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 * Map cloned requests (request-based multipath)
463 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
464 union map_info *map_context,
465 struct request **__clone)
467 struct multipath *m = ti->private;
468 size_t nr_bytes = blk_rq_bytes(rq);
469 struct pgpath *pgpath;
470 struct block_device *bdev;
471 struct dm_mpath_io *mpio = get_mpio(map_context);
472 struct request_queue *q;
473 struct request *clone;
475 /* Do we need to select a new pgpath? */
476 pgpath = READ_ONCE(m->current_pgpath);
477 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
478 pgpath = choose_pgpath(m, nr_bytes);
481 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
482 return DM_MAPIO_DELAY_REQUEUE;
483 dm_report_EIO(m); /* Failed */
484 return DM_MAPIO_KILL;
485 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
486 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
487 if (pg_init_all_paths(m))
488 return DM_MAPIO_DELAY_REQUEUE;
489 return DM_MAPIO_REQUEUE;
492 memset(mpio, 0, sizeof(*mpio));
493 mpio->pgpath = pgpath;
494 mpio->nr_bytes = nr_bytes;
496 bdev = pgpath->path.dev->bdev;
497 q = bdev_get_queue(bdev);
498 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
500 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501 bool queue_dying = blk_queue_dying(q);
502 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
503 PTR_ERR(clone), queue_dying ? " (path offline)" : "");
505 atomic_inc(&m->pg_init_in_progress);
506 activate_or_offline_path(pgpath);
508 return DM_MAPIO_DELAY_REQUEUE;
510 clone->bio = clone->biotail = NULL;
511 clone->rq_disk = bdev->bd_disk;
512 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
515 if (pgpath->pg->ps.type->start_io)
516 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
519 return DM_MAPIO_REMAPPED;
522 static void multipath_release_clone(struct request *clone)
524 blk_put_request(clone);
528 * Map cloned bios (bio-based multipath)
530 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
532 size_t nr_bytes = bio->bi_iter.bi_size;
533 struct pgpath *pgpath;
537 /* Do we need to select a new pgpath? */
538 pgpath = READ_ONCE(m->current_pgpath);
539 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
540 if (!pgpath || !queue_io)
541 pgpath = choose_pgpath(m, nr_bytes);
543 if ((pgpath && queue_io) ||
544 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
545 /* Queue for the daemon to resubmit */
546 spin_lock_irqsave(&m->lock, flags);
547 bio_list_add(&m->queued_bios, bio);
548 spin_unlock_irqrestore(&m->lock, flags);
549 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
550 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
551 pg_init_all_paths(m);
553 queue_work(kmultipathd, &m->process_queued_bios);
554 return DM_MAPIO_SUBMITTED;
558 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
559 return DM_MAPIO_REQUEUE;
561 return DM_MAPIO_KILL;
564 mpio->pgpath = pgpath;
565 mpio->nr_bytes = nr_bytes;
568 bio_set_dev(bio, pgpath->path.dev->bdev);
569 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
571 if (pgpath->pg->ps.type->start_io)
572 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
575 return DM_MAPIO_REMAPPED;
578 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
580 struct multipath *m = ti->private;
581 struct dm_mpath_io *mpio = NULL;
583 multipath_init_per_bio_data(bio, &mpio, NULL);
585 return __multipath_map_bio(m, bio, mpio);
588 static void process_queued_io_list(struct multipath *m)
590 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
591 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
592 else if (m->queue_mode == DM_TYPE_BIO_BASED)
593 queue_work(kmultipathd, &m->process_queued_bios);
596 static void process_queued_bios(struct work_struct *work)
601 struct bio_list bios;
602 struct blk_plug plug;
603 struct multipath *m =
604 container_of(work, struct multipath, process_queued_bios);
606 bio_list_init(&bios);
608 spin_lock_irqsave(&m->lock, flags);
610 if (bio_list_empty(&m->queued_bios)) {
611 spin_unlock_irqrestore(&m->lock, flags);
615 bio_list_merge(&bios, &m->queued_bios);
616 bio_list_init(&m->queued_bios);
618 spin_unlock_irqrestore(&m->lock, flags);
620 blk_start_plug(&plug);
621 while ((bio = bio_list_pop(&bios))) {
622 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
625 bio->bi_status = BLK_STS_IOERR;
628 case DM_MAPIO_REQUEUE:
629 bio->bi_status = BLK_STS_DM_REQUEUE;
632 case DM_MAPIO_REMAPPED:
633 generic_make_request(bio);
638 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
641 blk_finish_plug(&plug);
645 * If we run out of usable paths, should we queue I/O or error it?
647 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
652 spin_lock_irqsave(&m->lock, flags);
653 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
654 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
655 (!save_old_value && queue_if_no_path));
656 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
657 queue_if_no_path || dm_noflush_suspending(m->ti));
658 spin_unlock_irqrestore(&m->lock, flags);
660 if (!queue_if_no_path) {
661 dm_table_run_md_queue_async(m->ti->table);
662 process_queued_io_list(m);
669 * An event is triggered whenever a path is taken out of use.
670 * Includes path failure and PG bypass.
672 static void trigger_event(struct work_struct *work)
674 struct multipath *m =
675 container_of(work, struct multipath, trigger_event);
677 dm_table_event(m->ti->table);
680 /*-----------------------------------------------------------------
681 * Constructor/argument parsing:
682 * <#multipath feature args> [<arg>]*
683 * <#hw_handler args> [hw_handler [<arg>]*]
685 * <initial priority group>
686 * [<selector> <#selector args> [<arg>]*
687 * <#paths> <#per-path selector args>
688 * [<path> [<arg>]* ]+ ]+
689 *---------------------------------------------------------------*/
690 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
691 struct dm_target *ti)
694 struct path_selector_type *pst;
697 static const struct dm_arg _args[] = {
698 {0, 1024, "invalid number of path selector args"},
701 pst = dm_get_path_selector(dm_shift_arg(as));
703 ti->error = "unknown path selector type";
707 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
709 dm_put_path_selector(pst);
713 r = pst->create(&pg->ps, ps_argc, as->argv);
715 dm_put_path_selector(pst);
716 ti->error = "path selector constructor failed";
721 dm_consume_args(as, ps_argc);
726 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
727 struct dm_target *ti)
731 struct multipath *m = ti->private;
732 struct request_queue *q = NULL;
733 const char *attached_handler_name;
735 /* we need at least a path arg */
737 ti->error = "no device given";
738 return ERR_PTR(-EINVAL);
743 return ERR_PTR(-ENOMEM);
745 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
748 ti->error = "error getting device";
752 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
753 q = bdev_get_queue(p->path.dev->bdev);
755 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
757 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
758 if (attached_handler_name) {
760 * Clear any hw_handler_params associated with a
761 * handler that isn't already attached.
763 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
764 kfree(m->hw_handler_params);
765 m->hw_handler_params = NULL;
769 * Reset hw_handler_name to match the attached handler
771 * NB. This modifies the table line to show the actual
772 * handler instead of the original table passed in.
774 kfree(m->hw_handler_name);
775 m->hw_handler_name = attached_handler_name;
779 if (m->hw_handler_name) {
780 r = scsi_dh_attach(q, m->hw_handler_name);
782 char b[BDEVNAME_SIZE];
784 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
785 bdevname(p->path.dev->bdev, b));
789 ti->error = "error attaching hardware handler";
790 dm_put_device(ti, p->path.dev);
794 if (m->hw_handler_params) {
795 r = scsi_dh_set_params(q, m->hw_handler_params);
797 ti->error = "unable to set hardware "
798 "handler parameters";
799 dm_put_device(ti, p->path.dev);
805 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
807 dm_put_device(ti, p->path.dev);
818 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
821 static const struct dm_arg _args[] = {
822 {1, 1024, "invalid number of paths"},
823 {0, 1024, "invalid number of selector args"}
827 unsigned i, nr_selector_args, nr_args;
828 struct priority_group *pg;
829 struct dm_target *ti = m->ti;
833 ti->error = "not enough priority group arguments";
834 return ERR_PTR(-EINVAL);
837 pg = alloc_priority_group();
839 ti->error = "couldn't allocate priority group";
840 return ERR_PTR(-ENOMEM);
844 r = parse_path_selector(as, pg, ti);
851 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
855 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
859 nr_args = 1 + nr_selector_args;
860 for (i = 0; i < pg->nr_pgpaths; i++) {
861 struct pgpath *pgpath;
862 struct dm_arg_set path_args;
864 if (as->argc < nr_args) {
865 ti->error = "not enough path parameters";
870 path_args.argc = nr_args;
871 path_args.argv = as->argv;
873 pgpath = parse_path(&path_args, &pg->ps, ti);
874 if (IS_ERR(pgpath)) {
880 list_add_tail(&pgpath->list, &pg->pgpaths);
881 dm_consume_args(as, nr_args);
887 free_priority_group(pg, ti);
891 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
895 struct dm_target *ti = m->ti;
897 static const struct dm_arg _args[] = {
898 {0, 1024, "invalid number of hardware handler args"},
901 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
907 if (m->queue_mode == DM_TYPE_BIO_BASED) {
908 dm_consume_args(as, hw_argc);
909 DMERR("bio-based multipath doesn't allow hardware handler args");
913 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
914 if (!m->hw_handler_name)
921 for (i = 0; i <= hw_argc - 2; i++)
922 len += strlen(as->argv[i]) + 1;
923 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
925 ti->error = "memory allocation failed";
929 j = sprintf(p, "%d", hw_argc - 1);
930 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
931 j = sprintf(p, "%s", as->argv[i]);
933 dm_consume_args(as, hw_argc - 1);
937 kfree(m->hw_handler_name);
938 m->hw_handler_name = NULL;
942 static int parse_features(struct dm_arg_set *as, struct multipath *m)
946 struct dm_target *ti = m->ti;
947 const char *arg_name;
949 static const struct dm_arg _args[] = {
950 {0, 8, "invalid number of feature args"},
951 {1, 50, "pg_init_retries must be between 1 and 50"},
952 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
955 r = dm_read_arg_group(_args, as, &argc, &ti->error);
963 arg_name = dm_shift_arg(as);
966 if (!strcasecmp(arg_name, "queue_if_no_path")) {
967 r = queue_if_no_path(m, true, false);
971 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
972 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
976 if (!strcasecmp(arg_name, "pg_init_retries") &&
978 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
983 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
985 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
990 if (!strcasecmp(arg_name, "queue_mode") &&
992 const char *queue_mode_name = dm_shift_arg(as);
994 if (!strcasecmp(queue_mode_name, "bio"))
995 m->queue_mode = DM_TYPE_BIO_BASED;
996 else if (!strcasecmp(queue_mode_name, "rq"))
997 m->queue_mode = DM_TYPE_REQUEST_BASED;
998 else if (!strcasecmp(queue_mode_name, "mq"))
999 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1001 ti->error = "Unknown 'queue_mode' requested";
1008 ti->error = "Unrecognised multipath feature request";
1010 } while (argc && !r);
1015 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1017 /* target arguments */
1018 static const struct dm_arg _args[] = {
1019 {0, 1024, "invalid number of priority groups"},
1020 {0, 1024, "invalid initial priority group number"},
1024 struct multipath *m;
1025 struct dm_arg_set as;
1026 unsigned pg_count = 0;
1027 unsigned next_pg_num;
1032 m = alloc_multipath(ti);
1034 ti->error = "can't allocate multipath";
1038 r = parse_features(&as, m);
1042 r = alloc_multipath_stage2(ti, m);
1046 r = parse_hw_handler(&as, m);
1050 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1054 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1058 if ((!m->nr_priority_groups && next_pg_num) ||
1059 (m->nr_priority_groups && !next_pg_num)) {
1060 ti->error = "invalid initial priority group";
1065 /* parse the priority groups */
1067 struct priority_group *pg;
1068 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1070 pg = parse_priority_group(&as, m);
1076 nr_valid_paths += pg->nr_pgpaths;
1077 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1079 list_add_tail(&pg->list, &m->priority_groups);
1081 pg->pg_num = pg_count;
1086 if (pg_count != m->nr_priority_groups) {
1087 ti->error = "priority group count mismatch";
1092 ti->num_flush_bios = 1;
1093 ti->num_discard_bios = 1;
1094 ti->num_write_same_bios = 1;
1095 ti->num_write_zeroes_bios = 1;
1096 if (m->queue_mode == DM_TYPE_BIO_BASED)
1097 ti->per_io_data_size = multipath_per_bio_data_size();
1099 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1108 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1113 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1115 if (!atomic_read(&m->pg_init_in_progress))
1120 finish_wait(&m->pg_init_wait, &wait);
1123 static void flush_multipath_work(struct multipath *m)
1125 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1126 smp_mb__after_atomic();
1128 flush_workqueue(kmpath_handlerd);
1129 multipath_wait_for_pg_init_completion(m);
1130 flush_workqueue(kmultipathd);
1131 flush_work(&m->trigger_event);
1133 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1134 smp_mb__after_atomic();
1137 static void multipath_dtr(struct dm_target *ti)
1139 struct multipath *m = ti->private;
1141 flush_multipath_work(m);
1146 * Take a path out of use.
1148 static int fail_path(struct pgpath *pgpath)
1150 unsigned long flags;
1151 struct multipath *m = pgpath->pg->m;
1153 spin_lock_irqsave(&m->lock, flags);
1155 if (!pgpath->is_active)
1158 DMWARN("Failing path %s.", pgpath->path.dev->name);
1160 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1161 pgpath->is_active = false;
1162 pgpath->fail_count++;
1164 atomic_dec(&m->nr_valid_paths);
1166 if (pgpath == m->current_pgpath)
1167 m->current_pgpath = NULL;
1169 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1170 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1172 schedule_work(&m->trigger_event);
1175 spin_unlock_irqrestore(&m->lock, flags);
1181 * Reinstate a previously-failed path
1183 static int reinstate_path(struct pgpath *pgpath)
1185 int r = 0, run_queue = 0;
1186 unsigned long flags;
1187 struct multipath *m = pgpath->pg->m;
1188 unsigned nr_valid_paths;
1190 spin_lock_irqsave(&m->lock, flags);
1192 if (pgpath->is_active)
1195 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1197 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1201 pgpath->is_active = true;
1203 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1204 if (nr_valid_paths == 1) {
1205 m->current_pgpath = NULL;
1207 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1208 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1209 atomic_inc(&m->pg_init_in_progress);
1212 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1213 pgpath->path.dev->name, nr_valid_paths);
1215 schedule_work(&m->trigger_event);
1218 spin_unlock_irqrestore(&m->lock, flags);
1220 dm_table_run_md_queue_async(m->ti->table);
1221 process_queued_io_list(m);
1228 * Fail or reinstate all paths that match the provided struct dm_dev.
1230 static int action_dev(struct multipath *m, struct dm_dev *dev,
1234 struct pgpath *pgpath;
1235 struct priority_group *pg;
1237 list_for_each_entry(pg, &m->priority_groups, list) {
1238 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1239 if (pgpath->path.dev == dev)
1248 * Temporarily try to avoid having to use the specified PG
1250 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1253 unsigned long flags;
1255 spin_lock_irqsave(&m->lock, flags);
1257 pg->bypassed = bypassed;
1258 m->current_pgpath = NULL;
1259 m->current_pg = NULL;
1261 spin_unlock_irqrestore(&m->lock, flags);
1263 schedule_work(&m->trigger_event);
1267 * Switch to using the specified PG from the next I/O that gets mapped
1269 static int switch_pg_num(struct multipath *m, const char *pgstr)
1271 struct priority_group *pg;
1273 unsigned long flags;
1276 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1277 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1278 DMWARN("invalid PG number supplied to switch_pg_num");
1282 spin_lock_irqsave(&m->lock, flags);
1283 list_for_each_entry(pg, &m->priority_groups, list) {
1284 pg->bypassed = false;
1288 m->current_pgpath = NULL;
1289 m->current_pg = NULL;
1292 spin_unlock_irqrestore(&m->lock, flags);
1294 schedule_work(&m->trigger_event);
1299 * Set/clear bypassed status of a PG.
1300 * PGs are numbered upwards from 1 in the order they were declared.
1302 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1304 struct priority_group *pg;
1308 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1309 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1310 DMWARN("invalid PG number supplied to bypass_pg");
1314 list_for_each_entry(pg, &m->priority_groups, list) {
1319 bypass_pg(m, pg, bypassed);
1324 * Should we retry pg_init immediately?
1326 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1328 unsigned long flags;
1329 bool limit_reached = false;
1331 spin_lock_irqsave(&m->lock, flags);
1333 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1334 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1335 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1337 limit_reached = true;
1339 spin_unlock_irqrestore(&m->lock, flags);
1341 return limit_reached;
1344 static void pg_init_done(void *data, int errors)
1346 struct pgpath *pgpath = data;
1347 struct priority_group *pg = pgpath->pg;
1348 struct multipath *m = pg->m;
1349 unsigned long flags;
1350 bool delay_retry = false;
1352 /* device or driver problems */
1357 if (!m->hw_handler_name) {
1361 DMERR("Could not failover the device: Handler scsi_dh_%s "
1362 "Error %d.", m->hw_handler_name, errors);
1364 * Fail path for now, so we do not ping pong
1368 case SCSI_DH_DEV_TEMP_BUSY:
1370 * Probably doing something like FW upgrade on the
1371 * controller so try the other pg.
1373 bypass_pg(m, pg, true);
1376 /* Wait before retrying. */
1379 case SCSI_DH_IMM_RETRY:
1380 case SCSI_DH_RES_TEMP_UNAVAIL:
1381 if (pg_init_limit_reached(m, pgpath))
1385 case SCSI_DH_DEV_OFFLINED:
1388 * We probably do not want to fail the path for a device
1389 * error, but this is what the old dm did. In future
1390 * patches we can do more advanced handling.
1395 spin_lock_irqsave(&m->lock, flags);
1397 if (pgpath == m->current_pgpath) {
1398 DMERR("Could not failover device. Error %d.", errors);
1399 m->current_pgpath = NULL;
1400 m->current_pg = NULL;
1402 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1403 pg->bypassed = false;
1405 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1406 /* Activations of other paths are still on going */
1409 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1411 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1413 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1415 if (__pg_init_all_paths(m))
1418 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1420 process_queued_io_list(m);
1423 * Wake up any thread waiting to suspend.
1425 wake_up(&m->pg_init_wait);
1428 spin_unlock_irqrestore(&m->lock, flags);
1431 static void activate_or_offline_path(struct pgpath *pgpath)
1433 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1435 if (pgpath->is_active && !blk_queue_dying(q))
1436 scsi_dh_activate(q, pg_init_done, pgpath);
1438 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1441 static void activate_path_work(struct work_struct *work)
1443 struct pgpath *pgpath =
1444 container_of(work, struct pgpath, activate_path.work);
1446 activate_or_offline_path(pgpath);
1449 static int noretry_error(blk_status_t error)
1452 case BLK_STS_NOTSUPP:
1454 case BLK_STS_TARGET:
1456 case BLK_STS_MEDIUM:
1460 /* Anything else could be a path failure, so should be retried */
1464 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1465 blk_status_t error, union map_info *map_context)
1467 struct dm_mpath_io *mpio = get_mpio(map_context);
1468 struct pgpath *pgpath = mpio->pgpath;
1469 int r = DM_ENDIO_DONE;
1472 * We don't queue any clone request inside the multipath target
1473 * during end I/O handling, since those clone requests don't have
1474 * bio clones. If we queue them inside the multipath target,
1475 * we need to make bio clones, that requires memory allocation.
1476 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1477 * don't have bio clones.)
1478 * Instead of queueing the clone request here, we queue the original
1479 * request into dm core, which will remake a clone request and
1480 * clone bios for it and resubmit it later.
1482 if (error && !noretry_error(error)) {
1483 struct multipath *m = ti->private;
1485 r = DM_ENDIO_REQUEUE;
1490 if (atomic_read(&m->nr_valid_paths) == 0 &&
1491 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1492 if (error == BLK_STS_IOERR)
1494 /* complete with the original error */
1500 struct path_selector *ps = &pgpath->pg->ps;
1502 if (ps->type->end_io)
1503 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1509 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1510 blk_status_t *error)
1512 struct multipath *m = ti->private;
1513 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1514 struct pgpath *pgpath = mpio->pgpath;
1515 unsigned long flags;
1516 int r = DM_ENDIO_DONE;
1518 if (!*error || noretry_error(*error))
1524 if (atomic_read(&m->nr_valid_paths) == 0 &&
1525 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1527 *error = BLK_STS_IOERR;
1531 /* Queue for the daemon to resubmit */
1532 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1534 spin_lock_irqsave(&m->lock, flags);
1535 bio_list_add(&m->queued_bios, clone);
1536 spin_unlock_irqrestore(&m->lock, flags);
1537 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1538 queue_work(kmultipathd, &m->process_queued_bios);
1540 r = DM_ENDIO_INCOMPLETE;
1543 struct path_selector *ps = &pgpath->pg->ps;
1545 if (ps->type->end_io)
1546 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1553 * Suspend can't complete until all the I/O is processed so if
1554 * the last path fails we must error any remaining I/O.
1555 * Note that if the freeze_bdev fails while suspending, the
1556 * queue_if_no_path state is lost - userspace should reset it.
1558 static void multipath_presuspend(struct dm_target *ti)
1560 struct multipath *m = ti->private;
1562 queue_if_no_path(m, false, true);
1565 static void multipath_postsuspend(struct dm_target *ti)
1567 struct multipath *m = ti->private;
1569 mutex_lock(&m->work_mutex);
1570 flush_multipath_work(m);
1571 mutex_unlock(&m->work_mutex);
1575 * Restore the queue_if_no_path setting.
1577 static void multipath_resume(struct dm_target *ti)
1579 struct multipath *m = ti->private;
1580 unsigned long flags;
1582 spin_lock_irqsave(&m->lock, flags);
1583 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1584 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1585 spin_unlock_irqrestore(&m->lock, flags);
1589 * Info output has the following format:
1590 * num_multipath_feature_args [multipath_feature_args]*
1591 * num_handler_status_args [handler_status_args]*
1592 * num_groups init_group_number
1593 * [A|D|E num_ps_status_args [ps_status_args]*
1594 * num_paths num_selector_args
1595 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1597 * Table output has the following format (identical to the constructor string):
1598 * num_feature_args [features_args]*
1599 * num_handler_args hw_handler [hw_handler_args]*
1600 * num_groups init_group_number
1601 * [priority selector-name num_ps_args [ps_args]*
1602 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1604 static void multipath_status(struct dm_target *ti, status_type_t type,
1605 unsigned status_flags, char *result, unsigned maxlen)
1608 unsigned long flags;
1609 struct multipath *m = ti->private;
1610 struct priority_group *pg;
1615 spin_lock_irqsave(&m->lock, flags);
1618 if (type == STATUSTYPE_INFO)
1619 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1620 atomic_read(&m->pg_init_count));
1622 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1623 (m->pg_init_retries > 0) * 2 +
1624 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1625 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1626 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1628 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1629 DMEMIT("queue_if_no_path ");
1630 if (m->pg_init_retries)
1631 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1632 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1633 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1634 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1635 DMEMIT("retain_attached_hw_handler ");
1636 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1637 switch(m->queue_mode) {
1638 case DM_TYPE_BIO_BASED:
1639 DMEMIT("queue_mode bio ");
1641 case DM_TYPE_MQ_REQUEST_BASED:
1642 DMEMIT("queue_mode mq ");
1651 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1654 DMEMIT("1 %s ", m->hw_handler_name);
1656 DMEMIT("%u ", m->nr_priority_groups);
1659 pg_num = m->next_pg->pg_num;
1660 else if (m->current_pg)
1661 pg_num = m->current_pg->pg_num;
1663 pg_num = (m->nr_priority_groups ? 1 : 0);
1665 DMEMIT("%u ", pg_num);
1668 case STATUSTYPE_INFO:
1669 list_for_each_entry(pg, &m->priority_groups, list) {
1671 state = 'D'; /* Disabled */
1672 else if (pg == m->current_pg)
1673 state = 'A'; /* Currently Active */
1675 state = 'E'; /* Enabled */
1677 DMEMIT("%c ", state);
1679 if (pg->ps.type->status)
1680 sz += pg->ps.type->status(&pg->ps, NULL, type,
1686 DMEMIT("%u %u ", pg->nr_pgpaths,
1687 pg->ps.type->info_args);
1689 list_for_each_entry(p, &pg->pgpaths, list) {
1690 DMEMIT("%s %s %u ", p->path.dev->name,
1691 p->is_active ? "A" : "F",
1693 if (pg->ps.type->status)
1694 sz += pg->ps.type->status(&pg->ps,
1695 &p->path, type, result + sz,
1701 case STATUSTYPE_TABLE:
1702 list_for_each_entry(pg, &m->priority_groups, list) {
1703 DMEMIT("%s ", pg->ps.type->name);
1705 if (pg->ps.type->status)
1706 sz += pg->ps.type->status(&pg->ps, NULL, type,
1712 DMEMIT("%u %u ", pg->nr_pgpaths,
1713 pg->ps.type->table_args);
1715 list_for_each_entry(p, &pg->pgpaths, list) {
1716 DMEMIT("%s ", p->path.dev->name);
1717 if (pg->ps.type->status)
1718 sz += pg->ps.type->status(&pg->ps,
1719 &p->path, type, result + sz,
1726 spin_unlock_irqrestore(&m->lock, flags);
1729 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1733 struct multipath *m = ti->private;
1736 mutex_lock(&m->work_mutex);
1738 if (dm_suspended(ti)) {
1744 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1745 r = queue_if_no_path(m, true, false);
1747 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1748 r = queue_if_no_path(m, false, false);
1754 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1758 if (!strcasecmp(argv[0], "disable_group")) {
1759 r = bypass_pg_num(m, argv[1], true);
1761 } else if (!strcasecmp(argv[0], "enable_group")) {
1762 r = bypass_pg_num(m, argv[1], false);
1764 } else if (!strcasecmp(argv[0], "switch_group")) {
1765 r = switch_pg_num(m, argv[1]);
1767 } else if (!strcasecmp(argv[0], "reinstate_path"))
1768 action = reinstate_path;
1769 else if (!strcasecmp(argv[0], "fail_path"))
1772 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1776 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1778 DMWARN("message: error getting device %s",
1783 r = action_dev(m, dev, action);
1785 dm_put_device(ti, dev);
1788 mutex_unlock(&m->work_mutex);
1792 static int multipath_prepare_ioctl(struct dm_target *ti,
1793 struct block_device **bdev, fmode_t *mode)
1795 struct multipath *m = ti->private;
1796 struct pgpath *current_pgpath;
1799 current_pgpath = READ_ONCE(m->current_pgpath);
1800 if (!current_pgpath)
1801 current_pgpath = choose_pgpath(m, 0);
1803 if (current_pgpath) {
1804 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1805 *bdev = current_pgpath->path.dev->bdev;
1806 *mode = current_pgpath->path.dev->mode;
1809 /* pg_init has not started or completed */
1813 /* No path is available */
1814 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1820 if (r == -ENOTCONN) {
1821 if (!READ_ONCE(m->current_pg)) {
1822 /* Path status changed, redo selection */
1823 (void) choose_pgpath(m, 0);
1825 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1826 pg_init_all_paths(m);
1827 dm_table_run_md_queue_async(m->ti->table);
1828 process_queued_io_list(m);
1832 * Only pass ioctls through if the device sizes match exactly.
1834 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1839 static int multipath_iterate_devices(struct dm_target *ti,
1840 iterate_devices_callout_fn fn, void *data)
1842 struct multipath *m = ti->private;
1843 struct priority_group *pg;
1847 list_for_each_entry(pg, &m->priority_groups, list) {
1848 list_for_each_entry(p, &pg->pgpaths, list) {
1849 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1859 static int pgpath_busy(struct pgpath *pgpath)
1861 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1863 return blk_lld_busy(q);
1867 * We return "busy", only when we can map I/Os but underlying devices
1868 * are busy (so even if we map I/Os now, the I/Os will wait on
1869 * the underlying queue).
1870 * In other words, if we want to kill I/Os or queue them inside us
1871 * due to map unavailability, we don't return "busy". Otherwise,
1872 * dm core won't give us the I/Os and we can't do what we want.
1874 static int multipath_busy(struct dm_target *ti)
1876 bool busy = false, has_active = false;
1877 struct multipath *m = ti->private;
1878 struct priority_group *pg, *next_pg;
1879 struct pgpath *pgpath;
1881 /* pg_init in progress */
1882 if (atomic_read(&m->pg_init_in_progress))
1885 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1886 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1887 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1889 /* Guess which priority_group will be used at next mapping time */
1890 pg = READ_ONCE(m->current_pg);
1891 next_pg = READ_ONCE(m->next_pg);
1892 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1897 * We don't know which pg will be used at next mapping time.
1898 * We don't call choose_pgpath() here to avoid to trigger
1899 * pg_init just by busy checking.
1900 * So we don't know whether underlying devices we will be using
1901 * at next mapping time are busy or not. Just try mapping.
1907 * If there is one non-busy active path at least, the path selector
1908 * will be able to select it. So we consider such a pg as not busy.
1911 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1912 if (pgpath->is_active) {
1914 if (!pgpath_busy(pgpath)) {
1923 * No active path in this pg, so this pg won't be used and
1924 * the current_pg will be changed at next mapping time.
1925 * We need to try mapping to determine it.
1933 /*-----------------------------------------------------------------
1935 *---------------------------------------------------------------*/
1936 static struct target_type multipath_target = {
1937 .name = "multipath",
1938 .version = {1, 12, 0},
1939 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1940 .module = THIS_MODULE,
1941 .ctr = multipath_ctr,
1942 .dtr = multipath_dtr,
1943 .clone_and_map_rq = multipath_clone_and_map,
1944 .release_clone_rq = multipath_release_clone,
1945 .rq_end_io = multipath_end_io,
1946 .map = multipath_map_bio,
1947 .end_io = multipath_end_io_bio,
1948 .presuspend = multipath_presuspend,
1949 .postsuspend = multipath_postsuspend,
1950 .resume = multipath_resume,
1951 .status = multipath_status,
1952 .message = multipath_message,
1953 .prepare_ioctl = multipath_prepare_ioctl,
1954 .iterate_devices = multipath_iterate_devices,
1955 .busy = multipath_busy,
1958 static int __init dm_multipath_init(void)
1962 r = dm_register_target(&multipath_target);
1964 DMERR("request-based register failed %d", r);
1966 goto bad_register_target;
1969 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1971 DMERR("failed to create workqueue kmpathd");
1973 goto bad_alloc_kmultipathd;
1977 * A separate workqueue is used to handle the device handlers
1978 * to avoid overloading existing workqueue. Overloading the
1979 * old workqueue would also create a bottleneck in the
1980 * path of the storage hardware device activation.
1982 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1984 if (!kmpath_handlerd) {
1985 DMERR("failed to create workqueue kmpath_handlerd");
1987 goto bad_alloc_kmpath_handlerd;
1992 bad_alloc_kmpath_handlerd:
1993 destroy_workqueue(kmultipathd);
1994 bad_alloc_kmultipathd:
1995 dm_unregister_target(&multipath_target);
1996 bad_register_target:
2000 static void __exit dm_multipath_exit(void)
2002 destroy_workqueue(kmpath_handlerd);
2003 destroy_workqueue(kmultipathd);
2005 dm_unregister_target(&multipath_target);
2008 module_init(dm_multipath_init);
2009 module_exit(dm_multipath_exit);
2011 MODULE_DESCRIPTION(DM_NAME " multipath target");
2012 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2013 MODULE_LICENSE("GPL");