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_mpio(struct dm_mpath_io *mpio)
269 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270 void *bio_details = mpio + 1;
274 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
276 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
277 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
279 mpio->nr_bytes = bio->bi_iter.bi_size;
283 dm_bio_record(bio_details, bio);
286 /*-----------------------------------------------
288 *-----------------------------------------------*/
290 static int __pg_init_all_paths(struct multipath *m)
292 struct pgpath *pgpath;
293 unsigned long pg_init_delay = 0;
295 lockdep_assert_held(&m->lock);
297 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
300 atomic_inc(&m->pg_init_count);
301 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
303 /* Check here to reset pg_init_required */
307 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
308 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
309 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
310 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
311 /* Skip failed paths */
312 if (!pgpath->is_active)
314 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
316 atomic_inc(&m->pg_init_in_progress);
318 return atomic_read(&m->pg_init_in_progress);
321 static int pg_init_all_paths(struct multipath *m)
326 spin_lock_irqsave(&m->lock, flags);
327 ret = __pg_init_all_paths(m);
328 spin_unlock_irqrestore(&m->lock, flags);
333 static void __switch_pg(struct multipath *m, struct priority_group *pg)
337 /* Must we initialise the PG first, and queue I/O till it's ready? */
338 if (m->hw_handler_name) {
339 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
340 set_bit(MPATHF_QUEUE_IO, &m->flags);
342 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
343 clear_bit(MPATHF_QUEUE_IO, &m->flags);
346 atomic_set(&m->pg_init_count, 0);
349 static struct pgpath *choose_path_in_pg(struct multipath *m,
350 struct priority_group *pg,
354 struct dm_path *path;
355 struct pgpath *pgpath;
357 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
359 return ERR_PTR(-ENXIO);
361 pgpath = path_to_pgpath(path);
363 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
364 /* Only update current_pgpath if pg changed */
365 spin_lock_irqsave(&m->lock, flags);
366 m->current_pgpath = pgpath;
368 spin_unlock_irqrestore(&m->lock, flags);
374 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
377 struct priority_group *pg;
378 struct pgpath *pgpath;
379 unsigned bypassed = 1;
381 if (!atomic_read(&m->nr_valid_paths)) {
382 clear_bit(MPATHF_QUEUE_IO, &m->flags);
386 /* Were we instructed to switch PG? */
387 if (READ_ONCE(m->next_pg)) {
388 spin_lock_irqsave(&m->lock, flags);
391 spin_unlock_irqrestore(&m->lock, flags);
392 goto check_current_pg;
395 spin_unlock_irqrestore(&m->lock, flags);
396 pgpath = choose_path_in_pg(m, pg, nr_bytes);
397 if (!IS_ERR_OR_NULL(pgpath))
401 /* Don't change PG until it has no remaining paths */
403 pg = READ_ONCE(m->current_pg);
405 pgpath = choose_path_in_pg(m, pg, nr_bytes);
406 if (!IS_ERR_OR_NULL(pgpath))
411 * Loop through priority groups until we find a valid path.
412 * First time we skip PGs marked 'bypassed'.
413 * Second time we only try the ones we skipped, but set
414 * pg_init_delay_retry so we do not hammer controllers.
417 list_for_each_entry(pg, &m->priority_groups, list) {
418 if (pg->bypassed == !!bypassed)
420 pgpath = choose_path_in_pg(m, pg, nr_bytes);
421 if (!IS_ERR_OR_NULL(pgpath)) {
423 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
427 } while (bypassed--);
430 spin_lock_irqsave(&m->lock, flags);
431 m->current_pgpath = NULL;
432 m->current_pg = NULL;
433 spin_unlock_irqrestore(&m->lock, flags);
439 * dm_report_EIO() is a macro instead of a function to make pr_debug()
440 * report the function name and line number of the function from which
441 * it has been invoked.
443 #define dm_report_EIO(m) \
445 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
447 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
448 dm_device_name(md), \
449 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
450 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
451 dm_noflush_suspending((m)->ti)); \
455 * Check whether bios must be queued in the device-mapper core rather
456 * than here in the target.
458 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
459 * the same value then we are not between multipath_presuspend()
460 * and multipath_resume() calls and we have no need to check
461 * for the DMF_NOFLUSH_SUSPENDING flag.
463 static bool __must_push_back(struct multipath *m, unsigned long flags)
465 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
466 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
467 dm_noflush_suspending(m->ti));
471 * Following functions use READ_ONCE to get atomic access to
472 * all m->flags to avoid taking spinlock
474 static bool must_push_back_rq(struct multipath *m)
476 unsigned long flags = READ_ONCE(m->flags);
477 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
480 static bool must_push_back_bio(struct multipath *m)
482 unsigned long flags = READ_ONCE(m->flags);
483 return __must_push_back(m, flags);
487 * Map cloned requests (request-based multipath)
489 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
490 union map_info *map_context,
491 struct request **__clone)
493 struct multipath *m = ti->private;
494 size_t nr_bytes = blk_rq_bytes(rq);
495 struct pgpath *pgpath;
496 struct block_device *bdev;
497 struct dm_mpath_io *mpio = get_mpio(map_context);
498 struct request_queue *q;
499 struct request *clone;
501 /* Do we need to select a new pgpath? */
502 pgpath = READ_ONCE(m->current_pgpath);
503 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
504 pgpath = choose_pgpath(m, nr_bytes);
507 if (must_push_back_rq(m))
508 return DM_MAPIO_DELAY_REQUEUE;
509 dm_report_EIO(m); /* Failed */
510 return DM_MAPIO_KILL;
511 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
512 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
513 if (pg_init_all_paths(m))
514 return DM_MAPIO_DELAY_REQUEUE;
515 return DM_MAPIO_REQUEUE;
518 mpio->pgpath = pgpath;
519 mpio->nr_bytes = nr_bytes;
521 bdev = pgpath->path.dev->bdev;
522 q = bdev_get_queue(bdev);
523 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
525 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
526 bool queue_dying = blk_queue_dying(q);
528 atomic_inc(&m->pg_init_in_progress);
529 activate_or_offline_path(pgpath);
531 return DM_MAPIO_DELAY_REQUEUE;
533 clone->bio = clone->biotail = NULL;
534 clone->rq_disk = bdev->bd_disk;
535 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
538 if (pgpath->pg->ps.type->start_io)
539 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
542 return DM_MAPIO_REMAPPED;
545 static void multipath_release_clone(struct request *clone)
547 blk_put_request(clone);
551 * Map cloned bios (bio-based multipath)
553 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
555 struct pgpath *pgpath;
559 /* Do we need to select a new pgpath? */
560 pgpath = READ_ONCE(m->current_pgpath);
561 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
562 if (!pgpath || !queue_io)
563 pgpath = choose_pgpath(m, mpio->nr_bytes);
565 if ((pgpath && queue_io) ||
566 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
567 /* Queue for the daemon to resubmit */
568 spin_lock_irqsave(&m->lock, flags);
569 bio_list_add(&m->queued_bios, bio);
570 spin_unlock_irqrestore(&m->lock, flags);
571 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
572 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
573 pg_init_all_paths(m);
575 queue_work(kmultipathd, &m->process_queued_bios);
576 return DM_MAPIO_SUBMITTED;
580 if (must_push_back_bio(m))
581 return DM_MAPIO_REQUEUE;
583 return DM_MAPIO_KILL;
586 mpio->pgpath = pgpath;
589 bio_set_dev(bio, pgpath->path.dev->bdev);
590 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
592 if (pgpath->pg->ps.type->start_io)
593 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
596 return DM_MAPIO_REMAPPED;
599 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
601 struct multipath *m = ti->private;
602 struct dm_mpath_io *mpio = NULL;
604 multipath_init_per_bio_data(bio, &mpio);
605 return __multipath_map_bio(m, bio, mpio);
608 static void process_queued_io_list(struct multipath *m)
610 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
611 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
612 else if (m->queue_mode == DM_TYPE_BIO_BASED)
613 queue_work(kmultipathd, &m->process_queued_bios);
616 static void process_queued_bios(struct work_struct *work)
621 struct bio_list bios;
622 struct blk_plug plug;
623 struct multipath *m =
624 container_of(work, struct multipath, process_queued_bios);
626 bio_list_init(&bios);
628 spin_lock_irqsave(&m->lock, flags);
630 if (bio_list_empty(&m->queued_bios)) {
631 spin_unlock_irqrestore(&m->lock, flags);
635 bio_list_merge(&bios, &m->queued_bios);
636 bio_list_init(&m->queued_bios);
638 spin_unlock_irqrestore(&m->lock, flags);
640 blk_start_plug(&plug);
641 while ((bio = bio_list_pop(&bios))) {
642 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
643 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
644 r = __multipath_map_bio(m, bio, mpio);
647 bio->bi_status = BLK_STS_IOERR;
650 case DM_MAPIO_REQUEUE:
651 bio->bi_status = BLK_STS_DM_REQUEUE;
654 case DM_MAPIO_REMAPPED:
655 generic_make_request(bio);
660 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
663 blk_finish_plug(&plug);
667 * If we run out of usable paths, should we queue I/O or error it?
669 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
674 spin_lock_irqsave(&m->lock, flags);
675 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
676 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
677 (!save_old_value && queue_if_no_path));
678 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
679 spin_unlock_irqrestore(&m->lock, flags);
681 if (!queue_if_no_path) {
682 dm_table_run_md_queue_async(m->ti->table);
683 process_queued_io_list(m);
690 * An event is triggered whenever a path is taken out of use.
691 * Includes path failure and PG bypass.
693 static void trigger_event(struct work_struct *work)
695 struct multipath *m =
696 container_of(work, struct multipath, trigger_event);
698 dm_table_event(m->ti->table);
701 /*-----------------------------------------------------------------
702 * Constructor/argument parsing:
703 * <#multipath feature args> [<arg>]*
704 * <#hw_handler args> [hw_handler [<arg>]*]
706 * <initial priority group>
707 * [<selector> <#selector args> [<arg>]*
708 * <#paths> <#per-path selector args>
709 * [<path> [<arg>]* ]+ ]+
710 *---------------------------------------------------------------*/
711 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
712 struct dm_target *ti)
715 struct path_selector_type *pst;
718 static const struct dm_arg _args[] = {
719 {0, 1024, "invalid number of path selector args"},
722 pst = dm_get_path_selector(dm_shift_arg(as));
724 ti->error = "unknown path selector type";
728 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
730 dm_put_path_selector(pst);
734 r = pst->create(&pg->ps, ps_argc, as->argv);
736 dm_put_path_selector(pst);
737 ti->error = "path selector constructor failed";
742 dm_consume_args(as, ps_argc);
747 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
748 struct dm_target *ti)
752 struct multipath *m = ti->private;
753 struct request_queue *q = NULL;
754 const char *attached_handler_name;
756 /* we need at least a path arg */
758 ti->error = "no device given";
759 return ERR_PTR(-EINVAL);
764 return ERR_PTR(-ENOMEM);
766 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
769 ti->error = "error getting device";
773 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
774 q = bdev_get_queue(p->path.dev->bdev);
776 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
778 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
779 if (attached_handler_name) {
781 * Clear any hw_handler_params associated with a
782 * handler that isn't already attached.
784 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
785 kfree(m->hw_handler_params);
786 m->hw_handler_params = NULL;
790 * Reset hw_handler_name to match the attached handler
792 * NB. This modifies the table line to show the actual
793 * handler instead of the original table passed in.
795 kfree(m->hw_handler_name);
796 m->hw_handler_name = attached_handler_name;
800 if (m->hw_handler_name) {
801 r = scsi_dh_attach(q, m->hw_handler_name);
803 char b[BDEVNAME_SIZE];
805 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
806 bdevname(p->path.dev->bdev, b));
810 ti->error = "error attaching hardware handler";
811 dm_put_device(ti, p->path.dev);
815 if (m->hw_handler_params) {
816 r = scsi_dh_set_params(q, m->hw_handler_params);
818 ti->error = "unable to set hardware "
819 "handler parameters";
820 dm_put_device(ti, p->path.dev);
826 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
828 dm_put_device(ti, p->path.dev);
839 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
842 static const struct dm_arg _args[] = {
843 {1, 1024, "invalid number of paths"},
844 {0, 1024, "invalid number of selector args"}
848 unsigned i, nr_selector_args, nr_args;
849 struct priority_group *pg;
850 struct dm_target *ti = m->ti;
854 ti->error = "not enough priority group arguments";
855 return ERR_PTR(-EINVAL);
858 pg = alloc_priority_group();
860 ti->error = "couldn't allocate priority group";
861 return ERR_PTR(-ENOMEM);
865 r = parse_path_selector(as, pg, ti);
872 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
876 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
880 nr_args = 1 + nr_selector_args;
881 for (i = 0; i < pg->nr_pgpaths; i++) {
882 struct pgpath *pgpath;
883 struct dm_arg_set path_args;
885 if (as->argc < nr_args) {
886 ti->error = "not enough path parameters";
891 path_args.argc = nr_args;
892 path_args.argv = as->argv;
894 pgpath = parse_path(&path_args, &pg->ps, ti);
895 if (IS_ERR(pgpath)) {
901 list_add_tail(&pgpath->list, &pg->pgpaths);
902 dm_consume_args(as, nr_args);
908 free_priority_group(pg, ti);
912 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
916 struct dm_target *ti = m->ti;
918 static const struct dm_arg _args[] = {
919 {0, 1024, "invalid number of hardware handler args"},
922 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
928 if (m->queue_mode == DM_TYPE_BIO_BASED) {
929 dm_consume_args(as, hw_argc);
930 DMERR("bio-based multipath doesn't allow hardware handler args");
934 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
935 if (!m->hw_handler_name)
942 for (i = 0; i <= hw_argc - 2; i++)
943 len += strlen(as->argv[i]) + 1;
944 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
946 ti->error = "memory allocation failed";
950 j = sprintf(p, "%d", hw_argc - 1);
951 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
952 j = sprintf(p, "%s", as->argv[i]);
954 dm_consume_args(as, hw_argc - 1);
958 kfree(m->hw_handler_name);
959 m->hw_handler_name = NULL;
963 static int parse_features(struct dm_arg_set *as, struct multipath *m)
967 struct dm_target *ti = m->ti;
968 const char *arg_name;
970 static const struct dm_arg _args[] = {
971 {0, 8, "invalid number of feature args"},
972 {1, 50, "pg_init_retries must be between 1 and 50"},
973 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
976 r = dm_read_arg_group(_args, as, &argc, &ti->error);
984 arg_name = dm_shift_arg(as);
987 if (!strcasecmp(arg_name, "queue_if_no_path")) {
988 r = queue_if_no_path(m, true, false);
992 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
993 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
997 if (!strcasecmp(arg_name, "pg_init_retries") &&
999 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1004 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1006 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1011 if (!strcasecmp(arg_name, "queue_mode") &&
1013 const char *queue_mode_name = dm_shift_arg(as);
1015 if (!strcasecmp(queue_mode_name, "bio"))
1016 m->queue_mode = DM_TYPE_BIO_BASED;
1017 else if (!strcasecmp(queue_mode_name, "rq"))
1018 m->queue_mode = DM_TYPE_REQUEST_BASED;
1019 else if (!strcasecmp(queue_mode_name, "mq"))
1020 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1022 ti->error = "Unknown 'queue_mode' requested";
1029 ti->error = "Unrecognised multipath feature request";
1031 } while (argc && !r);
1036 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1038 /* target arguments */
1039 static const struct dm_arg _args[] = {
1040 {0, 1024, "invalid number of priority groups"},
1041 {0, 1024, "invalid initial priority group number"},
1045 struct multipath *m;
1046 struct dm_arg_set as;
1047 unsigned pg_count = 0;
1048 unsigned next_pg_num;
1053 m = alloc_multipath(ti);
1055 ti->error = "can't allocate multipath";
1059 r = parse_features(&as, m);
1063 r = alloc_multipath_stage2(ti, m);
1067 r = parse_hw_handler(&as, m);
1071 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1075 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1079 if ((!m->nr_priority_groups && next_pg_num) ||
1080 (m->nr_priority_groups && !next_pg_num)) {
1081 ti->error = "invalid initial priority group";
1086 /* parse the priority groups */
1088 struct priority_group *pg;
1089 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1091 pg = parse_priority_group(&as, m);
1097 nr_valid_paths += pg->nr_pgpaths;
1098 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1100 list_add_tail(&pg->list, &m->priority_groups);
1102 pg->pg_num = pg_count;
1107 if (pg_count != m->nr_priority_groups) {
1108 ti->error = "priority group count mismatch";
1113 ti->num_flush_bios = 1;
1114 ti->num_discard_bios = 1;
1115 ti->num_write_same_bios = 1;
1116 ti->num_write_zeroes_bios = 1;
1117 if (m->queue_mode == DM_TYPE_BIO_BASED)
1118 ti->per_io_data_size = multipath_per_bio_data_size();
1120 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1129 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1134 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1136 if (!atomic_read(&m->pg_init_in_progress))
1141 finish_wait(&m->pg_init_wait, &wait);
1144 static void flush_multipath_work(struct multipath *m)
1146 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1147 smp_mb__after_atomic();
1149 flush_workqueue(kmpath_handlerd);
1150 multipath_wait_for_pg_init_completion(m);
1151 flush_workqueue(kmultipathd);
1152 flush_work(&m->trigger_event);
1154 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1155 smp_mb__after_atomic();
1158 static void multipath_dtr(struct dm_target *ti)
1160 struct multipath *m = ti->private;
1162 flush_multipath_work(m);
1167 * Take a path out of use.
1169 static int fail_path(struct pgpath *pgpath)
1171 unsigned long flags;
1172 struct multipath *m = pgpath->pg->m;
1174 spin_lock_irqsave(&m->lock, flags);
1176 if (!pgpath->is_active)
1179 DMWARN("Failing path %s.", pgpath->path.dev->name);
1181 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1182 pgpath->is_active = false;
1183 pgpath->fail_count++;
1185 atomic_dec(&m->nr_valid_paths);
1187 if (pgpath == m->current_pgpath)
1188 m->current_pgpath = NULL;
1190 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1191 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1193 schedule_work(&m->trigger_event);
1196 spin_unlock_irqrestore(&m->lock, flags);
1202 * Reinstate a previously-failed path
1204 static int reinstate_path(struct pgpath *pgpath)
1206 int r = 0, run_queue = 0;
1207 unsigned long flags;
1208 struct multipath *m = pgpath->pg->m;
1209 unsigned nr_valid_paths;
1211 spin_lock_irqsave(&m->lock, flags);
1213 if (pgpath->is_active)
1216 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1218 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1222 pgpath->is_active = true;
1224 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1225 if (nr_valid_paths == 1) {
1226 m->current_pgpath = NULL;
1228 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1229 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1230 atomic_inc(&m->pg_init_in_progress);
1233 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1234 pgpath->path.dev->name, nr_valid_paths);
1236 schedule_work(&m->trigger_event);
1239 spin_unlock_irqrestore(&m->lock, flags);
1241 dm_table_run_md_queue_async(m->ti->table);
1242 process_queued_io_list(m);
1249 * Fail or reinstate all paths that match the provided struct dm_dev.
1251 static int action_dev(struct multipath *m, struct dm_dev *dev,
1255 struct pgpath *pgpath;
1256 struct priority_group *pg;
1258 list_for_each_entry(pg, &m->priority_groups, list) {
1259 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1260 if (pgpath->path.dev == dev)
1269 * Temporarily try to avoid having to use the specified PG
1271 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1274 unsigned long flags;
1276 spin_lock_irqsave(&m->lock, flags);
1278 pg->bypassed = bypassed;
1279 m->current_pgpath = NULL;
1280 m->current_pg = NULL;
1282 spin_unlock_irqrestore(&m->lock, flags);
1284 schedule_work(&m->trigger_event);
1288 * Switch to using the specified PG from the next I/O that gets mapped
1290 static int switch_pg_num(struct multipath *m, const char *pgstr)
1292 struct priority_group *pg;
1294 unsigned long flags;
1297 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1298 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1299 DMWARN("invalid PG number supplied to switch_pg_num");
1303 spin_lock_irqsave(&m->lock, flags);
1304 list_for_each_entry(pg, &m->priority_groups, list) {
1305 pg->bypassed = false;
1309 m->current_pgpath = NULL;
1310 m->current_pg = NULL;
1313 spin_unlock_irqrestore(&m->lock, flags);
1315 schedule_work(&m->trigger_event);
1320 * Set/clear bypassed status of a PG.
1321 * PGs are numbered upwards from 1 in the order they were declared.
1323 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1325 struct priority_group *pg;
1329 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1330 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1331 DMWARN("invalid PG number supplied to bypass_pg");
1335 list_for_each_entry(pg, &m->priority_groups, list) {
1340 bypass_pg(m, pg, bypassed);
1345 * Should we retry pg_init immediately?
1347 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1349 unsigned long flags;
1350 bool limit_reached = false;
1352 spin_lock_irqsave(&m->lock, flags);
1354 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1355 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1356 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1358 limit_reached = true;
1360 spin_unlock_irqrestore(&m->lock, flags);
1362 return limit_reached;
1365 static void pg_init_done(void *data, int errors)
1367 struct pgpath *pgpath = data;
1368 struct priority_group *pg = pgpath->pg;
1369 struct multipath *m = pg->m;
1370 unsigned long flags;
1371 bool delay_retry = false;
1373 /* device or driver problems */
1378 if (!m->hw_handler_name) {
1382 DMERR("Could not failover the device: Handler scsi_dh_%s "
1383 "Error %d.", m->hw_handler_name, errors);
1385 * Fail path for now, so we do not ping pong
1389 case SCSI_DH_DEV_TEMP_BUSY:
1391 * Probably doing something like FW upgrade on the
1392 * controller so try the other pg.
1394 bypass_pg(m, pg, true);
1397 /* Wait before retrying. */
1400 case SCSI_DH_IMM_RETRY:
1401 case SCSI_DH_RES_TEMP_UNAVAIL:
1402 if (pg_init_limit_reached(m, pgpath))
1406 case SCSI_DH_DEV_OFFLINED:
1409 * We probably do not want to fail the path for a device
1410 * error, but this is what the old dm did. In future
1411 * patches we can do more advanced handling.
1416 spin_lock_irqsave(&m->lock, flags);
1418 if (pgpath == m->current_pgpath) {
1419 DMERR("Could not failover device. Error %d.", errors);
1420 m->current_pgpath = NULL;
1421 m->current_pg = NULL;
1423 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1424 pg->bypassed = false;
1426 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1427 /* Activations of other paths are still on going */
1430 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1432 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1434 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1436 if (__pg_init_all_paths(m))
1439 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1441 process_queued_io_list(m);
1444 * Wake up any thread waiting to suspend.
1446 wake_up(&m->pg_init_wait);
1449 spin_unlock_irqrestore(&m->lock, flags);
1452 static void activate_or_offline_path(struct pgpath *pgpath)
1454 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1456 if (pgpath->is_active && !blk_queue_dying(q))
1457 scsi_dh_activate(q, pg_init_done, pgpath);
1459 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1462 static void activate_path_work(struct work_struct *work)
1464 struct pgpath *pgpath =
1465 container_of(work, struct pgpath, activate_path.work);
1467 activate_or_offline_path(pgpath);
1470 static int noretry_error(blk_status_t error)
1473 case BLK_STS_NOTSUPP:
1475 case BLK_STS_TARGET:
1477 case BLK_STS_MEDIUM:
1481 /* Anything else could be a path failure, so should be retried */
1485 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1486 blk_status_t error, union map_info *map_context)
1488 struct dm_mpath_io *mpio = get_mpio(map_context);
1489 struct pgpath *pgpath = mpio->pgpath;
1490 int r = DM_ENDIO_DONE;
1493 * We don't queue any clone request inside the multipath target
1494 * during end I/O handling, since those clone requests don't have
1495 * bio clones. If we queue them inside the multipath target,
1496 * we need to make bio clones, that requires memory allocation.
1497 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1498 * don't have bio clones.)
1499 * Instead of queueing the clone request here, we queue the original
1500 * request into dm core, which will remake a clone request and
1501 * clone bios for it and resubmit it later.
1503 if (error && !noretry_error(error)) {
1504 struct multipath *m = ti->private;
1506 r = DM_ENDIO_REQUEUE;
1511 if (atomic_read(&m->nr_valid_paths) == 0 &&
1512 !must_push_back_rq(m)) {
1513 if (error == BLK_STS_IOERR)
1515 /* complete with the original error */
1521 struct path_selector *ps = &pgpath->pg->ps;
1523 if (ps->type->end_io)
1524 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1530 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1531 blk_status_t *error)
1533 struct multipath *m = ti->private;
1534 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1535 struct pgpath *pgpath = mpio->pgpath;
1536 unsigned long flags;
1537 int r = DM_ENDIO_DONE;
1539 if (!*error || noretry_error(*error))
1545 if (atomic_read(&m->nr_valid_paths) == 0 &&
1546 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1547 if (must_push_back_bio(m)) {
1548 r = DM_ENDIO_REQUEUE;
1551 *error = BLK_STS_IOERR;
1556 spin_lock_irqsave(&m->lock, flags);
1557 bio_list_add(&m->queued_bios, clone);
1558 spin_unlock_irqrestore(&m->lock, flags);
1559 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1560 queue_work(kmultipathd, &m->process_queued_bios);
1562 r = DM_ENDIO_INCOMPLETE;
1565 struct path_selector *ps = &pgpath->pg->ps;
1567 if (ps->type->end_io)
1568 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1575 * Suspend can't complete until all the I/O is processed so if
1576 * the last path fails we must error any remaining I/O.
1577 * Note that if the freeze_bdev fails while suspending, the
1578 * queue_if_no_path state is lost - userspace should reset it.
1580 static void multipath_presuspend(struct dm_target *ti)
1582 struct multipath *m = ti->private;
1584 queue_if_no_path(m, false, true);
1587 static void multipath_postsuspend(struct dm_target *ti)
1589 struct multipath *m = ti->private;
1591 mutex_lock(&m->work_mutex);
1592 flush_multipath_work(m);
1593 mutex_unlock(&m->work_mutex);
1597 * Restore the queue_if_no_path setting.
1599 static void multipath_resume(struct dm_target *ti)
1601 struct multipath *m = ti->private;
1602 unsigned long flags;
1604 spin_lock_irqsave(&m->lock, flags);
1605 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1606 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1607 spin_unlock_irqrestore(&m->lock, flags);
1611 * Info output has the following format:
1612 * num_multipath_feature_args [multipath_feature_args]*
1613 * num_handler_status_args [handler_status_args]*
1614 * num_groups init_group_number
1615 * [A|D|E num_ps_status_args [ps_status_args]*
1616 * num_paths num_selector_args
1617 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1619 * Table output has the following format (identical to the constructor string):
1620 * num_feature_args [features_args]*
1621 * num_handler_args hw_handler [hw_handler_args]*
1622 * num_groups init_group_number
1623 * [priority selector-name num_ps_args [ps_args]*
1624 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1626 static void multipath_status(struct dm_target *ti, status_type_t type,
1627 unsigned status_flags, char *result, unsigned maxlen)
1630 unsigned long flags;
1631 struct multipath *m = ti->private;
1632 struct priority_group *pg;
1637 spin_lock_irqsave(&m->lock, flags);
1640 if (type == STATUSTYPE_INFO)
1641 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1642 atomic_read(&m->pg_init_count));
1644 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1645 (m->pg_init_retries > 0) * 2 +
1646 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1647 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1648 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1650 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1651 DMEMIT("queue_if_no_path ");
1652 if (m->pg_init_retries)
1653 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1654 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1655 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1656 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1657 DMEMIT("retain_attached_hw_handler ");
1658 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1659 switch(m->queue_mode) {
1660 case DM_TYPE_BIO_BASED:
1661 DMEMIT("queue_mode bio ");
1663 case DM_TYPE_MQ_REQUEST_BASED:
1664 DMEMIT("queue_mode mq ");
1673 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1676 DMEMIT("1 %s ", m->hw_handler_name);
1678 DMEMIT("%u ", m->nr_priority_groups);
1681 pg_num = m->next_pg->pg_num;
1682 else if (m->current_pg)
1683 pg_num = m->current_pg->pg_num;
1685 pg_num = (m->nr_priority_groups ? 1 : 0);
1687 DMEMIT("%u ", pg_num);
1690 case STATUSTYPE_INFO:
1691 list_for_each_entry(pg, &m->priority_groups, list) {
1693 state = 'D'; /* Disabled */
1694 else if (pg == m->current_pg)
1695 state = 'A'; /* Currently Active */
1697 state = 'E'; /* Enabled */
1699 DMEMIT("%c ", state);
1701 if (pg->ps.type->status)
1702 sz += pg->ps.type->status(&pg->ps, NULL, type,
1708 DMEMIT("%u %u ", pg->nr_pgpaths,
1709 pg->ps.type->info_args);
1711 list_for_each_entry(p, &pg->pgpaths, list) {
1712 DMEMIT("%s %s %u ", p->path.dev->name,
1713 p->is_active ? "A" : "F",
1715 if (pg->ps.type->status)
1716 sz += pg->ps.type->status(&pg->ps,
1717 &p->path, type, result + sz,
1723 case STATUSTYPE_TABLE:
1724 list_for_each_entry(pg, &m->priority_groups, list) {
1725 DMEMIT("%s ", pg->ps.type->name);
1727 if (pg->ps.type->status)
1728 sz += pg->ps.type->status(&pg->ps, NULL, type,
1734 DMEMIT("%u %u ", pg->nr_pgpaths,
1735 pg->ps.type->table_args);
1737 list_for_each_entry(p, &pg->pgpaths, list) {
1738 DMEMIT("%s ", p->path.dev->name);
1739 if (pg->ps.type->status)
1740 sz += pg->ps.type->status(&pg->ps,
1741 &p->path, type, result + sz,
1748 spin_unlock_irqrestore(&m->lock, flags);
1751 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1755 struct multipath *m = ti->private;
1758 mutex_lock(&m->work_mutex);
1760 if (dm_suspended(ti)) {
1766 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1767 r = queue_if_no_path(m, true, false);
1769 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1770 r = queue_if_no_path(m, false, false);
1776 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1780 if (!strcasecmp(argv[0], "disable_group")) {
1781 r = bypass_pg_num(m, argv[1], true);
1783 } else if (!strcasecmp(argv[0], "enable_group")) {
1784 r = bypass_pg_num(m, argv[1], false);
1786 } else if (!strcasecmp(argv[0], "switch_group")) {
1787 r = switch_pg_num(m, argv[1]);
1789 } else if (!strcasecmp(argv[0], "reinstate_path"))
1790 action = reinstate_path;
1791 else if (!strcasecmp(argv[0], "fail_path"))
1794 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1798 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1800 DMWARN("message: error getting device %s",
1805 r = action_dev(m, dev, action);
1807 dm_put_device(ti, dev);
1810 mutex_unlock(&m->work_mutex);
1814 static int multipath_prepare_ioctl(struct dm_target *ti,
1815 struct block_device **bdev, fmode_t *mode)
1817 struct multipath *m = ti->private;
1818 struct pgpath *current_pgpath;
1821 current_pgpath = READ_ONCE(m->current_pgpath);
1822 if (!current_pgpath)
1823 current_pgpath = choose_pgpath(m, 0);
1825 if (current_pgpath) {
1826 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1827 *bdev = current_pgpath->path.dev->bdev;
1828 *mode = current_pgpath->path.dev->mode;
1831 /* pg_init has not started or completed */
1835 /* No path is available */
1836 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1842 if (r == -ENOTCONN) {
1843 if (!READ_ONCE(m->current_pg)) {
1844 /* Path status changed, redo selection */
1845 (void) choose_pgpath(m, 0);
1847 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1848 pg_init_all_paths(m);
1849 dm_table_run_md_queue_async(m->ti->table);
1850 process_queued_io_list(m);
1854 * Only pass ioctls through if the device sizes match exactly.
1856 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1861 static int multipath_iterate_devices(struct dm_target *ti,
1862 iterate_devices_callout_fn fn, void *data)
1864 struct multipath *m = ti->private;
1865 struct priority_group *pg;
1869 list_for_each_entry(pg, &m->priority_groups, list) {
1870 list_for_each_entry(p, &pg->pgpaths, list) {
1871 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1881 static int pgpath_busy(struct pgpath *pgpath)
1883 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1885 return blk_lld_busy(q);
1889 * We return "busy", only when we can map I/Os but underlying devices
1890 * are busy (so even if we map I/Os now, the I/Os will wait on
1891 * the underlying queue).
1892 * In other words, if we want to kill I/Os or queue them inside us
1893 * due to map unavailability, we don't return "busy". Otherwise,
1894 * dm core won't give us the I/Os and we can't do what we want.
1896 static int multipath_busy(struct dm_target *ti)
1898 bool busy = false, has_active = false;
1899 struct multipath *m = ti->private;
1900 struct priority_group *pg, *next_pg;
1901 struct pgpath *pgpath;
1903 /* pg_init in progress */
1904 if (atomic_read(&m->pg_init_in_progress))
1907 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1908 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1909 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1911 /* Guess which priority_group will be used at next mapping time */
1912 pg = READ_ONCE(m->current_pg);
1913 next_pg = READ_ONCE(m->next_pg);
1914 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1919 * We don't know which pg will be used at next mapping time.
1920 * We don't call choose_pgpath() here to avoid to trigger
1921 * pg_init just by busy checking.
1922 * So we don't know whether underlying devices we will be using
1923 * at next mapping time are busy or not. Just try mapping.
1929 * If there is one non-busy active path at least, the path selector
1930 * will be able to select it. So we consider such a pg as not busy.
1933 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1934 if (pgpath->is_active) {
1936 if (!pgpath_busy(pgpath)) {
1945 * No active path in this pg, so this pg won't be used and
1946 * the current_pg will be changed at next mapping time.
1947 * We need to try mapping to determine it.
1955 /*-----------------------------------------------------------------
1957 *---------------------------------------------------------------*/
1958 static struct target_type multipath_target = {
1959 .name = "multipath",
1960 .version = {1, 12, 0},
1961 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1962 .module = THIS_MODULE,
1963 .ctr = multipath_ctr,
1964 .dtr = multipath_dtr,
1965 .clone_and_map_rq = multipath_clone_and_map,
1966 .release_clone_rq = multipath_release_clone,
1967 .rq_end_io = multipath_end_io,
1968 .map = multipath_map_bio,
1969 .end_io = multipath_end_io_bio,
1970 .presuspend = multipath_presuspend,
1971 .postsuspend = multipath_postsuspend,
1972 .resume = multipath_resume,
1973 .status = multipath_status,
1974 .message = multipath_message,
1975 .prepare_ioctl = multipath_prepare_ioctl,
1976 .iterate_devices = multipath_iterate_devices,
1977 .busy = multipath_busy,
1980 static int __init dm_multipath_init(void)
1984 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1986 DMERR("failed to create workqueue kmpathd");
1988 goto bad_alloc_kmultipathd;
1992 * A separate workqueue is used to handle the device handlers
1993 * to avoid overloading existing workqueue. Overloading the
1994 * old workqueue would also create a bottleneck in the
1995 * path of the storage hardware device activation.
1997 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1999 if (!kmpath_handlerd) {
2000 DMERR("failed to create workqueue kmpath_handlerd");
2002 goto bad_alloc_kmpath_handlerd;
2005 r = dm_register_target(&multipath_target);
2007 DMERR("request-based register failed %d", r);
2009 goto bad_register_target;
2014 bad_register_target:
2015 destroy_workqueue(kmpath_handlerd);
2016 bad_alloc_kmpath_handlerd:
2017 destroy_workqueue(kmultipathd);
2018 bad_alloc_kmultipathd:
2022 static void __exit dm_multipath_exit(void)
2024 destroy_workqueue(kmpath_handlerd);
2025 destroy_workqueue(kmultipathd);
2027 dm_unregister_target(&multipath_target);
2030 module_init(dm_multipath_init);
2031 module_exit(dm_multipath_exit);
2033 MODULE_DESCRIPTION(DM_NAME " multipath target");
2034 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2035 MODULE_LICENSE("GPL");