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
->bi_bdev
= 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
);
637 blk_finish_plug(&plug
);
640 static void assign_bit(bool value
, long nr
, unsigned long *addr
)
649 * If we run out of usable paths, should we queue I/O or error it?
651 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
656 spin_lock_irqsave(&m
->lock
, flags
);
657 assign_bit((save_old_value
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) ||
658 (!save_old_value
&& queue_if_no_path
),
659 MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
);
660 assign_bit(queue_if_no_path
|| dm_noflush_suspending(m
->ti
),
661 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
662 spin_unlock_irqrestore(&m
->lock
, flags
);
664 if (!queue_if_no_path
) {
665 dm_table_run_md_queue_async(m
->ti
->table
);
666 process_queued_io_list(m
);
673 * An event is triggered whenever a path is taken out of use.
674 * Includes path failure and PG bypass.
676 static void trigger_event(struct work_struct
*work
)
678 struct multipath
*m
=
679 container_of(work
, struct multipath
, trigger_event
);
681 dm_table_event(m
->ti
->table
);
684 /*-----------------------------------------------------------------
685 * Constructor/argument parsing:
686 * <#multipath feature args> [<arg>]*
687 * <#hw_handler args> [hw_handler [<arg>]*]
689 * <initial priority group>
690 * [<selector> <#selector args> [<arg>]*
691 * <#paths> <#per-path selector args>
692 * [<path> [<arg>]* ]+ ]+
693 *---------------------------------------------------------------*/
694 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
695 struct dm_target
*ti
)
698 struct path_selector_type
*pst
;
701 static struct dm_arg _args
[] = {
702 {0, 1024, "invalid number of path selector args"},
705 pst
= dm_get_path_selector(dm_shift_arg(as
));
707 ti
->error
= "unknown path selector type";
711 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
713 dm_put_path_selector(pst
);
717 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
719 dm_put_path_selector(pst
);
720 ti
->error
= "path selector constructor failed";
725 dm_consume_args(as
, ps_argc
);
730 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
731 struct dm_target
*ti
)
735 struct multipath
*m
= ti
->private;
736 struct request_queue
*q
= NULL
;
737 const char *attached_handler_name
;
739 /* we need at least a path arg */
741 ti
->error
= "no device given";
742 return ERR_PTR(-EINVAL
);
747 return ERR_PTR(-ENOMEM
);
749 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
752 ti
->error
= "error getting device";
756 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) || m
->hw_handler_name
)
757 q
= bdev_get_queue(p
->path
.dev
->bdev
);
759 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
)) {
761 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
762 if (attached_handler_name
) {
764 * Clear any hw_handler_params associated with a
765 * handler that isn't already attached.
767 if (m
->hw_handler_name
&& strcmp(attached_handler_name
, m
->hw_handler_name
)) {
768 kfree(m
->hw_handler_params
);
769 m
->hw_handler_params
= NULL
;
773 * Reset hw_handler_name to match the attached handler
775 * NB. This modifies the table line to show the actual
776 * handler instead of the original table passed in.
778 kfree(m
->hw_handler_name
);
779 m
->hw_handler_name
= attached_handler_name
;
783 if (m
->hw_handler_name
) {
784 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
786 char b
[BDEVNAME_SIZE
];
788 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
789 bdevname(p
->path
.dev
->bdev
, b
));
793 ti
->error
= "error attaching hardware handler";
794 dm_put_device(ti
, p
->path
.dev
);
798 if (m
->hw_handler_params
) {
799 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
801 ti
->error
= "unable to set hardware "
802 "handler parameters";
803 dm_put_device(ti
, p
->path
.dev
);
809 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
811 dm_put_device(ti
, p
->path
.dev
);
822 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
825 static struct dm_arg _args
[] = {
826 {1, 1024, "invalid number of paths"},
827 {0, 1024, "invalid number of selector args"}
831 unsigned i
, nr_selector_args
, nr_args
;
832 struct priority_group
*pg
;
833 struct dm_target
*ti
= m
->ti
;
837 ti
->error
= "not enough priority group arguments";
838 return ERR_PTR(-EINVAL
);
841 pg
= alloc_priority_group();
843 ti
->error
= "couldn't allocate priority group";
844 return ERR_PTR(-ENOMEM
);
848 r
= parse_path_selector(as
, pg
, ti
);
855 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
859 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
863 nr_args
= 1 + nr_selector_args
;
864 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
865 struct pgpath
*pgpath
;
866 struct dm_arg_set path_args
;
868 if (as
->argc
< nr_args
) {
869 ti
->error
= "not enough path parameters";
874 path_args
.argc
= nr_args
;
875 path_args
.argv
= as
->argv
;
877 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
878 if (IS_ERR(pgpath
)) {
884 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
885 dm_consume_args(as
, nr_args
);
891 free_priority_group(pg
, ti
);
895 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
899 struct dm_target
*ti
= m
->ti
;
901 static struct dm_arg _args
[] = {
902 {0, 1024, "invalid number of hardware handler args"},
905 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
911 if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
912 dm_consume_args(as
, hw_argc
);
913 DMERR("bio-based multipath doesn't allow hardware handler args");
917 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
918 if (!m
->hw_handler_name
)
925 for (i
= 0; i
<= hw_argc
- 2; i
++)
926 len
+= strlen(as
->argv
[i
]) + 1;
927 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
929 ti
->error
= "memory allocation failed";
933 j
= sprintf(p
, "%d", hw_argc
- 1);
934 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
935 j
= sprintf(p
, "%s", as
->argv
[i
]);
937 dm_consume_args(as
, hw_argc
- 1);
941 kfree(m
->hw_handler_name
);
942 m
->hw_handler_name
= NULL
;
946 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
950 struct dm_target
*ti
= m
->ti
;
951 const char *arg_name
;
953 static struct dm_arg _args
[] = {
954 {0, 8, "invalid number of feature args"},
955 {1, 50, "pg_init_retries must be between 1 and 50"},
956 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
959 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
967 arg_name
= dm_shift_arg(as
);
970 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
971 r
= queue_if_no_path(m
, true, false);
975 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
976 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
980 if (!strcasecmp(arg_name
, "pg_init_retries") &&
982 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
987 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
989 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
994 if (!strcasecmp(arg_name
, "queue_mode") &&
996 const char *queue_mode_name
= dm_shift_arg(as
);
998 if (!strcasecmp(queue_mode_name
, "bio"))
999 m
->queue_mode
= DM_TYPE_BIO_BASED
;
1000 else if (!strcasecmp(queue_mode_name
, "rq"))
1001 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
1002 else if (!strcasecmp(queue_mode_name
, "mq"))
1003 m
->queue_mode
= DM_TYPE_MQ_REQUEST_BASED
;
1005 ti
->error
= "Unknown 'queue_mode' requested";
1012 ti
->error
= "Unrecognised multipath feature request";
1014 } while (argc
&& !r
);
1019 static int multipath_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
1021 /* target arguments */
1022 static struct dm_arg _args
[] = {
1023 {0, 1024, "invalid number of priority groups"},
1024 {0, 1024, "invalid initial priority group number"},
1028 struct multipath
*m
;
1029 struct dm_arg_set as
;
1030 unsigned pg_count
= 0;
1031 unsigned next_pg_num
;
1036 m
= alloc_multipath(ti
);
1038 ti
->error
= "can't allocate multipath";
1042 r
= parse_features(&as
, m
);
1046 r
= alloc_multipath_stage2(ti
, m
);
1050 r
= parse_hw_handler(&as
, m
);
1054 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
1058 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
1062 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
1063 (m
->nr_priority_groups
&& !next_pg_num
)) {
1064 ti
->error
= "invalid initial priority group";
1069 /* parse the priority groups */
1071 struct priority_group
*pg
;
1072 unsigned nr_valid_paths
= atomic_read(&m
->nr_valid_paths
);
1074 pg
= parse_priority_group(&as
, m
);
1080 nr_valid_paths
+= pg
->nr_pgpaths
;
1081 atomic_set(&m
->nr_valid_paths
, nr_valid_paths
);
1083 list_add_tail(&pg
->list
, &m
->priority_groups
);
1085 pg
->pg_num
= pg_count
;
1090 if (pg_count
!= m
->nr_priority_groups
) {
1091 ti
->error
= "priority group count mismatch";
1096 ti
->num_flush_bios
= 1;
1097 ti
->num_discard_bios
= 1;
1098 ti
->num_write_same_bios
= 1;
1099 ti
->num_write_zeroes_bios
= 1;
1100 if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
1101 ti
->per_io_data_size
= multipath_per_bio_data_size();
1103 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
1112 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
1117 prepare_to_wait(&m
->pg_init_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1119 if (!atomic_read(&m
->pg_init_in_progress
))
1124 finish_wait(&m
->pg_init_wait
, &wait
);
1127 static void flush_multipath_work(struct multipath
*m
)
1129 set_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1130 smp_mb__after_atomic();
1132 flush_workqueue(kmpath_handlerd
);
1133 multipath_wait_for_pg_init_completion(m
);
1134 flush_workqueue(kmultipathd
);
1135 flush_work(&m
->trigger_event
);
1137 clear_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1138 smp_mb__after_atomic();
1141 static void multipath_dtr(struct dm_target
*ti
)
1143 struct multipath
*m
= ti
->private;
1145 flush_multipath_work(m
);
1150 * Take a path out of use.
1152 static int fail_path(struct pgpath
*pgpath
)
1154 unsigned long flags
;
1155 struct multipath
*m
= pgpath
->pg
->m
;
1157 spin_lock_irqsave(&m
->lock
, flags
);
1159 if (!pgpath
->is_active
)
1162 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1164 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1165 pgpath
->is_active
= false;
1166 pgpath
->fail_count
++;
1168 atomic_dec(&m
->nr_valid_paths
);
1170 if (pgpath
== m
->current_pgpath
)
1171 m
->current_pgpath
= NULL
;
1173 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1174 pgpath
->path
.dev
->name
, atomic_read(&m
->nr_valid_paths
));
1176 schedule_work(&m
->trigger_event
);
1179 spin_unlock_irqrestore(&m
->lock
, flags
);
1185 * Reinstate a previously-failed path
1187 static int reinstate_path(struct pgpath
*pgpath
)
1189 int r
= 0, run_queue
= 0;
1190 unsigned long flags
;
1191 struct multipath
*m
= pgpath
->pg
->m
;
1192 unsigned nr_valid_paths
;
1194 spin_lock_irqsave(&m
->lock
, flags
);
1196 if (pgpath
->is_active
)
1199 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1201 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1205 pgpath
->is_active
= true;
1207 nr_valid_paths
= atomic_inc_return(&m
->nr_valid_paths
);
1208 if (nr_valid_paths
== 1) {
1209 m
->current_pgpath
= NULL
;
1211 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1212 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1213 atomic_inc(&m
->pg_init_in_progress
);
1216 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1217 pgpath
->path
.dev
->name
, nr_valid_paths
);
1219 schedule_work(&m
->trigger_event
);
1222 spin_unlock_irqrestore(&m
->lock
, flags
);
1224 dm_table_run_md_queue_async(m
->ti
->table
);
1225 process_queued_io_list(m
);
1232 * Fail or reinstate all paths that match the provided struct dm_dev.
1234 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1238 struct pgpath
*pgpath
;
1239 struct priority_group
*pg
;
1241 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1242 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1243 if (pgpath
->path
.dev
== dev
)
1252 * Temporarily try to avoid having to use the specified PG
1254 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1257 unsigned long flags
;
1259 spin_lock_irqsave(&m
->lock
, flags
);
1261 pg
->bypassed
= bypassed
;
1262 m
->current_pgpath
= NULL
;
1263 m
->current_pg
= NULL
;
1265 spin_unlock_irqrestore(&m
->lock
, flags
);
1267 schedule_work(&m
->trigger_event
);
1271 * Switch to using the specified PG from the next I/O that gets mapped
1273 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1275 struct priority_group
*pg
;
1277 unsigned long flags
;
1280 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1281 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1282 DMWARN("invalid PG number supplied to switch_pg_num");
1286 spin_lock_irqsave(&m
->lock
, flags
);
1287 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1288 pg
->bypassed
= false;
1292 m
->current_pgpath
= NULL
;
1293 m
->current_pg
= NULL
;
1296 spin_unlock_irqrestore(&m
->lock
, flags
);
1298 schedule_work(&m
->trigger_event
);
1303 * Set/clear bypassed status of a PG.
1304 * PGs are numbered upwards from 1 in the order they were declared.
1306 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1308 struct priority_group
*pg
;
1312 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1313 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1314 DMWARN("invalid PG number supplied to bypass_pg");
1318 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1323 bypass_pg(m
, pg
, bypassed
);
1328 * Should we retry pg_init immediately?
1330 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1332 unsigned long flags
;
1333 bool limit_reached
= false;
1335 spin_lock_irqsave(&m
->lock
, flags
);
1337 if (atomic_read(&m
->pg_init_count
) <= m
->pg_init_retries
&&
1338 !test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
1339 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
1341 limit_reached
= true;
1343 spin_unlock_irqrestore(&m
->lock
, flags
);
1345 return limit_reached
;
1348 static void pg_init_done(void *data
, int errors
)
1350 struct pgpath
*pgpath
= data
;
1351 struct priority_group
*pg
= pgpath
->pg
;
1352 struct multipath
*m
= pg
->m
;
1353 unsigned long flags
;
1354 bool delay_retry
= false;
1356 /* device or driver problems */
1361 if (!m
->hw_handler_name
) {
1365 DMERR("Could not failover the device: Handler scsi_dh_%s "
1366 "Error %d.", m
->hw_handler_name
, errors
);
1368 * Fail path for now, so we do not ping pong
1372 case SCSI_DH_DEV_TEMP_BUSY
:
1374 * Probably doing something like FW upgrade on the
1375 * controller so try the other pg.
1377 bypass_pg(m
, pg
, true);
1380 /* Wait before retrying. */
1382 case SCSI_DH_IMM_RETRY
:
1383 case SCSI_DH_RES_TEMP_UNAVAIL
:
1384 if (pg_init_limit_reached(m
, pgpath
))
1388 case SCSI_DH_DEV_OFFLINED
:
1391 * We probably do not want to fail the path for a device
1392 * error, but this is what the old dm did. In future
1393 * patches we can do more advanced handling.
1398 spin_lock_irqsave(&m
->lock
, flags
);
1400 if (pgpath
== m
->current_pgpath
) {
1401 DMERR("Could not failover device. Error %d.", errors
);
1402 m
->current_pgpath
= NULL
;
1403 m
->current_pg
= NULL
;
1405 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1406 pg
->bypassed
= false;
1408 if (atomic_dec_return(&m
->pg_init_in_progress
) > 0)
1409 /* Activations of other paths are still on going */
1412 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
1414 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1416 clear_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1418 if (__pg_init_all_paths(m
))
1421 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
1423 process_queued_io_list(m
);
1426 * Wake up any thread waiting to suspend.
1428 wake_up(&m
->pg_init_wait
);
1431 spin_unlock_irqrestore(&m
->lock
, flags
);
1434 static void activate_or_offline_path(struct pgpath
*pgpath
)
1436 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1438 if (pgpath
->is_active
&& !blk_queue_dying(q
))
1439 scsi_dh_activate(q
, pg_init_done
, pgpath
);
1441 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1444 static void activate_path_work(struct work_struct
*work
)
1446 struct pgpath
*pgpath
=
1447 container_of(work
, struct pgpath
, activate_path
.work
);
1449 activate_or_offline_path(pgpath
);
1452 static int noretry_error(blk_status_t error
)
1455 case BLK_STS_NOTSUPP
:
1457 case BLK_STS_TARGET
:
1459 case BLK_STS_MEDIUM
:
1463 /* Anything else could be a path failure, so should be retried */
1467 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1468 blk_status_t error
, union map_info
*map_context
)
1470 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1471 struct pgpath
*pgpath
= mpio
->pgpath
;
1472 int r
= DM_ENDIO_DONE
;
1475 * We don't queue any clone request inside the multipath target
1476 * during end I/O handling, since those clone requests don't have
1477 * bio clones. If we queue them inside the multipath target,
1478 * we need to make bio clones, that requires memory allocation.
1479 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1480 * don't have bio clones.)
1481 * Instead of queueing the clone request here, we queue the original
1482 * request into dm core, which will remake a clone request and
1483 * clone bios for it and resubmit it later.
1485 if (error
&& !noretry_error(error
)) {
1486 struct multipath
*m
= ti
->private;
1488 r
= DM_ENDIO_REQUEUE
;
1493 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1494 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1495 if (error
== BLK_STS_IOERR
)
1497 /* complete with the original error */
1503 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1505 if (ps
->type
->end_io
)
1506 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1512 static int multipath_end_io_bio(struct dm_target
*ti
, struct bio
*clone
,
1513 blk_status_t
*error
)
1515 struct multipath
*m
= ti
->private;
1516 struct dm_mpath_io
*mpio
= get_mpio_from_bio(clone
);
1517 struct pgpath
*pgpath
= mpio
->pgpath
;
1518 unsigned long flags
;
1519 int r
= DM_ENDIO_DONE
;
1521 if (!*error
|| noretry_error(*error
))
1527 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1528 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1530 *error
= BLK_STS_IOERR
;
1534 /* Queue for the daemon to resubmit */
1535 dm_bio_restore(get_bio_details_from_bio(clone
), clone
);
1537 spin_lock_irqsave(&m
->lock
, flags
);
1538 bio_list_add(&m
->queued_bios
, clone
);
1539 spin_unlock_irqrestore(&m
->lock
, flags
);
1540 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
1541 queue_work(kmultipathd
, &m
->process_queued_bios
);
1543 r
= DM_ENDIO_INCOMPLETE
;
1546 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1548 if (ps
->type
->end_io
)
1549 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1556 * Suspend can't complete until all the I/O is processed so if
1557 * the last path fails we must error any remaining I/O.
1558 * Note that if the freeze_bdev fails while suspending, the
1559 * queue_if_no_path state is lost - userspace should reset it.
1561 static void multipath_presuspend(struct dm_target
*ti
)
1563 struct multipath
*m
= ti
->private;
1565 queue_if_no_path(m
, false, true);
1568 static void multipath_postsuspend(struct dm_target
*ti
)
1570 struct multipath
*m
= ti
->private;
1572 mutex_lock(&m
->work_mutex
);
1573 flush_multipath_work(m
);
1574 mutex_unlock(&m
->work_mutex
);
1578 * Restore the queue_if_no_path setting.
1580 static void multipath_resume(struct dm_target
*ti
)
1582 struct multipath
*m
= ti
->private;
1583 unsigned long flags
;
1585 spin_lock_irqsave(&m
->lock
, flags
);
1586 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
),
1587 MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
);
1588 spin_unlock_irqrestore(&m
->lock
, flags
);
1592 * Info output has the following format:
1593 * num_multipath_feature_args [multipath_feature_args]*
1594 * num_handler_status_args [handler_status_args]*
1595 * num_groups init_group_number
1596 * [A|D|E num_ps_status_args [ps_status_args]*
1597 * num_paths num_selector_args
1598 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1600 * Table output has the following format (identical to the constructor string):
1601 * num_feature_args [features_args]*
1602 * num_handler_args hw_handler [hw_handler_args]*
1603 * num_groups init_group_number
1604 * [priority selector-name num_ps_args [ps_args]*
1605 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1607 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1608 unsigned status_flags
, char *result
, unsigned maxlen
)
1611 unsigned long flags
;
1612 struct multipath
*m
= ti
->private;
1613 struct priority_group
*pg
;
1618 spin_lock_irqsave(&m
->lock
, flags
);
1621 if (type
== STATUSTYPE_INFO
)
1622 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO
, &m
->flags
),
1623 atomic_read(&m
->pg_init_count
));
1625 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) +
1626 (m
->pg_init_retries
> 0) * 2 +
1627 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1628 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) +
1629 (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) * 2);
1631 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1632 DMEMIT("queue_if_no_path ");
1633 if (m
->pg_init_retries
)
1634 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1635 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1636 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1637 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
))
1638 DMEMIT("retain_attached_hw_handler ");
1639 if (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) {
1640 switch(m
->queue_mode
) {
1641 case DM_TYPE_BIO_BASED
:
1642 DMEMIT("queue_mode bio ");
1644 case DM_TYPE_MQ_REQUEST_BASED
:
1645 DMEMIT("queue_mode mq ");
1654 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1657 DMEMIT("1 %s ", m
->hw_handler_name
);
1659 DMEMIT("%u ", m
->nr_priority_groups
);
1662 pg_num
= m
->next_pg
->pg_num
;
1663 else if (m
->current_pg
)
1664 pg_num
= m
->current_pg
->pg_num
;
1666 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1668 DMEMIT("%u ", pg_num
);
1671 case STATUSTYPE_INFO
:
1672 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1674 state
= 'D'; /* Disabled */
1675 else if (pg
== m
->current_pg
)
1676 state
= 'A'; /* Currently Active */
1678 state
= 'E'; /* Enabled */
1680 DMEMIT("%c ", state
);
1682 if (pg
->ps
.type
->status
)
1683 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1689 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1690 pg
->ps
.type
->info_args
);
1692 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1693 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1694 p
->is_active
? "A" : "F",
1696 if (pg
->ps
.type
->status
)
1697 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1698 &p
->path
, type
, result
+ sz
,
1704 case STATUSTYPE_TABLE
:
1705 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1706 DMEMIT("%s ", pg
->ps
.type
->name
);
1708 if (pg
->ps
.type
->status
)
1709 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1715 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1716 pg
->ps
.type
->table_args
);
1718 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1719 DMEMIT("%s ", p
->path
.dev
->name
);
1720 if (pg
->ps
.type
->status
)
1721 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1722 &p
->path
, type
, result
+ sz
,
1729 spin_unlock_irqrestore(&m
->lock
, flags
);
1732 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
1736 struct multipath
*m
= ti
->private;
1739 mutex_lock(&m
->work_mutex
);
1741 if (dm_suspended(ti
)) {
1747 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1748 r
= queue_if_no_path(m
, true, false);
1750 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1751 r
= queue_if_no_path(m
, false, false);
1757 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1761 if (!strcasecmp(argv
[0], "disable_group")) {
1762 r
= bypass_pg_num(m
, argv
[1], true);
1764 } else if (!strcasecmp(argv
[0], "enable_group")) {
1765 r
= bypass_pg_num(m
, argv
[1], false);
1767 } else if (!strcasecmp(argv
[0], "switch_group")) {
1768 r
= switch_pg_num(m
, argv
[1]);
1770 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1771 action
= reinstate_path
;
1772 else if (!strcasecmp(argv
[0], "fail_path"))
1775 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1779 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1781 DMWARN("message: error getting device %s",
1786 r
= action_dev(m
, dev
, action
);
1788 dm_put_device(ti
, dev
);
1791 mutex_unlock(&m
->work_mutex
);
1795 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1796 struct block_device
**bdev
, fmode_t
*mode
)
1798 struct multipath
*m
= ti
->private;
1799 struct pgpath
*current_pgpath
;
1802 current_pgpath
= lockless_dereference(m
->current_pgpath
);
1803 if (!current_pgpath
)
1804 current_pgpath
= choose_pgpath(m
, 0);
1806 if (current_pgpath
) {
1807 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
)) {
1808 *bdev
= current_pgpath
->path
.dev
->bdev
;
1809 *mode
= current_pgpath
->path
.dev
->mode
;
1812 /* pg_init has not started or completed */
1816 /* No path is available */
1817 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1823 if (r
== -ENOTCONN
) {
1824 if (!lockless_dereference(m
->current_pg
)) {
1825 /* Path status changed, redo selection */
1826 (void) choose_pgpath(m
, 0);
1828 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1829 pg_init_all_paths(m
);
1830 dm_table_run_md_queue_async(m
->ti
->table
);
1831 process_queued_io_list(m
);
1835 * Only pass ioctls through if the device sizes match exactly.
1837 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1842 static int multipath_iterate_devices(struct dm_target
*ti
,
1843 iterate_devices_callout_fn fn
, void *data
)
1845 struct multipath
*m
= ti
->private;
1846 struct priority_group
*pg
;
1850 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1851 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1852 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1862 static int pgpath_busy(struct pgpath
*pgpath
)
1864 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1866 return blk_lld_busy(q
);
1870 * We return "busy", only when we can map I/Os but underlying devices
1871 * are busy (so even if we map I/Os now, the I/Os will wait on
1872 * the underlying queue).
1873 * In other words, if we want to kill I/Os or queue them inside us
1874 * due to map unavailability, we don't return "busy". Otherwise,
1875 * dm core won't give us the I/Os and we can't do what we want.
1877 static int multipath_busy(struct dm_target
*ti
)
1879 bool busy
= false, has_active
= false;
1880 struct multipath
*m
= ti
->private;
1881 struct priority_group
*pg
, *next_pg
;
1882 struct pgpath
*pgpath
;
1884 /* pg_init in progress */
1885 if (atomic_read(&m
->pg_init_in_progress
))
1888 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1889 if (!atomic_read(&m
->nr_valid_paths
) && test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1890 return (m
->queue_mode
!= DM_TYPE_MQ_REQUEST_BASED
);
1892 /* Guess which priority_group will be used at next mapping time */
1893 pg
= lockless_dereference(m
->current_pg
);
1894 next_pg
= lockless_dereference(m
->next_pg
);
1895 if (unlikely(!lockless_dereference(m
->current_pgpath
) && next_pg
))
1900 * We don't know which pg will be used at next mapping time.
1901 * We don't call choose_pgpath() here to avoid to trigger
1902 * pg_init just by busy checking.
1903 * So we don't know whether underlying devices we will be using
1904 * at next mapping time are busy or not. Just try mapping.
1910 * If there is one non-busy active path at least, the path selector
1911 * will be able to select it. So we consider such a pg as not busy.
1914 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1915 if (pgpath
->is_active
) {
1917 if (!pgpath_busy(pgpath
)) {
1926 * No active path in this pg, so this pg won't be used and
1927 * the current_pg will be changed at next mapping time.
1928 * We need to try mapping to determine it.
1936 /*-----------------------------------------------------------------
1938 *---------------------------------------------------------------*/
1939 static struct target_type multipath_target
= {
1940 .name
= "multipath",
1941 .version
= {1, 12, 0},
1942 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
,
1943 .module
= THIS_MODULE
,
1944 .ctr
= multipath_ctr
,
1945 .dtr
= multipath_dtr
,
1946 .clone_and_map_rq
= multipath_clone_and_map
,
1947 .release_clone_rq
= multipath_release_clone
,
1948 .rq_end_io
= multipath_end_io
,
1949 .map
= multipath_map_bio
,
1950 .end_io
= multipath_end_io_bio
,
1951 .presuspend
= multipath_presuspend
,
1952 .postsuspend
= multipath_postsuspend
,
1953 .resume
= multipath_resume
,
1954 .status
= multipath_status
,
1955 .message
= multipath_message
,
1956 .prepare_ioctl
= multipath_prepare_ioctl
,
1957 .iterate_devices
= multipath_iterate_devices
,
1958 .busy
= multipath_busy
,
1961 static int __init
dm_multipath_init(void)
1965 r
= dm_register_target(&multipath_target
);
1967 DMERR("request-based register failed %d", r
);
1969 goto bad_register_target
;
1972 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
1974 DMERR("failed to create workqueue kmpathd");
1976 goto bad_alloc_kmultipathd
;
1980 * A separate workqueue is used to handle the device handlers
1981 * to avoid overloading existing workqueue. Overloading the
1982 * old workqueue would also create a bottleneck in the
1983 * path of the storage hardware device activation.
1985 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
1987 if (!kmpath_handlerd
) {
1988 DMERR("failed to create workqueue kmpath_handlerd");
1990 goto bad_alloc_kmpath_handlerd
;
1995 bad_alloc_kmpath_handlerd
:
1996 destroy_workqueue(kmultipathd
);
1997 bad_alloc_kmultipathd
:
1998 dm_unregister_target(&multipath_target
);
1999 bad_register_target
:
2003 static void __exit
dm_multipath_exit(void)
2005 destroy_workqueue(kmpath_handlerd
);
2006 destroy_workqueue(kmultipathd
);
2008 dm_unregister_target(&multipath_target
);
2011 module_init(dm_multipath_init
);
2012 module_exit(dm_multipath_exit
);
2014 MODULE_DESCRIPTION(DM_NAME
" multipath target");
2015 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2016 MODULE_LICENSE("GPL");