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>
26 #include <linux/blk-mq.h>
28 #define DM_MSG_PREFIX "multipath"
29 #define DM_PG_INIT_DELAY_MSECS 2000
30 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
34 struct list_head list
;
36 struct priority_group
*pg
; /* Owning PG */
37 unsigned fail_count
; /* Cumulative failure count */
40 struct delayed_work activate_path
;
42 bool is_active
:1; /* Path status */
45 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
48 * Paths are grouped into Priority Groups and numbered from 1 upwards.
49 * Each has a path selector which controls which path gets used.
51 struct priority_group
{
52 struct list_head list
;
54 struct multipath
*m
; /* Owning multipath instance */
55 struct path_selector ps
;
57 unsigned pg_num
; /* Reference number */
58 unsigned nr_pgpaths
; /* Number of paths in PG */
59 struct list_head pgpaths
;
61 bool bypassed
:1; /* Temporarily bypass this PG? */
64 /* Multipath context */
66 struct list_head list
;
69 const char *hw_handler_name
;
70 char *hw_handler_params
;
74 unsigned nr_priority_groups
;
75 struct list_head priority_groups
;
77 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
79 unsigned pg_init_in_progress
; /* Only one pg_init allowed at once */
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 */
86 bool queue_io
:1; /* Must we queue all I/O? */
87 bool queue_if_no_path
:1; /* Queue I/O if last path fails? */
88 bool saved_queue_if_no_path
:1; /* Saved state during suspension */
89 bool retain_attached_hw_handler
:1; /* If there's already a hw_handler present, don't change it. */
90 bool pg_init_disabled
:1; /* pg_init is not currently allowed */
91 bool pg_init_required
:1; /* pg_init needs calling? */
92 bool pg_init_delay_retry
:1; /* Delay pg_init retry? */
94 unsigned pg_init_retries
; /* Number of times to retry pg_init */
95 unsigned pg_init_count
; /* Number of times pg_init called */
96 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
98 struct work_struct trigger_event
;
101 * We must use a mempool of dm_mpath_io structs so that we
102 * can resubmit bios on error.
104 mempool_t
*mpio_pool
;
106 struct mutex work_mutex
;
110 * Context information attached to each bio we process.
113 struct pgpath
*pgpath
;
117 typedef int (*action_fn
) (struct pgpath
*pgpath
);
119 static struct kmem_cache
*_mpio_cache
;
121 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
122 static void trigger_event(struct work_struct
*work
);
123 static void activate_path(struct work_struct
*work
);
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
= true;
136 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path
);
142 static void free_pgpath(struct pgpath
*pgpath
)
147 static struct priority_group
*alloc_priority_group(void)
149 struct priority_group
*pg
;
151 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
154 INIT_LIST_HEAD(&pg
->pgpaths
);
159 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
161 struct pgpath
*pgpath
, *tmp
;
163 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
164 list_del(&pgpath
->list
);
165 dm_put_device(ti
, pgpath
->path
.dev
);
170 static void free_priority_group(struct priority_group
*pg
,
171 struct dm_target
*ti
)
173 struct path_selector
*ps
= &pg
->ps
;
176 ps
->type
->destroy(ps
);
177 dm_put_path_selector(ps
->type
);
180 free_pgpaths(&pg
->pgpaths
, ti
);
184 static struct multipath
*alloc_multipath(struct dm_target
*ti
, bool use_blk_mq
)
188 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
190 INIT_LIST_HEAD(&m
->priority_groups
);
191 spin_lock_init(&m
->lock
);
193 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
194 INIT_WORK(&m
->trigger_event
, trigger_event
);
195 init_waitqueue_head(&m
->pg_init_wait
);
196 mutex_init(&m
->work_mutex
);
200 unsigned min_ios
= dm_get_reserved_rq_based_ios();
202 m
->mpio_pool
= mempool_create_slab_pool(min_ios
, _mpio_cache
);
216 static void free_multipath(struct multipath
*m
)
218 struct priority_group
*pg
, *tmp
;
220 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
222 free_priority_group(pg
, m
->ti
);
225 kfree(m
->hw_handler_name
);
226 kfree(m
->hw_handler_params
);
227 mempool_destroy(m
->mpio_pool
);
231 static struct dm_mpath_io
*get_mpio(union map_info
*info
)
236 static struct dm_mpath_io
*set_mpio(struct multipath
*m
, union map_info
*info
)
238 struct dm_mpath_io
*mpio
;
241 /* Use blk-mq pdu memory requested via per_io_data_size */
242 mpio
= get_mpio(info
);
243 memset(mpio
, 0, sizeof(*mpio
));
247 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
251 memset(mpio
, 0, sizeof(*mpio
));
257 static void clear_request_fn_mpio(struct multipath
*m
, union map_info
*info
)
259 /* Only needed for non blk-mq (.request_fn) multipath */
261 struct dm_mpath_io
*mpio
= info
->ptr
;
264 mempool_free(mpio
, m
->mpio_pool
);
268 /*-----------------------------------------------
270 *-----------------------------------------------*/
272 static int __pg_init_all_paths(struct multipath
*m
)
274 struct pgpath
*pgpath
;
275 unsigned long pg_init_delay
= 0;
277 if (m
->pg_init_in_progress
|| m
->pg_init_disabled
)
281 m
->pg_init_required
= false;
283 /* Check here to reset pg_init_required */
287 if (m
->pg_init_delay_retry
)
288 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
289 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
290 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
291 /* Skip failed paths */
292 if (!pgpath
->is_active
)
294 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
296 m
->pg_init_in_progress
++;
298 return m
->pg_init_in_progress
;
301 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
303 m
->current_pg
= pgpath
->pg
;
305 /* Must we initialise the PG first, and queue I/O till it's ready? */
306 if (m
->hw_handler_name
) {
307 m
->pg_init_required
= true;
310 m
->pg_init_required
= false;
314 m
->pg_init_count
= 0;
317 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
320 struct dm_path
*path
;
322 path
= pg
->ps
.type
->select_path(&pg
->ps
, nr_bytes
);
326 m
->current_pgpath
= path_to_pgpath(path
);
328 if (m
->current_pg
!= pg
)
329 __switch_pg(m
, m
->current_pgpath
);
334 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
336 struct priority_group
*pg
;
337 bool bypassed
= true;
339 if (!m
->nr_valid_paths
) {
344 /* Were we instructed to switch PG? */
348 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
352 /* Don't change PG until it has no remaining paths */
353 if (m
->current_pg
&& !__choose_path_in_pg(m
, m
->current_pg
, nr_bytes
))
357 * Loop through priority groups until we find a valid path.
358 * First time we skip PGs marked 'bypassed'.
359 * Second time we only try the ones we skipped, but set
360 * pg_init_delay_retry so we do not hammer controllers.
363 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
364 if (pg
->bypassed
== bypassed
)
366 if (!__choose_path_in_pg(m
, pg
, nr_bytes
)) {
368 m
->pg_init_delay_retry
= true;
372 } while (bypassed
--);
375 m
->current_pgpath
= NULL
;
376 m
->current_pg
= NULL
;
380 * Check whether bios must be queued in the device-mapper core rather
381 * than here in the target.
383 * m->lock must be held on entry.
385 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
386 * same value then we are not between multipath_presuspend()
387 * and multipath_resume() calls and we have no need to check
388 * for the DMF_NOFLUSH_SUSPENDING flag.
390 static int __must_push_back(struct multipath
*m
)
392 return (m
->queue_if_no_path
||
393 (m
->queue_if_no_path
!= m
->saved_queue_if_no_path
&&
394 dm_noflush_suspending(m
->ti
)));
398 * Map cloned requests
400 static int __multipath_map(struct dm_target
*ti
, struct request
*clone
,
401 union map_info
*map_context
,
402 struct request
*rq
, struct request
**__clone
)
404 struct multipath
*m
= ti
->private;
405 int r
= DM_MAPIO_REQUEUE
;
406 size_t nr_bytes
= clone
? blk_rq_bytes(clone
) : blk_rq_bytes(rq
);
407 struct pgpath
*pgpath
;
408 struct block_device
*bdev
;
409 struct dm_mpath_io
*mpio
;
411 spin_lock_irq(&m
->lock
);
413 /* Do we need to select a new pgpath? */
414 if (!m
->current_pgpath
|| !m
->queue_io
)
415 __choose_pgpath(m
, nr_bytes
);
417 pgpath
= m
->current_pgpath
;
420 if (!__must_push_back(m
))
421 r
= -EIO
; /* Failed */
423 } else if (m
->queue_io
|| m
->pg_init_required
) {
424 __pg_init_all_paths(m
);
428 mpio
= set_mpio(m
, map_context
);
430 /* ENOMEM, requeue */
433 mpio
->pgpath
= pgpath
;
434 mpio
->nr_bytes
= nr_bytes
;
436 bdev
= pgpath
->path
.dev
->bdev
;
438 spin_unlock_irq(&m
->lock
);
442 * Old request-based interface: allocated clone is passed in.
443 * Used by: .request_fn stacked on .request_fn path(s).
445 clone
->q
= bdev_get_queue(bdev
);
446 clone
->rq_disk
= bdev
->bd_disk
;
447 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
450 * blk-mq request-based interface; used by both:
451 * .request_fn stacked on blk-mq path(s) and
452 * blk-mq stacked on blk-mq path(s).
454 *__clone
= blk_mq_alloc_request(bdev_get_queue(bdev
),
455 rq_data_dir(rq
), BLK_MQ_REQ_NOWAIT
);
456 if (IS_ERR(*__clone
)) {
457 /* ENOMEM, requeue */
458 clear_request_fn_mpio(m
, map_context
);
461 (*__clone
)->bio
= (*__clone
)->biotail
= NULL
;
462 (*__clone
)->rq_disk
= bdev
->bd_disk
;
463 (*__clone
)->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
466 if (pgpath
->pg
->ps
.type
->start_io
)
467 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
470 return DM_MAPIO_REMAPPED
;
473 spin_unlock_irq(&m
->lock
);
478 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
479 union map_info
*map_context
)
481 return __multipath_map(ti
, clone
, map_context
, NULL
, NULL
);
484 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
485 union map_info
*map_context
,
486 struct request
**clone
)
488 return __multipath_map(ti
, NULL
, map_context
, rq
, clone
);
491 static void multipath_release_clone(struct request
*clone
)
493 blk_mq_free_request(clone
);
497 * If we run out of usable paths, should we queue I/O or error it?
499 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
504 spin_lock_irqsave(&m
->lock
, flags
);
507 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
509 m
->saved_queue_if_no_path
= queue_if_no_path
;
510 m
->queue_if_no_path
= queue_if_no_path
;
511 spin_unlock_irqrestore(&m
->lock
, flags
);
513 if (!queue_if_no_path
)
514 dm_table_run_md_queue_async(m
->ti
->table
);
520 * An event is triggered whenever a path is taken out of use.
521 * Includes path failure and PG bypass.
523 static void trigger_event(struct work_struct
*work
)
525 struct multipath
*m
=
526 container_of(work
, struct multipath
, trigger_event
);
528 dm_table_event(m
->ti
->table
);
531 /*-----------------------------------------------------------------
532 * Constructor/argument parsing:
533 * <#multipath feature args> [<arg>]*
534 * <#hw_handler args> [hw_handler [<arg>]*]
536 * <initial priority group>
537 * [<selector> <#selector args> [<arg>]*
538 * <#paths> <#per-path selector args>
539 * [<path> [<arg>]* ]+ ]+
540 *---------------------------------------------------------------*/
541 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
542 struct dm_target
*ti
)
545 struct path_selector_type
*pst
;
548 static struct dm_arg _args
[] = {
549 {0, 1024, "invalid number of path selector args"},
552 pst
= dm_get_path_selector(dm_shift_arg(as
));
554 ti
->error
= "unknown path selector type";
558 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
560 dm_put_path_selector(pst
);
564 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
566 dm_put_path_selector(pst
);
567 ti
->error
= "path selector constructor failed";
572 dm_consume_args(as
, ps_argc
);
577 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
578 struct dm_target
*ti
)
582 struct multipath
*m
= ti
->private;
583 struct request_queue
*q
= NULL
;
584 const char *attached_handler_name
;
586 /* we need at least a path arg */
588 ti
->error
= "no device given";
589 return ERR_PTR(-EINVAL
);
594 return ERR_PTR(-ENOMEM
);
596 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
599 ti
->error
= "error getting device";
603 if (m
->retain_attached_hw_handler
|| m
->hw_handler_name
)
604 q
= bdev_get_queue(p
->path
.dev
->bdev
);
606 if (m
->retain_attached_hw_handler
) {
608 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
609 if (attached_handler_name
) {
611 * Reset hw_handler_name to match the attached handler
612 * and clear any hw_handler_params associated with the
615 * NB. This modifies the table line to show the actual
616 * handler instead of the original table passed in.
618 kfree(m
->hw_handler_name
);
619 m
->hw_handler_name
= attached_handler_name
;
621 kfree(m
->hw_handler_params
);
622 m
->hw_handler_params
= NULL
;
626 if (m
->hw_handler_name
) {
627 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
629 char b
[BDEVNAME_SIZE
];
631 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
632 bdevname(p
->path
.dev
->bdev
, b
));
636 ti
->error
= "error attaching hardware handler";
637 dm_put_device(ti
, p
->path
.dev
);
641 if (m
->hw_handler_params
) {
642 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
644 ti
->error
= "unable to set hardware "
645 "handler parameters";
646 dm_put_device(ti
, p
->path
.dev
);
652 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
654 dm_put_device(ti
, p
->path
.dev
);
665 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
668 static struct dm_arg _args
[] = {
669 {1, 1024, "invalid number of paths"},
670 {0, 1024, "invalid number of selector args"}
674 unsigned i
, nr_selector_args
, nr_args
;
675 struct priority_group
*pg
;
676 struct dm_target
*ti
= m
->ti
;
680 ti
->error
= "not enough priority group arguments";
681 return ERR_PTR(-EINVAL
);
684 pg
= alloc_priority_group();
686 ti
->error
= "couldn't allocate priority group";
687 return ERR_PTR(-ENOMEM
);
691 r
= parse_path_selector(as
, pg
, ti
);
698 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
702 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
706 nr_args
= 1 + nr_selector_args
;
707 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
708 struct pgpath
*pgpath
;
709 struct dm_arg_set path_args
;
711 if (as
->argc
< nr_args
) {
712 ti
->error
= "not enough path parameters";
717 path_args
.argc
= nr_args
;
718 path_args
.argv
= as
->argv
;
720 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
721 if (IS_ERR(pgpath
)) {
727 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
728 dm_consume_args(as
, nr_args
);
734 free_priority_group(pg
, ti
);
738 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
742 struct dm_target
*ti
= m
->ti
;
744 static struct dm_arg _args
[] = {
745 {0, 1024, "invalid number of hardware handler args"},
748 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
754 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
760 for (i
= 0; i
<= hw_argc
- 2; i
++)
761 len
+= strlen(as
->argv
[i
]) + 1;
762 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
764 ti
->error
= "memory allocation failed";
768 j
= sprintf(p
, "%d", hw_argc
- 1);
769 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
770 j
= sprintf(p
, "%s", as
->argv
[i
]);
772 dm_consume_args(as
, hw_argc
- 1);
776 kfree(m
->hw_handler_name
);
777 m
->hw_handler_name
= NULL
;
781 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
785 struct dm_target
*ti
= m
->ti
;
786 const char *arg_name
;
788 static struct dm_arg _args
[] = {
789 {0, 6, "invalid number of feature args"},
790 {1, 50, "pg_init_retries must be between 1 and 50"},
791 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
794 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
802 arg_name
= dm_shift_arg(as
);
805 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
806 r
= queue_if_no_path(m
, true, false);
810 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
811 m
->retain_attached_hw_handler
= true;
815 if (!strcasecmp(arg_name
, "pg_init_retries") &&
817 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
822 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
824 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
829 ti
->error
= "Unrecognised multipath feature request";
831 } while (argc
&& !r
);
836 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
839 /* target arguments */
840 static struct dm_arg _args
[] = {
841 {0, 1024, "invalid number of priority groups"},
842 {0, 1024, "invalid initial priority group number"},
847 struct dm_arg_set as
;
848 unsigned pg_count
= 0;
849 unsigned next_pg_num
;
850 bool use_blk_mq
= dm_use_blk_mq(dm_table_get_md(ti
->table
));
855 m
= alloc_multipath(ti
, use_blk_mq
);
857 ti
->error
= "can't allocate multipath";
861 r
= parse_features(&as
, m
);
865 r
= parse_hw_handler(&as
, m
);
869 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
873 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
877 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
878 (m
->nr_priority_groups
&& !next_pg_num
)) {
879 ti
->error
= "invalid initial priority group";
884 /* parse the priority groups */
886 struct priority_group
*pg
;
888 pg
= parse_priority_group(&as
, m
);
894 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
895 list_add_tail(&pg
->list
, &m
->priority_groups
);
897 pg
->pg_num
= pg_count
;
902 if (pg_count
!= m
->nr_priority_groups
) {
903 ti
->error
= "priority group count mismatch";
908 ti
->num_flush_bios
= 1;
909 ti
->num_discard_bios
= 1;
910 ti
->num_write_same_bios
= 1;
912 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
921 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
923 DECLARE_WAITQUEUE(wait
, current
);
926 add_wait_queue(&m
->pg_init_wait
, &wait
);
929 set_current_state(TASK_UNINTERRUPTIBLE
);
931 spin_lock_irqsave(&m
->lock
, flags
);
932 if (!m
->pg_init_in_progress
) {
933 spin_unlock_irqrestore(&m
->lock
, flags
);
936 spin_unlock_irqrestore(&m
->lock
, flags
);
940 set_current_state(TASK_RUNNING
);
942 remove_wait_queue(&m
->pg_init_wait
, &wait
);
945 static void flush_multipath_work(struct multipath
*m
)
949 spin_lock_irqsave(&m
->lock
, flags
);
950 m
->pg_init_disabled
= true;
951 spin_unlock_irqrestore(&m
->lock
, flags
);
953 flush_workqueue(kmpath_handlerd
);
954 multipath_wait_for_pg_init_completion(m
);
955 flush_workqueue(kmultipathd
);
956 flush_work(&m
->trigger_event
);
958 spin_lock_irqsave(&m
->lock
, flags
);
959 m
->pg_init_disabled
= false;
960 spin_unlock_irqrestore(&m
->lock
, flags
);
963 static void multipath_dtr(struct dm_target
*ti
)
965 struct multipath
*m
= ti
->private;
967 flush_multipath_work(m
);
972 * Take a path out of use.
974 static int fail_path(struct pgpath
*pgpath
)
977 struct multipath
*m
= pgpath
->pg
->m
;
979 spin_lock_irqsave(&m
->lock
, flags
);
981 if (!pgpath
->is_active
)
984 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
986 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
987 pgpath
->is_active
= false;
988 pgpath
->fail_count
++;
992 if (pgpath
== m
->current_pgpath
)
993 m
->current_pgpath
= NULL
;
995 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
996 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
998 schedule_work(&m
->trigger_event
);
1001 spin_unlock_irqrestore(&m
->lock
, flags
);
1007 * Reinstate a previously-failed path
1009 static int reinstate_path(struct pgpath
*pgpath
)
1011 int r
= 0, run_queue
= 0;
1012 unsigned long flags
;
1013 struct multipath
*m
= pgpath
->pg
->m
;
1015 spin_lock_irqsave(&m
->lock
, flags
);
1017 if (pgpath
->is_active
)
1020 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1022 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1026 pgpath
->is_active
= true;
1028 if (!m
->nr_valid_paths
++) {
1029 m
->current_pgpath
= NULL
;
1031 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1032 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1033 m
->pg_init_in_progress
++;
1036 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1037 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1039 schedule_work(&m
->trigger_event
);
1042 spin_unlock_irqrestore(&m
->lock
, flags
);
1044 dm_table_run_md_queue_async(m
->ti
->table
);
1050 * Fail or reinstate all paths that match the provided struct dm_dev.
1052 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1056 struct pgpath
*pgpath
;
1057 struct priority_group
*pg
;
1059 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1060 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1061 if (pgpath
->path
.dev
== dev
)
1070 * Temporarily try to avoid having to use the specified PG
1072 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1075 unsigned long flags
;
1077 spin_lock_irqsave(&m
->lock
, flags
);
1079 pg
->bypassed
= bypassed
;
1080 m
->current_pgpath
= NULL
;
1081 m
->current_pg
= NULL
;
1083 spin_unlock_irqrestore(&m
->lock
, flags
);
1085 schedule_work(&m
->trigger_event
);
1089 * Switch to using the specified PG from the next I/O that gets mapped
1091 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1093 struct priority_group
*pg
;
1095 unsigned long flags
;
1098 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1099 (pgnum
> m
->nr_priority_groups
)) {
1100 DMWARN("invalid PG number supplied to switch_pg_num");
1104 spin_lock_irqsave(&m
->lock
, flags
);
1105 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1106 pg
->bypassed
= false;
1110 m
->current_pgpath
= NULL
;
1111 m
->current_pg
= NULL
;
1114 spin_unlock_irqrestore(&m
->lock
, flags
);
1116 schedule_work(&m
->trigger_event
);
1121 * Set/clear bypassed status of a PG.
1122 * PGs are numbered upwards from 1 in the order they were declared.
1124 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1126 struct priority_group
*pg
;
1130 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1131 (pgnum
> m
->nr_priority_groups
)) {
1132 DMWARN("invalid PG number supplied to bypass_pg");
1136 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1141 bypass_pg(m
, pg
, bypassed
);
1146 * Should we retry pg_init immediately?
1148 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1150 unsigned long flags
;
1151 bool limit_reached
= false;
1153 spin_lock_irqsave(&m
->lock
, flags
);
1155 if (m
->pg_init_count
<= m
->pg_init_retries
&& !m
->pg_init_disabled
)
1156 m
->pg_init_required
= true;
1158 limit_reached
= true;
1160 spin_unlock_irqrestore(&m
->lock
, flags
);
1162 return limit_reached
;
1165 static void pg_init_done(void *data
, int errors
)
1167 struct pgpath
*pgpath
= data
;
1168 struct priority_group
*pg
= pgpath
->pg
;
1169 struct multipath
*m
= pg
->m
;
1170 unsigned long flags
;
1171 bool delay_retry
= false;
1173 /* device or driver problems */
1178 if (!m
->hw_handler_name
) {
1182 DMERR("Could not failover the device: Handler scsi_dh_%s "
1183 "Error %d.", m
->hw_handler_name
, errors
);
1185 * Fail path for now, so we do not ping pong
1189 case SCSI_DH_DEV_TEMP_BUSY
:
1191 * Probably doing something like FW upgrade on the
1192 * controller so try the other pg.
1194 bypass_pg(m
, pg
, true);
1197 /* Wait before retrying. */
1199 case SCSI_DH_IMM_RETRY
:
1200 case SCSI_DH_RES_TEMP_UNAVAIL
:
1201 if (pg_init_limit_reached(m
, pgpath
))
1205 case SCSI_DH_DEV_OFFLINED
:
1208 * We probably do not want to fail the path for a device
1209 * error, but this is what the old dm did. In future
1210 * patches we can do more advanced handling.
1215 spin_lock_irqsave(&m
->lock
, flags
);
1217 if (pgpath
== m
->current_pgpath
) {
1218 DMERR("Could not failover device. Error %d.", errors
);
1219 m
->current_pgpath
= NULL
;
1220 m
->current_pg
= NULL
;
1222 } else if (!m
->pg_init_required
)
1223 pg
->bypassed
= false;
1225 if (--m
->pg_init_in_progress
)
1226 /* Activations of other paths are still on going */
1229 if (m
->pg_init_required
) {
1230 m
->pg_init_delay_retry
= delay_retry
;
1231 if (__pg_init_all_paths(m
))
1234 m
->queue_io
= false;
1237 * Wake up any thread waiting to suspend.
1239 wake_up(&m
->pg_init_wait
);
1242 spin_unlock_irqrestore(&m
->lock
, flags
);
1245 static void activate_path(struct work_struct
*work
)
1247 struct pgpath
*pgpath
=
1248 container_of(work
, struct pgpath
, activate_path
.work
);
1250 if (pgpath
->is_active
)
1251 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1252 pg_init_done
, pgpath
);
1254 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1257 static int noretry_error(int error
)
1268 /* Anything else could be a path failure, so should be retried */
1275 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1276 int error
, struct dm_mpath_io
*mpio
)
1279 * We don't queue any clone request inside the multipath target
1280 * during end I/O handling, since those clone requests don't have
1281 * bio clones. If we queue them inside the multipath target,
1282 * we need to make bio clones, that requires memory allocation.
1283 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1284 * don't have bio clones.)
1285 * Instead of queueing the clone request here, we queue the original
1286 * request into dm core, which will remake a clone request and
1287 * clone bios for it and resubmit it later.
1289 int r
= DM_ENDIO_REQUEUE
;
1290 unsigned long flags
;
1292 if (!error
&& !clone
->errors
)
1293 return 0; /* I/O complete */
1295 if (noretry_error(error
))
1299 fail_path(mpio
->pgpath
);
1301 spin_lock_irqsave(&m
->lock
, flags
);
1302 if (!m
->nr_valid_paths
) {
1303 if (!m
->queue_if_no_path
) {
1304 if (!__must_push_back(m
))
1307 if (error
== -EBADE
)
1311 spin_unlock_irqrestore(&m
->lock
, flags
);
1316 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1317 int error
, union map_info
*map_context
)
1319 struct multipath
*m
= ti
->private;
1320 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1321 struct pgpath
*pgpath
;
1322 struct path_selector
*ps
;
1327 r
= do_end_io(m
, clone
, error
, mpio
);
1328 pgpath
= mpio
->pgpath
;
1330 ps
= &pgpath
->pg
->ps
;
1331 if (ps
->type
->end_io
)
1332 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1334 clear_request_fn_mpio(m
, map_context
);
1340 * Suspend can't complete until all the I/O is processed so if
1341 * the last path fails we must error any remaining I/O.
1342 * Note that if the freeze_bdev fails while suspending, the
1343 * queue_if_no_path state is lost - userspace should reset it.
1345 static void multipath_presuspend(struct dm_target
*ti
)
1347 struct multipath
*m
= ti
->private;
1349 queue_if_no_path(m
, false, true);
1352 static void multipath_postsuspend(struct dm_target
*ti
)
1354 struct multipath
*m
= ti
->private;
1356 mutex_lock(&m
->work_mutex
);
1357 flush_multipath_work(m
);
1358 mutex_unlock(&m
->work_mutex
);
1362 * Restore the queue_if_no_path setting.
1364 static void multipath_resume(struct dm_target
*ti
)
1366 struct multipath
*m
= ti
->private;
1367 unsigned long flags
;
1369 spin_lock_irqsave(&m
->lock
, flags
);
1370 m
->queue_if_no_path
= m
->saved_queue_if_no_path
;
1371 spin_unlock_irqrestore(&m
->lock
, flags
);
1375 * Info output has the following format:
1376 * num_multipath_feature_args [multipath_feature_args]*
1377 * num_handler_status_args [handler_status_args]*
1378 * num_groups init_group_number
1379 * [A|D|E num_ps_status_args [ps_status_args]*
1380 * num_paths num_selector_args
1381 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1383 * Table output has the following format (identical to the constructor string):
1384 * num_feature_args [features_args]*
1385 * num_handler_args hw_handler [hw_handler_args]*
1386 * num_groups init_group_number
1387 * [priority selector-name num_ps_args [ps_args]*
1388 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1390 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1391 unsigned status_flags
, char *result
, unsigned maxlen
)
1394 unsigned long flags
;
1395 struct multipath
*m
= ti
->private;
1396 struct priority_group
*pg
;
1401 spin_lock_irqsave(&m
->lock
, flags
);
1404 if (type
== STATUSTYPE_INFO
)
1405 DMEMIT("2 %u %u ", m
->queue_io
, m
->pg_init_count
);
1407 DMEMIT("%u ", m
->queue_if_no_path
+
1408 (m
->pg_init_retries
> 0) * 2 +
1409 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1410 m
->retain_attached_hw_handler
);
1411 if (m
->queue_if_no_path
)
1412 DMEMIT("queue_if_no_path ");
1413 if (m
->pg_init_retries
)
1414 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1415 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1416 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1417 if (m
->retain_attached_hw_handler
)
1418 DMEMIT("retain_attached_hw_handler ");
1421 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1424 DMEMIT("1 %s ", m
->hw_handler_name
);
1426 DMEMIT("%u ", m
->nr_priority_groups
);
1429 pg_num
= m
->next_pg
->pg_num
;
1430 else if (m
->current_pg
)
1431 pg_num
= m
->current_pg
->pg_num
;
1433 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1435 DMEMIT("%u ", pg_num
);
1438 case STATUSTYPE_INFO
:
1439 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1441 state
= 'D'; /* Disabled */
1442 else if (pg
== m
->current_pg
)
1443 state
= 'A'; /* Currently Active */
1445 state
= 'E'; /* Enabled */
1447 DMEMIT("%c ", state
);
1449 if (pg
->ps
.type
->status
)
1450 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1456 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1457 pg
->ps
.type
->info_args
);
1459 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1460 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1461 p
->is_active
? "A" : "F",
1463 if (pg
->ps
.type
->status
)
1464 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1465 &p
->path
, type
, result
+ sz
,
1471 case STATUSTYPE_TABLE
:
1472 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1473 DMEMIT("%s ", pg
->ps
.type
->name
);
1475 if (pg
->ps
.type
->status
)
1476 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1482 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1483 pg
->ps
.type
->table_args
);
1485 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1486 DMEMIT("%s ", p
->path
.dev
->name
);
1487 if (pg
->ps
.type
->status
)
1488 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1489 &p
->path
, type
, result
+ sz
,
1496 spin_unlock_irqrestore(&m
->lock
, flags
);
1499 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1503 struct multipath
*m
= ti
->private;
1506 mutex_lock(&m
->work_mutex
);
1508 if (dm_suspended(ti
)) {
1514 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1515 r
= queue_if_no_path(m
, true, false);
1517 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1518 r
= queue_if_no_path(m
, false, false);
1524 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1528 if (!strcasecmp(argv
[0], "disable_group")) {
1529 r
= bypass_pg_num(m
, argv
[1], true);
1531 } else if (!strcasecmp(argv
[0], "enable_group")) {
1532 r
= bypass_pg_num(m
, argv
[1], false);
1534 } else if (!strcasecmp(argv
[0], "switch_group")) {
1535 r
= switch_pg_num(m
, argv
[1]);
1537 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1538 action
= reinstate_path
;
1539 else if (!strcasecmp(argv
[0], "fail_path"))
1542 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1546 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1548 DMWARN("message: error getting device %s",
1553 r
= action_dev(m
, dev
, action
);
1555 dm_put_device(ti
, dev
);
1558 mutex_unlock(&m
->work_mutex
);
1562 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1563 struct block_device
**bdev
, fmode_t
*mode
)
1565 struct multipath
*m
= ti
->private;
1566 unsigned long flags
;
1569 spin_lock_irqsave(&m
->lock
, flags
);
1571 if (!m
->current_pgpath
)
1572 __choose_pgpath(m
, 0);
1574 if (m
->current_pgpath
) {
1576 *bdev
= m
->current_pgpath
->path
.dev
->bdev
;
1577 *mode
= m
->current_pgpath
->path
.dev
->mode
;
1580 /* pg_init has not started or completed */
1584 /* No path is available */
1585 if (m
->queue_if_no_path
)
1591 spin_unlock_irqrestore(&m
->lock
, flags
);
1593 if (r
== -ENOTCONN
) {
1594 spin_lock_irqsave(&m
->lock
, flags
);
1595 if (!m
->current_pg
) {
1596 /* Path status changed, redo selection */
1597 __choose_pgpath(m
, 0);
1599 if (m
->pg_init_required
)
1600 __pg_init_all_paths(m
);
1601 spin_unlock_irqrestore(&m
->lock
, flags
);
1602 dm_table_run_md_queue_async(m
->ti
->table
);
1606 * Only pass ioctls through if the device sizes match exactly.
1608 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1613 static int multipath_iterate_devices(struct dm_target
*ti
,
1614 iterate_devices_callout_fn fn
, void *data
)
1616 struct multipath
*m
= ti
->private;
1617 struct priority_group
*pg
;
1621 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1622 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1623 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1633 static int pgpath_busy(struct pgpath
*pgpath
)
1635 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1637 return blk_lld_busy(q
);
1641 * We return "busy", only when we can map I/Os but underlying devices
1642 * are busy (so even if we map I/Os now, the I/Os will wait on
1643 * the underlying queue).
1644 * In other words, if we want to kill I/Os or queue them inside us
1645 * due to map unavailability, we don't return "busy". Otherwise,
1646 * dm core won't give us the I/Os and we can't do what we want.
1648 static int multipath_busy(struct dm_target
*ti
)
1650 bool busy
= false, has_active
= false;
1651 struct multipath
*m
= ti
->private;
1652 struct priority_group
*pg
;
1653 struct pgpath
*pgpath
;
1654 unsigned long flags
;
1656 spin_lock_irqsave(&m
->lock
, flags
);
1658 /* pg_init in progress or no paths available */
1659 if (m
->pg_init_in_progress
||
1660 (!m
->nr_valid_paths
&& m
->queue_if_no_path
)) {
1664 /* Guess which priority_group will be used at next mapping time */
1665 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1667 else if (likely(m
->current_pg
))
1671 * We don't know which pg will be used at next mapping time.
1672 * We don't call __choose_pgpath() here to avoid to trigger
1673 * pg_init just by busy checking.
1674 * So we don't know whether underlying devices we will be using
1675 * at next mapping time are busy or not. Just try mapping.
1680 * If there is one non-busy active path at least, the path selector
1681 * will be able to select it. So we consider such a pg as not busy.
1684 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1685 if (pgpath
->is_active
) {
1687 if (!pgpath_busy(pgpath
)) {
1695 * No active path in this pg, so this pg won't be used and
1696 * the current_pg will be changed at next mapping time.
1697 * We need to try mapping to determine it.
1702 spin_unlock_irqrestore(&m
->lock
, flags
);
1707 /*-----------------------------------------------------------------
1709 *---------------------------------------------------------------*/
1710 static struct target_type multipath_target
= {
1711 .name
= "multipath",
1712 .version
= {1, 11, 0},
1713 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1714 .module
= THIS_MODULE
,
1715 .ctr
= multipath_ctr
,
1716 .dtr
= multipath_dtr
,
1717 .map_rq
= multipath_map
,
1718 .clone_and_map_rq
= multipath_clone_and_map
,
1719 .release_clone_rq
= multipath_release_clone
,
1720 .rq_end_io
= multipath_end_io
,
1721 .presuspend
= multipath_presuspend
,
1722 .postsuspend
= multipath_postsuspend
,
1723 .resume
= multipath_resume
,
1724 .status
= multipath_status
,
1725 .message
= multipath_message
,
1726 .prepare_ioctl
= multipath_prepare_ioctl
,
1727 .iterate_devices
= multipath_iterate_devices
,
1728 .busy
= multipath_busy
,
1731 static int __init
dm_multipath_init(void)
1735 /* allocate a slab for the dm_ios */
1736 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1740 r
= dm_register_target(&multipath_target
);
1742 DMERR("register failed %d", r
);
1744 goto bad_register_target
;
1747 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1749 DMERR("failed to create workqueue kmpathd");
1751 goto bad_alloc_kmultipathd
;
1755 * A separate workqueue is used to handle the device handlers
1756 * to avoid overloading existing workqueue. Overloading the
1757 * old workqueue would also create a bottleneck in the
1758 * path of the storage hardware device activation.
1760 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1762 if (!kmpath_handlerd
) {
1763 DMERR("failed to create workqueue kmpath_handlerd");
1765 goto bad_alloc_kmpath_handlerd
;
1768 DMINFO("version %u.%u.%u loaded",
1769 multipath_target
.version
[0], multipath_target
.version
[1],
1770 multipath_target
.version
[2]);
1774 bad_alloc_kmpath_handlerd
:
1775 destroy_workqueue(kmultipathd
);
1776 bad_alloc_kmultipathd
:
1777 dm_unregister_target(&multipath_target
);
1778 bad_register_target
:
1779 kmem_cache_destroy(_mpio_cache
);
1784 static void __exit
dm_multipath_exit(void)
1786 destroy_workqueue(kmpath_handlerd
);
1787 destroy_workqueue(kmultipathd
);
1789 dm_unregister_target(&multipath_target
);
1790 kmem_cache_destroy(_mpio_cache
);
1793 module_init(dm_multipath_init
);
1794 module_exit(dm_multipath_exit
);
1796 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1797 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1798 MODULE_LICENSE("GPL");