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)
26 #define DM_PG_INIT_DELAY_MSECS 2000
27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
31 struct list_head list
;
33 struct priority_group
*pg
; /* Owning PG */
34 unsigned is_active
; /* Path status */
35 unsigned fail_count
; /* Cumulative failure count */
38 struct delayed_work activate_path
;
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
44 * Paths are grouped into Priority Groups and numbered from 1 upwards.
45 * Each has a path selector which controls which path gets used.
47 struct priority_group
{
48 struct list_head list
;
50 struct multipath
*m
; /* Owning multipath instance */
51 struct path_selector ps
;
53 unsigned pg_num
; /* Reference number */
54 unsigned bypassed
; /* Temporarily bypass this PG? */
56 unsigned nr_pgpaths
; /* Number of paths in PG */
57 struct list_head pgpaths
;
60 /* Multipath context */
62 struct list_head list
;
67 const char *hw_handler_name
;
68 char *hw_handler_params
;
70 unsigned nr_priority_groups
;
71 struct list_head priority_groups
;
73 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
75 unsigned pg_init_required
; /* pg_init needs calling? */
76 unsigned pg_init_in_progress
; /* Only one pg_init allowed at once */
77 unsigned pg_init_delay_retry
; /* Delay pg_init retry? */
79 unsigned nr_valid_paths
; /* Total number of usable paths */
80 struct pgpath
*current_pgpath
;
81 struct priority_group
*current_pg
;
82 struct priority_group
*next_pg
; /* Switch to this PG if set */
83 unsigned repeat_count
; /* I/Os left before calling PS again */
85 unsigned queue_io
; /* Must we queue all I/O? */
86 unsigned queue_if_no_path
; /* Queue I/O if last path fails? */
87 unsigned saved_queue_if_no_path
;/* Saved state during suspension */
88 unsigned pg_init_retries
; /* Number of times to retry pg_init */
89 unsigned pg_init_count
; /* Number of times pg_init called */
90 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
92 struct work_struct process_queued_ios
;
93 struct list_head queued_ios
;
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 #define MIN_IOS 256 /* Mempool size */
119 static struct kmem_cache
*_mpio_cache
;
121 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
122 static void process_queued_ios(struct work_struct
*work
);
123 static void trigger_event(struct work_struct
*work
);
124 static void activate_path(struct work_struct
*work
);
127 /*-----------------------------------------------
128 * Allocation routines
129 *-----------------------------------------------*/
131 static struct pgpath
*alloc_pgpath(void)
133 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
136 pgpath
->is_active
= 1;
137 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path
);
143 static void free_pgpath(struct pgpath
*pgpath
)
148 static struct priority_group
*alloc_priority_group(void)
150 struct priority_group
*pg
;
152 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
155 INIT_LIST_HEAD(&pg
->pgpaths
);
160 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
162 struct pgpath
*pgpath
, *tmp
;
163 struct multipath
*m
= ti
->private;
165 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
166 list_del(&pgpath
->list
);
167 if (m
->hw_handler_name
)
168 scsi_dh_detach(bdev_get_queue(pgpath
->path
.dev
->bdev
));
169 dm_put_device(ti
, pgpath
->path
.dev
);
174 static void free_priority_group(struct priority_group
*pg
,
175 struct dm_target
*ti
)
177 struct path_selector
*ps
= &pg
->ps
;
180 ps
->type
->destroy(ps
);
181 dm_put_path_selector(ps
->type
);
184 free_pgpaths(&pg
->pgpaths
, ti
);
188 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
192 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
194 INIT_LIST_HEAD(&m
->priority_groups
);
195 INIT_LIST_HEAD(&m
->queued_ios
);
196 spin_lock_init(&m
->lock
);
198 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
199 INIT_WORK(&m
->process_queued_ios
, process_queued_ios
);
200 INIT_WORK(&m
->trigger_event
, trigger_event
);
201 init_waitqueue_head(&m
->pg_init_wait
);
202 mutex_init(&m
->work_mutex
);
203 m
->mpio_pool
= mempool_create_slab_pool(MIN_IOS
, _mpio_cache
);
215 static void free_multipath(struct multipath
*m
)
217 struct priority_group
*pg
, *tmp
;
219 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
221 free_priority_group(pg
, m
->ti
);
224 kfree(m
->hw_handler_name
);
225 kfree(m
->hw_handler_params
);
226 mempool_destroy(m
->mpio_pool
);
231 /*-----------------------------------------------
233 *-----------------------------------------------*/
235 static void __pg_init_all_paths(struct multipath
*m
)
237 struct pgpath
*pgpath
;
238 unsigned long pg_init_delay
= 0;
241 m
->pg_init_required
= 0;
242 if (m
->pg_init_delay_retry
)
243 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
244 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
245 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
246 /* Skip failed paths */
247 if (!pgpath
->is_active
)
249 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
251 m
->pg_init_in_progress
++;
255 static void __switch_pg(struct multipath
*m
, struct pgpath
*pgpath
)
257 m
->current_pg
= pgpath
->pg
;
259 /* Must we initialise the PG first, and queue I/O till it's ready? */
260 if (m
->hw_handler_name
) {
261 m
->pg_init_required
= 1;
264 m
->pg_init_required
= 0;
268 m
->pg_init_count
= 0;
271 static int __choose_path_in_pg(struct multipath
*m
, struct priority_group
*pg
,
274 struct dm_path
*path
;
276 path
= pg
->ps
.type
->select_path(&pg
->ps
, &m
->repeat_count
, nr_bytes
);
280 m
->current_pgpath
= path_to_pgpath(path
);
282 if (m
->current_pg
!= pg
)
283 __switch_pg(m
, m
->current_pgpath
);
288 static void __choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
290 struct priority_group
*pg
;
291 unsigned bypassed
= 1;
293 if (!m
->nr_valid_paths
)
296 /* Were we instructed to switch PG? */
300 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
304 /* Don't change PG until it has no remaining paths */
305 if (m
->current_pg
&& !__choose_path_in_pg(m
, m
->current_pg
, nr_bytes
))
309 * Loop through priority groups until we find a valid path.
310 * First time we skip PGs marked 'bypassed'.
311 * Second time we only try the ones we skipped.
314 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
315 if (pg
->bypassed
== bypassed
)
317 if (!__choose_path_in_pg(m
, pg
, nr_bytes
))
320 } while (bypassed
--);
323 m
->current_pgpath
= NULL
;
324 m
->current_pg
= NULL
;
328 * Check whether bios must be queued in the device-mapper core rather
329 * than here in the target.
331 * m->lock must be held on entry.
333 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
334 * same value then we are not between multipath_presuspend()
335 * and multipath_resume() calls and we have no need to check
336 * for the DMF_NOFLUSH_SUSPENDING flag.
338 static int __must_push_back(struct multipath
*m
)
340 return (m
->queue_if_no_path
!= m
->saved_queue_if_no_path
&&
341 dm_noflush_suspending(m
->ti
));
344 static int map_io(struct multipath
*m
, struct request
*clone
,
345 struct dm_mpath_io
*mpio
, unsigned was_queued
)
347 int r
= DM_MAPIO_REMAPPED
;
348 size_t nr_bytes
= blk_rq_bytes(clone
);
350 struct pgpath
*pgpath
;
351 struct block_device
*bdev
;
353 spin_lock_irqsave(&m
->lock
, flags
);
355 /* Do we need to select a new pgpath? */
356 if (!m
->current_pgpath
||
357 (!m
->queue_io
&& (m
->repeat_count
&& --m
->repeat_count
== 0)))
358 __choose_pgpath(m
, nr_bytes
);
360 pgpath
= m
->current_pgpath
;
365 if ((pgpath
&& m
->queue_io
) ||
366 (!pgpath
&& m
->queue_if_no_path
)) {
367 /* Queue for the daemon to resubmit */
368 list_add_tail(&clone
->queuelist
, &m
->queued_ios
);
370 if ((m
->pg_init_required
&& !m
->pg_init_in_progress
) ||
372 queue_work(kmultipathd
, &m
->process_queued_ios
);
374 r
= DM_MAPIO_SUBMITTED
;
376 bdev
= pgpath
->path
.dev
->bdev
;
377 clone
->q
= bdev_get_queue(bdev
);
378 clone
->rq_disk
= bdev
->bd_disk
;
379 } else if (__must_push_back(m
))
380 r
= DM_MAPIO_REQUEUE
;
382 r
= -EIO
; /* Failed */
384 mpio
->pgpath
= pgpath
;
385 mpio
->nr_bytes
= nr_bytes
;
387 if (r
== DM_MAPIO_REMAPPED
&& pgpath
->pg
->ps
.type
->start_io
)
388 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
, &pgpath
->path
,
391 spin_unlock_irqrestore(&m
->lock
, flags
);
397 * If we run out of usable paths, should we queue I/O or error it?
399 static int queue_if_no_path(struct multipath
*m
, unsigned queue_if_no_path
,
400 unsigned save_old_value
)
404 spin_lock_irqsave(&m
->lock
, flags
);
407 m
->saved_queue_if_no_path
= m
->queue_if_no_path
;
409 m
->saved_queue_if_no_path
= queue_if_no_path
;
410 m
->queue_if_no_path
= queue_if_no_path
;
411 if (!m
->queue_if_no_path
&& m
->queue_size
)
412 queue_work(kmultipathd
, &m
->process_queued_ios
);
414 spin_unlock_irqrestore(&m
->lock
, flags
);
419 /*-----------------------------------------------------------------
420 * The multipath daemon is responsible for resubmitting queued ios.
421 *---------------------------------------------------------------*/
423 static void dispatch_queued_ios(struct multipath
*m
)
427 struct dm_mpath_io
*mpio
;
428 union map_info
*info
;
429 struct request
*clone
, *n
;
432 spin_lock_irqsave(&m
->lock
, flags
);
433 list_splice_init(&m
->queued_ios
, &cl
);
434 spin_unlock_irqrestore(&m
->lock
, flags
);
436 list_for_each_entry_safe(clone
, n
, &cl
, queuelist
) {
437 list_del_init(&clone
->queuelist
);
439 info
= dm_get_rq_mapinfo(clone
);
442 r
= map_io(m
, clone
, mpio
, 1);
444 mempool_free(mpio
, m
->mpio_pool
);
445 dm_kill_unmapped_request(clone
, r
);
446 } else if (r
== DM_MAPIO_REMAPPED
)
447 dm_dispatch_request(clone
);
448 else if (r
== DM_MAPIO_REQUEUE
) {
449 mempool_free(mpio
, m
->mpio_pool
);
450 dm_requeue_unmapped_request(clone
);
455 static void process_queued_ios(struct work_struct
*work
)
457 struct multipath
*m
=
458 container_of(work
, struct multipath
, process_queued_ios
);
459 struct pgpath
*pgpath
= NULL
;
460 unsigned must_queue
= 1;
463 spin_lock_irqsave(&m
->lock
, flags
);
468 if (!m
->current_pgpath
)
469 __choose_pgpath(m
, 0);
471 pgpath
= m
->current_pgpath
;
473 if ((pgpath
&& !m
->queue_io
) ||
474 (!pgpath
&& !m
->queue_if_no_path
))
477 if (m
->pg_init_required
&& !m
->pg_init_in_progress
&& pgpath
)
478 __pg_init_all_paths(m
);
481 spin_unlock_irqrestore(&m
->lock
, flags
);
483 dispatch_queued_ios(m
);
487 * An event is triggered whenever a path is taken out of use.
488 * Includes path failure and PG bypass.
490 static void trigger_event(struct work_struct
*work
)
492 struct multipath
*m
=
493 container_of(work
, struct multipath
, trigger_event
);
495 dm_table_event(m
->ti
->table
);
498 /*-----------------------------------------------------------------
499 * Constructor/argument parsing:
500 * <#multipath feature args> [<arg>]*
501 * <#hw_handler args> [hw_handler [<arg>]*]
503 * <initial priority group>
504 * [<selector> <#selector args> [<arg>]*
505 * <#paths> <#per-path selector args>
506 * [<path> [<arg>]* ]+ ]+
507 *---------------------------------------------------------------*/
514 static int read_param(struct param
*param
, char *str
, unsigned *v
, char **error
)
517 (sscanf(str
, "%u", v
) != 1) ||
520 *error
= param
->error
;
532 static char *shift(struct arg_set
*as
)
546 static void consume(struct arg_set
*as
, unsigned n
)
548 BUG_ON (as
->argc
< n
);
553 static int parse_path_selector(struct arg_set
*as
, struct priority_group
*pg
,
554 struct dm_target
*ti
)
557 struct path_selector_type
*pst
;
560 static struct param _params
[] = {
561 {0, 1024, "invalid number of path selector args"},
564 pst
= dm_get_path_selector(shift(as
));
566 ti
->error
= "unknown path selector type";
570 r
= read_param(_params
, shift(as
), &ps_argc
, &ti
->error
);
572 dm_put_path_selector(pst
);
576 if (ps_argc
> as
->argc
) {
577 dm_put_path_selector(pst
);
578 ti
->error
= "not enough arguments for path selector";
582 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
584 dm_put_path_selector(pst
);
585 ti
->error
= "path selector constructor failed";
590 consume(as
, ps_argc
);
595 static struct pgpath
*parse_path(struct arg_set
*as
, struct path_selector
*ps
,
596 struct dm_target
*ti
)
600 struct multipath
*m
= ti
->private;
602 /* we need at least a path arg */
604 ti
->error
= "no device given";
605 return ERR_PTR(-EINVAL
);
610 return ERR_PTR(-ENOMEM
);
612 r
= dm_get_device(ti
, shift(as
), dm_table_get_mode(ti
->table
),
615 ti
->error
= "error getting device";
619 if (m
->hw_handler_name
) {
620 struct request_queue
*q
= bdev_get_queue(p
->path
.dev
->bdev
);
622 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
625 * Already attached to different hw_handler,
626 * try to reattach with correct one.
629 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
633 ti
->error
= "error attaching hardware handler";
634 dm_put_device(ti
, p
->path
.dev
);
638 if (m
->hw_handler_params
) {
639 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
641 ti
->error
= "unable to set hardware "
642 "handler parameters";
644 dm_put_device(ti
, p
->path
.dev
);
650 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
652 dm_put_device(ti
, p
->path
.dev
);
663 static struct priority_group
*parse_priority_group(struct arg_set
*as
,
666 static struct param _params
[] = {
667 {1, 1024, "invalid number of paths"},
668 {0, 1024, "invalid number of selector args"}
672 unsigned i
, nr_selector_args
, nr_params
;
673 struct priority_group
*pg
;
674 struct dm_target
*ti
= m
->ti
;
678 ti
->error
= "not enough priority group arguments";
679 return ERR_PTR(-EINVAL
);
682 pg
= alloc_priority_group();
684 ti
->error
= "couldn't allocate priority group";
685 return ERR_PTR(-ENOMEM
);
689 r
= parse_path_selector(as
, pg
, ti
);
696 r
= read_param(_params
, shift(as
), &pg
->nr_pgpaths
, &ti
->error
);
700 r
= read_param(_params
+ 1, shift(as
), &nr_selector_args
, &ti
->error
);
704 nr_params
= 1 + nr_selector_args
;
705 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
706 struct pgpath
*pgpath
;
707 struct arg_set path_args
;
709 if (as
->argc
< nr_params
) {
710 ti
->error
= "not enough path parameters";
715 path_args
.argc
= nr_params
;
716 path_args
.argv
= as
->argv
;
718 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
719 if (IS_ERR(pgpath
)) {
725 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
726 consume(as
, nr_params
);
732 free_priority_group(pg
, ti
);
736 static int parse_hw_handler(struct arg_set
*as
, struct multipath
*m
)
740 struct dm_target
*ti
= m
->ti
;
742 static struct param _params
[] = {
743 {0, 1024, "invalid number of hardware handler args"},
746 if (read_param(_params
, shift(as
), &hw_argc
, &ti
->error
))
752 if (hw_argc
> as
->argc
) {
753 ti
->error
= "not enough arguments for hardware handler";
757 m
->hw_handler_name
= kstrdup(shift(as
), GFP_KERNEL
);
758 request_module("scsi_dh_%s", m
->hw_handler_name
);
759 if (scsi_dh_handler_exist(m
->hw_handler_name
) == 0) {
760 ti
->error
= "unknown hardware handler type";
769 for (i
= 0; i
<= hw_argc
- 2; i
++)
770 len
+= strlen(as
->argv
[i
]) + 1;
771 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
773 ti
->error
= "memory allocation failed";
777 j
= sprintf(p
, "%d", hw_argc
- 1);
778 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
779 j
= sprintf(p
, "%s", as
->argv
[i
]);
781 consume(as
, hw_argc
- 1);
785 kfree(m
->hw_handler_name
);
786 m
->hw_handler_name
= NULL
;
790 static int parse_features(struct arg_set
*as
, struct multipath
*m
)
794 struct dm_target
*ti
= m
->ti
;
795 const char *param_name
;
797 static struct param _params
[] = {
798 {0, 5, "invalid number of feature args"},
799 {1, 50, "pg_init_retries must be between 1 and 50"},
800 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
803 r
= read_param(_params
, shift(as
), &argc
, &ti
->error
);
810 if (argc
> as
->argc
) {
811 ti
->error
= "not enough arguments for features";
816 param_name
= shift(as
);
819 if (!strnicmp(param_name
, MESG_STR("queue_if_no_path"))) {
820 r
= queue_if_no_path(m
, 1, 0);
824 if (!strnicmp(param_name
, MESG_STR("pg_init_retries")) &&
826 r
= read_param(_params
+ 1, shift(as
),
827 &m
->pg_init_retries
, &ti
->error
);
832 if (!strnicmp(param_name
, MESG_STR("pg_init_delay_msecs")) &&
834 r
= read_param(_params
+ 2, shift(as
),
835 &m
->pg_init_delay_msecs
, &ti
->error
);
840 ti
->error
= "Unrecognised multipath feature request";
842 } while (argc
&& !r
);
847 static int multipath_ctr(struct dm_target
*ti
, unsigned int argc
,
850 /* target parameters */
851 static struct param _params
[] = {
852 {0, 1024, "invalid number of priority groups"},
853 {0, 1024, "invalid initial priority group number"},
859 unsigned pg_count
= 0;
860 unsigned next_pg_num
;
865 m
= alloc_multipath(ti
);
867 ti
->error
= "can't allocate multipath";
871 r
= parse_features(&as
, m
);
875 r
= parse_hw_handler(&as
, m
);
879 r
= read_param(_params
, shift(&as
), &m
->nr_priority_groups
, &ti
->error
);
883 r
= read_param(_params
+ 1, shift(&as
), &next_pg_num
, &ti
->error
);
887 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
888 (m
->nr_priority_groups
&& !next_pg_num
)) {
889 ti
->error
= "invalid initial priority group";
894 /* parse the priority groups */
896 struct priority_group
*pg
;
898 pg
= parse_priority_group(&as
, m
);
904 m
->nr_valid_paths
+= pg
->nr_pgpaths
;
905 list_add_tail(&pg
->list
, &m
->priority_groups
);
907 pg
->pg_num
= pg_count
;
912 if (pg_count
!= m
->nr_priority_groups
) {
913 ti
->error
= "priority group count mismatch";
918 ti
->num_flush_requests
= 1;
919 ti
->num_discard_requests
= 1;
928 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
930 DECLARE_WAITQUEUE(wait
, current
);
933 add_wait_queue(&m
->pg_init_wait
, &wait
);
936 set_current_state(TASK_UNINTERRUPTIBLE
);
938 spin_lock_irqsave(&m
->lock
, flags
);
939 if (!m
->pg_init_in_progress
) {
940 spin_unlock_irqrestore(&m
->lock
, flags
);
943 spin_unlock_irqrestore(&m
->lock
, flags
);
947 set_current_state(TASK_RUNNING
);
949 remove_wait_queue(&m
->pg_init_wait
, &wait
);
952 static void flush_multipath_work(struct multipath
*m
)
954 flush_workqueue(kmpath_handlerd
);
955 multipath_wait_for_pg_init_completion(m
);
956 flush_workqueue(kmultipathd
);
957 flush_work_sync(&m
->trigger_event
);
960 static void multipath_dtr(struct dm_target
*ti
)
962 struct multipath
*m
= ti
->private;
964 flush_multipath_work(m
);
969 * Map cloned requests
971 static int multipath_map(struct dm_target
*ti
, struct request
*clone
,
972 union map_info
*map_context
)
975 struct dm_mpath_io
*mpio
;
976 struct multipath
*m
= (struct multipath
*) ti
->private;
978 mpio
= mempool_alloc(m
->mpio_pool
, GFP_ATOMIC
);
980 /* ENOMEM, requeue */
981 return DM_MAPIO_REQUEUE
;
982 memset(mpio
, 0, sizeof(*mpio
));
984 map_context
->ptr
= mpio
;
985 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
986 r
= map_io(m
, clone
, mpio
, 0);
987 if (r
< 0 || r
== DM_MAPIO_REQUEUE
)
988 mempool_free(mpio
, m
->mpio_pool
);
994 * Take a path out of use.
996 static int fail_path(struct pgpath
*pgpath
)
999 struct multipath
*m
= pgpath
->pg
->m
;
1001 spin_lock_irqsave(&m
->lock
, flags
);
1003 if (!pgpath
->is_active
)
1006 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1008 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1009 pgpath
->is_active
= 0;
1010 pgpath
->fail_count
++;
1012 m
->nr_valid_paths
--;
1014 if (pgpath
== m
->current_pgpath
)
1015 m
->current_pgpath
= NULL
;
1017 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1018 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1020 schedule_work(&m
->trigger_event
);
1023 spin_unlock_irqrestore(&m
->lock
, flags
);
1029 * Reinstate a previously-failed path
1031 static int reinstate_path(struct pgpath
*pgpath
)
1034 unsigned long flags
;
1035 struct multipath
*m
= pgpath
->pg
->m
;
1037 spin_lock_irqsave(&m
->lock
, flags
);
1039 if (pgpath
->is_active
)
1042 if (!pgpath
->pg
->ps
.type
->reinstate_path
) {
1043 DMWARN("Reinstate path not supported by path selector %s",
1044 pgpath
->pg
->ps
.type
->name
);
1049 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1053 pgpath
->is_active
= 1;
1055 if (!m
->nr_valid_paths
++ && m
->queue_size
) {
1056 m
->current_pgpath
= NULL
;
1057 queue_work(kmultipathd
, &m
->process_queued_ios
);
1058 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1059 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1060 m
->pg_init_in_progress
++;
1063 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1064 pgpath
->path
.dev
->name
, m
->nr_valid_paths
);
1066 schedule_work(&m
->trigger_event
);
1069 spin_unlock_irqrestore(&m
->lock
, flags
);
1075 * Fail or reinstate all paths that match the provided struct dm_dev.
1077 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1081 struct pgpath
*pgpath
;
1082 struct priority_group
*pg
;
1084 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1085 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1086 if (pgpath
->path
.dev
== dev
)
1095 * Temporarily try to avoid having to use the specified PG
1097 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1100 unsigned long flags
;
1102 spin_lock_irqsave(&m
->lock
, flags
);
1104 pg
->bypassed
= bypassed
;
1105 m
->current_pgpath
= NULL
;
1106 m
->current_pg
= NULL
;
1108 spin_unlock_irqrestore(&m
->lock
, flags
);
1110 schedule_work(&m
->trigger_event
);
1114 * Switch to using the specified PG from the next I/O that gets mapped
1116 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1118 struct priority_group
*pg
;
1120 unsigned long flags
;
1122 if (!pgstr
|| (sscanf(pgstr
, "%u", &pgnum
) != 1) || !pgnum
||
1123 (pgnum
> m
->nr_priority_groups
)) {
1124 DMWARN("invalid PG number supplied to switch_pg_num");
1128 spin_lock_irqsave(&m
->lock
, flags
);
1129 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1134 m
->current_pgpath
= NULL
;
1135 m
->current_pg
= NULL
;
1138 spin_unlock_irqrestore(&m
->lock
, flags
);
1140 schedule_work(&m
->trigger_event
);
1145 * Set/clear bypassed status of a PG.
1146 * PGs are numbered upwards from 1 in the order they were declared.
1148 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, int bypassed
)
1150 struct priority_group
*pg
;
1153 if (!pgstr
|| (sscanf(pgstr
, "%u", &pgnum
) != 1) || !pgnum
||
1154 (pgnum
> m
->nr_priority_groups
)) {
1155 DMWARN("invalid PG number supplied to bypass_pg");
1159 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1164 bypass_pg(m
, pg
, bypassed
);
1169 * Should we retry pg_init immediately?
1171 static int pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1173 unsigned long flags
;
1174 int limit_reached
= 0;
1176 spin_lock_irqsave(&m
->lock
, flags
);
1178 if (m
->pg_init_count
<= m
->pg_init_retries
)
1179 m
->pg_init_required
= 1;
1183 spin_unlock_irqrestore(&m
->lock
, flags
);
1185 return limit_reached
;
1188 static void pg_init_done(void *data
, int errors
)
1190 struct pgpath
*pgpath
= data
;
1191 struct priority_group
*pg
= pgpath
->pg
;
1192 struct multipath
*m
= pg
->m
;
1193 unsigned long flags
;
1194 unsigned delay_retry
= 0;
1196 /* device or driver problems */
1201 if (!m
->hw_handler_name
) {
1205 DMERR("Could not failover the device: Handler scsi_dh_%s "
1206 "Error %d.", m
->hw_handler_name
, errors
);
1208 * Fail path for now, so we do not ping pong
1212 case SCSI_DH_DEV_TEMP_BUSY
:
1214 * Probably doing something like FW upgrade on the
1215 * controller so try the other pg.
1217 bypass_pg(m
, pg
, 1);
1220 /* Wait before retrying. */
1222 case SCSI_DH_IMM_RETRY
:
1223 case SCSI_DH_RES_TEMP_UNAVAIL
:
1224 if (pg_init_limit_reached(m
, pgpath
))
1230 * We probably do not want to fail the path for a device
1231 * error, but this is what the old dm did. In future
1232 * patches we can do more advanced handling.
1237 spin_lock_irqsave(&m
->lock
, flags
);
1239 if (pgpath
== m
->current_pgpath
) {
1240 DMERR("Could not failover device. Error %d.", errors
);
1241 m
->current_pgpath
= NULL
;
1242 m
->current_pg
= NULL
;
1244 } else if (!m
->pg_init_required
)
1247 if (--m
->pg_init_in_progress
)
1248 /* Activations of other paths are still on going */
1251 if (!m
->pg_init_required
)
1254 m
->pg_init_delay_retry
= delay_retry
;
1255 queue_work(kmultipathd
, &m
->process_queued_ios
);
1258 * Wake up any thread waiting to suspend.
1260 wake_up(&m
->pg_init_wait
);
1263 spin_unlock_irqrestore(&m
->lock
, flags
);
1266 static void activate_path(struct work_struct
*work
)
1268 struct pgpath
*pgpath
=
1269 container_of(work
, struct pgpath
, activate_path
.work
);
1271 scsi_dh_activate(bdev_get_queue(pgpath
->path
.dev
->bdev
),
1272 pg_init_done
, pgpath
);
1278 static int do_end_io(struct multipath
*m
, struct request
*clone
,
1279 int error
, struct dm_mpath_io
*mpio
)
1282 * We don't queue any clone request inside the multipath target
1283 * during end I/O handling, since those clone requests don't have
1284 * bio clones. If we queue them inside the multipath target,
1285 * we need to make bio clones, that requires memory allocation.
1286 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1287 * don't have bio clones.)
1288 * Instead of queueing the clone request here, we queue the original
1289 * request into dm core, which will remake a clone request and
1290 * clone bios for it and resubmit it later.
1292 int r
= DM_ENDIO_REQUEUE
;
1293 unsigned long flags
;
1295 if (!error
&& !clone
->errors
)
1296 return 0; /* I/O complete */
1298 if (error
== -EOPNOTSUPP
|| error
== -EREMOTEIO
|| error
== -EILSEQ
)
1302 fail_path(mpio
->pgpath
);
1304 spin_lock_irqsave(&m
->lock
, flags
);
1305 if (!m
->nr_valid_paths
) {
1306 if (!m
->queue_if_no_path
) {
1307 if (!__must_push_back(m
))
1310 if (error
== -EBADE
)
1314 spin_unlock_irqrestore(&m
->lock
, flags
);
1319 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1320 int error
, union map_info
*map_context
)
1322 struct multipath
*m
= ti
->private;
1323 struct dm_mpath_io
*mpio
= map_context
->ptr
;
1324 struct pgpath
*pgpath
= mpio
->pgpath
;
1325 struct path_selector
*ps
;
1328 r
= do_end_io(m
, clone
, error
, mpio
);
1330 ps
= &pgpath
->pg
->ps
;
1331 if (ps
->type
->end_io
)
1332 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1334 mempool_free(mpio
, m
->mpio_pool
);
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
= (struct multipath
*) ti
->private;
1349 queue_if_no_path(m
, 0, 1);
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
= (struct multipath
*) 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 int multipath_status(struct dm_target
*ti
, status_type_t type
,
1391 char *result
, unsigned int maxlen
)
1394 unsigned long flags
;
1395 struct multipath
*m
= (struct multipath
*) 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_size
, 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 if (m
->queue_if_no_path
)
1411 DMEMIT("queue_if_no_path ");
1412 if (m
->pg_init_retries
)
1413 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1414 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1415 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1418 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1421 DMEMIT("1 %s ", m
->hw_handler_name
);
1423 DMEMIT("%u ", m
->nr_priority_groups
);
1426 pg_num
= m
->next_pg
->pg_num
;
1427 else if (m
->current_pg
)
1428 pg_num
= m
->current_pg
->pg_num
;
1430 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1432 DMEMIT("%u ", pg_num
);
1435 case STATUSTYPE_INFO
:
1436 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1438 state
= 'D'; /* Disabled */
1439 else if (pg
== m
->current_pg
)
1440 state
= 'A'; /* Currently Active */
1442 state
= 'E'; /* Enabled */
1444 DMEMIT("%c ", state
);
1446 if (pg
->ps
.type
->status
)
1447 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1453 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1454 pg
->ps
.type
->info_args
);
1456 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1457 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1458 p
->is_active
? "A" : "F",
1460 if (pg
->ps
.type
->status
)
1461 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1462 &p
->path
, type
, result
+ sz
,
1468 case STATUSTYPE_TABLE
:
1469 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1470 DMEMIT("%s ", pg
->ps
.type
->name
);
1472 if (pg
->ps
.type
->status
)
1473 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1479 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1480 pg
->ps
.type
->table_args
);
1482 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1483 DMEMIT("%s ", p
->path
.dev
->name
);
1484 if (pg
->ps
.type
->status
)
1485 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1486 &p
->path
, type
, result
+ sz
,
1493 spin_unlock_irqrestore(&m
->lock
, flags
);
1498 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1502 struct multipath
*m
= (struct multipath
*) ti
->private;
1505 mutex_lock(&m
->work_mutex
);
1507 if (dm_suspended(ti
)) {
1513 if (!strnicmp(argv
[0], MESG_STR("queue_if_no_path"))) {
1514 r
= queue_if_no_path(m
, 1, 0);
1516 } else if (!strnicmp(argv
[0], MESG_STR("fail_if_no_path"))) {
1517 r
= queue_if_no_path(m
, 0, 0);
1523 DMWARN("Unrecognised multipath message received.");
1527 if (!strnicmp(argv
[0], MESG_STR("disable_group"))) {
1528 r
= bypass_pg_num(m
, argv
[1], 1);
1530 } else if (!strnicmp(argv
[0], MESG_STR("enable_group"))) {
1531 r
= bypass_pg_num(m
, argv
[1], 0);
1533 } else if (!strnicmp(argv
[0], MESG_STR("switch_group"))) {
1534 r
= switch_pg_num(m
, argv
[1]);
1536 } else if (!strnicmp(argv
[0], MESG_STR("reinstate_path")))
1537 action
= reinstate_path
;
1538 else if (!strnicmp(argv
[0], MESG_STR("fail_path")))
1541 DMWARN("Unrecognised multipath message received.");
1545 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1547 DMWARN("message: error getting device %s",
1552 r
= action_dev(m
, dev
, action
);
1554 dm_put_device(ti
, dev
);
1557 mutex_unlock(&m
->work_mutex
);
1561 static int multipath_ioctl(struct dm_target
*ti
, unsigned int cmd
,
1564 struct multipath
*m
= (struct multipath
*) ti
->private;
1565 struct block_device
*bdev
= NULL
;
1567 unsigned long flags
;
1570 spin_lock_irqsave(&m
->lock
, flags
);
1572 if (!m
->current_pgpath
)
1573 __choose_pgpath(m
, 0);
1575 if (m
->current_pgpath
) {
1576 bdev
= m
->current_pgpath
->path
.dev
->bdev
;
1577 mode
= m
->current_pgpath
->path
.dev
->mode
;
1585 spin_unlock_irqrestore(&m
->lock
, flags
);
1588 * Only pass ioctls through if the device sizes match exactly.
1590 if (!r
&& ti
->len
!= i_size_read(bdev
->bd_inode
) >> SECTOR_SHIFT
)
1591 r
= scsi_verify_blk_ioctl(NULL
, cmd
);
1593 return r
? : __blkdev_driver_ioctl(bdev
, mode
, cmd
, arg
);
1596 static int multipath_iterate_devices(struct dm_target
*ti
,
1597 iterate_devices_callout_fn fn
, void *data
)
1599 struct multipath
*m
= ti
->private;
1600 struct priority_group
*pg
;
1604 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1605 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1606 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1616 static int __pgpath_busy(struct pgpath
*pgpath
)
1618 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1620 return dm_underlying_device_busy(q
);
1624 * We return "busy", only when we can map I/Os but underlying devices
1625 * are busy (so even if we map I/Os now, the I/Os will wait on
1626 * the underlying queue).
1627 * In other words, if we want to kill I/Os or queue them inside us
1628 * due to map unavailability, we don't return "busy". Otherwise,
1629 * dm core won't give us the I/Os and we can't do what we want.
1631 static int multipath_busy(struct dm_target
*ti
)
1633 int busy
= 0, has_active
= 0;
1634 struct multipath
*m
= ti
->private;
1635 struct priority_group
*pg
;
1636 struct pgpath
*pgpath
;
1637 unsigned long flags
;
1639 spin_lock_irqsave(&m
->lock
, flags
);
1641 /* Guess which priority_group will be used at next mapping time */
1642 if (unlikely(!m
->current_pgpath
&& m
->next_pg
))
1644 else if (likely(m
->current_pg
))
1648 * We don't know which pg will be used at next mapping time.
1649 * We don't call __choose_pgpath() here to avoid to trigger
1650 * pg_init just by busy checking.
1651 * So we don't know whether underlying devices we will be using
1652 * at next mapping time are busy or not. Just try mapping.
1657 * If there is one non-busy active path at least, the path selector
1658 * will be able to select it. So we consider such a pg as not busy.
1661 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
)
1662 if (pgpath
->is_active
) {
1665 if (!__pgpath_busy(pgpath
)) {
1673 * No active path in this pg, so this pg won't be used and
1674 * the current_pg will be changed at next mapping time.
1675 * We need to try mapping to determine it.
1680 spin_unlock_irqrestore(&m
->lock
, flags
);
1685 /*-----------------------------------------------------------------
1687 *---------------------------------------------------------------*/
1688 static struct target_type multipath_target
= {
1689 .name
= "multipath",
1690 .version
= {1, 3, 0},
1691 .module
= THIS_MODULE
,
1692 .ctr
= multipath_ctr
,
1693 .dtr
= multipath_dtr
,
1694 .map_rq
= multipath_map
,
1695 .rq_end_io
= multipath_end_io
,
1696 .presuspend
= multipath_presuspend
,
1697 .postsuspend
= multipath_postsuspend
,
1698 .resume
= multipath_resume
,
1699 .status
= multipath_status
,
1700 .message
= multipath_message
,
1701 .ioctl
= multipath_ioctl
,
1702 .iterate_devices
= multipath_iterate_devices
,
1703 .busy
= multipath_busy
,
1706 static int __init
dm_multipath_init(void)
1710 /* allocate a slab for the dm_ios */
1711 _mpio_cache
= KMEM_CACHE(dm_mpath_io
, 0);
1715 r
= dm_register_target(&multipath_target
);
1717 DMERR("register failed %d", r
);
1718 kmem_cache_destroy(_mpio_cache
);
1722 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1724 DMERR("failed to create workqueue kmpathd");
1725 dm_unregister_target(&multipath_target
);
1726 kmem_cache_destroy(_mpio_cache
);
1731 * A separate workqueue is used to handle the device handlers
1732 * to avoid overloading existing workqueue. Overloading the
1733 * old workqueue would also create a bottleneck in the
1734 * path of the storage hardware device activation.
1736 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1738 if (!kmpath_handlerd
) {
1739 DMERR("failed to create workqueue kmpath_handlerd");
1740 destroy_workqueue(kmultipathd
);
1741 dm_unregister_target(&multipath_target
);
1742 kmem_cache_destroy(_mpio_cache
);
1746 DMINFO("version %u.%u.%u loaded",
1747 multipath_target
.version
[0], multipath_target
.version
[1],
1748 multipath_target
.version
[2]);
1753 static void __exit
dm_multipath_exit(void)
1755 destroy_workqueue(kmpath_handlerd
);
1756 destroy_workqueue(kmultipathd
);
1758 dm_unregister_target(&multipath_target
);
1759 kmem_cache_destroy(_mpio_cache
);
1762 module_init(dm_multipath_init
);
1763 module_exit(dm_multipath_exit
);
1765 MODULE_DESCRIPTION(DM_NAME
" multipath target");
1766 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1767 MODULE_LICENSE("GPL");