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 */
96 * We must use a mempool of dm_mpath_io structs so that we
97 * can resubmit bios on error.
101 struct mutex work_mutex
;
102 struct work_struct trigger_event
;
104 struct work_struct process_queued_bios
;
105 struct bio_list queued_bios
;
109 * Context information attached to each io we process.
112 struct pgpath
*pgpath
;
116 typedef int (*action_fn
) (struct pgpath
*pgpath
);
118 static struct kmem_cache
*_mpio_cache
;
120 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
121 static void trigger_event(struct work_struct
*work
);
122 static void activate_path(struct work_struct
*work
);
123 static void process_queued_bios(struct work_struct
*work
);
125 /*-----------------------------------------------
126 * Multipath state flags.
127 *-----------------------------------------------*/
129 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
130 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
131 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
132 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
133 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
134 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
135 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
137 /*-----------------------------------------------
138 * Allocation routines
139 *-----------------------------------------------*/
141 static struct pgpath
*alloc_pgpath(void)
143 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
146 pgpath
->is_active
= true;
147 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path
);
153 static void free_pgpath(struct pgpath
*pgpath
)
158 static struct priority_group
*alloc_priority_group(void)
160 struct priority_group
*pg
;
162 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
165 INIT_LIST_HEAD(&pg
->pgpaths
);
170 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
172 struct pgpath
*pgpath
, *tmp
;
174 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
175 list_del(&pgpath
->list
);
176 dm_put_device(ti
, pgpath
->path
.dev
);
181 static void free_priority_group(struct priority_group
*pg
,
182 struct dm_target
*ti
)
184 struct path_selector
*ps
= &pg
->ps
;
187 ps
->type
->destroy(ps
);
188 dm_put_path_selector(ps
->type
);
191 free_pgpaths(&pg
->pgpaths
, ti
);
195 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
199 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
201 INIT_LIST_HEAD(&m
->priority_groups
);
202 spin_lock_init(&m
->lock
);
203 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
204 atomic_set(&m
->nr_valid_paths
, 0);
205 atomic_set(&m
->pg_init_in_progress
, 0);
206 atomic_set(&m
->pg_init_count
, 0);
207 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
208 INIT_WORK(&m
->trigger_event
, trigger_event
);
209 init_waitqueue_head(&m
->pg_init_wait
);
210 mutex_init(&m
->work_mutex
);
213 m
->queue_mode
= DM_TYPE_NONE
;
222 static int alloc_multipath_stage2(struct dm_target
*ti
, struct multipath
*m
)
224 if (m
->queue_mode
== DM_TYPE_NONE
) {
226 * Default to request-based.
228 if (dm_use_blk_mq(dm_table_get_md(ti
->table
)))
229 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
231 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
234 if (m
->queue_mode
== DM_TYPE_REQUEST_BASED
) {
235 unsigned min_ios
= dm_get_reserved_rq_based_ios();
237 m
->mpio_pool
= mempool_create_slab_pool(min_ios
, _mpio_cache
);
241 else if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
242 INIT_WORK(&m
->process_queued_bios
, process_queued_bios
);
244 * bio-based doesn't support any direct scsi_dh management;
245 * it just discovers if a scsi_dh is attached.
247 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
250 dm_table_set_type(ti
->table
, m
->queue_mode
);
255 static void free_multipath(struct multipath
*m
)
257 struct priority_group
*pg
, *tmp
;
259 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
261 free_priority_group(pg
, m
->ti
);
264 kfree(m
->hw_handler_name
);
265 kfree(m
->hw_handler_params
);
266 mempool_destroy(m
->mpio_pool
);
270 static struct dm_mpath_io
*get_mpio(union map_info
*info
)
275 static struct dm_mpath_io
*set_mpio(struct multipath
*m
, union map_info
*info
)
277 struct dm_mpath_io
*mpio
;
280 /* Use blk-mq pdu memory requested via per_io_data_size */
281 mpio
= get_mpio(info
);
282 memset(mpio
, 0, sizeof(*mpio
));
286 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
290 memset(mpio
, 0, sizeof(*mpio
));
296 static void clear_request_fn_mpio(struct multipath
*m
, union map_info
*info
)
298 /* Only needed for non blk-mq (.request_fn) multipath */
300 struct dm_mpath_io
*mpio
= info
->ptr
;
303 mempool_free(mpio
, m
->mpio_pool
);
307 static size_t multipath_per_bio_data_size(void)
309 return sizeof(struct dm_mpath_io
) + sizeof(struct dm_bio_details
);
312 static struct dm_mpath_io
*get_mpio_from_bio(struct bio
*bio
)
314 return dm_per_bio_data(bio
, multipath_per_bio_data_size());
317 static struct dm_bio_details
*get_bio_details_from_bio(struct bio
*bio
)
319 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
320 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
321 void *bio_details
= mpio
+ 1;
326 static void multipath_init_per_bio_data(struct bio
*bio
, struct dm_mpath_io
**mpio_p
,
327 struct dm_bio_details
**bio_details_p
)
329 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
330 struct dm_bio_details
*bio_details
= get_bio_details_from_bio(bio
);
332 memset(mpio
, 0, sizeof(*mpio
));
333 memset(bio_details
, 0, sizeof(*bio_details
));
334 dm_bio_record(bio_details
, bio
);
339 *bio_details_p
= bio_details
;
342 /*-----------------------------------------------
344 *-----------------------------------------------*/
346 static int __pg_init_all_paths(struct multipath
*m
)
348 struct pgpath
*pgpath
;
349 unsigned long pg_init_delay
= 0;
351 if (atomic_read(&m
->pg_init_in_progress
) || test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
354 atomic_inc(&m
->pg_init_count
);
355 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
357 /* Check here to reset pg_init_required */
361 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
))
362 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
363 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
364 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
365 /* Skip failed paths */
366 if (!pgpath
->is_active
)
368 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
370 atomic_inc(&m
->pg_init_in_progress
);
372 return atomic_read(&m
->pg_init_in_progress
);
375 static int pg_init_all_paths(struct multipath
*m
)
380 spin_lock_irqsave(&m
->lock
, flags
);
381 r
= __pg_init_all_paths(m
);
382 spin_unlock_irqrestore(&m
->lock
, flags
);
387 static void __switch_pg(struct multipath
*m
, struct priority_group
*pg
)
391 /* Must we initialise the PG first, and queue I/O till it's ready? */
392 if (m
->hw_handler_name
) {
393 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
394 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
396 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
397 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
400 atomic_set(&m
->pg_init_count
, 0);
403 static struct pgpath
*choose_path_in_pg(struct multipath
*m
,
404 struct priority_group
*pg
,
408 struct dm_path
*path
;
409 struct pgpath
*pgpath
;
411 path
= pg
->ps
.type
->select_path(&pg
->ps
, nr_bytes
);
413 return ERR_PTR(-ENXIO
);
415 pgpath
= path_to_pgpath(path
);
417 if (unlikely(lockless_dereference(m
->current_pg
) != pg
)) {
418 /* Only update current_pgpath if pg changed */
419 spin_lock_irqsave(&m
->lock
, flags
);
420 m
->current_pgpath
= pgpath
;
422 spin_unlock_irqrestore(&m
->lock
, flags
);
428 static struct pgpath
*choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
431 struct priority_group
*pg
;
432 struct pgpath
*pgpath
;
433 bool bypassed
= true;
435 if (!atomic_read(&m
->nr_valid_paths
)) {
436 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
440 /* Were we instructed to switch PG? */
441 if (lockless_dereference(m
->next_pg
)) {
442 spin_lock_irqsave(&m
->lock
, flags
);
445 spin_unlock_irqrestore(&m
->lock
, flags
);
446 goto check_current_pg
;
449 spin_unlock_irqrestore(&m
->lock
, flags
);
450 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
451 if (!IS_ERR_OR_NULL(pgpath
))
455 /* Don't change PG until it has no remaining paths */
457 pg
= lockless_dereference(m
->current_pg
);
459 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
460 if (!IS_ERR_OR_NULL(pgpath
))
465 * Loop through priority groups until we find a valid path.
466 * First time we skip PGs marked 'bypassed'.
467 * Second time we only try the ones we skipped, but set
468 * pg_init_delay_retry so we do not hammer controllers.
471 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
472 if (pg
->bypassed
== bypassed
)
474 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
475 if (!IS_ERR_OR_NULL(pgpath
)) {
477 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
481 } while (bypassed
--);
484 spin_lock_irqsave(&m
->lock
, flags
);
485 m
->current_pgpath
= NULL
;
486 m
->current_pg
= NULL
;
487 spin_unlock_irqrestore(&m
->lock
, flags
);
493 * Check whether bios must be queued in the device-mapper core rather
494 * than here in the target.
496 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
497 * same value then we are not between multipath_presuspend()
498 * and multipath_resume() calls and we have no need to check
499 * for the DMF_NOFLUSH_SUSPENDING flag.
501 static bool __must_push_back(struct multipath
*m
)
503 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) !=
504 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
)) &&
505 dm_noflush_suspending(m
->ti
));
508 static bool must_push_back_rq(struct multipath
*m
)
513 spin_lock_irqsave(&m
->lock
, flags
);
514 r
= (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) ||
515 __must_push_back(m
));
516 spin_unlock_irqrestore(&m
->lock
, flags
);
521 static bool must_push_back_bio(struct multipath
*m
)
526 spin_lock_irqsave(&m
->lock
, flags
);
527 r
= __must_push_back(m
);
528 spin_unlock_irqrestore(&m
->lock
, flags
);
534 * Map cloned requests (request-based multipath)
536 static int __multipath_map(struct dm_target
*ti
, struct request
*clone
,
537 union map_info
*map_context
,
538 struct request
*rq
, struct request
**__clone
)
540 struct multipath
*m
= ti
->private;
541 int r
= DM_MAPIO_REQUEUE
;
542 size_t nr_bytes
= clone
? blk_rq_bytes(clone
) : blk_rq_bytes(rq
);
543 struct pgpath
*pgpath
;
544 struct block_device
*bdev
;
545 struct dm_mpath_io
*mpio
;
547 /* Do we need to select a new pgpath? */
548 pgpath
= lockless_dereference(m
->current_pgpath
);
549 if (!pgpath
|| !test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
550 pgpath
= choose_pgpath(m
, nr_bytes
);
553 if (must_push_back_rq(m
))
554 return DM_MAPIO_DELAY_REQUEUE
;
555 return -EIO
; /* Failed */
556 } else if (test_bit(MPATHF_QUEUE_IO
, &m
->flags
) ||
557 test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
558 pg_init_all_paths(m
);
562 mpio
= set_mpio(m
, map_context
);
564 /* ENOMEM, requeue */
567 mpio
->pgpath
= pgpath
;
568 mpio
->nr_bytes
= nr_bytes
;
570 bdev
= pgpath
->path
.dev
->bdev
;
574 * Old request-based interface: allocated clone is passed in.
575 * Used by: .request_fn stacked on .request_fn path(s).
577 clone
->q
= bdev_get_queue(bdev
);
578 clone
->rq_disk
= bdev
->bd_disk
;
579 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
582 * blk-mq request-based interface; used by both:
583 * .request_fn stacked on blk-mq path(s) and
584 * blk-mq stacked on blk-mq path(s).
586 *__clone
= blk_mq_alloc_request(bdev_get_queue(bdev
),
587 rq_data_dir(rq
), BLK_MQ_REQ_NOWAIT
);
588 if (IS_ERR(*__clone
)) {
589 /* ENOMEM, requeue */
590 clear_request_fn_mpio(m
, map_context
);
593 (*__clone
)->bio
= (*__clone
)->biotail
= NULL
;
594 (*__clone
)->rq_disk
= bdev
->bd_disk
;
595 (*__clone
)->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
598 if (pgpath
->pg
->ps
.type
->start_io
)
599 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
602 return DM_MAPIO_REMAPPED
;
605 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
606 union map_info
*map_context
)
608 return __multipath_map(ti
, clone
, map_context
, NULL
, NULL
);
611 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
612 union map_info
*map_context
,
613 struct request
**clone
)
615 return __multipath_map(ti
, NULL
, map_context
, rq
, clone
);
618 static void multipath_release_clone(struct request
*clone
)
620 blk_mq_free_request(clone
);
624 * Map cloned bios (bio-based multipath)
626 static int __multipath_map_bio(struct multipath
*m
, struct bio
*bio
, struct dm_mpath_io
*mpio
)
628 size_t nr_bytes
= bio
->bi_iter
.bi_size
;
629 struct pgpath
*pgpath
;
633 /* Do we need to select a new pgpath? */
634 pgpath
= lockless_dereference(m
->current_pgpath
);
635 queue_io
= test_bit(MPATHF_QUEUE_IO
, &m
->flags
);
636 if (!pgpath
|| !queue_io
)
637 pgpath
= choose_pgpath(m
, nr_bytes
);
639 if ((pgpath
&& queue_io
) ||
640 (!pgpath
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))) {
641 /* Queue for the daemon to resubmit */
642 spin_lock_irqsave(&m
->lock
, flags
);
643 bio_list_add(&m
->queued_bios
, bio
);
644 spin_unlock_irqrestore(&m
->lock
, flags
);
645 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
646 if (queue_io
|| test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
647 pg_init_all_paths(m
);
649 queue_work(kmultipathd
, &m
->process_queued_bios
);
650 return DM_MAPIO_SUBMITTED
;
654 if (!must_push_back_bio(m
))
656 return DM_MAPIO_REQUEUE
;
659 mpio
->pgpath
= pgpath
;
660 mpio
->nr_bytes
= nr_bytes
;
663 bio
->bi_bdev
= pgpath
->path
.dev
->bdev
;
664 bio
->bi_opf
|= REQ_FAILFAST_TRANSPORT
;
666 if (pgpath
->pg
->ps
.type
->start_io
)
667 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
670 return DM_MAPIO_REMAPPED
;
673 static int multipath_map_bio(struct dm_target
*ti
, struct bio
*bio
)
675 struct multipath
*m
= ti
->private;
676 struct dm_mpath_io
*mpio
= NULL
;
678 multipath_init_per_bio_data(bio
, &mpio
, NULL
);
680 return __multipath_map_bio(m
, bio
, mpio
);
683 static void process_queued_io_list(struct multipath
*m
)
685 if (m
->queue_mode
== DM_TYPE_MQ_REQUEST_BASED
)
686 dm_mq_kick_requeue_list(dm_table_get_md(m
->ti
->table
));
687 else if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
688 queue_work(kmultipathd
, &m
->process_queued_bios
);
691 static void process_queued_bios(struct work_struct
*work
)
696 struct bio_list bios
;
697 struct blk_plug plug
;
698 struct multipath
*m
=
699 container_of(work
, struct multipath
, process_queued_bios
);
701 bio_list_init(&bios
);
703 spin_lock_irqsave(&m
->lock
, flags
);
705 if (bio_list_empty(&m
->queued_bios
)) {
706 spin_unlock_irqrestore(&m
->lock
, flags
);
710 bio_list_merge(&bios
, &m
->queued_bios
);
711 bio_list_init(&m
->queued_bios
);
713 spin_unlock_irqrestore(&m
->lock
, flags
);
715 blk_start_plug(&plug
);
716 while ((bio
= bio_list_pop(&bios
))) {
717 r
= __multipath_map_bio(m
, bio
, get_mpio_from_bio(bio
));
718 if (r
< 0 || r
== DM_MAPIO_REQUEUE
) {
721 } else if (r
== DM_MAPIO_REMAPPED
)
722 generic_make_request(bio
);
724 blk_finish_plug(&plug
);
728 * If we run out of usable paths, should we queue I/O or error it?
730 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
735 spin_lock_irqsave(&m
->lock
, flags
);
737 if (save_old_value
) {
738 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
739 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
741 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
743 if (queue_if_no_path
)
744 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
746 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
748 if (queue_if_no_path
)
749 set_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
751 clear_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
753 spin_unlock_irqrestore(&m
->lock
, flags
);
755 if (!queue_if_no_path
) {
756 dm_table_run_md_queue_async(m
->ti
->table
);
757 process_queued_io_list(m
);
764 * An event is triggered whenever a path is taken out of use.
765 * Includes path failure and PG bypass.
767 static void trigger_event(struct work_struct
*work
)
769 struct multipath
*m
=
770 container_of(work
, struct multipath
, trigger_event
);
772 dm_table_event(m
->ti
->table
);
775 /*-----------------------------------------------------------------
776 * Constructor/argument parsing:
777 * <#multipath feature args> [<arg>]*
778 * <#hw_handler args> [hw_handler [<arg>]*]
780 * <initial priority group>
781 * [<selector> <#selector args> [<arg>]*
782 * <#paths> <#per-path selector args>
783 * [<path> [<arg>]* ]+ ]+
784 *---------------------------------------------------------------*/
785 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
786 struct dm_target
*ti
)
789 struct path_selector_type
*pst
;
792 static struct dm_arg _args
[] = {
793 {0, 1024, "invalid number of path selector args"},
796 pst
= dm_get_path_selector(dm_shift_arg(as
));
798 ti
->error
= "unknown path selector type";
802 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
804 dm_put_path_selector(pst
);
808 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
810 dm_put_path_selector(pst
);
811 ti
->error
= "path selector constructor failed";
816 dm_consume_args(as
, ps_argc
);
821 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
822 struct dm_target
*ti
)
826 struct multipath
*m
= ti
->private;
827 struct request_queue
*q
= NULL
;
828 const char *attached_handler_name
;
830 /* we need at least a path arg */
832 ti
->error
= "no device given";
833 return ERR_PTR(-EINVAL
);
838 return ERR_PTR(-ENOMEM
);
840 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
843 ti
->error
= "error getting device";
847 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) || m
->hw_handler_name
)
848 q
= bdev_get_queue(p
->path
.dev
->bdev
);
850 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
)) {
852 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
853 if (attached_handler_name
) {
855 * Reset hw_handler_name to match the attached handler
856 * and clear any hw_handler_params associated with the
859 * NB. This modifies the table line to show the actual
860 * handler instead of the original table passed in.
862 kfree(m
->hw_handler_name
);
863 m
->hw_handler_name
= attached_handler_name
;
865 kfree(m
->hw_handler_params
);
866 m
->hw_handler_params
= NULL
;
870 if (m
->hw_handler_name
) {
871 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
873 char b
[BDEVNAME_SIZE
];
875 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
876 bdevname(p
->path
.dev
->bdev
, b
));
880 ti
->error
= "error attaching hardware handler";
881 dm_put_device(ti
, p
->path
.dev
);
885 if (m
->hw_handler_params
) {
886 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
888 ti
->error
= "unable to set hardware "
889 "handler parameters";
890 dm_put_device(ti
, p
->path
.dev
);
896 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
898 dm_put_device(ti
, p
->path
.dev
);
909 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
912 static struct dm_arg _args
[] = {
913 {1, 1024, "invalid number of paths"},
914 {0, 1024, "invalid number of selector args"}
918 unsigned i
, nr_selector_args
, nr_args
;
919 struct priority_group
*pg
;
920 struct dm_target
*ti
= m
->ti
;
924 ti
->error
= "not enough priority group arguments";
925 return ERR_PTR(-EINVAL
);
928 pg
= alloc_priority_group();
930 ti
->error
= "couldn't allocate priority group";
931 return ERR_PTR(-ENOMEM
);
935 r
= parse_path_selector(as
, pg
, ti
);
942 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
946 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
950 nr_args
= 1 + nr_selector_args
;
951 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
952 struct pgpath
*pgpath
;
953 struct dm_arg_set path_args
;
955 if (as
->argc
< nr_args
) {
956 ti
->error
= "not enough path parameters";
961 path_args
.argc
= nr_args
;
962 path_args
.argv
= as
->argv
;
964 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
965 if (IS_ERR(pgpath
)) {
971 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
972 dm_consume_args(as
, nr_args
);
978 free_priority_group(pg
, ti
);
982 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
986 struct dm_target
*ti
= m
->ti
;
988 static struct dm_arg _args
[] = {
989 {0, 1024, "invalid number of hardware handler args"},
992 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
998 if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
999 dm_consume_args(as
, hw_argc
);
1000 DMERR("bio-based multipath doesn't allow hardware handler args");
1004 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
1010 for (i
= 0; i
<= hw_argc
- 2; i
++)
1011 len
+= strlen(as
->argv
[i
]) + 1;
1012 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
1014 ti
->error
= "memory allocation failed";
1018 j
= sprintf(p
, "%d", hw_argc
- 1);
1019 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
1020 j
= sprintf(p
, "%s", as
->argv
[i
]);
1022 dm_consume_args(as
, hw_argc
- 1);
1026 kfree(m
->hw_handler_name
);
1027 m
->hw_handler_name
= NULL
;
1031 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
1035 struct dm_target
*ti
= m
->ti
;
1036 const char *arg_name
;
1038 static struct dm_arg _args
[] = {
1039 {0, 8, "invalid number of feature args"},
1040 {1, 50, "pg_init_retries must be between 1 and 50"},
1041 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1044 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
1052 arg_name
= dm_shift_arg(as
);
1055 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
1056 r
= queue_if_no_path(m
, true, false);
1060 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
1061 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
1065 if (!strcasecmp(arg_name
, "pg_init_retries") &&
1067 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
1072 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
1074 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
1079 if (!strcasecmp(arg_name
, "queue_mode") &&
1081 const char *queue_mode_name
= dm_shift_arg(as
);
1083 if (!strcasecmp(queue_mode_name
, "bio"))
1084 m
->queue_mode
= DM_TYPE_BIO_BASED
;
1085 else if (!strcasecmp(queue_mode_name
, "rq"))
1086 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
1087 else if (!strcasecmp(queue_mode_name
, "mq"))
1088 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
1090 ti
->error
= "Unknown 'queue_mode' requested";
1097 ti
->error
= "Unrecognised multipath feature request";
1099 } while (argc
&& !r
);
1104 static int multipath_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
1106 /* target arguments */
1107 static struct dm_arg _args
[] = {
1108 {0, 1024, "invalid number of priority groups"},
1109 {0, 1024, "invalid initial priority group number"},
1113 struct multipath
*m
;
1114 struct dm_arg_set as
;
1115 unsigned pg_count
= 0;
1116 unsigned next_pg_num
;
1121 m
= alloc_multipath(ti
);
1123 ti
->error
= "can't allocate multipath";
1127 r
= parse_features(&as
, m
);
1131 r
= alloc_multipath_stage2(ti
, m
);
1135 r
= parse_hw_handler(&as
, m
);
1139 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
1143 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
1147 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
1148 (m
->nr_priority_groups
&& !next_pg_num
)) {
1149 ti
->error
= "invalid initial priority group";
1154 /* parse the priority groups */
1156 struct priority_group
*pg
;
1157 unsigned nr_valid_paths
= atomic_read(&m
->nr_valid_paths
);
1159 pg
= parse_priority_group(&as
, m
);
1165 nr_valid_paths
+= pg
->nr_pgpaths
;
1166 atomic_set(&m
->nr_valid_paths
, nr_valid_paths
);
1168 list_add_tail(&pg
->list
, &m
->priority_groups
);
1170 pg
->pg_num
= pg_count
;
1175 if (pg_count
!= m
->nr_priority_groups
) {
1176 ti
->error
= "priority group count mismatch";
1181 ti
->num_flush_bios
= 1;
1182 ti
->num_discard_bios
= 1;
1183 ti
->num_write_same_bios
= 1;
1184 if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
1185 ti
->per_io_data_size
= multipath_per_bio_data_size();
1186 else if (m
->queue_mode
== DM_TYPE_MQ_REQUEST_BASED
)
1187 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
1196 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
1201 prepare_to_wait(&m
->pg_init_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1203 if (!atomic_read(&m
->pg_init_in_progress
))
1208 finish_wait(&m
->pg_init_wait
, &wait
);
1211 static void flush_multipath_work(struct multipath
*m
)
1213 set_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1214 smp_mb__after_atomic();
1216 flush_workqueue(kmpath_handlerd
);
1217 multipath_wait_for_pg_init_completion(m
);
1218 flush_workqueue(kmultipathd
);
1219 flush_work(&m
->trigger_event
);
1221 clear_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1222 smp_mb__after_atomic();
1225 static void multipath_dtr(struct dm_target
*ti
)
1227 struct multipath
*m
= ti
->private;
1229 flush_multipath_work(m
);
1234 * Take a path out of use.
1236 static int fail_path(struct pgpath
*pgpath
)
1238 unsigned long flags
;
1239 struct multipath
*m
= pgpath
->pg
->m
;
1241 spin_lock_irqsave(&m
->lock
, flags
);
1243 if (!pgpath
->is_active
)
1246 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1248 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1249 pgpath
->is_active
= false;
1250 pgpath
->fail_count
++;
1252 atomic_dec(&m
->nr_valid_paths
);
1254 if (pgpath
== m
->current_pgpath
)
1255 m
->current_pgpath
= NULL
;
1257 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1258 pgpath
->path
.dev
->name
, atomic_read(&m
->nr_valid_paths
));
1260 schedule_work(&m
->trigger_event
);
1263 spin_unlock_irqrestore(&m
->lock
, flags
);
1269 * Reinstate a previously-failed path
1271 static int reinstate_path(struct pgpath
*pgpath
)
1273 int r
= 0, run_queue
= 0;
1274 unsigned long flags
;
1275 struct multipath
*m
= pgpath
->pg
->m
;
1276 unsigned nr_valid_paths
;
1278 spin_lock_irqsave(&m
->lock
, flags
);
1280 if (pgpath
->is_active
)
1283 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1285 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1289 pgpath
->is_active
= true;
1291 nr_valid_paths
= atomic_inc_return(&m
->nr_valid_paths
);
1292 if (nr_valid_paths
== 1) {
1293 m
->current_pgpath
= NULL
;
1295 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1296 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1297 atomic_inc(&m
->pg_init_in_progress
);
1300 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1301 pgpath
->path
.dev
->name
, nr_valid_paths
);
1303 schedule_work(&m
->trigger_event
);
1306 spin_unlock_irqrestore(&m
->lock
, flags
);
1308 dm_table_run_md_queue_async(m
->ti
->table
);
1309 process_queued_io_list(m
);
1316 * Fail or reinstate all paths that match the provided struct dm_dev.
1318 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1322 struct pgpath
*pgpath
;
1323 struct priority_group
*pg
;
1325 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1326 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1327 if (pgpath
->path
.dev
== dev
)
1336 * Temporarily try to avoid having to use the specified PG
1338 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1341 unsigned long flags
;
1343 spin_lock_irqsave(&m
->lock
, flags
);
1345 pg
->bypassed
= bypassed
;
1346 m
->current_pgpath
= NULL
;
1347 m
->current_pg
= NULL
;
1349 spin_unlock_irqrestore(&m
->lock
, flags
);
1351 schedule_work(&m
->trigger_event
);
1355 * Switch to using the specified PG from the next I/O that gets mapped
1357 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1359 struct priority_group
*pg
;
1361 unsigned long flags
;
1364 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1365 (pgnum
> m
->nr_priority_groups
)) {
1366 DMWARN("invalid PG number supplied to switch_pg_num");
1370 spin_lock_irqsave(&m
->lock
, flags
);
1371 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1372 pg
->bypassed
= false;
1376 m
->current_pgpath
= NULL
;
1377 m
->current_pg
= NULL
;
1380 spin_unlock_irqrestore(&m
->lock
, flags
);
1382 schedule_work(&m
->trigger_event
);
1387 * Set/clear bypassed status of a PG.
1388 * PGs are numbered upwards from 1 in the order they were declared.
1390 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1392 struct priority_group
*pg
;
1396 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1397 (pgnum
> m
->nr_priority_groups
)) {
1398 DMWARN("invalid PG number supplied to bypass_pg");
1402 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1407 bypass_pg(m
, pg
, bypassed
);
1412 * Should we retry pg_init immediately?
1414 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1416 unsigned long flags
;
1417 bool limit_reached
= false;
1419 spin_lock_irqsave(&m
->lock
, flags
);
1421 if (atomic_read(&m
->pg_init_count
) <= m
->pg_init_retries
&&
1422 !test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
1423 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
1425 limit_reached
= true;
1427 spin_unlock_irqrestore(&m
->lock
, flags
);
1429 return limit_reached
;
1432 static void pg_init_done(void *data
, int errors
)
1434 struct pgpath
*pgpath
= data
;
1435 struct priority_group
*pg
= pgpath
->pg
;
1436 struct multipath
*m
= pg
->m
;
1437 unsigned long flags
;
1438 bool delay_retry
= false;
1440 /* device or driver problems */
1445 if (!m
->hw_handler_name
) {
1449 DMERR("Could not failover the device: Handler scsi_dh_%s "
1450 "Error %d.", m
->hw_handler_name
, errors
);
1452 * Fail path for now, so we do not ping pong
1456 case SCSI_DH_DEV_TEMP_BUSY
:
1458 * Probably doing something like FW upgrade on the
1459 * controller so try the other pg.
1461 bypass_pg(m
, pg
, true);
1464 /* Wait before retrying. */
1466 case SCSI_DH_IMM_RETRY
:
1467 case SCSI_DH_RES_TEMP_UNAVAIL
:
1468 if (pg_init_limit_reached(m
, pgpath
))
1472 case SCSI_DH_DEV_OFFLINED
:
1475 * We probably do not want to fail the path for a device
1476 * error, but this is what the old dm did. In future
1477 * patches we can do more advanced handling.
1482 spin_lock_irqsave(&m
->lock
, flags
);
1484 if (pgpath
== m
->current_pgpath
) {
1485 DMERR("Could not failover device. Error %d.", errors
);
1486 m
->current_pgpath
= NULL
;
1487 m
->current_pg
= NULL
;
1489 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1490 pg
->bypassed
= false;
1492 if (atomic_dec_return(&m
->pg_init_in_progress
) > 0)
1493 /* Activations of other paths are still on going */
1496 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
1498 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1500 clear_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1502 if (__pg_init_all_paths(m
))
1505 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
1507 process_queued_io_list(m
);
1510 * Wake up any thread waiting to suspend.
1512 wake_up(&m
->pg_init_wait
);
1515 spin_unlock_irqrestore(&m
->lock
, flags
);
1518 static void activate_path(struct work_struct
*work
)
1520 struct pgpath
*pgpath
=
1521 container_of(work
, struct pgpath
, activate_path
.work
);
1522 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1524 if (pgpath
->is_active
&& !blk_queue_dying(q
))
1525 scsi_dh_activate(q
, pg_init_done
, pgpath
);
1527 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1530 static int noretry_error(int error
)
1535 * EBADE signals an reservation conflict.
1536 * We shouldn't fail the path here as we can communicate with
1537 * the target. We should failover to the next path, but in
1538 * doing so we might be causing a ping-pong between paths.
1539 * So just return the reservation conflict error.
1549 /* Anything else could be a path failure, so should be retried */
1556 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1557 int error
, struct dm_mpath_io
*mpio
)
1560 * We don't queue any clone request inside the multipath target
1561 * during end I/O handling, since those clone requests don't have
1562 * bio clones. If we queue them inside the multipath target,
1563 * we need to make bio clones, that requires memory allocation.
1564 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1565 * don't have bio clones.)
1566 * Instead of queueing the clone request here, we queue the original
1567 * request into dm core, which will remake a clone request and
1568 * clone bios for it and resubmit it later.
1570 int r
= DM_ENDIO_REQUEUE
;
1572 if (!error
&& !clone
->errors
)
1573 return 0; /* I/O complete */
1575 if (noretry_error(error
))
1579 fail_path(mpio
->pgpath
);
1581 if (!atomic_read(&m
->nr_valid_paths
)) {
1582 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1583 if (!must_push_back_rq(m
))
1591 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1592 int error
, union map_info
*map_context
)
1594 struct multipath
*m
= ti
->private;
1595 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1596 struct pgpath
*pgpath
;
1597 struct path_selector
*ps
;
1602 r
= do_end_io(m
, clone
, error
, mpio
);
1603 pgpath
= mpio
->pgpath
;
1605 ps
= &pgpath
->pg
->ps
;
1606 if (ps
->type
->end_io
)
1607 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1609 clear_request_fn_mpio(m
, map_context
);
1614 static int do_end_io_bio(struct multipath
*m
, struct bio
*clone
,
1615 int error
, struct dm_mpath_io
*mpio
)
1617 unsigned long flags
;
1620 return 0; /* I/O complete */
1622 if (noretry_error(error
))
1626 fail_path(mpio
->pgpath
);
1628 if (!atomic_read(&m
->nr_valid_paths
)) {
1629 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1630 if (!must_push_back_bio(m
))
1632 return DM_ENDIO_REQUEUE
;
1636 /* Queue for the daemon to resubmit */
1637 dm_bio_restore(get_bio_details_from_bio(clone
), clone
);
1639 spin_lock_irqsave(&m
->lock
, flags
);
1640 bio_list_add(&m
->queued_bios
, clone
);
1641 spin_unlock_irqrestore(&m
->lock
, flags
);
1642 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
1643 queue_work(kmultipathd
, &m
->process_queued_bios
);
1645 return DM_ENDIO_INCOMPLETE
;
1648 static int multipath_end_io_bio(struct dm_target
*ti
, struct bio
*clone
, int error
)
1650 struct multipath
*m
= ti
->private;
1651 struct dm_mpath_io
*mpio
= get_mpio_from_bio(clone
);
1652 struct pgpath
*pgpath
;
1653 struct path_selector
*ps
;
1658 r
= do_end_io_bio(m
, clone
, error
, mpio
);
1659 pgpath
= mpio
->pgpath
;
1661 ps
= &pgpath
->pg
->ps
;
1662 if (ps
->type
->end_io
)
1663 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1670 * Suspend can't complete until all the I/O is processed so if
1671 * the last path fails we must error any remaining I/O.
1672 * Note that if the freeze_bdev fails while suspending, the
1673 * queue_if_no_path state is lost - userspace should reset it.
1675 static void multipath_presuspend(struct dm_target
*ti
)
1677 struct multipath
*m
= ti
->private;
1679 queue_if_no_path(m
, false, true);
1682 static void multipath_postsuspend(struct dm_target
*ti
)
1684 struct multipath
*m
= ti
->private;
1686 mutex_lock(&m
->work_mutex
);
1687 flush_multipath_work(m
);
1688 mutex_unlock(&m
->work_mutex
);
1692 * Restore the queue_if_no_path setting.
1694 static void multipath_resume(struct dm_target
*ti
)
1696 struct multipath
*m
= ti
->private;
1697 unsigned long flags
;
1699 spin_lock_irqsave(&m
->lock
, flags
);
1700 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
))
1701 set_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
1703 clear_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
1704 spin_unlock_irqrestore(&m
->lock
, flags
);
1708 * Info output has the following format:
1709 * num_multipath_feature_args [multipath_feature_args]*
1710 * num_handler_status_args [handler_status_args]*
1711 * num_groups init_group_number
1712 * [A|D|E num_ps_status_args [ps_status_args]*
1713 * num_paths num_selector_args
1714 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1716 * Table output has the following format (identical to the constructor string):
1717 * num_feature_args [features_args]*
1718 * num_handler_args hw_handler [hw_handler_args]*
1719 * num_groups init_group_number
1720 * [priority selector-name num_ps_args [ps_args]*
1721 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1723 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1724 unsigned status_flags
, char *result
, unsigned maxlen
)
1727 unsigned long flags
;
1728 struct multipath
*m
= ti
->private;
1729 struct priority_group
*pg
;
1734 spin_lock_irqsave(&m
->lock
, flags
);
1737 if (type
== STATUSTYPE_INFO
)
1738 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO
, &m
->flags
),
1739 atomic_read(&m
->pg_init_count
));
1741 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) +
1742 (m
->pg_init_retries
> 0) * 2 +
1743 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1744 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) +
1745 (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) * 2);
1747 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1748 DMEMIT("queue_if_no_path ");
1749 if (m
->pg_init_retries
)
1750 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1751 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1752 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1753 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
))
1754 DMEMIT("retain_attached_hw_handler ");
1755 if (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) {
1756 switch(m
->queue_mode
) {
1757 case DM_TYPE_BIO_BASED
:
1758 DMEMIT("queue_mode bio ");
1760 case DM_TYPE_MQ_REQUEST_BASED
:
1761 DMEMIT("queue_mode mq ");
1767 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1770 DMEMIT("1 %s ", m
->hw_handler_name
);
1772 DMEMIT("%u ", m
->nr_priority_groups
);
1775 pg_num
= m
->next_pg
->pg_num
;
1776 else if (m
->current_pg
)
1777 pg_num
= m
->current_pg
->pg_num
;
1779 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1781 DMEMIT("%u ", pg_num
);
1784 case STATUSTYPE_INFO
:
1785 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1787 state
= 'D'; /* Disabled */
1788 else if (pg
== m
->current_pg
)
1789 state
= 'A'; /* Currently Active */
1791 state
= 'E'; /* Enabled */
1793 DMEMIT("%c ", state
);
1795 if (pg
->ps
.type
->status
)
1796 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1802 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1803 pg
->ps
.type
->info_args
);
1805 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1806 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1807 p
->is_active
? "A" : "F",
1809 if (pg
->ps
.type
->status
)
1810 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1811 &p
->path
, type
, result
+ sz
,
1817 case STATUSTYPE_TABLE
:
1818 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1819 DMEMIT("%s ", pg
->ps
.type
->name
);
1821 if (pg
->ps
.type
->status
)
1822 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1828 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1829 pg
->ps
.type
->table_args
);
1831 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1832 DMEMIT("%s ", p
->path
.dev
->name
);
1833 if (pg
->ps
.type
->status
)
1834 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1835 &p
->path
, type
, result
+ sz
,
1842 spin_unlock_irqrestore(&m
->lock
, flags
);
1845 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1849 struct multipath
*m
= ti
->private;
1852 mutex_lock(&m
->work_mutex
);
1854 if (dm_suspended(ti
)) {
1860 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1861 r
= queue_if_no_path(m
, true, false);
1863 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1864 r
= queue_if_no_path(m
, false, false);
1870 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1874 if (!strcasecmp(argv
[0], "disable_group")) {
1875 r
= bypass_pg_num(m
, argv
[1], true);
1877 } else if (!strcasecmp(argv
[0], "enable_group")) {
1878 r
= bypass_pg_num(m
, argv
[1], false);
1880 } else if (!strcasecmp(argv
[0], "switch_group")) {
1881 r
= switch_pg_num(m
, argv
[1]);
1883 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1884 action
= reinstate_path
;
1885 else if (!strcasecmp(argv
[0], "fail_path"))
1888 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1892 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1894 DMWARN("message: error getting device %s",
1899 r
= action_dev(m
, dev
, action
);
1901 dm_put_device(ti
, dev
);
1904 mutex_unlock(&m
->work_mutex
);
1908 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1909 struct block_device
**bdev
, fmode_t
*mode
)
1911 struct multipath
*m
= ti
->private;
1912 struct pgpath
*current_pgpath
;
1915 current_pgpath
= lockless_dereference(m
->current_pgpath
);
1916 if (!current_pgpath
)
1917 current_pgpath
= choose_pgpath(m
, 0);
1919 if (current_pgpath
) {
1920 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
)) {
1921 *bdev
= current_pgpath
->path
.dev
->bdev
;
1922 *mode
= current_pgpath
->path
.dev
->mode
;
1925 /* pg_init has not started or completed */
1929 /* No path is available */
1930 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1936 if (r
== -ENOTCONN
) {
1937 if (!lockless_dereference(m
->current_pg
)) {
1938 /* Path status changed, redo selection */
1939 (void) choose_pgpath(m
, 0);
1941 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1942 pg_init_all_paths(m
);
1943 dm_table_run_md_queue_async(m
->ti
->table
);
1944 process_queued_io_list(m
);
1948 * Only pass ioctls through if the device sizes match exactly.
1950 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1955 static int multipath_iterate_devices(struct dm_target
*ti
,
1956 iterate_devices_callout_fn fn
, void *data
)
1958 struct multipath
*m
= ti
->private;
1959 struct priority_group
*pg
;
1963 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1964 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1965 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1975 static int pgpath_busy(struct pgpath
*pgpath
)
1977 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1979 return blk_lld_busy(q
);
1983 * We return "busy", only when we can map I/Os but underlying devices
1984 * are busy (so even if we map I/Os now, the I/Os will wait on
1985 * the underlying queue).
1986 * In other words, if we want to kill I/Os or queue them inside us
1987 * due to map unavailability, we don't return "busy". Otherwise,
1988 * dm core won't give us the I/Os and we can't do what we want.
1990 static int multipath_busy(struct dm_target
*ti
)
1992 bool busy
= false, has_active
= false;
1993 struct multipath
*m
= ti
->private;
1994 struct priority_group
*pg
, *next_pg
;
1995 struct pgpath
*pgpath
;
1997 /* pg_init in progress */
1998 if (atomic_read(&m
->pg_init_in_progress
))
2001 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2002 if (!atomic_read(&m
->nr_valid_paths
) && test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
2003 return (m
->queue_mode
!= DM_TYPE_MQ_REQUEST_BASED
);
2005 /* Guess which priority_group will be used at next mapping time */
2006 pg
= lockless_dereference(m
->current_pg
);
2007 next_pg
= lockless_dereference(m
->next_pg
);
2008 if (unlikely(!lockless_dereference(m
->current_pgpath
) && next_pg
))
2013 * We don't know which pg will be used at next mapping time.
2014 * We don't call choose_pgpath() here to avoid to trigger
2015 * pg_init just by busy checking.
2016 * So we don't know whether underlying devices we will be using
2017 * at next mapping time are busy or not. Just try mapping.
2023 * If there is one non-busy active path at least, the path selector
2024 * will be able to select it. So we consider such a pg as not busy.
2027 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
2028 if (pgpath
->is_active
) {
2030 if (!pgpath_busy(pgpath
)) {
2039 * No active path in this pg, so this pg won't be used and
2040 * the current_pg will be changed at next mapping time.
2041 * We need to try mapping to determine it.
2049 /*-----------------------------------------------------------------
2051 *---------------------------------------------------------------*/
2052 static struct target_type multipath_target
= {
2053 .name
= "multipath",
2054 .version
= {1, 12, 0},
2055 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
2056 .module
= THIS_MODULE
,
2057 .ctr
= multipath_ctr
,
2058 .dtr
= multipath_dtr
,
2059 .map_rq
= multipath_map
,
2060 .clone_and_map_rq
= multipath_clone_and_map
,
2061 .release_clone_rq
= multipath_release_clone
,
2062 .rq_end_io
= multipath_end_io
,
2063 .map
= multipath_map_bio
,
2064 .end_io
= multipath_end_io_bio
,
2065 .presuspend
= multipath_presuspend
,
2066 .postsuspend
= multipath_postsuspend
,
2067 .resume
= multipath_resume
,
2068 .status
= multipath_status
,
2069 .message
= multipath_message
,
2070 .prepare_ioctl
= multipath_prepare_ioctl
,
2071 .iterate_devices
= multipath_iterate_devices
,
2072 .busy
= multipath_busy
,
2075 static int __init
dm_multipath_init(void)
2079 /* allocate a slab for the dm_mpath_ios */
2080 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
2084 r
= dm_register_target(&multipath_target
);
2086 DMERR("request-based register failed %d", r
);
2088 goto bad_register_target
;
2091 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
2093 DMERR("failed to create workqueue kmpathd");
2095 goto bad_alloc_kmultipathd
;
2099 * A separate workqueue is used to handle the device handlers
2100 * to avoid overloading existing workqueue. Overloading the
2101 * old workqueue would also create a bottleneck in the
2102 * path of the storage hardware device activation.
2104 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
2106 if (!kmpath_handlerd
) {
2107 DMERR("failed to create workqueue kmpath_handlerd");
2109 goto bad_alloc_kmpath_handlerd
;
2114 bad_alloc_kmpath_handlerd
:
2115 destroy_workqueue(kmultipathd
);
2116 bad_alloc_kmultipathd
:
2117 dm_unregister_target(&multipath_target
);
2118 bad_register_target
:
2119 kmem_cache_destroy(_mpio_cache
);
2124 static void __exit
dm_multipath_exit(void)
2126 destroy_workqueue(kmpath_handlerd
);
2127 destroy_workqueue(kmultipathd
);
2129 dm_unregister_target(&multipath_target
);
2130 kmem_cache_destroy(_mpio_cache
);
2133 module_init(dm_multipath_init
);
2134 module_exit(dm_multipath_exit
);
2136 MODULE_DESCRIPTION(DM_NAME
" multipath target");
2137 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2138 MODULE_LICENSE("GPL");