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
);
508 return DM_MAPIO_DELAY_REQUEUE
;
510 clone
->bio
= clone
->biotail
= NULL
;
511 clone
->rq_disk
= bdev
->bd_disk
;
512 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
515 if (pgpath
->pg
->ps
.type
->start_io
)
516 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
519 return DM_MAPIO_REMAPPED
;
522 static void multipath_release_clone(struct request
*clone
)
524 blk_put_request(clone
);
528 * Map cloned bios (bio-based multipath)
530 static int __multipath_map_bio(struct multipath
*m
, struct bio
*bio
, struct dm_mpath_io
*mpio
)
532 size_t nr_bytes
= bio
->bi_iter
.bi_size
;
533 struct pgpath
*pgpath
;
537 /* Do we need to select a new pgpath? */
538 pgpath
= lockless_dereference(m
->current_pgpath
);
539 queue_io
= test_bit(MPATHF_QUEUE_IO
, &m
->flags
);
540 if (!pgpath
|| !queue_io
)
541 pgpath
= choose_pgpath(m
, nr_bytes
);
543 if ((pgpath
&& queue_io
) ||
544 (!pgpath
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))) {
545 /* Queue for the daemon to resubmit */
546 spin_lock_irqsave(&m
->lock
, flags
);
547 bio_list_add(&m
->queued_bios
, bio
);
548 spin_unlock_irqrestore(&m
->lock
, flags
);
549 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
550 if (queue_io
|| test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
551 pg_init_all_paths(m
);
553 queue_work(kmultipathd
, &m
->process_queued_bios
);
554 return DM_MAPIO_SUBMITTED
;
558 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
559 return DM_MAPIO_REQUEUE
;
561 return DM_MAPIO_KILL
;
564 mpio
->pgpath
= pgpath
;
565 mpio
->nr_bytes
= nr_bytes
;
568 bio_set_dev(bio
, pgpath
->path
.dev
->bdev
);
569 bio
->bi_opf
|= REQ_FAILFAST_TRANSPORT
;
571 if (pgpath
->pg
->ps
.type
->start_io
)
572 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
575 return DM_MAPIO_REMAPPED
;
578 static int multipath_map_bio(struct dm_target
*ti
, struct bio
*bio
)
580 struct multipath
*m
= ti
->private;
581 struct dm_mpath_io
*mpio
= NULL
;
583 multipath_init_per_bio_data(bio
, &mpio
, NULL
);
585 return __multipath_map_bio(m
, bio
, mpio
);
588 static void process_queued_io_list(struct multipath
*m
)
590 if (m
->queue_mode
== DM_TYPE_MQ_REQUEST_BASED
)
591 dm_mq_kick_requeue_list(dm_table_get_md(m
->ti
->table
));
592 else if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
593 queue_work(kmultipathd
, &m
->process_queued_bios
);
596 static void process_queued_bios(struct work_struct
*work
)
601 struct bio_list bios
;
602 struct blk_plug plug
;
603 struct multipath
*m
=
604 container_of(work
, struct multipath
, process_queued_bios
);
606 bio_list_init(&bios
);
608 spin_lock_irqsave(&m
->lock
, flags
);
610 if (bio_list_empty(&m
->queued_bios
)) {
611 spin_unlock_irqrestore(&m
->lock
, flags
);
615 bio_list_merge(&bios
, &m
->queued_bios
);
616 bio_list_init(&m
->queued_bios
);
618 spin_unlock_irqrestore(&m
->lock
, flags
);
620 blk_start_plug(&plug
);
621 while ((bio
= bio_list_pop(&bios
))) {
622 r
= __multipath_map_bio(m
, bio
, get_mpio_from_bio(bio
));
625 bio
->bi_status
= BLK_STS_IOERR
;
628 case DM_MAPIO_REQUEUE
:
629 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
632 case DM_MAPIO_REMAPPED
:
633 generic_make_request(bio
);
638 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r
);
641 blk_finish_plug(&plug
);
644 static void assign_bit(bool value
, long nr
, unsigned long *addr
)
653 * If we run out of usable paths, should we queue I/O or error it?
655 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
660 spin_lock_irqsave(&m
->lock
, flags
);
661 assign_bit((save_old_value
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) ||
662 (!save_old_value
&& queue_if_no_path
),
663 MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
664 assign_bit(queue_if_no_path
|| dm_noflush_suspending(m
->ti
),
665 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
666 spin_unlock_irqrestore(&m
->lock
, flags
);
668 if (!queue_if_no_path
) {
669 dm_table_run_md_queue_async(m
->ti
->table
);
670 process_queued_io_list(m
);
677 * An event is triggered whenever a path is taken out of use.
678 * Includes path failure and PG bypass.
680 static void trigger_event(struct work_struct
*work
)
682 struct multipath
*m
=
683 container_of(work
, struct multipath
, trigger_event
);
685 dm_table_event(m
->ti
->table
);
688 /*-----------------------------------------------------------------
689 * Constructor/argument parsing:
690 * <#multipath feature args> [<arg>]*
691 * <#hw_handler args> [hw_handler [<arg>]*]
693 * <initial priority group>
694 * [<selector> <#selector args> [<arg>]*
695 * <#paths> <#per-path selector args>
696 * [<path> [<arg>]* ]+ ]+
697 *---------------------------------------------------------------*/
698 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
699 struct dm_target
*ti
)
702 struct path_selector_type
*pst
;
705 static const struct dm_arg _args
[] = {
706 {0, 1024, "invalid number of path selector args"},
709 pst
= dm_get_path_selector(dm_shift_arg(as
));
711 ti
->error
= "unknown path selector type";
715 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
717 dm_put_path_selector(pst
);
721 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
723 dm_put_path_selector(pst
);
724 ti
->error
= "path selector constructor failed";
729 dm_consume_args(as
, ps_argc
);
734 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
735 struct dm_target
*ti
)
739 struct multipath
*m
= ti
->private;
740 struct request_queue
*q
= NULL
;
741 const char *attached_handler_name
;
743 /* we need at least a path arg */
745 ti
->error
= "no device given";
746 return ERR_PTR(-EINVAL
);
751 return ERR_PTR(-ENOMEM
);
753 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
756 ti
->error
= "error getting device";
760 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) || m
->hw_handler_name
)
761 q
= bdev_get_queue(p
->path
.dev
->bdev
);
763 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
)) {
765 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
766 if (attached_handler_name
) {
768 * Clear any hw_handler_params associated with a
769 * handler that isn't already attached.
771 if (m
->hw_handler_name
&& strcmp(attached_handler_name
, m
->hw_handler_name
)) {
772 kfree(m
->hw_handler_params
);
773 m
->hw_handler_params
= NULL
;
777 * Reset hw_handler_name to match the attached handler
779 * NB. This modifies the table line to show the actual
780 * handler instead of the original table passed in.
782 kfree(m
->hw_handler_name
);
783 m
->hw_handler_name
= attached_handler_name
;
787 if (m
->hw_handler_name
) {
788 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
790 char b
[BDEVNAME_SIZE
];
792 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
793 bdevname(p
->path
.dev
->bdev
, b
));
797 ti
->error
= "error attaching hardware handler";
798 dm_put_device(ti
, p
->path
.dev
);
802 if (m
->hw_handler_params
) {
803 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
805 ti
->error
= "unable to set hardware "
806 "handler parameters";
807 dm_put_device(ti
, p
->path
.dev
);
813 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
815 dm_put_device(ti
, p
->path
.dev
);
826 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
829 static const struct dm_arg _args
[] = {
830 {1, 1024, "invalid number of paths"},
831 {0, 1024, "invalid number of selector args"}
835 unsigned i
, nr_selector_args
, nr_args
;
836 struct priority_group
*pg
;
837 struct dm_target
*ti
= m
->ti
;
841 ti
->error
= "not enough priority group arguments";
842 return ERR_PTR(-EINVAL
);
845 pg
= alloc_priority_group();
847 ti
->error
= "couldn't allocate priority group";
848 return ERR_PTR(-ENOMEM
);
852 r
= parse_path_selector(as
, pg
, ti
);
859 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
863 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
867 nr_args
= 1 + nr_selector_args
;
868 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
869 struct pgpath
*pgpath
;
870 struct dm_arg_set path_args
;
872 if (as
->argc
< nr_args
) {
873 ti
->error
= "not enough path parameters";
878 path_args
.argc
= nr_args
;
879 path_args
.argv
= as
->argv
;
881 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
882 if (IS_ERR(pgpath
)) {
888 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
889 dm_consume_args(as
, nr_args
);
895 free_priority_group(pg
, ti
);
899 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
903 struct dm_target
*ti
= m
->ti
;
905 static const struct dm_arg _args
[] = {
906 {0, 1024, "invalid number of hardware handler args"},
909 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
915 if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
916 dm_consume_args(as
, hw_argc
);
917 DMERR("bio-based multipath doesn't allow hardware handler args");
921 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
922 if (!m
->hw_handler_name
)
929 for (i
= 0; i
<= hw_argc
- 2; i
++)
930 len
+= strlen(as
->argv
[i
]) + 1;
931 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
933 ti
->error
= "memory allocation failed";
937 j
= sprintf(p
, "%d", hw_argc
- 1);
938 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
939 j
= sprintf(p
, "%s", as
->argv
[i
]);
941 dm_consume_args(as
, hw_argc
- 1);
945 kfree(m
->hw_handler_name
);
946 m
->hw_handler_name
= NULL
;
950 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
954 struct dm_target
*ti
= m
->ti
;
955 const char *arg_name
;
957 static const struct dm_arg _args
[] = {
958 {0, 8, "invalid number of feature args"},
959 {1, 50, "pg_init_retries must be between 1 and 50"},
960 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
963 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
971 arg_name
= dm_shift_arg(as
);
974 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
975 r
= queue_if_no_path(m
, true, false);
979 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
980 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
984 if (!strcasecmp(arg_name
, "pg_init_retries") &&
986 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
991 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
993 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
998 if (!strcasecmp(arg_name
, "queue_mode") &&
1000 const char *queue_mode_name
= dm_shift_arg(as
);
1002 if (!strcasecmp(queue_mode_name
, "bio"))
1003 m
->queue_mode
= DM_TYPE_BIO_BASED
;
1004 else if (!strcasecmp(queue_mode_name
, "rq"))
1005 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
1006 else if (!strcasecmp(queue_mode_name
, "mq"))
1007 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
1009 ti
->error
= "Unknown 'queue_mode' requested";
1016 ti
->error
= "Unrecognised multipath feature request";
1018 } while (argc
&& !r
);
1023 static int multipath_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
1025 /* target arguments */
1026 static const struct dm_arg _args
[] = {
1027 {0, 1024, "invalid number of priority groups"},
1028 {0, 1024, "invalid initial priority group number"},
1032 struct multipath
*m
;
1033 struct dm_arg_set as
;
1034 unsigned pg_count
= 0;
1035 unsigned next_pg_num
;
1040 m
= alloc_multipath(ti
);
1042 ti
->error
= "can't allocate multipath";
1046 r
= parse_features(&as
, m
);
1050 r
= alloc_multipath_stage2(ti
, m
);
1054 r
= parse_hw_handler(&as
, m
);
1058 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
1062 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
1066 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
1067 (m
->nr_priority_groups
&& !next_pg_num
)) {
1068 ti
->error
= "invalid initial priority group";
1073 /* parse the priority groups */
1075 struct priority_group
*pg
;
1076 unsigned nr_valid_paths
= atomic_read(&m
->nr_valid_paths
);
1078 pg
= parse_priority_group(&as
, m
);
1084 nr_valid_paths
+= pg
->nr_pgpaths
;
1085 atomic_set(&m
->nr_valid_paths
, nr_valid_paths
);
1087 list_add_tail(&pg
->list
, &m
->priority_groups
);
1089 pg
->pg_num
= pg_count
;
1094 if (pg_count
!= m
->nr_priority_groups
) {
1095 ti
->error
= "priority group count mismatch";
1100 ti
->num_flush_bios
= 1;
1101 ti
->num_discard_bios
= 1;
1102 ti
->num_write_same_bios
= 1;
1103 ti
->num_write_zeroes_bios
= 1;
1104 if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
1105 ti
->per_io_data_size
= multipath_per_bio_data_size();
1107 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
1116 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
1121 prepare_to_wait(&m
->pg_init_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1123 if (!atomic_read(&m
->pg_init_in_progress
))
1128 finish_wait(&m
->pg_init_wait
, &wait
);
1131 static void flush_multipath_work(struct multipath
*m
)
1133 set_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1134 smp_mb__after_atomic();
1136 flush_workqueue(kmpath_handlerd
);
1137 multipath_wait_for_pg_init_completion(m
);
1138 flush_workqueue(kmultipathd
);
1139 flush_work(&m
->trigger_event
);
1141 clear_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1142 smp_mb__after_atomic();
1145 static void multipath_dtr(struct dm_target
*ti
)
1147 struct multipath
*m
= ti
->private;
1149 flush_multipath_work(m
);
1154 * Take a path out of use.
1156 static int fail_path(struct pgpath
*pgpath
)
1158 unsigned long flags
;
1159 struct multipath
*m
= pgpath
->pg
->m
;
1161 spin_lock_irqsave(&m
->lock
, flags
);
1163 if (!pgpath
->is_active
)
1166 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1168 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1169 pgpath
->is_active
= false;
1170 pgpath
->fail_count
++;
1172 atomic_dec(&m
->nr_valid_paths
);
1174 if (pgpath
== m
->current_pgpath
)
1175 m
->current_pgpath
= NULL
;
1177 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1178 pgpath
->path
.dev
->name
, atomic_read(&m
->nr_valid_paths
));
1180 schedule_work(&m
->trigger_event
);
1183 spin_unlock_irqrestore(&m
->lock
, flags
);
1189 * Reinstate a previously-failed path
1191 static int reinstate_path(struct pgpath
*pgpath
)
1193 int r
= 0, run_queue
= 0;
1194 unsigned long flags
;
1195 struct multipath
*m
= pgpath
->pg
->m
;
1196 unsigned nr_valid_paths
;
1198 spin_lock_irqsave(&m
->lock
, flags
);
1200 if (pgpath
->is_active
)
1203 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1205 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1209 pgpath
->is_active
= true;
1211 nr_valid_paths
= atomic_inc_return(&m
->nr_valid_paths
);
1212 if (nr_valid_paths
== 1) {
1213 m
->current_pgpath
= NULL
;
1215 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1216 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1217 atomic_inc(&m
->pg_init_in_progress
);
1220 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1221 pgpath
->path
.dev
->name
, nr_valid_paths
);
1223 schedule_work(&m
->trigger_event
);
1226 spin_unlock_irqrestore(&m
->lock
, flags
);
1228 dm_table_run_md_queue_async(m
->ti
->table
);
1229 process_queued_io_list(m
);
1236 * Fail or reinstate all paths that match the provided struct dm_dev.
1238 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1242 struct pgpath
*pgpath
;
1243 struct priority_group
*pg
;
1245 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1246 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1247 if (pgpath
->path
.dev
== dev
)
1256 * Temporarily try to avoid having to use the specified PG
1258 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1261 unsigned long flags
;
1263 spin_lock_irqsave(&m
->lock
, flags
);
1265 pg
->bypassed
= bypassed
;
1266 m
->current_pgpath
= NULL
;
1267 m
->current_pg
= NULL
;
1269 spin_unlock_irqrestore(&m
->lock
, flags
);
1271 schedule_work(&m
->trigger_event
);
1275 * Switch to using the specified PG from the next I/O that gets mapped
1277 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1279 struct priority_group
*pg
;
1281 unsigned long flags
;
1284 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1285 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1286 DMWARN("invalid PG number supplied to switch_pg_num");
1290 spin_lock_irqsave(&m
->lock
, flags
);
1291 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1292 pg
->bypassed
= false;
1296 m
->current_pgpath
= NULL
;
1297 m
->current_pg
= NULL
;
1300 spin_unlock_irqrestore(&m
->lock
, flags
);
1302 schedule_work(&m
->trigger_event
);
1307 * Set/clear bypassed status of a PG.
1308 * PGs are numbered upwards from 1 in the order they were declared.
1310 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1312 struct priority_group
*pg
;
1316 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1317 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1318 DMWARN("invalid PG number supplied to bypass_pg");
1322 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1327 bypass_pg(m
, pg
, bypassed
);
1332 * Should we retry pg_init immediately?
1334 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1336 unsigned long flags
;
1337 bool limit_reached
= false;
1339 spin_lock_irqsave(&m
->lock
, flags
);
1341 if (atomic_read(&m
->pg_init_count
) <= m
->pg_init_retries
&&
1342 !test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
1343 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
1345 limit_reached
= true;
1347 spin_unlock_irqrestore(&m
->lock
, flags
);
1349 return limit_reached
;
1352 static void pg_init_done(void *data
, int errors
)
1354 struct pgpath
*pgpath
= data
;
1355 struct priority_group
*pg
= pgpath
->pg
;
1356 struct multipath
*m
= pg
->m
;
1357 unsigned long flags
;
1358 bool delay_retry
= false;
1360 /* device or driver problems */
1365 if (!m
->hw_handler_name
) {
1369 DMERR("Could not failover the device: Handler scsi_dh_%s "
1370 "Error %d.", m
->hw_handler_name
, errors
);
1372 * Fail path for now, so we do not ping pong
1376 case SCSI_DH_DEV_TEMP_BUSY
:
1378 * Probably doing something like FW upgrade on the
1379 * controller so try the other pg.
1381 bypass_pg(m
, pg
, true);
1384 /* Wait before retrying. */
1387 case SCSI_DH_IMM_RETRY
:
1388 case SCSI_DH_RES_TEMP_UNAVAIL
:
1389 if (pg_init_limit_reached(m
, pgpath
))
1393 case SCSI_DH_DEV_OFFLINED
:
1396 * We probably do not want to fail the path for a device
1397 * error, but this is what the old dm did. In future
1398 * patches we can do more advanced handling.
1403 spin_lock_irqsave(&m
->lock
, flags
);
1405 if (pgpath
== m
->current_pgpath
) {
1406 DMERR("Could not failover device. Error %d.", errors
);
1407 m
->current_pgpath
= NULL
;
1408 m
->current_pg
= NULL
;
1410 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1411 pg
->bypassed
= false;
1413 if (atomic_dec_return(&m
->pg_init_in_progress
) > 0)
1414 /* Activations of other paths are still on going */
1417 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
1419 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1421 clear_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1423 if (__pg_init_all_paths(m
))
1426 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
1428 process_queued_io_list(m
);
1431 * Wake up any thread waiting to suspend.
1433 wake_up(&m
->pg_init_wait
);
1436 spin_unlock_irqrestore(&m
->lock
, flags
);
1439 static void activate_or_offline_path(struct pgpath
*pgpath
)
1441 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1443 if (pgpath
->is_active
&& !blk_queue_dying(q
))
1444 scsi_dh_activate(q
, pg_init_done
, pgpath
);
1446 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1449 static void activate_path_work(struct work_struct
*work
)
1451 struct pgpath
*pgpath
=
1452 container_of(work
, struct pgpath
, activate_path
.work
);
1454 activate_or_offline_path(pgpath
);
1457 static int noretry_error(blk_status_t error
)
1460 case BLK_STS_NOTSUPP
:
1462 case BLK_STS_TARGET
:
1464 case BLK_STS_MEDIUM
:
1468 /* Anything else could be a path failure, so should be retried */
1472 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1473 blk_status_t error
, union map_info
*map_context
)
1475 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1476 struct pgpath
*pgpath
= mpio
->pgpath
;
1477 int r
= DM_ENDIO_DONE
;
1480 * We don't queue any clone request inside the multipath target
1481 * during end I/O handling, since those clone requests don't have
1482 * bio clones. If we queue them inside the multipath target,
1483 * we need to make bio clones, that requires memory allocation.
1484 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1485 * don't have bio clones.)
1486 * Instead of queueing the clone request here, we queue the original
1487 * request into dm core, which will remake a clone request and
1488 * clone bios for it and resubmit it later.
1490 if (error
&& !noretry_error(error
)) {
1491 struct multipath
*m
= ti
->private;
1493 r
= DM_ENDIO_REQUEUE
;
1498 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1499 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1500 if (error
== BLK_STS_IOERR
)
1502 /* complete with the original error */
1508 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1510 if (ps
->type
->end_io
)
1511 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1517 static int multipath_end_io_bio(struct dm_target
*ti
, struct bio
*clone
,
1518 blk_status_t
*error
)
1520 struct multipath
*m
= ti
->private;
1521 struct dm_mpath_io
*mpio
= get_mpio_from_bio(clone
);
1522 struct pgpath
*pgpath
= mpio
->pgpath
;
1523 unsigned long flags
;
1524 int r
= DM_ENDIO_DONE
;
1526 if (!*error
|| noretry_error(*error
))
1532 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1533 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1535 *error
= BLK_STS_IOERR
;
1539 /* Queue for the daemon to resubmit */
1540 dm_bio_restore(get_bio_details_from_bio(clone
), clone
);
1542 spin_lock_irqsave(&m
->lock
, flags
);
1543 bio_list_add(&m
->queued_bios
, clone
);
1544 spin_unlock_irqrestore(&m
->lock
, flags
);
1545 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
1546 queue_work(kmultipathd
, &m
->process_queued_bios
);
1548 r
= DM_ENDIO_INCOMPLETE
;
1551 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1553 if (ps
->type
->end_io
)
1554 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1561 * Suspend can't complete until all the I/O is processed so if
1562 * the last path fails we must error any remaining I/O.
1563 * Note that if the freeze_bdev fails while suspending, the
1564 * queue_if_no_path state is lost - userspace should reset it.
1566 static void multipath_presuspend(struct dm_target
*ti
)
1568 struct multipath
*m
= ti
->private;
1570 queue_if_no_path(m
, false, true);
1573 static void multipath_postsuspend(struct dm_target
*ti
)
1575 struct multipath
*m
= ti
->private;
1577 mutex_lock(&m
->work_mutex
);
1578 flush_multipath_work(m
);
1579 mutex_unlock(&m
->work_mutex
);
1583 * Restore the queue_if_no_path setting.
1585 static void multipath_resume(struct dm_target
*ti
)
1587 struct multipath
*m
= ti
->private;
1588 unsigned long flags
;
1590 spin_lock_irqsave(&m
->lock
, flags
);
1591 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
),
1592 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
1593 spin_unlock_irqrestore(&m
->lock
, flags
);
1597 * Info output has the following format:
1598 * num_multipath_feature_args [multipath_feature_args]*
1599 * num_handler_status_args [handler_status_args]*
1600 * num_groups init_group_number
1601 * [A|D|E num_ps_status_args [ps_status_args]*
1602 * num_paths num_selector_args
1603 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1605 * Table output has the following format (identical to the constructor string):
1606 * num_feature_args [features_args]*
1607 * num_handler_args hw_handler [hw_handler_args]*
1608 * num_groups init_group_number
1609 * [priority selector-name num_ps_args [ps_args]*
1610 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1612 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1613 unsigned status_flags
, char *result
, unsigned maxlen
)
1616 unsigned long flags
;
1617 struct multipath
*m
= ti
->private;
1618 struct priority_group
*pg
;
1623 spin_lock_irqsave(&m
->lock
, flags
);
1626 if (type
== STATUSTYPE_INFO
)
1627 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO
, &m
->flags
),
1628 atomic_read(&m
->pg_init_count
));
1630 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) +
1631 (m
->pg_init_retries
> 0) * 2 +
1632 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1633 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) +
1634 (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) * 2);
1636 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1637 DMEMIT("queue_if_no_path ");
1638 if (m
->pg_init_retries
)
1639 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1640 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1641 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1642 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
))
1643 DMEMIT("retain_attached_hw_handler ");
1644 if (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) {
1645 switch(m
->queue_mode
) {
1646 case DM_TYPE_BIO_BASED
:
1647 DMEMIT("queue_mode bio ");
1649 case DM_TYPE_MQ_REQUEST_BASED
:
1650 DMEMIT("queue_mode mq ");
1659 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1662 DMEMIT("1 %s ", m
->hw_handler_name
);
1664 DMEMIT("%u ", m
->nr_priority_groups
);
1667 pg_num
= m
->next_pg
->pg_num
;
1668 else if (m
->current_pg
)
1669 pg_num
= m
->current_pg
->pg_num
;
1671 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1673 DMEMIT("%u ", pg_num
);
1676 case STATUSTYPE_INFO
:
1677 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1679 state
= 'D'; /* Disabled */
1680 else if (pg
== m
->current_pg
)
1681 state
= 'A'; /* Currently Active */
1683 state
= 'E'; /* Enabled */
1685 DMEMIT("%c ", state
);
1687 if (pg
->ps
.type
->status
)
1688 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1694 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1695 pg
->ps
.type
->info_args
);
1697 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1698 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1699 p
->is_active
? "A" : "F",
1701 if (pg
->ps
.type
->status
)
1702 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1703 &p
->path
, type
, result
+ sz
,
1709 case STATUSTYPE_TABLE
:
1710 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1711 DMEMIT("%s ", pg
->ps
.type
->name
);
1713 if (pg
->ps
.type
->status
)
1714 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1720 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1721 pg
->ps
.type
->table_args
);
1723 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1724 DMEMIT("%s ", p
->path
.dev
->name
);
1725 if (pg
->ps
.type
->status
)
1726 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1727 &p
->path
, type
, result
+ sz
,
1734 spin_unlock_irqrestore(&m
->lock
, flags
);
1737 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1741 struct multipath
*m
= ti
->private;
1744 mutex_lock(&m
->work_mutex
);
1746 if (dm_suspended(ti
)) {
1752 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1753 r
= queue_if_no_path(m
, true, false);
1755 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1756 r
= queue_if_no_path(m
, false, false);
1762 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1766 if (!strcasecmp(argv
[0], "disable_group")) {
1767 r
= bypass_pg_num(m
, argv
[1], true);
1769 } else if (!strcasecmp(argv
[0], "enable_group")) {
1770 r
= bypass_pg_num(m
, argv
[1], false);
1772 } else if (!strcasecmp(argv
[0], "switch_group")) {
1773 r
= switch_pg_num(m
, argv
[1]);
1775 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1776 action
= reinstate_path
;
1777 else if (!strcasecmp(argv
[0], "fail_path"))
1780 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1784 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1786 DMWARN("message: error getting device %s",
1791 r
= action_dev(m
, dev
, action
);
1793 dm_put_device(ti
, dev
);
1796 mutex_unlock(&m
->work_mutex
);
1800 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1801 struct block_device
**bdev
, fmode_t
*mode
)
1803 struct multipath
*m
= ti
->private;
1804 struct pgpath
*current_pgpath
;
1807 current_pgpath
= lockless_dereference(m
->current_pgpath
);
1808 if (!current_pgpath
)
1809 current_pgpath
= choose_pgpath(m
, 0);
1811 if (current_pgpath
) {
1812 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
)) {
1813 *bdev
= current_pgpath
->path
.dev
->bdev
;
1814 *mode
= current_pgpath
->path
.dev
->mode
;
1817 /* pg_init has not started or completed */
1821 /* No path is available */
1822 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1828 if (r
== -ENOTCONN
) {
1829 if (!lockless_dereference(m
->current_pg
)) {
1830 /* Path status changed, redo selection */
1831 (void) choose_pgpath(m
, 0);
1833 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1834 pg_init_all_paths(m
);
1835 dm_table_run_md_queue_async(m
->ti
->table
);
1836 process_queued_io_list(m
);
1840 * Only pass ioctls through if the device sizes match exactly.
1842 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1847 static int multipath_iterate_devices(struct dm_target
*ti
,
1848 iterate_devices_callout_fn fn
, void *data
)
1850 struct multipath
*m
= ti
->private;
1851 struct priority_group
*pg
;
1855 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1856 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1857 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1867 static int pgpath_busy(struct pgpath
*pgpath
)
1869 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1871 return blk_lld_busy(q
);
1875 * We return "busy", only when we can map I/Os but underlying devices
1876 * are busy (so even if we map I/Os now, the I/Os will wait on
1877 * the underlying queue).
1878 * In other words, if we want to kill I/Os or queue them inside us
1879 * due to map unavailability, we don't return "busy". Otherwise,
1880 * dm core won't give us the I/Os and we can't do what we want.
1882 static int multipath_busy(struct dm_target
*ti
)
1884 bool busy
= false, has_active
= false;
1885 struct multipath
*m
= ti
->private;
1886 struct priority_group
*pg
, *next_pg
;
1887 struct pgpath
*pgpath
;
1889 /* pg_init in progress */
1890 if (atomic_read(&m
->pg_init_in_progress
))
1893 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1894 if (!atomic_read(&m
->nr_valid_paths
) && test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1895 return (m
->queue_mode
!= DM_TYPE_MQ_REQUEST_BASED
);
1897 /* Guess which priority_group will be used at next mapping time */
1898 pg
= lockless_dereference(m
->current_pg
);
1899 next_pg
= lockless_dereference(m
->next_pg
);
1900 if (unlikely(!lockless_dereference(m
->current_pgpath
) && next_pg
))
1905 * We don't know which pg will be used at next mapping time.
1906 * We don't call choose_pgpath() here to avoid to trigger
1907 * pg_init just by busy checking.
1908 * So we don't know whether underlying devices we will be using
1909 * at next mapping time are busy or not. Just try mapping.
1915 * If there is one non-busy active path at least, the path selector
1916 * will be able to select it. So we consider such a pg as not busy.
1919 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1920 if (pgpath
->is_active
) {
1922 if (!pgpath_busy(pgpath
)) {
1931 * No active path in this pg, so this pg won't be used and
1932 * the current_pg will be changed at next mapping time.
1933 * We need to try mapping to determine it.
1941 /*-----------------------------------------------------------------
1943 *---------------------------------------------------------------*/
1944 static struct target_type multipath_target
= {
1945 .name
= "multipath",
1946 .version
= {1, 12, 0},
1947 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1948 .module
= THIS_MODULE
,
1949 .ctr
= multipath_ctr
,
1950 .dtr
= multipath_dtr
,
1951 .clone_and_map_rq
= multipath_clone_and_map
,
1952 .release_clone_rq
= multipath_release_clone
,
1953 .rq_end_io
= multipath_end_io
,
1954 .map
= multipath_map_bio
,
1955 .end_io
= multipath_end_io_bio
,
1956 .presuspend
= multipath_presuspend
,
1957 .postsuspend
= multipath_postsuspend
,
1958 .resume
= multipath_resume
,
1959 .status
= multipath_status
,
1960 .message
= multipath_message
,
1961 .prepare_ioctl
= multipath_prepare_ioctl
,
1962 .iterate_devices
= multipath_iterate_devices
,
1963 .busy
= multipath_busy
,
1966 static int __init
dm_multipath_init(void)
1970 r
= dm_register_target(&multipath_target
);
1972 DMERR("request-based register failed %d", r
);
1974 goto bad_register_target
;
1977 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1979 DMERR("failed to create workqueue kmpathd");
1981 goto bad_alloc_kmultipathd
;
1985 * A separate workqueue is used to handle the device handlers
1986 * to avoid overloading existing workqueue. Overloading the
1987 * old workqueue would also create a bottleneck in the
1988 * path of the storage hardware device activation.
1990 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1992 if (!kmpath_handlerd
) {
1993 DMERR("failed to create workqueue kmpath_handlerd");
1995 goto bad_alloc_kmpath_handlerd
;
2000 bad_alloc_kmpath_handlerd
:
2001 destroy_workqueue(kmultipathd
);
2002 bad_alloc_kmultipathd
:
2003 dm_unregister_target(&multipath_target
);
2004 bad_register_target
:
2008 static void __exit
dm_multipath_exit(void)
2010 destroy_workqueue(kmpath_handlerd
);
2011 destroy_workqueue(kmultipathd
);
2013 dm_unregister_target(&multipath_target
);
2016 module_init(dm_multipath_init
);
2017 module_exit(dm_multipath_exit
);
2019 MODULE_DESCRIPTION(DM_NAME
" multipath target");
2020 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2021 MODULE_LICENSE("GPL");