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
;
162 struct multipath
*m
= ti
->private;
164 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
165 list_del(&pgpath
->list
);
166 if (m
->hw_handler_name
)
167 scsi_dh_detach(bdev_get_queue(pgpath
->path
.dev
->bdev
));
168 dm_put_device(ti
, pgpath
->path
.dev
);
173 static void free_priority_group(struct priority_group
*pg
,
174 struct dm_target
*ti
)
176 struct path_selector
*ps
= &pg
->ps
;
179 ps
->type
->destroy(ps
);
180 dm_put_path_selector(ps
->type
);
183 free_pgpaths(&pg
->pgpaths
, ti
);
187 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
190 unsigned min_ios
= dm_get_reserved_rq_based_ios();
192 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
194 INIT_LIST_HEAD(&m
->priority_groups
);
195 spin_lock_init(&m
->lock
);
197 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
198 INIT_WORK(&m
->trigger_event
, trigger_event
);
199 init_waitqueue_head(&m
->pg_init_wait
);
200 mutex_init(&m
->work_mutex
);
201 m
->mpio_pool
= mempool_create_slab_pool(min_ios
, _mpio_cache
);
213 static void free_multipath(struct multipath
*m
)
215 struct priority_group
*pg
, *tmp
;
217 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
219 free_priority_group(pg
, m
->ti
);
222 kfree(m
->hw_handler_name
);
223 kfree(m
->hw_handler_params
);
224 mempool_destroy(m
->mpio_pool
);
228 static int set_mapinfo(struct multipath
*m
, union map_info
*info
)
230 struct dm_mpath_io
*mpio
;
232 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
236 memset(mpio
, 0, sizeof(*mpio
));
242 static void clear_mapinfo(struct multipath
*m
, union map_info
*info
)
244 struct dm_mpath_io
*mpio
= info
->ptr
;
247 mempool_free(mpio
, m
->mpio_pool
);
250 /*-----------------------------------------------
252 *-----------------------------------------------*/
254 static int __pg_init_all_paths(struct multipath
*m
)
256 struct pgpath
*pgpath
;
257 unsigned long pg_init_delay
= 0;
259 if (m
->pg_init_in_progress
|| m
->pg_init_disabled
)
263 m
->pg_init_required
= 0;
265 /* Check here to reset pg_init_required */
269 if (m
->pg_init_delay_retry
)
270 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
271 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
272 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
273 /* Skip failed paths */
274 if (!pgpath
->is_active
)
276 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
278 m
->pg_init_in_progress
++;
280 return m
->pg_init_in_progress
;
283 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
285 m
->current_pg
= pgpath
->pg
;
287 /* Must we initialise the PG first, and queue I/O till it's ready? */
288 if (m
->hw_handler_name
) {
289 m
->pg_init_required
= 1;
292 m
->pg_init_required
= 0;
296 m
->pg_init_count
= 0;
299 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
302 struct dm_path
*path
;
304 path
= pg
->ps
.type
->select_path(&pg
->ps
, &m
->repeat_count
, nr_bytes
);
308 m
->current_pgpath
= path_to_pgpath(path
);
310 if (m
->current_pg
!= pg
)
311 __switch_pg(m
, m
->current_pgpath
);
316 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
318 struct priority_group
*pg
;
319 unsigned bypassed
= 1;
321 if (!m
->nr_valid_paths
) {
326 /* Were we instructed to switch PG? */
330 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
334 /* Don't change PG until it has no remaining paths */
335 if (m
->current_pg
&& !__choose_path_in_pg(m
, m
->current_pg
, nr_bytes
))
339 * Loop through priority groups until we find a valid path.
340 * First time we skip PGs marked 'bypassed'.
341 * Second time we only try the ones we skipped, but set
342 * pg_init_delay_retry so we do not hammer controllers.
345 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
346 if (pg
->bypassed
== bypassed
)
348 if (!__choose_path_in_pg(m
, pg
, nr_bytes
)) {
350 m
->pg_init_delay_retry
= 1;
354 } while (bypassed
--);
357 m
->current_pgpath
= NULL
;
358 m
->current_pg
= NULL
;
362 * Check whether bios must be queued in the device-mapper core rather
363 * than here in the target.
365 * m->lock must be held on entry.
367 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
368 * same value then we are not between multipath_presuspend()
369 * and multipath_resume() calls and we have no need to check
370 * for the DMF_NOFLUSH_SUSPENDING flag.
372 static int __must_push_back(struct multipath
*m
)
374 return (m
->queue_if_no_path
||
375 (m
->queue_if_no_path
!= m
->saved_queue_if_no_path
&&
376 dm_noflush_suspending(m
->ti
)));
380 * Map cloned requests
382 static int __multipath_map(struct dm_target
*ti
, struct request
*clone
,
383 union map_info
*map_context
,
384 struct request
*rq
, struct request
**__clone
)
386 struct multipath
*m
= (struct multipath
*) ti
->private;
387 int r
= DM_MAPIO_REQUEUE
;
388 size_t nr_bytes
= clone
? blk_rq_bytes(clone
) : blk_rq_bytes(rq
);
389 struct pgpath
*pgpath
;
390 struct block_device
*bdev
;
391 struct dm_mpath_io
*mpio
;
393 spin_lock_irq(&m
->lock
);
395 /* Do we need to select a new pgpath? */
396 if (!m
->current_pgpath
||
397 (!m
->queue_io
&& (m
->repeat_count
&& --m
->repeat_count
== 0)))
398 __choose_pgpath(m
, nr_bytes
);
400 pgpath
= m
->current_pgpath
;
403 if (!__must_push_back(m
))
404 r
= -EIO
; /* Failed */
406 } else if (m
->queue_io
|| m
->pg_init_required
) {
407 __pg_init_all_paths(m
);
411 if (set_mapinfo(m
, map_context
) < 0)
412 /* ENOMEM, requeue */
415 mpio
= map_context
->ptr
;
416 mpio
->pgpath
= pgpath
;
417 mpio
->nr_bytes
= nr_bytes
;
419 bdev
= pgpath
->path
.dev
->bdev
;
421 spin_unlock_irq(&m
->lock
);
424 /* Old request-based interface: allocated clone is passed in */
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 */
435 (*__clone
)->bio
= (*__clone
)->biotail
= NULL
;
436 (*__clone
)->rq_disk
= bdev
->bd_disk
;
437 (*__clone
)->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
440 if (pgpath
->pg
->ps
.type
->start_io
)
441 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
444 return DM_MAPIO_REMAPPED
;
447 spin_unlock_irq(&m
->lock
);
452 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
453 union map_info
*map_context
)
455 return __multipath_map(ti
, clone
, map_context
, NULL
, NULL
);
458 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
459 union map_info
*map_context
,
460 struct request
**clone
)
462 return __multipath_map(ti
, NULL
, map_context
, rq
, clone
);
465 static void multipath_release_clone(struct request
*clone
)
467 blk_put_request(clone
);
471 * If we run out of usable paths, should we queue I/O or error it?
473 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
474 unsigned save_old_value
)
478 spin_lock_irqsave(&m
->lock
, flags
);
481 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
483 m
->saved_queue_if_no_path
= queue_if_no_path
;
484 m
->queue_if_no_path
= queue_if_no_path
;
485 spin_unlock_irqrestore(&m
->lock
, flags
);
487 if (!queue_if_no_path
)
488 dm_table_run_md_queue_async(m
->ti
->table
);
494 * An event is triggered whenever a path is taken out of use.
495 * Includes path failure and PG bypass.
497 static void trigger_event(struct work_struct
*work
)
499 struct multipath
*m
=
500 container_of(work
, struct multipath
, trigger_event
);
502 dm_table_event(m
->ti
->table
);
505 /*-----------------------------------------------------------------
506 * Constructor/argument parsing:
507 * <#multipath feature args> [<arg>]*
508 * <#hw_handler args> [hw_handler [<arg>]*]
510 * <initial priority group>
511 * [<selector> <#selector args> [<arg>]*
512 * <#paths> <#per-path selector args>
513 * [<path> [<arg>]* ]+ ]+
514 *---------------------------------------------------------------*/
515 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
516 struct dm_target
*ti
)
519 struct path_selector_type
*pst
;
522 static struct dm_arg _args
[] = {
523 {0, 1024, "invalid number of path selector args"},
526 pst
= dm_get_path_selector(dm_shift_arg(as
));
528 ti
->error
= "unknown path selector type";
532 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
534 dm_put_path_selector(pst
);
538 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
540 dm_put_path_selector(pst
);
541 ti
->error
= "path selector constructor failed";
546 dm_consume_args(as
, ps_argc
);
551 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
552 struct dm_target
*ti
)
556 struct multipath
*m
= ti
->private;
557 struct request_queue
*q
= NULL
;
558 const char *attached_handler_name
;
560 /* we need at least a path arg */
562 ti
->error
= "no device given";
563 return ERR_PTR(-EINVAL
);
568 return ERR_PTR(-ENOMEM
);
570 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
573 ti
->error
= "error getting device";
577 if (m
->retain_attached_hw_handler
|| m
->hw_handler_name
)
578 q
= bdev_get_queue(p
->path
.dev
->bdev
);
580 if (m
->retain_attached_hw_handler
) {
581 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
582 if (attached_handler_name
) {
584 * Reset hw_handler_name to match the attached handler
585 * and clear any hw_handler_params associated with the
588 * NB. This modifies the table line to show the actual
589 * handler instead of the original table passed in.
591 kfree(m
->hw_handler_name
);
592 m
->hw_handler_name
= attached_handler_name
;
594 kfree(m
->hw_handler_params
);
595 m
->hw_handler_params
= NULL
;
599 if (m
->hw_handler_name
) {
601 * Increments scsi_dh reference, even when using an
602 * already-attached handler.
604 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
607 * Already attached to different hw_handler:
608 * try to reattach with correct one.
611 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
615 ti
->error
= "error attaching hardware handler";
616 dm_put_device(ti
, p
->path
.dev
);
620 if (m
->hw_handler_params
) {
621 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
623 ti
->error
= "unable to set hardware "
624 "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
);
735 if (!try_then_request_module(scsi_dh_handler_exist(m
->hw_handler_name
),
736 "scsi_dh_%s", m
->hw_handler_name
)) {
737 ti
->error
= "unknown hardware handler type";
746 for (i
= 0; i
<= hw_argc
- 2; i
++)
747 len
+= strlen(as
->argv
[i
]) + 1;
748 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
750 ti
->error
= "memory allocation failed";
754 j
= sprintf(p
, "%d", hw_argc
- 1);
755 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
756 j
= sprintf(p
, "%s", as
->argv
[i
]);
758 dm_consume_args(as
, hw_argc
- 1);
762 kfree(m
->hw_handler_name
);
763 m
->hw_handler_name
= NULL
;
767 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
771 struct dm_target
*ti
= m
->ti
;
772 const char *arg_name
;
774 static struct dm_arg _args
[] = {
775 {0, 6, "invalid number of feature args"},
776 {1, 50, "pg_init_retries must be between 1 and 50"},
777 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
780 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
788 arg_name
= dm_shift_arg(as
);
791 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
792 r
= queue_if_no_path(m
, 1, 0);
796 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
797 m
->retain_attached_hw_handler
= 1;
801 if (!strcasecmp(arg_name
, "pg_init_retries") &&
803 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
808 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
810 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
815 ti
->error
= "Unrecognised multipath feature request";
817 } while (argc
&& !r
);
822 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
825 /* target arguments */
826 static struct dm_arg _args
[] = {
827 {0, 1024, "invalid number of priority groups"},
828 {0, 1024, "invalid initial priority group number"},
833 struct dm_arg_set as
;
834 unsigned pg_count
= 0;
835 unsigned next_pg_num
;
840 m
= alloc_multipath(ti
);
842 ti
->error
= "can't allocate multipath";
846 r
= parse_features(&as
, m
);
850 r
= parse_hw_handler(&as
, m
);
854 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
858 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
862 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
863 (m
->nr_priority_groups
&& !next_pg_num
)) {
864 ti
->error
= "invalid initial priority group";
869 /* parse the priority groups */
871 struct priority_group
*pg
;
873 pg
= parse_priority_group(&as
, m
);
879 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
880 list_add_tail(&pg
->list
, &m
->priority_groups
);
882 pg
->pg_num
= pg_count
;
887 if (pg_count
!= m
->nr_priority_groups
) {
888 ti
->error
= "priority group count mismatch";
893 ti
->num_flush_bios
= 1;
894 ti
->num_discard_bios
= 1;
895 ti
->num_write_same_bios
= 1;
904 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
906 DECLARE_WAITQUEUE(wait
, current
);
909 add_wait_queue(&m
->pg_init_wait
, &wait
);
912 set_current_state(TASK_UNINTERRUPTIBLE
);
914 spin_lock_irqsave(&m
->lock
, flags
);
915 if (!m
->pg_init_in_progress
) {
916 spin_unlock_irqrestore(&m
->lock
, flags
);
919 spin_unlock_irqrestore(&m
->lock
, flags
);
923 set_current_state(TASK_RUNNING
);
925 remove_wait_queue(&m
->pg_init_wait
, &wait
);
928 static void flush_multipath_work(struct multipath
*m
)
932 spin_lock_irqsave(&m
->lock
, flags
);
933 m
->pg_init_disabled
= 1;
934 spin_unlock_irqrestore(&m
->lock
, flags
);
936 flush_workqueue(kmpath_handlerd
);
937 multipath_wait_for_pg_init_completion(m
);
938 flush_workqueue(kmultipathd
);
939 flush_work(&m
->trigger_event
);
941 spin_lock_irqsave(&m
->lock
, flags
);
942 m
->pg_init_disabled
= 0;
943 spin_unlock_irqrestore(&m
->lock
, flags
);
946 static void multipath_dtr(struct dm_target
*ti
)
948 struct multipath
*m
= ti
->private;
950 flush_multipath_work(m
);
955 * Take a path out of use.
957 static int fail_path(struct pgpath
*pgpath
)
960 struct multipath
*m
= pgpath
->pg
->m
;
962 spin_lock_irqsave(&m
->lock
, flags
);
964 if (!pgpath
->is_active
)
967 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
969 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
970 pgpath
->is_active
= 0;
971 pgpath
->fail_count
++;
975 if (pgpath
== m
->current_pgpath
)
976 m
->current_pgpath
= NULL
;
978 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
979 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
981 schedule_work(&m
->trigger_event
);
984 spin_unlock_irqrestore(&m
->lock
, flags
);
990 * Reinstate a previously-failed path
992 static int reinstate_path(struct pgpath
*pgpath
)
994 int r
= 0, run_queue
= 0;
996 struct multipath
*m
= pgpath
->pg
->m
;
998 spin_lock_irqsave(&m
->lock
, flags
);
1000 if (pgpath
->is_active
)
1003 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
1004 DMWARN("Reinstate path not supported by path selector %s",
1005 pgpath
->pg
->ps
.type
->name
);
1010 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1014 pgpath
->is_active
= 1;
1016 if (!m
->nr_valid_paths
++) {
1017 m
->current_pgpath
= NULL
;
1019 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1020 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1021 m
->pg_init_in_progress
++;
1024 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1025 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1027 schedule_work(&m
->trigger_event
);
1030 spin_unlock_irqrestore(&m
->lock
, flags
);
1032 dm_table_run_md_queue_async(m
->ti
->table
);
1038 * Fail or reinstate all paths that match the provided struct dm_dev.
1040 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1044 struct pgpath
*pgpath
;
1045 struct priority_group
*pg
;
1047 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1048 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1049 if (pgpath
->path
.dev
== dev
)
1058 * Temporarily try to avoid having to use the specified PG
1060 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1063 unsigned long flags
;
1065 spin_lock_irqsave(&m
->lock
, flags
);
1067 pg
->bypassed
= bypassed
;
1068 m
->current_pgpath
= NULL
;
1069 m
->current_pg
= NULL
;
1071 spin_unlock_irqrestore(&m
->lock
, flags
);
1073 schedule_work(&m
->trigger_event
);
1077 * Switch to using the specified PG from the next I/O that gets mapped
1079 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1081 struct priority_group
*pg
;
1083 unsigned long flags
;
1086 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1087 (pgnum
> m
->nr_priority_groups
)) {
1088 DMWARN("invalid PG number supplied to switch_pg_num");
1092 spin_lock_irqsave(&m
->lock
, flags
);
1093 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1098 m
->current_pgpath
= NULL
;
1099 m
->current_pg
= NULL
;
1102 spin_unlock_irqrestore(&m
->lock
, flags
);
1104 schedule_work(&m
->trigger_event
);
1109 * Set/clear bypassed status of a PG.
1110 * PGs are numbered upwards from 1 in the order they were declared.
1112 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1114 struct priority_group
*pg
;
1118 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1119 (pgnum
> m
->nr_priority_groups
)) {
1120 DMWARN("invalid PG number supplied to bypass_pg");
1124 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1129 bypass_pg(m
, pg
, bypassed
);
1134 * Should we retry pg_init immediately?
1136 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1138 unsigned long flags
;
1139 int limit_reached
= 0;
1141 spin_lock_irqsave(&m
->lock
, flags
);
1143 if (m
->pg_init_count
<= m
->pg_init_retries
&& !m
->pg_init_disabled
)
1144 m
->pg_init_required
= 1;
1148 spin_unlock_irqrestore(&m
->lock
, flags
);
1150 return limit_reached
;
1153 static void pg_init_done(void *data
, int errors
)
1155 struct pgpath
*pgpath
= data
;
1156 struct priority_group
*pg
= pgpath
->pg
;
1157 struct multipath
*m
= pg
->m
;
1158 unsigned long flags
;
1159 unsigned delay_retry
= 0;
1161 /* device or driver problems */
1166 if (!m
->hw_handler_name
) {
1170 DMERR("Could not failover the device: Handler scsi_dh_%s "
1171 "Error %d.", m
->hw_handler_name
, errors
);
1173 * Fail path for now, so we do not ping pong
1177 case SCSI_DH_DEV_TEMP_BUSY
:
1179 * Probably doing something like FW upgrade on the
1180 * controller so try the other pg.
1182 bypass_pg(m
, pg
, 1);
1185 /* Wait before retrying. */
1187 case SCSI_DH_IMM_RETRY
:
1188 case SCSI_DH_RES_TEMP_UNAVAIL
:
1189 if (pg_init_limit_reached(m
, pgpath
))
1195 * We probably do not want to fail the path for a device
1196 * error, but this is what the old dm did. In future
1197 * patches we can do more advanced handling.
1202 spin_lock_irqsave(&m
->lock
, flags
);
1204 if (pgpath
== m
->current_pgpath
) {
1205 DMERR("Could not failover device. Error %d.", errors
);
1206 m
->current_pgpath
= NULL
;
1207 m
->current_pg
= NULL
;
1209 } else if (!m
->pg_init_required
)
1212 if (--m
->pg_init_in_progress
)
1213 /* Activations of other paths are still on going */
1216 if (m
->pg_init_required
) {
1217 m
->pg_init_delay_retry
= delay_retry
;
1218 if (__pg_init_all_paths(m
))
1224 * Wake up any thread waiting to suspend.
1226 wake_up(&m
->pg_init_wait
);
1229 spin_unlock_irqrestore(&m
->lock
, flags
);
1232 static void activate_path(struct work_struct
*work
)
1234 struct pgpath
*pgpath
=
1235 container_of(work
, struct pgpath
, activate_path
.work
);
1237 if (pgpath
->is_active
)
1238 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1239 pg_init_done
, pgpath
);
1241 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1244 static int noretry_error(int error
)
1255 /* Anything else could be a path failure, so should be retried */
1262 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1263 int error
, struct dm_mpath_io
*mpio
)
1266 * We don't queue any clone request inside the multipath target
1267 * during end I/O handling, since those clone requests don't have
1268 * bio clones. If we queue them inside the multipath target,
1269 * we need to make bio clones, that requires memory allocation.
1270 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1271 * don't have bio clones.)
1272 * Instead of queueing the clone request here, we queue the original
1273 * request into dm core, which will remake a clone request and
1274 * clone bios for it and resubmit it later.
1276 int r
= DM_ENDIO_REQUEUE
;
1277 unsigned long flags
;
1279 if (!error
&& !clone
->errors
)
1280 return 0; /* I/O complete */
1282 if (noretry_error(error
))
1286 fail_path(mpio
->pgpath
);
1288 spin_lock_irqsave(&m
->lock
, flags
);
1289 if (!m
->nr_valid_paths
) {
1290 if (!m
->queue_if_no_path
) {
1291 if (!__must_push_back(m
))
1294 if (error
== -EBADE
)
1298 spin_unlock_irqrestore(&m
->lock
, flags
);
1303 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1304 int error
, union map_info
*map_context
)
1306 struct multipath
*m
= ti
->private;
1307 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1308 struct pgpath
*pgpath
;
1309 struct path_selector
*ps
;
1314 r
= do_end_io(m
, clone
, error
, mpio
);
1315 pgpath
= mpio
->pgpath
;
1317 ps
= &pgpath
->pg
->ps
;
1318 if (ps
->type
->end_io
)
1319 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1321 clear_mapinfo(m
, map_context
);
1327 * Suspend can't complete until all the I/O is processed so if
1328 * the last path fails we must error any remaining I/O.
1329 * Note that if the freeze_bdev fails while suspending, the
1330 * queue_if_no_path state is lost - userspace should reset it.
1332 static void multipath_presuspend(struct dm_target
*ti
)
1334 struct multipath
*m
= (struct multipath
*) ti
->private;
1336 queue_if_no_path(m
, 0, 1);
1339 static void multipath_postsuspend(struct dm_target
*ti
)
1341 struct multipath
*m
= ti
->private;
1343 mutex_lock(&m
->work_mutex
);
1344 flush_multipath_work(m
);
1345 mutex_unlock(&m
->work_mutex
);
1349 * Restore the queue_if_no_path setting.
1351 static void multipath_resume(struct dm_target
*ti
)
1353 struct multipath
*m
= (struct multipath
*) ti
->private;
1354 unsigned long flags
;
1356 spin_lock_irqsave(&m
->lock
, flags
);
1357 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1358 spin_unlock_irqrestore(&m
->lock
, flags
);
1362 * Info output has the following format:
1363 * num_multipath_feature_args [multipath_feature_args]*
1364 * num_handler_status_args [handler_status_args]*
1365 * num_groups init_group_number
1366 * [A|D|E num_ps_status_args [ps_status_args]*
1367 * num_paths num_selector_args
1368 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1370 * Table output has the following format (identical to the constructor string):
1371 * num_feature_args [features_args]*
1372 * num_handler_args hw_handler [hw_handler_args]*
1373 * num_groups init_group_number
1374 * [priority selector-name num_ps_args [ps_args]*
1375 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1377 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1378 unsigned status_flags
, char *result
, unsigned maxlen
)
1381 unsigned long flags
;
1382 struct multipath
*m
= (struct multipath
*) ti
->private;
1383 struct priority_group
*pg
;
1388 spin_lock_irqsave(&m
->lock
, flags
);
1391 if (type
== STATUSTYPE_INFO
)
1392 DMEMIT("2 %u %u ", m
->queue_io
, m
->pg_init_count
);
1394 DMEMIT("%u ", m
->queue_if_no_path
+
1395 (m
->pg_init_retries
> 0) * 2 +
1396 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1397 m
->retain_attached_hw_handler
);
1398 if (m
->queue_if_no_path
)
1399 DMEMIT("queue_if_no_path ");
1400 if (m
->pg_init_retries
)
1401 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1402 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1403 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1404 if (m
->retain_attached_hw_handler
)
1405 DMEMIT("retain_attached_hw_handler ");
1408 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1411 DMEMIT("1 %s ", m
->hw_handler_name
);
1413 DMEMIT("%u ", m
->nr_priority_groups
);
1416 pg_num
= m
->next_pg
->pg_num
;
1417 else if (m
->current_pg
)
1418 pg_num
= m
->current_pg
->pg_num
;
1420 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1422 DMEMIT("%u ", pg_num
);
1425 case STATUSTYPE_INFO
:
1426 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1428 state
= 'D'; /* Disabled */
1429 else if (pg
== m
->current_pg
)
1430 state
= 'A'; /* Currently Active */
1432 state
= 'E'; /* Enabled */
1434 DMEMIT("%c ", state
);
1436 if (pg
->ps
.type
->status
)
1437 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1443 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1444 pg
->ps
.type
->info_args
);
1446 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1447 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1448 p
->is_active
? "A" : "F",
1450 if (pg
->ps
.type
->status
)
1451 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1452 &p
->path
, type
, result
+ sz
,
1458 case STATUSTYPE_TABLE
:
1459 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1460 DMEMIT("%s ", pg
->ps
.type
->name
);
1462 if (pg
->ps
.type
->status
)
1463 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1469 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1470 pg
->ps
.type
->table_args
);
1472 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1473 DMEMIT("%s ", p
->path
.dev
->name
);
1474 if (pg
->ps
.type
->status
)
1475 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1476 &p
->path
, type
, result
+ sz
,
1483 spin_unlock_irqrestore(&m
->lock
, flags
);
1486 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1490 struct multipath
*m
= (struct multipath
*) ti
->private;
1493 mutex_lock(&m
->work_mutex
);
1495 if (dm_suspended(ti
)) {
1501 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1502 r
= queue_if_no_path(m
, 1, 0);
1504 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1505 r
= queue_if_no_path(m
, 0, 0);
1511 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1515 if (!strcasecmp(argv
[0], "disable_group")) {
1516 r
= bypass_pg_num(m
, argv
[1], 1);
1518 } else if (!strcasecmp(argv
[0], "enable_group")) {
1519 r
= bypass_pg_num(m
, argv
[1], 0);
1521 } else if (!strcasecmp(argv
[0], "switch_group")) {
1522 r
= switch_pg_num(m
, argv
[1]);
1524 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1525 action
= reinstate_path
;
1526 else if (!strcasecmp(argv
[0], "fail_path"))
1529 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1533 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1535 DMWARN("message: error getting device %s",
1540 r
= action_dev(m
, dev
, action
);
1542 dm_put_device(ti
, dev
);
1545 mutex_unlock(&m
->work_mutex
);
1549 static int multipath_ioctl(struct dm_target
*ti
, unsigned int cmd
,
1552 struct multipath
*m
= ti
->private;
1553 struct pgpath
*pgpath
;
1554 struct block_device
*bdev
;
1556 unsigned long flags
;
1563 spin_lock_irqsave(&m
->lock
, flags
);
1565 if (!m
->current_pgpath
)
1566 __choose_pgpath(m
, 0);
1568 pgpath
= m
->current_pgpath
;
1571 bdev
= pgpath
->path
.dev
->bdev
;
1572 mode
= pgpath
->path
.dev
->mode
;
1575 if ((pgpath
&& m
->queue_io
) || (!pgpath
&& m
->queue_if_no_path
))
1580 spin_unlock_irqrestore(&m
->lock
, flags
);
1583 * Only pass ioctls through if the device sizes match exactly.
1585 if (!bdev
|| ti
->len
!= i_size_read(bdev
->bd_inode
) >> SECTOR_SHIFT
) {
1586 int err
= scsi_verify_blk_ioctl(NULL
, cmd
);
1591 if (r
== -ENOTCONN
&& !fatal_signal_pending(current
)) {
1592 spin_lock_irqsave(&m
->lock
, flags
);
1593 if (!m
->current_pg
) {
1594 /* Path status changed, redo selection */
1595 __choose_pgpath(m
, 0);
1597 if (m
->pg_init_required
)
1598 __pg_init_all_paths(m
);
1599 spin_unlock_irqrestore(&m
->lock
, flags
);
1600 dm_table_run_md_queue_async(m
->ti
->table
);
1603 return r
? : __blkdev_driver_ioctl(bdev
, mode
, cmd
, arg
);
1606 static int multipath_iterate_devices(struct dm_target
*ti
,
1607 iterate_devices_callout_fn fn
, void *data
)
1609 struct multipath
*m
= ti
->private;
1610 struct priority_group
*pg
;
1614 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1615 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1616 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1626 static int __pgpath_busy(struct pgpath
*pgpath
)
1628 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1630 return blk_lld_busy(q
);
1634 * We return "busy", only when we can map I/Os but underlying devices
1635 * are busy (so even if we map I/Os now, the I/Os will wait on
1636 * the underlying queue).
1637 * In other words, if we want to kill I/Os or queue them inside us
1638 * due to map unavailability, we don't return "busy". Otherwise,
1639 * dm core won't give us the I/Os and we can't do what we want.
1641 static int multipath_busy(struct dm_target
*ti
)
1643 int busy
= 0, has_active
= 0;
1644 struct multipath
*m
= ti
->private;
1645 struct priority_group
*pg
;
1646 struct pgpath
*pgpath
;
1647 unsigned long flags
;
1649 spin_lock_irqsave(&m
->lock
, flags
);
1651 /* pg_init in progress or no paths available */
1652 if (m
->pg_init_in_progress
||
1653 (!m
->nr_valid_paths
&& m
->queue_if_no_path
)) {
1657 /* Guess which priority_group will be used at next mapping time */
1658 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1660 else if (likely(m
->current_pg
))
1664 * We don't know which pg will be used at next mapping time.
1665 * We don't call __choose_pgpath() here to avoid to trigger
1666 * pg_init just by busy checking.
1667 * So we don't know whether underlying devices we will be using
1668 * at next mapping time are busy or not. Just try mapping.
1673 * If there is one non-busy active path at least, the path selector
1674 * will be able to select it. So we consider such a pg as not busy.
1677 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1678 if (pgpath
->is_active
) {
1681 if (!__pgpath_busy(pgpath
)) {
1689 * No active path in this pg, so this pg won't be used and
1690 * the current_pg will be changed at next mapping time.
1691 * We need to try mapping to determine it.
1696 spin_unlock_irqrestore(&m
->lock
, flags
);
1701 /*-----------------------------------------------------------------
1703 *---------------------------------------------------------------*/
1704 static struct target_type multipath_target
= {
1705 .name
= "multipath",
1706 .version
= {1, 9, 0},
1707 .module
= THIS_MODULE
,
1708 .ctr
= multipath_ctr
,
1709 .dtr
= multipath_dtr
,
1710 .map_rq
= multipath_map
,
1711 .clone_and_map_rq
= multipath_clone_and_map
,
1712 .release_clone_rq
= multipath_release_clone
,
1713 .rq_end_io
= multipath_end_io
,
1714 .presuspend
= multipath_presuspend
,
1715 .postsuspend
= multipath_postsuspend
,
1716 .resume
= multipath_resume
,
1717 .status
= multipath_status
,
1718 .message
= multipath_message
,
1719 .ioctl
= multipath_ioctl
,
1720 .iterate_devices
= multipath_iterate_devices
,
1721 .busy
= multipath_busy
,
1724 static int __init
dm_multipath_init(void)
1728 /* allocate a slab for the dm_ios */
1729 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1733 r
= dm_register_target(&multipath_target
);
1735 DMERR("register failed %d", r
);
1737 goto bad_register_target
;
1740 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1742 DMERR("failed to create workqueue kmpathd");
1744 goto bad_alloc_kmultipathd
;
1748 * A separate workqueue is used to handle the device handlers
1749 * to avoid overloading existing workqueue. Overloading the
1750 * old workqueue would also create a bottleneck in the
1751 * path of the storage hardware device activation.
1753 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1755 if (!kmpath_handlerd
) {
1756 DMERR("failed to create workqueue kmpath_handlerd");
1758 goto bad_alloc_kmpath_handlerd
;
1761 DMINFO("version %u.%u.%u loaded",
1762 multipath_target
.version
[0], multipath_target
.version
[1],
1763 multipath_target
.version
[2]);
1767 bad_alloc_kmpath_handlerd
:
1768 destroy_workqueue(kmultipathd
);
1769 bad_alloc_kmultipathd
:
1770 dm_unregister_target(&multipath_target
);
1771 bad_register_target
:
1772 kmem_cache_destroy(_mpio_cache
);
1777 static void __exit
dm_multipath_exit(void)
1779 destroy_workqueue(kmpath_handlerd
);
1780 destroy_workqueue(kmultipathd
);
1782 dm_unregister_target(&multipath_target
);
1783 kmem_cache_destroy(_mpio_cache
);
1786 module_init(dm_multipath_init
);
1787 module_exit(dm_multipath_exit
);
1789 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1790 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1791 MODULE_LICENSE("GPL");