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 unsigned long flags
; /* Multipath state flags */
70 enum dm_queue_mode queue_mode
;
72 struct pgpath
*current_pgpath
;
73 struct priority_group
*current_pg
;
74 struct priority_group
*next_pg
; /* Switch to this PG if set */
76 atomic_t nr_valid_paths
; /* Total number of usable paths */
77 unsigned nr_priority_groups
;
78 struct list_head priority_groups
;
80 const char *hw_handler_name
;
81 char *hw_handler_params
;
82 wait_queue_head_t pg_init_wait
; /* Wait for pg_init completion */
83 unsigned pg_init_retries
; /* Number of times to retry pg_init */
84 unsigned pg_init_delay_msecs
; /* Number of msecs before pg_init retry */
85 atomic_t pg_init_in_progress
; /* Only one pg_init allowed at once */
86 atomic_t pg_init_count
; /* Number of times pg_init called */
88 struct mutex work_mutex
;
89 struct work_struct trigger_event
;
92 struct work_struct process_queued_bios
;
93 struct bio_list queued_bios
;
97 * Context information attached to each io we process.
100 struct pgpath
*pgpath
;
104 typedef int (*action_fn
) (struct pgpath
*pgpath
);
106 static struct workqueue_struct
*kmultipathd
, *kmpath_handlerd
;
107 static void trigger_event(struct work_struct
*work
);
108 static void activate_or_offline_path(struct pgpath
*pgpath
);
109 static void activate_path_work(struct work_struct
*work
);
110 static void process_queued_bios(struct work_struct
*work
);
112 /*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
116 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
117 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
120 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
121 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
122 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
124 /*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
128 static struct pgpath
*alloc_pgpath(void)
130 struct pgpath
*pgpath
= kzalloc(sizeof(*pgpath
), GFP_KERNEL
);
135 pgpath
->is_active
= true;
140 static void free_pgpath(struct pgpath
*pgpath
)
145 static struct priority_group
*alloc_priority_group(void)
147 struct priority_group
*pg
;
149 pg
= kzalloc(sizeof(*pg
), GFP_KERNEL
);
152 INIT_LIST_HEAD(&pg
->pgpaths
);
157 static void free_pgpaths(struct list_head
*pgpaths
, struct dm_target
*ti
)
159 struct pgpath
*pgpath
, *tmp
;
161 list_for_each_entry_safe(pgpath
, tmp
, pgpaths
, list
) {
162 list_del(&pgpath
->list
);
163 dm_put_device(ti
, pgpath
->path
.dev
);
168 static void free_priority_group(struct priority_group
*pg
,
169 struct dm_target
*ti
)
171 struct path_selector
*ps
= &pg
->ps
;
174 ps
->type
->destroy(ps
);
175 dm_put_path_selector(ps
->type
);
178 free_pgpaths(&pg
->pgpaths
, ti
);
182 static struct multipath
*alloc_multipath(struct dm_target
*ti
)
186 m
= kzalloc(sizeof(*m
), GFP_KERNEL
);
188 INIT_LIST_HEAD(&m
->priority_groups
);
189 spin_lock_init(&m
->lock
);
190 atomic_set(&m
->nr_valid_paths
, 0);
191 INIT_WORK(&m
->trigger_event
, trigger_event
);
192 mutex_init(&m
->work_mutex
);
194 m
->queue_mode
= DM_TYPE_NONE
;
203 static int alloc_multipath_stage2(struct dm_target
*ti
, struct multipath
*m
)
205 if (m
->queue_mode
== DM_TYPE_NONE
) {
206 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
207 } else if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
208 INIT_WORK(&m
->process_queued_bios
, process_queued_bios
);
210 * bio-based doesn't support any direct scsi_dh management;
211 * it just discovers if a scsi_dh is attached.
213 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
216 dm_table_set_type(ti
->table
, m
->queue_mode
);
219 * Init fields that are only used when a scsi_dh is attached
220 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
222 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
223 atomic_set(&m
->pg_init_in_progress
, 0);
224 atomic_set(&m
->pg_init_count
, 0);
225 m
->pg_init_delay_msecs
= DM_PG_INIT_DELAY_DEFAULT
;
226 init_waitqueue_head(&m
->pg_init_wait
);
231 static void free_multipath(struct multipath
*m
)
233 struct priority_group
*pg
, *tmp
;
235 list_for_each_entry_safe(pg
, tmp
, &m
->priority_groups
, list
) {
237 free_priority_group(pg
, m
->ti
);
240 kfree(m
->hw_handler_name
);
241 kfree(m
->hw_handler_params
);
242 mutex_destroy(&m
->work_mutex
);
246 static struct dm_mpath_io
*get_mpio(union map_info
*info
)
251 static size_t multipath_per_bio_data_size(void)
253 return sizeof(struct dm_mpath_io
) + sizeof(struct dm_bio_details
);
256 static struct dm_mpath_io
*get_mpio_from_bio(struct bio
*bio
)
258 return dm_per_bio_data(bio
, multipath_per_bio_data_size());
261 static struct dm_bio_details
*get_bio_details_from_mpio(struct dm_mpath_io
*mpio
)
263 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
264 void *bio_details
= mpio
+ 1;
268 static void multipath_init_per_bio_data(struct bio
*bio
, struct dm_mpath_io
**mpio_p
)
270 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
271 struct dm_bio_details
*bio_details
= get_bio_details_from_mpio(mpio
);
273 mpio
->nr_bytes
= bio
->bi_iter
.bi_size
;
277 dm_bio_record(bio_details
, bio
);
280 /*-----------------------------------------------
282 *-----------------------------------------------*/
284 static int __pg_init_all_paths(struct multipath
*m
)
286 struct pgpath
*pgpath
;
287 unsigned long pg_init_delay
= 0;
289 lockdep_assert_held(&m
->lock
);
291 if (atomic_read(&m
->pg_init_in_progress
) || test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
294 atomic_inc(&m
->pg_init_count
);
295 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
297 /* Check here to reset pg_init_required */
301 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
))
302 pg_init_delay
= msecs_to_jiffies(m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
?
303 m
->pg_init_delay_msecs
: DM_PG_INIT_DELAY_MSECS
);
304 list_for_each_entry(pgpath
, &m
->current_pg
->pgpaths
, list
) {
305 /* Skip failed paths */
306 if (!pgpath
->is_active
)
308 if (queue_delayed_work(kmpath_handlerd
, &pgpath
->activate_path
,
310 atomic_inc(&m
->pg_init_in_progress
);
312 return atomic_read(&m
->pg_init_in_progress
);
315 static int pg_init_all_paths(struct multipath
*m
)
320 spin_lock_irqsave(&m
->lock
, flags
);
321 ret
= __pg_init_all_paths(m
);
322 spin_unlock_irqrestore(&m
->lock
, flags
);
327 static void __switch_pg(struct multipath
*m
, struct priority_group
*pg
)
331 /* Must we initialise the PG first, and queue I/O till it's ready? */
332 if (m
->hw_handler_name
) {
333 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
334 set_bit(MPATHF_QUEUE_IO
, &m
->flags
);
336 clear_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
337 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
340 atomic_set(&m
->pg_init_count
, 0);
343 static struct pgpath
*choose_path_in_pg(struct multipath
*m
,
344 struct priority_group
*pg
,
348 struct dm_path
*path
;
349 struct pgpath
*pgpath
;
351 path
= pg
->ps
.type
->select_path(&pg
->ps
, nr_bytes
);
353 return ERR_PTR(-ENXIO
);
355 pgpath
= path_to_pgpath(path
);
357 if (unlikely(READ_ONCE(m
->current_pg
) != pg
)) {
358 /* Only update current_pgpath if pg changed */
359 spin_lock_irqsave(&m
->lock
, flags
);
360 m
->current_pgpath
= pgpath
;
362 spin_unlock_irqrestore(&m
->lock
, flags
);
368 static struct pgpath
*choose_pgpath(struct multipath
*m
, size_t nr_bytes
)
371 struct priority_group
*pg
;
372 struct pgpath
*pgpath
;
373 unsigned bypassed
= 1;
375 if (!atomic_read(&m
->nr_valid_paths
)) {
376 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
380 /* Were we instructed to switch PG? */
381 if (READ_ONCE(m
->next_pg
)) {
382 spin_lock_irqsave(&m
->lock
, flags
);
385 spin_unlock_irqrestore(&m
->lock
, flags
);
386 goto check_current_pg
;
389 spin_unlock_irqrestore(&m
->lock
, flags
);
390 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
391 if (!IS_ERR_OR_NULL(pgpath
))
395 /* Don't change PG until it has no remaining paths */
397 pg
= READ_ONCE(m
->current_pg
);
399 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
400 if (!IS_ERR_OR_NULL(pgpath
))
405 * Loop through priority groups until we find a valid path.
406 * First time we skip PGs marked 'bypassed'.
407 * Second time we only try the ones we skipped, but set
408 * pg_init_delay_retry so we do not hammer controllers.
411 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
412 if (pg
->bypassed
== !!bypassed
)
414 pgpath
= choose_path_in_pg(m
, pg
, nr_bytes
);
415 if (!IS_ERR_OR_NULL(pgpath
)) {
417 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
421 } while (bypassed
--);
424 spin_lock_irqsave(&m
->lock
, flags
);
425 m
->current_pgpath
= NULL
;
426 m
->current_pg
= NULL
;
427 spin_unlock_irqrestore(&m
->lock
, flags
);
433 * dm_report_EIO() is a macro instead of a function to make pr_debug()
434 * report the function name and line number of the function from which
435 * it has been invoked.
437 #define dm_report_EIO(m) \
439 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
441 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
442 dm_device_name(md), \
443 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
444 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
445 dm_noflush_suspending((m)->ti)); \
449 * Check whether bios must be queued in the device-mapper core rather
450 * than here in the target.
452 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
453 * the same value then we are not between multipath_presuspend()
454 * and multipath_resume() calls and we have no need to check
455 * for the DMF_NOFLUSH_SUSPENDING flag.
457 static bool __must_push_back(struct multipath
*m
, unsigned long flags
)
459 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH
, &flags
) !=
460 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &flags
)) &&
461 dm_noflush_suspending(m
->ti
));
465 * Following functions use READ_ONCE to get atomic access to
466 * all m->flags to avoid taking spinlock
468 static bool must_push_back_rq(struct multipath
*m
)
470 unsigned long flags
= READ_ONCE(m
->flags
);
471 return test_bit(MPATHF_QUEUE_IF_NO_PATH
, &flags
) || __must_push_back(m
, flags
);
474 static bool must_push_back_bio(struct multipath
*m
)
476 unsigned long flags
= READ_ONCE(m
->flags
);
477 return __must_push_back(m
, flags
);
481 * Map cloned requests (request-based multipath)
483 static int multipath_clone_and_map(struct dm_target
*ti
, struct request
*rq
,
484 union map_info
*map_context
,
485 struct request
**__clone
)
487 struct multipath
*m
= ti
->private;
488 size_t nr_bytes
= blk_rq_bytes(rq
);
489 struct pgpath
*pgpath
;
490 struct block_device
*bdev
;
491 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
492 struct request_queue
*q
;
493 struct request
*clone
;
495 /* Do we need to select a new pgpath? */
496 pgpath
= READ_ONCE(m
->current_pgpath
);
497 if (!pgpath
|| !test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
498 pgpath
= choose_pgpath(m
, nr_bytes
);
501 if (must_push_back_rq(m
))
502 return DM_MAPIO_DELAY_REQUEUE
;
503 dm_report_EIO(m
); /* Failed */
504 return DM_MAPIO_KILL
;
505 } else if (test_bit(MPATHF_QUEUE_IO
, &m
->flags
) ||
506 test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
507 pg_init_all_paths(m
);
508 return DM_MAPIO_DELAY_REQUEUE
;
511 mpio
->pgpath
= pgpath
;
512 mpio
->nr_bytes
= nr_bytes
;
514 bdev
= pgpath
->path
.dev
->bdev
;
515 q
= bdev_get_queue(bdev
);
516 clone
= blk_get_request(q
, rq
->cmd_flags
| REQ_NOMERGE
,
519 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
520 if (blk_queue_dying(q
)) {
521 atomic_inc(&m
->pg_init_in_progress
);
522 activate_or_offline_path(pgpath
);
523 return DM_MAPIO_DELAY_REQUEUE
;
527 * blk-mq's SCHED_RESTART can cover this requeue, so we
528 * needn't deal with it by DELAY_REQUEUE. More importantly,
529 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
530 * get the queue busy feedback (via BLK_STS_RESOURCE),
531 * otherwise I/O merging can suffer.
533 return DM_MAPIO_REQUEUE
;
535 clone
->bio
= clone
->biotail
= NULL
;
536 clone
->rq_disk
= bdev
->bd_disk
;
537 clone
->cmd_flags
|= REQ_FAILFAST_TRANSPORT
;
540 if (pgpath
->pg
->ps
.type
->start_io
)
541 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
544 return DM_MAPIO_REMAPPED
;
547 static void multipath_release_clone(struct request
*clone
)
549 blk_put_request(clone
);
553 * Map cloned bios (bio-based multipath)
556 static struct pgpath
*__map_bio(struct multipath
*m
, struct bio
*bio
)
558 struct pgpath
*pgpath
;
562 /* Do we need to select a new pgpath? */
563 pgpath
= READ_ONCE(m
->current_pgpath
);
564 queue_io
= test_bit(MPATHF_QUEUE_IO
, &m
->flags
);
565 if (!pgpath
|| !queue_io
)
566 pgpath
= choose_pgpath(m
, bio
->bi_iter
.bi_size
);
568 if ((pgpath
&& queue_io
) ||
569 (!pgpath
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))) {
570 /* Queue for the daemon to resubmit */
571 spin_lock_irqsave(&m
->lock
, flags
);
572 bio_list_add(&m
->queued_bios
, bio
);
573 spin_unlock_irqrestore(&m
->lock
, flags
);
575 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
576 if (queue_io
|| test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
577 pg_init_all_paths(m
);
579 queue_work(kmultipathd
, &m
->process_queued_bios
);
581 return ERR_PTR(-EAGAIN
);
587 static struct pgpath
*__map_bio_fast(struct multipath
*m
, struct bio
*bio
)
589 struct pgpath
*pgpath
;
592 /* Do we need to select a new pgpath? */
594 * FIXME: currently only switching path if no path (due to failure, etc)
595 * - which negates the point of using a path selector
597 pgpath
= READ_ONCE(m
->current_pgpath
);
599 pgpath
= choose_pgpath(m
, bio
->bi_iter
.bi_size
);
602 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
603 /* Queue for the daemon to resubmit */
604 spin_lock_irqsave(&m
->lock
, flags
);
605 bio_list_add(&m
->queued_bios
, bio
);
606 spin_unlock_irqrestore(&m
->lock
, flags
);
607 queue_work(kmultipathd
, &m
->process_queued_bios
);
609 return ERR_PTR(-EAGAIN
);
617 static int __multipath_map_bio(struct multipath
*m
, struct bio
*bio
,
618 struct dm_mpath_io
*mpio
)
620 struct pgpath
*pgpath
;
622 if (!m
->hw_handler_name
)
623 pgpath
= __map_bio_fast(m
, bio
);
625 pgpath
= __map_bio(m
, bio
);
628 return DM_MAPIO_SUBMITTED
;
631 if (must_push_back_bio(m
))
632 return DM_MAPIO_REQUEUE
;
634 return DM_MAPIO_KILL
;
637 mpio
->pgpath
= pgpath
;
640 bio_set_dev(bio
, pgpath
->path
.dev
->bdev
);
641 bio
->bi_opf
|= REQ_FAILFAST_TRANSPORT
;
643 if (pgpath
->pg
->ps
.type
->start_io
)
644 pgpath
->pg
->ps
.type
->start_io(&pgpath
->pg
->ps
,
647 return DM_MAPIO_REMAPPED
;
650 static int multipath_map_bio(struct dm_target
*ti
, struct bio
*bio
)
652 struct multipath
*m
= ti
->private;
653 struct dm_mpath_io
*mpio
= NULL
;
655 multipath_init_per_bio_data(bio
, &mpio
);
656 return __multipath_map_bio(m
, bio
, mpio
);
659 static void process_queued_io_list(struct multipath
*m
)
661 if (m
->queue_mode
== DM_TYPE_REQUEST_BASED
)
662 dm_mq_kick_requeue_list(dm_table_get_md(m
->ti
->table
));
663 else if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
664 queue_work(kmultipathd
, &m
->process_queued_bios
);
667 static void process_queued_bios(struct work_struct
*work
)
672 struct bio_list bios
;
673 struct blk_plug plug
;
674 struct multipath
*m
=
675 container_of(work
, struct multipath
, process_queued_bios
);
677 bio_list_init(&bios
);
679 spin_lock_irqsave(&m
->lock
, flags
);
681 if (bio_list_empty(&m
->queued_bios
)) {
682 spin_unlock_irqrestore(&m
->lock
, flags
);
686 bio_list_merge(&bios
, &m
->queued_bios
);
687 bio_list_init(&m
->queued_bios
);
689 spin_unlock_irqrestore(&m
->lock
, flags
);
691 blk_start_plug(&plug
);
692 while ((bio
= bio_list_pop(&bios
))) {
693 struct dm_mpath_io
*mpio
= get_mpio_from_bio(bio
);
694 dm_bio_restore(get_bio_details_from_mpio(mpio
), bio
);
695 r
= __multipath_map_bio(m
, bio
, mpio
);
698 bio
->bi_status
= BLK_STS_IOERR
;
701 case DM_MAPIO_REQUEUE
:
702 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
705 case DM_MAPIO_REMAPPED
:
706 generic_make_request(bio
);
708 case DM_MAPIO_SUBMITTED
:
711 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r
);
714 blk_finish_plug(&plug
);
718 * If we run out of usable paths, should we queue I/O or error it?
720 static int queue_if_no_path(struct multipath
*m
, bool queue_if_no_path
,
725 spin_lock_irqsave(&m
->lock
, flags
);
726 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
,
727 (save_old_value
&& test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) ||
728 (!save_old_value
&& queue_if_no_path
));
729 assign_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
, queue_if_no_path
);
730 spin_unlock_irqrestore(&m
->lock
, flags
);
732 if (!queue_if_no_path
) {
733 dm_table_run_md_queue_async(m
->ti
->table
);
734 process_queued_io_list(m
);
741 * An event is triggered whenever a path is taken out of use.
742 * Includes path failure and PG bypass.
744 static void trigger_event(struct work_struct
*work
)
746 struct multipath
*m
=
747 container_of(work
, struct multipath
, trigger_event
);
749 dm_table_event(m
->ti
->table
);
752 /*-----------------------------------------------------------------
753 * Constructor/argument parsing:
754 * <#multipath feature args> [<arg>]*
755 * <#hw_handler args> [hw_handler [<arg>]*]
757 * <initial priority group>
758 * [<selector> <#selector args> [<arg>]*
759 * <#paths> <#per-path selector args>
760 * [<path> [<arg>]* ]+ ]+
761 *---------------------------------------------------------------*/
762 static int parse_path_selector(struct dm_arg_set
*as
, struct priority_group
*pg
,
763 struct dm_target
*ti
)
766 struct path_selector_type
*pst
;
769 static const struct dm_arg _args
[] = {
770 {0, 1024, "invalid number of path selector args"},
773 pst
= dm_get_path_selector(dm_shift_arg(as
));
775 ti
->error
= "unknown path selector type";
779 r
= dm_read_arg_group(_args
, as
, &ps_argc
, &ti
->error
);
781 dm_put_path_selector(pst
);
785 r
= pst
->create(&pg
->ps
, ps_argc
, as
->argv
);
787 dm_put_path_selector(pst
);
788 ti
->error
= "path selector constructor failed";
793 dm_consume_args(as
, ps_argc
);
798 static int setup_scsi_dh(struct block_device
*bdev
, struct multipath
*m
,
799 const char **attached_handler_name
, char **error
)
801 struct request_queue
*q
= bdev_get_queue(bdev
);
804 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
)) {
806 if (*attached_handler_name
) {
808 * Clear any hw_handler_params associated with a
809 * handler that isn't already attached.
811 if (m
->hw_handler_name
&& strcmp(*attached_handler_name
, m
->hw_handler_name
)) {
812 kfree(m
->hw_handler_params
);
813 m
->hw_handler_params
= NULL
;
817 * Reset hw_handler_name to match the attached handler
819 * NB. This modifies the table line to show the actual
820 * handler instead of the original table passed in.
822 kfree(m
->hw_handler_name
);
823 m
->hw_handler_name
= *attached_handler_name
;
824 *attached_handler_name
= NULL
;
828 if (m
->hw_handler_name
) {
829 r
= scsi_dh_attach(q
, m
->hw_handler_name
);
831 char b
[BDEVNAME_SIZE
];
833 printk(KERN_INFO
"dm-mpath: retaining handler on device %s\n",
838 *error
= "error attaching hardware handler";
842 if (m
->hw_handler_params
) {
843 r
= scsi_dh_set_params(q
, m
->hw_handler_params
);
845 *error
= "unable to set hardware handler parameters";
854 static struct pgpath
*parse_path(struct dm_arg_set
*as
, struct path_selector
*ps
,
855 struct dm_target
*ti
)
859 struct multipath
*m
= ti
->private;
860 struct request_queue
*q
;
861 const char *attached_handler_name
= NULL
;
863 /* we need at least a path arg */
865 ti
->error
= "no device given";
866 return ERR_PTR(-EINVAL
);
871 return ERR_PTR(-ENOMEM
);
873 r
= dm_get_device(ti
, dm_shift_arg(as
), dm_table_get_mode(ti
->table
),
876 ti
->error
= "error getting device";
880 q
= bdev_get_queue(p
->path
.dev
->bdev
);
881 attached_handler_name
= scsi_dh_attached_handler_name(q
, GFP_KERNEL
);
882 if (attached_handler_name
|| m
->hw_handler_name
) {
883 INIT_DELAYED_WORK(&p
->activate_path
, activate_path_work
);
884 r
= setup_scsi_dh(p
->path
.dev
->bdev
, m
, &attached_handler_name
, &ti
->error
);
886 dm_put_device(ti
, p
->path
.dev
);
891 r
= ps
->type
->add_path(ps
, &p
->path
, as
->argc
, as
->argv
, &ti
->error
);
893 dm_put_device(ti
, p
->path
.dev
);
899 kfree(attached_handler_name
);
904 static struct priority_group
*parse_priority_group(struct dm_arg_set
*as
,
907 static const struct dm_arg _args
[] = {
908 {1, 1024, "invalid number of paths"},
909 {0, 1024, "invalid number of selector args"}
913 unsigned i
, nr_selector_args
, nr_args
;
914 struct priority_group
*pg
;
915 struct dm_target
*ti
= m
->ti
;
919 ti
->error
= "not enough priority group arguments";
920 return ERR_PTR(-EINVAL
);
923 pg
= alloc_priority_group();
925 ti
->error
= "couldn't allocate priority group";
926 return ERR_PTR(-ENOMEM
);
930 r
= parse_path_selector(as
, pg
, ti
);
937 r
= dm_read_arg(_args
, as
, &pg
->nr_pgpaths
, &ti
->error
);
941 r
= dm_read_arg(_args
+ 1, as
, &nr_selector_args
, &ti
->error
);
945 nr_args
= 1 + nr_selector_args
;
946 for (i
= 0; i
< pg
->nr_pgpaths
; i
++) {
947 struct pgpath
*pgpath
;
948 struct dm_arg_set path_args
;
950 if (as
->argc
< nr_args
) {
951 ti
->error
= "not enough path parameters";
956 path_args
.argc
= nr_args
;
957 path_args
.argv
= as
->argv
;
959 pgpath
= parse_path(&path_args
, &pg
->ps
, ti
);
960 if (IS_ERR(pgpath
)) {
966 list_add_tail(&pgpath
->list
, &pg
->pgpaths
);
967 dm_consume_args(as
, nr_args
);
973 free_priority_group(pg
, ti
);
977 static int parse_hw_handler(struct dm_arg_set
*as
, struct multipath
*m
)
981 struct dm_target
*ti
= m
->ti
;
983 static const struct dm_arg _args
[] = {
984 {0, 1024, "invalid number of hardware handler args"},
987 if (dm_read_arg_group(_args
, as
, &hw_argc
, &ti
->error
))
993 if (m
->queue_mode
== DM_TYPE_BIO_BASED
) {
994 dm_consume_args(as
, hw_argc
);
995 DMERR("bio-based multipath doesn't allow hardware handler args");
999 m
->hw_handler_name
= kstrdup(dm_shift_arg(as
), GFP_KERNEL
);
1000 if (!m
->hw_handler_name
)
1007 for (i
= 0; i
<= hw_argc
- 2; i
++)
1008 len
+= strlen(as
->argv
[i
]) + 1;
1009 p
= m
->hw_handler_params
= kzalloc(len
, GFP_KERNEL
);
1011 ti
->error
= "memory allocation failed";
1015 j
= sprintf(p
, "%d", hw_argc
- 1);
1016 for (i
= 0, p
+=j
+1; i
<= hw_argc
- 2; i
++, p
+=j
+1)
1017 j
= sprintf(p
, "%s", as
->argv
[i
]);
1019 dm_consume_args(as
, hw_argc
- 1);
1023 kfree(m
->hw_handler_name
);
1024 m
->hw_handler_name
= NULL
;
1028 static int parse_features(struct dm_arg_set
*as
, struct multipath
*m
)
1032 struct dm_target
*ti
= m
->ti
;
1033 const char *arg_name
;
1035 static const struct dm_arg _args
[] = {
1036 {0, 8, "invalid number of feature args"},
1037 {1, 50, "pg_init_retries must be between 1 and 50"},
1038 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1041 r
= dm_read_arg_group(_args
, as
, &argc
, &ti
->error
);
1049 arg_name
= dm_shift_arg(as
);
1052 if (!strcasecmp(arg_name
, "queue_if_no_path")) {
1053 r
= queue_if_no_path(m
, true, false);
1057 if (!strcasecmp(arg_name
, "retain_attached_hw_handler")) {
1058 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
);
1062 if (!strcasecmp(arg_name
, "pg_init_retries") &&
1064 r
= dm_read_arg(_args
+ 1, as
, &m
->pg_init_retries
, &ti
->error
);
1069 if (!strcasecmp(arg_name
, "pg_init_delay_msecs") &&
1071 r
= dm_read_arg(_args
+ 2, as
, &m
->pg_init_delay_msecs
, &ti
->error
);
1076 if (!strcasecmp(arg_name
, "queue_mode") &&
1078 const char *queue_mode_name
= dm_shift_arg(as
);
1080 if (!strcasecmp(queue_mode_name
, "bio"))
1081 m
->queue_mode
= DM_TYPE_BIO_BASED
;
1082 else if (!strcasecmp(queue_mode_name
, "rq") ||
1083 !strcasecmp(queue_mode_name
, "mq"))
1084 m
->queue_mode
= DM_TYPE_REQUEST_BASED
;
1086 ti
->error
= "Unknown 'queue_mode' requested";
1093 ti
->error
= "Unrecognised multipath feature request";
1095 } while (argc
&& !r
);
1100 static int multipath_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
1102 /* target arguments */
1103 static const struct dm_arg _args
[] = {
1104 {0, 1024, "invalid number of priority groups"},
1105 {0, 1024, "invalid initial priority group number"},
1109 struct multipath
*m
;
1110 struct dm_arg_set as
;
1111 unsigned pg_count
= 0;
1112 unsigned next_pg_num
;
1117 m
= alloc_multipath(ti
);
1119 ti
->error
= "can't allocate multipath";
1123 r
= parse_features(&as
, m
);
1127 r
= alloc_multipath_stage2(ti
, m
);
1131 r
= parse_hw_handler(&as
, m
);
1135 r
= dm_read_arg(_args
, &as
, &m
->nr_priority_groups
, &ti
->error
);
1139 r
= dm_read_arg(_args
+ 1, &as
, &next_pg_num
, &ti
->error
);
1143 if ((!m
->nr_priority_groups
&& next_pg_num
) ||
1144 (m
->nr_priority_groups
&& !next_pg_num
)) {
1145 ti
->error
= "invalid initial priority group";
1150 /* parse the priority groups */
1152 struct priority_group
*pg
;
1153 unsigned nr_valid_paths
= atomic_read(&m
->nr_valid_paths
);
1155 pg
= parse_priority_group(&as
, m
);
1161 nr_valid_paths
+= pg
->nr_pgpaths
;
1162 atomic_set(&m
->nr_valid_paths
, nr_valid_paths
);
1164 list_add_tail(&pg
->list
, &m
->priority_groups
);
1166 pg
->pg_num
= pg_count
;
1171 if (pg_count
!= m
->nr_priority_groups
) {
1172 ti
->error
= "priority group count mismatch";
1177 ti
->num_flush_bios
= 1;
1178 ti
->num_discard_bios
= 1;
1179 ti
->num_write_same_bios
= 1;
1180 ti
->num_write_zeroes_bios
= 1;
1181 if (m
->queue_mode
== DM_TYPE_BIO_BASED
)
1182 ti
->per_io_data_size
= multipath_per_bio_data_size();
1184 ti
->per_io_data_size
= sizeof(struct dm_mpath_io
);
1193 static void multipath_wait_for_pg_init_completion(struct multipath
*m
)
1198 prepare_to_wait(&m
->pg_init_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1200 if (!atomic_read(&m
->pg_init_in_progress
))
1205 finish_wait(&m
->pg_init_wait
, &wait
);
1208 static void flush_multipath_work(struct multipath
*m
)
1210 if (m
->hw_handler_name
) {
1211 set_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1212 smp_mb__after_atomic();
1214 flush_workqueue(kmpath_handlerd
);
1215 multipath_wait_for_pg_init_completion(m
);
1217 clear_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
);
1218 smp_mb__after_atomic();
1221 flush_workqueue(kmultipathd
);
1222 flush_work(&m
->trigger_event
);
1225 static void multipath_dtr(struct dm_target
*ti
)
1227 struct multipath
*m
= ti
->private;
1229 flush_multipath_work(m
);
1234 * Take a path out of use.
1236 static int fail_path(struct pgpath
*pgpath
)
1238 unsigned long flags
;
1239 struct multipath
*m
= pgpath
->pg
->m
;
1241 spin_lock_irqsave(&m
->lock
, flags
);
1243 if (!pgpath
->is_active
)
1246 DMWARN("Failing path %s.", pgpath
->path
.dev
->name
);
1248 pgpath
->pg
->ps
.type
->fail_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1249 pgpath
->is_active
= false;
1250 pgpath
->fail_count
++;
1252 atomic_dec(&m
->nr_valid_paths
);
1254 if (pgpath
== m
->current_pgpath
)
1255 m
->current_pgpath
= NULL
;
1257 dm_path_uevent(DM_UEVENT_PATH_FAILED
, m
->ti
,
1258 pgpath
->path
.dev
->name
, atomic_read(&m
->nr_valid_paths
));
1260 schedule_work(&m
->trigger_event
);
1263 spin_unlock_irqrestore(&m
->lock
, flags
);
1269 * Reinstate a previously-failed path
1271 static int reinstate_path(struct pgpath
*pgpath
)
1273 int r
= 0, run_queue
= 0;
1274 unsigned long flags
;
1275 struct multipath
*m
= pgpath
->pg
->m
;
1276 unsigned nr_valid_paths
;
1278 spin_lock_irqsave(&m
->lock
, flags
);
1280 if (pgpath
->is_active
)
1283 DMWARN("Reinstating path %s.", pgpath
->path
.dev
->name
);
1285 r
= pgpath
->pg
->ps
.type
->reinstate_path(&pgpath
->pg
->ps
, &pgpath
->path
);
1289 pgpath
->is_active
= true;
1291 nr_valid_paths
= atomic_inc_return(&m
->nr_valid_paths
);
1292 if (nr_valid_paths
== 1) {
1293 m
->current_pgpath
= NULL
;
1295 } else if (m
->hw_handler_name
&& (m
->current_pg
== pgpath
->pg
)) {
1296 if (queue_work(kmpath_handlerd
, &pgpath
->activate_path
.work
))
1297 atomic_inc(&m
->pg_init_in_progress
);
1300 dm_path_uevent(DM_UEVENT_PATH_REINSTATED
, m
->ti
,
1301 pgpath
->path
.dev
->name
, nr_valid_paths
);
1303 schedule_work(&m
->trigger_event
);
1306 spin_unlock_irqrestore(&m
->lock
, flags
);
1308 dm_table_run_md_queue_async(m
->ti
->table
);
1309 process_queued_io_list(m
);
1316 * Fail or reinstate all paths that match the provided struct dm_dev.
1318 static int action_dev(struct multipath
*m
, struct dm_dev
*dev
,
1322 struct pgpath
*pgpath
;
1323 struct priority_group
*pg
;
1325 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1326 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1327 if (pgpath
->path
.dev
== dev
)
1336 * Temporarily try to avoid having to use the specified PG
1338 static void bypass_pg(struct multipath
*m
, struct priority_group
*pg
,
1341 unsigned long flags
;
1343 spin_lock_irqsave(&m
->lock
, flags
);
1345 pg
->bypassed
= bypassed
;
1346 m
->current_pgpath
= NULL
;
1347 m
->current_pg
= NULL
;
1349 spin_unlock_irqrestore(&m
->lock
, flags
);
1351 schedule_work(&m
->trigger_event
);
1355 * Switch to using the specified PG from the next I/O that gets mapped
1357 static int switch_pg_num(struct multipath
*m
, const char *pgstr
)
1359 struct priority_group
*pg
;
1361 unsigned long flags
;
1364 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1365 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1366 DMWARN("invalid PG number supplied to switch_pg_num");
1370 spin_lock_irqsave(&m
->lock
, flags
);
1371 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1372 pg
->bypassed
= false;
1376 m
->current_pgpath
= NULL
;
1377 m
->current_pg
= NULL
;
1380 spin_unlock_irqrestore(&m
->lock
, flags
);
1382 schedule_work(&m
->trigger_event
);
1387 * Set/clear bypassed status of a PG.
1388 * PGs are numbered upwards from 1 in the order they were declared.
1390 static int bypass_pg_num(struct multipath
*m
, const char *pgstr
, bool bypassed
)
1392 struct priority_group
*pg
;
1396 if (!pgstr
|| (sscanf(pgstr
, "%u%c", &pgnum
, &dummy
) != 1) || !pgnum
||
1397 !m
->nr_priority_groups
|| (pgnum
> m
->nr_priority_groups
)) {
1398 DMWARN("invalid PG number supplied to bypass_pg");
1402 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1407 bypass_pg(m
, pg
, bypassed
);
1412 * Should we retry pg_init immediately?
1414 static bool pg_init_limit_reached(struct multipath
*m
, struct pgpath
*pgpath
)
1416 unsigned long flags
;
1417 bool limit_reached
= false;
1419 spin_lock_irqsave(&m
->lock
, flags
);
1421 if (atomic_read(&m
->pg_init_count
) <= m
->pg_init_retries
&&
1422 !test_bit(MPATHF_PG_INIT_DISABLED
, &m
->flags
))
1423 set_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
);
1425 limit_reached
= true;
1427 spin_unlock_irqrestore(&m
->lock
, flags
);
1429 return limit_reached
;
1432 static void pg_init_done(void *data
, int errors
)
1434 struct pgpath
*pgpath
= data
;
1435 struct priority_group
*pg
= pgpath
->pg
;
1436 struct multipath
*m
= pg
->m
;
1437 unsigned long flags
;
1438 bool delay_retry
= false;
1440 /* device or driver problems */
1445 if (!m
->hw_handler_name
) {
1449 DMERR("Could not failover the device: Handler scsi_dh_%s "
1450 "Error %d.", m
->hw_handler_name
, errors
);
1452 * Fail path for now, so we do not ping pong
1456 case SCSI_DH_DEV_TEMP_BUSY
:
1458 * Probably doing something like FW upgrade on the
1459 * controller so try the other pg.
1461 bypass_pg(m
, pg
, true);
1464 /* Wait before retrying. */
1467 case SCSI_DH_IMM_RETRY
:
1468 case SCSI_DH_RES_TEMP_UNAVAIL
:
1469 if (pg_init_limit_reached(m
, pgpath
))
1473 case SCSI_DH_DEV_OFFLINED
:
1476 * We probably do not want to fail the path for a device
1477 * error, but this is what the old dm did. In future
1478 * patches we can do more advanced handling.
1483 spin_lock_irqsave(&m
->lock
, flags
);
1485 if (pgpath
== m
->current_pgpath
) {
1486 DMERR("Could not failover device. Error %d.", errors
);
1487 m
->current_pgpath
= NULL
;
1488 m
->current_pg
= NULL
;
1490 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1491 pg
->bypassed
= false;
1493 if (atomic_dec_return(&m
->pg_init_in_progress
) > 0)
1494 /* Activations of other paths are still on going */
1497 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
)) {
1499 set_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1501 clear_bit(MPATHF_PG_INIT_DELAY_RETRY
, &m
->flags
);
1503 if (__pg_init_all_paths(m
))
1506 clear_bit(MPATHF_QUEUE_IO
, &m
->flags
);
1508 process_queued_io_list(m
);
1511 * Wake up any thread waiting to suspend.
1513 wake_up(&m
->pg_init_wait
);
1516 spin_unlock_irqrestore(&m
->lock
, flags
);
1519 static void activate_or_offline_path(struct pgpath
*pgpath
)
1521 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1523 if (pgpath
->is_active
&& !blk_queue_dying(q
))
1524 scsi_dh_activate(q
, pg_init_done
, pgpath
);
1526 pg_init_done(pgpath
, SCSI_DH_DEV_OFFLINED
);
1529 static void activate_path_work(struct work_struct
*work
)
1531 struct pgpath
*pgpath
=
1532 container_of(work
, struct pgpath
, activate_path
.work
);
1534 activate_or_offline_path(pgpath
);
1537 static int multipath_end_io(struct dm_target
*ti
, struct request
*clone
,
1538 blk_status_t error
, union map_info
*map_context
)
1540 struct dm_mpath_io
*mpio
= get_mpio(map_context
);
1541 struct pgpath
*pgpath
= mpio
->pgpath
;
1542 int r
= DM_ENDIO_DONE
;
1545 * We don't queue any clone request inside the multipath target
1546 * during end I/O handling, since those clone requests don't have
1547 * bio clones. If we queue them inside the multipath target,
1548 * we need to make bio clones, that requires memory allocation.
1549 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1550 * don't have bio clones.)
1551 * Instead of queueing the clone request here, we queue the original
1552 * request into dm core, which will remake a clone request and
1553 * clone bios for it and resubmit it later.
1555 if (error
&& blk_path_error(error
)) {
1556 struct multipath
*m
= ti
->private;
1558 if (error
== BLK_STS_RESOURCE
)
1559 r
= DM_ENDIO_DELAY_REQUEUE
;
1561 r
= DM_ENDIO_REQUEUE
;
1566 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1567 !must_push_back_rq(m
)) {
1568 if (error
== BLK_STS_IOERR
)
1570 /* complete with the original error */
1576 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1578 if (ps
->type
->end_io
)
1579 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1585 static int multipath_end_io_bio(struct dm_target
*ti
, struct bio
*clone
,
1586 blk_status_t
*error
)
1588 struct multipath
*m
= ti
->private;
1589 struct dm_mpath_io
*mpio
= get_mpio_from_bio(clone
);
1590 struct pgpath
*pgpath
= mpio
->pgpath
;
1591 unsigned long flags
;
1592 int r
= DM_ENDIO_DONE
;
1594 if (!*error
|| !blk_path_error(*error
))
1600 if (atomic_read(&m
->nr_valid_paths
) == 0 &&
1601 !test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
)) {
1602 if (must_push_back_bio(m
)) {
1603 r
= DM_ENDIO_REQUEUE
;
1606 *error
= BLK_STS_IOERR
;
1611 spin_lock_irqsave(&m
->lock
, flags
);
1612 bio_list_add(&m
->queued_bios
, clone
);
1613 spin_unlock_irqrestore(&m
->lock
, flags
);
1614 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
))
1615 queue_work(kmultipathd
, &m
->process_queued_bios
);
1617 r
= DM_ENDIO_INCOMPLETE
;
1620 struct path_selector
*ps
= &pgpath
->pg
->ps
;
1622 if (ps
->type
->end_io
)
1623 ps
->type
->end_io(ps
, &pgpath
->path
, mpio
->nr_bytes
);
1630 * Suspend can't complete until all the I/O is processed so if
1631 * the last path fails we must error any remaining I/O.
1632 * Note that if the freeze_bdev fails while suspending, the
1633 * queue_if_no_path state is lost - userspace should reset it.
1635 static void multipath_presuspend(struct dm_target
*ti
)
1637 struct multipath
*m
= ti
->private;
1639 queue_if_no_path(m
, false, true);
1642 static void multipath_postsuspend(struct dm_target
*ti
)
1644 struct multipath
*m
= ti
->private;
1646 mutex_lock(&m
->work_mutex
);
1647 flush_multipath_work(m
);
1648 mutex_unlock(&m
->work_mutex
);
1652 * Restore the queue_if_no_path setting.
1654 static void multipath_resume(struct dm_target
*ti
)
1656 struct multipath
*m
= ti
->private;
1657 unsigned long flags
;
1659 spin_lock_irqsave(&m
->lock
, flags
);
1660 assign_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
,
1661 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH
, &m
->flags
));
1662 spin_unlock_irqrestore(&m
->lock
, flags
);
1666 * Info output has the following format:
1667 * num_multipath_feature_args [multipath_feature_args]*
1668 * num_handler_status_args [handler_status_args]*
1669 * num_groups init_group_number
1670 * [A|D|E num_ps_status_args [ps_status_args]*
1671 * num_paths num_selector_args
1672 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1674 * Table output has the following format (identical to the constructor string):
1675 * num_feature_args [features_args]*
1676 * num_handler_args hw_handler [hw_handler_args]*
1677 * num_groups init_group_number
1678 * [priority selector-name num_ps_args [ps_args]*
1679 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1681 static void multipath_status(struct dm_target
*ti
, status_type_t type
,
1682 unsigned status_flags
, char *result
, unsigned maxlen
)
1685 unsigned long flags
;
1686 struct multipath
*m
= ti
->private;
1687 struct priority_group
*pg
;
1692 spin_lock_irqsave(&m
->lock
, flags
);
1695 if (type
== STATUSTYPE_INFO
)
1696 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO
, &m
->flags
),
1697 atomic_read(&m
->pg_init_count
));
1699 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
) +
1700 (m
->pg_init_retries
> 0) * 2 +
1701 (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
) * 2 +
1702 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
) +
1703 (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) * 2);
1705 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1706 DMEMIT("queue_if_no_path ");
1707 if (m
->pg_init_retries
)
1708 DMEMIT("pg_init_retries %u ", m
->pg_init_retries
);
1709 if (m
->pg_init_delay_msecs
!= DM_PG_INIT_DELAY_DEFAULT
)
1710 DMEMIT("pg_init_delay_msecs %u ", m
->pg_init_delay_msecs
);
1711 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER
, &m
->flags
))
1712 DMEMIT("retain_attached_hw_handler ");
1713 if (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
) {
1714 switch(m
->queue_mode
) {
1715 case DM_TYPE_BIO_BASED
:
1716 DMEMIT("queue_mode bio ");
1725 if (!m
->hw_handler_name
|| type
== STATUSTYPE_INFO
)
1728 DMEMIT("1 %s ", m
->hw_handler_name
);
1730 DMEMIT("%u ", m
->nr_priority_groups
);
1733 pg_num
= m
->next_pg
->pg_num
;
1734 else if (m
->current_pg
)
1735 pg_num
= m
->current_pg
->pg_num
;
1737 pg_num
= (m
->nr_priority_groups
? 1 : 0);
1739 DMEMIT("%u ", pg_num
);
1742 case STATUSTYPE_INFO
:
1743 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1745 state
= 'D'; /* Disabled */
1746 else if (pg
== m
->current_pg
)
1747 state
= 'A'; /* Currently Active */
1749 state
= 'E'; /* Enabled */
1751 DMEMIT("%c ", state
);
1753 if (pg
->ps
.type
->status
)
1754 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1760 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1761 pg
->ps
.type
->info_args
);
1763 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1764 DMEMIT("%s %s %u ", p
->path
.dev
->name
,
1765 p
->is_active
? "A" : "F",
1767 if (pg
->ps
.type
->status
)
1768 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1769 &p
->path
, type
, result
+ sz
,
1775 case STATUSTYPE_TABLE
:
1776 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1777 DMEMIT("%s ", pg
->ps
.type
->name
);
1779 if (pg
->ps
.type
->status
)
1780 sz
+= pg
->ps
.type
->status(&pg
->ps
, NULL
, type
,
1786 DMEMIT("%u %u ", pg
->nr_pgpaths
,
1787 pg
->ps
.type
->table_args
);
1789 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1790 DMEMIT("%s ", p
->path
.dev
->name
);
1791 if (pg
->ps
.type
->status
)
1792 sz
+= pg
->ps
.type
->status(&pg
->ps
,
1793 &p
->path
, type
, result
+ sz
,
1800 spin_unlock_irqrestore(&m
->lock
, flags
);
1803 static int multipath_message(struct dm_target
*ti
, unsigned argc
, char **argv
,
1804 char *result
, unsigned maxlen
)
1808 struct multipath
*m
= ti
->private;
1811 mutex_lock(&m
->work_mutex
);
1813 if (dm_suspended(ti
)) {
1819 if (!strcasecmp(argv
[0], "queue_if_no_path")) {
1820 r
= queue_if_no_path(m
, true, false);
1822 } else if (!strcasecmp(argv
[0], "fail_if_no_path")) {
1823 r
= queue_if_no_path(m
, false, false);
1829 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc
);
1833 if (!strcasecmp(argv
[0], "disable_group")) {
1834 r
= bypass_pg_num(m
, argv
[1], true);
1836 } else if (!strcasecmp(argv
[0], "enable_group")) {
1837 r
= bypass_pg_num(m
, argv
[1], false);
1839 } else if (!strcasecmp(argv
[0], "switch_group")) {
1840 r
= switch_pg_num(m
, argv
[1]);
1842 } else if (!strcasecmp(argv
[0], "reinstate_path"))
1843 action
= reinstate_path
;
1844 else if (!strcasecmp(argv
[0], "fail_path"))
1847 DMWARN("Unrecognised multipath message received: %s", argv
[0]);
1851 r
= dm_get_device(ti
, argv
[1], dm_table_get_mode(ti
->table
), &dev
);
1853 DMWARN("message: error getting device %s",
1858 r
= action_dev(m
, dev
, action
);
1860 dm_put_device(ti
, dev
);
1863 mutex_unlock(&m
->work_mutex
);
1867 static int multipath_prepare_ioctl(struct dm_target
*ti
,
1868 struct block_device
**bdev
)
1870 struct multipath
*m
= ti
->private;
1871 struct pgpath
*current_pgpath
;
1874 current_pgpath
= READ_ONCE(m
->current_pgpath
);
1875 if (!current_pgpath
)
1876 current_pgpath
= choose_pgpath(m
, 0);
1878 if (current_pgpath
) {
1879 if (!test_bit(MPATHF_QUEUE_IO
, &m
->flags
)) {
1880 *bdev
= current_pgpath
->path
.dev
->bdev
;
1883 /* pg_init has not started or completed */
1887 /* No path is available */
1888 if (test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1894 if (r
== -ENOTCONN
) {
1895 if (!READ_ONCE(m
->current_pg
)) {
1896 /* Path status changed, redo selection */
1897 (void) choose_pgpath(m
, 0);
1899 if (test_bit(MPATHF_PG_INIT_REQUIRED
, &m
->flags
))
1900 pg_init_all_paths(m
);
1901 dm_table_run_md_queue_async(m
->ti
->table
);
1902 process_queued_io_list(m
);
1906 * Only pass ioctls through if the device sizes match exactly.
1908 if (!r
&& ti
->len
!= i_size_read((*bdev
)->bd_inode
) >> SECTOR_SHIFT
)
1913 static int multipath_iterate_devices(struct dm_target
*ti
,
1914 iterate_devices_callout_fn fn
, void *data
)
1916 struct multipath
*m
= ti
->private;
1917 struct priority_group
*pg
;
1921 list_for_each_entry(pg
, &m
->priority_groups
, list
) {
1922 list_for_each_entry(p
, &pg
->pgpaths
, list
) {
1923 ret
= fn(ti
, p
->path
.dev
, ti
->begin
, ti
->len
, data
);
1933 static int pgpath_busy(struct pgpath
*pgpath
)
1935 struct request_queue
*q
= bdev_get_queue(pgpath
->path
.dev
->bdev
);
1937 return blk_lld_busy(q
);
1941 * We return "busy", only when we can map I/Os but underlying devices
1942 * are busy (so even if we map I/Os now, the I/Os will wait on
1943 * the underlying queue).
1944 * In other words, if we want to kill I/Os or queue them inside us
1945 * due to map unavailability, we don't return "busy". Otherwise,
1946 * dm core won't give us the I/Os and we can't do what we want.
1948 static int multipath_busy(struct dm_target
*ti
)
1950 bool busy
= false, has_active
= false;
1951 struct multipath
*m
= ti
->private;
1952 struct priority_group
*pg
, *next_pg
;
1953 struct pgpath
*pgpath
;
1955 /* pg_init in progress */
1956 if (atomic_read(&m
->pg_init_in_progress
))
1959 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1960 if (!atomic_read(&m
->nr_valid_paths
) && test_bit(MPATHF_QUEUE_IF_NO_PATH
, &m
->flags
))
1961 return (m
->queue_mode
!= DM_TYPE_REQUEST_BASED
);
1963 /* Guess which priority_group will be used at next mapping time */
1964 pg
= READ_ONCE(m
->current_pg
);
1965 next_pg
= READ_ONCE(m
->next_pg
);
1966 if (unlikely(!READ_ONCE(m
->current_pgpath
) && next_pg
))
1971 * We don't know which pg will be used at next mapping time.
1972 * We don't call choose_pgpath() here to avoid to trigger
1973 * pg_init just by busy checking.
1974 * So we don't know whether underlying devices we will be using
1975 * at next mapping time are busy or not. Just try mapping.
1981 * If there is one non-busy active path at least, the path selector
1982 * will be able to select it. So we consider such a pg as not busy.
1985 list_for_each_entry(pgpath
, &pg
->pgpaths
, list
) {
1986 if (pgpath
->is_active
) {
1988 if (!pgpath_busy(pgpath
)) {
1997 * No active path in this pg, so this pg won't be used and
1998 * the current_pg will be changed at next mapping time.
1999 * We need to try mapping to determine it.
2007 /*-----------------------------------------------------------------
2009 *---------------------------------------------------------------*/
2010 static struct target_type multipath_target
= {
2011 .name
= "multipath",
2012 .version
= {1, 13, 0},
2013 .features
= DM_TARGET_SINGLETON
| DM_TARGET_IMMUTABLE
|
2014 DM_TARGET_PASSES_INTEGRITY
,
2015 .module
= THIS_MODULE
,
2016 .ctr
= multipath_ctr
,
2017 .dtr
= multipath_dtr
,
2018 .clone_and_map_rq
= multipath_clone_and_map
,
2019 .release_clone_rq
= multipath_release_clone
,
2020 .rq_end_io
= multipath_end_io
,
2021 .map
= multipath_map_bio
,
2022 .end_io
= multipath_end_io_bio
,
2023 .presuspend
= multipath_presuspend
,
2024 .postsuspend
= multipath_postsuspend
,
2025 .resume
= multipath_resume
,
2026 .status
= multipath_status
,
2027 .message
= multipath_message
,
2028 .prepare_ioctl
= multipath_prepare_ioctl
,
2029 .iterate_devices
= multipath_iterate_devices
,
2030 .busy
= multipath_busy
,
2033 static int __init
dm_multipath_init(void)
2037 kmultipathd
= alloc_workqueue("kmpathd", WQ_MEM_RECLAIM
, 0);
2039 DMERR("failed to create workqueue kmpathd");
2041 goto bad_alloc_kmultipathd
;
2045 * A separate workqueue is used to handle the device handlers
2046 * to avoid overloading existing workqueue. Overloading the
2047 * old workqueue would also create a bottleneck in the
2048 * path of the storage hardware device activation.
2050 kmpath_handlerd
= alloc_ordered_workqueue("kmpath_handlerd",
2052 if (!kmpath_handlerd
) {
2053 DMERR("failed to create workqueue kmpath_handlerd");
2055 goto bad_alloc_kmpath_handlerd
;
2058 r
= dm_register_target(&multipath_target
);
2060 DMERR("request-based register failed %d", r
);
2062 goto bad_register_target
;
2067 bad_register_target
:
2068 destroy_workqueue(kmpath_handlerd
);
2069 bad_alloc_kmpath_handlerd
:
2070 destroy_workqueue(kmultipathd
);
2071 bad_alloc_kmultipathd
:
2075 static void __exit
dm_multipath_exit(void)
2077 destroy_workqueue(kmpath_handlerd
);
2078 destroy_workqueue(kmultipathd
);
2080 dm_unregister_target(&multipath_target
);
2083 module_init(dm_multipath_init
);
2084 module_exit(dm_multipath_exit
);
2086 MODULE_DESCRIPTION(DM_NAME
" multipath target");
2087 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2088 MODULE_LICENSE("GPL");