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
;
430 * blk-mq request-based interface; used by both:
431 * .request_fn stacked on blk-mq path(s) and
432 * blk-mq stacked on blk-mq path(s).
434 *__clone
= blk_get_request(bdev_get_queue(bdev
),
435 rq_data_dir(rq
), GFP_ATOMIC
);
436 if (IS_ERR(*__clone
)) {
437 /* ENOMEM, requeue */
438 clear_mapinfo(m
, map_context
);
441 (*__clone
)->bio
= (*__clone
)->biotail
= NULL
;
442 (*__clone
)->rq_disk
= bdev
->bd_disk
;
443 (*__clone
)->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
446 if (pgpath
->pg
->ps
.type
->start_io
)
447 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
450 return DM_MAPIO_REMAPPED
;
453 spin_unlock_irq(&m
->lock
);
458 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
459 union map_info
*map_context
)
461 return __multipath_map(ti
, clone
, map_context
, NULL
, NULL
);
464 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
465 union map_info
*map_context
,
466 struct request
**clone
)
468 return __multipath_map(ti
, NULL
, map_context
, rq
, clone
);
471 static void multipath_release_clone(struct request
*clone
)
473 blk_put_request(clone
);
477 * If we run out of usable paths, should we queue I/O or error it?
479 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
480 unsigned save_old_value
)
484 spin_lock_irqsave(&m
->lock
, flags
);
487 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
489 m
->saved_queue_if_no_path
= queue_if_no_path
;
490 m
->queue_if_no_path
= queue_if_no_path
;
491 spin_unlock_irqrestore(&m
->lock
, flags
);
493 if (!queue_if_no_path
)
494 dm_table_run_md_queue_async(m
->ti
->table
);
500 * An event is triggered whenever a path is taken out of use.
501 * Includes path failure and PG bypass.
503 static void trigger_event(struct work_struct
*work
)
505 struct multipath
*m
=
506 container_of(work
, struct multipath
, trigger_event
);
508 dm_table_event(m
->ti
->table
);
511 /*-----------------------------------------------------------------
512 * Constructor/argument parsing:
513 * <#multipath feature args> [<arg>]*
514 * <#hw_handler args> [hw_handler [<arg>]*]
516 * <initial priority group>
517 * [<selector> <#selector args> [<arg>]*
518 * <#paths> <#per-path selector args>
519 * [<path> [<arg>]* ]+ ]+
520 *---------------------------------------------------------------*/
521 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
522 struct dm_target
*ti
)
525 struct path_selector_type
*pst
;
528 static struct dm_arg _args
[] = {
529 {0, 1024, "invalid number of path selector args"},
532 pst
= dm_get_path_selector(dm_shift_arg(as
));
534 ti
->error
= "unknown path selector type";
538 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
540 dm_put_path_selector(pst
);
544 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
546 dm_put_path_selector(pst
);
547 ti
->error
= "path selector constructor failed";
552 dm_consume_args(as
, ps_argc
);
557 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
558 struct dm_target
*ti
)
562 struct multipath
*m
= ti
->private;
563 struct request_queue
*q
= NULL
;
564 const char *attached_handler_name
;
566 /* we need at least a path arg */
568 ti
->error
= "no device given";
569 return ERR_PTR(-EINVAL
);
574 return ERR_PTR(-ENOMEM
);
576 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
579 ti
->error
= "error getting device";
583 if (m
->retain_attached_hw_handler
|| m
->hw_handler_name
)
584 q
= bdev_get_queue(p
->path
.dev
->bdev
);
586 if (m
->retain_attached_hw_handler
) {
588 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
589 if (attached_handler_name
) {
591 * Reset hw_handler_name to match the attached handler
592 * and clear any hw_handler_params associated with the
595 * NB. This modifies the table line to show the actual
596 * handler instead of the original table passed in.
598 kfree(m
->hw_handler_name
);
599 m
->hw_handler_name
= attached_handler_name
;
601 kfree(m
->hw_handler_params
);
602 m
->hw_handler_params
= NULL
;
606 if (m
->hw_handler_name
) {
607 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
609 char b
[BDEVNAME_SIZE
];
611 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
612 bdevname(p
->path
.dev
->bdev
, b
));
616 ti
->error
= "error attaching hardware handler";
617 dm_put_device(ti
, p
->path
.dev
);
621 if (m
->hw_handler_params
) {
622 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
624 ti
->error
= "unable to set hardware "
625 "handler parameters";
626 dm_put_device(ti
, p
->path
.dev
);
632 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
634 dm_put_device(ti
, p
->path
.dev
);
645 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
648 static struct dm_arg _args
[] = {
649 {1, 1024, "invalid number of paths"},
650 {0, 1024, "invalid number of selector args"}
654 unsigned i
, nr_selector_args
, nr_args
;
655 struct priority_group
*pg
;
656 struct dm_target
*ti
= m
->ti
;
660 ti
->error
= "not enough priority group arguments";
661 return ERR_PTR(-EINVAL
);
664 pg
= alloc_priority_group();
666 ti
->error
= "couldn't allocate priority group";
667 return ERR_PTR(-ENOMEM
);
671 r
= parse_path_selector(as
, pg
, ti
);
678 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
682 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
686 nr_args
= 1 + nr_selector_args
;
687 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
688 struct pgpath
*pgpath
;
689 struct dm_arg_set path_args
;
691 if (as
->argc
< nr_args
) {
692 ti
->error
= "not enough path parameters";
697 path_args
.argc
= nr_args
;
698 path_args
.argv
= as
->argv
;
700 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
701 if (IS_ERR(pgpath
)) {
707 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
708 dm_consume_args(as
, nr_args
);
714 free_priority_group(pg
, ti
);
718 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
722 struct dm_target
*ti
= m
->ti
;
724 static struct dm_arg _args
[] = {
725 {0, 1024, "invalid number of hardware handler args"},
728 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
734 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
740 for (i
= 0; i
<= hw_argc
- 2; i
++)
741 len
+= strlen(as
->argv
[i
]) + 1;
742 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
744 ti
->error
= "memory allocation failed";
748 j
= sprintf(p
, "%d", hw_argc
- 1);
749 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
750 j
= sprintf(p
, "%s", as
->argv
[i
]);
752 dm_consume_args(as
, hw_argc
- 1);
756 kfree(m
->hw_handler_name
);
757 m
->hw_handler_name
= NULL
;
761 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
765 struct dm_target
*ti
= m
->ti
;
766 const char *arg_name
;
768 static struct dm_arg _args
[] = {
769 {0, 6, "invalid number of feature args"},
770 {1, 50, "pg_init_retries must be between 1 and 50"},
771 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
774 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
782 arg_name
= dm_shift_arg(as
);
785 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
786 r
= queue_if_no_path(m
, 1, 0);
790 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
791 m
->retain_attached_hw_handler
= 1;
795 if (!strcasecmp(arg_name
, "pg_init_retries") &&
797 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
802 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
804 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
809 ti
->error
= "Unrecognised multipath feature request";
811 } while (argc
&& !r
);
816 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
819 /* target arguments */
820 static struct dm_arg _args
[] = {
821 {0, 1024, "invalid number of priority groups"},
822 {0, 1024, "invalid initial priority group number"},
827 struct dm_arg_set as
;
828 unsigned pg_count
= 0;
829 unsigned next_pg_num
;
834 m
= alloc_multipath(ti
);
836 ti
->error
= "can't allocate multipath";
840 r
= parse_features(&as
, m
);
844 r
= parse_hw_handler(&as
, m
);
848 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
852 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
856 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
857 (m
->nr_priority_groups
&& !next_pg_num
)) {
858 ti
->error
= "invalid initial priority group";
863 /* parse the priority groups */
865 struct priority_group
*pg
;
867 pg
= parse_priority_group(&as
, m
);
873 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
874 list_add_tail(&pg
->list
, &m
->priority_groups
);
876 pg
->pg_num
= pg_count
;
881 if (pg_count
!= m
->nr_priority_groups
) {
882 ti
->error
= "priority group count mismatch";
887 ti
->num_flush_bios
= 1;
888 ti
->num_discard_bios
= 1;
889 ti
->num_write_same_bios
= 1;
898 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
900 DECLARE_WAITQUEUE(wait
, current
);
903 add_wait_queue(&m
->pg_init_wait
, &wait
);
906 set_current_state(TASK_UNINTERRUPTIBLE
);
908 spin_lock_irqsave(&m
->lock
, flags
);
909 if (!m
->pg_init_in_progress
) {
910 spin_unlock_irqrestore(&m
->lock
, flags
);
913 spin_unlock_irqrestore(&m
->lock
, flags
);
917 set_current_state(TASK_RUNNING
);
919 remove_wait_queue(&m
->pg_init_wait
, &wait
);
922 static void flush_multipath_work(struct multipath
*m
)
926 spin_lock_irqsave(&m
->lock
, flags
);
927 m
->pg_init_disabled
= 1;
928 spin_unlock_irqrestore(&m
->lock
, flags
);
930 flush_workqueue(kmpath_handlerd
);
931 multipath_wait_for_pg_init_completion(m
);
932 flush_workqueue(kmultipathd
);
933 flush_work(&m
->trigger_event
);
935 spin_lock_irqsave(&m
->lock
, flags
);
936 m
->pg_init_disabled
= 0;
937 spin_unlock_irqrestore(&m
->lock
, flags
);
940 static void multipath_dtr(struct dm_target
*ti
)
942 struct multipath
*m
= ti
->private;
944 flush_multipath_work(m
);
949 * Take a path out of use.
951 static int fail_path(struct pgpath
*pgpath
)
954 struct multipath
*m
= pgpath
->pg
->m
;
956 spin_lock_irqsave(&m
->lock
, flags
);
958 if (!pgpath
->is_active
)
961 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
963 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
964 pgpath
->is_active
= 0;
965 pgpath
->fail_count
++;
969 if (pgpath
== m
->current_pgpath
)
970 m
->current_pgpath
= NULL
;
972 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
973 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
975 schedule_work(&m
->trigger_event
);
978 spin_unlock_irqrestore(&m
->lock
, flags
);
984 * Reinstate a previously-failed path
986 static int reinstate_path(struct pgpath
*pgpath
)
988 int r
= 0, run_queue
= 0;
990 struct multipath
*m
= pgpath
->pg
->m
;
992 spin_lock_irqsave(&m
->lock
, flags
);
994 if (pgpath
->is_active
)
997 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
998 DMWARN("Reinstate path not supported by path selector %s",
999 pgpath
->pg
->ps
.type
->name
);
1004 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1008 pgpath
->is_active
= 1;
1010 if (!m
->nr_valid_paths
++) {
1011 m
->current_pgpath
= NULL
;
1013 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1014 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1015 m
->pg_init_in_progress
++;
1018 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1019 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1021 schedule_work(&m
->trigger_event
);
1024 spin_unlock_irqrestore(&m
->lock
, flags
);
1026 dm_table_run_md_queue_async(m
->ti
->table
);
1032 * Fail or reinstate all paths that match the provided struct dm_dev.
1034 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1038 struct pgpath
*pgpath
;
1039 struct priority_group
*pg
;
1041 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1042 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1043 if (pgpath
->path
.dev
== dev
)
1052 * Temporarily try to avoid having to use the specified PG
1054 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1057 unsigned long flags
;
1059 spin_lock_irqsave(&m
->lock
, flags
);
1061 pg
->bypassed
= bypassed
;
1062 m
->current_pgpath
= NULL
;
1063 m
->current_pg
= NULL
;
1065 spin_unlock_irqrestore(&m
->lock
, flags
);
1067 schedule_work(&m
->trigger_event
);
1071 * Switch to using the specified PG from the next I/O that gets mapped
1073 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1075 struct priority_group
*pg
;
1077 unsigned long flags
;
1080 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1081 (pgnum
> m
->nr_priority_groups
)) {
1082 DMWARN("invalid PG number supplied to switch_pg_num");
1086 spin_lock_irqsave(&m
->lock
, flags
);
1087 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1092 m
->current_pgpath
= NULL
;
1093 m
->current_pg
= NULL
;
1096 spin_unlock_irqrestore(&m
->lock
, flags
);
1098 schedule_work(&m
->trigger_event
);
1103 * Set/clear bypassed status of a PG.
1104 * PGs are numbered upwards from 1 in the order they were declared.
1106 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1108 struct priority_group
*pg
;
1112 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1113 (pgnum
> m
->nr_priority_groups
)) {
1114 DMWARN("invalid PG number supplied to bypass_pg");
1118 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1123 bypass_pg(m
, pg
, bypassed
);
1128 * Should we retry pg_init immediately?
1130 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1132 unsigned long flags
;
1133 int limit_reached
= 0;
1135 spin_lock_irqsave(&m
->lock
, flags
);
1137 if (m
->pg_init_count
<= m
->pg_init_retries
&& !m
->pg_init_disabled
)
1138 m
->pg_init_required
= 1;
1142 spin_unlock_irqrestore(&m
->lock
, flags
);
1144 return limit_reached
;
1147 static void pg_init_done(void *data
, int errors
)
1149 struct pgpath
*pgpath
= data
;
1150 struct priority_group
*pg
= pgpath
->pg
;
1151 struct multipath
*m
= pg
->m
;
1152 unsigned long flags
;
1153 unsigned delay_retry
= 0;
1155 /* device or driver problems */
1160 if (!m
->hw_handler_name
) {
1164 DMERR("Could not failover the device: Handler scsi_dh_%s "
1165 "Error %d.", m
->hw_handler_name
, errors
);
1167 * Fail path for now, so we do not ping pong
1171 case SCSI_DH_DEV_TEMP_BUSY
:
1173 * Probably doing something like FW upgrade on the
1174 * controller so try the other pg.
1176 bypass_pg(m
, pg
, 1);
1179 /* Wait before retrying. */
1181 case SCSI_DH_IMM_RETRY
:
1182 case SCSI_DH_RES_TEMP_UNAVAIL
:
1183 if (pg_init_limit_reached(m
, pgpath
))
1189 * We probably do not want to fail the path for a device
1190 * error, but this is what the old dm did. In future
1191 * patches we can do more advanced handling.
1196 spin_lock_irqsave(&m
->lock
, flags
);
1198 if (pgpath
== m
->current_pgpath
) {
1199 DMERR("Could not failover device. Error %d.", errors
);
1200 m
->current_pgpath
= NULL
;
1201 m
->current_pg
= NULL
;
1203 } else if (!m
->pg_init_required
)
1206 if (--m
->pg_init_in_progress
)
1207 /* Activations of other paths are still on going */
1210 if (m
->pg_init_required
) {
1211 m
->pg_init_delay_retry
= delay_retry
;
1212 if (__pg_init_all_paths(m
))
1218 * Wake up any thread waiting to suspend.
1220 wake_up(&m
->pg_init_wait
);
1223 spin_unlock_irqrestore(&m
->lock
, flags
);
1226 static void activate_path(struct work_struct
*work
)
1228 struct pgpath
*pgpath
=
1229 container_of(work
, struct pgpath
, activate_path
.work
);
1231 if (pgpath
->is_active
)
1232 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1233 pg_init_done
, pgpath
);
1235 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1238 static int noretry_error(int error
)
1249 /* Anything else could be a path failure, so should be retried */
1256 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1257 int error
, struct dm_mpath_io
*mpio
)
1260 * We don't queue any clone request inside the multipath target
1261 * during end I/O handling, since those clone requests don't have
1262 * bio clones. If we queue them inside the multipath target,
1263 * we need to make bio clones, that requires memory allocation.
1264 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1265 * don't have bio clones.)
1266 * Instead of queueing the clone request here, we queue the original
1267 * request into dm core, which will remake a clone request and
1268 * clone bios for it and resubmit it later.
1270 int r
= DM_ENDIO_REQUEUE
;
1271 unsigned long flags
;
1273 if (!error
&& !clone
->errors
)
1274 return 0; /* I/O complete */
1276 if (noretry_error(error
))
1280 fail_path(mpio
->pgpath
);
1282 spin_lock_irqsave(&m
->lock
, flags
);
1283 if (!m
->nr_valid_paths
) {
1284 if (!m
->queue_if_no_path
) {
1285 if (!__must_push_back(m
))
1288 if (error
== -EBADE
)
1292 spin_unlock_irqrestore(&m
->lock
, flags
);
1297 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1298 int error
, union map_info
*map_context
)
1300 struct multipath
*m
= ti
->private;
1301 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1302 struct pgpath
*pgpath
;
1303 struct path_selector
*ps
;
1308 r
= do_end_io(m
, clone
, error
, mpio
);
1309 pgpath
= mpio
->pgpath
;
1311 ps
= &pgpath
->pg
->ps
;
1312 if (ps
->type
->end_io
)
1313 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1315 clear_mapinfo(m
, map_context
);
1321 * Suspend can't complete until all the I/O is processed so if
1322 * the last path fails we must error any remaining I/O.
1323 * Note that if the freeze_bdev fails while suspending, the
1324 * queue_if_no_path state is lost - userspace should reset it.
1326 static void multipath_presuspend(struct dm_target
*ti
)
1328 struct multipath
*m
= (struct multipath
*) ti
->private;
1330 queue_if_no_path(m
, 0, 1);
1333 static void multipath_postsuspend(struct dm_target
*ti
)
1335 struct multipath
*m
= ti
->private;
1337 mutex_lock(&m
->work_mutex
);
1338 flush_multipath_work(m
);
1339 mutex_unlock(&m
->work_mutex
);
1343 * Restore the queue_if_no_path setting.
1345 static void multipath_resume(struct dm_target
*ti
)
1347 struct multipath
*m
= (struct multipath
*) ti
->private;
1348 unsigned long flags
;
1350 spin_lock_irqsave(&m
->lock
, flags
);
1351 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1352 spin_unlock_irqrestore(&m
->lock
, flags
);
1356 * Info output has the following format:
1357 * num_multipath_feature_args [multipath_feature_args]*
1358 * num_handler_status_args [handler_status_args]*
1359 * num_groups init_group_number
1360 * [A|D|E num_ps_status_args [ps_status_args]*
1361 * num_paths num_selector_args
1362 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1364 * Table output has the following format (identical to the constructor string):
1365 * num_feature_args [features_args]*
1366 * num_handler_args hw_handler [hw_handler_args]*
1367 * num_groups init_group_number
1368 * [priority selector-name num_ps_args [ps_args]*
1369 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1371 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1372 unsigned status_flags
, char *result
, unsigned maxlen
)
1375 unsigned long flags
;
1376 struct multipath
*m
= (struct multipath
*) ti
->private;
1377 struct priority_group
*pg
;
1382 spin_lock_irqsave(&m
->lock
, flags
);
1385 if (type
== STATUSTYPE_INFO
)
1386 DMEMIT("2 %u %u ", m
->queue_io
, m
->pg_init_count
);
1388 DMEMIT("%u ", m
->queue_if_no_path
+
1389 (m
->pg_init_retries
> 0) * 2 +
1390 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1391 m
->retain_attached_hw_handler
);
1392 if (m
->queue_if_no_path
)
1393 DMEMIT("queue_if_no_path ");
1394 if (m
->pg_init_retries
)
1395 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1396 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1397 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1398 if (m
->retain_attached_hw_handler
)
1399 DMEMIT("retain_attached_hw_handler ");
1402 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1405 DMEMIT("1 %s ", m
->hw_handler_name
);
1407 DMEMIT("%u ", m
->nr_priority_groups
);
1410 pg_num
= m
->next_pg
->pg_num
;
1411 else if (m
->current_pg
)
1412 pg_num
= m
->current_pg
->pg_num
;
1414 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1416 DMEMIT("%u ", pg_num
);
1419 case STATUSTYPE_INFO
:
1420 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1422 state
= 'D'; /* Disabled */
1423 else if (pg
== m
->current_pg
)
1424 state
= 'A'; /* Currently Active */
1426 state
= 'E'; /* Enabled */
1428 DMEMIT("%c ", state
);
1430 if (pg
->ps
.type
->status
)
1431 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1437 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1438 pg
->ps
.type
->info_args
);
1440 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1441 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1442 p
->is_active
? "A" : "F",
1444 if (pg
->ps
.type
->status
)
1445 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1446 &p
->path
, type
, result
+ sz
,
1452 case STATUSTYPE_TABLE
:
1453 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1454 DMEMIT("%s ", pg
->ps
.type
->name
);
1456 if (pg
->ps
.type
->status
)
1457 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1463 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1464 pg
->ps
.type
->table_args
);
1466 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1467 DMEMIT("%s ", p
->path
.dev
->name
);
1468 if (pg
->ps
.type
->status
)
1469 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1470 &p
->path
, type
, result
+ sz
,
1477 spin_unlock_irqrestore(&m
->lock
, flags
);
1480 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1484 struct multipath
*m
= (struct multipath
*) ti
->private;
1487 mutex_lock(&m
->work_mutex
);
1489 if (dm_suspended(ti
)) {
1495 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1496 r
= queue_if_no_path(m
, 1, 0);
1498 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1499 r
= queue_if_no_path(m
, 0, 0);
1505 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1509 if (!strcasecmp(argv
[0], "disable_group")) {
1510 r
= bypass_pg_num(m
, argv
[1], 1);
1512 } else if (!strcasecmp(argv
[0], "enable_group")) {
1513 r
= bypass_pg_num(m
, argv
[1], 0);
1515 } else if (!strcasecmp(argv
[0], "switch_group")) {
1516 r
= switch_pg_num(m
, argv
[1]);
1518 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1519 action
= reinstate_path
;
1520 else if (!strcasecmp(argv
[0], "fail_path"))
1523 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1527 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1529 DMWARN("message: error getting device %s",
1534 r
= action_dev(m
, dev
, action
);
1536 dm_put_device(ti
, dev
);
1539 mutex_unlock(&m
->work_mutex
);
1543 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1544 struct block_device
**bdev
, fmode_t
*mode
)
1546 struct multipath
*m
= ti
->private;
1547 unsigned long flags
;
1550 spin_lock_irqsave(&m
->lock
, flags
);
1552 if (!m
->current_pgpath
)
1553 __choose_pgpath(m
, 0);
1555 if (m
->current_pgpath
) {
1557 *bdev
= m
->current_pgpath
->path
.dev
->bdev
;
1558 *mode
= m
->current_pgpath
->path
.dev
->mode
;
1561 /* pg_init has not started or completed */
1565 /* No path is available */
1566 if (m
->queue_if_no_path
)
1572 spin_unlock_irqrestore(&m
->lock
, flags
);
1574 if (r
== -ENOTCONN
) {
1575 spin_lock_irqsave(&m
->lock
, flags
);
1576 if (!m
->current_pg
) {
1577 /* Path status changed, redo selection */
1578 __choose_pgpath(m
, 0);
1580 if (m
->pg_init_required
)
1581 __pg_init_all_paths(m
);
1582 spin_unlock_irqrestore(&m
->lock
, flags
);
1583 dm_table_run_md_queue_async(m
->ti
->table
);
1587 * Only pass ioctls through if the device sizes match exactly.
1589 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1594 static int multipath_iterate_devices(struct dm_target
*ti
,
1595 iterate_devices_callout_fn fn
, void *data
)
1597 struct multipath
*m
= ti
->private;
1598 struct priority_group
*pg
;
1602 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1603 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1604 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1614 static int __pgpath_busy(struct pgpath
*pgpath
)
1616 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1618 return blk_lld_busy(q
);
1622 * We return "busy", only when we can map I/Os but underlying devices
1623 * are busy (so even if we map I/Os now, the I/Os will wait on
1624 * the underlying queue).
1625 * In other words, if we want to kill I/Os or queue them inside us
1626 * due to map unavailability, we don't return "busy". Otherwise,
1627 * dm core won't give us the I/Os and we can't do what we want.
1629 static int multipath_busy(struct dm_target
*ti
)
1631 int busy
= 0, has_active
= 0;
1632 struct multipath
*m
= ti
->private;
1633 struct priority_group
*pg
;
1634 struct pgpath
*pgpath
;
1635 unsigned long flags
;
1637 spin_lock_irqsave(&m
->lock
, flags
);
1639 /* pg_init in progress or no paths available */
1640 if (m
->pg_init_in_progress
||
1641 (!m
->nr_valid_paths
&& m
->queue_if_no_path
)) {
1645 /* Guess which priority_group will be used at next mapping time */
1646 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1648 else if (likely(m
->current_pg
))
1652 * We don't know which pg will be used at next mapping time.
1653 * We don't call __choose_pgpath() here to avoid to trigger
1654 * pg_init just by busy checking.
1655 * So we don't know whether underlying devices we will be using
1656 * at next mapping time are busy or not. Just try mapping.
1661 * If there is one non-busy active path at least, the path selector
1662 * will be able to select it. So we consider such a pg as not busy.
1665 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1666 if (pgpath
->is_active
) {
1669 if (!__pgpath_busy(pgpath
)) {
1677 * No active path in this pg, so this pg won't be used and
1678 * the current_pg will be changed at next mapping time.
1679 * We need to try mapping to determine it.
1684 spin_unlock_irqrestore(&m
->lock
, flags
);
1689 /*-----------------------------------------------------------------
1691 *---------------------------------------------------------------*/
1692 static struct target_type multipath_target
= {
1693 .name
= "multipath",
1694 .version
= {1, 11, 0},
1695 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1696 .module
= THIS_MODULE
,
1697 .ctr
= multipath_ctr
,
1698 .dtr
= multipath_dtr
,
1699 .map_rq
= multipath_map
,
1700 .clone_and_map_rq
= multipath_clone_and_map
,
1701 .release_clone_rq
= multipath_release_clone
,
1702 .rq_end_io
= multipath_end_io
,
1703 .presuspend
= multipath_presuspend
,
1704 .postsuspend
= multipath_postsuspend
,
1705 .resume
= multipath_resume
,
1706 .status
= multipath_status
,
1707 .message
= multipath_message
,
1708 .prepare_ioctl
= multipath_prepare_ioctl
,
1709 .iterate_devices
= multipath_iterate_devices
,
1710 .busy
= multipath_busy
,
1713 static int __init
dm_multipath_init(void)
1717 /* allocate a slab for the dm_ios */
1718 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1722 r
= dm_register_target(&multipath_target
);
1724 DMERR("register failed %d", r
);
1726 goto bad_register_target
;
1729 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1731 DMERR("failed to create workqueue kmpathd");
1733 goto bad_alloc_kmultipathd
;
1737 * A separate workqueue is used to handle the device handlers
1738 * to avoid overloading existing workqueue. Overloading the
1739 * old workqueue would also create a bottleneck in the
1740 * path of the storage hardware device activation.
1742 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1744 if (!kmpath_handlerd
) {
1745 DMERR("failed to create workqueue kmpath_handlerd");
1747 goto bad_alloc_kmpath_handlerd
;
1750 DMINFO("version %u.%u.%u loaded",
1751 multipath_target
.version
[0], multipath_target
.version
[1],
1752 multipath_target
.version
[2]);
1756 bad_alloc_kmpath_handlerd
:
1757 destroy_workqueue(kmultipathd
);
1758 bad_alloc_kmultipathd
:
1759 dm_unregister_target(&multipath_target
);
1760 bad_register_target
:
1761 kmem_cache_destroy(_mpio_cache
);
1766 static void __exit
dm_multipath_exit(void)
1768 destroy_workqueue(kmpath_handlerd
);
1769 destroy_workqueue(kmultipathd
);
1771 dm_unregister_target(&multipath_target
);
1772 kmem_cache_destroy(_mpio_cache
);
1775 module_init(dm_multipath_init
);
1776 module_exit(dm_multipath_exit
);
1778 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1779 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1780 MODULE_LICENSE("GPL");