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>
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <asm/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
29 struct list_head list
;
31 struct priority_group
*pg
; /* Owning PG */
32 unsigned is_active
; /* Path status */
33 unsigned fail_count
; /* Cumulative failure count */
36 struct work_struct activate_path
;
39 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
42 * Paths are grouped into Priority Groups and numbered from 1 upwards.
43 * Each has a path selector which controls which path gets used.
45 struct priority_group
{
46 struct list_head list
;
48 struct multipath
*m
; /* Owning multipath instance */
49 struct path_selector ps
;
51 unsigned pg_num
; /* Reference number */
52 unsigned bypassed
; /* Temporarily bypass this PG? */
54 unsigned nr_pgpaths
; /* Number of paths in PG */
55 struct list_head pgpaths
;
58 /* Multipath context */
60 struct list_head list
;
65 const char *hw_handler_name
;
66 char *hw_handler_params
;
67 unsigned nr_priority_groups
;
68 struct list_head priority_groups
;
69 unsigned pg_init_required
; /* pg_init needs calling? */
70 unsigned pg_init_in_progress
; /* Only one pg_init allowed at once */
72 unsigned nr_valid_paths
; /* Total number of usable paths */
73 struct pgpath
*current_pgpath
;
74 struct priority_group
*current_pg
;
75 struct priority_group
*next_pg
; /* Switch to this PG if set */
76 unsigned repeat_count
; /* I/Os left before calling PS again */
78 unsigned queue_io
; /* Must we queue all I/O? */
79 unsigned queue_if_no_path
; /* Queue I/O if last path fails? */
80 unsigned saved_queue_if_no_path
;/* Saved state during suspension */
81 unsigned pg_init_retries
; /* Number of times to retry pg_init */
82 unsigned pg_init_count
; /* Number of times pg_init called */
84 struct work_struct process_queued_ios
;
85 struct list_head queued_ios
;
88 struct work_struct trigger_event
;
91 * We must use a mempool of dm_mpath_io structs so that we
92 * can resubmit bios on error.
96 struct mutex work_mutex
;
98 unsigned suspended
; /* Don't create new I/O internally when set. */
102 * Context information attached to each bio we process.
105 struct pgpath
*pgpath
;
109 typedef int (*action_fn
) (struct pgpath
*pgpath
);
111 #define MIN_IOS 256 /* Mempool size */
113 static struct kmem_cache
*_mpio_cache
;
115 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
116 static void process_queued_ios(struct work_struct
*work
);
117 static void trigger_event(struct work_struct
*work
);
118 static void activate_path(struct work_struct
*work
);
121 /*-----------------------------------------------
122 * Allocation routines
123 *-----------------------------------------------*/
125 static struct pgpath
*alloc_pgpath(void)
127 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
130 pgpath
->is_active
= 1;
131 INIT_WORK(&pgpath
->activate_path
, activate_path
);
137 static void free_pgpath(struct pgpath
*pgpath
)
142 static struct priority_group
*alloc_priority_group(void)
144 struct priority_group
*pg
;
146 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
149 INIT_LIST_HEAD(&pg
->pgpaths
);
154 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
156 struct pgpath
*pgpath
, *tmp
;
157 struct multipath
*m
= ti
->private;
159 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
160 list_del(&pgpath
->list
);
161 if (m
->hw_handler_name
)
162 scsi_dh_detach(bdev_get_queue(pgpath
->path
.dev
->bdev
));
163 dm_put_device(ti
, pgpath
->path
.dev
);
168 static void free_priority_group(struct priority_group
*pg
,
169 struct dm_target
*ti
)
171 struct path_selector
*ps
= &pg
->ps
;
174 ps
->type
->destroy(ps
);
175 dm_put_path_selector(ps
->type
);
178 free_pgpaths(&pg
->pgpaths
, ti
);
182 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
186 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
188 INIT_LIST_HEAD(&m
->priority_groups
);
189 INIT_LIST_HEAD(&m
->queued_ios
);
190 spin_lock_init(&m
->lock
);
192 INIT_WORK(&m
->process_queued_ios
, process_queued_ios
);
193 INIT_WORK(&m
->trigger_event
, trigger_event
);
194 mutex_init(&m
->work_mutex
);
195 m
->mpio_pool
= mempool_create_slab_pool(MIN_IOS
, _mpio_cache
);
207 static void free_multipath(struct multipath
*m
)
209 struct priority_group
*pg
, *tmp
;
211 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
213 free_priority_group(pg
, m
->ti
);
216 kfree(m
->hw_handler_name
);
217 kfree(m
->hw_handler_params
);
218 mempool_destroy(m
->mpio_pool
);
223 /*-----------------------------------------------
225 *-----------------------------------------------*/
227 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
229 m
->current_pg
= pgpath
->pg
;
231 /* Must we initialise the PG first, and queue I/O till it's ready? */
232 if (m
->hw_handler_name
) {
233 m
->pg_init_required
= 1;
236 m
->pg_init_required
= 0;
240 m
->pg_init_count
= 0;
243 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
246 struct dm_path
*path
;
248 path
= pg
->ps
.type
->select_path(&pg
->ps
, &m
->repeat_count
, nr_bytes
);
252 m
->current_pgpath
= path_to_pgpath(path
);
254 if (m
->current_pg
!= pg
)
255 __switch_pg(m
, m
->current_pgpath
);
260 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
262 struct priority_group
*pg
;
263 unsigned bypassed
= 1;
265 if (!m
->nr_valid_paths
)
268 /* Were we instructed to switch PG? */
272 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
276 /* Don't change PG until it has no remaining paths */
277 if (m
->current_pg
&& !__choose_path_in_pg(m
, m
->current_pg
, nr_bytes
))
281 * Loop through priority groups until we find a valid path.
282 * First time we skip PGs marked 'bypassed'.
283 * Second time we only try the ones we skipped.
286 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
287 if (pg
->bypassed
== bypassed
)
289 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
292 } while (bypassed
--);
295 m
->current_pgpath
= NULL
;
296 m
->current_pg
= NULL
;
300 * Check whether bios must be queued in the device-mapper core rather
301 * than here in the target.
303 * m->lock must be held on entry.
305 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
306 * same value then we are not between multipath_presuspend()
307 * and multipath_resume() calls and we have no need to check
308 * for the DMF_NOFLUSH_SUSPENDING flag.
310 static int __must_push_back(struct multipath
*m
)
312 return (m
->queue_if_no_path
!= m
->saved_queue_if_no_path
&&
313 dm_noflush_suspending(m
->ti
));
316 static int map_io(struct multipath
*m
, struct request
*clone
,
317 struct dm_mpath_io
*mpio
, unsigned was_queued
)
319 int r
= DM_MAPIO_REMAPPED
;
320 size_t nr_bytes
= blk_rq_bytes(clone
);
322 struct pgpath
*pgpath
;
323 struct block_device
*bdev
;
325 spin_lock_irqsave(&m
->lock
, flags
);
327 /* Do we need to select a new pgpath? */
328 if (!m
->current_pgpath
||
329 (!m
->queue_io
&& (m
->repeat_count
&& --m
->repeat_count
== 0)))
330 __choose_pgpath(m
, nr_bytes
);
332 pgpath
= m
->current_pgpath
;
337 if ((pgpath
&& m
->queue_io
) ||
338 (!pgpath
&& m
->queue_if_no_path
)) {
339 /* Queue for the daemon to resubmit */
340 list_add_tail(&clone
->queuelist
, &m
->queued_ios
);
342 if ((m
->pg_init_required
&& !m
->pg_init_in_progress
) ||
344 queue_work(kmultipathd
, &m
->process_queued_ios
);
346 r
= DM_MAPIO_SUBMITTED
;
348 bdev
= pgpath
->path
.dev
->bdev
;
349 clone
->q
= bdev_get_queue(bdev
);
350 clone
->rq_disk
= bdev
->bd_disk
;
351 } else if (__must_push_back(m
))
352 r
= DM_MAPIO_REQUEUE
;
354 r
= -EIO
; /* Failed */
356 mpio
->pgpath
= pgpath
;
357 mpio
->nr_bytes
= nr_bytes
;
359 if (r
== DM_MAPIO_REMAPPED
&& pgpath
->pg
->ps
.type
->start_io
)
360 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
, &pgpath
->path
,
363 spin_unlock_irqrestore(&m
->lock
, flags
);
369 * If we run out of usable paths, should we queue I/O or error it?
371 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
372 unsigned save_old_value
)
376 spin_lock_irqsave(&m
->lock
, flags
);
379 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
381 m
->saved_queue_if_no_path
= queue_if_no_path
;
382 m
->queue_if_no_path
= queue_if_no_path
;
383 if (!m
->queue_if_no_path
&& m
->queue_size
)
384 queue_work(kmultipathd
, &m
->process_queued_ios
);
386 spin_unlock_irqrestore(&m
->lock
, flags
);
391 /*-----------------------------------------------------------------
392 * The multipath daemon is responsible for resubmitting queued ios.
393 *---------------------------------------------------------------*/
395 static void dispatch_queued_ios(struct multipath
*m
)
399 struct dm_mpath_io
*mpio
;
400 union map_info
*info
;
401 struct request
*clone
, *n
;
404 spin_lock_irqsave(&m
->lock
, flags
);
405 list_splice_init(&m
->queued_ios
, &cl
);
406 spin_unlock_irqrestore(&m
->lock
, flags
);
408 list_for_each_entry_safe(clone
, n
, &cl
, queuelist
) {
409 list_del_init(&clone
->queuelist
);
411 info
= dm_get_rq_mapinfo(clone
);
414 r
= map_io(m
, clone
, mpio
, 1);
416 mempool_free(mpio
, m
->mpio_pool
);
417 dm_kill_unmapped_request(clone
, r
);
418 } else if (r
== DM_MAPIO_REMAPPED
)
419 dm_dispatch_request(clone
);
420 else if (r
== DM_MAPIO_REQUEUE
) {
421 mempool_free(mpio
, m
->mpio_pool
);
422 dm_requeue_unmapped_request(clone
);
427 static void process_queued_ios(struct work_struct
*work
)
429 struct multipath
*m
=
430 container_of(work
, struct multipath
, process_queued_ios
);
431 struct pgpath
*pgpath
= NULL
, *tmp
;
432 unsigned must_queue
= 1;
435 spin_lock_irqsave(&m
->lock
, flags
);
440 if (!m
->current_pgpath
)
441 __choose_pgpath(m
, 0);
443 pgpath
= m
->current_pgpath
;
445 if ((pgpath
&& !m
->queue_io
) ||
446 (!pgpath
&& !m
->queue_if_no_path
))
449 if (m
->pg_init_required
&& !m
->pg_init_in_progress
&& pgpath
) {
451 m
->pg_init_required
= 0;
452 list_for_each_entry(tmp
, &pgpath
->pg
->pgpaths
, list
) {
453 if (queue_work(kmpath_handlerd
, &tmp
->activate_path
))
454 m
->pg_init_in_progress
++;
458 spin_unlock_irqrestore(&m
->lock
, flags
);
460 dispatch_queued_ios(m
);
464 * An event is triggered whenever a path is taken out of use.
465 * Includes path failure and PG bypass.
467 static void trigger_event(struct work_struct
*work
)
469 struct multipath
*m
=
470 container_of(work
, struct multipath
, trigger_event
);
472 dm_table_event(m
->ti
->table
);
475 /*-----------------------------------------------------------------
476 * Constructor/argument parsing:
477 * <#multipath feature args> [<arg>]*
478 * <#hw_handler args> [hw_handler [<arg>]*]
480 * <initial priority group>
481 * [<selector> <#selector args> [<arg>]*
482 * <#paths> <#per-path selector args>
483 * [<path> [<arg>]* ]+ ]+
484 *---------------------------------------------------------------*/
491 static int read_param(struct param
*param
, char *str
, unsigned *v
, char **error
)
494 (sscanf(str
, "%u", v
) != 1) ||
497 *error
= param
->error
;
509 static char *shift(struct arg_set
*as
)
523 static void consume(struct arg_set
*as
, unsigned n
)
525 BUG_ON (as
->argc
< n
);
530 static int parse_path_selector(struct arg_set
*as
, struct priority_group
*pg
,
531 struct dm_target
*ti
)
534 struct path_selector_type
*pst
;
537 static struct param _params
[] = {
538 {0, 1024, "invalid number of path selector args"},
541 pst
= dm_get_path_selector(shift(as
));
543 ti
->error
= "unknown path selector type";
547 r
= read_param(_params
, shift(as
), &ps_argc
, &ti
->error
);
549 dm_put_path_selector(pst
);
553 if (ps_argc
> as
->argc
) {
554 dm_put_path_selector(pst
);
555 ti
->error
= "not enough arguments for path selector";
559 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
561 dm_put_path_selector(pst
);
562 ti
->error
= "path selector constructor failed";
567 consume(as
, ps_argc
);
572 static struct pgpath
*parse_path(struct arg_set
*as
, struct path_selector
*ps
,
573 struct dm_target
*ti
)
577 struct multipath
*m
= ti
->private;
579 /* we need at least a path arg */
581 ti
->error
= "no device given";
582 return ERR_PTR(-EINVAL
);
587 return ERR_PTR(-ENOMEM
);
589 r
= dm_get_device(ti
, shift(as
), ti
->begin
, ti
->len
,
590 dm_table_get_mode(ti
->table
), &p
->path
.dev
);
592 ti
->error
= "error getting device";
596 if (m
->hw_handler_name
) {
597 struct request_queue
*q
= bdev_get_queue(p
->path
.dev
->bdev
);
599 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
602 * Already attached to different hw_handler,
603 * try to reattach with correct one.
606 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
610 ti
->error
= "error attaching hardware handler";
611 dm_put_device(ti
, p
->path
.dev
);
615 if (m
->hw_handler_params
) {
616 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
618 ti
->error
= "unable to set hardware "
619 "handler parameters";
621 dm_put_device(ti
, p
->path
.dev
);
627 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
629 dm_put_device(ti
, p
->path
.dev
);
640 static struct priority_group
*parse_priority_group(struct arg_set
*as
,
643 static struct param _params
[] = {
644 {1, 1024, "invalid number of paths"},
645 {0, 1024, "invalid number of selector args"}
649 unsigned i
, nr_selector_args
, nr_params
;
650 struct priority_group
*pg
;
651 struct dm_target
*ti
= m
->ti
;
655 ti
->error
= "not enough priority group arguments";
656 return ERR_PTR(-EINVAL
);
659 pg
= alloc_priority_group();
661 ti
->error
= "couldn't allocate priority group";
662 return ERR_PTR(-ENOMEM
);
666 r
= parse_path_selector(as
, pg
, ti
);
673 r
= read_param(_params
, shift(as
), &pg
->nr_pgpaths
, &ti
->error
);
677 r
= read_param(_params
+ 1, shift(as
), &nr_selector_args
, &ti
->error
);
681 nr_params
= 1 + nr_selector_args
;
682 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
683 struct pgpath
*pgpath
;
684 struct arg_set path_args
;
686 if (as
->argc
< nr_params
) {
687 ti
->error
= "not enough path parameters";
691 path_args
.argc
= nr_params
;
692 path_args
.argv
= as
->argv
;
694 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
695 if (IS_ERR(pgpath
)) {
701 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
702 consume(as
, nr_params
);
708 free_priority_group(pg
, ti
);
712 static int parse_hw_handler(struct arg_set
*as
, struct multipath
*m
)
716 struct dm_target
*ti
= m
->ti
;
718 static struct param _params
[] = {
719 {0, 1024, "invalid number of hardware handler args"},
722 if (read_param(_params
, shift(as
), &hw_argc
, &ti
->error
))
728 if (hw_argc
> as
->argc
) {
729 ti
->error
= "not enough arguments for hardware handler";
733 m
->hw_handler_name
= kstrdup(shift(as
), GFP_KERNEL
);
734 request_module("scsi_dh_%s", m
->hw_handler_name
);
735 if (scsi_dh_handler_exist(m
->hw_handler_name
) == 0) {
736 ti
->error
= "unknown hardware handler type";
745 for (i
= 0; i
<= hw_argc
- 2; i
++)
746 len
+= strlen(as
->argv
[i
]) + 1;
747 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
749 ti
->error
= "memory allocation failed";
753 j
= sprintf(p
, "%d", hw_argc
- 1);
754 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
755 j
= sprintf(p
, "%s", as
->argv
[i
]);
757 consume(as
, hw_argc
- 1);
761 kfree(m
->hw_handler_name
);
762 m
->hw_handler_name
= NULL
;
766 static int parse_features(struct arg_set
*as
, struct multipath
*m
)
770 struct dm_target
*ti
= m
->ti
;
771 const char *param_name
;
773 static struct param _params
[] = {
774 {0, 3, "invalid number of feature args"},
775 {1, 50, "pg_init_retries must be between 1 and 50"},
778 r
= read_param(_params
, shift(as
), &argc
, &ti
->error
);
786 param_name
= shift(as
);
789 if (!strnicmp(param_name
, MESG_STR("queue_if_no_path"))) {
790 r
= queue_if_no_path(m
, 1, 0);
794 if (!strnicmp(param_name
, MESG_STR("pg_init_retries")) &&
796 r
= read_param(_params
+ 1, shift(as
),
797 &m
->pg_init_retries
, &ti
->error
);
802 ti
->error
= "Unrecognised multipath feature request";
804 } while (argc
&& !r
);
809 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
812 /* target parameters */
813 static struct param _params
[] = {
814 {1, 1024, "invalid number of priority groups"},
815 {1, 1024, "invalid initial priority group number"},
821 unsigned pg_count
= 0;
822 unsigned next_pg_num
;
827 m
= alloc_multipath(ti
);
829 ti
->error
= "can't allocate multipath";
833 r
= parse_features(&as
, m
);
837 r
= parse_hw_handler(&as
, m
);
841 r
= read_param(_params
, shift(&as
), &m
->nr_priority_groups
, &ti
->error
);
845 r
= read_param(_params
+ 1, shift(&as
), &next_pg_num
, &ti
->error
);
849 /* parse the priority groups */
851 struct priority_group
*pg
;
853 pg
= parse_priority_group(&as
, m
);
859 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
860 list_add_tail(&pg
->list
, &m
->priority_groups
);
862 pg
->pg_num
= pg_count
;
867 if (pg_count
!= m
->nr_priority_groups
) {
868 ti
->error
= "priority group count mismatch";
873 ti
->num_flush_requests
= 1;
882 static void flush_multipath_work(void)
884 flush_workqueue(kmpath_handlerd
);
885 flush_workqueue(kmultipathd
);
886 flush_scheduled_work();
889 static void multipath_dtr(struct dm_target
*ti
)
891 struct multipath
*m
= ti
->private;
893 flush_multipath_work();
898 * Map cloned requests
900 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
901 union map_info
*map_context
)
904 struct dm_mpath_io
*mpio
;
905 struct multipath
*m
= (struct multipath
*) ti
->private;
907 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
909 /* ENOMEM, requeue */
910 return DM_MAPIO_REQUEUE
;
911 memset(mpio
, 0, sizeof(*mpio
));
913 map_context
->ptr
= mpio
;
914 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
915 r
= map_io(m
, clone
, mpio
, 0);
916 if (r
< 0 || r
== DM_MAPIO_REQUEUE
)
917 mempool_free(mpio
, m
->mpio_pool
);
923 * Take a path out of use.
925 static int fail_path(struct pgpath
*pgpath
)
928 struct multipath
*m
= pgpath
->pg
->m
;
930 spin_lock_irqsave(&m
->lock
, flags
);
932 if (!pgpath
->is_active
)
935 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
937 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
938 pgpath
->is_active
= 0;
939 pgpath
->fail_count
++;
943 if (pgpath
== m
->current_pgpath
)
944 m
->current_pgpath
= NULL
;
946 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
947 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
949 schedule_work(&m
->trigger_event
);
952 spin_unlock_irqrestore(&m
->lock
, flags
);
958 * Reinstate a previously-failed path
960 static int reinstate_path(struct pgpath
*pgpath
)
964 struct multipath
*m
= pgpath
->pg
->m
;
966 spin_lock_irqsave(&m
->lock
, flags
);
968 if (pgpath
->is_active
)
971 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
972 DMWARN("Reinstate path not supported by path selector %s",
973 pgpath
->pg
->ps
.type
->name
);
978 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
982 pgpath
->is_active
= 1;
984 if (!m
->nr_valid_paths
++ && m
->queue_size
) {
985 m
->current_pgpath
= NULL
;
986 queue_work(kmultipathd
, &m
->process_queued_ios
);
987 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
988 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
))
989 m
->pg_init_in_progress
++;
992 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
993 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
995 schedule_work(&m
->trigger_event
);
998 spin_unlock_irqrestore(&m
->lock
, flags
);
1004 * Fail or reinstate all paths that match the provided struct dm_dev.
1006 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1010 struct pgpath
*pgpath
;
1011 struct priority_group
*pg
;
1013 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1014 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1015 if (pgpath
->path
.dev
== dev
)
1024 * Temporarily try to avoid having to use the specified PG
1026 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1029 unsigned long flags
;
1031 spin_lock_irqsave(&m
->lock
, flags
);
1033 pg
->bypassed
= bypassed
;
1034 m
->current_pgpath
= NULL
;
1035 m
->current_pg
= NULL
;
1037 spin_unlock_irqrestore(&m
->lock
, flags
);
1039 schedule_work(&m
->trigger_event
);
1043 * Switch to using the specified PG from the next I/O that gets mapped
1045 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1047 struct priority_group
*pg
;
1049 unsigned long flags
;
1051 if (!pgstr
|| (sscanf(pgstr
, "%u", &pgnum
) != 1) || !pgnum
||
1052 (pgnum
> m
->nr_priority_groups
)) {
1053 DMWARN("invalid PG number supplied to switch_pg_num");
1057 spin_lock_irqsave(&m
->lock
, flags
);
1058 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1063 m
->current_pgpath
= NULL
;
1064 m
->current_pg
= NULL
;
1067 spin_unlock_irqrestore(&m
->lock
, flags
);
1069 schedule_work(&m
->trigger_event
);
1074 * Set/clear bypassed status of a PG.
1075 * PGs are numbered upwards from 1 in the order they were declared.
1077 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1079 struct priority_group
*pg
;
1082 if (!pgstr
|| (sscanf(pgstr
, "%u", &pgnum
) != 1) || !pgnum
||
1083 (pgnum
> m
->nr_priority_groups
)) {
1084 DMWARN("invalid PG number supplied to bypass_pg");
1088 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1093 bypass_pg(m
, pg
, bypassed
);
1098 * Should we retry pg_init immediately?
1100 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1102 unsigned long flags
;
1103 int limit_reached
= 0;
1105 spin_lock_irqsave(&m
->lock
, flags
);
1107 if (m
->pg_init_count
<= m
->pg_init_retries
)
1108 m
->pg_init_required
= 1;
1112 spin_unlock_irqrestore(&m
->lock
, flags
);
1114 return limit_reached
;
1117 static void pg_init_done(void *data
, int errors
)
1119 struct dm_path
*path
= data
;
1120 struct pgpath
*pgpath
= path_to_pgpath(path
);
1121 struct priority_group
*pg
= pgpath
->pg
;
1122 struct multipath
*m
= pg
->m
;
1123 unsigned long flags
;
1125 /* device or driver problems */
1130 if (!m
->hw_handler_name
) {
1134 DMERR("Cannot failover device because scsi_dh_%s was not "
1135 "loaded.", m
->hw_handler_name
);
1137 * Fail path for now, so we do not ping pong
1141 case SCSI_DH_DEV_TEMP_BUSY
:
1143 * Probably doing something like FW upgrade on the
1144 * controller so try the other pg.
1146 bypass_pg(m
, pg
, 1);
1148 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1150 case SCSI_DH_IMM_RETRY
:
1151 case SCSI_DH_RES_TEMP_UNAVAIL
:
1152 if (pg_init_limit_reached(m
, pgpath
))
1158 * We probably do not want to fail the path for a device
1159 * error, but this is what the old dm did. In future
1160 * patches we can do more advanced handling.
1165 spin_lock_irqsave(&m
->lock
, flags
);
1167 if (pgpath
== m
->current_pgpath
) {
1168 DMERR("Could not failover device. Error %d.", errors
);
1169 m
->current_pgpath
= NULL
;
1170 m
->current_pg
= NULL
;
1172 } else if (!m
->pg_init_required
) {
1177 m
->pg_init_in_progress
--;
1178 if (!m
->pg_init_in_progress
)
1179 queue_work(kmultipathd
, &m
->process_queued_ios
);
1180 spin_unlock_irqrestore(&m
->lock
, flags
);
1183 static void activate_path(struct work_struct
*work
)
1185 struct pgpath
*pgpath
=
1186 container_of(work
, struct pgpath
, activate_path
);
1188 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1189 pg_init_done
, &pgpath
->path
);
1195 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1196 int error
, struct dm_mpath_io
*mpio
)
1199 * We don't queue any clone request inside the multipath target
1200 * during end I/O handling, since those clone requests don't have
1201 * bio clones. If we queue them inside the multipath target,
1202 * we need to make bio clones, that requires memory allocation.
1203 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1204 * don't have bio clones.)
1205 * Instead of queueing the clone request here, we queue the original
1206 * request into dm core, which will remake a clone request and
1207 * clone bios for it and resubmit it later.
1209 int r
= DM_ENDIO_REQUEUE
;
1210 unsigned long flags
;
1212 if (!error
&& !clone
->errors
)
1213 return 0; /* I/O complete */
1215 if (error
== -EOPNOTSUPP
)
1219 fail_path(mpio
->pgpath
);
1221 spin_lock_irqsave(&m
->lock
, flags
);
1222 if (!m
->nr_valid_paths
&& !m
->queue_if_no_path
&& !__must_push_back(m
))
1224 spin_unlock_irqrestore(&m
->lock
, flags
);
1229 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1230 int error
, union map_info
*map_context
)
1232 struct multipath
*m
= ti
->private;
1233 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1234 struct pgpath
*pgpath
= mpio
->pgpath
;
1235 struct path_selector
*ps
;
1238 r
= do_end_io(m
, clone
, error
, mpio
);
1240 ps
= &pgpath
->pg
->ps
;
1241 if (ps
->type
->end_io
)
1242 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1244 mempool_free(mpio
, m
->mpio_pool
);
1250 * Suspend can't complete until all the I/O is processed so if
1251 * the last path fails we must error any remaining I/O.
1252 * Note that if the freeze_bdev fails while suspending, the
1253 * queue_if_no_path state is lost - userspace should reset it.
1255 static void multipath_presuspend(struct dm_target
*ti
)
1257 struct multipath
*m
= (struct multipath
*) ti
->private;
1259 queue_if_no_path(m
, 0, 1);
1262 static void multipath_postsuspend(struct dm_target
*ti
)
1264 struct multipath
*m
= ti
->private;
1266 mutex_lock(&m
->work_mutex
);
1268 flush_multipath_work();
1269 mutex_unlock(&m
->work_mutex
);
1273 * Restore the queue_if_no_path setting.
1275 static void multipath_resume(struct dm_target
*ti
)
1277 struct multipath
*m
= (struct multipath
*) ti
->private;
1278 unsigned long flags
;
1280 mutex_lock(&m
->work_mutex
);
1282 mutex_unlock(&m
->work_mutex
);
1284 spin_lock_irqsave(&m
->lock
, flags
);
1285 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1286 spin_unlock_irqrestore(&m
->lock
, flags
);
1290 * Info output has the following format:
1291 * num_multipath_feature_args [multipath_feature_args]*
1292 * num_handler_status_args [handler_status_args]*
1293 * num_groups init_group_number
1294 * [A|D|E num_ps_status_args [ps_status_args]*
1295 * num_paths num_selector_args
1296 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1298 * Table output has the following format (identical to the constructor string):
1299 * num_feature_args [features_args]*
1300 * num_handler_args hw_handler [hw_handler_args]*
1301 * num_groups init_group_number
1302 * [priority selector-name num_ps_args [ps_args]*
1303 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1305 static int multipath_status(struct dm_target
*ti
, status_type_t type
,
1306 char *result
, unsigned int maxlen
)
1309 unsigned long flags
;
1310 struct multipath
*m
= (struct multipath
*) ti
->private;
1311 struct priority_group
*pg
;
1316 spin_lock_irqsave(&m
->lock
, flags
);
1319 if (type
== STATUSTYPE_INFO
)
1320 DMEMIT("2 %u %u ", m
->queue_size
, m
->pg_init_count
);
1322 DMEMIT("%u ", m
->queue_if_no_path
+
1323 (m
->pg_init_retries
> 0) * 2);
1324 if (m
->queue_if_no_path
)
1325 DMEMIT("queue_if_no_path ");
1326 if (m
->pg_init_retries
)
1327 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1330 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1333 DMEMIT("1 %s ", m
->hw_handler_name
);
1335 DMEMIT("%u ", m
->nr_priority_groups
);
1338 pg_num
= m
->next_pg
->pg_num
;
1339 else if (m
->current_pg
)
1340 pg_num
= m
->current_pg
->pg_num
;
1344 DMEMIT("%u ", pg_num
);
1347 case STATUSTYPE_INFO
:
1348 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1350 state
= 'D'; /* Disabled */
1351 else if (pg
== m
->current_pg
)
1352 state
= 'A'; /* Currently Active */
1354 state
= 'E'; /* Enabled */
1356 DMEMIT("%c ", state
);
1358 if (pg
->ps
.type
->status
)
1359 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1365 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1366 pg
->ps
.type
->info_args
);
1368 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1369 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1370 p
->is_active
? "A" : "F",
1372 if (pg
->ps
.type
->status
)
1373 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1374 &p
->path
, type
, result
+ sz
,
1380 case STATUSTYPE_TABLE
:
1381 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1382 DMEMIT("%s ", pg
->ps
.type
->name
);
1384 if (pg
->ps
.type
->status
)
1385 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1391 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1392 pg
->ps
.type
->table_args
);
1394 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1395 DMEMIT("%s ", p
->path
.dev
->name
);
1396 if (pg
->ps
.type
->status
)
1397 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1398 &p
->path
, type
, result
+ sz
,
1405 spin_unlock_irqrestore(&m
->lock
, flags
);
1410 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1414 struct multipath
*m
= (struct multipath
*) ti
->private;
1417 mutex_lock(&m
->work_mutex
);
1424 if (dm_suspended(ti
)) {
1430 if (!strnicmp(argv
[0], MESG_STR("queue_if_no_path"))) {
1431 r
= queue_if_no_path(m
, 1, 0);
1433 } else if (!strnicmp(argv
[0], MESG_STR("fail_if_no_path"))) {
1434 r
= queue_if_no_path(m
, 0, 0);
1440 DMWARN("Unrecognised multipath message received.");
1444 if (!strnicmp(argv
[0], MESG_STR("disable_group"))) {
1445 r
= bypass_pg_num(m
, argv
[1], 1);
1447 } else if (!strnicmp(argv
[0], MESG_STR("enable_group"))) {
1448 r
= bypass_pg_num(m
, argv
[1], 0);
1450 } else if (!strnicmp(argv
[0], MESG_STR("switch_group"))) {
1451 r
= switch_pg_num(m
, argv
[1]);
1453 } else if (!strnicmp(argv
[0], MESG_STR("reinstate_path")))
1454 action
= reinstate_path
;
1455 else if (!strnicmp(argv
[0], MESG_STR("fail_path")))
1458 DMWARN("Unrecognised multipath message received.");
1462 r
= dm_get_device(ti
, argv
[1], ti
->begin
, ti
->len
,
1463 dm_table_get_mode(ti
->table
), &dev
);
1465 DMWARN("message: error getting device %s",
1470 r
= action_dev(m
, dev
, action
);
1472 dm_put_device(ti
, dev
);
1475 mutex_unlock(&m
->work_mutex
);
1479 static int multipath_ioctl(struct dm_target
*ti
, unsigned int cmd
,
1482 struct multipath
*m
= (struct multipath
*) ti
->private;
1483 struct block_device
*bdev
= NULL
;
1485 unsigned long flags
;
1488 spin_lock_irqsave(&m
->lock
, flags
);
1490 if (!m
->current_pgpath
)
1491 __choose_pgpath(m
, 0);
1493 if (m
->current_pgpath
) {
1494 bdev
= m
->current_pgpath
->path
.dev
->bdev
;
1495 mode
= m
->current_pgpath
->path
.dev
->mode
;
1503 spin_unlock_irqrestore(&m
->lock
, flags
);
1505 return r
? : __blkdev_driver_ioctl(bdev
, mode
, cmd
, arg
);
1508 static int multipath_iterate_devices(struct dm_target
*ti
,
1509 iterate_devices_callout_fn fn
, void *data
)
1511 struct multipath
*m
= ti
->private;
1512 struct priority_group
*pg
;
1516 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1517 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1518 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1528 static int __pgpath_busy(struct pgpath
*pgpath
)
1530 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1532 return dm_underlying_device_busy(q
);
1536 * We return "busy", only when we can map I/Os but underlying devices
1537 * are busy (so even if we map I/Os now, the I/Os will wait on
1538 * the underlying queue).
1539 * In other words, if we want to kill I/Os or queue them inside us
1540 * due to map unavailability, we don't return "busy". Otherwise,
1541 * dm core won't give us the I/Os and we can't do what we want.
1543 static int multipath_busy(struct dm_target
*ti
)
1545 int busy
= 0, has_active
= 0;
1546 struct multipath
*m
= ti
->private;
1547 struct priority_group
*pg
;
1548 struct pgpath
*pgpath
;
1549 unsigned long flags
;
1551 spin_lock_irqsave(&m
->lock
, flags
);
1553 /* Guess which priority_group will be used at next mapping time */
1554 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1556 else if (likely(m
->current_pg
))
1560 * We don't know which pg will be used at next mapping time.
1561 * We don't call __choose_pgpath() here to avoid to trigger
1562 * pg_init just by busy checking.
1563 * So we don't know whether underlying devices we will be using
1564 * at next mapping time are busy or not. Just try mapping.
1569 * If there is one non-busy active path at least, the path selector
1570 * will be able to select it. So we consider such a pg as not busy.
1573 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1574 if (pgpath
->is_active
) {
1577 if (!__pgpath_busy(pgpath
)) {
1585 * No active path in this pg, so this pg won't be used and
1586 * the current_pg will be changed at next mapping time.
1587 * We need to try mapping to determine it.
1592 spin_unlock_irqrestore(&m
->lock
, flags
);
1597 /*-----------------------------------------------------------------
1599 *---------------------------------------------------------------*/
1600 static struct target_type multipath_target
= {
1601 .name
= "multipath",
1602 .version
= {1, 1, 1},
1603 .module
= THIS_MODULE
,
1604 .ctr
= multipath_ctr
,
1605 .dtr
= multipath_dtr
,
1606 .map_rq
= multipath_map
,
1607 .rq_end_io
= multipath_end_io
,
1608 .presuspend
= multipath_presuspend
,
1609 .postsuspend
= multipath_postsuspend
,
1610 .resume
= multipath_resume
,
1611 .status
= multipath_status
,
1612 .message
= multipath_message
,
1613 .ioctl
= multipath_ioctl
,
1614 .iterate_devices
= multipath_iterate_devices
,
1615 .busy
= multipath_busy
,
1618 static int __init
dm_multipath_init(void)
1622 /* allocate a slab for the dm_ios */
1623 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1627 r
= dm_register_target(&multipath_target
);
1629 DMERR("register failed %d", r
);
1630 kmem_cache_destroy(_mpio_cache
);
1634 kmultipathd
= create_workqueue("kmpathd");
1636 DMERR("failed to create workqueue kmpathd");
1637 dm_unregister_target(&multipath_target
);
1638 kmem_cache_destroy(_mpio_cache
);
1643 * A separate workqueue is used to handle the device handlers
1644 * to avoid overloading existing workqueue. Overloading the
1645 * old workqueue would also create a bottleneck in the
1646 * path of the storage hardware device activation.
1648 kmpath_handlerd
= create_singlethread_workqueue("kmpath_handlerd");
1649 if (!kmpath_handlerd
) {
1650 DMERR("failed to create workqueue kmpath_handlerd");
1651 destroy_workqueue(kmultipathd
);
1652 dm_unregister_target(&multipath_target
);
1653 kmem_cache_destroy(_mpio_cache
);
1657 DMINFO("version %u.%u.%u loaded",
1658 multipath_target
.version
[0], multipath_target
.version
[1],
1659 multipath_target
.version
[2]);
1664 static void __exit
dm_multipath_exit(void)
1666 destroy_workqueue(kmpath_handlerd
);
1667 destroy_workqueue(kmultipathd
);
1669 dm_unregister_target(&multipath_target
);
1670 kmem_cache_destroy(_mpio_cache
);
1673 module_init(dm_multipath_init
);
1674 module_exit(dm_multipath_exit
);
1676 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1677 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1678 MODULE_LICENSE("GPL");