2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
35 struct list_head list;
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
41 struct delayed_work activate_path;
43 bool is_active:1; /* Path status */
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
52 struct priority_group {
53 struct list_head list;
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
62 bool bypassed:1; /* Temporarily bypass this PG? */
65 /* Multipath context */
67 struct list_head list;
70 const char *hw_handler_name;
71 char *hw_handler_params;
75 unsigned nr_priority_groups;
76 struct list_head priority_groups;
78 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned long flags; /* Multipath state flags */
86 unsigned pg_init_retries; /* Number of times to retry pg_init */
87 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
89 atomic_t nr_valid_paths; /* Total number of usable paths */
90 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
91 atomic_t pg_init_count; /* Number of times pg_init called */
93 enum dm_queue_mode queue_mode;
95 struct mutex work_mutex;
96 struct work_struct trigger_event;
98 struct work_struct process_queued_bios;
99 struct bio_list queued_bios;
103 * Context information attached to each io we process.
106 struct pgpath *pgpath;
110 typedef int (*action_fn) (struct pgpath *pgpath);
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void trigger_event(struct work_struct *work);
114 static void activate_or_offline_path(struct pgpath *pgpath);
115 static void activate_path_work(struct work_struct *work);
116 static void process_queued_bios(struct work_struct *work);
118 /*-----------------------------------------------
119 * Multipath state flags.
120 *-----------------------------------------------*/
122 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
134 static struct pgpath *alloc_pgpath(void)
136 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
139 pgpath->is_active = true;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
146 static void free_pgpath(struct pgpath *pgpath)
151 static struct priority_group *alloc_priority_group(void)
153 struct priority_group *pg;
155 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
158 INIT_LIST_HEAD(&pg->pgpaths);
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
165 struct pgpath *pgpath, *tmp;
167 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168 list_del(&pgpath->list);
169 dm_put_device(ti, pgpath->path.dev);
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
177 struct path_selector *ps = &pg->ps;
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
184 free_pgpaths(&pg->pgpaths, ti);
188 static struct multipath *alloc_multipath(struct dm_target *ti)
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
194 INIT_LIST_HEAD(&m->priority_groups);
195 spin_lock_init(&m->lock);
196 set_bit(MPATHF_QUEUE_IO, &m->flags);
197 atomic_set(&m->nr_valid_paths, 0);
198 atomic_set(&m->pg_init_in_progress, 0);
199 atomic_set(&m->pg_init_count, 0);
200 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
201 INIT_WORK(&m->trigger_event, trigger_event);
202 init_waitqueue_head(&m->pg_init_wait);
203 mutex_init(&m->work_mutex);
205 m->queue_mode = DM_TYPE_NONE;
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
216 if (m->queue_mode == DM_TYPE_NONE) {
218 * Default to request-based.
220 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
223 m->queue_mode = DM_TYPE_REQUEST_BASED;
224 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225 INIT_WORK(&m->process_queued_bios, process_queued_bios);
227 * bio-based doesn't support any direct scsi_dh management;
228 * it just discovers if a scsi_dh is attached.
230 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
233 dm_table_set_type(ti->table, m->queue_mode);
238 static void free_multipath(struct multipath *m)
240 struct priority_group *pg, *tmp;
242 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
244 free_priority_group(pg, m->ti);
247 kfree(m->hw_handler_name);
248 kfree(m->hw_handler_params);
252 static struct dm_mpath_io *get_mpio(union map_info *info)
257 static size_t multipath_per_bio_data_size(void)
259 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
264 return dm_per_bio_data(bio, multipath_per_bio_data_size());
267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
269 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271 void *bio_details = mpio + 1;
276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
277 struct dm_bio_details **bio_details_p)
279 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
282 memset(mpio, 0, sizeof(*mpio));
283 memset(bio_details, 0, sizeof(*bio_details));
284 dm_bio_record(bio_details, bio);
289 *bio_details_p = bio_details;
292 /*-----------------------------------------------
294 *-----------------------------------------------*/
296 static int __pg_init_all_paths(struct multipath *m)
298 struct pgpath *pgpath;
299 unsigned long pg_init_delay = 0;
301 lockdep_assert_held(&m->lock);
303 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
306 atomic_inc(&m->pg_init_count);
307 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
309 /* Check here to reset pg_init_required */
313 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
314 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
315 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
316 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
317 /* Skip failed paths */
318 if (!pgpath->is_active)
320 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
322 atomic_inc(&m->pg_init_in_progress);
324 return atomic_read(&m->pg_init_in_progress);
327 static int pg_init_all_paths(struct multipath *m)
332 spin_lock_irqsave(&m->lock, flags);
333 ret = __pg_init_all_paths(m);
334 spin_unlock_irqrestore(&m->lock, flags);
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
343 /* Must we initialise the PG first, and queue I/O till it's ready? */
344 if (m->hw_handler_name) {
345 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346 set_bit(MPATHF_QUEUE_IO, &m->flags);
348 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349 clear_bit(MPATHF_QUEUE_IO, &m->flags);
352 atomic_set(&m->pg_init_count, 0);
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356 struct priority_group *pg,
360 struct dm_path *path;
361 struct pgpath *pgpath;
363 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
365 return ERR_PTR(-ENXIO);
367 pgpath = path_to_pgpath(path);
369 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
370 /* Only update current_pgpath if pg changed */
371 spin_lock_irqsave(&m->lock, flags);
372 m->current_pgpath = pgpath;
374 spin_unlock_irqrestore(&m->lock, flags);
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
383 struct priority_group *pg;
384 struct pgpath *pgpath;
385 unsigned bypassed = 1;
387 if (!atomic_read(&m->nr_valid_paths)) {
388 clear_bit(MPATHF_QUEUE_IO, &m->flags);
392 /* Were we instructed to switch PG? */
393 if (READ_ONCE(m->next_pg)) {
394 spin_lock_irqsave(&m->lock, flags);
397 spin_unlock_irqrestore(&m->lock, flags);
398 goto check_current_pg;
401 spin_unlock_irqrestore(&m->lock, flags);
402 pgpath = choose_path_in_pg(m, pg, nr_bytes);
403 if (!IS_ERR_OR_NULL(pgpath))
407 /* Don't change PG until it has no remaining paths */
409 pg = READ_ONCE(m->current_pg);
411 pgpath = choose_path_in_pg(m, pg, nr_bytes);
412 if (!IS_ERR_OR_NULL(pgpath))
417 * Loop through priority groups until we find a valid path.
418 * First time we skip PGs marked 'bypassed'.
419 * Second time we only try the ones we skipped, but set
420 * pg_init_delay_retry so we do not hammer controllers.
423 list_for_each_entry(pg, &m->priority_groups, list) {
424 if (pg->bypassed == !!bypassed)
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath)) {
429 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
433 } while (bypassed--);
436 spin_lock_irqsave(&m->lock, flags);
437 m->current_pgpath = NULL;
438 m->current_pg = NULL;
439 spin_unlock_irqrestore(&m->lock, flags);
445 * dm_report_EIO() is a macro instead of a function to make pr_debug()
446 * report the function name and line number of the function from which
447 * it has been invoked.
449 #define dm_report_EIO(m) \
451 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
453 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454 dm_device_name(md), \
455 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
456 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
457 dm_noflush_suspending((m)->ti)); \
461 * Check whether bios must be queued in the device-mapper core rather
462 * than here in the target.
464 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
465 * the same value then we are not between multipath_presuspend()
466 * and multipath_resume() calls and we have no need to check
467 * for the DMF_NOFLUSH_SUSPENDING flag.
469 static bool __must_push_back(struct multipath *m, unsigned long flags)
471 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
472 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
473 dm_noflush_suspending(m->ti));
477 * Following functions use READ_ONCE to get atomic access to
478 * all m->flags to avoid taking spinlock
480 static bool must_push_back_rq(struct multipath *m)
482 unsigned long flags = READ_ONCE(m->flags);
483 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
486 static bool must_push_back_bio(struct multipath *m)
488 unsigned long flags = READ_ONCE(m->flags);
489 return __must_push_back(m, flags);
493 * Map cloned requests (request-based multipath)
495 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
496 union map_info *map_context,
497 struct request **__clone)
499 struct multipath *m = ti->private;
500 size_t nr_bytes = blk_rq_bytes(rq);
501 struct pgpath *pgpath;
502 struct block_device *bdev;
503 struct dm_mpath_io *mpio = get_mpio(map_context);
504 struct request_queue *q;
505 struct request *clone;
507 /* Do we need to select a new pgpath? */
508 pgpath = READ_ONCE(m->current_pgpath);
509 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
510 pgpath = choose_pgpath(m, nr_bytes);
513 if (must_push_back_rq(m))
514 return DM_MAPIO_DELAY_REQUEUE;
515 dm_report_EIO(m); /* Failed */
516 return DM_MAPIO_KILL;
517 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
518 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
519 if (pg_init_all_paths(m))
520 return DM_MAPIO_DELAY_REQUEUE;
521 return DM_MAPIO_REQUEUE;
524 memset(mpio, 0, sizeof(*mpio));
525 mpio->pgpath = pgpath;
526 mpio->nr_bytes = nr_bytes;
528 bdev = pgpath->path.dev->bdev;
529 q = bdev_get_queue(bdev);
530 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
532 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
533 bool queue_dying = blk_queue_dying(q);
535 atomic_inc(&m->pg_init_in_progress);
536 activate_or_offline_path(pgpath);
538 return DM_MAPIO_DELAY_REQUEUE;
540 clone->bio = clone->biotail = NULL;
541 clone->rq_disk = bdev->bd_disk;
542 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
545 if (pgpath->pg->ps.type->start_io)
546 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
549 return DM_MAPIO_REMAPPED;
552 static void multipath_release_clone(struct request *clone)
554 blk_put_request(clone);
558 * Map cloned bios (bio-based multipath)
560 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
562 size_t nr_bytes = bio->bi_iter.bi_size;
563 struct pgpath *pgpath;
567 /* Do we need to select a new pgpath? */
568 pgpath = READ_ONCE(m->current_pgpath);
569 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
570 if (!pgpath || !queue_io)
571 pgpath = choose_pgpath(m, nr_bytes);
573 if ((pgpath && queue_io) ||
574 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
575 /* Queue for the daemon to resubmit */
576 spin_lock_irqsave(&m->lock, flags);
577 bio_list_add(&m->queued_bios, bio);
578 spin_unlock_irqrestore(&m->lock, flags);
579 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
580 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
581 pg_init_all_paths(m);
583 queue_work(kmultipathd, &m->process_queued_bios);
584 return DM_MAPIO_SUBMITTED;
588 if (must_push_back_bio(m))
589 return DM_MAPIO_REQUEUE;
591 return DM_MAPIO_KILL;
594 mpio->pgpath = pgpath;
595 mpio->nr_bytes = nr_bytes;
598 bio_set_dev(bio, pgpath->path.dev->bdev);
599 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
601 if (pgpath->pg->ps.type->start_io)
602 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
605 return DM_MAPIO_REMAPPED;
608 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
610 struct multipath *m = ti->private;
611 struct dm_mpath_io *mpio = NULL;
613 multipath_init_per_bio_data(bio, &mpio, NULL);
615 return __multipath_map_bio(m, bio, mpio);
618 static void process_queued_io_list(struct multipath *m)
620 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
621 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
622 else if (m->queue_mode == DM_TYPE_BIO_BASED)
623 queue_work(kmultipathd, &m->process_queued_bios);
626 static void process_queued_bios(struct work_struct *work)
631 struct bio_list bios;
632 struct blk_plug plug;
633 struct multipath *m =
634 container_of(work, struct multipath, process_queued_bios);
636 bio_list_init(&bios);
638 spin_lock_irqsave(&m->lock, flags);
640 if (bio_list_empty(&m->queued_bios)) {
641 spin_unlock_irqrestore(&m->lock, flags);
645 bio_list_merge(&bios, &m->queued_bios);
646 bio_list_init(&m->queued_bios);
648 spin_unlock_irqrestore(&m->lock, flags);
650 blk_start_plug(&plug);
651 while ((bio = bio_list_pop(&bios))) {
652 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
655 bio->bi_status = BLK_STS_IOERR;
658 case DM_MAPIO_REQUEUE:
659 bio->bi_status = BLK_STS_DM_REQUEUE;
662 case DM_MAPIO_REMAPPED:
663 generic_make_request(bio);
668 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
671 blk_finish_plug(&plug);
675 * If we run out of usable paths, should we queue I/O or error it?
677 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
682 spin_lock_irqsave(&m->lock, flags);
683 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
684 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
685 (!save_old_value && queue_if_no_path));
686 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
687 spin_unlock_irqrestore(&m->lock, flags);
689 if (!queue_if_no_path) {
690 dm_table_run_md_queue_async(m->ti->table);
691 process_queued_io_list(m);
698 * An event is triggered whenever a path is taken out of use.
699 * Includes path failure and PG bypass.
701 static void trigger_event(struct work_struct *work)
703 struct multipath *m =
704 container_of(work, struct multipath, trigger_event);
706 dm_table_event(m->ti->table);
709 /*-----------------------------------------------------------------
710 * Constructor/argument parsing:
711 * <#multipath feature args> [<arg>]*
712 * <#hw_handler args> [hw_handler [<arg>]*]
714 * <initial priority group>
715 * [<selector> <#selector args> [<arg>]*
716 * <#paths> <#per-path selector args>
717 * [<path> [<arg>]* ]+ ]+
718 *---------------------------------------------------------------*/
719 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
720 struct dm_target *ti)
723 struct path_selector_type *pst;
726 static const struct dm_arg _args[] = {
727 {0, 1024, "invalid number of path selector args"},
730 pst = dm_get_path_selector(dm_shift_arg(as));
732 ti->error = "unknown path selector type";
736 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
738 dm_put_path_selector(pst);
742 r = pst->create(&pg->ps, ps_argc, as->argv);
744 dm_put_path_selector(pst);
745 ti->error = "path selector constructor failed";
750 dm_consume_args(as, ps_argc);
755 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
756 struct dm_target *ti)
760 struct multipath *m = ti->private;
761 struct request_queue *q = NULL;
762 const char *attached_handler_name;
764 /* we need at least a path arg */
766 ti->error = "no device given";
767 return ERR_PTR(-EINVAL);
772 return ERR_PTR(-ENOMEM);
774 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
777 ti->error = "error getting device";
781 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
782 q = bdev_get_queue(p->path.dev->bdev);
784 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
786 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
787 if (attached_handler_name) {
789 * Clear any hw_handler_params associated with a
790 * handler that isn't already attached.
792 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
793 kfree(m->hw_handler_params);
794 m->hw_handler_params = NULL;
798 * Reset hw_handler_name to match the attached handler
800 * NB. This modifies the table line to show the actual
801 * handler instead of the original table passed in.
803 kfree(m->hw_handler_name);
804 m->hw_handler_name = attached_handler_name;
808 if (m->hw_handler_name) {
809 r = scsi_dh_attach(q, m->hw_handler_name);
811 char b[BDEVNAME_SIZE];
813 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
814 bdevname(p->path.dev->bdev, b));
818 ti->error = "error attaching hardware handler";
819 dm_put_device(ti, p->path.dev);
823 if (m->hw_handler_params) {
824 r = scsi_dh_set_params(q, m->hw_handler_params);
826 ti->error = "unable to set hardware "
827 "handler parameters";
828 dm_put_device(ti, p->path.dev);
834 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
836 dm_put_device(ti, p->path.dev);
847 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
850 static const struct dm_arg _args[] = {
851 {1, 1024, "invalid number of paths"},
852 {0, 1024, "invalid number of selector args"}
856 unsigned i, nr_selector_args, nr_args;
857 struct priority_group *pg;
858 struct dm_target *ti = m->ti;
862 ti->error = "not enough priority group arguments";
863 return ERR_PTR(-EINVAL);
866 pg = alloc_priority_group();
868 ti->error = "couldn't allocate priority group";
869 return ERR_PTR(-ENOMEM);
873 r = parse_path_selector(as, pg, ti);
880 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
884 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
888 nr_args = 1 + nr_selector_args;
889 for (i = 0; i < pg->nr_pgpaths; i++) {
890 struct pgpath *pgpath;
891 struct dm_arg_set path_args;
893 if (as->argc < nr_args) {
894 ti->error = "not enough path parameters";
899 path_args.argc = nr_args;
900 path_args.argv = as->argv;
902 pgpath = parse_path(&path_args, &pg->ps, ti);
903 if (IS_ERR(pgpath)) {
909 list_add_tail(&pgpath->list, &pg->pgpaths);
910 dm_consume_args(as, nr_args);
916 free_priority_group(pg, ti);
920 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
924 struct dm_target *ti = m->ti;
926 static const struct dm_arg _args[] = {
927 {0, 1024, "invalid number of hardware handler args"},
930 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
936 if (m->queue_mode == DM_TYPE_BIO_BASED) {
937 dm_consume_args(as, hw_argc);
938 DMERR("bio-based multipath doesn't allow hardware handler args");
942 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
943 if (!m->hw_handler_name)
950 for (i = 0; i <= hw_argc - 2; i++)
951 len += strlen(as->argv[i]) + 1;
952 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
954 ti->error = "memory allocation failed";
958 j = sprintf(p, "%d", hw_argc - 1);
959 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
960 j = sprintf(p, "%s", as->argv[i]);
962 dm_consume_args(as, hw_argc - 1);
966 kfree(m->hw_handler_name);
967 m->hw_handler_name = NULL;
971 static int parse_features(struct dm_arg_set *as, struct multipath *m)
975 struct dm_target *ti = m->ti;
976 const char *arg_name;
978 static const struct dm_arg _args[] = {
979 {0, 8, "invalid number of feature args"},
980 {1, 50, "pg_init_retries must be between 1 and 50"},
981 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
984 r = dm_read_arg_group(_args, as, &argc, &ti->error);
992 arg_name = dm_shift_arg(as);
995 if (!strcasecmp(arg_name, "queue_if_no_path")) {
996 r = queue_if_no_path(m, true, false);
1000 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1001 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1005 if (!strcasecmp(arg_name, "pg_init_retries") &&
1007 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1012 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1014 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1019 if (!strcasecmp(arg_name, "queue_mode") &&
1021 const char *queue_mode_name = dm_shift_arg(as);
1023 if (!strcasecmp(queue_mode_name, "bio"))
1024 m->queue_mode = DM_TYPE_BIO_BASED;
1025 else if (!strcasecmp(queue_mode_name, "rq"))
1026 m->queue_mode = DM_TYPE_REQUEST_BASED;
1027 else if (!strcasecmp(queue_mode_name, "mq"))
1028 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1030 ti->error = "Unknown 'queue_mode' requested";
1037 ti->error = "Unrecognised multipath feature request";
1039 } while (argc && !r);
1044 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1046 /* target arguments */
1047 static const struct dm_arg _args[] = {
1048 {0, 1024, "invalid number of priority groups"},
1049 {0, 1024, "invalid initial priority group number"},
1053 struct multipath *m;
1054 struct dm_arg_set as;
1055 unsigned pg_count = 0;
1056 unsigned next_pg_num;
1061 m = alloc_multipath(ti);
1063 ti->error = "can't allocate multipath";
1067 r = parse_features(&as, m);
1071 r = alloc_multipath_stage2(ti, m);
1075 r = parse_hw_handler(&as, m);
1079 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1083 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1087 if ((!m->nr_priority_groups && next_pg_num) ||
1088 (m->nr_priority_groups && !next_pg_num)) {
1089 ti->error = "invalid initial priority group";
1094 /* parse the priority groups */
1096 struct priority_group *pg;
1097 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1099 pg = parse_priority_group(&as, m);
1105 nr_valid_paths += pg->nr_pgpaths;
1106 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1108 list_add_tail(&pg->list, &m->priority_groups);
1110 pg->pg_num = pg_count;
1115 if (pg_count != m->nr_priority_groups) {
1116 ti->error = "priority group count mismatch";
1121 ti->num_flush_bios = 1;
1122 ti->num_discard_bios = 1;
1123 ti->num_write_same_bios = 1;
1124 ti->num_write_zeroes_bios = 1;
1125 if (m->queue_mode == DM_TYPE_BIO_BASED)
1126 ti->per_io_data_size = multipath_per_bio_data_size();
1128 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1137 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1142 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1144 if (!atomic_read(&m->pg_init_in_progress))
1149 finish_wait(&m->pg_init_wait, &wait);
1152 static void flush_multipath_work(struct multipath *m)
1154 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1155 smp_mb__after_atomic();
1157 flush_workqueue(kmpath_handlerd);
1158 multipath_wait_for_pg_init_completion(m);
1159 flush_workqueue(kmultipathd);
1160 flush_work(&m->trigger_event);
1162 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1163 smp_mb__after_atomic();
1166 static void multipath_dtr(struct dm_target *ti)
1168 struct multipath *m = ti->private;
1170 flush_multipath_work(m);
1175 * Take a path out of use.
1177 static int fail_path(struct pgpath *pgpath)
1179 unsigned long flags;
1180 struct multipath *m = pgpath->pg->m;
1182 spin_lock_irqsave(&m->lock, flags);
1184 if (!pgpath->is_active)
1187 DMWARN("Failing path %s.", pgpath->path.dev->name);
1189 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1190 pgpath->is_active = false;
1191 pgpath->fail_count++;
1193 atomic_dec(&m->nr_valid_paths);
1195 if (pgpath == m->current_pgpath)
1196 m->current_pgpath = NULL;
1198 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1199 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1201 schedule_work(&m->trigger_event);
1204 spin_unlock_irqrestore(&m->lock, flags);
1210 * Reinstate a previously-failed path
1212 static int reinstate_path(struct pgpath *pgpath)
1214 int r = 0, run_queue = 0;
1215 unsigned long flags;
1216 struct multipath *m = pgpath->pg->m;
1217 unsigned nr_valid_paths;
1219 spin_lock_irqsave(&m->lock, flags);
1221 if (pgpath->is_active)
1224 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1226 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1230 pgpath->is_active = true;
1232 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1233 if (nr_valid_paths == 1) {
1234 m->current_pgpath = NULL;
1236 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1237 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1238 atomic_inc(&m->pg_init_in_progress);
1241 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1242 pgpath->path.dev->name, nr_valid_paths);
1244 schedule_work(&m->trigger_event);
1247 spin_unlock_irqrestore(&m->lock, flags);
1249 dm_table_run_md_queue_async(m->ti->table);
1250 process_queued_io_list(m);
1257 * Fail or reinstate all paths that match the provided struct dm_dev.
1259 static int action_dev(struct multipath *m, struct dm_dev *dev,
1263 struct pgpath *pgpath;
1264 struct priority_group *pg;
1266 list_for_each_entry(pg, &m->priority_groups, list) {
1267 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1268 if (pgpath->path.dev == dev)
1277 * Temporarily try to avoid having to use the specified PG
1279 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1282 unsigned long flags;
1284 spin_lock_irqsave(&m->lock, flags);
1286 pg->bypassed = bypassed;
1287 m->current_pgpath = NULL;
1288 m->current_pg = NULL;
1290 spin_unlock_irqrestore(&m->lock, flags);
1292 schedule_work(&m->trigger_event);
1296 * Switch to using the specified PG from the next I/O that gets mapped
1298 static int switch_pg_num(struct multipath *m, const char *pgstr)
1300 struct priority_group *pg;
1302 unsigned long flags;
1305 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1306 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1307 DMWARN("invalid PG number supplied to switch_pg_num");
1311 spin_lock_irqsave(&m->lock, flags);
1312 list_for_each_entry(pg, &m->priority_groups, list) {
1313 pg->bypassed = false;
1317 m->current_pgpath = NULL;
1318 m->current_pg = NULL;
1321 spin_unlock_irqrestore(&m->lock, flags);
1323 schedule_work(&m->trigger_event);
1328 * Set/clear bypassed status of a PG.
1329 * PGs are numbered upwards from 1 in the order they were declared.
1331 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1333 struct priority_group *pg;
1337 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1338 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1339 DMWARN("invalid PG number supplied to bypass_pg");
1343 list_for_each_entry(pg, &m->priority_groups, list) {
1348 bypass_pg(m, pg, bypassed);
1353 * Should we retry pg_init immediately?
1355 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1357 unsigned long flags;
1358 bool limit_reached = false;
1360 spin_lock_irqsave(&m->lock, flags);
1362 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1363 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1364 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1366 limit_reached = true;
1368 spin_unlock_irqrestore(&m->lock, flags);
1370 return limit_reached;
1373 static void pg_init_done(void *data, int errors)
1375 struct pgpath *pgpath = data;
1376 struct priority_group *pg = pgpath->pg;
1377 struct multipath *m = pg->m;
1378 unsigned long flags;
1379 bool delay_retry = false;
1381 /* device or driver problems */
1386 if (!m->hw_handler_name) {
1390 DMERR("Could not failover the device: Handler scsi_dh_%s "
1391 "Error %d.", m->hw_handler_name, errors);
1393 * Fail path for now, so we do not ping pong
1397 case SCSI_DH_DEV_TEMP_BUSY:
1399 * Probably doing something like FW upgrade on the
1400 * controller so try the other pg.
1402 bypass_pg(m, pg, true);
1405 /* Wait before retrying. */
1408 case SCSI_DH_IMM_RETRY:
1409 case SCSI_DH_RES_TEMP_UNAVAIL:
1410 if (pg_init_limit_reached(m, pgpath))
1414 case SCSI_DH_DEV_OFFLINED:
1417 * We probably do not want to fail the path for a device
1418 * error, but this is what the old dm did. In future
1419 * patches we can do more advanced handling.
1424 spin_lock_irqsave(&m->lock, flags);
1426 if (pgpath == m->current_pgpath) {
1427 DMERR("Could not failover device. Error %d.", errors);
1428 m->current_pgpath = NULL;
1429 m->current_pg = NULL;
1431 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1432 pg->bypassed = false;
1434 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1435 /* Activations of other paths are still on going */
1438 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1440 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1442 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1444 if (__pg_init_all_paths(m))
1447 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1449 process_queued_io_list(m);
1452 * Wake up any thread waiting to suspend.
1454 wake_up(&m->pg_init_wait);
1457 spin_unlock_irqrestore(&m->lock, flags);
1460 static void activate_or_offline_path(struct pgpath *pgpath)
1462 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1464 if (pgpath->is_active && !blk_queue_dying(q))
1465 scsi_dh_activate(q, pg_init_done, pgpath);
1467 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1470 static void activate_path_work(struct work_struct *work)
1472 struct pgpath *pgpath =
1473 container_of(work, struct pgpath, activate_path.work);
1475 activate_or_offline_path(pgpath);
1478 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1479 blk_status_t error, union map_info *map_context)
1481 struct dm_mpath_io *mpio = get_mpio(map_context);
1482 struct pgpath *pgpath = mpio->pgpath;
1483 int r = DM_ENDIO_DONE;
1486 * We don't queue any clone request inside the multipath target
1487 * during end I/O handling, since those clone requests don't have
1488 * bio clones. If we queue them inside the multipath target,
1489 * we need to make bio clones, that requires memory allocation.
1490 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1491 * don't have bio clones.)
1492 * Instead of queueing the clone request here, we queue the original
1493 * request into dm core, which will remake a clone request and
1494 * clone bios for it and resubmit it later.
1496 if (error && blk_path_error(error)) {
1497 struct multipath *m = ti->private;
1499 r = DM_ENDIO_REQUEUE;
1504 if (atomic_read(&m->nr_valid_paths) == 0 &&
1505 !must_push_back_rq(m)) {
1506 if (error == BLK_STS_IOERR)
1508 /* complete with the original error */
1514 struct path_selector *ps = &pgpath->pg->ps;
1516 if (ps->type->end_io)
1517 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1523 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1524 blk_status_t *error)
1526 struct multipath *m = ti->private;
1527 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1528 struct pgpath *pgpath = mpio->pgpath;
1529 unsigned long flags;
1530 int r = DM_ENDIO_DONE;
1532 if (!*error || !blk_path_error(*error))
1538 if (atomic_read(&m->nr_valid_paths) == 0 &&
1539 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1540 if (must_push_back_bio(m)) {
1541 r = DM_ENDIO_REQUEUE;
1544 *error = BLK_STS_IOERR;
1549 /* Queue for the daemon to resubmit */
1550 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1552 spin_lock_irqsave(&m->lock, flags);
1553 bio_list_add(&m->queued_bios, clone);
1554 spin_unlock_irqrestore(&m->lock, flags);
1555 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1556 queue_work(kmultipathd, &m->process_queued_bios);
1558 r = DM_ENDIO_INCOMPLETE;
1561 struct path_selector *ps = &pgpath->pg->ps;
1563 if (ps->type->end_io)
1564 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1571 * Suspend can't complete until all the I/O is processed so if
1572 * the last path fails we must error any remaining I/O.
1573 * Note that if the freeze_bdev fails while suspending, the
1574 * queue_if_no_path state is lost - userspace should reset it.
1576 static void multipath_presuspend(struct dm_target *ti)
1578 struct multipath *m = ti->private;
1580 queue_if_no_path(m, false, true);
1583 static void multipath_postsuspend(struct dm_target *ti)
1585 struct multipath *m = ti->private;
1587 mutex_lock(&m->work_mutex);
1588 flush_multipath_work(m);
1589 mutex_unlock(&m->work_mutex);
1593 * Restore the queue_if_no_path setting.
1595 static void multipath_resume(struct dm_target *ti)
1597 struct multipath *m = ti->private;
1598 unsigned long flags;
1600 spin_lock_irqsave(&m->lock, flags);
1601 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1602 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1603 spin_unlock_irqrestore(&m->lock, flags);
1607 * Info output has the following format:
1608 * num_multipath_feature_args [multipath_feature_args]*
1609 * num_handler_status_args [handler_status_args]*
1610 * num_groups init_group_number
1611 * [A|D|E num_ps_status_args [ps_status_args]*
1612 * num_paths num_selector_args
1613 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1615 * Table output has the following format (identical to the constructor string):
1616 * num_feature_args [features_args]*
1617 * num_handler_args hw_handler [hw_handler_args]*
1618 * num_groups init_group_number
1619 * [priority selector-name num_ps_args [ps_args]*
1620 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1622 static void multipath_status(struct dm_target *ti, status_type_t type,
1623 unsigned status_flags, char *result, unsigned maxlen)
1626 unsigned long flags;
1627 struct multipath *m = ti->private;
1628 struct priority_group *pg;
1633 spin_lock_irqsave(&m->lock, flags);
1636 if (type == STATUSTYPE_INFO)
1637 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1638 atomic_read(&m->pg_init_count));
1640 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1641 (m->pg_init_retries > 0) * 2 +
1642 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1643 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1644 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1646 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1647 DMEMIT("queue_if_no_path ");
1648 if (m->pg_init_retries)
1649 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1650 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1651 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1652 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1653 DMEMIT("retain_attached_hw_handler ");
1654 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1655 switch(m->queue_mode) {
1656 case DM_TYPE_BIO_BASED:
1657 DMEMIT("queue_mode bio ");
1659 case DM_TYPE_MQ_REQUEST_BASED:
1660 DMEMIT("queue_mode mq ");
1669 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1672 DMEMIT("1 %s ", m->hw_handler_name);
1674 DMEMIT("%u ", m->nr_priority_groups);
1677 pg_num = m->next_pg->pg_num;
1678 else if (m->current_pg)
1679 pg_num = m->current_pg->pg_num;
1681 pg_num = (m->nr_priority_groups ? 1 : 0);
1683 DMEMIT("%u ", pg_num);
1686 case STATUSTYPE_INFO:
1687 list_for_each_entry(pg, &m->priority_groups, list) {
1689 state = 'D'; /* Disabled */
1690 else if (pg == m->current_pg)
1691 state = 'A'; /* Currently Active */
1693 state = 'E'; /* Enabled */
1695 DMEMIT("%c ", state);
1697 if (pg->ps.type->status)
1698 sz += pg->ps.type->status(&pg->ps, NULL, type,
1704 DMEMIT("%u %u ", pg->nr_pgpaths,
1705 pg->ps.type->info_args);
1707 list_for_each_entry(p, &pg->pgpaths, list) {
1708 DMEMIT("%s %s %u ", p->path.dev->name,
1709 p->is_active ? "A" : "F",
1711 if (pg->ps.type->status)
1712 sz += pg->ps.type->status(&pg->ps,
1713 &p->path, type, result + sz,
1719 case STATUSTYPE_TABLE:
1720 list_for_each_entry(pg, &m->priority_groups, list) {
1721 DMEMIT("%s ", pg->ps.type->name);
1723 if (pg->ps.type->status)
1724 sz += pg->ps.type->status(&pg->ps, NULL, type,
1730 DMEMIT("%u %u ", pg->nr_pgpaths,
1731 pg->ps.type->table_args);
1733 list_for_each_entry(p, &pg->pgpaths, list) {
1734 DMEMIT("%s ", p->path.dev->name);
1735 if (pg->ps.type->status)
1736 sz += pg->ps.type->status(&pg->ps,
1737 &p->path, type, result + sz,
1744 spin_unlock_irqrestore(&m->lock, flags);
1747 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1751 struct multipath *m = ti->private;
1754 mutex_lock(&m->work_mutex);
1756 if (dm_suspended(ti)) {
1762 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1763 r = queue_if_no_path(m, true, false);
1765 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1766 r = queue_if_no_path(m, false, false);
1772 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1776 if (!strcasecmp(argv[0], "disable_group")) {
1777 r = bypass_pg_num(m, argv[1], true);
1779 } else if (!strcasecmp(argv[0], "enable_group")) {
1780 r = bypass_pg_num(m, argv[1], false);
1782 } else if (!strcasecmp(argv[0], "switch_group")) {
1783 r = switch_pg_num(m, argv[1]);
1785 } else if (!strcasecmp(argv[0], "reinstate_path"))
1786 action = reinstate_path;
1787 else if (!strcasecmp(argv[0], "fail_path"))
1790 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1794 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1796 DMWARN("message: error getting device %s",
1801 r = action_dev(m, dev, action);
1803 dm_put_device(ti, dev);
1806 mutex_unlock(&m->work_mutex);
1810 static int multipath_prepare_ioctl(struct dm_target *ti,
1811 struct block_device **bdev, fmode_t *mode)
1813 struct multipath *m = ti->private;
1814 struct pgpath *current_pgpath;
1817 current_pgpath = READ_ONCE(m->current_pgpath);
1818 if (!current_pgpath)
1819 current_pgpath = choose_pgpath(m, 0);
1821 if (current_pgpath) {
1822 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1823 *bdev = current_pgpath->path.dev->bdev;
1824 *mode = current_pgpath->path.dev->mode;
1827 /* pg_init has not started or completed */
1831 /* No path is available */
1832 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1838 if (r == -ENOTCONN) {
1839 if (!READ_ONCE(m->current_pg)) {
1840 /* Path status changed, redo selection */
1841 (void) choose_pgpath(m, 0);
1843 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1844 pg_init_all_paths(m);
1845 dm_table_run_md_queue_async(m->ti->table);
1846 process_queued_io_list(m);
1850 * Only pass ioctls through if the device sizes match exactly.
1852 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1857 static int multipath_iterate_devices(struct dm_target *ti,
1858 iterate_devices_callout_fn fn, void *data)
1860 struct multipath *m = ti->private;
1861 struct priority_group *pg;
1865 list_for_each_entry(pg, &m->priority_groups, list) {
1866 list_for_each_entry(p, &pg->pgpaths, list) {
1867 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1877 static int pgpath_busy(struct pgpath *pgpath)
1879 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1881 return blk_lld_busy(q);
1885 * We return "busy", only when we can map I/Os but underlying devices
1886 * are busy (so even if we map I/Os now, the I/Os will wait on
1887 * the underlying queue).
1888 * In other words, if we want to kill I/Os or queue them inside us
1889 * due to map unavailability, we don't return "busy". Otherwise,
1890 * dm core won't give us the I/Os and we can't do what we want.
1892 static int multipath_busy(struct dm_target *ti)
1894 bool busy = false, has_active = false;
1895 struct multipath *m = ti->private;
1896 struct priority_group *pg, *next_pg;
1897 struct pgpath *pgpath;
1899 /* pg_init in progress */
1900 if (atomic_read(&m->pg_init_in_progress))
1903 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1904 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1905 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1907 /* Guess which priority_group will be used at next mapping time */
1908 pg = READ_ONCE(m->current_pg);
1909 next_pg = READ_ONCE(m->next_pg);
1910 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1915 * We don't know which pg will be used at next mapping time.
1916 * We don't call choose_pgpath() here to avoid to trigger
1917 * pg_init just by busy checking.
1918 * So we don't know whether underlying devices we will be using
1919 * at next mapping time are busy or not. Just try mapping.
1925 * If there is one non-busy active path at least, the path selector
1926 * will be able to select it. So we consider such a pg as not busy.
1929 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1930 if (pgpath->is_active) {
1932 if (!pgpath_busy(pgpath)) {
1941 * No active path in this pg, so this pg won't be used and
1942 * the current_pg will be changed at next mapping time.
1943 * We need to try mapping to determine it.
1951 /*-----------------------------------------------------------------
1953 *---------------------------------------------------------------*/
1954 static struct target_type multipath_target = {
1955 .name = "multipath",
1956 .version = {1, 12, 0},
1957 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1958 .module = THIS_MODULE,
1959 .ctr = multipath_ctr,
1960 .dtr = multipath_dtr,
1961 .clone_and_map_rq = multipath_clone_and_map,
1962 .release_clone_rq = multipath_release_clone,
1963 .rq_end_io = multipath_end_io,
1964 .map = multipath_map_bio,
1965 .end_io = multipath_end_io_bio,
1966 .presuspend = multipath_presuspend,
1967 .postsuspend = multipath_postsuspend,
1968 .resume = multipath_resume,
1969 .status = multipath_status,
1970 .message = multipath_message,
1971 .prepare_ioctl = multipath_prepare_ioctl,
1972 .iterate_devices = multipath_iterate_devices,
1973 .busy = multipath_busy,
1976 static int __init dm_multipath_init(void)
1980 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1982 DMERR("failed to create workqueue kmpathd");
1984 goto bad_alloc_kmultipathd;
1988 * A separate workqueue is used to handle the device handlers
1989 * to avoid overloading existing workqueue. Overloading the
1990 * old workqueue would also create a bottleneck in the
1991 * path of the storage hardware device activation.
1993 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1995 if (!kmpath_handlerd) {
1996 DMERR("failed to create workqueue kmpath_handlerd");
1998 goto bad_alloc_kmpath_handlerd;
2001 r = dm_register_target(&multipath_target);
2003 DMERR("request-based register failed %d", r);
2005 goto bad_register_target;
2010 bad_register_target:
2011 destroy_workqueue(kmpath_handlerd);
2012 bad_alloc_kmpath_handlerd:
2013 destroy_workqueue(kmultipathd);
2014 bad_alloc_kmultipathd:
2018 static void __exit dm_multipath_exit(void)
2020 destroy_workqueue(kmpath_handlerd);
2021 destroy_workqueue(kmultipathd);
2023 dm_unregister_target(&multipath_target);
2026 module_init(dm_multipath_init);
2027 module_exit(dm_multipath_exit);
2029 MODULE_DESCRIPTION(DM_NAME " multipath target");
2030 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2031 MODULE_LICENSE("GPL");