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 unsigned long flags; /* Multipath state flags */
70 enum dm_queue_mode queue_mode;
72 struct pgpath *current_pgpath;
73 struct priority_group *current_pg;
74 struct priority_group *next_pg; /* Switch to this PG if set */
76 atomic_t nr_valid_paths; /* Total number of usable paths */
77 unsigned nr_priority_groups;
78 struct list_head priority_groups;
80 const char *hw_handler_name;
81 char *hw_handler_params;
82 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
83 unsigned pg_init_retries; /* Number of times to retry pg_init */
84 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
85 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
86 atomic_t pg_init_count; /* Number of times pg_init called */
88 struct mutex work_mutex;
89 struct work_struct trigger_event;
92 struct work_struct process_queued_bios;
93 struct bio_list queued_bios;
97 * Context information attached to each io we process.
100 struct pgpath *pgpath;
104 typedef int (*action_fn) (struct pgpath *pgpath);
106 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
107 static void trigger_event(struct work_struct *work);
108 static void activate_or_offline_path(struct pgpath *pgpath);
109 static void activate_path_work(struct work_struct *work);
110 static void process_queued_bios(struct work_struct *work);
112 /*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
116 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
117 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
120 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
121 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
122 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
124 /*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
128 static struct pgpath *alloc_pgpath(void)
130 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
135 pgpath->is_active = true;
140 static void free_pgpath(struct pgpath *pgpath)
145 static struct priority_group *alloc_priority_group(void)
147 struct priority_group *pg;
149 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
152 INIT_LIST_HEAD(&pg->pgpaths);
157 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
159 struct pgpath *pgpath, *tmp;
161 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
162 list_del(&pgpath->list);
163 dm_put_device(ti, pgpath->path.dev);
168 static void free_priority_group(struct priority_group *pg,
169 struct dm_target *ti)
171 struct path_selector *ps = &pg->ps;
174 ps->type->destroy(ps);
175 dm_put_path_selector(ps->type);
178 free_pgpaths(&pg->pgpaths, ti);
182 static struct multipath *alloc_multipath(struct dm_target *ti)
186 m = kzalloc(sizeof(*m), GFP_KERNEL);
188 INIT_LIST_HEAD(&m->priority_groups);
189 spin_lock_init(&m->lock);
190 atomic_set(&m->nr_valid_paths, 0);
191 INIT_WORK(&m->trigger_event, trigger_event);
192 mutex_init(&m->work_mutex);
194 m->queue_mode = DM_TYPE_NONE;
203 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
205 if (m->queue_mode == DM_TYPE_NONE) {
207 * Default to request-based.
209 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
210 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
212 m->queue_mode = DM_TYPE_REQUEST_BASED;
214 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
215 INIT_WORK(&m->process_queued_bios, process_queued_bios);
217 * bio-based doesn't support any direct scsi_dh management;
218 * it just discovers if a scsi_dh is attached.
220 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
223 dm_table_set_type(ti->table, m->queue_mode);
226 * Init fields that are only used when a scsi_dh is attached
227 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
229 set_bit(MPATHF_QUEUE_IO, &m->flags);
230 atomic_set(&m->pg_init_in_progress, 0);
231 atomic_set(&m->pg_init_count, 0);
232 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
233 init_waitqueue_head(&m->pg_init_wait);
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);
249 mutex_destroy(&m->work_mutex);
253 static struct dm_mpath_io *get_mpio(union map_info *info)
258 static size_t multipath_per_bio_data_size(void)
260 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
263 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
265 return dm_per_bio_data(bio, multipath_per_bio_data_size());
268 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
270 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
271 void *bio_details = mpio + 1;
275 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
277 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
278 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
280 mpio->nr_bytes = bio->bi_iter.bi_size;
284 dm_bio_record(bio_details, bio);
287 /*-----------------------------------------------
289 *-----------------------------------------------*/
291 static int __pg_init_all_paths(struct multipath *m)
293 struct pgpath *pgpath;
294 unsigned long pg_init_delay = 0;
296 lockdep_assert_held(&m->lock);
298 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
301 atomic_inc(&m->pg_init_count);
302 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
304 /* Check here to reset pg_init_required */
308 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
309 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
310 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
311 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
312 /* Skip failed paths */
313 if (!pgpath->is_active)
315 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
317 atomic_inc(&m->pg_init_in_progress);
319 return atomic_read(&m->pg_init_in_progress);
322 static int pg_init_all_paths(struct multipath *m)
327 spin_lock_irqsave(&m->lock, flags);
328 ret = __pg_init_all_paths(m);
329 spin_unlock_irqrestore(&m->lock, flags);
334 static void __switch_pg(struct multipath *m, struct priority_group *pg)
338 /* Must we initialise the PG first, and queue I/O till it's ready? */
339 if (m->hw_handler_name) {
340 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
341 set_bit(MPATHF_QUEUE_IO, &m->flags);
343 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
344 clear_bit(MPATHF_QUEUE_IO, &m->flags);
347 atomic_set(&m->pg_init_count, 0);
350 static struct pgpath *choose_path_in_pg(struct multipath *m,
351 struct priority_group *pg,
355 struct dm_path *path;
356 struct pgpath *pgpath;
358 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
360 return ERR_PTR(-ENXIO);
362 pgpath = path_to_pgpath(path);
364 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
365 /* Only update current_pgpath if pg changed */
366 spin_lock_irqsave(&m->lock, flags);
367 m->current_pgpath = pgpath;
369 spin_unlock_irqrestore(&m->lock, flags);
375 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
378 struct priority_group *pg;
379 struct pgpath *pgpath;
380 unsigned bypassed = 1;
382 if (!atomic_read(&m->nr_valid_paths)) {
383 clear_bit(MPATHF_QUEUE_IO, &m->flags);
387 /* Were we instructed to switch PG? */
388 if (READ_ONCE(m->next_pg)) {
389 spin_lock_irqsave(&m->lock, flags);
392 spin_unlock_irqrestore(&m->lock, flags);
393 goto check_current_pg;
396 spin_unlock_irqrestore(&m->lock, flags);
397 pgpath = choose_path_in_pg(m, pg, nr_bytes);
398 if (!IS_ERR_OR_NULL(pgpath))
402 /* Don't change PG until it has no remaining paths */
404 pg = READ_ONCE(m->current_pg);
406 pgpath = choose_path_in_pg(m, pg, nr_bytes);
407 if (!IS_ERR_OR_NULL(pgpath))
412 * Loop through priority groups until we find a valid path.
413 * First time we skip PGs marked 'bypassed'.
414 * Second time we only try the ones we skipped, but set
415 * pg_init_delay_retry so we do not hammer controllers.
418 list_for_each_entry(pg, &m->priority_groups, list) {
419 if (pg->bypassed == !!bypassed)
421 pgpath = choose_path_in_pg(m, pg, nr_bytes);
422 if (!IS_ERR_OR_NULL(pgpath)) {
424 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
428 } while (bypassed--);
431 spin_lock_irqsave(&m->lock, flags);
432 m->current_pgpath = NULL;
433 m->current_pg = NULL;
434 spin_unlock_irqrestore(&m->lock, flags);
440 * dm_report_EIO() is a macro instead of a function to make pr_debug()
441 * report the function name and line number of the function from which
442 * it has been invoked.
444 #define dm_report_EIO(m) \
446 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
448 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
449 dm_device_name(md), \
450 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
451 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
452 dm_noflush_suspending((m)->ti)); \
456 * Check whether bios must be queued in the device-mapper core rather
457 * than here in the target.
459 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
460 * the same value then we are not between multipath_presuspend()
461 * and multipath_resume() calls and we have no need to check
462 * for the DMF_NOFLUSH_SUSPENDING flag.
464 static bool __must_push_back(struct multipath *m, unsigned long flags)
466 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
467 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
468 dm_noflush_suspending(m->ti));
472 * Following functions use READ_ONCE to get atomic access to
473 * all m->flags to avoid taking spinlock
475 static bool must_push_back_rq(struct multipath *m)
477 unsigned long flags = READ_ONCE(m->flags);
478 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
481 static bool must_push_back_bio(struct multipath *m)
483 unsigned long flags = READ_ONCE(m->flags);
484 return __must_push_back(m, flags);
488 * Map cloned requests (request-based multipath)
490 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
491 union map_info *map_context,
492 struct request **__clone)
494 struct multipath *m = ti->private;
495 size_t nr_bytes = blk_rq_bytes(rq);
496 struct pgpath *pgpath;
497 struct block_device *bdev;
498 struct dm_mpath_io *mpio = get_mpio(map_context);
499 struct request_queue *q;
500 struct request *clone;
502 /* Do we need to select a new pgpath? */
503 pgpath = READ_ONCE(m->current_pgpath);
504 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
505 pgpath = choose_pgpath(m, nr_bytes);
508 if (must_push_back_rq(m))
509 return DM_MAPIO_DELAY_REQUEUE;
510 dm_report_EIO(m); /* Failed */
511 return DM_MAPIO_KILL;
512 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
513 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
514 pg_init_all_paths(m);
515 return DM_MAPIO_DELAY_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 if (blk_queue_dying(q)) {
527 atomic_inc(&m->pg_init_in_progress);
528 activate_or_offline_path(pgpath);
529 return DM_MAPIO_DELAY_REQUEUE;
533 * blk-mq's SCHED_RESTART can cover this requeue, so we
534 * needn't deal with it by DELAY_REQUEUE. More importantly,
535 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
536 * get the queue busy feedback (via BLK_STS_RESOURCE),
537 * otherwise I/O merging can suffer.
540 return DM_MAPIO_REQUEUE;
542 return DM_MAPIO_DELAY_REQUEUE;
544 clone->bio = clone->biotail = NULL;
545 clone->rq_disk = bdev->bd_disk;
546 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
549 if (pgpath->pg->ps.type->start_io)
550 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
553 return DM_MAPIO_REMAPPED;
556 static void multipath_release_clone(struct request *clone)
558 blk_put_request(clone);
562 * Map cloned bios (bio-based multipath)
565 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
567 struct pgpath *pgpath;
571 /* Do we need to select a new pgpath? */
572 pgpath = READ_ONCE(m->current_pgpath);
573 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
574 if (!pgpath || !queue_io)
575 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
577 if ((pgpath && queue_io) ||
578 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
579 /* Queue for the daemon to resubmit */
580 spin_lock_irqsave(&m->lock, flags);
581 bio_list_add(&m->queued_bios, bio);
582 spin_unlock_irqrestore(&m->lock, flags);
584 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
585 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
586 pg_init_all_paths(m);
588 queue_work(kmultipathd, &m->process_queued_bios);
590 return ERR_PTR(-EAGAIN);
596 static struct pgpath *__map_bio_fast(struct multipath *m, struct bio *bio)
598 struct pgpath *pgpath;
601 /* Do we need to select a new pgpath? */
603 * FIXME: currently only switching path if no path (due to failure, etc)
604 * - which negates the point of using a path selector
606 pgpath = READ_ONCE(m->current_pgpath);
608 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
611 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
612 /* Queue for the daemon to resubmit */
613 spin_lock_irqsave(&m->lock, flags);
614 bio_list_add(&m->queued_bios, bio);
615 spin_unlock_irqrestore(&m->lock, flags);
616 queue_work(kmultipathd, &m->process_queued_bios);
618 return ERR_PTR(-EAGAIN);
626 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
627 struct dm_mpath_io *mpio)
629 struct pgpath *pgpath;
631 if (!m->hw_handler_name)
632 pgpath = __map_bio_fast(m, bio);
634 pgpath = __map_bio(m, bio);
637 return DM_MAPIO_SUBMITTED;
640 if (must_push_back_bio(m))
641 return DM_MAPIO_REQUEUE;
643 return DM_MAPIO_KILL;
646 mpio->pgpath = pgpath;
649 bio_set_dev(bio, pgpath->path.dev->bdev);
650 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
652 if (pgpath->pg->ps.type->start_io)
653 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
656 return DM_MAPIO_REMAPPED;
659 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
661 struct multipath *m = ti->private;
662 struct dm_mpath_io *mpio = NULL;
664 multipath_init_per_bio_data(bio, &mpio);
665 return __multipath_map_bio(m, bio, mpio);
668 static void process_queued_io_list(struct multipath *m)
670 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
671 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
672 else if (m->queue_mode == DM_TYPE_BIO_BASED)
673 queue_work(kmultipathd, &m->process_queued_bios);
676 static void process_queued_bios(struct work_struct *work)
681 struct bio_list bios;
682 struct blk_plug plug;
683 struct multipath *m =
684 container_of(work, struct multipath, process_queued_bios);
686 bio_list_init(&bios);
688 spin_lock_irqsave(&m->lock, flags);
690 if (bio_list_empty(&m->queued_bios)) {
691 spin_unlock_irqrestore(&m->lock, flags);
695 bio_list_merge(&bios, &m->queued_bios);
696 bio_list_init(&m->queued_bios);
698 spin_unlock_irqrestore(&m->lock, flags);
700 blk_start_plug(&plug);
701 while ((bio = bio_list_pop(&bios))) {
702 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
703 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
704 r = __multipath_map_bio(m, bio, mpio);
707 bio->bi_status = BLK_STS_IOERR;
710 case DM_MAPIO_REQUEUE:
711 bio->bi_status = BLK_STS_DM_REQUEUE;
714 case DM_MAPIO_REMAPPED:
715 generic_make_request(bio);
717 case DM_MAPIO_SUBMITTED:
720 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
723 blk_finish_plug(&plug);
727 * If we run out of usable paths, should we queue I/O or error it?
729 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
734 spin_lock_irqsave(&m->lock, flags);
735 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
736 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
737 (!save_old_value && queue_if_no_path));
738 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
739 spin_unlock_irqrestore(&m->lock, flags);
741 if (!queue_if_no_path) {
742 dm_table_run_md_queue_async(m->ti->table);
743 process_queued_io_list(m);
750 * An event is triggered whenever a path is taken out of use.
751 * Includes path failure and PG bypass.
753 static void trigger_event(struct work_struct *work)
755 struct multipath *m =
756 container_of(work, struct multipath, trigger_event);
758 dm_table_event(m->ti->table);
761 /*-----------------------------------------------------------------
762 * Constructor/argument parsing:
763 * <#multipath feature args> [<arg>]*
764 * <#hw_handler args> [hw_handler [<arg>]*]
766 * <initial priority group>
767 * [<selector> <#selector args> [<arg>]*
768 * <#paths> <#per-path selector args>
769 * [<path> [<arg>]* ]+ ]+
770 *---------------------------------------------------------------*/
771 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
772 struct dm_target *ti)
775 struct path_selector_type *pst;
778 static const struct dm_arg _args[] = {
779 {0, 1024, "invalid number of path selector args"},
782 pst = dm_get_path_selector(dm_shift_arg(as));
784 ti->error = "unknown path selector type";
788 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
790 dm_put_path_selector(pst);
794 r = pst->create(&pg->ps, ps_argc, as->argv);
796 dm_put_path_selector(pst);
797 ti->error = "path selector constructor failed";
802 dm_consume_args(as, ps_argc);
807 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
808 const char *attached_handler_name, char **error)
810 struct request_queue *q = bdev_get_queue(bdev);
813 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
815 if (attached_handler_name) {
817 * Clear any hw_handler_params associated with a
818 * handler that isn't already attached.
820 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
821 kfree(m->hw_handler_params);
822 m->hw_handler_params = NULL;
826 * Reset hw_handler_name to match the attached handler
828 * NB. This modifies the table line to show the actual
829 * handler instead of the original table passed in.
831 kfree(m->hw_handler_name);
832 m->hw_handler_name = attached_handler_name;
836 if (m->hw_handler_name) {
837 r = scsi_dh_attach(q, m->hw_handler_name);
839 char b[BDEVNAME_SIZE];
841 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
846 *error = "error attaching hardware handler";
850 if (m->hw_handler_params) {
851 r = scsi_dh_set_params(q, m->hw_handler_params);
853 *error = "unable to set hardware handler parameters";
862 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
863 struct dm_target *ti)
867 struct multipath *m = ti->private;
868 struct request_queue *q;
869 const char *attached_handler_name;
871 /* we need at least a path arg */
873 ti->error = "no device given";
874 return ERR_PTR(-EINVAL);
879 return ERR_PTR(-ENOMEM);
881 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
884 ti->error = "error getting device";
888 q = bdev_get_queue(p->path.dev->bdev);
889 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
890 if (attached_handler_name || m->hw_handler_name) {
891 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
892 r = setup_scsi_dh(p->path.dev->bdev, m, attached_handler_name, &ti->error);
894 dm_put_device(ti, p->path.dev);
899 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
901 dm_put_device(ti, p->path.dev);
911 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
914 static const struct dm_arg _args[] = {
915 {1, 1024, "invalid number of paths"},
916 {0, 1024, "invalid number of selector args"}
920 unsigned i, nr_selector_args, nr_args;
921 struct priority_group *pg;
922 struct dm_target *ti = m->ti;
926 ti->error = "not enough priority group arguments";
927 return ERR_PTR(-EINVAL);
930 pg = alloc_priority_group();
932 ti->error = "couldn't allocate priority group";
933 return ERR_PTR(-ENOMEM);
937 r = parse_path_selector(as, pg, ti);
944 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
948 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
952 nr_args = 1 + nr_selector_args;
953 for (i = 0; i < pg->nr_pgpaths; i++) {
954 struct pgpath *pgpath;
955 struct dm_arg_set path_args;
957 if (as->argc < nr_args) {
958 ti->error = "not enough path parameters";
963 path_args.argc = nr_args;
964 path_args.argv = as->argv;
966 pgpath = parse_path(&path_args, &pg->ps, ti);
967 if (IS_ERR(pgpath)) {
973 list_add_tail(&pgpath->list, &pg->pgpaths);
974 dm_consume_args(as, nr_args);
980 free_priority_group(pg, ti);
984 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
988 struct dm_target *ti = m->ti;
990 static const struct dm_arg _args[] = {
991 {0, 1024, "invalid number of hardware handler args"},
994 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1000 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1001 dm_consume_args(as, hw_argc);
1002 DMERR("bio-based multipath doesn't allow hardware handler args");
1006 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1007 if (!m->hw_handler_name)
1014 for (i = 0; i <= hw_argc - 2; i++)
1015 len += strlen(as->argv[i]) + 1;
1016 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1018 ti->error = "memory allocation failed";
1022 j = sprintf(p, "%d", hw_argc - 1);
1023 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1024 j = sprintf(p, "%s", as->argv[i]);
1026 dm_consume_args(as, hw_argc - 1);
1030 kfree(m->hw_handler_name);
1031 m->hw_handler_name = NULL;
1035 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1039 struct dm_target *ti = m->ti;
1040 const char *arg_name;
1042 static const struct dm_arg _args[] = {
1043 {0, 8, "invalid number of feature args"},
1044 {1, 50, "pg_init_retries must be between 1 and 50"},
1045 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1048 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1056 arg_name = dm_shift_arg(as);
1059 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1060 r = queue_if_no_path(m, true, false);
1064 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1065 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1069 if (!strcasecmp(arg_name, "pg_init_retries") &&
1071 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1076 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1078 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1083 if (!strcasecmp(arg_name, "queue_mode") &&
1085 const char *queue_mode_name = dm_shift_arg(as);
1087 if (!strcasecmp(queue_mode_name, "bio"))
1088 m->queue_mode = DM_TYPE_BIO_BASED;
1089 else if (!strcasecmp(queue_mode_name, "rq"))
1090 m->queue_mode = DM_TYPE_REQUEST_BASED;
1091 else if (!strcasecmp(queue_mode_name, "mq"))
1092 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1094 ti->error = "Unknown 'queue_mode' requested";
1101 ti->error = "Unrecognised multipath feature request";
1103 } while (argc && !r);
1108 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1110 /* target arguments */
1111 static const struct dm_arg _args[] = {
1112 {0, 1024, "invalid number of priority groups"},
1113 {0, 1024, "invalid initial priority group number"},
1117 struct multipath *m;
1118 struct dm_arg_set as;
1119 unsigned pg_count = 0;
1120 unsigned next_pg_num;
1125 m = alloc_multipath(ti);
1127 ti->error = "can't allocate multipath";
1131 r = parse_features(&as, m);
1135 r = alloc_multipath_stage2(ti, m);
1139 r = parse_hw_handler(&as, m);
1143 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1147 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1151 if ((!m->nr_priority_groups && next_pg_num) ||
1152 (m->nr_priority_groups && !next_pg_num)) {
1153 ti->error = "invalid initial priority group";
1158 /* parse the priority groups */
1160 struct priority_group *pg;
1161 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1163 pg = parse_priority_group(&as, m);
1169 nr_valid_paths += pg->nr_pgpaths;
1170 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1172 list_add_tail(&pg->list, &m->priority_groups);
1174 pg->pg_num = pg_count;
1179 if (pg_count != m->nr_priority_groups) {
1180 ti->error = "priority group count mismatch";
1185 ti->num_flush_bios = 1;
1186 ti->num_discard_bios = 1;
1187 ti->num_write_same_bios = 1;
1188 ti->num_write_zeroes_bios = 1;
1189 if (m->queue_mode == DM_TYPE_BIO_BASED)
1190 ti->per_io_data_size = multipath_per_bio_data_size();
1192 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1201 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1206 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1208 if (!atomic_read(&m->pg_init_in_progress))
1213 finish_wait(&m->pg_init_wait, &wait);
1216 static void flush_multipath_work(struct multipath *m)
1218 if (m->hw_handler_name) {
1219 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1220 smp_mb__after_atomic();
1222 flush_workqueue(kmpath_handlerd);
1223 multipath_wait_for_pg_init_completion(m);
1225 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1226 smp_mb__after_atomic();
1229 flush_workqueue(kmultipathd);
1230 flush_work(&m->trigger_event);
1233 static void multipath_dtr(struct dm_target *ti)
1235 struct multipath *m = ti->private;
1237 flush_multipath_work(m);
1242 * Take a path out of use.
1244 static int fail_path(struct pgpath *pgpath)
1246 unsigned long flags;
1247 struct multipath *m = pgpath->pg->m;
1249 spin_lock_irqsave(&m->lock, flags);
1251 if (!pgpath->is_active)
1254 DMWARN("Failing path %s.", pgpath->path.dev->name);
1256 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1257 pgpath->is_active = false;
1258 pgpath->fail_count++;
1260 atomic_dec(&m->nr_valid_paths);
1262 if (pgpath == m->current_pgpath)
1263 m->current_pgpath = NULL;
1265 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1266 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1268 schedule_work(&m->trigger_event);
1271 spin_unlock_irqrestore(&m->lock, flags);
1277 * Reinstate a previously-failed path
1279 static int reinstate_path(struct pgpath *pgpath)
1281 int r = 0, run_queue = 0;
1282 unsigned long flags;
1283 struct multipath *m = pgpath->pg->m;
1284 unsigned nr_valid_paths;
1286 spin_lock_irqsave(&m->lock, flags);
1288 if (pgpath->is_active)
1291 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1293 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1297 pgpath->is_active = true;
1299 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1300 if (nr_valid_paths == 1) {
1301 m->current_pgpath = NULL;
1303 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1304 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1305 atomic_inc(&m->pg_init_in_progress);
1308 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1309 pgpath->path.dev->name, nr_valid_paths);
1311 schedule_work(&m->trigger_event);
1314 spin_unlock_irqrestore(&m->lock, flags);
1316 dm_table_run_md_queue_async(m->ti->table);
1317 process_queued_io_list(m);
1324 * Fail or reinstate all paths that match the provided struct dm_dev.
1326 static int action_dev(struct multipath *m, struct dm_dev *dev,
1330 struct pgpath *pgpath;
1331 struct priority_group *pg;
1333 list_for_each_entry(pg, &m->priority_groups, list) {
1334 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1335 if (pgpath->path.dev == dev)
1344 * Temporarily try to avoid having to use the specified PG
1346 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1349 unsigned long flags;
1351 spin_lock_irqsave(&m->lock, flags);
1353 pg->bypassed = bypassed;
1354 m->current_pgpath = NULL;
1355 m->current_pg = NULL;
1357 spin_unlock_irqrestore(&m->lock, flags);
1359 schedule_work(&m->trigger_event);
1363 * Switch to using the specified PG from the next I/O that gets mapped
1365 static int switch_pg_num(struct multipath *m, const char *pgstr)
1367 struct priority_group *pg;
1369 unsigned long flags;
1372 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1373 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1374 DMWARN("invalid PG number supplied to switch_pg_num");
1378 spin_lock_irqsave(&m->lock, flags);
1379 list_for_each_entry(pg, &m->priority_groups, list) {
1380 pg->bypassed = false;
1384 m->current_pgpath = NULL;
1385 m->current_pg = NULL;
1388 spin_unlock_irqrestore(&m->lock, flags);
1390 schedule_work(&m->trigger_event);
1395 * Set/clear bypassed status of a PG.
1396 * PGs are numbered upwards from 1 in the order they were declared.
1398 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1400 struct priority_group *pg;
1404 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1405 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1406 DMWARN("invalid PG number supplied to bypass_pg");
1410 list_for_each_entry(pg, &m->priority_groups, list) {
1415 bypass_pg(m, pg, bypassed);
1420 * Should we retry pg_init immediately?
1422 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1424 unsigned long flags;
1425 bool limit_reached = false;
1427 spin_lock_irqsave(&m->lock, flags);
1429 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1430 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1431 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1433 limit_reached = true;
1435 spin_unlock_irqrestore(&m->lock, flags);
1437 return limit_reached;
1440 static void pg_init_done(void *data, int errors)
1442 struct pgpath *pgpath = data;
1443 struct priority_group *pg = pgpath->pg;
1444 struct multipath *m = pg->m;
1445 unsigned long flags;
1446 bool delay_retry = false;
1448 /* device or driver problems */
1453 if (!m->hw_handler_name) {
1457 DMERR("Could not failover the device: Handler scsi_dh_%s "
1458 "Error %d.", m->hw_handler_name, errors);
1460 * Fail path for now, so we do not ping pong
1464 case SCSI_DH_DEV_TEMP_BUSY:
1466 * Probably doing something like FW upgrade on the
1467 * controller so try the other pg.
1469 bypass_pg(m, pg, true);
1472 /* Wait before retrying. */
1475 case SCSI_DH_IMM_RETRY:
1476 case SCSI_DH_RES_TEMP_UNAVAIL:
1477 if (pg_init_limit_reached(m, pgpath))
1481 case SCSI_DH_DEV_OFFLINED:
1484 * We probably do not want to fail the path for a device
1485 * error, but this is what the old dm did. In future
1486 * patches we can do more advanced handling.
1491 spin_lock_irqsave(&m->lock, flags);
1493 if (pgpath == m->current_pgpath) {
1494 DMERR("Could not failover device. Error %d.", errors);
1495 m->current_pgpath = NULL;
1496 m->current_pg = NULL;
1498 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1499 pg->bypassed = false;
1501 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1502 /* Activations of other paths are still on going */
1505 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1507 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1509 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1511 if (__pg_init_all_paths(m))
1514 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1516 process_queued_io_list(m);
1519 * Wake up any thread waiting to suspend.
1521 wake_up(&m->pg_init_wait);
1524 spin_unlock_irqrestore(&m->lock, flags);
1527 static void activate_or_offline_path(struct pgpath *pgpath)
1529 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1531 if (pgpath->is_active && !blk_queue_dying(q))
1532 scsi_dh_activate(q, pg_init_done, pgpath);
1534 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1537 static void activate_path_work(struct work_struct *work)
1539 struct pgpath *pgpath =
1540 container_of(work, struct pgpath, activate_path.work);
1542 activate_or_offline_path(pgpath);
1545 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1546 blk_status_t error, union map_info *map_context)
1548 struct dm_mpath_io *mpio = get_mpio(map_context);
1549 struct pgpath *pgpath = mpio->pgpath;
1550 int r = DM_ENDIO_DONE;
1553 * We don't queue any clone request inside the multipath target
1554 * during end I/O handling, since those clone requests don't have
1555 * bio clones. If we queue them inside the multipath target,
1556 * we need to make bio clones, that requires memory allocation.
1557 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1558 * don't have bio clones.)
1559 * Instead of queueing the clone request here, we queue the original
1560 * request into dm core, which will remake a clone request and
1561 * clone bios for it and resubmit it later.
1563 if (error && blk_path_error(error)) {
1564 struct multipath *m = ti->private;
1566 if (error == BLK_STS_RESOURCE)
1567 r = DM_ENDIO_DELAY_REQUEUE;
1569 r = DM_ENDIO_REQUEUE;
1574 if (atomic_read(&m->nr_valid_paths) == 0 &&
1575 !must_push_back_rq(m)) {
1576 if (error == BLK_STS_IOERR)
1578 /* complete with the original error */
1584 struct path_selector *ps = &pgpath->pg->ps;
1586 if (ps->type->end_io)
1587 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1593 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1594 blk_status_t *error)
1596 struct multipath *m = ti->private;
1597 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1598 struct pgpath *pgpath = mpio->pgpath;
1599 unsigned long flags;
1600 int r = DM_ENDIO_DONE;
1602 if (!*error || !blk_path_error(*error))
1608 if (atomic_read(&m->nr_valid_paths) == 0 &&
1609 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1610 if (must_push_back_bio(m)) {
1611 r = DM_ENDIO_REQUEUE;
1614 *error = BLK_STS_IOERR;
1619 spin_lock_irqsave(&m->lock, flags);
1620 bio_list_add(&m->queued_bios, clone);
1621 spin_unlock_irqrestore(&m->lock, flags);
1622 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1623 queue_work(kmultipathd, &m->process_queued_bios);
1625 r = DM_ENDIO_INCOMPLETE;
1628 struct path_selector *ps = &pgpath->pg->ps;
1630 if (ps->type->end_io)
1631 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1638 * Suspend can't complete until all the I/O is processed so if
1639 * the last path fails we must error any remaining I/O.
1640 * Note that if the freeze_bdev fails while suspending, the
1641 * queue_if_no_path state is lost - userspace should reset it.
1643 static void multipath_presuspend(struct dm_target *ti)
1645 struct multipath *m = ti->private;
1647 queue_if_no_path(m, false, true);
1650 static void multipath_postsuspend(struct dm_target *ti)
1652 struct multipath *m = ti->private;
1654 mutex_lock(&m->work_mutex);
1655 flush_multipath_work(m);
1656 mutex_unlock(&m->work_mutex);
1660 * Restore the queue_if_no_path setting.
1662 static void multipath_resume(struct dm_target *ti)
1664 struct multipath *m = ti->private;
1665 unsigned long flags;
1667 spin_lock_irqsave(&m->lock, flags);
1668 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1669 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1670 spin_unlock_irqrestore(&m->lock, flags);
1674 * Info output has the following format:
1675 * num_multipath_feature_args [multipath_feature_args]*
1676 * num_handler_status_args [handler_status_args]*
1677 * num_groups init_group_number
1678 * [A|D|E num_ps_status_args [ps_status_args]*
1679 * num_paths num_selector_args
1680 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1682 * Table output has the following format (identical to the constructor string):
1683 * num_feature_args [features_args]*
1684 * num_handler_args hw_handler [hw_handler_args]*
1685 * num_groups init_group_number
1686 * [priority selector-name num_ps_args [ps_args]*
1687 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1689 static void multipath_status(struct dm_target *ti, status_type_t type,
1690 unsigned status_flags, char *result, unsigned maxlen)
1693 unsigned long flags;
1694 struct multipath *m = ti->private;
1695 struct priority_group *pg;
1700 spin_lock_irqsave(&m->lock, flags);
1703 if (type == STATUSTYPE_INFO)
1704 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1705 atomic_read(&m->pg_init_count));
1707 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1708 (m->pg_init_retries > 0) * 2 +
1709 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1710 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1711 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1713 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1714 DMEMIT("queue_if_no_path ");
1715 if (m->pg_init_retries)
1716 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1717 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1718 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1719 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1720 DMEMIT("retain_attached_hw_handler ");
1721 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1722 switch(m->queue_mode) {
1723 case DM_TYPE_BIO_BASED:
1724 DMEMIT("queue_mode bio ");
1726 case DM_TYPE_MQ_REQUEST_BASED:
1727 DMEMIT("queue_mode mq ");
1736 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1739 DMEMIT("1 %s ", m->hw_handler_name);
1741 DMEMIT("%u ", m->nr_priority_groups);
1744 pg_num = m->next_pg->pg_num;
1745 else if (m->current_pg)
1746 pg_num = m->current_pg->pg_num;
1748 pg_num = (m->nr_priority_groups ? 1 : 0);
1750 DMEMIT("%u ", pg_num);
1753 case STATUSTYPE_INFO:
1754 list_for_each_entry(pg, &m->priority_groups, list) {
1756 state = 'D'; /* Disabled */
1757 else if (pg == m->current_pg)
1758 state = 'A'; /* Currently Active */
1760 state = 'E'; /* Enabled */
1762 DMEMIT("%c ", state);
1764 if (pg->ps.type->status)
1765 sz += pg->ps.type->status(&pg->ps, NULL, type,
1771 DMEMIT("%u %u ", pg->nr_pgpaths,
1772 pg->ps.type->info_args);
1774 list_for_each_entry(p, &pg->pgpaths, list) {
1775 DMEMIT("%s %s %u ", p->path.dev->name,
1776 p->is_active ? "A" : "F",
1778 if (pg->ps.type->status)
1779 sz += pg->ps.type->status(&pg->ps,
1780 &p->path, type, result + sz,
1786 case STATUSTYPE_TABLE:
1787 list_for_each_entry(pg, &m->priority_groups, list) {
1788 DMEMIT("%s ", pg->ps.type->name);
1790 if (pg->ps.type->status)
1791 sz += pg->ps.type->status(&pg->ps, NULL, type,
1797 DMEMIT("%u %u ", pg->nr_pgpaths,
1798 pg->ps.type->table_args);
1800 list_for_each_entry(p, &pg->pgpaths, list) {
1801 DMEMIT("%s ", p->path.dev->name);
1802 if (pg->ps.type->status)
1803 sz += pg->ps.type->status(&pg->ps,
1804 &p->path, type, result + sz,
1811 spin_unlock_irqrestore(&m->lock, flags);
1814 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1815 char *result, unsigned maxlen)
1819 struct multipath *m = ti->private;
1822 mutex_lock(&m->work_mutex);
1824 if (dm_suspended(ti)) {
1830 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1831 r = queue_if_no_path(m, true, false);
1833 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1834 r = queue_if_no_path(m, false, false);
1840 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1844 if (!strcasecmp(argv[0], "disable_group")) {
1845 r = bypass_pg_num(m, argv[1], true);
1847 } else if (!strcasecmp(argv[0], "enable_group")) {
1848 r = bypass_pg_num(m, argv[1], false);
1850 } else if (!strcasecmp(argv[0], "switch_group")) {
1851 r = switch_pg_num(m, argv[1]);
1853 } else if (!strcasecmp(argv[0], "reinstate_path"))
1854 action = reinstate_path;
1855 else if (!strcasecmp(argv[0], "fail_path"))
1858 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1862 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1864 DMWARN("message: error getting device %s",
1869 r = action_dev(m, dev, action);
1871 dm_put_device(ti, dev);
1874 mutex_unlock(&m->work_mutex);
1878 static int multipath_prepare_ioctl(struct dm_target *ti,
1879 struct block_device **bdev, fmode_t *mode)
1881 struct multipath *m = ti->private;
1882 struct pgpath *current_pgpath;
1885 current_pgpath = READ_ONCE(m->current_pgpath);
1886 if (!current_pgpath)
1887 current_pgpath = choose_pgpath(m, 0);
1889 if (current_pgpath) {
1890 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1891 *bdev = current_pgpath->path.dev->bdev;
1892 *mode = current_pgpath->path.dev->mode;
1895 /* pg_init has not started or completed */
1899 /* No path is available */
1900 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1906 if (r == -ENOTCONN) {
1907 if (!READ_ONCE(m->current_pg)) {
1908 /* Path status changed, redo selection */
1909 (void) choose_pgpath(m, 0);
1911 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1912 pg_init_all_paths(m);
1913 dm_table_run_md_queue_async(m->ti->table);
1914 process_queued_io_list(m);
1918 * Only pass ioctls through if the device sizes match exactly.
1920 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1925 static int multipath_iterate_devices(struct dm_target *ti,
1926 iterate_devices_callout_fn fn, void *data)
1928 struct multipath *m = ti->private;
1929 struct priority_group *pg;
1933 list_for_each_entry(pg, &m->priority_groups, list) {
1934 list_for_each_entry(p, &pg->pgpaths, list) {
1935 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1945 static int pgpath_busy(struct pgpath *pgpath)
1947 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1949 return blk_lld_busy(q);
1953 * We return "busy", only when we can map I/Os but underlying devices
1954 * are busy (so even if we map I/Os now, the I/Os will wait on
1955 * the underlying queue).
1956 * In other words, if we want to kill I/Os or queue them inside us
1957 * due to map unavailability, we don't return "busy". Otherwise,
1958 * dm core won't give us the I/Os and we can't do what we want.
1960 static int multipath_busy(struct dm_target *ti)
1962 bool busy = false, has_active = false;
1963 struct multipath *m = ti->private;
1964 struct priority_group *pg, *next_pg;
1965 struct pgpath *pgpath;
1967 /* pg_init in progress */
1968 if (atomic_read(&m->pg_init_in_progress))
1971 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1972 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1973 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1975 /* Guess which priority_group will be used at next mapping time */
1976 pg = READ_ONCE(m->current_pg);
1977 next_pg = READ_ONCE(m->next_pg);
1978 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1983 * We don't know which pg will be used at next mapping time.
1984 * We don't call choose_pgpath() here to avoid to trigger
1985 * pg_init just by busy checking.
1986 * So we don't know whether underlying devices we will be using
1987 * at next mapping time are busy or not. Just try mapping.
1993 * If there is one non-busy active path at least, the path selector
1994 * will be able to select it. So we consider such a pg as not busy.
1997 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1998 if (pgpath->is_active) {
2000 if (!pgpath_busy(pgpath)) {
2009 * No active path in this pg, so this pg won't be used and
2010 * the current_pg will be changed at next mapping time.
2011 * We need to try mapping to determine it.
2019 /*-----------------------------------------------------------------
2021 *---------------------------------------------------------------*/
2022 static struct target_type multipath_target = {
2023 .name = "multipath",
2024 .version = {1, 13, 0},
2025 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2026 DM_TARGET_PASSES_INTEGRITY,
2027 .module = THIS_MODULE,
2028 .ctr = multipath_ctr,
2029 .dtr = multipath_dtr,
2030 .clone_and_map_rq = multipath_clone_and_map,
2031 .release_clone_rq = multipath_release_clone,
2032 .rq_end_io = multipath_end_io,
2033 .map = multipath_map_bio,
2034 .end_io = multipath_end_io_bio,
2035 .presuspend = multipath_presuspend,
2036 .postsuspend = multipath_postsuspend,
2037 .resume = multipath_resume,
2038 .status = multipath_status,
2039 .message = multipath_message,
2040 .prepare_ioctl = multipath_prepare_ioctl,
2041 .iterate_devices = multipath_iterate_devices,
2042 .busy = multipath_busy,
2045 static int __init dm_multipath_init(void)
2049 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2051 DMERR("failed to create workqueue kmpathd");
2053 goto bad_alloc_kmultipathd;
2057 * A separate workqueue is used to handle the device handlers
2058 * to avoid overloading existing workqueue. Overloading the
2059 * old workqueue would also create a bottleneck in the
2060 * path of the storage hardware device activation.
2062 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2064 if (!kmpath_handlerd) {
2065 DMERR("failed to create workqueue kmpath_handlerd");
2067 goto bad_alloc_kmpath_handlerd;
2070 r = dm_register_target(&multipath_target);
2072 DMERR("request-based register failed %d", r);
2074 goto bad_register_target;
2079 bad_register_target:
2080 destroy_workqueue(kmpath_handlerd);
2081 bad_alloc_kmpath_handlerd:
2082 destroy_workqueue(kmultipathd);
2083 bad_alloc_kmultipathd:
2087 static void __exit dm_multipath_exit(void)
2089 destroy_workqueue(kmpath_handlerd);
2090 destroy_workqueue(kmultipathd);
2092 dm_unregister_target(&multipath_target);
2095 module_init(dm_multipath_init);
2096 module_exit(dm_multipath_exit);
2098 MODULE_DESCRIPTION(DM_NAME " multipath target");
2099 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2100 MODULE_LICENSE("GPL");