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/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <linux/delay.h>
23 #include <scsi/scsi_dh.h>
24 #include <linux/atomic.h>
26 #define DM_MSG_PREFIX "multipath"
27 #define DM_PG_INIT_DELAY_MSECS 2000
28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32 struct list_head list
;
34 struct priority_group
*pg
; /* Owning PG */
35 unsigned is_active
; /* Path status */
36 unsigned fail_count
; /* Cumulative failure count */
39 struct delayed_work activate_path
;
42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
45 * Paths are grouped into Priority Groups and numbered from 1 upwards.
46 * Each has a path selector which controls which path gets used.
48 struct priority_group
{
49 struct list_head list
;
51 struct multipath
*m
; /* Owning multipath instance */
52 struct path_selector ps
;
54 unsigned pg_num
; /* Reference number */
55 unsigned bypassed
; /* Temporarily bypass this PG? */
57 unsigned nr_pgpaths
; /* Number of paths in PG */
58 struct list_head pgpaths
;
61 /* Multipath context */
63 struct list_head list
;
66 const char *hw_handler_name
;
67 char *hw_handler_params
;
71 unsigned nr_priority_groups
;
72 struct list_head priority_groups
;
74 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
76 unsigned pg_init_required
; /* pg_init needs calling? */
77 unsigned pg_init_in_progress
; /* Only one pg_init allowed at once */
78 unsigned pg_init_delay_retry
; /* Delay pg_init retry? */
80 unsigned nr_valid_paths
; /* Total number of usable paths */
81 struct pgpath
*current_pgpath
;
82 struct priority_group
*current_pg
;
83 struct priority_group
*next_pg
; /* Switch to this PG if set */
84 unsigned repeat_count
; /* I/Os left before calling PS again */
86 unsigned queue_io
:1; /* Must we queue all I/O? */
87 unsigned queue_if_no_path
:1; /* Queue I/O if last path fails? */
88 unsigned saved_queue_if_no_path
:1; /* Saved state during suspension */
89 unsigned retain_attached_hw_handler
:1; /* If there's already a hw_handler present, don't change it. */
90 unsigned pg_init_disabled
:1; /* pg_init is not currently allowed */
92 unsigned pg_init_retries
; /* Number of times to retry pg_init */
93 unsigned pg_init_count
; /* Number of times pg_init called */
94 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
96 struct work_struct trigger_event
;
99 * We must use a mempool of dm_mpath_io structs so that we
100 * can resubmit bios on error.
102 mempool_t
*mpio_pool
;
104 struct mutex work_mutex
;
108 * Context information attached to each bio we process.
111 struct pgpath
*pgpath
;
115 typedef int (*action_fn
) (struct pgpath
*pgpath
);
117 static struct kmem_cache
*_mpio_cache
;
119 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
120 static void trigger_event(struct work_struct
*work
);
121 static void activate_path(struct work_struct
*work
);
122 static int __pgpath_busy(struct pgpath
*pgpath
);
125 /*-----------------------------------------------
126 * Allocation routines
127 *-----------------------------------------------*/
129 static struct pgpath
*alloc_pgpath(void)
131 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
134 pgpath
->is_active
= 1;
135 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path
);
141 static void free_pgpath(struct pgpath
*pgpath
)
146 static struct priority_group
*alloc_priority_group(void)
148 struct priority_group
*pg
;
150 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
153 INIT_LIST_HEAD(&pg
->pgpaths
);
158 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
160 struct pgpath
*pgpath
, *tmp
;
161 struct multipath
*m
= ti
->private;
163 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
164 list_del(&pgpath
->list
);
165 if (m
->hw_handler_name
)
166 scsi_dh_detach(bdev_get_queue(pgpath
->path
.dev
->bdev
));
167 dm_put_device(ti
, pgpath
->path
.dev
);
172 static void free_priority_group(struct priority_group
*pg
,
173 struct dm_target
*ti
)
175 struct path_selector
*ps
= &pg
->ps
;
178 ps
->type
->destroy(ps
);
179 dm_put_path_selector(ps
->type
);
182 free_pgpaths(&pg
->pgpaths
, ti
);
186 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
189 unsigned min_ios
= dm_get_reserved_rq_based_ios();
191 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
193 INIT_LIST_HEAD(&m
->priority_groups
);
194 spin_lock_init(&m
->lock
);
196 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
197 INIT_WORK(&m
->trigger_event
, trigger_event
);
198 init_waitqueue_head(&m
->pg_init_wait
);
199 mutex_init(&m
->work_mutex
);
200 m
->mpio_pool
= mempool_create_slab_pool(min_ios
, _mpio_cache
);
212 static void free_multipath(struct multipath
*m
)
214 struct priority_group
*pg
, *tmp
;
216 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
218 free_priority_group(pg
, m
->ti
);
221 kfree(m
->hw_handler_name
);
222 kfree(m
->hw_handler_params
);
223 mempool_destroy(m
->mpio_pool
);
227 static int set_mapinfo(struct multipath
*m
, union map_info
*info
)
229 struct dm_mpath_io
*mpio
;
231 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
235 memset(mpio
, 0, sizeof(*mpio
));
241 static void clear_mapinfo(struct multipath
*m
, union map_info
*info
)
243 struct dm_mpath_io
*mpio
= info
->ptr
;
246 mempool_free(mpio
, m
->mpio_pool
);
249 /*-----------------------------------------------
251 *-----------------------------------------------*/
253 static int __pg_init_all_paths(struct multipath
*m
)
255 struct pgpath
*pgpath
;
256 unsigned long pg_init_delay
= 0;
258 if (m
->pg_init_in_progress
|| m
->pg_init_disabled
)
262 m
->pg_init_required
= 0;
264 /* Check here to reset pg_init_required */
268 if (m
->pg_init_delay_retry
)
269 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
270 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
271 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
272 /* Skip failed paths */
273 if (!pgpath
->is_active
)
275 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
277 m
->pg_init_in_progress
++;
279 return m
->pg_init_in_progress
;
282 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
284 m
->current_pg
= pgpath
->pg
;
286 /* Must we initialise the PG first, and queue I/O till it's ready? */
287 if (m
->hw_handler_name
) {
288 m
->pg_init_required
= 1;
291 m
->pg_init_required
= 0;
295 m
->pg_init_count
= 0;
298 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
301 struct dm_path
*path
;
303 path
= pg
->ps
.type
->select_path(&pg
->ps
, &m
->repeat_count
, nr_bytes
);
307 m
->current_pgpath
= path_to_pgpath(path
);
309 if (m
->current_pg
!= pg
)
310 __switch_pg(m
, m
->current_pgpath
);
315 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
317 struct priority_group
*pg
;
318 unsigned bypassed
= 1;
320 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
)));
376 #define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
379 * Map cloned requests
381 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
382 union map_info
*map_context
)
384 struct multipath
*m
= (struct multipath
*) ti
->private;
385 int r
= DM_MAPIO_REQUEUE
;
386 size_t nr_bytes
= blk_rq_bytes(clone
);
388 struct pgpath
*pgpath
;
389 struct block_device
*bdev
;
390 struct dm_mpath_io
*mpio
;
392 spin_lock_irqsave(&m
->lock
, flags
);
394 /* Do we need to select a new pgpath? */
395 if (!m
->current_pgpath
||
396 (!m
->queue_io
&& (m
->repeat_count
&& --m
->repeat_count
== 0)))
397 __choose_pgpath(m
, nr_bytes
);
399 pgpath
= m
->current_pgpath
;
402 if (!__must_push_back(m
))
403 r
= -EIO
; /* Failed */
407 __pg_init_all_paths(m
);
410 if (set_mapinfo(m
, map_context
) < 0)
411 /* ENOMEM, requeue */
414 bdev
= pgpath
->path
.dev
->bdev
;
415 clone
->q
= bdev_get_queue(bdev
);
416 clone
->rq_disk
= bdev
->bd_disk
;
417 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
418 mpio
= map_context
->ptr
;
419 mpio
->pgpath
= pgpath
;
420 mpio
->nr_bytes
= nr_bytes
;
421 if (pgpath
->pg
->ps
.type
->start_io
)
422 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
425 r
= DM_MAPIO_REMAPPED
;
428 spin_unlock_irqrestore(&m
->lock
, flags
);
434 * If we run out of usable paths, should we queue I/O or error it?
436 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
437 unsigned save_old_value
)
441 spin_lock_irqsave(&m
->lock
, flags
);
444 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
446 m
->saved_queue_if_no_path
= queue_if_no_path
;
447 m
->queue_if_no_path
= queue_if_no_path
;
448 if (!m
->queue_if_no_path
)
449 dm_table_run_md_queue_async(m
->ti
->table
);
451 spin_unlock_irqrestore(&m
->lock
, flags
);
457 * An event is triggered whenever a path is taken out of use.
458 * Includes path failure and PG bypass.
460 static void trigger_event(struct work_struct
*work
)
462 struct multipath
*m
=
463 container_of(work
, struct multipath
, trigger_event
);
465 dm_table_event(m
->ti
->table
);
468 /*-----------------------------------------------------------------
469 * Constructor/argument parsing:
470 * <#multipath feature args> [<arg>]*
471 * <#hw_handler args> [hw_handler [<arg>]*]
473 * <initial priority group>
474 * [<selector> <#selector args> [<arg>]*
475 * <#paths> <#per-path selector args>
476 * [<path> [<arg>]* ]+ ]+
477 *---------------------------------------------------------------*/
478 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
479 struct dm_target
*ti
)
482 struct path_selector_type
*pst
;
485 static struct dm_arg _args
[] = {
486 {0, 1024, "invalid number of path selector args"},
489 pst
= dm_get_path_selector(dm_shift_arg(as
));
491 ti
->error
= "unknown path selector type";
495 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
497 dm_put_path_selector(pst
);
501 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
503 dm_put_path_selector(pst
);
504 ti
->error
= "path selector constructor failed";
509 dm_consume_args(as
, ps_argc
);
514 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
515 struct dm_target
*ti
)
519 struct multipath
*m
= ti
->private;
520 struct request_queue
*q
= NULL
;
521 const char *attached_handler_name
;
523 /* we need at least a path arg */
525 ti
->error
= "no device given";
526 return ERR_PTR(-EINVAL
);
531 return ERR_PTR(-ENOMEM
);
533 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
536 ti
->error
= "error getting device";
540 if (m
->retain_attached_hw_handler
|| m
->hw_handler_name
)
541 q
= bdev_get_queue(p
->path
.dev
->bdev
);
543 if (m
->retain_attached_hw_handler
) {
544 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
545 if (attached_handler_name
) {
547 * Reset hw_handler_name to match the attached handler
548 * and clear any hw_handler_params associated with the
551 * NB. This modifies the table line to show the actual
552 * handler instead of the original table passed in.
554 kfree(m
->hw_handler_name
);
555 m
->hw_handler_name
= attached_handler_name
;
557 kfree(m
->hw_handler_params
);
558 m
->hw_handler_params
= NULL
;
562 if (m
->hw_handler_name
) {
564 * Increments scsi_dh reference, even when using an
565 * already-attached handler.
567 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
570 * Already attached to different hw_handler:
571 * try to reattach with correct one.
574 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
578 ti
->error
= "error attaching hardware handler";
579 dm_put_device(ti
, p
->path
.dev
);
583 if (m
->hw_handler_params
) {
584 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
586 ti
->error
= "unable to set hardware "
587 "handler parameters";
589 dm_put_device(ti
, p
->path
.dev
);
595 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
597 dm_put_device(ti
, p
->path
.dev
);
608 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
611 static struct dm_arg _args
[] = {
612 {1, 1024, "invalid number of paths"},
613 {0, 1024, "invalid number of selector args"}
617 unsigned i
, nr_selector_args
, nr_args
;
618 struct priority_group
*pg
;
619 struct dm_target
*ti
= m
->ti
;
623 ti
->error
= "not enough priority group arguments";
624 return ERR_PTR(-EINVAL
);
627 pg
= alloc_priority_group();
629 ti
->error
= "couldn't allocate priority group";
630 return ERR_PTR(-ENOMEM
);
634 r
= parse_path_selector(as
, pg
, ti
);
641 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
645 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
649 nr_args
= 1 + nr_selector_args
;
650 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
651 struct pgpath
*pgpath
;
652 struct dm_arg_set path_args
;
654 if (as
->argc
< nr_args
) {
655 ti
->error
= "not enough path parameters";
660 path_args
.argc
= nr_args
;
661 path_args
.argv
= as
->argv
;
663 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
664 if (IS_ERR(pgpath
)) {
670 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
671 dm_consume_args(as
, nr_args
);
677 free_priority_group(pg
, ti
);
681 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
685 struct dm_target
*ti
= m
->ti
;
687 static struct dm_arg _args
[] = {
688 {0, 1024, "invalid number of hardware handler args"},
691 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
697 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
698 if (!try_then_request_module(scsi_dh_handler_exist(m
->hw_handler_name
),
699 "scsi_dh_%s", m
->hw_handler_name
)) {
700 ti
->error
= "unknown hardware handler type";
709 for (i
= 0; i
<= hw_argc
- 2; i
++)
710 len
+= strlen(as
->argv
[i
]) + 1;
711 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
713 ti
->error
= "memory allocation failed";
717 j
= sprintf(p
, "%d", hw_argc
- 1);
718 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
719 j
= sprintf(p
, "%s", as
->argv
[i
]);
721 dm_consume_args(as
, hw_argc
- 1);
725 kfree(m
->hw_handler_name
);
726 m
->hw_handler_name
= NULL
;
730 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
734 struct dm_target
*ti
= m
->ti
;
735 const char *arg_name
;
737 static struct dm_arg _args
[] = {
738 {0, 6, "invalid number of feature args"},
739 {1, 50, "pg_init_retries must be between 1 and 50"},
740 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
743 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
751 arg_name
= dm_shift_arg(as
);
754 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
755 r
= queue_if_no_path(m
, 1, 0);
759 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
760 m
->retain_attached_hw_handler
= 1;
764 if (!strcasecmp(arg_name
, "pg_init_retries") &&
766 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
771 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
773 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
778 ti
->error
= "Unrecognised multipath feature request";
780 } while (argc
&& !r
);
785 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
788 /* target arguments */
789 static struct dm_arg _args
[] = {
790 {0, 1024, "invalid number of priority groups"},
791 {0, 1024, "invalid initial priority group number"},
796 struct dm_arg_set as
;
797 unsigned pg_count
= 0;
798 unsigned next_pg_num
;
803 m
= alloc_multipath(ti
);
805 ti
->error
= "can't allocate multipath";
809 r
= parse_features(&as
, m
);
813 r
= parse_hw_handler(&as
, m
);
817 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
821 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
825 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
826 (m
->nr_priority_groups
&& !next_pg_num
)) {
827 ti
->error
= "invalid initial priority group";
832 /* parse the priority groups */
834 struct priority_group
*pg
;
836 pg
= parse_priority_group(&as
, m
);
842 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
843 list_add_tail(&pg
->list
, &m
->priority_groups
);
845 pg
->pg_num
= pg_count
;
850 if (pg_count
!= m
->nr_priority_groups
) {
851 ti
->error
= "priority group count mismatch";
856 ti
->num_flush_bios
= 1;
857 ti
->num_discard_bios
= 1;
858 ti
->num_write_same_bios
= 1;
867 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
869 DECLARE_WAITQUEUE(wait
, current
);
872 add_wait_queue(&m
->pg_init_wait
, &wait
);
875 set_current_state(TASK_UNINTERRUPTIBLE
);
877 spin_lock_irqsave(&m
->lock
, flags
);
878 if (!m
->pg_init_in_progress
) {
879 spin_unlock_irqrestore(&m
->lock
, flags
);
882 spin_unlock_irqrestore(&m
->lock
, flags
);
886 set_current_state(TASK_RUNNING
);
888 remove_wait_queue(&m
->pg_init_wait
, &wait
);
891 static void flush_multipath_work(struct multipath
*m
)
895 spin_lock_irqsave(&m
->lock
, flags
);
896 m
->pg_init_disabled
= 1;
897 spin_unlock_irqrestore(&m
->lock
, flags
);
899 flush_workqueue(kmpath_handlerd
);
900 multipath_wait_for_pg_init_completion(m
);
901 flush_workqueue(kmultipathd
);
902 flush_work(&m
->trigger_event
);
904 spin_lock_irqsave(&m
->lock
, flags
);
905 m
->pg_init_disabled
= 0;
906 spin_unlock_irqrestore(&m
->lock
, flags
);
909 static void multipath_dtr(struct dm_target
*ti
)
911 struct multipath
*m
= ti
->private;
913 flush_multipath_work(m
);
918 * Take a path out of use.
920 static int fail_path(struct pgpath
*pgpath
)
923 struct multipath
*m
= pgpath
->pg
->m
;
925 spin_lock_irqsave(&m
->lock
, flags
);
927 if (!pgpath
->is_active
)
930 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
932 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
933 pgpath
->is_active
= 0;
934 pgpath
->fail_count
++;
938 if (pgpath
== m
->current_pgpath
)
939 m
->current_pgpath
= NULL
;
941 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
942 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
944 schedule_work(&m
->trigger_event
);
947 spin_unlock_irqrestore(&m
->lock
, flags
);
953 * Reinstate a previously-failed path
955 static int reinstate_path(struct pgpath
*pgpath
)
959 struct multipath
*m
= pgpath
->pg
->m
;
961 spin_lock_irqsave(&m
->lock
, flags
);
963 if (pgpath
->is_active
)
966 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
967 DMWARN("Reinstate path not supported by path selector %s",
968 pgpath
->pg
->ps
.type
->name
);
973 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
977 pgpath
->is_active
= 1;
979 if (!m
->nr_valid_paths
++) {
980 m
->current_pgpath
= NULL
;
981 dm_table_run_md_queue_async(m
->ti
->table
);
982 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
983 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
984 m
->pg_init_in_progress
++;
987 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
988 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
990 schedule_work(&m
->trigger_event
);
993 spin_unlock_irqrestore(&m
->lock
, flags
);
999 * Fail or reinstate all paths that match the provided struct dm_dev.
1001 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1005 struct pgpath
*pgpath
;
1006 struct priority_group
*pg
;
1008 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1009 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1010 if (pgpath
->path
.dev
== dev
)
1019 * Temporarily try to avoid having to use the specified PG
1021 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1024 unsigned long flags
;
1026 spin_lock_irqsave(&m
->lock
, flags
);
1028 pg
->bypassed
= bypassed
;
1029 m
->current_pgpath
= NULL
;
1030 m
->current_pg
= NULL
;
1032 spin_unlock_irqrestore(&m
->lock
, flags
);
1034 schedule_work(&m
->trigger_event
);
1038 * Switch to using the specified PG from the next I/O that gets mapped
1040 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1042 struct priority_group
*pg
;
1044 unsigned long flags
;
1047 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1048 (pgnum
> m
->nr_priority_groups
)) {
1049 DMWARN("invalid PG number supplied to switch_pg_num");
1053 spin_lock_irqsave(&m
->lock
, flags
);
1054 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1059 m
->current_pgpath
= NULL
;
1060 m
->current_pg
= NULL
;
1063 spin_unlock_irqrestore(&m
->lock
, flags
);
1065 schedule_work(&m
->trigger_event
);
1070 * Set/clear bypassed status of a PG.
1071 * PGs are numbered upwards from 1 in the order they were declared.
1073 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1075 struct priority_group
*pg
;
1079 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1080 (pgnum
> m
->nr_priority_groups
)) {
1081 DMWARN("invalid PG number supplied to bypass_pg");
1085 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1090 bypass_pg(m
, pg
, bypassed
);
1095 * Should we retry pg_init immediately?
1097 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1099 unsigned long flags
;
1100 int limit_reached
= 0;
1102 spin_lock_irqsave(&m
->lock
, flags
);
1104 if (m
->pg_init_count
<= m
->pg_init_retries
&& !m
->pg_init_disabled
)
1105 m
->pg_init_required
= 1;
1109 spin_unlock_irqrestore(&m
->lock
, flags
);
1111 return limit_reached
;
1114 static void pg_init_done(void *data
, int errors
)
1116 struct pgpath
*pgpath
= data
;
1117 struct priority_group
*pg
= pgpath
->pg
;
1118 struct multipath
*m
= pg
->m
;
1119 unsigned long flags
;
1120 unsigned delay_retry
= 0;
1122 /* device or driver problems */
1127 if (!m
->hw_handler_name
) {
1131 DMERR("Could not failover the device: Handler scsi_dh_%s "
1132 "Error %d.", m
->hw_handler_name
, errors
);
1134 * Fail path for now, so we do not ping pong
1138 case SCSI_DH_DEV_TEMP_BUSY
:
1140 * Probably doing something like FW upgrade on the
1141 * controller so try the other pg.
1143 bypass_pg(m
, pg
, 1);
1146 /* Wait before retrying. */
1148 case SCSI_DH_IMM_RETRY
:
1149 case SCSI_DH_RES_TEMP_UNAVAIL
:
1150 if (pg_init_limit_reached(m
, pgpath
))
1156 * We probably do not want to fail the path for a device
1157 * error, but this is what the old dm did. In future
1158 * patches we can do more advanced handling.
1163 spin_lock_irqsave(&m
->lock
, flags
);
1165 if (pgpath
== m
->current_pgpath
) {
1166 DMERR("Could not failover device. Error %d.", errors
);
1167 m
->current_pgpath
= NULL
;
1168 m
->current_pg
= NULL
;
1170 } else if (!m
->pg_init_required
)
1173 if (--m
->pg_init_in_progress
)
1174 /* Activations of other paths are still on going */
1177 if (m
->pg_init_required
) {
1178 m
->pg_init_delay_retry
= delay_retry
;
1179 if (__pg_init_all_paths(m
))
1185 * Wake up any thread waiting to suspend.
1187 wake_up(&m
->pg_init_wait
);
1190 spin_unlock_irqrestore(&m
->lock
, flags
);
1193 static void activate_path(struct work_struct
*work
)
1195 struct pgpath
*pgpath
=
1196 container_of(work
, struct pgpath
, activate_path
.work
);
1198 if (pgpath
->is_active
)
1199 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1200 pg_init_done
, pgpath
);
1202 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1205 static int noretry_error(int error
)
1216 /* Anything else could be a path failure, so should be retried */
1223 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1224 int error
, struct dm_mpath_io
*mpio
)
1227 * We don't queue any clone request inside the multipath target
1228 * during end I/O handling, since those clone requests don't have
1229 * bio clones. If we queue them inside the multipath target,
1230 * we need to make bio clones, that requires memory allocation.
1231 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1232 * don't have bio clones.)
1233 * Instead of queueing the clone request here, we queue the original
1234 * request into dm core, which will remake a clone request and
1235 * clone bios for it and resubmit it later.
1237 int r
= DM_ENDIO_REQUEUE
;
1238 unsigned long flags
;
1240 if (!error
&& !clone
->errors
)
1241 return 0; /* I/O complete */
1243 if (noretry_error(error
)) {
1244 if ((clone
->cmd_flags
& REQ_WRITE_SAME
) &&
1245 !clone
->q
->limits
.max_write_same_sectors
) {
1246 struct queue_limits
*limits
;
1248 /* device doesn't really support WRITE SAME, disable it */
1249 limits
= dm_get_queue_limits(dm_table_get_md(m
->ti
->table
));
1250 limits
->max_write_same_sectors
= 0;
1256 fail_path(mpio
->pgpath
);
1258 spin_lock_irqsave(&m
->lock
, flags
);
1259 if (!m
->nr_valid_paths
) {
1260 if (!m
->queue_if_no_path
) {
1261 if (!__must_push_back(m
))
1264 if (error
== -EBADE
)
1268 spin_unlock_irqrestore(&m
->lock
, flags
);
1273 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1274 int error
, union map_info
*map_context
)
1276 struct multipath
*m
= ti
->private;
1277 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1278 struct pgpath
*pgpath
;
1279 struct path_selector
*ps
;
1284 r
= do_end_io(m
, clone
, error
, mpio
);
1285 pgpath
= mpio
->pgpath
;
1287 ps
= &pgpath
->pg
->ps
;
1288 if (ps
->type
->end_io
)
1289 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1291 clear_mapinfo(m
, map_context
);
1297 * Suspend can't complete until all the I/O is processed so if
1298 * the last path fails we must error any remaining I/O.
1299 * Note that if the freeze_bdev fails while suspending, the
1300 * queue_if_no_path state is lost - userspace should reset it.
1302 static void multipath_presuspend(struct dm_target
*ti
)
1304 struct multipath
*m
= (struct multipath
*) ti
->private;
1306 queue_if_no_path(m
, 0, 1);
1309 static void multipath_postsuspend(struct dm_target
*ti
)
1311 struct multipath
*m
= ti
->private;
1313 mutex_lock(&m
->work_mutex
);
1314 flush_multipath_work(m
);
1315 mutex_unlock(&m
->work_mutex
);
1319 * Restore the queue_if_no_path setting.
1321 static void multipath_resume(struct dm_target
*ti
)
1323 struct multipath
*m
= (struct multipath
*) ti
->private;
1324 unsigned long flags
;
1326 spin_lock_irqsave(&m
->lock
, flags
);
1327 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1328 spin_unlock_irqrestore(&m
->lock
, flags
);
1332 * Info output has the following format:
1333 * num_multipath_feature_args [multipath_feature_args]*
1334 * num_handler_status_args [handler_status_args]*
1335 * num_groups init_group_number
1336 * [A|D|E num_ps_status_args [ps_status_args]*
1337 * num_paths num_selector_args
1338 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1340 * Table output has the following format (identical to the constructor string):
1341 * num_feature_args [features_args]*
1342 * num_handler_args hw_handler [hw_handler_args]*
1343 * num_groups init_group_number
1344 * [priority selector-name num_ps_args [ps_args]*
1345 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1347 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1348 unsigned status_flags
, char *result
, unsigned maxlen
)
1351 unsigned long flags
;
1352 struct multipath
*m
= (struct multipath
*) ti
->private;
1353 struct priority_group
*pg
;
1358 spin_lock_irqsave(&m
->lock
, flags
);
1361 if (type
== STATUSTYPE_INFO
)
1362 DMEMIT("2 %u %u ", m
->queue_io
, m
->pg_init_count
);
1364 DMEMIT("%u ", m
->queue_if_no_path
+
1365 (m
->pg_init_retries
> 0) * 2 +
1366 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1367 m
->retain_attached_hw_handler
);
1368 if (m
->queue_if_no_path
)
1369 DMEMIT("queue_if_no_path ");
1370 if (m
->pg_init_retries
)
1371 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1372 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1373 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1374 if (m
->retain_attached_hw_handler
)
1375 DMEMIT("retain_attached_hw_handler ");
1378 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1381 DMEMIT("1 %s ", m
->hw_handler_name
);
1383 DMEMIT("%u ", m
->nr_priority_groups
);
1386 pg_num
= m
->next_pg
->pg_num
;
1387 else if (m
->current_pg
)
1388 pg_num
= m
->current_pg
->pg_num
;
1390 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1392 DMEMIT("%u ", pg_num
);
1395 case STATUSTYPE_INFO
:
1396 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1398 state
= 'D'; /* Disabled */
1399 else if (pg
== m
->current_pg
)
1400 state
= 'A'; /* Currently Active */
1402 state
= 'E'; /* Enabled */
1404 DMEMIT("%c ", state
);
1406 if (pg
->ps
.type
->status
)
1407 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1413 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1414 pg
->ps
.type
->info_args
);
1416 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1417 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1418 p
->is_active
? "A" : "F",
1420 if (pg
->ps
.type
->status
)
1421 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1422 &p
->path
, type
, result
+ sz
,
1428 case STATUSTYPE_TABLE
:
1429 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1430 DMEMIT("%s ", pg
->ps
.type
->name
);
1432 if (pg
->ps
.type
->status
)
1433 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1439 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1440 pg
->ps
.type
->table_args
);
1442 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1443 DMEMIT("%s ", p
->path
.dev
->name
);
1444 if (pg
->ps
.type
->status
)
1445 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1446 &p
->path
, type
, result
+ sz
,
1453 spin_unlock_irqrestore(&m
->lock
, flags
);
1456 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1460 struct multipath
*m
= (struct multipath
*) ti
->private;
1463 mutex_lock(&m
->work_mutex
);
1465 if (dm_suspended(ti
)) {
1471 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1472 r
= queue_if_no_path(m
, 1, 0);
1474 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1475 r
= queue_if_no_path(m
, 0, 0);
1481 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1485 if (!strcasecmp(argv
[0], "disable_group")) {
1486 r
= bypass_pg_num(m
, argv
[1], 1);
1488 } else if (!strcasecmp(argv
[0], "enable_group")) {
1489 r
= bypass_pg_num(m
, argv
[1], 0);
1491 } else if (!strcasecmp(argv
[0], "switch_group")) {
1492 r
= switch_pg_num(m
, argv
[1]);
1494 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1495 action
= reinstate_path
;
1496 else if (!strcasecmp(argv
[0], "fail_path"))
1499 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1503 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1505 DMWARN("message: error getting device %s",
1510 r
= action_dev(m
, dev
, action
);
1512 dm_put_device(ti
, dev
);
1515 mutex_unlock(&m
->work_mutex
);
1519 static int multipath_ioctl(struct dm_target
*ti
, unsigned int cmd
,
1522 struct multipath
*m
= ti
->private;
1523 struct pgpath
*pgpath
;
1524 struct block_device
*bdev
;
1526 unsigned long flags
;
1533 spin_lock_irqsave(&m
->lock
, flags
);
1535 if (!m
->current_pgpath
)
1536 __choose_pgpath(m
, 0);
1538 pgpath
= m
->current_pgpath
;
1541 bdev
= pgpath
->path
.dev
->bdev
;
1542 mode
= pgpath
->path
.dev
->mode
;
1545 if ((pgpath
&& m
->queue_io
) || (!pgpath
&& m
->queue_if_no_path
))
1550 spin_unlock_irqrestore(&m
->lock
, flags
);
1553 * Only pass ioctls through if the device sizes match exactly.
1555 if (!bdev
|| ti
->len
!= i_size_read(bdev
->bd_inode
) >> SECTOR_SHIFT
) {
1556 int err
= scsi_verify_blk_ioctl(NULL
, cmd
);
1561 if (r
== -ENOTCONN
&& !fatal_signal_pending(current
)) {
1562 spin_lock_irqsave(&m
->lock
, flags
);
1563 if (!m
->current_pg
) {
1564 /* Path status changed, redo selection */
1565 __choose_pgpath(m
, 0);
1567 if (m
->pg_init_required
)
1568 __pg_init_all_paths(m
);
1569 spin_unlock_irqrestore(&m
->lock
, flags
);
1570 dm_table_run_md_queue_async(m
->ti
->table
);
1573 return r
? : __blkdev_driver_ioctl(bdev
, mode
, cmd
, arg
);
1576 static int multipath_iterate_devices(struct dm_target
*ti
,
1577 iterate_devices_callout_fn fn
, void *data
)
1579 struct multipath
*m
= ti
->private;
1580 struct priority_group
*pg
;
1584 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1585 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1586 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1596 static int __pgpath_busy(struct pgpath
*pgpath
)
1598 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1600 return dm_underlying_device_busy(q
);
1604 * We return "busy", only when we can map I/Os but underlying devices
1605 * are busy (so even if we map I/Os now, the I/Os will wait on
1606 * the underlying queue).
1607 * In other words, if we want to kill I/Os or queue them inside us
1608 * due to map unavailability, we don't return "busy". Otherwise,
1609 * dm core won't give us the I/Os and we can't do what we want.
1611 static int multipath_busy(struct dm_target
*ti
)
1613 int busy
= 0, has_active
= 0;
1614 struct multipath
*m
= ti
->private;
1615 struct priority_group
*pg
;
1616 struct pgpath
*pgpath
;
1617 unsigned long flags
;
1619 spin_lock_irqsave(&m
->lock
, flags
);
1621 /* pg_init in progress, requeue until done */
1626 /* Guess which priority_group will be used at next mapping time */
1627 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1629 else if (likely(m
->current_pg
))
1633 * We don't know which pg will be used at next mapping time.
1634 * We don't call __choose_pgpath() here to avoid to trigger
1635 * pg_init just by busy checking.
1636 * So we don't know whether underlying devices we will be using
1637 * at next mapping time are busy or not. Just try mapping.
1642 * If there is one non-busy active path at least, the path selector
1643 * will be able to select it. So we consider such a pg as not busy.
1646 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1647 if (pgpath
->is_active
) {
1650 if (!__pgpath_busy(pgpath
)) {
1658 * No active path in this pg, so this pg won't be used and
1659 * the current_pg will be changed at next mapping time.
1660 * We need to try mapping to determine it.
1665 spin_unlock_irqrestore(&m
->lock
, flags
);
1670 /*-----------------------------------------------------------------
1672 *---------------------------------------------------------------*/
1673 static struct target_type multipath_target
= {
1674 .name
= "multipath",
1675 .version
= {1, 7, 0},
1676 .module
= THIS_MODULE
,
1677 .ctr
= multipath_ctr
,
1678 .dtr
= multipath_dtr
,
1679 .map_rq
= multipath_map
,
1680 .rq_end_io
= multipath_end_io
,
1681 .presuspend
= multipath_presuspend
,
1682 .postsuspend
= multipath_postsuspend
,
1683 .resume
= multipath_resume
,
1684 .status
= multipath_status
,
1685 .message
= multipath_message
,
1686 .ioctl
= multipath_ioctl
,
1687 .iterate_devices
= multipath_iterate_devices
,
1688 .busy
= multipath_busy
,
1691 static int __init
dm_multipath_init(void)
1695 /* allocate a slab for the dm_ios */
1696 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1700 r
= dm_register_target(&multipath_target
);
1702 DMERR("register failed %d", r
);
1703 kmem_cache_destroy(_mpio_cache
);
1707 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1709 DMERR("failed to create workqueue kmpathd");
1710 dm_unregister_target(&multipath_target
);
1711 kmem_cache_destroy(_mpio_cache
);
1716 * A separate workqueue is used to handle the device handlers
1717 * to avoid overloading existing workqueue. Overloading the
1718 * old workqueue would also create a bottleneck in the
1719 * path of the storage hardware device activation.
1721 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1723 if (!kmpath_handlerd
) {
1724 DMERR("failed to create workqueue kmpath_handlerd");
1725 destroy_workqueue(kmultipathd
);
1726 dm_unregister_target(&multipath_target
);
1727 kmem_cache_destroy(_mpio_cache
);
1731 DMINFO("version %u.%u.%u loaded",
1732 multipath_target
.version
[0], multipath_target
.version
[1],
1733 multipath_target
.version
[2]);
1738 static void __exit
dm_multipath_exit(void)
1740 destroy_workqueue(kmpath_handlerd
);
1741 destroy_workqueue(kmultipathd
);
1743 dm_unregister_target(&multipath_target
);
1744 kmem_cache_destroy(_mpio_cache
);
1747 module_init(dm_multipath_init
);
1748 module_exit(dm_multipath_exit
);
1750 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1751 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1752 MODULE_LICENSE("GPL");