m->pg_init_in_progress);
1209 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1210 pgpath->path.dev->name, nr_valid_paths);
1212 schedule_work(&m->trigger_event);
1215 spin_unlock_irqrestore(&m->lock, flags);
1217 dm_table_run_md_queue_async(m->ti->table);
1218 process_queued_io_list(m);
1225 * Fail or reinstate all paths that match the provided struct dm_dev.
1227 static int action_dev(struct multipath *m, struct dm_dev *dev,
1231 struct pgpath *pgpath;
1232 struct priority_group *pg;
1234 list_for_each_entry(pg, &m->priority_groups, list) {
1235 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1236 if (pgpath->path.dev == dev)
1245 * Temporarily try to avoid having to use the specified PG
1247 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1250 unsigned long flags;
1252 spin_lock_irqsave(&m->lock, flags);
1254 pg->bypassed = bypassed;
1255 m->current_pgpath = NULL;
1256 m->current_pg = NULL;
1258 spin_unlock_irqrestore(&m->lock, flags);
1260 schedule_work(&m->trigger_event);
1264 * Switch to using the specified PG from the next I/O that gets mapped
1266 static int switch_pg_num(struct multipath *m, const char *pgstr)
1268 struct priority_group *pg;
1270 unsigned long flags;
1273 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1274 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1275 DMWARN("invalid PG number supplied to switch_pg_num");
1279 spin_lock_irqsave(&m->lock, flags);
1280 list_for_each_entry(pg, &m->priority_groups, list) {
1281 pg->bypassed = false;
1285 m->current_pgpath = NULL;
1286 m->current_pg = NULL;
1289 spin_unlock_irqrestore(&m->lock, flags);
1291 schedule_work(&m->trigger_event);
1296 * Set/clear bypassed status of a PG.
1297 * PGs are numbered upwards from 1 in the order they were declared.
1299 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1301 struct priority_group *pg;
1305 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1306 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1307 DMWARN("invalid PG number supplied to bypass_pg");
1311 list_for_each_entry(pg, &m->priority_groups, list) {
1316 bypass_pg(m, pg, bypassed);
1321 * Should we retry pg_init immediately?
1323 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1325 unsigned long flags;
1326 bool limit_reached = false;
1328 spin_lock_irqsave(&m->lock, flags);
1330 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1331 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1332 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1334 limit_reached = true;
1336 spin_unlock_irqrestore(&m->lock, flags);
1338 return limit_reached;
1341 static void pg_init_done(void *data, int errors)
1343 struct pgpath *pgpath = data;
1344 struct priority_group *pg = pgpath->pg;
1345 struct multipath *m = pg->m;
1346 unsigned long flags;
1347 bool delay_retry = false;
1349 /* device or driver problems */
1354 if (!m->hw_handler_name) {
1358 DMERR("Could not failover the device: Handler scsi_dh_%s "
1359 "Error %d.", m->hw_handler_name, errors);
1361 * Fail path for now, so we do not ping pong
1365 case SCSI_DH_DEV_TEMP_BUSY:
1367 * Probably doing something like FW upgrade on the
1368 * controller so try the other pg.
1370 bypass_pg(m, pg, true);
1373 /* Wait before retrying. */
1375 case SCSI_DH_IMM_RETRY:
1376 case SCSI_DH_RES_TEMP_UNAVAIL:
1377 if (pg_init_limit_reached(m, pgpath))
1381 case SCSI_DH_DEV_OFFLINED:
1384 * We probably do not want to fail the path for a device
1385 * error, but this is what the old dm did. In future
1386 * patches we can do more advanced handling.
1391 spin_lock_irqsave(&m->lock, flags);
1393 if (pgpath == m->current_pgpath) {
1394 DMERR("Could not failover device. Error %d.", errors);
1395 m->current_pgpath = NULL;
1396 m->current_pg = NULL;
1398 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1399 pg->bypassed = false;
1401 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1402 /* Activations of other paths are still on going */
1405 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1407 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1409 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1411 if (__pg_init_all_paths(m))
1414 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1416 process_queued_io_list(m);
1419 * Wake up any thread waiting to suspend.
1421 wake_up(&m->pg_init_wait);
1424 spin_unlock_irqrestore(&m->lock, flags);
1427 static void activate_or_offline_path(struct pgpath *pgpath)
1429 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1431 if (pgpath->is_active && !blk_queue_dying(q))
1432 scsi_dh_activate(q, pg_init_done, pgpath);
1434 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1437 static void activate_path_work(struct work_struct *work)
1439 struct pgpath *pgpath =
1440 container_of(work, struct pgpath, activate_path.work);
1442 activate_or_offline_path(pgpath);
1445 static int noretry_error(int error)
1450 * EBADE signals an reservation conflict.
1451 * We shouldn't fail the path here as we can communicate with
1452 * the target. We should failover to the next path, but in
1453 * doing so we might be causing a ping-pong between paths.
1454 * So just return the reservation conflict error.
1464 /* Anything else could be a path failure, so should be retried */
1468 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1469 int error, union map_info *map_context)
1471 struct dm_mpath_io *mpio = get_mpio(map_context);
1472 struct pgpath *pgpath = mpio->pgpath;
1473 int r = DM_ENDIO_DONE;
1476 * We don't queue any clone request inside the multipath target
1477 * during end I/O handling, since those clone requests don't have
1478 * bio clones. If we queue them inside the multipath target,
1479 * we need to make bio clones, that requires memory allocation.
1480 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1481 * don't have bio clones.)
1482 * Instead of queueing the clone request here, we queue the original
1483 * request into dm core, which will remake a clone request and
1484 * clone bios for it and resubmit it later.
1486 if (error && !noretry_error(error)) {
1487 struct multipath *m = ti->private;
1489 r = DM_ENDIO_REQUEUE;
1494 if (atomic_read(&m->nr_valid_paths) == 0 &&
1495 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1498 /* complete with the original error */
1504 struct path_selector *ps = &pgpath->pg->ps;
1506 if (ps->type->end_io)
1507 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1513 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1515 struct multipath *m = ti->private;
1516 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1517 struct pgpath *pgpath = mpio->pgpath;
1518 unsigned long flags;
1520 if (!error || noretry_error(error))
1526 if (atomic_read(&m->nr_valid_paths) == 0 &&
1527 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1533 /* Queue for the daemon to resubmit */
1534 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1536 spin_lock_irqsave(&m->lock, flags);
1537 bio_list_add(&m->queued_bios, clone);
1538 spin_unlock_irqrestore(&m->lock, flags);
1539 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1540 queue_work(kmultipathd, &m->process_queued_bios);
1542 error = DM_ENDIO_INCOMPLETE;
1545 struct path_selector *ps = &pgpath->pg->ps;
1547 if (ps->type->end_io)
1548 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1555 * Suspend can't complete until all the I/O is processed so if
1556 * the last path fails we must error any remaining I/O.
1557 * Note that if the freeze_bdev fails while suspending, the
1558 * queue_if_no_path state is lost - userspace should reset it.
1560 static void multipath_presuspend(struct dm_target *ti)
1562 struct multipath *m = ti->private;
1564 queue_if_no_path(m, false, true);
1567 static void multipath_postsuspend(struct dm_target *ti)
1569 struct multipath *m = ti->private;
1571 mutex_lock(&m->work_mutex);
1572 flush_multipath_work(m);
1573 mutex_unlock(&m->work_mutex);
1577 * Restore the queue_if_no_path setting.
1579 static void multipath_resume(struct dm_target *ti)
1581 struct multipath *m = ti->private;
1582 unsigned long flags;
1584 spin_lock_irqsave(&m->lock, flags);
1585 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1586 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1587 spin_unlock_irqrestore(&m->lock, flags);
1591 * Info output has the following format:
1592 * num_multipath_feature_args [multipath_feature_args]*
1593 * num_handler_status_args [handler_status_args]*
1594 * num_groups init_group_number
1595 * [A|D|E num_ps_status_args [ps_status_args]*
1596 * num_paths num_selector_args
1597 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1599 * Table output has the following format (identical to the constructor string):
1600 * num_feature_args [features_args]*
1601 * num_handler_args hw_handler [hw_handler_args]*
1602 * num_groups init_group_number
1603 * [priority selector-name num_ps_args [ps_args]*
1604 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1606 static void multipath_status(struct dm_target *ti, status_type_t type,
1607 unsigned status_flags, char *result, unsigned maxlen)
1610 unsigned long flags;
1611 struct multipath *m = ti->private;
1612 struct priority_group *pg;
1617 spin_lock_irqsave(&m->lock, flags);
1620 if (type == STATUSTYPE_INFO)
1621 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1622 atomic_read(&m->pg_init_count));
1624 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1625 (m->pg_init_retries > 0) * 2 +
1626 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1627 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1628 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1630 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1631 DMEMIT("queue_if_no_path ");
1632 if (m->pg_init_retries)
1633 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1634 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1635 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1636 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1637 DMEMIT("retain_attached_hw_handler ");
1638 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1639 switch(m->queue_mode) {
1640 case DM_TYPE_BIO_BASED:
1641 DMEMIT("queue_mode bio ");
1643 case DM_TYPE_MQ_REQUEST_BASED:
1644 DMEMIT("queue_mode mq ");
1653 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1656 DMEMIT("1 %s ", m->hw_handler_name);
1658 DMEMIT("%u ", m->nr_priority_groups);
1661 pg_num = m->next_pg->pg_num;
1662 else if (m->current_pg)
1663 pg_num = m->current_pg->pg_num;
1665 pg_num = (m->nr_priority_groups ? 1 : 0);
1667 DMEMIT("%u ", pg_num);
1670 case STATUSTYPE_INFO:
1671 list_for_each_entry(pg, &m->priority_groups, list) {
1673 state = 'D'; /* Disabled */
1674 else if (pg == m->current_pg)
1675 state = 'A'; /* Currently Active */
1677 state = 'E'; /* Enabled */
1679 DMEMIT("%c ", state);
1681 if (pg->ps.type->status)
1682 sz += pg->ps.type->status(&pg->ps, NULL, type,
1688 DMEMIT("%u %u ", pg->nr_pgpaths,
1689 pg->ps.type->info_args);
1691 list_for_each_entry(p, &pg->pgpaths, list) {
1692 DMEMIT("%s %s %u ", p->path.dev->name,
1693 p->is_active ? "A" : "F",
1695 if (pg->ps.type->status)
1696 sz += pg->ps.type->status(&pg->ps,
1697 &p->path, type, result + sz,
1703 case STATUSTYPE_TABLE:
1704 list_for_each_entry(pg, &m->priority_groups, list) {
1705 DMEMIT("%s ", pg->ps.type->name);
1707 if (pg->ps.type->status)
1708 sz += pg->ps.type->status(&pg->ps, NULL, type,
1714 DMEMIT("%u %u ", pg->nr_pgpaths,
1715 pg->ps.type->table_args);
1717 list_for_each_entry(p, &pg->pgpaths, list) {
1718 DMEMIT("%s ", p->path.dev->name);
1719 if (pg->ps.type->status)
1720 sz += pg->ps.type->status(&pg->ps,
1721 &p->path, type, result + sz,
1728 spin_unlock_irqrestore(&m->lock, flags);
1731 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1735 struct multipath *m = ti->private;
1738 mutex_lock(&m->work_mutex);
1740 if (dm_suspended(ti)) {
1746 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1747 r = queue_if_no_path(m, true, false);
1749 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1750 r = queue_if_no_path(m, false, false);
1756 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1760 if (!strcasecmp(argv[0], "disable_group")) {
1761 r = bypass_pg_num(m, argv[1], true);
1763 } else if (!strcasecmp(argv[0], "enable_group")) {
1764 r = bypass_pg_num(m, argv[1], false);
1766 } else if (!strcasecmp(argv[0], "switch_group")) {
1767 r = switch_pg_num(m, argv[1]);
1769 } else if (!strcasecmp(argv[0], "reinstate_path"))
1770 action = reinstate_path;
1771 else if (!strcasecmp(argv[0], "fail_path"))
1774 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1778 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1780 DMWARN("message: error getting device %s",
1785 r = action_dev(m, dev, action);
1787 dm_put_device(ti, dev);
1790 mutex_unlock(&m->work_mutex);
1794 static int multipath_prepare_ioctl(struct dm_target *ti,
1795 struct block_device **bdev, fmode_t *mode)
1797 struct multipath *m = ti->private;
1798 struct pgpath *current_pgpath;
1801 current_pgpath = lockless_dereference(m->current_pgpath);
1802 if (!current_pgpath)
1803 current_pgpath = choose_pgpath(m, 0);
1805 if (current_pgpath) {
1806 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1807 *bdev = current_pgpath->path.dev->bdev;
1808 *mode = current_pgpath->path.dev->mode;
1811 /* pg_init has not started or completed */
1815 /* No path is available */
1816 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1822 if (r == -ENOTCONN) {
1823 if (!lockless_dereference(m->current_pg)) {
1824 /* Path status changed, redo selection */
1825 (void) choose_pgpath(m, 0);
1827 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1828 pg_init_all_paths(m);
1829 dm_table_run_md_queue_async(m->ti->table);
1830 process_queued_io_list(m);
1834 * Only pass ioctls through if the device sizes match exactly.
1836 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1841 static int multipath_iterate_devices(struct dm_target *ti,
1842 iterate_devices_callout_fn fn, void *data)
1844 struct multipath *m = ti->private;
1845 struct priority_group *pg;
1849 list_for_each_entry(pg, &m->priority_groups, list) {
1850 list_for_each_entry(p, &pg->pgpaths, list) {
1851 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1861 static int pgpath_busy(struct pgpath *pgpath)
1863 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1865 return blk_lld_busy(q);
1869 * We return "busy", only when we can map I/Os but underlying devices
1870 * are busy (so even if we map I/Os now, the I/Os will wait on
1871 * the underlying queue).
1872 * In other words, if we want to kill I/Os or queue them inside us
1873 * due to map unavailability, we don't return "busy". Otherwise,
1874 * dm core won't give us the I/Os and we can't do what we want.
1876 static int multipath_busy(struct dm_target *ti)
1878 bool busy = false, has_active = false;
1879 struct multipath *m = ti->private;
1880 struct priority_group *pg, *next_pg;
1881 struct pgpath *pgpath;
1883 /* pg_init in progress */
1884 if (atomic_read(&m->pg_init_in_progress))
1887 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1888 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1889 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1891 /* Guess which priority_group will be used at next mapping time */
1892 pg = lockless_dereference(m->current_pg);
1893 next_pg = lockless_dereference(m->next_pg);
1894 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1899 * We don't know which pg will be used at next mapping time.
1900 * We don't call choose_pgpath() here to avoid to trigger
1901 * pg_init just by busy checking.
1902 * So we don't know whether underlying devices we will be using
1903 * at next mapping time are busy or not. Just try mapping.
1909 * If there is one non-busy active path at least, the path selector
1910 * will be able to select it. So we consider such a pg as not busy.
1913 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1914 if (pgpath->is_active) {
1916 if (!pgpath_busy(pgpath)) {
1925 * No active path in this pg, so this pg won't be used and
1926 * the current_pg will be changed at next mapping time.
1927 * We need to try mapping to determine it.
1935 /*-----------------------------------------------------------------
1937 *---------------------------------------------------------------*/
1938 static struct target_type multipath_target = {
1939 .name = "multipath",
1940 .version = {1, 12, 0},
1941 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1942 .module = THIS_MODULE,
1943 .ctr = multipath_ctr,
1944 .dtr = multipath_dtr,
1945 .clone_and_map_rq = multipath_clone_and_map,
1946 .release_clone_rq = multipath_release_clone,
1947 .rq_end_io = multipath_end_io,
1948 .map = multipath_map_bio,
1949 .end_io = multipath_end_io_bio,
1950 .presuspend = multipath_presuspend,
1951 .postsuspend = multipath_postsuspend,
1952 .resume = multipath_resume,
1953 .status = multipath_status,
1954 .message = multipath_message,
1955 .prepare_ioctl = multipath_prepare_ioctl,
1956 .iterate_devices = multipath_iterate_devices,
1957 .busy = multipath_busy,
1960 static int __init dm_multipath_init(void)
1964 r = dm_register_target(&multipath_target);
1966 DMERR("request-based register failed %d", r);
1968 goto bad_register_target;
1971 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1973 DMERR("failed to create workqueue kmpathd");
1975 goto bad_alloc_kmultipathd;
1979 * A separate workqueue is used to handle the device handlers
1980 * to avoid overloading existing workqueue. Overloading the
1981 * old workqueue would also create a bottleneck in the
1982 * path of the storage hardware device activation.
1984 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1986 if (!kmpath_handlerd) {
1987 DMERR("failed to create workqueue kmpath_handlerd");
1989 goto bad_alloc_kmpath_handlerd;
1994 bad_alloc_kmpath_handlerd:
1995 destroy_workqueue(kmultipathd);
1996 bad_alloc_kmultipathd:
1997 dm_unregister_target(&multipath_target);
1998 bad_register_target:
2002 static void __exit dm_multipath_exit(void)
2004 destroy_workqueue(kmpath_handlerd);
2005 destroy_workqueue(kmultipathd);
2007 dm_unregister_target(&multipath_target);
2010 module_init(dm_multipath_init);
2011 module_exit(dm_multipath_exit);
2013 MODULE_DESCRIPTION(DM_NAME " multipath target");
2014 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2015 MODULE_LICENSE("GPL");