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 */
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_path(struct work_struct *work);
115 static void process_queued_bios(struct work_struct *work);
117 /*-----------------------------------------------
118 * Multipath state flags.
119 *-----------------------------------------------*/
121 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
122 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
123 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
124 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
125 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
126 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
127 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
129 /*-----------------------------------------------
130 * Allocation routines
131 *-----------------------------------------------*/
133 static struct pgpath *alloc_pgpath(void)
135 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
138 pgpath->is_active = true;
139 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
145 static void free_pgpath(struct pgpath *pgpath)
150 static struct priority_group *alloc_priority_group(void)
152 struct priority_group *pg;
154 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
157 INIT_LIST_HEAD(&pg->pgpaths);
162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
164 struct pgpath *pgpath, *tmp;
166 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
167 list_del(&pgpath->list);
168 dm_put_device(ti, pgpath->path.dev);
173 static void free_priority_group(struct priority_group *pg,
174 struct dm_target *ti)
176 struct path_selector *ps = &pg->ps;
179 ps->type->destroy(ps);
180 dm_put_path_selector(ps->type);
183 free_pgpaths(&pg->pgpaths, ti);
187 static struct multipath *alloc_multipath(struct dm_target *ti)
191 m = kzalloc(sizeof(*m), GFP_KERNEL);
193 INIT_LIST_HEAD(&m->priority_groups);
194 spin_lock_init(&m->lock);
195 set_bit(MPATHF_QUEUE_IO, &m->flags);
196 atomic_set(&m->nr_valid_paths, 0);
197 atomic_set(&m->pg_init_in_progress, 0);
198 atomic_set(&m->pg_init_count, 0);
199 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
200 INIT_WORK(&m->trigger_event, trigger_event);
201 init_waitqueue_head(&m->pg_init_wait);
202 mutex_init(&m->work_mutex);
204 m->queue_mode = DM_TYPE_NONE;
213 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
215 if (m->queue_mode == DM_TYPE_NONE) {
217 * Default to request-based.
219 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
220 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
222 m->queue_mode = DM_TYPE_REQUEST_BASED;
223 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
224 INIT_WORK(&m->process_queued_bios, process_queued_bios);
226 * bio-based doesn't support any direct scsi_dh management;
227 * it just discovers if a scsi_dh is attached.
229 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
232 dm_table_set_type(ti->table, m->queue_mode);
237 static void free_multipath(struct multipath *m)
239 struct priority_group *pg, *tmp;
241 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
243 free_priority_group(pg, m->ti);
246 kfree(m->hw_handler_name);
247 kfree(m->hw_handler_params);
251 static struct dm_mpath_io *get_mpio(union map_info *info)
256 static size_t multipath_per_bio_data_size(void)
258 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
261 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
263 return dm_per_bio_data(bio, multipath_per_bio_data_size());
266 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
268 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
269 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
270 void *bio_details = mpio + 1;
275 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
276 struct dm_bio_details **bio_details_p)
278 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
279 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
281 memset(mpio, 0, sizeof(*mpio));
282 memset(bio_details, 0, sizeof(*bio_details));
283 dm_bio_record(bio_details, bio);
288 *bio_details_p = bio_details;
291 /*-----------------------------------------------
293 *-----------------------------------------------*/
295 static int __pg_init_all_paths(struct multipath *m)
297 struct pgpath *pgpath;
298 unsigned long pg_init_delay = 0;
300 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
303 atomic_inc(&m->pg_init_count);
304 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
306 /* Check here to reset pg_init_required */
310 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
311 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
312 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
313 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
314 /* Skip failed paths */
315 if (!pgpath->is_active)
317 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
319 atomic_inc(&m->pg_init_in_progress);
321 return atomic_read(&m->pg_init_in_progress);
324 static void pg_init_all_paths(struct multipath *m)
328 spin_lock_irqsave(&m->lock, flags);
329 __pg_init_all_paths(m);
330 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(lockless_dereference(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 (lockless_dereference(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 = lockless_dereference(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 * Check whether bios must be queued in the device-mapper core rather
440 * than here in the target.
442 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
443 * same value then we are not between multipath_presuspend()
444 * and multipath_resume() calls and we have no need to check
445 * for the DMF_NOFLUSH_SUSPENDING flag.
447 static bool __must_push_back(struct multipath *m)
449 return ((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));
454 static bool must_push_back_rq(struct multipath *m)
459 spin_lock_irqsave(&m->lock, flags);
460 r = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
461 __must_push_back(m));
462 spin_unlock_irqrestore(&m->lock, flags);
467 static bool must_push_back_bio(struct multipath *m)
472 spin_lock_irqsave(&m->lock, flags);
473 r = __must_push_back(m);
474 spin_unlock_irqrestore(&m->lock, flags);
480 * Map cloned requests (request-based multipath)
482 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
483 union map_info *map_context,
484 struct request **__clone)
486 struct multipath *m = ti->private;
487 int r = DM_MAPIO_REQUEUE;
488 size_t nr_bytes = blk_rq_bytes(rq);
489 struct pgpath *pgpath;
490 struct block_device *bdev;
491 struct dm_mpath_io *mpio = get_mpio(map_context);
492 struct request *clone;
494 /* Do we need to select a new pgpath? */
495 pgpath = lockless_dereference(m->current_pgpath);
496 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
497 pgpath = choose_pgpath(m, nr_bytes);
500 if (must_push_back_rq(m))
501 return DM_MAPIO_DELAY_REQUEUE;
502 return -EIO; /* Failed */
503 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
504 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
505 pg_init_all_paths(m);
509 memset(mpio, 0, sizeof(*mpio));
510 mpio->pgpath = pgpath;
511 mpio->nr_bytes = nr_bytes;
513 bdev = pgpath->path.dev->bdev;
515 clone = blk_get_request(bdev_get_queue(bdev),
516 rq->cmd_flags | REQ_NOMERGE,
519 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
522 clone->bio = clone->biotail = NULL;
523 clone->rq_disk = bdev->bd_disk;
524 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
527 if (pgpath->pg->ps.type->start_io)
528 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
531 return DM_MAPIO_REMAPPED;
534 static void multipath_release_clone(struct request *clone)
536 blk_put_request(clone);
540 * Map cloned bios (bio-based multipath)
542 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
544 size_t nr_bytes = bio->bi_iter.bi_size;
545 struct pgpath *pgpath;
549 /* Do we need to select a new pgpath? */
550 pgpath = lockless_dereference(m->current_pgpath);
551 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
552 if (!pgpath || !queue_io)
553 pgpath = choose_pgpath(m, nr_bytes);
555 if ((pgpath && queue_io) ||
556 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
557 /* Queue for the daemon to resubmit */
558 spin_lock_irqsave(&m->lock, flags);
559 bio_list_add(&m->queued_bios, bio);
560 spin_unlock_irqrestore(&m->lock, flags);
561 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
562 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
563 pg_init_all_paths(m);
565 queue_work(kmultipathd, &m->process_queued_bios);
566 return DM_MAPIO_SUBMITTED;
570 if (!must_push_back_bio(m))
572 return DM_MAPIO_REQUEUE;
575 mpio->pgpath = pgpath;
576 mpio->nr_bytes = nr_bytes;
579 bio->bi_bdev = pgpath->path.dev->bdev;
580 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
582 if (pgpath->pg->ps.type->start_io)
583 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
586 return DM_MAPIO_REMAPPED;
589 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
591 struct multipath *m = ti->private;
592 struct dm_mpath_io *mpio = NULL;
594 multipath_init_per_bio_data(bio, &mpio, NULL);
596 return __multipath_map_bio(m, bio, mpio);
599 static void process_queued_io_list(struct multipath *m)
601 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
602 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
603 else if (m->queue_mode == DM_TYPE_BIO_BASED)
604 queue_work(kmultipathd, &m->process_queued_bios);
607 static void process_queued_bios(struct work_struct *work)
612 struct bio_list bios;
613 struct blk_plug plug;
614 struct multipath *m =
615 container_of(work, struct multipath, process_queued_bios);
617 bio_list_init(&bios);
619 spin_lock_irqsave(&m->lock, flags);
621 if (bio_list_empty(&m->queued_bios)) {
622 spin_unlock_irqrestore(&m->lock, flags);
626 bio_list_merge(&bios, &m->queued_bios);
627 bio_list_init(&m->queued_bios);
629 spin_unlock_irqrestore(&m->lock, flags);
631 blk_start_plug(&plug);
632 while ((bio = bio_list_pop(&bios))) {
633 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
634 if (r < 0 || r == DM_MAPIO_REQUEUE) {
637 } else if (r == DM_MAPIO_REMAPPED)
638 generic_make_request(bio);
640 blk_finish_plug(&plug);
644 * If we run out of usable paths, should we queue I/O or error it?
646 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
651 spin_lock_irqsave(&m->lock, flags);
653 if (save_old_value) {
654 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
655 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
657 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
659 if (queue_if_no_path)
660 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
662 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
664 if (queue_if_no_path)
665 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
667 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
669 spin_unlock_irqrestore(&m->lock, flags);
671 if (!queue_if_no_path) {
672 dm_table_run_md_queue_async(m->ti->table);
673 process_queued_io_list(m);
680 * An event is triggered whenever a path is taken out of use.
681 * Includes path failure and PG bypass.
683 static void trigger_event(struct work_struct *work)
685 struct multipath *m =
686 container_of(work, struct multipath, trigger_event);
688 dm_table_event(m->ti->table);
691 /*-----------------------------------------------------------------
692 * Constructor/argument parsing:
693 * <#multipath feature args> [<arg>]*
694 * <#hw_handler args> [hw_handler [<arg>]*]
696 * <initial priority group>
697 * [<selector> <#selector args> [<arg>]*
698 * <#paths> <#per-path selector args>
699 * [<path> [<arg>]* ]+ ]+
700 *---------------------------------------------------------------*/
701 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
702 struct dm_target *ti)
705 struct path_selector_type *pst;
708 static struct dm_arg _args[] = {
709 {0, 1024, "invalid number of path selector args"},
712 pst = dm_get_path_selector(dm_shift_arg(as));
714 ti->error = "unknown path selector type";
718 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
720 dm_put_path_selector(pst);
724 r = pst->create(&pg->ps, ps_argc, as->argv);
726 dm_put_path_selector(pst);
727 ti->error = "path selector constructor failed";
732 dm_consume_args(as, ps_argc);
737 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
738 struct dm_target *ti)
742 struct multipath *m = ti->private;
743 struct request_queue *q = NULL;
744 const char *attached_handler_name;
746 /* we need at least a path arg */
748 ti->error = "no device given";
749 return ERR_PTR(-EINVAL);
754 return ERR_PTR(-ENOMEM);
756 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
759 ti->error = "error getting device";
763 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
764 q = bdev_get_queue(p->path.dev->bdev);
766 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
768 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
769 if (attached_handler_name) {
771 * Clear any hw_handler_params associated with a
772 * handler that isn't already attached.
774 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
775 kfree(m->hw_handler_params);
776 m->hw_handler_params = NULL;
780 * Reset hw_handler_name to match the attached handler
782 * NB. This modifies the table line to show the actual
783 * handler instead of the original table passed in.
785 kfree(m->hw_handler_name);
786 m->hw_handler_name = attached_handler_name;
790 if (m->hw_handler_name) {
791 r = scsi_dh_attach(q, m->hw_handler_name);
793 char b[BDEVNAME_SIZE];
795 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
796 bdevname(p->path.dev->bdev, b));
800 ti->error = "error attaching hardware handler";
801 dm_put_device(ti, p->path.dev);
805 if (m->hw_handler_params) {
806 r = scsi_dh_set_params(q, m->hw_handler_params);
808 ti->error = "unable to set hardware "
809 "handler parameters";
810 dm_put_device(ti, p->path.dev);
816 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
818 dm_put_device(ti, p->path.dev);
829 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
832 static struct dm_arg _args[] = {
833 {1, 1024, "invalid number of paths"},
834 {0, 1024, "invalid number of selector args"}
838 unsigned i, nr_selector_args, nr_args;
839 struct priority_group *pg;
840 struct dm_target *ti = m->ti;
844 ti->error = "not enough priority group arguments";
845 return ERR_PTR(-EINVAL);
848 pg = alloc_priority_group();
850 ti->error = "couldn't allocate priority group";
851 return ERR_PTR(-ENOMEM);
855 r = parse_path_selector(as, pg, ti);
862 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
866 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
870 nr_args = 1 + nr_selector_args;
871 for (i = 0; i < pg->nr_pgpaths; i++) {
872 struct pgpath *pgpath;
873 struct dm_arg_set path_args;
875 if (as->argc < nr_args) {
876 ti->error = "not enough path parameters";
881 path_args.argc = nr_args;
882 path_args.argv = as->argv;
884 pgpath = parse_path(&path_args, &pg->ps, ti);
885 if (IS_ERR(pgpath)) {
891 list_add_tail(&pgpath->list, &pg->pgpaths);
892 dm_consume_args(as, nr_args);
898 free_priority_group(pg, ti);
902 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
906 struct dm_target *ti = m->ti;
908 static struct dm_arg _args[] = {
909 {0, 1024, "invalid number of hardware handler args"},
912 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
918 if (m->queue_mode == DM_TYPE_BIO_BASED) {
919 dm_consume_args(as, hw_argc);
920 DMERR("bio-based multipath doesn't allow hardware handler args");
924 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
925 if (!m->hw_handler_name)
932 for (i = 0; i <= hw_argc - 2; i++)
933 len += strlen(as->argv[i]) + 1;
934 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
936 ti->error = "memory allocation failed";
940 j = sprintf(p, "%d", hw_argc - 1);
941 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
942 j = sprintf(p, "%s", as->argv[i]);
944 dm_consume_args(as, hw_argc - 1);
948 kfree(m->hw_handler_name);
949 m->hw_handler_name = NULL;
953 static int parse_features(struct dm_arg_set *as, struct multipath *m)
957 struct dm_target *ti = m->ti;
958 const char *arg_name;
960 static struct dm_arg _args[] = {
961 {0, 8, "invalid number of feature args"},
962 {1, 50, "pg_init_retries must be between 1 and 50"},
963 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
966 r = dm_read_arg_group(_args, as, &argc, &ti->error);
974 arg_name = dm_shift_arg(as);
977 if (!strcasecmp(arg_name, "queue_if_no_path")) {
978 r = queue_if_no_path(m, true, false);
982 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
983 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
987 if (!strcasecmp(arg_name, "pg_init_retries") &&
989 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
994 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
996 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1001 if (!strcasecmp(arg_name, "queue_mode") &&
1003 const char *queue_mode_name = dm_shift_arg(as);
1005 if (!strcasecmp(queue_mode_name, "bio"))
1006 m->queue_mode = DM_TYPE_BIO_BASED;
1007 else if (!strcasecmp(queue_mode_name, "rq"))
1008 m->queue_mode = DM_TYPE_REQUEST_BASED;
1009 else if (!strcasecmp(queue_mode_name, "mq"))
1010 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1012 ti->error = "Unknown 'queue_mode' requested";
1019 ti->error = "Unrecognised multipath feature request";
1021 } while (argc && !r);
1026 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1028 /* target arguments */
1029 static struct dm_arg _args[] = {
1030 {0, 1024, "invalid number of priority groups"},
1031 {0, 1024, "invalid initial priority group number"},
1035 struct multipath *m;
1036 struct dm_arg_set as;
1037 unsigned pg_count = 0;
1038 unsigned next_pg_num;
1043 m = alloc_multipath(ti);
1045 ti->error = "can't allocate multipath";
1049 r = parse_features(&as, m);
1053 r = alloc_multipath_stage2(ti, m);
1057 r = parse_hw_handler(&as, m);
1061 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1065 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1069 if ((!m->nr_priority_groups && next_pg_num) ||
1070 (m->nr_priority_groups && !next_pg_num)) {
1071 ti->error = "invalid initial priority group";
1076 /* parse the priority groups */
1078 struct priority_group *pg;
1079 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1081 pg = parse_priority_group(&as, m);
1087 nr_valid_paths += pg->nr_pgpaths;
1088 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1090 list_add_tail(&pg->list, &m->priority_groups);
1092 pg->pg_num = pg_count;
1097 if (pg_count != m->nr_priority_groups) {
1098 ti->error = "priority group count mismatch";
1103 ti->num_flush_bios = 1;
1104 ti->num_discard_bios = 1;
1105 ti->num_write_same_bios = 1;
1106 ti->num_write_zeroes_bios = 1;
1107 if (m->queue_mode == DM_TYPE_BIO_BASED)
1108 ti->per_io_data_size = multipath_per_bio_data_size();
1110 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1119 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1124 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1126 if (!atomic_read(&m->pg_init_in_progress))
1131 finish_wait(&m->pg_init_wait, &wait);
1134 static void flush_multipath_work(struct multipath *m)
1136 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1137 smp_mb__after_atomic();
1139 flush_workqueue(kmpath_handlerd);
1140 multipath_wait_for_pg_init_completion(m);
1141 flush_workqueue(kmultipathd);
1142 flush_work(&m->trigger_event);
1144 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1145 smp_mb__after_atomic();
1148 static void multipath_dtr(struct dm_target *ti)
1150 struct multipath *m = ti->private;
1152 flush_multipath_work(m);
1157 * Take a path out of use.
1159 static int fail_path(struct pgpath *pgpath)
1161 unsigned long flags;
1162 struct multipath *m = pgpath->pg->m;
1164 spin_lock_irqsave(&m->lock, flags);
1166 if (!pgpath->is_active)
1169 DMWARN("Failing path %s.", pgpath->path.dev->name);
1171 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1172 pgpath->is_active = false;
1173 pgpath->fail_count++;
1175 atomic_dec(&m->nr_valid_paths);
1177 if (pgpath == m->current_pgpath)
1178 m->current_pgpath = NULL;
1180 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1181 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1183 schedule_work(&m->trigger_event);
1186 spin_unlock_irqrestore(&m->lock, flags);
1192 * Reinstate a previously-failed path
1194 static int reinstate_path(struct pgpath *pgpath)
1196 int r = 0, run_queue = 0;
1197 unsigned long flags;
1198 struct multipath *m = pgpath->pg->m;
1199 unsigned nr_valid_paths;
1201 spin_lock_irqsave(&m->lock, flags);
1203 if (pgpath->is_active)
1206 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1208 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1212 pgpath->is_active = true;
1214 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1215 if (nr_valid_paths == 1) {
1216 m->current_pgpath = NULL;
1218 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1219 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1220 atomic_inc(&m->pg_init_in_progress);
1223 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1224 pgpath->path.dev->name, nr_valid_paths);
1226 schedule_work(&m->trigger_event);
1229 spin_unlock_irqrestore(&m->lock, flags);
1231 dm_table_run_md_queue_async(m->ti->table);
1232 process_queued_io_list(m);
1239 * Fail or reinstate all paths that match the provided struct dm_dev.
1241 static int action_dev(struct multipath *m, struct dm_dev *dev,
1245 struct pgpath *pgpath;
1246 struct priority_group *pg;
1248 list_for_each_entry(pg, &m->priority_groups, list) {
1249 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1250 if (pgpath->path.dev == dev)
1259 * Temporarily try to avoid having to use the specified PG
1261 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1264 unsigned long flags;
1266 spin_lock_irqsave(&m->lock, flags);
1268 pg->bypassed = bypassed;
1269 m->current_pgpath = NULL;
1270 m->current_pg = NULL;
1272 spin_unlock_irqrestore(&m->lock, flags);
1274 schedule_work(&m->trigger_event);
1278 * Switch to using the specified PG from the next I/O that gets mapped
1280 static int switch_pg_num(struct multipath *m, const char *pgstr)
1282 struct priority_group *pg;
1284 unsigned long flags;
1287 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1288 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1289 DMWARN("invalid PG number supplied to switch_pg_num");
1293 spin_lock_irqsave(&m->lock, flags);
1294 list_for_each_entry(pg, &m->priority_groups, list) {
1295 pg->bypassed = false;
1299 m->current_pgpath = NULL;
1300 m->current_pg = NULL;
1303 spin_unlock_irqrestore(&m->lock, flags);
1305 schedule_work(&m->trigger_event);
1310 * Set/clear bypassed status of a PG.
1311 * PGs are numbered upwards from 1 in the order they were declared.
1313 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1315 struct priority_group *pg;
1319 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1320 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1321 DMWARN("invalid PG number supplied to bypass_pg");
1325 list_for_each_entry(pg, &m->priority_groups, list) {
1330 bypass_pg(m, pg, bypassed);
1335 * Should we retry pg_init immediately?
1337 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1339 unsigned long flags;
1340 bool limit_reached = false;
1342 spin_lock_irqsave(&m->lock, flags);
1344 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1345 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1346 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1348 limit_reached = true;
1350 spin_unlock_irqrestore(&m->lock, flags);
1352 return limit_reached;
1355 static void pg_init_done(void *data, int errors)
1357 struct pgpath *pgpath = data;
1358 struct priority_group *pg = pgpath->pg;
1359 struct multipath *m = pg->m;
1360 unsigned long flags;
1361 bool delay_retry = false;
1363 /* device or driver problems */
1368 if (!m->hw_handler_name) {
1372 DMERR("Could not failover the device: Handler scsi_dh_%s "
1373 "Error %d.", m->hw_handler_name, errors);
1375 * Fail path for now, so we do not ping pong
1379 case SCSI_DH_DEV_TEMP_BUSY:
1381 * Probably doing something like FW upgrade on the
1382 * controller so try the other pg.
1384 bypass_pg(m, pg, true);
1387 /* Wait before retrying. */
1389 case SCSI_DH_IMM_RETRY:
1390 case SCSI_DH_RES_TEMP_UNAVAIL:
1391 if (pg_init_limit_reached(m, pgpath))
1395 case SCSI_DH_DEV_OFFLINED:
1398 * We probably do not want to fail the path for a device
1399 * error, but this is what the old dm did. In future
1400 * patches we can do more advanced handling.
1405 spin_lock_irqsave(&m->lock, flags);
1407 if (pgpath == m->current_pgpath) {
1408 DMERR("Could not failover device. Error %d.", errors);
1409 m->current_pgpath = NULL;
1410 m->current_pg = NULL;
1412 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1413 pg->bypassed = false;
1415 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1416 /* Activations of other paths are still on going */
1419 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1421 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1423 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1425 if (__pg_init_all_paths(m))
1428 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1430 process_queued_io_list(m);
1433 * Wake up any thread waiting to suspend.
1435 wake_up(&m->pg_init_wait);
1438 spin_unlock_irqrestore(&m->lock, flags);
1441 static void activate_path(struct work_struct *work)
1443 struct pgpath *pgpath =
1444 container_of(work, struct pgpath, activate_path.work);
1445 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1447 if (pgpath->is_active && !blk_queue_dying(q))
1448 scsi_dh_activate(q, pg_init_done, pgpath);
1450 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1453 static int noretry_error(int error)
1458 * EBADE signals an reservation conflict.
1459 * We shouldn't fail the path here as we can communicate with
1460 * the target. We should failover to the next path, but in
1461 * doing so we might be causing a ping-pong between paths.
1462 * So just return the reservation conflict error.
1472 /* Anything else could be a path failure, so should be retried */
1479 static int do_end_io(struct multipath *m, struct request *clone,
1480 int error, struct dm_mpath_io *mpio)
1483 * We don't queue any clone request inside the multipath target
1484 * during end I/O handling, since those clone requests don't have
1485 * bio clones. If we queue them inside the multipath target,
1486 * we need to make bio clones, that requires memory allocation.
1487 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1488 * don't have bio clones.)
1489 * Instead of queueing the clone request here, we queue the original
1490 * request into dm core, which will remake a clone request and
1491 * clone bios for it and resubmit it later.
1493 int r = DM_ENDIO_REQUEUE;
1496 return 0; /* I/O complete */
1498 if (noretry_error(error))
1502 fail_path(mpio->pgpath);
1504 if (!atomic_read(&m->nr_valid_paths)) {
1505 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1506 if (!must_push_back_rq(m))
1514 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1515 int error, union map_info *map_context)
1517 struct multipath *m = ti->private;
1518 struct dm_mpath_io *mpio = get_mpio(map_context);
1519 struct pgpath *pgpath;
1520 struct path_selector *ps;
1525 r = do_end_io(m, clone, error, mpio);
1526 pgpath = mpio->pgpath;
1528 ps = &pgpath->pg->ps;
1529 if (ps->type->end_io)
1530 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1536 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1537 int error, struct dm_mpath_io *mpio)
1539 unsigned long flags;
1542 return 0; /* I/O complete */
1544 if (noretry_error(error))
1548 fail_path(mpio->pgpath);
1550 if (!atomic_read(&m->nr_valid_paths)) {
1551 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1552 if (!must_push_back_bio(m))
1554 return DM_ENDIO_REQUEUE;
1558 /* Queue for the daemon to resubmit */
1559 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1561 spin_lock_irqsave(&m->lock, flags);
1562 bio_list_add(&m->queued_bios, clone);
1563 spin_unlock_irqrestore(&m->lock, flags);
1564 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1565 queue_work(kmultipathd, &m->process_queued_bios);
1567 return DM_ENDIO_INCOMPLETE;
1570 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1572 struct multipath *m = ti->private;
1573 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1574 struct pgpath *pgpath;
1575 struct path_selector *ps;
1580 r = do_end_io_bio(m, clone, error, mpio);
1581 pgpath = mpio->pgpath;
1583 ps = &pgpath->pg->ps;
1584 if (ps->type->end_io)
1585 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1592 * Suspend can't complete until all the I/O is processed so if
1593 * the last path fails we must error any remaining I/O.
1594 * Note that if the freeze_bdev fails while suspending, the
1595 * queue_if_no_path state is lost - userspace should reset it.
1597 static void multipath_presuspend(struct dm_target *ti)
1599 struct multipath *m = ti->private;
1601 queue_if_no_path(m, false, true);
1604 static void multipath_postsuspend(struct dm_target *ti)
1606 struct multipath *m = ti->private;
1608 mutex_lock(&m->work_mutex);
1609 flush_multipath_work(m);
1610 mutex_unlock(&m->work_mutex);
1614 * Restore the queue_if_no_path setting.
1616 static void multipath_resume(struct dm_target *ti)
1618 struct multipath *m = ti->private;
1619 unsigned long flags;
1621 spin_lock_irqsave(&m->lock, flags);
1622 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1623 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1625 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1626 spin_unlock_irqrestore(&m->lock, flags);
1630 * Info output has the following format:
1631 * num_multipath_feature_args [multipath_feature_args]*
1632 * num_handler_status_args [handler_status_args]*
1633 * num_groups init_group_number
1634 * [A|D|E num_ps_status_args [ps_status_args]*
1635 * num_paths num_selector_args
1636 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1638 * Table output has the following format (identical to the constructor string):
1639 * num_feature_args [features_args]*
1640 * num_handler_args hw_handler [hw_handler_args]*
1641 * num_groups init_group_number
1642 * [priority selector-name num_ps_args [ps_args]*
1643 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1645 static void multipath_status(struct dm_target *ti, status_type_t type,
1646 unsigned status_flags, char *result, unsigned maxlen)
1649 unsigned long flags;
1650 struct multipath *m = ti->private;
1651 struct priority_group *pg;
1656 spin_lock_irqsave(&m->lock, flags);
1659 if (type == STATUSTYPE_INFO)
1660 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1661 atomic_read(&m->pg_init_count));
1663 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1664 (m->pg_init_retries > 0) * 2 +
1665 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1666 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1667 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1669 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1670 DMEMIT("queue_if_no_path ");
1671 if (m->pg_init_retries)
1672 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1673 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1674 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1675 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1676 DMEMIT("retain_attached_hw_handler ");
1677 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1678 switch(m->queue_mode) {
1679 case DM_TYPE_BIO_BASED:
1680 DMEMIT("queue_mode bio ");
1682 case DM_TYPE_MQ_REQUEST_BASED:
1683 DMEMIT("queue_mode mq ");
1689 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1692 DMEMIT("1 %s ", m->hw_handler_name);
1694 DMEMIT("%u ", m->nr_priority_groups);
1697 pg_num = m->next_pg->pg_num;
1698 else if (m->current_pg)
1699 pg_num = m->current_pg->pg_num;
1701 pg_num = (m->nr_priority_groups ? 1 : 0);
1703 DMEMIT("%u ", pg_num);
1706 case STATUSTYPE_INFO:
1707 list_for_each_entry(pg, &m->priority_groups, list) {
1709 state = 'D'; /* Disabled */
1710 else if (pg == m->current_pg)
1711 state = 'A'; /* Currently Active */
1713 state = 'E'; /* Enabled */
1715 DMEMIT("%c ", state);
1717 if (pg->ps.type->status)
1718 sz += pg->ps.type->status(&pg->ps, NULL, type,
1724 DMEMIT("%u %u ", pg->nr_pgpaths,
1725 pg->ps.type->info_args);
1727 list_for_each_entry(p, &pg->pgpaths, list) {
1728 DMEMIT("%s %s %u ", p->path.dev->name,
1729 p->is_active ? "A" : "F",
1731 if (pg->ps.type->status)
1732 sz += pg->ps.type->status(&pg->ps,
1733 &p->path, type, result + sz,
1739 case STATUSTYPE_TABLE:
1740 list_for_each_entry(pg, &m->priority_groups, list) {
1741 DMEMIT("%s ", pg->ps.type->name);
1743 if (pg->ps.type->status)
1744 sz += pg->ps.type->status(&pg->ps, NULL, type,
1750 DMEMIT("%u %u ", pg->nr_pgpaths,
1751 pg->ps.type->table_args);
1753 list_for_each_entry(p, &pg->pgpaths, list) {
1754 DMEMIT("%s ", p->path.dev->name);
1755 if (pg->ps.type->status)
1756 sz += pg->ps.type->status(&pg->ps,
1757 &p->path, type, result + sz,
1764 spin_unlock_irqrestore(&m->lock, flags);
1767 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1771 struct multipath *m = ti->private;
1774 mutex_lock(&m->work_mutex);
1776 if (dm_suspended(ti)) {
1782 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1783 r = queue_if_no_path(m, true, false);
1785 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1786 r = queue_if_no_path(m, false, false);
1792 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1796 if (!strcasecmp(argv[0], "disable_group")) {
1797 r = bypass_pg_num(m, argv[1], true);
1799 } else if (!strcasecmp(argv[0], "enable_group")) {
1800 r = bypass_pg_num(m, argv[1], false);
1802 } else if (!strcasecmp(argv[0], "switch_group")) {
1803 r = switch_pg_num(m, argv[1]);
1805 } else if (!strcasecmp(argv[0], "reinstate_path"))
1806 action = reinstate_path;
1807 else if (!strcasecmp(argv[0], "fail_path"))
1810 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1814 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1816 DMWARN("message: error getting device %s",
1821 r = action_dev(m, dev, action);
1823 dm_put_device(ti, dev);
1826 mutex_unlock(&m->work_mutex);
1830 static int multipath_prepare_ioctl(struct dm_target *ti,
1831 struct block_device **bdev, fmode_t *mode)
1833 struct multipath *m = ti->private;
1834 struct pgpath *current_pgpath;
1837 current_pgpath = lockless_dereference(m->current_pgpath);
1838 if (!current_pgpath)
1839 current_pgpath = choose_pgpath(m, 0);
1841 if (current_pgpath) {
1842 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1843 *bdev = current_pgpath->path.dev->bdev;
1844 *mode = current_pgpath->path.dev->mode;
1847 /* pg_init has not started or completed */
1851 /* No path is available */
1852 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1858 if (r == -ENOTCONN) {
1859 if (!lockless_dereference(m->current_pg)) {
1860 /* Path status changed, redo selection */
1861 (void) choose_pgpath(m, 0);
1863 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1864 pg_init_all_paths(m);
1865 dm_table_run_md_queue_async(m->ti->table);
1866 process_queued_io_list(m);
1870 * Only pass ioctls through if the device sizes match exactly.
1872 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1877 static int multipath_iterate_devices(struct dm_target *ti,
1878 iterate_devices_callout_fn fn, void *data)
1880 struct multipath *m = ti->private;
1881 struct priority_group *pg;
1885 list_for_each_entry(pg, &m->priority_groups, list) {
1886 list_for_each_entry(p, &pg->pgpaths, list) {
1887 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1897 static int pgpath_busy(struct pgpath *pgpath)
1899 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1901 return blk_lld_busy(q);
1905 * We return "busy", only when we can map I/Os but underlying devices
1906 * are busy (so even if we map I/Os now, the I/Os will wait on
1907 * the underlying queue).
1908 * In other words, if we want to kill I/Os or queue them inside us
1909 * due to map unavailability, we don't return "busy". Otherwise,
1910 * dm core won't give us the I/Os and we can't do what we want.
1912 static int multipath_busy(struct dm_target *ti)
1914 bool busy = false, has_active = false;
1915 struct multipath *m = ti->private;
1916 struct priority_group *pg, *next_pg;
1917 struct pgpath *pgpath;
1919 /* pg_init in progress */
1920 if (atomic_read(&m->pg_init_in_progress))
1923 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1924 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1925 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1927 /* Guess which priority_group will be used at next mapping time */
1928 pg = lockless_dereference(m->current_pg);
1929 next_pg = lockless_dereference(m->next_pg);
1930 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1935 * We don't know which pg will be used at next mapping time.
1936 * We don't call choose_pgpath() here to avoid to trigger
1937 * pg_init just by busy checking.
1938 * So we don't know whether underlying devices we will be using
1939 * at next mapping time are busy or not. Just try mapping.
1945 * If there is one non-busy active path at least, the path selector
1946 * will be able to select it. So we consider such a pg as not busy.
1949 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1950 if (pgpath->is_active) {
1952 if (!pgpath_busy(pgpath)) {
1961 * No active path in this pg, so this pg won't be used and
1962 * the current_pg will be changed at next mapping time.
1963 * We need to try mapping to determine it.
1971 /*-----------------------------------------------------------------
1973 *---------------------------------------------------------------*/
1974 static struct target_type multipath_target = {
1975 .name = "multipath",
1976 .version = {1, 12, 0},
1977 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1978 .module = THIS_MODULE,
1979 .ctr = multipath_ctr,
1980 .dtr = multipath_dtr,
1981 .clone_and_map_rq = multipath_clone_and_map,
1982 .release_clone_rq = multipath_release_clone,
1983 .rq_end_io = multipath_end_io,
1984 .map = multipath_map_bio,
1985 .end_io = multipath_end_io_bio,
1986 .presuspend = multipath_presuspend,
1987 .postsuspend = multipath_postsuspend,
1988 .resume = multipath_resume,
1989 .status = multipath_status,
1990 .message = multipath_message,
1991 .prepare_ioctl = multipath_prepare_ioctl,
1992 .iterate_devices = multipath_iterate_devices,
1993 .busy = multipath_busy,
1996 static int __init dm_multipath_init(void)
2000 r = dm_register_target(&multipath_target);
2002 DMERR("request-based register failed %d", r);
2004 goto bad_register_target;
2007 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2009 DMERR("failed to create workqueue kmpathd");
2011 goto bad_alloc_kmultipathd;
2015 * A separate workqueue is used to handle the device handlers
2016 * to avoid overloading existing workqueue. Overloading the
2017 * old workqueue would also create a bottleneck in the
2018 * path of the storage hardware device activation.
2020 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2022 if (!kmpath_handlerd) {
2023 DMERR("failed to create workqueue kmpath_handlerd");
2025 goto bad_alloc_kmpath_handlerd;
2030 bad_alloc_kmpath_handlerd:
2031 destroy_workqueue(kmultipathd);
2032 bad_alloc_kmultipathd:
2033 dm_unregister_target(&multipath_target);
2034 bad_register_target:
2038 static void __exit dm_multipath_exit(void)
2040 destroy_workqueue(kmpath_handlerd);
2041 destroy_workqueue(kmultipathd);
2043 dm_unregister_target(&multipath_target);
2046 module_init(dm_multipath_init);
2047 module_exit(dm_multipath_exit);
2049 MODULE_DESCRIPTION(DM_NAME " multipath target");
2050 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2051 MODULE_LICENSE("GPL");