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,
548 union map_info *map_context)
550 if (unlikely(map_context)) {
552 * non-NULL map_context means caller is still map
553 * method; must undo multipath_clone_and_map()
555 struct dm_mpath_io *mpio = get_mpio(map_context);
556 struct pgpath *pgpath = mpio->pgpath;
558 if (pgpath && pgpath->pg->ps.type->end_io)
559 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
564 blk_put_request(clone);
568 * Map cloned bios (bio-based multipath)
571 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
573 struct pgpath *pgpath;
577 /* Do we need to select a new pgpath? */
578 pgpath = READ_ONCE(m->current_pgpath);
579 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
580 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
582 /* MPATHF_QUEUE_IO might have been cleared by choose_pgpath. */
583 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
585 if ((pgpath && queue_io) ||
586 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
587 /* Queue for the daemon to resubmit */
588 spin_lock_irqsave(&m->lock, flags);
589 bio_list_add(&m->queued_bios, bio);
590 spin_unlock_irqrestore(&m->lock, flags);
592 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
593 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
594 pg_init_all_paths(m);
596 queue_work(kmultipathd, &m->process_queued_bios);
598 return ERR_PTR(-EAGAIN);
604 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
605 struct dm_mpath_io *mpio)
607 struct pgpath *pgpath = __map_bio(m, bio);
610 return DM_MAPIO_SUBMITTED;
613 if (must_push_back_bio(m))
614 return DM_MAPIO_REQUEUE;
616 return DM_MAPIO_KILL;
619 mpio->pgpath = pgpath;
622 bio_set_dev(bio, pgpath->path.dev->bdev);
623 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
625 if (pgpath->pg->ps.type->start_io)
626 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
629 return DM_MAPIO_REMAPPED;
632 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
634 struct multipath *m = ti->private;
635 struct dm_mpath_io *mpio = NULL;
637 multipath_init_per_bio_data(bio, &mpio);
638 return __multipath_map_bio(m, bio, mpio);
641 static void process_queued_io_list(struct multipath *m)
643 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
644 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
645 else if (m->queue_mode == DM_TYPE_BIO_BASED)
646 queue_work(kmultipathd, &m->process_queued_bios);
649 static void process_queued_bios(struct work_struct *work)
654 struct bio_list bios;
655 struct blk_plug plug;
656 struct multipath *m =
657 container_of(work, struct multipath, process_queued_bios);
659 bio_list_init(&bios);
661 spin_lock_irqsave(&m->lock, flags);
663 if (bio_list_empty(&m->queued_bios)) {
664 spin_unlock_irqrestore(&m->lock, flags);
668 bio_list_merge(&bios, &m->queued_bios);
669 bio_list_init(&m->queued_bios);
671 spin_unlock_irqrestore(&m->lock, flags);
673 blk_start_plug(&plug);
674 while ((bio = bio_list_pop(&bios))) {
675 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
676 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
677 r = __multipath_map_bio(m, bio, mpio);
680 bio->bi_status = BLK_STS_IOERR;
683 case DM_MAPIO_REQUEUE:
684 bio->bi_status = BLK_STS_DM_REQUEUE;
687 case DM_MAPIO_REMAPPED:
688 generic_make_request(bio);
690 case DM_MAPIO_SUBMITTED:
693 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
696 blk_finish_plug(&plug);
700 * If we run out of usable paths, should we queue I/O or error it?
702 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
707 spin_lock_irqsave(&m->lock, flags);
708 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
709 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
710 (!save_old_value && queue_if_no_path));
711 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
712 spin_unlock_irqrestore(&m->lock, flags);
714 if (!queue_if_no_path) {
715 dm_table_run_md_queue_async(m->ti->table);
716 process_queued_io_list(m);
723 * An event is triggered whenever a path is taken out of use.
724 * Includes path failure and PG bypass.
726 static void trigger_event(struct work_struct *work)
728 struct multipath *m =
729 container_of(work, struct multipath, trigger_event);
731 dm_table_event(m->ti->table);
734 /*-----------------------------------------------------------------
735 * Constructor/argument parsing:
736 * <#multipath feature args> [<arg>]*
737 * <#hw_handler args> [hw_handler [<arg>]*]
739 * <initial priority group>
740 * [<selector> <#selector args> [<arg>]*
741 * <#paths> <#per-path selector args>
742 * [<path> [<arg>]* ]+ ]+
743 *---------------------------------------------------------------*/
744 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
745 struct dm_target *ti)
748 struct path_selector_type *pst;
751 static const struct dm_arg _args[] = {
752 {0, 1024, "invalid number of path selector args"},
755 pst = dm_get_path_selector(dm_shift_arg(as));
757 ti->error = "unknown path selector type";
761 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
763 dm_put_path_selector(pst);
767 r = pst->create(&pg->ps, ps_argc, as->argv);
769 dm_put_path_selector(pst);
770 ti->error = "path selector constructor failed";
775 dm_consume_args(as, ps_argc);
780 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
781 const char **attached_handler_name, char **error)
783 struct request_queue *q = bdev_get_queue(bdev);
786 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
788 if (*attached_handler_name) {
790 * Clear any hw_handler_params associated with a
791 * handler that isn't already attached.
793 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
794 kfree(m->hw_handler_params);
795 m->hw_handler_params = NULL;
799 * Reset hw_handler_name to match the attached handler
801 * NB. This modifies the table line to show the actual
802 * handler instead of the original table passed in.
804 kfree(m->hw_handler_name);
805 m->hw_handler_name = *attached_handler_name;
806 *attached_handler_name = NULL;
810 if (m->hw_handler_name) {
811 r = scsi_dh_attach(q, m->hw_handler_name);
813 char b[BDEVNAME_SIZE];
815 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
820 *error = "error attaching hardware handler";
824 if (m->hw_handler_params) {
825 r = scsi_dh_set_params(q, m->hw_handler_params);
827 *error = "unable to set hardware handler parameters";
836 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
837 struct dm_target *ti)
841 struct multipath *m = ti->private;
842 struct request_queue *q;
843 const char *attached_handler_name = NULL;
845 /* we need at least a path arg */
847 ti->error = "no device given";
848 return ERR_PTR(-EINVAL);
853 return ERR_PTR(-ENOMEM);
855 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
858 ti->error = "error getting device";
862 q = bdev_get_queue(p->path.dev->bdev);
863 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
864 if (attached_handler_name || m->hw_handler_name) {
865 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
866 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
867 kfree(attached_handler_name);
869 dm_put_device(ti, p->path.dev);
874 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
876 dm_put_device(ti, p->path.dev);
886 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
889 static const struct dm_arg _args[] = {
890 {1, 1024, "invalid number of paths"},
891 {0, 1024, "invalid number of selector args"}
895 unsigned i, nr_selector_args, nr_args;
896 struct priority_group *pg;
897 struct dm_target *ti = m->ti;
901 ti->error = "not enough priority group arguments";
902 return ERR_PTR(-EINVAL);
905 pg = alloc_priority_group();
907 ti->error = "couldn't allocate priority group";
908 return ERR_PTR(-ENOMEM);
912 r = parse_path_selector(as, pg, ti);
919 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
923 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
927 nr_args = 1 + nr_selector_args;
928 for (i = 0; i < pg->nr_pgpaths; i++) {
929 struct pgpath *pgpath;
930 struct dm_arg_set path_args;
932 if (as->argc < nr_args) {
933 ti->error = "not enough path parameters";
938 path_args.argc = nr_args;
939 path_args.argv = as->argv;
941 pgpath = parse_path(&path_args, &pg->ps, ti);
942 if (IS_ERR(pgpath)) {
948 list_add_tail(&pgpath->list, &pg->pgpaths);
949 dm_consume_args(as, nr_args);
955 free_priority_group(pg, ti);
959 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
963 struct dm_target *ti = m->ti;
965 static const struct dm_arg _args[] = {
966 {0, 1024, "invalid number of hardware handler args"},
969 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
975 if (m->queue_mode == DM_TYPE_BIO_BASED) {
976 dm_consume_args(as, hw_argc);
977 DMERR("bio-based multipath doesn't allow hardware handler args");
981 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
982 if (!m->hw_handler_name)
989 for (i = 0; i <= hw_argc - 2; i++)
990 len += strlen(as->argv[i]) + 1;
991 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
993 ti->error = "memory allocation failed";
997 j = sprintf(p, "%d", hw_argc - 1);
998 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
999 j = sprintf(p, "%s", as->argv[i]);
1001 dm_consume_args(as, hw_argc - 1);
1005 kfree(m->hw_handler_name);
1006 m->hw_handler_name = NULL;
1010 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1014 struct dm_target *ti = m->ti;
1015 const char *arg_name;
1017 static const struct dm_arg _args[] = {
1018 {0, 8, "invalid number of feature args"},
1019 {1, 50, "pg_init_retries must be between 1 and 50"},
1020 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1023 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1031 arg_name = dm_shift_arg(as);
1034 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1035 r = queue_if_no_path(m, true, false);
1039 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1040 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1044 if (!strcasecmp(arg_name, "pg_init_retries") &&
1046 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1051 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1053 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1058 if (!strcasecmp(arg_name, "queue_mode") &&
1060 const char *queue_mode_name = dm_shift_arg(as);
1062 if (!strcasecmp(queue_mode_name, "bio"))
1063 m->queue_mode = DM_TYPE_BIO_BASED;
1064 else if (!strcasecmp(queue_mode_name, "rq") ||
1065 !strcasecmp(queue_mode_name, "mq"))
1066 m->queue_mode = DM_TYPE_REQUEST_BASED;
1068 ti->error = "Unknown 'queue_mode' requested";
1075 ti->error = "Unrecognised multipath feature request";
1077 } while (argc && !r);
1082 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1084 /* target arguments */
1085 static const struct dm_arg _args[] = {
1086 {0, 1024, "invalid number of priority groups"},
1087 {0, 1024, "invalid initial priority group number"},
1091 struct multipath *m;
1092 struct dm_arg_set as;
1093 unsigned pg_count = 0;
1094 unsigned next_pg_num;
1099 m = alloc_multipath(ti);
1101 ti->error = "can't allocate multipath";
1105 r = parse_features(&as, m);
1109 r = alloc_multipath_stage2(ti, m);
1113 r = parse_hw_handler(&as, m);
1117 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1121 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1125 if ((!m->nr_priority_groups && next_pg_num) ||
1126 (m->nr_priority_groups && !next_pg_num)) {
1127 ti->error = "invalid initial priority group";
1132 /* parse the priority groups */
1134 struct priority_group *pg;
1135 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1137 pg = parse_priority_group(&as, m);
1143 nr_valid_paths += pg->nr_pgpaths;
1144 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1146 list_add_tail(&pg->list, &m->priority_groups);
1148 pg->pg_num = pg_count;
1153 if (pg_count != m->nr_priority_groups) {
1154 ti->error = "priority group count mismatch";
1159 ti->num_flush_bios = 1;
1160 ti->num_discard_bios = 1;
1161 ti->num_write_same_bios = 1;
1162 ti->num_write_zeroes_bios = 1;
1163 if (m->queue_mode == DM_TYPE_BIO_BASED)
1164 ti->per_io_data_size = multipath_per_bio_data_size();
1166 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1175 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1180 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1182 if (!atomic_read(&m->pg_init_in_progress))
1187 finish_wait(&m->pg_init_wait, &wait);
1190 static void flush_multipath_work(struct multipath *m)
1192 if (m->hw_handler_name) {
1193 unsigned long flags;
1195 if (!atomic_read(&m->pg_init_in_progress))
1198 spin_lock_irqsave(&m->lock, flags);
1199 if (atomic_read(&m->pg_init_in_progress) &&
1200 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1201 spin_unlock_irqrestore(&m->lock, flags);
1203 flush_workqueue(kmpath_handlerd);
1204 multipath_wait_for_pg_init_completion(m);
1206 spin_lock_irqsave(&m->lock, flags);
1207 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1209 spin_unlock_irqrestore(&m->lock, flags);
1212 if (m->queue_mode == DM_TYPE_BIO_BASED)
1213 flush_work(&m->process_queued_bios);
1214 flush_work(&m->trigger_event);
1217 static void multipath_dtr(struct dm_target *ti)
1219 struct multipath *m = ti->private;
1221 flush_multipath_work(m);
1226 * Take a path out of use.
1228 static int fail_path(struct pgpath *pgpath)
1230 unsigned long flags;
1231 struct multipath *m = pgpath->pg->m;
1233 spin_lock_irqsave(&m->lock, flags);
1235 if (!pgpath->is_active)
1238 DMWARN("Failing path %s.", pgpath->path.dev->name);
1240 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1241 pgpath->is_active = false;
1242 pgpath->fail_count++;
1244 atomic_dec(&m->nr_valid_paths);
1246 if (pgpath == m->current_pgpath)
1247 m->current_pgpath = NULL;
1249 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1250 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1252 schedule_work(&m->trigger_event);
1255 spin_unlock_irqrestore(&m->lock, flags);
1261 * Reinstate a previously-failed path
1263 static int reinstate_path(struct pgpath *pgpath)
1265 int r = 0, run_queue = 0;
1266 unsigned long flags;
1267 struct multipath *m = pgpath->pg->m;
1268 unsigned nr_valid_paths;
1270 spin_lock_irqsave(&m->lock, flags);
1272 if (pgpath->is_active)
1275 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1277 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1281 pgpath->is_active = true;
1283 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1284 if (nr_valid_paths == 1) {
1285 m->current_pgpath = NULL;
1287 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1288 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1289 atomic_inc(&m->pg_init_in_progress);
1292 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1293 pgpath->path.dev->name, nr_valid_paths);
1295 schedule_work(&m->trigger_event);
1298 spin_unlock_irqrestore(&m->lock, flags);
1300 dm_table_run_md_queue_async(m->ti->table);
1301 process_queued_io_list(m);
1308 * Fail or reinstate all paths that match the provided struct dm_dev.
1310 static int action_dev(struct multipath *m, struct dm_dev *dev,
1314 struct pgpath *pgpath;
1315 struct priority_group *pg;
1317 list_for_each_entry(pg, &m->priority_groups, list) {
1318 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1319 if (pgpath->path.dev == dev)
1328 * Temporarily try to avoid having to use the specified PG
1330 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1333 unsigned long flags;
1335 spin_lock_irqsave(&m->lock, flags);
1337 pg->bypassed = bypassed;
1338 m->current_pgpath = NULL;
1339 m->current_pg = NULL;
1341 spin_unlock_irqrestore(&m->lock, flags);
1343 schedule_work(&m->trigger_event);
1347 * Switch to using the specified PG from the next I/O that gets mapped
1349 static int switch_pg_num(struct multipath *m, const char *pgstr)
1351 struct priority_group *pg;
1353 unsigned long flags;
1356 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1357 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1358 DMWARN("invalid PG number supplied to switch_pg_num");
1362 spin_lock_irqsave(&m->lock, flags);
1363 list_for_each_entry(pg, &m->priority_groups, list) {
1364 pg->bypassed = false;
1368 m->current_pgpath = NULL;
1369 m->current_pg = NULL;
1372 spin_unlock_irqrestore(&m->lock, flags);
1374 schedule_work(&m->trigger_event);
1379 * Set/clear bypassed status of a PG.
1380 * PGs are numbered upwards from 1 in the order they were declared.
1382 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1384 struct priority_group *pg;
1388 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1389 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1390 DMWARN("invalid PG number supplied to bypass_pg");
1394 list_for_each_entry(pg, &m->priority_groups, list) {
1399 bypass_pg(m, pg, bypassed);
1404 * Should we retry pg_init immediately?
1406 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1408 unsigned long flags;
1409 bool limit_reached = false;
1411 spin_lock_irqsave(&m->lock, flags);
1413 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1414 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1415 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1417 limit_reached = true;
1419 spin_unlock_irqrestore(&m->lock, flags);
1421 return limit_reached;
1424 static void pg_init_done(void *data, int errors)
1426 struct pgpath *pgpath = data;
1427 struct priority_group *pg = pgpath->pg;
1428 struct multipath *m = pg->m;
1429 unsigned long flags;
1430 bool delay_retry = false;
1432 /* device or driver problems */
1437 if (!m->hw_handler_name) {
1441 DMERR("Could not failover the device: Handler scsi_dh_%s "
1442 "Error %d.", m->hw_handler_name, errors);
1444 * Fail path for now, so we do not ping pong
1448 case SCSI_DH_DEV_TEMP_BUSY:
1450 * Probably doing something like FW upgrade on the
1451 * controller so try the other pg.
1453 bypass_pg(m, pg, true);
1456 /* Wait before retrying. */
1459 case SCSI_DH_IMM_RETRY:
1460 case SCSI_DH_RES_TEMP_UNAVAIL:
1461 if (pg_init_limit_reached(m, pgpath))
1465 case SCSI_DH_DEV_OFFLINED:
1468 * We probably do not want to fail the path for a device
1469 * error, but this is what the old dm did. In future
1470 * patches we can do more advanced handling.
1475 spin_lock_irqsave(&m->lock, flags);
1477 if (pgpath == m->current_pgpath) {
1478 DMERR("Could not failover device. Error %d.", errors);
1479 m->current_pgpath = NULL;
1480 m->current_pg = NULL;
1482 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1483 pg->bypassed = false;
1485 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1486 /* Activations of other paths are still on going */
1489 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1491 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1493 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1495 if (__pg_init_all_paths(m))
1498 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1500 process_queued_io_list(m);
1503 * Wake up any thread waiting to suspend.
1505 wake_up(&m->pg_init_wait);
1508 spin_unlock_irqrestore(&m->lock, flags);
1511 static void activate_or_offline_path(struct pgpath *pgpath)
1513 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1515 if (pgpath->is_active && !blk_queue_dying(q))
1516 scsi_dh_activate(q, pg_init_done, pgpath);
1518 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1521 static void activate_path_work(struct work_struct *work)
1523 struct pgpath *pgpath =
1524 container_of(work, struct pgpath, activate_path.work);
1526 activate_or_offline_path(pgpath);
1529 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1530 blk_status_t error, union map_info *map_context)
1532 struct dm_mpath_io *mpio = get_mpio(map_context);
1533 struct pgpath *pgpath = mpio->pgpath;
1534 int r = DM_ENDIO_DONE;
1537 * We don't queue any clone request inside the multipath target
1538 * during end I/O handling, since those clone requests don't have
1539 * bio clones. If we queue them inside the multipath target,
1540 * we need to make bio clones, that requires memory allocation.
1541 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1542 * don't have bio clones.)
1543 * Instead of queueing the clone request here, we queue the original
1544 * request into dm core, which will remake a clone request and
1545 * clone bios for it and resubmit it later.
1547 if (error && blk_path_error(error)) {
1548 struct multipath *m = ti->private;
1550 if (error == BLK_STS_RESOURCE)
1551 r = DM_ENDIO_DELAY_REQUEUE;
1553 r = DM_ENDIO_REQUEUE;
1558 if (atomic_read(&m->nr_valid_paths) == 0 &&
1559 !must_push_back_rq(m)) {
1560 if (error == BLK_STS_IOERR)
1562 /* complete with the original error */
1568 struct path_selector *ps = &pgpath->pg->ps;
1570 if (ps->type->end_io)
1571 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1577 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1578 blk_status_t *error)
1580 struct multipath *m = ti->private;
1581 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1582 struct pgpath *pgpath = mpio->pgpath;
1583 unsigned long flags;
1584 int r = DM_ENDIO_DONE;
1586 if (!*error || !blk_path_error(*error))
1592 if (atomic_read(&m->nr_valid_paths) == 0 &&
1593 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1594 if (must_push_back_bio(m)) {
1595 r = DM_ENDIO_REQUEUE;
1598 *error = BLK_STS_IOERR;
1603 spin_lock_irqsave(&m->lock, flags);
1604 bio_list_add(&m->queued_bios, clone);
1605 spin_unlock_irqrestore(&m->lock, flags);
1606 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1607 queue_work(kmultipathd, &m->process_queued_bios);
1609 r = DM_ENDIO_INCOMPLETE;
1612 struct path_selector *ps = &pgpath->pg->ps;
1614 if (ps->type->end_io)
1615 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1622 * Suspend can't complete until all the I/O is processed so if
1623 * the last path fails we must error any remaining I/O.
1624 * Note that if the freeze_bdev fails while suspending, the
1625 * queue_if_no_path state is lost - userspace should reset it.
1627 static void multipath_presuspend(struct dm_target *ti)
1629 struct multipath *m = ti->private;
1631 queue_if_no_path(m, false, true);
1634 static void multipath_postsuspend(struct dm_target *ti)
1636 struct multipath *m = ti->private;
1638 mutex_lock(&m->work_mutex);
1639 flush_multipath_work(m);
1640 mutex_unlock(&m->work_mutex);
1644 * Restore the queue_if_no_path setting.
1646 static void multipath_resume(struct dm_target *ti)
1648 struct multipath *m = ti->private;
1649 unsigned long flags;
1651 spin_lock_irqsave(&m->lock, flags);
1652 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1653 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1654 spin_unlock_irqrestore(&m->lock, flags);
1658 * Info output has the following format:
1659 * num_multipath_feature_args [multipath_feature_args]*
1660 * num_handler_status_args [handler_status_args]*
1661 * num_groups init_group_number
1662 * [A|D|E num_ps_status_args [ps_status_args]*
1663 * num_paths num_selector_args
1664 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1666 * Table output has the following format (identical to the constructor string):
1667 * num_feature_args [features_args]*
1668 * num_handler_args hw_handler [hw_handler_args]*
1669 * num_groups init_group_number
1670 * [priority selector-name num_ps_args [ps_args]*
1671 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1673 static void multipath_status(struct dm_target *ti, status_type_t type,
1674 unsigned status_flags, char *result, unsigned maxlen)
1677 unsigned long flags;
1678 struct multipath *m = ti->private;
1679 struct priority_group *pg;
1684 spin_lock_irqsave(&m->lock, flags);
1687 if (type == STATUSTYPE_INFO)
1688 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1689 atomic_read(&m->pg_init_count));
1691 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1692 (m->pg_init_retries > 0) * 2 +
1693 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1694 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1695 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1697 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1698 DMEMIT("queue_if_no_path ");
1699 if (m->pg_init_retries)
1700 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1701 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1702 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1703 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1704 DMEMIT("retain_attached_hw_handler ");
1705 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1706 switch(m->queue_mode) {
1707 case DM_TYPE_BIO_BASED:
1708 DMEMIT("queue_mode bio ");
1717 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1720 DMEMIT("1 %s ", m->hw_handler_name);
1722 DMEMIT("%u ", m->nr_priority_groups);
1725 pg_num = m->next_pg->pg_num;
1726 else if (m->current_pg)
1727 pg_num = m->current_pg->pg_num;
1729 pg_num = (m->nr_priority_groups ? 1 : 0);
1731 DMEMIT("%u ", pg_num);
1734 case STATUSTYPE_INFO:
1735 list_for_each_entry(pg, &m->priority_groups, list) {
1737 state = 'D'; /* Disabled */
1738 else if (pg == m->current_pg)
1739 state = 'A'; /* Currently Active */
1741 state = 'E'; /* Enabled */
1743 DMEMIT("%c ", state);
1745 if (pg->ps.type->status)
1746 sz += pg->ps.type->status(&pg->ps, NULL, type,
1752 DMEMIT("%u %u ", pg->nr_pgpaths,
1753 pg->ps.type->info_args);
1755 list_for_each_entry(p, &pg->pgpaths, list) {
1756 DMEMIT("%s %s %u ", p->path.dev->name,
1757 p->is_active ? "A" : "F",
1759 if (pg->ps.type->status)
1760 sz += pg->ps.type->status(&pg->ps,
1761 &p->path, type, result + sz,
1767 case STATUSTYPE_TABLE:
1768 list_for_each_entry(pg, &m->priority_groups, list) {
1769 DMEMIT("%s ", pg->ps.type->name);
1771 if (pg->ps.type->status)
1772 sz += pg->ps.type->status(&pg->ps, NULL, type,
1778 DMEMIT("%u %u ", pg->nr_pgpaths,
1779 pg->ps.type->table_args);
1781 list_for_each_entry(p, &pg->pgpaths, list) {
1782 DMEMIT("%s ", p->path.dev->name);
1783 if (pg->ps.type->status)
1784 sz += pg->ps.type->status(&pg->ps,
1785 &p->path, type, result + sz,
1792 spin_unlock_irqrestore(&m->lock, flags);
1795 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1796 char *result, unsigned maxlen)
1800 struct multipath *m = ti->private;
1803 mutex_lock(&m->work_mutex);
1805 if (dm_suspended(ti)) {
1811 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1812 r = queue_if_no_path(m, true, false);
1814 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1815 r = queue_if_no_path(m, false, false);
1821 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1825 if (!strcasecmp(argv[0], "disable_group")) {
1826 r = bypass_pg_num(m, argv[1], true);
1828 } else if (!strcasecmp(argv[0], "enable_group")) {
1829 r = bypass_pg_num(m, argv[1], false);
1831 } else if (!strcasecmp(argv[0], "switch_group")) {
1832 r = switch_pg_num(m, argv[1]);
1834 } else if (!strcasecmp(argv[0], "reinstate_path"))
1835 action = reinstate_path;
1836 else if (!strcasecmp(argv[0], "fail_path"))
1839 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1843 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1845 DMWARN("message: error getting device %s",
1850 r = action_dev(m, dev, action);
1852 dm_put_device(ti, dev);
1855 mutex_unlock(&m->work_mutex);
1859 static int multipath_prepare_ioctl(struct dm_target *ti,
1860 struct block_device **bdev)
1862 struct multipath *m = ti->private;
1863 struct pgpath *current_pgpath;
1866 current_pgpath = READ_ONCE(m->current_pgpath);
1867 if (!current_pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
1868 current_pgpath = choose_pgpath(m, 0);
1870 if (current_pgpath) {
1871 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1872 *bdev = current_pgpath->path.dev->bdev;
1875 /* pg_init has not started or completed */
1879 /* No path is available */
1880 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1886 if (r == -ENOTCONN) {
1887 if (!READ_ONCE(m->current_pg)) {
1888 /* Path status changed, redo selection */
1889 (void) choose_pgpath(m, 0);
1891 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1892 pg_init_all_paths(m);
1893 dm_table_run_md_queue_async(m->ti->table);
1894 process_queued_io_list(m);
1898 * Only pass ioctls through if the device sizes match exactly.
1900 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1905 static int multipath_iterate_devices(struct dm_target *ti,
1906 iterate_devices_callout_fn fn, void *data)
1908 struct multipath *m = ti->private;
1909 struct priority_group *pg;
1913 list_for_each_entry(pg, &m->priority_groups, list) {
1914 list_for_each_entry(p, &pg->pgpaths, list) {
1915 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1925 static int pgpath_busy(struct pgpath *pgpath)
1927 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1929 return blk_lld_busy(q);
1933 * We return "busy", only when we can map I/Os but underlying devices
1934 * are busy (so even if we map I/Os now, the I/Os will wait on
1935 * the underlying queue).
1936 * In other words, if we want to kill I/Os or queue them inside us
1937 * due to map unavailability, we don't return "busy". Otherwise,
1938 * dm core won't give us the I/Os and we can't do what we want.
1940 static int multipath_busy(struct dm_target *ti)
1942 bool busy = false, has_active = false;
1943 struct multipath *m = ti->private;
1944 struct priority_group *pg, *next_pg;
1945 struct pgpath *pgpath;
1947 /* pg_init in progress */
1948 if (atomic_read(&m->pg_init_in_progress))
1951 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1952 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1953 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
1955 /* Guess which priority_group will be used at next mapping time */
1956 pg = READ_ONCE(m->current_pg);
1957 next_pg = READ_ONCE(m->next_pg);
1958 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1963 * We don't know which pg will be used at next mapping time.
1964 * We don't call choose_pgpath() here to avoid to trigger
1965 * pg_init just by busy checking.
1966 * So we don't know whether underlying devices we will be using
1967 * at next mapping time are busy or not. Just try mapping.
1973 * If there is one non-busy active path at least, the path selector
1974 * will be able to select it. So we consider such a pg as not busy.
1977 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1978 if (pgpath->is_active) {
1980 if (!pgpath_busy(pgpath)) {
1989 * No active path in this pg, so this pg won't be used and
1990 * the current_pg will be changed at next mapping time.
1991 * We need to try mapping to determine it.
1999 /*-----------------------------------------------------------------
2001 *---------------------------------------------------------------*/
2002 static struct target_type multipath_target = {
2003 .name = "multipath",
2004 .version = {1, 13, 0},
2005 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2006 DM_TARGET_PASSES_INTEGRITY,
2007 .module = THIS_MODULE,
2008 .ctr = multipath_ctr,
2009 .dtr = multipath_dtr,
2010 .clone_and_map_rq = multipath_clone_and_map,
2011 .release_clone_rq = multipath_release_clone,
2012 .rq_end_io = multipath_end_io,
2013 .map = multipath_map_bio,
2014 .end_io = multipath_end_io_bio,
2015 .presuspend = multipath_presuspend,
2016 .postsuspend = multipath_postsuspend,
2017 .resume = multipath_resume,
2018 .status = multipath_status,
2019 .message = multipath_message,
2020 .prepare_ioctl = multipath_prepare_ioctl,
2021 .iterate_devices = multipath_iterate_devices,
2022 .busy = multipath_busy,
2025 static int __init dm_multipath_init(void)
2029 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2031 DMERR("failed to create workqueue kmpathd");
2033 goto bad_alloc_kmultipathd;
2037 * A separate workqueue is used to handle the device handlers
2038 * to avoid overloading existing workqueue. Overloading the
2039 * old workqueue would also create a bottleneck in the
2040 * path of the storage hardware device activation.
2042 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2044 if (!kmpath_handlerd) {
2045 DMERR("failed to create workqueue kmpath_handlerd");
2047 goto bad_alloc_kmpath_handlerd;
2050 r = dm_register_target(&multipath_target);
2052 DMERR("request-based register failed %d", r);
2054 goto bad_register_target;
2059 bad_register_target:
2060 destroy_workqueue(kmpath_handlerd);
2061 bad_alloc_kmpath_handlerd:
2062 destroy_workqueue(kmultipathd);
2063 bad_alloc_kmultipathd:
2067 static void __exit dm_multipath_exit(void)
2069 destroy_workqueue(kmpath_handlerd);
2070 destroy_workqueue(kmultipathd);
2072 dm_unregister_target(&multipath_target);
2075 module_init(dm_multipath_init);
2076 module_exit(dm_multipath_exit);
2078 MODULE_DESCRIPTION(DM_NAME " multipath target");
2079 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2080 MODULE_LICENSE("GPL");