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-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
35 struct list_head list
;
37 struct priority_group
*pg
; /* Owning PG */
38 unsigned fail_count
; /* Cumulative failure count */
41 struct delayed_work activate_path
;
43 bool is_active
:1; /* Path status */
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
52 struct priority_group
{
53 struct list_head list
;
55 struct multipath
*m
; /* Owning multipath instance */
56 struct path_selector ps
;
58 unsigned pg_num
; /* Reference number */
59 unsigned nr_pgpaths
; /* Number of paths in PG */
60 struct list_head pgpaths
;
62 bool bypassed
:1; /* Temporarily bypass this PG? */
65 /* Multipath context */
67 struct list_head list
;
70 const char *hw_handler_name
;
71 char *hw_handler_params
;
75 unsigned nr_priority_groups
;
76 struct list_head priority_groups
;
78 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
80 struct pgpath
*current_pgpath
;
81 struct priority_group
*current_pg
;
82 struct priority_group
*next_pg
; /* Switch to this PG if set */
84 unsigned long flags
; /* Multipath state flags */
86 unsigned pg_init_retries
; /* Number of times to retry pg_init */
87 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
89 atomic_t nr_valid_paths
; /* Total number of usable paths */
90 atomic_t pg_init_in_progress
; /* Only one pg_init allowed at once */
91 atomic_t pg_init_count
; /* Number of times pg_init called */
93 enum dm_queue_mode queue_mode
;
95 struct mutex work_mutex
;
96 struct work_struct trigger_event
;
98 struct work_struct process_queued_bios
;
99 struct bio_list queued_bios
;
103 * Context information attached to each io we process.
106 struct pgpath
*pgpath
;
110 typedef int (*action_fn
) (struct pgpath
*pgpath
);
112 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
113 static void trigger_event(struct work_struct
*work
);
114 static void activate_or_offline_path(struct pgpath
*pgpath
);
115 static void activate_path_work(struct work_struct
*work
);
116 static void process_queued_bios(struct work_struct
*work
);
118 /*-----------------------------------------------
119 * Multipath state flags.
120 *-----------------------------------------------*/
122 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
134 static struct pgpath
*alloc_pgpath(void)
136 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
139 pgpath
->is_active
= true;
140 INIT_DELAYED_WORK(&pgpath
->activate_path
, activate_path_work
);
146 static void free_pgpath(struct pgpath
*pgpath
)
151 static struct priority_group
*alloc_priority_group(void)
153 struct priority_group
*pg
;
155 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
158 INIT_LIST_HEAD(&pg
->pgpaths
);
163 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
165 struct pgpath
*pgpath
, *tmp
;
167 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
168 list_del(&pgpath
->list
);
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 spin_lock_init(&m
->lock
);
196 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
197 atomic_set(&m
->nr_valid_paths
, 0);
198 atomic_set(&m
->pg_init_in_progress
, 0);
199 atomic_set(&m
->pg_init_count
, 0);
200 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
201 INIT_WORK(&m
->trigger_event
, trigger_event
);
202 init_waitqueue_head(&m
->pg_init_wait
);
203 mutex_init(&m
->work_mutex
);
205 m
->queue_mode
= DM_TYPE_NONE
;
214 static int alloc_multipath_stage2(struct dm_target
*ti
, struct multipath
*m
)
216 if (m
->queue_mode
== DM_TYPE_NONE
) {
218 * Default to request-based.
220 if (dm_use_blk_mq(dm_table_get_md(ti
->table
)))
221 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
223 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
224 } else if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
225 INIT_WORK(&m
->process_queued_bios
, process_queued_bios
);
227 * bio-based doesn't support any direct scsi_dh management;
228 * it just discovers if a scsi_dh is attached.
230 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
233 dm_table_set_type(ti
->table
, m
->queue_mode
);
238 static void free_multipath(struct multipath
*m
)
240 struct priority_group
*pg
, *tmp
;
242 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
244 free_priority_group(pg
, m
->ti
);
247 kfree(m
->hw_handler_name
);
248 kfree(m
->hw_handler_params
);
252 static struct dm_mpath_io
*get_mpio(union map_info
*info
)
257 static size_t multipath_per_bio_data_size(void)
259 return sizeof(struct dm_mpath_io
) + sizeof(struct dm_bio_details
);
262 static struct dm_mpath_io
*get_mpio_from_bio(struct bio
*bio
)
264 return dm_per_bio_data(bio
, multipath_per_bio_data_size());
267 static struct dm_bio_details
*get_bio_details_from_bio(struct bio
*bio
)
269 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
271 void *bio_details
= mpio
+ 1;
276 static void multipath_init_per_bio_data(struct bio
*bio
, struct dm_mpath_io
**mpio_p
,
277 struct dm_bio_details
**bio_details_p
)
279 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
280 struct dm_bio_details
*bio_details
= get_bio_details_from_bio(bio
);
282 memset(mpio
, 0, sizeof(*mpio
));
283 memset(bio_details
, 0, sizeof(*bio_details
));
284 dm_bio_record(bio_details
, bio
);
289 *bio_details_p
= bio_details
;
292 /*-----------------------------------------------
294 *-----------------------------------------------*/
296 static int __pg_init_all_paths(struct multipath
*m
)
298 struct pgpath
*pgpath
;
299 unsigned long pg_init_delay
= 0;
301 lockdep_assert_held(&m
->lock
);
303 if (atomic_read(&m
->pg_init_in_progress
) || test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
306 atomic_inc(&m
->pg_init_count
);
307 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
309 /* Check here to reset pg_init_required */
313 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
))
314 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
315 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
316 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
317 /* Skip failed paths */
318 if (!pgpath
->is_active
)
320 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
322 atomic_inc(&m
->pg_init_in_progress
);
324 return atomic_read(&m
->pg_init_in_progress
);
327 static int pg_init_all_paths(struct multipath
*m
)
332 spin_lock_irqsave(&m
->lock
, flags
);
333 ret
= __pg_init_all_paths(m
);
334 spin_unlock_irqrestore(&m
->lock
, flags
);
339 static void __switch_pg(struct multipath
*m
, struct priority_group
*pg
)
343 /* Must we initialise the PG first, and queue I/O till it's ready? */
344 if (m
->hw_handler_name
) {
345 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
346 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
348 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
349 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
352 atomic_set(&m
->pg_init_count
, 0);
355 static struct pgpath
*choose_path_in_pg(struct multipath
*m
,
356 struct priority_group
*pg
,
360 struct dm_path
*path
;
361 struct pgpath
*pgpath
;
363 path
= pg
->ps
.type
->select_path(&pg
->ps
, nr_bytes
);
365 return ERR_PTR(-ENXIO
);
367 pgpath
= path_to_pgpath(path
);
369 if (unlikely(lockless_dereference(m
->current_pg
) != pg
)) {
370 /* Only update current_pgpath if pg changed */
371 spin_lock_irqsave(&m
->lock
, flags
);
372 m
->current_pgpath
= pgpath
;
374 spin_unlock_irqrestore(&m
->lock
, flags
);
380 static struct pgpath
*choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
383 struct priority_group
*pg
;
384 struct pgpath
*pgpath
;
385 unsigned bypassed
= 1;
387 if (!atomic_read(&m
->nr_valid_paths
)) {
388 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
392 /* Were we instructed to switch PG? */
393 if (lockless_dereference(m
->next_pg
)) {
394 spin_lock_irqsave(&m
->lock
, flags
);
397 spin_unlock_irqrestore(&m
->lock
, flags
);
398 goto check_current_pg
;
401 spin_unlock_irqrestore(&m
->lock
, flags
);
402 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
403 if (!IS_ERR_OR_NULL(pgpath
))
407 /* Don't change PG until it has no remaining paths */
409 pg
= lockless_dereference(m
->current_pg
);
411 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
412 if (!IS_ERR_OR_NULL(pgpath
))
417 * Loop through priority groups until we find a valid path.
418 * First time we skip PGs marked 'bypassed'.
419 * Second time we only try the ones we skipped, but set
420 * pg_init_delay_retry so we do not hammer controllers.
423 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
424 if (pg
->bypassed
== !!bypassed
)
426 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
427 if (!IS_ERR_OR_NULL(pgpath
)) {
429 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
433 } while (bypassed
--);
436 spin_lock_irqsave(&m
->lock
, flags
);
437 m
->current_pgpath
= NULL
;
438 m
->current_pg
= NULL
;
439 spin_unlock_irqrestore(&m
->lock
, flags
);
445 * dm_report_EIO() is a macro instead of a function to make pr_debug()
446 * report the function name and line number of the function from which
447 * it has been invoked.
449 #define dm_report_EIO(m) \
451 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
453 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454 dm_device_name(md), \
455 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
456 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
457 dm_noflush_suspending((m)->ti)); \
461 * Map cloned requests (request-based multipath)
463 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
464 union map_info
*map_context
,
465 struct request
**__clone
)
467 struct multipath
*m
= ti
->private;
468 size_t nr_bytes
= blk_rq_bytes(rq
);
469 struct pgpath
*pgpath
;
470 struct block_device
*bdev
;
471 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
472 struct request_queue
*q
;
473 struct request
*clone
;
475 /* Do we need to select a new pgpath? */
476 pgpath
= lockless_dereference(m
->current_pgpath
);
477 if (!pgpath
|| !test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
478 pgpath
= choose_pgpath(m
, nr_bytes
);
481 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
482 return DM_MAPIO_DELAY_REQUEUE
;
483 dm_report_EIO(m
); /* Failed */
484 return DM_MAPIO_KILL
;
485 } else if (test_bit(MPATHF_QUEUE_IO
, &m
->flags
) ||
486 test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
487 if (pg_init_all_paths(m
))
488 return DM_MAPIO_DELAY_REQUEUE
;
489 return DM_MAPIO_REQUEUE
;
492 memset(mpio
, 0, sizeof(*mpio
));
493 mpio
->pgpath
= pgpath
;
494 mpio
->nr_bytes
= nr_bytes
;
496 bdev
= pgpath
->path
.dev
->bdev
;
497 q
= bdev_get_queue(bdev
);
498 clone
= blk_get_request(q
, rq
->cmd_flags
| REQ_NOMERGE
, GFP_ATOMIC
);
500 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501 bool queue_dying
= blk_queue_dying(q
);
502 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
503 PTR_ERR(clone
), queue_dying
? " (path offline)" : "");
505 atomic_inc(&m
->pg_init_in_progress
);
506 activate_or_offline_path(pgpath
);
507 return DM_MAPIO_REQUEUE
;
509 return DM_MAPIO_DELAY_REQUEUE
;
511 clone
->bio
= clone
->biotail
= NULL
;
512 clone
->rq_disk
= bdev
->bd_disk
;
513 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
516 if (pgpath
->pg
->ps
.type
->start_io
)
517 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
520 return DM_MAPIO_REMAPPED
;
523 static void multipath_release_clone(struct request
*clone
)
525 blk_put_request(clone
);
529 * Map cloned bios (bio-based multipath)
531 static int __multipath_map_bio(struct multipath
*m
, struct bio
*bio
, struct dm_mpath_io
*mpio
)
533 size_t nr_bytes
= bio
->bi_iter
.bi_size
;
534 struct pgpath
*pgpath
;
538 /* Do we need to select a new pgpath? */
539 pgpath
= lockless_dereference(m
->current_pgpath
);
540 queue_io
= test_bit(MPATHF_QUEUE_IO
, &m
->flags
);
541 if (!pgpath
|| !queue_io
)
542 pgpath
= choose_pgpath(m
, nr_bytes
);
544 if ((pgpath
&& queue_io
) ||
545 (!pgpath
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))) {
546 /* Queue for the daemon to resubmit */
547 spin_lock_irqsave(&m
->lock
, flags
);
548 bio_list_add(&m
->queued_bios
, bio
);
549 spin_unlock_irqrestore(&m
->lock
, flags
);
550 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
551 if (queue_io
|| test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
552 pg_init_all_paths(m
);
554 queue_work(kmultipathd
, &m
->process_queued_bios
);
555 return DM_MAPIO_SUBMITTED
;
559 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
560 return DM_MAPIO_REQUEUE
;
562 return DM_MAPIO_KILL
;
565 mpio
->pgpath
= pgpath
;
566 mpio
->nr_bytes
= nr_bytes
;
569 bio
->bi_bdev
= pgpath
->path
.dev
->bdev
;
570 bio
->bi_opf
|= REQ_FAILFAST_TRANSPORT
;
572 if (pgpath
->pg
->ps
.type
->start_io
)
573 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
576 return DM_MAPIO_REMAPPED
;
579 static int multipath_map_bio(struct dm_target
*ti
, struct bio
*bio
)
581 struct multipath
*m
= ti
->private;
582 struct dm_mpath_io
*mpio
= NULL
;
584 multipath_init_per_bio_data(bio
, &mpio
, NULL
);
586 return __multipath_map_bio(m
, bio
, mpio
);
589 static void process_queued_io_list(struct multipath
*m
)
591 if (m
->queue_mode
== DM_TYPE_MQ_REQUEST_BASED
)
592 dm_mq_kick_requeue_list(dm_table_get_md(m
->ti
->table
));
593 else if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
594 queue_work(kmultipathd
, &m
->process_queued_bios
);
597 static void process_queued_bios(struct work_struct
*work
)
602 struct bio_list bios
;
603 struct blk_plug plug
;
604 struct multipath
*m
=
605 container_of(work
, struct multipath
, process_queued_bios
);
607 bio_list_init(&bios
);
609 spin_lock_irqsave(&m
->lock
, flags
);
611 if (bio_list_empty(&m
->queued_bios
)) {
612 spin_unlock_irqrestore(&m
->lock
, flags
);
616 bio_list_merge(&bios
, &m
->queued_bios
);
617 bio_list_init(&m
->queued_bios
);
619 spin_unlock_irqrestore(&m
->lock
, flags
);
621 blk_start_plug(&plug
);
622 while ((bio
= bio_list_pop(&bios
))) {
623 r
= __multipath_map_bio(m
, bio
, get_mpio_from_bio(bio
));
626 bio
->bi_status
= BLK_STS_IOERR
;
629 case DM_MAPIO_REQUEUE
:
630 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
633 case DM_MAPIO_REMAPPED
:
634 generic_make_request(bio
);
638 blk_finish_plug(&plug
);
641 static void assign_bit(bool value
, long nr
, unsigned long *addr
)
650 * If we run out of usable paths, should we queue I/O or error it?
652 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
657 spin_lock_irqsave(&m
->lock
, flags
);
658 assign_bit((save_old_value
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) ||
659 (!save_old_value
&& queue_if_no_path
),
660 MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
661 assign_bit(queue_if_no_path
|| dm_noflush_suspending(m
->ti
),
662 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
663 spin_unlock_irqrestore(&m
->lock
, flags
);
665 if (!queue_if_no_path
) {
666 dm_table_run_md_queue_async(m
->ti
->table
);
667 process_queued_io_list(m
);
674 * An event is triggered whenever a path is taken out of use.
675 * Includes path failure and PG bypass.
677 static void trigger_event(struct work_struct
*work
)
679 struct multipath
*m
=
680 container_of(work
, struct multipath
, trigger_event
);
682 dm_table_event(m
->ti
->table
);
685 /*-----------------------------------------------------------------
686 * Constructor/argument parsing:
687 * <#multipath feature args> [<arg>]*
688 * <#hw_handler args> [hw_handler [<arg>]*]
690 * <initial priority group>
691 * [<selector> <#selector args> [<arg>]*
692 * <#paths> <#per-path selector args>
693 * [<path> [<arg>]* ]+ ]+
694 *---------------------------------------------------------------*/
695 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
696 struct dm_target
*ti
)
699 struct path_selector_type
*pst
;
702 static struct dm_arg _args
[] = {
703 {0, 1024, "invalid number of path selector args"},
706 pst
= dm_get_path_selector(dm_shift_arg(as
));
708 ti
->error
= "unknown path selector type";
712 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
714 dm_put_path_selector(pst
);
718 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
720 dm_put_path_selector(pst
);
721 ti
->error
= "path selector constructor failed";
726 dm_consume_args(as
, ps_argc
);
731 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
732 struct dm_target
*ti
)
736 struct multipath
*m
= ti
->private;
737 struct request_queue
*q
= NULL
;
738 const char *attached_handler_name
;
740 /* we need at least a path arg */
742 ti
->error
= "no device given";
743 return ERR_PTR(-EINVAL
);
748 return ERR_PTR(-ENOMEM
);
750 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
753 ti
->error
= "error getting device";
757 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) || m
->hw_handler_name
)
758 q
= bdev_get_queue(p
->path
.dev
->bdev
);
760 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
)) {
762 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
763 if (attached_handler_name
) {
765 * Clear any hw_handler_params associated with a
766 * handler that isn't already attached.
768 if (m
->hw_handler_name
&& strcmp(attached_handler_name
, m
->hw_handler_name
)) {
769 kfree(m
->hw_handler_params
);
770 m
->hw_handler_params
= NULL
;
774 * Reset hw_handler_name to match the attached handler
776 * NB. This modifies the table line to show the actual
777 * handler instead of the original table passed in.
779 kfree(m
->hw_handler_name
);
780 m
->hw_handler_name
= attached_handler_name
;
784 if (m
->hw_handler_name
) {
785 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
787 char b
[BDEVNAME_SIZE
];
789 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
790 bdevname(p
->path
.dev
->bdev
, b
));
794 ti
->error
= "error attaching hardware handler";
795 dm_put_device(ti
, p
->path
.dev
);
799 if (m
->hw_handler_params
) {
800 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
802 ti
->error
= "unable to set hardware "
803 "handler parameters";
804 dm_put_device(ti
, p
->path
.dev
);
810 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
812 dm_put_device(ti
, p
->path
.dev
);
823 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
826 static struct dm_arg _args
[] = {
827 {1, 1024, "invalid number of paths"},
828 {0, 1024, "invalid number of selector args"}
832 unsigned i
, nr_selector_args
, nr_args
;
833 struct priority_group
*pg
;
834 struct dm_target
*ti
= m
->ti
;
838 ti
->error
= "not enough priority group arguments";
839 return ERR_PTR(-EINVAL
);
842 pg
= alloc_priority_group();
844 ti
->error
= "couldn't allocate priority group";
845 return ERR_PTR(-ENOMEM
);
849 r
= parse_path_selector(as
, pg
, ti
);
856 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
860 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
864 nr_args
= 1 + nr_selector_args
;
865 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
866 struct pgpath
*pgpath
;
867 struct dm_arg_set path_args
;
869 if (as
->argc
< nr_args
) {
870 ti
->error
= "not enough path parameters";
875 path_args
.argc
= nr_args
;
876 path_args
.argv
= as
->argv
;
878 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
879 if (IS_ERR(pgpath
)) {
885 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
886 dm_consume_args(as
, nr_args
);
892 free_priority_group(pg
, ti
);
896 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
900 struct dm_target
*ti
= m
->ti
;
902 static struct dm_arg _args
[] = {
903 {0, 1024, "invalid number of hardware handler args"},
906 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
912 if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
913 dm_consume_args(as
, hw_argc
);
914 DMERR("bio-based multipath doesn't allow hardware handler args");
918 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
919 if (!m
->hw_handler_name
)
926 for (i
= 0; i
<= hw_argc
- 2; i
++)
927 len
+= strlen(as
->argv
[i
]) + 1;
928 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
930 ti
->error
= "memory allocation failed";
934 j
= sprintf(p
, "%d", hw_argc
- 1);
935 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
936 j
= sprintf(p
, "%s", as
->argv
[i
]);
938 dm_consume_args(as
, hw_argc
- 1);
942 kfree(m
->hw_handler_name
);
943 m
->hw_handler_name
= NULL
;
947 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
951 struct dm_target
*ti
= m
->ti
;
952 const char *arg_name
;
954 static struct dm_arg _args
[] = {
955 {0, 8, "invalid number of feature args"},
956 {1, 50, "pg_init_retries must be between 1 and 50"},
957 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
960 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
968 arg_name
= dm_shift_arg(as
);
971 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
972 r
= queue_if_no_path(m
, true, false);
976 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
977 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
981 if (!strcasecmp(arg_name
, "pg_init_retries") &&
983 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
988 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
990 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
995 if (!strcasecmp(arg_name
, "queue_mode") &&
997 const char *queue_mode_name
= dm_shift_arg(as
);
999 if (!strcasecmp(queue_mode_name
, "bio"))
1000 m
->queue_mode
= DM_TYPE_BIO_BASED
;
1001 else if (!strcasecmp(queue_mode_name
, "rq"))
1002 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
1003 else if (!strcasecmp(queue_mode_name
, "mq"))
1004 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
1006 ti
->error
= "Unknown 'queue_mode' requested";
1013 ti
->error
= "Unrecognised multipath feature request";
1015 } while (argc
&& !r
);
1020 static int multipath_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
1022 /* target arguments */
1023 static struct dm_arg _args
[] = {
1024 {0, 1024, "invalid number of priority groups"},
1025 {0, 1024, "invalid initial priority group number"},
1029 struct multipath
*m
;
1030 struct dm_arg_set as
;
1031 unsigned pg_count
= 0;
1032 unsigned next_pg_num
;
1037 m
= alloc_multipath(ti
);
1039 ti
->error
= "can't allocate multipath";
1043 r
= parse_features(&as
, m
);
1047 r
= alloc_multipath_stage2(ti
, m
);
1051 r
= parse_hw_handler(&as
, m
);
1055 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
1059 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
1063 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
1064 (m
->nr_priority_groups
&& !next_pg_num
)) {
1065 ti
->error
= "invalid initial priority group";
1070 /* parse the priority groups */
1072 struct priority_group
*pg
;
1073 unsigned nr_valid_paths
= atomic_read(&m
->nr_valid_paths
);
1075 pg
= parse_priority_group(&as
, m
);
1081 nr_valid_paths
+= pg
->nr_pgpaths
;
1082 atomic_set(&m
->nr_valid_paths
, nr_valid_paths
);
1084 list_add_tail(&pg
->list
, &m
->priority_groups
);
1086 pg
->pg_num
= pg_count
;
1091 if (pg_count
!= m
->nr_priority_groups
) {
1092 ti
->error
= "priority group count mismatch";
1097 ti
->num_flush_bios
= 1;
1098 ti
->num_discard_bios
= 1;
1099 ti
->num_write_same_bios
= 1;
1100 ti
->num_write_zeroes_bios
= 1;
1101 if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
1102 ti
->per_io_data_size
= multipath_per_bio_data_size();
1104 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
1113 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
1118 prepare_to_wait(&m
->pg_init_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1120 if (!atomic_read(&m
->pg_init_in_progress
))
1125 finish_wait(&m
->pg_init_wait
, &wait
);
1128 static void flush_multipath_work(struct multipath
*m
)
1130 set_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1131 smp_mb__after_atomic();
1133 flush_workqueue(kmpath_handlerd
);
1134 multipath_wait_for_pg_init_completion(m
);
1135 flush_workqueue(kmultipathd
);
1136 flush_work(&m
->trigger_event
);
1138 clear_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1139 smp_mb__after_atomic();
1142 static void multipath_dtr(struct dm_target
*ti
)
1144 struct multipath
*m
= ti
->private;
1146 flush_multipath_work(m
);
1151 * Take a path out of use.
1153 static int fail_path(struct pgpath
*pgpath
)
1155 unsigned long flags
;
1156 struct multipath
*m
= pgpath
->pg
->m
;
1158 spin_lock_irqsave(&m
->lock
, flags
);
1160 if (!pgpath
->is_active
)
1163 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1165 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1166 pgpath
->is_active
= false;
1167 pgpath
->fail_count
++;
1169 atomic_dec(&m
->nr_valid_paths
);
1171 if (pgpath
== m
->current_pgpath
)
1172 m
->current_pgpath
= NULL
;
1174 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1175 pgpath
->path
.dev
->name
, atomic_read(&m
->nr_valid_paths
));
1177 schedule_work(&m
->trigger_event
);
1180 spin_unlock_irqrestore(&m
->lock
, flags
);
1186 * Reinstate a previously-failed path
1188 static int reinstate_path(struct pgpath
*pgpath
)
1190 int r
= 0, run_queue
= 0;
1191 unsigned long flags
;
1192 struct multipath
*m
= pgpath
->pg
->m
;
1193 unsigned nr_valid_paths
;
1195 spin_lock_irqsave(&m
->lock
, flags
);
1197 if (pgpath
->is_active
)
1200 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1202 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1206 pgpath
->is_active
= true;
1208 nr_valid_paths
= atomic_inc_return(&m
->nr_valid_paths
);
1209 if (nr_valid_paths
== 1) {
1210 m
->current_pgpath
= NULL
;
1212 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1213 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1214 atomic_inc(&m
->pg_init_in_progress
);
1217 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1218 pgpath
->path
.dev
->name
, nr_valid_paths
);
1220 schedule_work(&m
->trigger_event
);
1223 spin_unlock_irqrestore(&m
->lock
, flags
);
1225 dm_table_run_md_queue_async(m
->ti
->table
);
1226 process_queued_io_list(m
);
1233 * Fail or reinstate all paths that match the provided struct dm_dev.
1235 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1239 struct pgpath
*pgpath
;
1240 struct priority_group
*pg
;
1242 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1243 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1244 if (pgpath
->path
.dev
== dev
)
1253 * Temporarily try to avoid having to use the specified PG
1255 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1258 unsigned long flags
;
1260 spin_lock_irqsave(&m
->lock
, flags
);
1262 pg
->bypassed
= bypassed
;
1263 m
->current_pgpath
= NULL
;
1264 m
->current_pg
= NULL
;
1266 spin_unlock_irqrestore(&m
->lock
, flags
);
1268 schedule_work(&m
->trigger_event
);
1272 * Switch to using the specified PG from the next I/O that gets mapped
1274 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1276 struct priority_group
*pg
;
1278 unsigned long flags
;
1281 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1282 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1283 DMWARN("invalid PG number supplied to switch_pg_num");
1287 spin_lock_irqsave(&m
->lock
, flags
);
1288 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1289 pg
->bypassed
= false;
1293 m
->current_pgpath
= NULL
;
1294 m
->current_pg
= NULL
;
1297 spin_unlock_irqrestore(&m
->lock
, flags
);
1299 schedule_work(&m
->trigger_event
);
1304 * Set/clear bypassed status of a PG.
1305 * PGs are numbered upwards from 1 in the order they were declared.
1307 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1309 struct priority_group
*pg
;
1313 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1314 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1315 DMWARN("invalid PG number supplied to bypass_pg");
1319 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1324 bypass_pg(m
, pg
, bypassed
);
1329 * Should we retry pg_init immediately?
1331 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1333 unsigned long flags
;
1334 bool limit_reached
= false;
1336 spin_lock_irqsave(&m
->lock
, flags
);
1338 if (atomic_read(&m
->pg_init_count
) <= m
->pg_init_retries
&&
1339 !test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
1340 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
1342 limit_reached
= true;
1344 spin_unlock_irqrestore(&m
->lock
, flags
);
1346 return limit_reached
;
1349 static void pg_init_done(void *data
, int errors
)
1351 struct pgpath
*pgpath
= data
;
1352 struct priority_group
*pg
= pgpath
->pg
;
1353 struct multipath
*m
= pg
->m
;
1354 unsigned long flags
;
1355 bool delay_retry
= false;
1357 /* device or driver problems */
1362 if (!m
->hw_handler_name
) {
1366 DMERR("Could not failover the device: Handler scsi_dh_%s "
1367 "Error %d.", m
->hw_handler_name
, errors
);
1369 * Fail path for now, so we do not ping pong
1373 case SCSI_DH_DEV_TEMP_BUSY
:
1375 * Probably doing something like FW upgrade on the
1376 * controller so try the other pg.
1378 bypass_pg(m
, pg
, true);
1381 /* Wait before retrying. */
1383 case SCSI_DH_IMM_RETRY
:
1384 case SCSI_DH_RES_TEMP_UNAVAIL
:
1385 if (pg_init_limit_reached(m
, pgpath
))
1389 case SCSI_DH_DEV_OFFLINED
:
1392 * We probably do not want to fail the path for a device
1393 * error, but this is what the old dm did. In future
1394 * patches we can do more advanced handling.
1399 spin_lock_irqsave(&m
->lock
, flags
);
1401 if (pgpath
== m
->current_pgpath
) {
1402 DMERR("Could not failover device. Error %d.", errors
);
1403 m
->current_pgpath
= NULL
;
1404 m
->current_pg
= NULL
;
1406 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1407 pg
->bypassed
= false;
1409 if (atomic_dec_return(&m
->pg_init_in_progress
) > 0)
1410 /* Activations of other paths are still on going */
1413 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
1415 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1417 clear_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1419 if (__pg_init_all_paths(m
))
1422 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
1424 process_queued_io_list(m
);
1427 * Wake up any thread waiting to suspend.
1429 wake_up(&m
->pg_init_wait
);
1432 spin_unlock_irqrestore(&m
->lock
, flags
);
1435 static void activate_or_offline_path(struct pgpath
*pgpath
)
1437 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1439 if (pgpath
->is_active
&& !blk_queue_dying(q
))
1440 scsi_dh_activate(q
, pg_init_done
, pgpath
);
1442 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1445 static void activate_path_work(struct work_struct
*work
)
1447 struct pgpath
*pgpath
=
1448 container_of(work
, struct pgpath
, activate_path
.work
);
1450 activate_or_offline_path(pgpath
);
1453 static int noretry_error(blk_status_t error
)
1456 case BLK_STS_NOTSUPP
:
1458 case BLK_STS_TARGET
:
1460 case BLK_STS_MEDIUM
:
1461 case BLK_STS_RESOURCE
:
1465 /* Anything else could be a path failure, so should be retried */
1469 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1470 blk_status_t error
, union map_info
*map_context
)
1472 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1473 struct pgpath
*pgpath
= mpio
->pgpath
;
1474 int r
= DM_ENDIO_DONE
;
1477 * We don't queue any clone request inside the multipath target
1478 * during end I/O handling, since those clone requests don't have
1479 * bio clones. If we queue them inside the multipath target,
1480 * we need to make bio clones, that requires memory allocation.
1481 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1482 * don't have bio clones.)
1483 * Instead of queueing the clone request here, we queue the original
1484 * request into dm core, which will remake a clone request and
1485 * clone bios for it and resubmit it later.
1487 if (error
&& !noretry_error(error
)) {
1488 struct multipath
*m
= ti
->private;
1490 r
= DM_ENDIO_REQUEUE
;
1495 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1496 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1497 if (error
== BLK_STS_IOERR
)
1499 /* complete with the original error */
1505 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1507 if (ps
->type
->end_io
)
1508 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1514 static int multipath_end_io_bio(struct dm_target
*ti
, struct bio
*clone
,
1515 blk_status_t
*error
)
1517 struct multipath
*m
= ti
->private;
1518 struct dm_mpath_io
*mpio
= get_mpio_from_bio(clone
);
1519 struct pgpath
*pgpath
= mpio
->pgpath
;
1520 unsigned long flags
;
1521 int r
= DM_ENDIO_DONE
;
1523 if (!*error
|| noretry_error(*error
))
1529 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1530 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1532 *error
= BLK_STS_IOERR
;
1536 /* Queue for the daemon to resubmit */
1537 dm_bio_restore(get_bio_details_from_bio(clone
), clone
);
1539 spin_lock_irqsave(&m
->lock
, flags
);
1540 bio_list_add(&m
->queued_bios
, clone
);
1541 spin_unlock_irqrestore(&m
->lock
, flags
);
1542 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
1543 queue_work(kmultipathd
, &m
->process_queued_bios
);
1545 r
= DM_ENDIO_INCOMPLETE
;
1548 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1550 if (ps
->type
->end_io
)
1551 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1558 * Suspend can't complete until all the I/O is processed so if
1559 * the last path fails we must error any remaining I/O.
1560 * Note that if the freeze_bdev fails while suspending, the
1561 * queue_if_no_path state is lost - userspace should reset it.
1563 static void multipath_presuspend(struct dm_target
*ti
)
1565 struct multipath
*m
= ti
->private;
1567 queue_if_no_path(m
, false, true);
1570 static void multipath_postsuspend(struct dm_target
*ti
)
1572 struct multipath
*m
= ti
->private;
1574 mutex_lock(&m
->work_mutex
);
1575 flush_multipath_work(m
);
1576 mutex_unlock(&m
->work_mutex
);
1580 * Restore the queue_if_no_path setting.
1582 static void multipath_resume(struct dm_target
*ti
)
1584 struct multipath
*m
= ti
->private;
1585 unsigned long flags
;
1587 spin_lock_irqsave(&m
->lock
, flags
);
1588 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
),
1589 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
1590 spin_unlock_irqrestore(&m
->lock
, flags
);
1594 * Info output has the following format:
1595 * num_multipath_feature_args [multipath_feature_args]*
1596 * num_handler_status_args [handler_status_args]*
1597 * num_groups init_group_number
1598 * [A|D|E num_ps_status_args [ps_status_args]*
1599 * num_paths num_selector_args
1600 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1602 * Table output has the following format (identical to the constructor string):
1603 * num_feature_args [features_args]*
1604 * num_handler_args hw_handler [hw_handler_args]*
1605 * num_groups init_group_number
1606 * [priority selector-name num_ps_args [ps_args]*
1607 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1609 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1610 unsigned status_flags
, char *result
, unsigned maxlen
)
1613 unsigned long flags
;
1614 struct multipath
*m
= ti
->private;
1615 struct priority_group
*pg
;
1620 spin_lock_irqsave(&m
->lock
, flags
);
1623 if (type
== STATUSTYPE_INFO
)
1624 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO
, &m
->flags
),
1625 atomic_read(&m
->pg_init_count
));
1627 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) +
1628 (m
->pg_init_retries
> 0) * 2 +
1629 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1630 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) +
1631 (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) * 2);
1633 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1634 DMEMIT("queue_if_no_path ");
1635 if (m
->pg_init_retries
)
1636 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1637 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1638 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1639 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
))
1640 DMEMIT("retain_attached_hw_handler ");
1641 if (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) {
1642 switch(m
->queue_mode
) {
1643 case DM_TYPE_BIO_BASED
:
1644 DMEMIT("queue_mode bio ");
1646 case DM_TYPE_MQ_REQUEST_BASED
:
1647 DMEMIT("queue_mode mq ");
1656 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1659 DMEMIT("1 %s ", m
->hw_handler_name
);
1661 DMEMIT("%u ", m
->nr_priority_groups
);
1664 pg_num
= m
->next_pg
->pg_num
;
1665 else if (m
->current_pg
)
1666 pg_num
= m
->current_pg
->pg_num
;
1668 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1670 DMEMIT("%u ", pg_num
);
1673 case STATUSTYPE_INFO
:
1674 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1676 state
= 'D'; /* Disabled */
1677 else if (pg
== m
->current_pg
)
1678 state
= 'A'; /* Currently Active */
1680 state
= 'E'; /* Enabled */
1682 DMEMIT("%c ", state
);
1684 if (pg
->ps
.type
->status
)
1685 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1691 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1692 pg
->ps
.type
->info_args
);
1694 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1695 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1696 p
->is_active
? "A" : "F",
1698 if (pg
->ps
.type
->status
)
1699 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1700 &p
->path
, type
, result
+ sz
,
1706 case STATUSTYPE_TABLE
:
1707 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1708 DMEMIT("%s ", pg
->ps
.type
->name
);
1710 if (pg
->ps
.type
->status
)
1711 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1717 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1718 pg
->ps
.type
->table_args
);
1720 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1721 DMEMIT("%s ", p
->path
.dev
->name
);
1722 if (pg
->ps
.type
->status
)
1723 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1724 &p
->path
, type
, result
+ sz
,
1731 spin_unlock_irqrestore(&m
->lock
, flags
);
1734 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1738 struct multipath
*m
= ti
->private;
1741 mutex_lock(&m
->work_mutex
);
1743 if (dm_suspended(ti
)) {
1749 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1750 r
= queue_if_no_path(m
, true, false);
1752 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1753 r
= queue_if_no_path(m
, false, false);
1759 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1763 if (!strcasecmp(argv
[0], "disable_group")) {
1764 r
= bypass_pg_num(m
, argv
[1], true);
1766 } else if (!strcasecmp(argv
[0], "enable_group")) {
1767 r
= bypass_pg_num(m
, argv
[1], false);
1769 } else if (!strcasecmp(argv
[0], "switch_group")) {
1770 r
= switch_pg_num(m
, argv
[1]);
1772 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1773 action
= reinstate_path
;
1774 else if (!strcasecmp(argv
[0], "fail_path"))
1777 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1781 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1783 DMWARN("message: error getting device %s",
1788 r
= action_dev(m
, dev
, action
);
1790 dm_put_device(ti
, dev
);
1793 mutex_unlock(&m
->work_mutex
);
1797 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1798 struct block_device
**bdev
, fmode_t
*mode
)
1800 struct multipath
*m
= ti
->private;
1801 struct pgpath
*current_pgpath
;
1804 current_pgpath
= lockless_dereference(m
->current_pgpath
);
1805 if (!current_pgpath
)
1806 current_pgpath
= choose_pgpath(m
, 0);
1808 if (current_pgpath
) {
1809 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
)) {
1810 *bdev
= current_pgpath
->path
.dev
->bdev
;
1811 *mode
= current_pgpath
->path
.dev
->mode
;
1814 /* pg_init has not started or completed */
1818 /* No path is available */
1819 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1825 if (r
== -ENOTCONN
) {
1826 if (!lockless_dereference(m
->current_pg
)) {
1827 /* Path status changed, redo selection */
1828 (void) choose_pgpath(m
, 0);
1830 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1831 pg_init_all_paths(m
);
1832 dm_table_run_md_queue_async(m
->ti
->table
);
1833 process_queued_io_list(m
);
1837 * Only pass ioctls through if the device sizes match exactly.
1839 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1844 static int multipath_iterate_devices(struct dm_target
*ti
,
1845 iterate_devices_callout_fn fn
, void *data
)
1847 struct multipath
*m
= ti
->private;
1848 struct priority_group
*pg
;
1852 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1853 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1854 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1864 static int pgpath_busy(struct pgpath
*pgpath
)
1866 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1868 return blk_lld_busy(q
);
1872 * We return "busy", only when we can map I/Os but underlying devices
1873 * are busy (so even if we map I/Os now, the I/Os will wait on
1874 * the underlying queue).
1875 * In other words, if we want to kill I/Os or queue them inside us
1876 * due to map unavailability, we don't return "busy". Otherwise,
1877 * dm core won't give us the I/Os and we can't do what we want.
1879 static int multipath_busy(struct dm_target
*ti
)
1881 bool busy
= false, has_active
= false;
1882 struct multipath
*m
= ti
->private;
1883 struct priority_group
*pg
, *next_pg
;
1884 struct pgpath
*pgpath
;
1886 /* pg_init in progress */
1887 if (atomic_read(&m
->pg_init_in_progress
))
1890 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1891 if (!atomic_read(&m
->nr_valid_paths
) && test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1892 return (m
->queue_mode
!= DM_TYPE_MQ_REQUEST_BASED
);
1894 /* Guess which priority_group will be used at next mapping time */
1895 pg
= lockless_dereference(m
->current_pg
);
1896 next_pg
= lockless_dereference(m
->next_pg
);
1897 if (unlikely(!lockless_dereference(m
->current_pgpath
) && next_pg
))
1902 * We don't know which pg will be used at next mapping time.
1903 * We don't call choose_pgpath() here to avoid to trigger
1904 * pg_init just by busy checking.
1905 * So we don't know whether underlying devices we will be using
1906 * at next mapping time are busy or not. Just try mapping.
1912 * If there is one non-busy active path at least, the path selector
1913 * will be able to select it. So we consider such a pg as not busy.
1916 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1917 if (pgpath
->is_active
) {
1919 if (!pgpath_busy(pgpath
)) {
1928 * No active path in this pg, so this pg won't be used and
1929 * the current_pg will be changed at next mapping time.
1930 * We need to try mapping to determine it.
1938 /*-----------------------------------------------------------------
1940 *---------------------------------------------------------------*/
1941 static struct target_type multipath_target
= {
1942 .name
= "multipath",
1943 .version
= {1, 12, 0},
1944 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1945 .module
= THIS_MODULE
,
1946 .ctr
= multipath_ctr
,
1947 .dtr
= multipath_dtr
,
1948 .clone_and_map_rq
= multipath_clone_and_map
,
1949 .release_clone_rq
= multipath_release_clone
,
1950 .rq_end_io
= multipath_end_io
,
1951 .map
= multipath_map_bio
,
1952 .end_io
= multipath_end_io_bio
,
1953 .presuspend
= multipath_presuspend
,
1954 .postsuspend
= multipath_postsuspend
,
1955 .resume
= multipath_resume
,
1956 .status
= multipath_status
,
1957 .message
= multipath_message
,
1958 .prepare_ioctl
= multipath_prepare_ioctl
,
1959 .iterate_devices
= multipath_iterate_devices
,
1960 .busy
= multipath_busy
,
1963 static int __init
dm_multipath_init(void)
1967 r
= dm_register_target(&multipath_target
);
1969 DMERR("request-based register failed %d", r
);
1971 goto bad_register_target
;
1974 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1976 DMERR("failed to create workqueue kmpathd");
1978 goto bad_alloc_kmultipathd
;
1982 * A separate workqueue is used to handle the device handlers
1983 * to avoid overloading existing workqueue. Overloading the
1984 * old workqueue would also create a bottleneck in the
1985 * path of the storage hardware device activation.
1987 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1989 if (!kmpath_handlerd
) {
1990 DMERR("failed to create workqueue kmpath_handlerd");
1992 goto bad_alloc_kmpath_handlerd
;
1997 bad_alloc_kmpath_handlerd
:
1998 destroy_workqueue(kmultipathd
);
1999 bad_alloc_kmultipathd
:
2000 dm_unregister_target(&multipath_target
);
2001 bad_register_target
:
2005 static void __exit
dm_multipath_exit(void)
2007 destroy_workqueue(kmpath_handlerd
);
2008 destroy_workqueue(kmultipathd
);
2010 dm_unregister_target(&multipath_target
);
2013 module_init(dm_multipath_init
);
2014 module_exit(dm_multipath_exit
);
2016 MODULE_DESCRIPTION(DM_NAME
" multipath target");
2017 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2018 MODULE_LICENSE("GPL");