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-path-selector.h"
12 #include "dm-uevent.h"
14 #include <linux/blkdev.h>
15 #include <linux/ctype.h>
16 #include <linux/init.h>
17 #include <linux/mempool.h>
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
20 #include <linux/slab.h>
21 #include <linux/time.h>
22 #include <linux/workqueue.h>
23 #include <linux/delay.h>
24 #include <scsi/scsi_dh.h>
25 #include <linux/atomic.h>
27 #define DM_MSG_PREFIX "multipath"
28 #define DM_PG_INIT_DELAY_MSECS 2000
29 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
33 struct list_head list
;
35 struct priority_group
*pg
; /* Owning PG */
36 unsigned is_active
; /* Path status */
37 unsigned fail_count
; /* Cumulative failure count */
40 struct delayed_work activate_path
;
43 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
46 * Paths are grouped into Priority Groups and numbered from 1 upwards.
47 * Each has a path selector which controls which path gets used.
49 struct priority_group
{
50 struct list_head list
;
52 struct multipath
*m
; /* Owning multipath instance */
53 struct path_selector ps
;
55 unsigned pg_num
; /* Reference number */
56 unsigned bypassed
; /* Temporarily bypass this PG? */
58 unsigned nr_pgpaths
; /* Number of paths in PG */
59 struct list_head pgpaths
;
62 /* Multipath context */
64 struct list_head list
;
67 const char *hw_handler_name
;
68 char *hw_handler_params
;
72 unsigned nr_priority_groups
;
73 struct list_head priority_groups
;
75 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
77 unsigned pg_init_required
; /* pg_init needs calling? */
78 unsigned pg_init_in_progress
; /* Only one pg_init allowed at once */
79 unsigned pg_init_delay_retry
; /* Delay pg_init retry? */
81 unsigned nr_valid_paths
; /* Total number of usable paths */
82 struct pgpath
*current_pgpath
;
83 struct priority_group
*current_pg
;
84 struct priority_group
*next_pg
; /* Switch to this PG if set */
85 unsigned repeat_count
; /* I/Os left before calling PS again */
87 unsigned queue_io
:1; /* Must we queue all I/O? */
88 unsigned queue_if_no_path
:1; /* Queue I/O if last path fails? */
89 unsigned saved_queue_if_no_path
:1; /* Saved state during suspension */
90 unsigned retain_attached_hw_handler
:1; /* If there's already a hw_handler present, don't change it. */
91 unsigned pg_init_disabled
:1; /* pg_init is not currently allowed */
93 unsigned pg_init_retries
; /* Number of times to retry pg_init */
94 unsigned pg_init_count
; /* Number of times pg_init called */
95 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
97 struct work_struct trigger_event
;
100 * We must use a mempool of dm_mpath_io structs so that we
101 * can resubmit bios on error.
103 mempool_t
*mpio_pool
;
105 struct mutex work_mutex
;
109 * Context information attached to each bio 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 int __pgpath_busy(struct pgpath
*pgpath
);
126 /*-----------------------------------------------
127 * Allocation routines
128 *-----------------------------------------------*/
130 static struct pgpath
*alloc_pgpath(void)
132 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
135 pgpath
->is_active
= 1;
136 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path
);
142 static void free_pgpath(struct pgpath
*pgpath
)
147 static struct priority_group
*alloc_priority_group(void)
149 struct priority_group
*pg
;
151 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
154 INIT_LIST_HEAD(&pg
->pgpaths
);
159 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
161 struct pgpath
*pgpath
, *tmp
;
163 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
164 list_del(&pgpath
->list
);
165 dm_put_device(ti
, pgpath
->path
.dev
);
170 static void free_priority_group(struct priority_group
*pg
,
171 struct dm_target
*ti
)
173 struct path_selector
*ps
= &pg
->ps
;
176 ps
->type
->destroy(ps
);
177 dm_put_path_selector(ps
->type
);
180 free_pgpaths(&pg
->pgpaths
, ti
);
184 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
187 unsigned min_ios
= dm_get_reserved_rq_based_ios();
189 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
191 INIT_LIST_HEAD(&m
->priority_groups
);
192 spin_lock_init(&m
->lock
);
194 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
195 INIT_WORK(&m
->trigger_event
, trigger_event
);
196 init_waitqueue_head(&m
->pg_init_wait
);
197 mutex_init(&m
->work_mutex
);
198 m
->mpio_pool
= mempool_create_slab_pool(min_ios
, _mpio_cache
);
210 static void free_multipath(struct multipath
*m
)
212 struct priority_group
*pg
, *tmp
;
214 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
216 free_priority_group(pg
, m
->ti
);
219 kfree(m
->hw_handler_name
);
220 kfree(m
->hw_handler_params
);
221 mempool_destroy(m
->mpio_pool
);
225 static int set_mapinfo(struct multipath
*m
, union map_info
*info
)
227 struct dm_mpath_io
*mpio
;
229 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
233 memset(mpio
, 0, sizeof(*mpio
));
239 static void clear_mapinfo(struct multipath
*m
, union map_info
*info
)
241 struct dm_mpath_io
*mpio
= info
->ptr
;
244 mempool_free(mpio
, m
->mpio_pool
);
247 /*-----------------------------------------------
249 *-----------------------------------------------*/
251 static int __pg_init_all_paths(struct multipath
*m
)
253 struct pgpath
*pgpath
;
254 unsigned long pg_init_delay
= 0;
256 if (m
->pg_init_in_progress
|| m
->pg_init_disabled
)
260 m
->pg_init_required
= 0;
262 /* Check here to reset pg_init_required */
266 if (m
->pg_init_delay_retry
)
267 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
268 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
269 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
270 /* Skip failed paths */
271 if (!pgpath
->is_active
)
273 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
275 m
->pg_init_in_progress
++;
277 return m
->pg_init_in_progress
;
280 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
282 m
->current_pg
= pgpath
->pg
;
284 /* Must we initialise the PG first, and queue I/O till it's ready? */
285 if (m
->hw_handler_name
) {
286 m
->pg_init_required
= 1;
289 m
->pg_init_required
= 0;
293 m
->pg_init_count
= 0;
296 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
299 struct dm_path
*path
;
301 path
= pg
->ps
.type
->select_path(&pg
->ps
, &m
->repeat_count
, nr_bytes
);
305 m
->current_pgpath
= path_to_pgpath(path
);
307 if (m
->current_pg
!= pg
)
308 __switch_pg(m
, m
->current_pgpath
);
313 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
315 struct priority_group
*pg
;
316 unsigned bypassed
= 1;
318 if (!m
->nr_valid_paths
) {
323 /* Were we instructed to switch PG? */
327 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
331 /* Don't change PG until it has no remaining paths */
332 if (m
->current_pg
&& !__choose_path_in_pg(m
, m
->current_pg
, nr_bytes
))
336 * Loop through priority groups until we find a valid path.
337 * First time we skip PGs marked 'bypassed'.
338 * Second time we only try the ones we skipped, but set
339 * pg_init_delay_retry so we do not hammer controllers.
342 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
343 if (pg
->bypassed
== bypassed
)
345 if (!__choose_path_in_pg(m
, pg
, nr_bytes
)) {
347 m
->pg_init_delay_retry
= 1;
351 } while (bypassed
--);
354 m
->current_pgpath
= NULL
;
355 m
->current_pg
= NULL
;
359 * Check whether bios must be queued in the device-mapper core rather
360 * than here in the target.
362 * m->lock must be held on entry.
364 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
365 * same value then we are not between multipath_presuspend()
366 * and multipath_resume() calls and we have no need to check
367 * for the DMF_NOFLUSH_SUSPENDING flag.
369 static int __must_push_back(struct multipath
*m
)
371 return (m
->queue_if_no_path
||
372 (m
->queue_if_no_path
!= m
->saved_queue_if_no_path
&&
373 dm_noflush_suspending(m
->ti
)));
377 * Map cloned requests
379 static int __multipath_map(struct dm_target
*ti
, struct request
*clone
,
380 union map_info
*map_context
,
381 struct request
*rq
, struct request
**__clone
)
383 struct multipath
*m
= (struct multipath
*) ti
->private;
384 int r
= DM_MAPIO_REQUEUE
;
385 size_t nr_bytes
= clone
? blk_rq_bytes(clone
) : blk_rq_bytes(rq
);
386 struct pgpath
*pgpath
;
387 struct block_device
*bdev
;
388 struct dm_mpath_io
*mpio
;
390 spin_lock_irq(&m
->lock
);
392 /* Do we need to select a new pgpath? */
393 if (!m
->current_pgpath
||
394 (!m
->queue_io
&& (m
->repeat_count
&& --m
->repeat_count
== 0)))
395 __choose_pgpath(m
, nr_bytes
);
397 pgpath
= m
->current_pgpath
;
400 if (!__must_push_back(m
))
401 r
= -EIO
; /* Failed */
403 } else if (m
->queue_io
|| m
->pg_init_required
) {
404 __pg_init_all_paths(m
);
408 if (set_mapinfo(m
, map_context
) < 0)
409 /* ENOMEM, requeue */
412 mpio
= map_context
->ptr
;
413 mpio
->pgpath
= pgpath
;
414 mpio
->nr_bytes
= nr_bytes
;
416 bdev
= pgpath
->path
.dev
->bdev
;
418 spin_unlock_irq(&m
->lock
);
422 * Old request-based interface: allocated clone is passed in.
423 * Used by: .request_fn stacked on .request_fn path(s).
425 clone
->q
= bdev_get_queue(bdev
);
426 clone
->rq_disk
= bdev
->bd_disk
;
427 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
429 /* blk-mq request-based interface */
430 *__clone
= blk_get_request(bdev_get_queue(bdev
),
431 rq_data_dir(rq
), GFP_ATOMIC
);
432 if (IS_ERR(*__clone
)) {
433 /* ENOMEM, requeue */
434 clear_mapinfo(m
, map_context
);
437 (*__clone
)->bio
= (*__clone
)->biotail
= NULL
;
438 (*__clone
)->rq_disk
= bdev
->bd_disk
;
439 (*__clone
)->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
442 if (pgpath
->pg
->ps
.type
->start_io
)
443 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
446 return DM_MAPIO_REMAPPED
;
449 spin_unlock_irq(&m
->lock
);
454 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
455 union map_info
*map_context
)
457 return __multipath_map(ti
, clone
, map_context
, NULL
, NULL
);
460 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
461 union map_info
*map_context
,
462 struct request
**clone
)
464 return __multipath_map(ti
, NULL
, map_context
, rq
, clone
);
467 static void multipath_release_clone(struct request
*clone
)
469 blk_put_request(clone
);
473 * If we run out of usable paths, should we queue I/O or error it?
475 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
476 unsigned save_old_value
)
480 spin_lock_irqsave(&m
->lock
, flags
);
483 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
485 m
->saved_queue_if_no_path
= queue_if_no_path
;
486 m
->queue_if_no_path
= queue_if_no_path
;
487 spin_unlock_irqrestore(&m
->lock
, flags
);
489 if (!queue_if_no_path
)
490 dm_table_run_md_queue_async(m
->ti
->table
);
496 * An event is triggered whenever a path is taken out of use.
497 * Includes path failure and PG bypass.
499 static void trigger_event(struct work_struct
*work
)
501 struct multipath
*m
=
502 container_of(work
, struct multipath
, trigger_event
);
504 dm_table_event(m
->ti
->table
);
507 /*-----------------------------------------------------------------
508 * Constructor/argument parsing:
509 * <#multipath feature args> [<arg>]*
510 * <#hw_handler args> [hw_handler [<arg>]*]
512 * <initial priority group>
513 * [<selector> <#selector args> [<arg>]*
514 * <#paths> <#per-path selector args>
515 * [<path> [<arg>]* ]+ ]+
516 *---------------------------------------------------------------*/
517 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
518 struct dm_target
*ti
)
521 struct path_selector_type
*pst
;
524 static struct dm_arg _args
[] = {
525 {0, 1024, "invalid number of path selector args"},
528 pst
= dm_get_path_selector(dm_shift_arg(as
));
530 ti
->error
= "unknown path selector type";
534 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
536 dm_put_path_selector(pst
);
540 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
542 dm_put_path_selector(pst
);
543 ti
->error
= "path selector constructor failed";
548 dm_consume_args(as
, ps_argc
);
553 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
554 struct dm_target
*ti
)
558 struct multipath
*m
= ti
->private;
559 struct request_queue
*q
= NULL
;
560 const char *attached_handler_name
;
562 /* we need at least a path arg */
564 ti
->error
= "no device given";
565 return ERR_PTR(-EINVAL
);
570 return ERR_PTR(-ENOMEM
);
572 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
575 ti
->error
= "error getting device";
579 if (m
->retain_attached_hw_handler
|| m
->hw_handler_name
)
580 q
= bdev_get_queue(p
->path
.dev
->bdev
);
582 if (m
->retain_attached_hw_handler
) {
584 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
585 if (attached_handler_name
) {
587 * Reset hw_handler_name to match the attached handler
588 * and clear any hw_handler_params associated with the
591 * NB. This modifies the table line to show the actual
592 * handler instead of the original table passed in.
594 kfree(m
->hw_handler_name
);
595 m
->hw_handler_name
= attached_handler_name
;
597 kfree(m
->hw_handler_params
);
598 m
->hw_handler_params
= NULL
;
602 if (m
->hw_handler_name
) {
603 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
605 char b
[BDEVNAME_SIZE
];
607 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
608 bdevname(p
->path
.dev
->bdev
, b
));
612 ti
->error
= "error attaching hardware handler";
613 dm_put_device(ti
, p
->path
.dev
);
617 if (m
->hw_handler_params
) {
618 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
620 ti
->error
= "unable to set hardware "
621 "handler parameters";
622 dm_put_device(ti
, p
->path
.dev
);
628 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
630 dm_put_device(ti
, p
->path
.dev
);
641 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
644 static struct dm_arg _args
[] = {
645 {1, 1024, "invalid number of paths"},
646 {0, 1024, "invalid number of selector args"}
650 unsigned i
, nr_selector_args
, nr_args
;
651 struct priority_group
*pg
;
652 struct dm_target
*ti
= m
->ti
;
656 ti
->error
= "not enough priority group arguments";
657 return ERR_PTR(-EINVAL
);
660 pg
= alloc_priority_group();
662 ti
->error
= "couldn't allocate priority group";
663 return ERR_PTR(-ENOMEM
);
667 r
= parse_path_selector(as
, pg
, ti
);
674 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
678 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
682 nr_args
= 1 + nr_selector_args
;
683 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
684 struct pgpath
*pgpath
;
685 struct dm_arg_set path_args
;
687 if (as
->argc
< nr_args
) {
688 ti
->error
= "not enough path parameters";
693 path_args
.argc
= nr_args
;
694 path_args
.argv
= as
->argv
;
696 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
697 if (IS_ERR(pgpath
)) {
703 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
704 dm_consume_args(as
, nr_args
);
710 free_priority_group(pg
, ti
);
714 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
718 struct dm_target
*ti
= m
->ti
;
720 static struct dm_arg _args
[] = {
721 {0, 1024, "invalid number of hardware handler args"},
724 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
730 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
736 for (i
= 0; i
<= hw_argc
- 2; i
++)
737 len
+= strlen(as
->argv
[i
]) + 1;
738 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
740 ti
->error
= "memory allocation failed";
744 j
= sprintf(p
, "%d", hw_argc
- 1);
745 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
746 j
= sprintf(p
, "%s", as
->argv
[i
]);
748 dm_consume_args(as
, hw_argc
- 1);
752 kfree(m
->hw_handler_name
);
753 m
->hw_handler_name
= NULL
;
757 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
761 struct dm_target
*ti
= m
->ti
;
762 const char *arg_name
;
764 static struct dm_arg _args
[] = {
765 {0, 6, "invalid number of feature args"},
766 {1, 50, "pg_init_retries must be between 1 and 50"},
767 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
770 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
778 arg_name
= dm_shift_arg(as
);
781 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
782 r
= queue_if_no_path(m
, 1, 0);
786 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
787 m
->retain_attached_hw_handler
= 1;
791 if (!strcasecmp(arg_name
, "pg_init_retries") &&
793 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
798 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
800 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
805 ti
->error
= "Unrecognised multipath feature request";
807 } while (argc
&& !r
);
812 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
815 /* target arguments */
816 static struct dm_arg _args
[] = {
817 {0, 1024, "invalid number of priority groups"},
818 {0, 1024, "invalid initial priority group number"},
823 struct dm_arg_set as
;
824 unsigned pg_count
= 0;
825 unsigned next_pg_num
;
830 m
= alloc_multipath(ti
);
832 ti
->error
= "can't allocate multipath";
836 r
= parse_features(&as
, m
);
840 r
= parse_hw_handler(&as
, m
);
844 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
848 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
852 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
853 (m
->nr_priority_groups
&& !next_pg_num
)) {
854 ti
->error
= "invalid initial priority group";
859 /* parse the priority groups */
861 struct priority_group
*pg
;
863 pg
= parse_priority_group(&as
, m
);
869 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
870 list_add_tail(&pg
->list
, &m
->priority_groups
);
872 pg
->pg_num
= pg_count
;
877 if (pg_count
!= m
->nr_priority_groups
) {
878 ti
->error
= "priority group count mismatch";
883 ti
->num_flush_bios
= 1;
884 ti
->num_discard_bios
= 1;
885 ti
->num_write_same_bios
= 1;
894 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
896 DECLARE_WAITQUEUE(wait
, current
);
899 add_wait_queue(&m
->pg_init_wait
, &wait
);
902 set_current_state(TASK_UNINTERRUPTIBLE
);
904 spin_lock_irqsave(&m
->lock
, flags
);
905 if (!m
->pg_init_in_progress
) {
906 spin_unlock_irqrestore(&m
->lock
, flags
);
909 spin_unlock_irqrestore(&m
->lock
, flags
);
913 set_current_state(TASK_RUNNING
);
915 remove_wait_queue(&m
->pg_init_wait
, &wait
);
918 static void flush_multipath_work(struct multipath
*m
)
922 spin_lock_irqsave(&m
->lock
, flags
);
923 m
->pg_init_disabled
= 1;
924 spin_unlock_irqrestore(&m
->lock
, flags
);
926 flush_workqueue(kmpath_handlerd
);
927 multipath_wait_for_pg_init_completion(m
);
928 flush_workqueue(kmultipathd
);
929 flush_work(&m
->trigger_event
);
931 spin_lock_irqsave(&m
->lock
, flags
);
932 m
->pg_init_disabled
= 0;
933 spin_unlock_irqrestore(&m
->lock
, flags
);
936 static void multipath_dtr(struct dm_target
*ti
)
938 struct multipath
*m
= ti
->private;
940 flush_multipath_work(m
);
945 * Take a path out of use.
947 static int fail_path(struct pgpath
*pgpath
)
950 struct multipath
*m
= pgpath
->pg
->m
;
952 spin_lock_irqsave(&m
->lock
, flags
);
954 if (!pgpath
->is_active
)
957 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
959 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
960 pgpath
->is_active
= 0;
961 pgpath
->fail_count
++;
965 if (pgpath
== m
->current_pgpath
)
966 m
->current_pgpath
= NULL
;
968 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
969 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
971 schedule_work(&m
->trigger_event
);
974 spin_unlock_irqrestore(&m
->lock
, flags
);
980 * Reinstate a previously-failed path
982 static int reinstate_path(struct pgpath
*pgpath
)
984 int r
= 0, run_queue
= 0;
986 struct multipath
*m
= pgpath
->pg
->m
;
988 spin_lock_irqsave(&m
->lock
, flags
);
990 if (pgpath
->is_active
)
993 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
994 DMWARN("Reinstate path not supported by path selector %s",
995 pgpath
->pg
->ps
.type
->name
);
1000 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1004 pgpath
->is_active
= 1;
1006 if (!m
->nr_valid_paths
++) {
1007 m
->current_pgpath
= NULL
;
1009 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1010 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1011 m
->pg_init_in_progress
++;
1014 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1015 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1017 schedule_work(&m
->trigger_event
);
1020 spin_unlock_irqrestore(&m
->lock
, flags
);
1022 dm_table_run_md_queue_async(m
->ti
->table
);
1028 * Fail or reinstate all paths that match the provided struct dm_dev.
1030 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1034 struct pgpath
*pgpath
;
1035 struct priority_group
*pg
;
1037 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1038 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1039 if (pgpath
->path
.dev
== dev
)
1048 * Temporarily try to avoid having to use the specified PG
1050 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1053 unsigned long flags
;
1055 spin_lock_irqsave(&m
->lock
, flags
);
1057 pg
->bypassed
= bypassed
;
1058 m
->current_pgpath
= NULL
;
1059 m
->current_pg
= NULL
;
1061 spin_unlock_irqrestore(&m
->lock
, flags
);
1063 schedule_work(&m
->trigger_event
);
1067 * Switch to using the specified PG from the next I/O that gets mapped
1069 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1071 struct priority_group
*pg
;
1073 unsigned long flags
;
1076 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1077 (pgnum
> m
->nr_priority_groups
)) {
1078 DMWARN("invalid PG number supplied to switch_pg_num");
1082 spin_lock_irqsave(&m
->lock
, flags
);
1083 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1088 m
->current_pgpath
= NULL
;
1089 m
->current_pg
= NULL
;
1092 spin_unlock_irqrestore(&m
->lock
, flags
);
1094 schedule_work(&m
->trigger_event
);
1099 * Set/clear bypassed status of a PG.
1100 * PGs are numbered upwards from 1 in the order they were declared.
1102 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1104 struct priority_group
*pg
;
1108 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1109 (pgnum
> m
->nr_priority_groups
)) {
1110 DMWARN("invalid PG number supplied to bypass_pg");
1114 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1119 bypass_pg(m
, pg
, bypassed
);
1124 * Should we retry pg_init immediately?
1126 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1128 unsigned long flags
;
1129 int limit_reached
= 0;
1131 spin_lock_irqsave(&m
->lock
, flags
);
1133 if (m
->pg_init_count
<= m
->pg_init_retries
&& !m
->pg_init_disabled
)
1134 m
->pg_init_required
= 1;
1138 spin_unlock_irqrestore(&m
->lock
, flags
);
1140 return limit_reached
;
1143 static void pg_init_done(void *data
, int errors
)
1145 struct pgpath
*pgpath
= data
;
1146 struct priority_group
*pg
= pgpath
->pg
;
1147 struct multipath
*m
= pg
->m
;
1148 unsigned long flags
;
1149 unsigned delay_retry
= 0;
1151 /* device or driver problems */
1156 if (!m
->hw_handler_name
) {
1160 DMERR("Could not failover the device: Handler scsi_dh_%s "
1161 "Error %d.", m
->hw_handler_name
, errors
);
1163 * Fail path for now, so we do not ping pong
1167 case SCSI_DH_DEV_TEMP_BUSY
:
1169 * Probably doing something like FW upgrade on the
1170 * controller so try the other pg.
1172 bypass_pg(m
, pg
, 1);
1175 /* Wait before retrying. */
1177 case SCSI_DH_IMM_RETRY
:
1178 case SCSI_DH_RES_TEMP_UNAVAIL
:
1179 if (pg_init_limit_reached(m
, pgpath
))
1185 * We probably do not want to fail the path for a device
1186 * error, but this is what the old dm did. In future
1187 * patches we can do more advanced handling.
1192 spin_lock_irqsave(&m
->lock
, flags
);
1194 if (pgpath
== m
->current_pgpath
) {
1195 DMERR("Could not failover device. Error %d.", errors
);
1196 m
->current_pgpath
= NULL
;
1197 m
->current_pg
= NULL
;
1199 } else if (!m
->pg_init_required
)
1202 if (--m
->pg_init_in_progress
)
1203 /* Activations of other paths are still on going */
1206 if (m
->pg_init_required
) {
1207 m
->pg_init_delay_retry
= delay_retry
;
1208 if (__pg_init_all_paths(m
))
1214 * Wake up any thread waiting to suspend.
1216 wake_up(&m
->pg_init_wait
);
1219 spin_unlock_irqrestore(&m
->lock
, flags
);
1222 static void activate_path(struct work_struct
*work
)
1224 struct pgpath
*pgpath
=
1225 container_of(work
, struct pgpath
, activate_path
.work
);
1227 if (pgpath
->is_active
)
1228 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1229 pg_init_done
, pgpath
);
1231 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1234 static int noretry_error(int error
)
1245 /* Anything else could be a path failure, so should be retried */
1252 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1253 int error
, struct dm_mpath_io
*mpio
)
1256 * We don't queue any clone request inside the multipath target
1257 * during end I/O handling, since those clone requests don't have
1258 * bio clones. If we queue them inside the multipath target,
1259 * we need to make bio clones, that requires memory allocation.
1260 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1261 * don't have bio clones.)
1262 * Instead of queueing the clone request here, we queue the original
1263 * request into dm core, which will remake a clone request and
1264 * clone bios for it and resubmit it later.
1266 int r
= DM_ENDIO_REQUEUE
;
1267 unsigned long flags
;
1269 if (!error
&& !clone
->errors
)
1270 return 0; /* I/O complete */
1272 if (noretry_error(error
))
1276 fail_path(mpio
->pgpath
);
1278 spin_lock_irqsave(&m
->lock
, flags
);
1279 if (!m
->nr_valid_paths
) {
1280 if (!m
->queue_if_no_path
) {
1281 if (!__must_push_back(m
))
1284 if (error
== -EBADE
)
1288 spin_unlock_irqrestore(&m
->lock
, flags
);
1293 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1294 int error
, union map_info
*map_context
)
1296 struct multipath
*m
= ti
->private;
1297 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1298 struct pgpath
*pgpath
;
1299 struct path_selector
*ps
;
1304 r
= do_end_io(m
, clone
, error
, mpio
);
1305 pgpath
= mpio
->pgpath
;
1307 ps
= &pgpath
->pg
->ps
;
1308 if (ps
->type
->end_io
)
1309 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1311 clear_mapinfo(m
, map_context
);
1317 * Suspend can't complete until all the I/O is processed so if
1318 * the last path fails we must error any remaining I/O.
1319 * Note that if the freeze_bdev fails while suspending, the
1320 * queue_if_no_path state is lost - userspace should reset it.
1322 static void multipath_presuspend(struct dm_target
*ti
)
1324 struct multipath
*m
= (struct multipath
*) ti
->private;
1326 queue_if_no_path(m
, 0, 1);
1329 static void multipath_postsuspend(struct dm_target
*ti
)
1331 struct multipath
*m
= ti
->private;
1333 mutex_lock(&m
->work_mutex
);
1334 flush_multipath_work(m
);
1335 mutex_unlock(&m
->work_mutex
);
1339 * Restore the queue_if_no_path setting.
1341 static void multipath_resume(struct dm_target
*ti
)
1343 struct multipath
*m
= (struct multipath
*) ti
->private;
1344 unsigned long flags
;
1346 spin_lock_irqsave(&m
->lock
, flags
);
1347 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1348 spin_unlock_irqrestore(&m
->lock
, flags
);
1352 * Info output has the following format:
1353 * num_multipath_feature_args [multipath_feature_args]*
1354 * num_handler_status_args [handler_status_args]*
1355 * num_groups init_group_number
1356 * [A|D|E num_ps_status_args [ps_status_args]*
1357 * num_paths num_selector_args
1358 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1360 * Table output has the following format (identical to the constructor string):
1361 * num_feature_args [features_args]*
1362 * num_handler_args hw_handler [hw_handler_args]*
1363 * num_groups init_group_number
1364 * [priority selector-name num_ps_args [ps_args]*
1365 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1367 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1368 unsigned status_flags
, char *result
, unsigned maxlen
)
1371 unsigned long flags
;
1372 struct multipath
*m
= (struct multipath
*) ti
->private;
1373 struct priority_group
*pg
;
1378 spin_lock_irqsave(&m
->lock
, flags
);
1381 if (type
== STATUSTYPE_INFO
)
1382 DMEMIT("2 %u %u ", m
->queue_io
, m
->pg_init_count
);
1384 DMEMIT("%u ", m
->queue_if_no_path
+
1385 (m
->pg_init_retries
> 0) * 2 +
1386 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1387 m
->retain_attached_hw_handler
);
1388 if (m
->queue_if_no_path
)
1389 DMEMIT("queue_if_no_path ");
1390 if (m
->pg_init_retries
)
1391 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1392 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1393 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1394 if (m
->retain_attached_hw_handler
)
1395 DMEMIT("retain_attached_hw_handler ");
1398 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1401 DMEMIT("1 %s ", m
->hw_handler_name
);
1403 DMEMIT("%u ", m
->nr_priority_groups
);
1406 pg_num
= m
->next_pg
->pg_num
;
1407 else if (m
->current_pg
)
1408 pg_num
= m
->current_pg
->pg_num
;
1410 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1412 DMEMIT("%u ", pg_num
);
1415 case STATUSTYPE_INFO
:
1416 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1418 state
= 'D'; /* Disabled */
1419 else if (pg
== m
->current_pg
)
1420 state
= 'A'; /* Currently Active */
1422 state
= 'E'; /* Enabled */
1424 DMEMIT("%c ", state
);
1426 if (pg
->ps
.type
->status
)
1427 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1433 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1434 pg
->ps
.type
->info_args
);
1436 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1437 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1438 p
->is_active
? "A" : "F",
1440 if (pg
->ps
.type
->status
)
1441 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1442 &p
->path
, type
, result
+ sz
,
1448 case STATUSTYPE_TABLE
:
1449 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1450 DMEMIT("%s ", pg
->ps
.type
->name
);
1452 if (pg
->ps
.type
->status
)
1453 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1459 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1460 pg
->ps
.type
->table_args
);
1462 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1463 DMEMIT("%s ", p
->path
.dev
->name
);
1464 if (pg
->ps
.type
->status
)
1465 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1466 &p
->path
, type
, result
+ sz
,
1473 spin_unlock_irqrestore(&m
->lock
, flags
);
1476 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1480 struct multipath
*m
= (struct multipath
*) ti
->private;
1483 mutex_lock(&m
->work_mutex
);
1485 if (dm_suspended(ti
)) {
1491 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1492 r
= queue_if_no_path(m
, 1, 0);
1494 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1495 r
= queue_if_no_path(m
, 0, 0);
1501 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1505 if (!strcasecmp(argv
[0], "disable_group")) {
1506 r
= bypass_pg_num(m
, argv
[1], 1);
1508 } else if (!strcasecmp(argv
[0], "enable_group")) {
1509 r
= bypass_pg_num(m
, argv
[1], 0);
1511 } else if (!strcasecmp(argv
[0], "switch_group")) {
1512 r
= switch_pg_num(m
, argv
[1]);
1514 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1515 action
= reinstate_path
;
1516 else if (!strcasecmp(argv
[0], "fail_path"))
1519 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1523 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1525 DMWARN("message: error getting device %s",
1530 r
= action_dev(m
, dev
, action
);
1532 dm_put_device(ti
, dev
);
1535 mutex_unlock(&m
->work_mutex
);
1539 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1540 struct block_device
**bdev
, fmode_t
*mode
)
1542 struct multipath
*m
= ti
->private;
1543 unsigned long flags
;
1546 spin_lock_irqsave(&m
->lock
, flags
);
1548 if (!m
->current_pgpath
)
1549 __choose_pgpath(m
, 0);
1551 if (m
->current_pgpath
) {
1553 *bdev
= m
->current_pgpath
->path
.dev
->bdev
;
1554 *mode
= m
->current_pgpath
->path
.dev
->mode
;
1557 /* pg_init has not started or completed */
1561 /* No path is available */
1562 if (m
->queue_if_no_path
)
1568 spin_unlock_irqrestore(&m
->lock
, flags
);
1570 if (r
== -ENOTCONN
) {
1571 spin_lock_irqsave(&m
->lock
, flags
);
1572 if (!m
->current_pg
) {
1573 /* Path status changed, redo selection */
1574 __choose_pgpath(m
, 0);
1576 if (m
->pg_init_required
)
1577 __pg_init_all_paths(m
);
1578 spin_unlock_irqrestore(&m
->lock
, flags
);
1579 dm_table_run_md_queue_async(m
->ti
->table
);
1583 * Only pass ioctls through if the device sizes match exactly.
1585 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1590 static int multipath_iterate_devices(struct dm_target
*ti
,
1591 iterate_devices_callout_fn fn
, void *data
)
1593 struct multipath
*m
= ti
->private;
1594 struct priority_group
*pg
;
1598 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1599 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1600 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1610 static int __pgpath_busy(struct pgpath
*pgpath
)
1612 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1614 return blk_lld_busy(q
);
1618 * We return "busy", only when we can map I/Os but underlying devices
1619 * are busy (so even if we map I/Os now, the I/Os will wait on
1620 * the underlying queue).
1621 * In other words, if we want to kill I/Os or queue them inside us
1622 * due to map unavailability, we don't return "busy". Otherwise,
1623 * dm core won't give us the I/Os and we can't do what we want.
1625 static int multipath_busy(struct dm_target
*ti
)
1627 int busy
= 0, has_active
= 0;
1628 struct multipath
*m
= ti
->private;
1629 struct priority_group
*pg
;
1630 struct pgpath
*pgpath
;
1631 unsigned long flags
;
1633 spin_lock_irqsave(&m
->lock
, flags
);
1635 /* pg_init in progress or no paths available */
1636 if (m
->pg_init_in_progress
||
1637 (!m
->nr_valid_paths
&& m
->queue_if_no_path
)) {
1641 /* Guess which priority_group will be used at next mapping time */
1642 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1644 else if (likely(m
->current_pg
))
1648 * We don't know which pg will be used at next mapping time.
1649 * We don't call __choose_pgpath() here to avoid to trigger
1650 * pg_init just by busy checking.
1651 * So we don't know whether underlying devices we will be using
1652 * at next mapping time are busy or not. Just try mapping.
1657 * If there is one non-busy active path at least, the path selector
1658 * will be able to select it. So we consider such a pg as not busy.
1661 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1662 if (pgpath
->is_active
) {
1665 if (!__pgpath_busy(pgpath
)) {
1673 * No active path in this pg, so this pg won't be used and
1674 * the current_pg will be changed at next mapping time.
1675 * We need to try mapping to determine it.
1680 spin_unlock_irqrestore(&m
->lock
, flags
);
1685 /*-----------------------------------------------------------------
1687 *---------------------------------------------------------------*/
1688 static struct target_type multipath_target
= {
1689 .name
= "multipath",
1690 .version
= {1, 11, 0},
1691 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1692 .module
= THIS_MODULE
,
1693 .ctr
= multipath_ctr
,
1694 .dtr
= multipath_dtr
,
1695 .map_rq
= multipath_map
,
1696 .clone_and_map_rq
= multipath_clone_and_map
,
1697 .release_clone_rq
= multipath_release_clone
,
1698 .rq_end_io
= multipath_end_io
,
1699 .presuspend
= multipath_presuspend
,
1700 .postsuspend
= multipath_postsuspend
,
1701 .resume
= multipath_resume
,
1702 .status
= multipath_status
,
1703 .message
= multipath_message
,
1704 .prepare_ioctl
= multipath_prepare_ioctl
,
1705 .iterate_devices
= multipath_iterate_devices
,
1706 .busy
= multipath_busy
,
1709 static int __init
dm_multipath_init(void)
1713 /* allocate a slab for the dm_ios */
1714 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1718 r
= dm_register_target(&multipath_target
);
1720 DMERR("register failed %d", r
);
1722 goto bad_register_target
;
1725 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1727 DMERR("failed to create workqueue kmpathd");
1729 goto bad_alloc_kmultipathd
;
1733 * A separate workqueue is used to handle the device handlers
1734 * to avoid overloading existing workqueue. Overloading the
1735 * old workqueue would also create a bottleneck in the
1736 * path of the storage hardware device activation.
1738 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1740 if (!kmpath_handlerd
) {
1741 DMERR("failed to create workqueue kmpath_handlerd");
1743 goto bad_alloc_kmpath_handlerd
;
1746 DMINFO("version %u.%u.%u loaded",
1747 multipath_target
.version
[0], multipath_target
.version
[1],
1748 multipath_target
.version
[2]);
1752 bad_alloc_kmpath_handlerd
:
1753 destroy_workqueue(kmultipathd
);
1754 bad_alloc_kmultipathd
:
1755 dm_unregister_target(&multipath_target
);
1756 bad_register_target
:
1757 kmem_cache_destroy(_mpio_cache
);
1762 static void __exit
dm_multipath_exit(void)
1764 destroy_workqueue(kmpath_handlerd
);
1765 destroy_workqueue(kmultipathd
);
1767 dm_unregister_target(&multipath_target
);
1768 kmem_cache_destroy(_mpio_cache
);
1771 module_init(dm_multipath_init
);
1772 module_exit(dm_multipath_exit
);
1774 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1775 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1776 MODULE_LICENSE("GPL");