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/timer.h>
24 #include <linux/workqueue.h>
25 #include <linux/delay.h>
26 #include <scsi/scsi_dh.h>
27 #include <linux/atomic.h>
28 #include <linux/blk-mq.h>
30 #define DM_MSG_PREFIX "multipath"
31 #define DM_PG_INIT_DELAY_MSECS 2000
32 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
33 #define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
35 static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
39 struct list_head list;
41 struct priority_group *pg; /* Owning PG */
42 unsigned fail_count; /* Cumulative failure count */
45 struct delayed_work activate_path;
47 bool is_active:1; /* Path status */
50 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
53 * Paths are grouped into Priority Groups and numbered from 1 upwards.
54 * Each has a path selector which controls which path gets used.
56 struct priority_group {
57 struct list_head list;
59 struct multipath *m; /* Owning multipath instance */
60 struct path_selector ps;
62 unsigned pg_num; /* Reference number */
63 unsigned nr_pgpaths; /* Number of paths in PG */
64 struct list_head pgpaths;
66 bool bypassed:1; /* Temporarily bypass this PG? */
69 /* Multipath context */
71 unsigned long flags; /* Multipath state flags */
74 enum dm_queue_mode queue_mode;
76 struct pgpath *current_pgpath;
77 struct priority_group *current_pg;
78 struct priority_group *next_pg; /* Switch to this PG if set */
80 atomic_t nr_valid_paths; /* Total number of usable paths */
81 unsigned nr_priority_groups;
82 struct list_head priority_groups;
84 const char *hw_handler_name;
85 char *hw_handler_params;
86 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
87 unsigned pg_init_retries; /* Number of times to retry pg_init */
88 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
89 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
90 atomic_t pg_init_count; /* Number of times pg_init called */
92 struct mutex work_mutex;
93 struct work_struct trigger_event;
96 struct work_struct process_queued_bios;
97 struct bio_list queued_bios;
99 struct timer_list nopath_timer; /* Timeout for queue_if_no_path */
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);
117 static void queue_if_no_path_timeout_work(struct timer_list *t);
119 /*-----------------------------------------------
120 * Multipath state flags.
121 *-----------------------------------------------*/
123 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
124 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
125 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
126 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
127 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
128 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
129 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
131 static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
133 bool r = test_bit(MPATHF_bit, &m->flags);
137 spin_lock_irqsave(&m->lock, flags);
138 r = test_bit(MPATHF_bit, &m->flags);
139 spin_unlock_irqrestore(&m->lock, flags);
145 /*-----------------------------------------------
146 * Allocation routines
147 *-----------------------------------------------*/
149 static struct pgpath *alloc_pgpath(void)
151 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
156 pgpath->is_active = true;
161 static void free_pgpath(struct pgpath *pgpath)
166 static struct priority_group *alloc_priority_group(void)
168 struct priority_group *pg;
170 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
173 INIT_LIST_HEAD(&pg->pgpaths);
178 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
180 struct pgpath *pgpath, *tmp;
182 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
183 list_del(&pgpath->list);
184 dm_put_device(ti, pgpath->path.dev);
189 static void free_priority_group(struct priority_group *pg,
190 struct dm_target *ti)
192 struct path_selector *ps = &pg->ps;
195 ps->type->destroy(ps);
196 dm_put_path_selector(ps->type);
199 free_pgpaths(&pg->pgpaths, ti);
203 static struct multipath *alloc_multipath(struct dm_target *ti)
207 m = kzalloc(sizeof(*m), GFP_KERNEL);
209 INIT_LIST_HEAD(&m->priority_groups);
210 spin_lock_init(&m->lock);
211 atomic_set(&m->nr_valid_paths, 0);
212 INIT_WORK(&m->trigger_event, trigger_event);
213 mutex_init(&m->work_mutex);
215 m->queue_mode = DM_TYPE_NONE;
220 timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
226 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
228 if (m->queue_mode == DM_TYPE_NONE) {
229 m->queue_mode = DM_TYPE_REQUEST_BASED;
230 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
231 INIT_WORK(&m->process_queued_bios, process_queued_bios);
233 * bio-based doesn't support any direct scsi_dh management;
234 * it just discovers if a scsi_dh is attached.
236 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
239 dm_table_set_type(ti->table, m->queue_mode);
242 * Init fields that are only used when a scsi_dh is attached
243 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
245 set_bit(MPATHF_QUEUE_IO, &m->flags);
246 atomic_set(&m->pg_init_in_progress, 0);
247 atomic_set(&m->pg_init_count, 0);
248 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
249 init_waitqueue_head(&m->pg_init_wait);
254 static void free_multipath(struct multipath *m)
256 struct priority_group *pg, *tmp;
258 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
260 free_priority_group(pg, m->ti);
263 kfree(m->hw_handler_name);
264 kfree(m->hw_handler_params);
265 mutex_destroy(&m->work_mutex);
269 static struct dm_mpath_io *get_mpio(union map_info *info)
274 static size_t multipath_per_bio_data_size(void)
276 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
279 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
281 return dm_per_bio_data(bio, multipath_per_bio_data_size());
284 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
286 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
287 void *bio_details = mpio + 1;
291 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
293 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
294 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
296 mpio->nr_bytes = bio->bi_iter.bi_size;
300 dm_bio_record(bio_details, bio);
303 /*-----------------------------------------------
305 *-----------------------------------------------*/
307 static int __pg_init_all_paths(struct multipath *m)
309 struct pgpath *pgpath;
310 unsigned long pg_init_delay = 0;
312 lockdep_assert_held(&m->lock);
314 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
317 atomic_inc(&m->pg_init_count);
318 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
320 /* Check here to reset pg_init_required */
324 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
325 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
326 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
327 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
328 /* Skip failed paths */
329 if (!pgpath->is_active)
331 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
333 atomic_inc(&m->pg_init_in_progress);
335 return atomic_read(&m->pg_init_in_progress);
338 static int pg_init_all_paths(struct multipath *m)
343 spin_lock_irqsave(&m->lock, flags);
344 ret = __pg_init_all_paths(m);
345 spin_unlock_irqrestore(&m->lock, flags);
350 static void __switch_pg(struct multipath *m, struct priority_group *pg)
352 lockdep_assert_held(&m->lock);
356 /* Must we initialise the PG first, and queue I/O till it's ready? */
357 if (m->hw_handler_name) {
358 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
359 set_bit(MPATHF_QUEUE_IO, &m->flags);
361 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
362 clear_bit(MPATHF_QUEUE_IO, &m->flags);
365 atomic_set(&m->pg_init_count, 0);
368 static struct pgpath *choose_path_in_pg(struct multipath *m,
369 struct priority_group *pg,
373 struct dm_path *path;
374 struct pgpath *pgpath;
376 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
378 return ERR_PTR(-ENXIO);
380 pgpath = path_to_pgpath(path);
382 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
383 /* Only update current_pgpath if pg changed */
384 spin_lock_irqsave(&m->lock, flags);
385 m->current_pgpath = pgpath;
387 spin_unlock_irqrestore(&m->lock, flags);
393 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
396 struct priority_group *pg;
397 struct pgpath *pgpath;
398 unsigned bypassed = 1;
400 if (!atomic_read(&m->nr_valid_paths)) {
401 spin_lock_irqsave(&m->lock, flags);
402 clear_bit(MPATHF_QUEUE_IO, &m->flags);
403 spin_unlock_irqrestore(&m->lock, flags);
407 /* Were we instructed to switch PG? */
408 if (READ_ONCE(m->next_pg)) {
409 spin_lock_irqsave(&m->lock, flags);
412 spin_unlock_irqrestore(&m->lock, flags);
413 goto check_current_pg;
416 spin_unlock_irqrestore(&m->lock, flags);
417 pgpath = choose_path_in_pg(m, pg, nr_bytes);
418 if (!IS_ERR_OR_NULL(pgpath))
422 /* Don't change PG until it has no remaining paths */
424 pg = READ_ONCE(m->current_pg);
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath))
432 * Loop through priority groups until we find a valid path.
433 * First time we skip PGs marked 'bypassed'.
434 * Second time we only try the ones we skipped, but set
435 * pg_init_delay_retry so we do not hammer controllers.
438 list_for_each_entry(pg, &m->priority_groups, list) {
439 if (pg->bypassed == !!bypassed)
441 pgpath = choose_path_in_pg(m, pg, nr_bytes);
442 if (!IS_ERR_OR_NULL(pgpath)) {
444 spin_lock_irqsave(&m->lock, flags);
445 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
446 spin_unlock_irqrestore(&m->lock, flags);
451 } while (bypassed--);
454 spin_lock_irqsave(&m->lock, flags);
455 m->current_pgpath = NULL;
456 m->current_pg = NULL;
457 spin_unlock_irqrestore(&m->lock, flags);
463 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
464 * report the function name and line number of the function from which
465 * it has been invoked.
467 #define dm_report_EIO(m) \
469 DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
470 dm_table_device_name((m)->ti->table), \
471 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
472 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
473 dm_noflush_suspending((m)->ti)); \
477 * Check whether bios must be queued in the device-mapper core rather
478 * than here in the target.
480 static bool __must_push_back(struct multipath *m)
482 return dm_noflush_suspending(m->ti);
485 static bool must_push_back_rq(struct multipath *m)
490 spin_lock_irqsave(&m->lock, flags);
491 ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
492 spin_unlock_irqrestore(&m->lock, flags);
498 * Map cloned requests (request-based multipath)
500 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
501 union map_info *map_context,
502 struct request **__clone)
504 struct multipath *m = ti->private;
505 size_t nr_bytes = blk_rq_bytes(rq);
506 struct pgpath *pgpath;
507 struct block_device *bdev;
508 struct dm_mpath_io *mpio = get_mpio(map_context);
509 struct request_queue *q;
510 struct request *clone;
512 /* Do we need to select a new pgpath? */
513 pgpath = READ_ONCE(m->current_pgpath);
514 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
515 pgpath = choose_pgpath(m, nr_bytes);
518 if (must_push_back_rq(m))
519 return DM_MAPIO_DELAY_REQUEUE;
520 dm_report_EIO(m); /* Failed */
521 return DM_MAPIO_KILL;
522 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
523 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
524 pg_init_all_paths(m);
525 return DM_MAPIO_DELAY_REQUEUE;
528 mpio->pgpath = pgpath;
529 mpio->nr_bytes = nr_bytes;
531 bdev = pgpath->path.dev->bdev;
532 q = bdev_get_queue(bdev);
533 clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
536 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
537 if (blk_queue_dying(q)) {
538 atomic_inc(&m->pg_init_in_progress);
539 activate_or_offline_path(pgpath);
540 return DM_MAPIO_DELAY_REQUEUE;
544 * blk-mq's SCHED_RESTART can cover this requeue, so we
545 * needn't deal with it by DELAY_REQUEUE. More importantly,
546 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
547 * get the queue busy feedback (via BLK_STS_RESOURCE),
548 * otherwise I/O merging can suffer.
550 return DM_MAPIO_REQUEUE;
552 clone->bio = clone->biotail = NULL;
553 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
556 if (pgpath->pg->ps.type->start_io)
557 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
560 return DM_MAPIO_REMAPPED;
563 static void multipath_release_clone(struct request *clone,
564 union map_info *map_context)
566 if (unlikely(map_context)) {
568 * non-NULL map_context means caller is still map
569 * method; must undo multipath_clone_and_map()
571 struct dm_mpath_io *mpio = get_mpio(map_context);
572 struct pgpath *pgpath = mpio->pgpath;
574 if (pgpath && pgpath->pg->ps.type->end_io)
575 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
578 clone->io_start_time_ns);
581 blk_mq_free_request(clone);
585 * Map cloned bios (bio-based multipath)
588 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
590 /* Queue for the daemon to resubmit */
591 bio_list_add(&m->queued_bios, bio);
592 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
593 queue_work(kmultipathd, &m->process_queued_bios);
596 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
600 spin_lock_irqsave(&m->lock, flags);
601 __multipath_queue_bio(m, bio);
602 spin_unlock_irqrestore(&m->lock, flags);
605 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
607 struct pgpath *pgpath;
610 /* Do we need to select a new pgpath? */
611 pgpath = READ_ONCE(m->current_pgpath);
612 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
613 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
616 spin_lock_irqsave(&m->lock, flags);
617 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
618 __multipath_queue_bio(m, bio);
619 pgpath = ERR_PTR(-EAGAIN);
621 spin_unlock_irqrestore(&m->lock, flags);
623 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
624 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
625 multipath_queue_bio(m, bio);
626 pg_init_all_paths(m);
627 return ERR_PTR(-EAGAIN);
633 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
634 struct dm_mpath_io *mpio)
636 struct pgpath *pgpath = __map_bio(m, bio);
639 return DM_MAPIO_SUBMITTED;
642 if (__must_push_back(m))
643 return DM_MAPIO_REQUEUE;
645 return DM_MAPIO_KILL;
648 mpio->pgpath = pgpath;
651 bio_set_dev(bio, pgpath->path.dev->bdev);
652 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
654 if (pgpath->pg->ps.type->start_io)
655 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
658 return DM_MAPIO_REMAPPED;
661 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
663 struct multipath *m = ti->private;
664 struct dm_mpath_io *mpio = NULL;
666 multipath_init_per_bio_data(bio, &mpio);
667 return __multipath_map_bio(m, bio, mpio);
670 static void process_queued_io_list(struct multipath *m)
672 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
673 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
674 else if (m->queue_mode == DM_TYPE_BIO_BASED)
675 queue_work(kmultipathd, &m->process_queued_bios);
678 static void process_queued_bios(struct work_struct *work)
683 struct bio_list bios;
684 struct blk_plug plug;
685 struct multipath *m =
686 container_of(work, struct multipath, process_queued_bios);
688 bio_list_init(&bios);
690 spin_lock_irqsave(&m->lock, flags);
692 if (bio_list_empty(&m->queued_bios)) {
693 spin_unlock_irqrestore(&m->lock, flags);
697 bio_list_merge(&bios, &m->queued_bios);
698 bio_list_init(&m->queued_bios);
700 spin_unlock_irqrestore(&m->lock, flags);
702 blk_start_plug(&plug);
703 while ((bio = bio_list_pop(&bios))) {
704 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
705 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
706 r = __multipath_map_bio(m, bio, mpio);
709 bio->bi_status = BLK_STS_IOERR;
712 case DM_MAPIO_REQUEUE:
713 bio->bi_status = BLK_STS_DM_REQUEUE;
716 case DM_MAPIO_REMAPPED:
717 submit_bio_noacct(bio);
719 case DM_MAPIO_SUBMITTED:
722 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
725 blk_finish_plug(&plug);
729 * If we run out of usable paths, should we queue I/O or error it?
731 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
732 bool save_old_value, const char *caller)
735 bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
736 const char *dm_dev_name = dm_table_device_name(m->ti->table);
738 DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
739 dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
741 spin_lock_irqsave(&m->lock, flags);
743 queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
744 saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
746 if (save_old_value) {
747 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
748 DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
751 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
752 } else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
753 /* due to "fail_if_no_path" message, need to honor it. */
754 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
756 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
758 DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
759 dm_dev_name, __func__,
760 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
761 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
762 dm_noflush_suspending(m->ti));
764 spin_unlock_irqrestore(&m->lock, flags);
766 if (!queue_if_no_path) {
767 dm_table_run_md_queue_async(m->ti->table);
768 process_queued_io_list(m);
775 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
776 * process any queued I/O.
778 static void queue_if_no_path_timeout_work(struct timer_list *t)
780 struct multipath *m = from_timer(m, t, nopath_timer);
782 DMWARN("queue_if_no_path timeout on %s, failing queued IO",
783 dm_table_device_name(m->ti->table));
784 queue_if_no_path(m, false, false, __func__);
788 * Enable the queue_if_no_path timeout if necessary.
789 * Called with m->lock held.
791 static void enable_nopath_timeout(struct multipath *m)
793 unsigned long queue_if_no_path_timeout =
794 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
796 lockdep_assert_held(&m->lock);
798 if (queue_if_no_path_timeout > 0 &&
799 atomic_read(&m->nr_valid_paths) == 0 &&
800 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
801 mod_timer(&m->nopath_timer,
802 jiffies + queue_if_no_path_timeout);
806 static void disable_nopath_timeout(struct multipath *m)
808 del_timer_sync(&m->nopath_timer);
812 * An event is triggered whenever a path is taken out of use.
813 * Includes path failure and PG bypass.
815 static void trigger_event(struct work_struct *work)
817 struct multipath *m =
818 container_of(work, struct multipath, trigger_event);
820 dm_table_event(m->ti->table);
823 /*-----------------------------------------------------------------
824 * Constructor/argument parsing:
825 * <#multipath feature args> [<arg>]*
826 * <#hw_handler args> [hw_handler [<arg>]*]
828 * <initial priority group>
829 * [<selector> <#selector args> [<arg>]*
830 * <#paths> <#per-path selector args>
831 * [<path> [<arg>]* ]+ ]+
832 *---------------------------------------------------------------*/
833 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
834 struct dm_target *ti)
837 struct path_selector_type *pst;
840 static const struct dm_arg _args[] = {
841 {0, 1024, "invalid number of path selector args"},
844 pst = dm_get_path_selector(dm_shift_arg(as));
846 ti->error = "unknown path selector type";
850 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
852 dm_put_path_selector(pst);
856 r = pst->create(&pg->ps, ps_argc, as->argv);
858 dm_put_path_selector(pst);
859 ti->error = "path selector constructor failed";
864 dm_consume_args(as, ps_argc);
869 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
870 const char **attached_handler_name, char **error)
872 struct request_queue *q = bdev_get_queue(bdev);
875 if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
877 if (*attached_handler_name) {
879 * Clear any hw_handler_params associated with a
880 * handler that isn't already attached.
882 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
883 kfree(m->hw_handler_params);
884 m->hw_handler_params = NULL;
888 * Reset hw_handler_name to match the attached handler
890 * NB. This modifies the table line to show the actual
891 * handler instead of the original table passed in.
893 kfree(m->hw_handler_name);
894 m->hw_handler_name = *attached_handler_name;
895 *attached_handler_name = NULL;
899 if (m->hw_handler_name) {
900 r = scsi_dh_attach(q, m->hw_handler_name);
902 char b[BDEVNAME_SIZE];
904 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
909 *error = "error attaching hardware handler";
913 if (m->hw_handler_params) {
914 r = scsi_dh_set_params(q, m->hw_handler_params);
916 *error = "unable to set hardware handler parameters";
925 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
926 struct dm_target *ti)
930 struct multipath *m = ti->private;
931 struct request_queue *q;
932 const char *attached_handler_name = NULL;
934 /* we need at least a path arg */
936 ti->error = "no device given";
937 return ERR_PTR(-EINVAL);
942 return ERR_PTR(-ENOMEM);
944 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
947 ti->error = "error getting device";
951 q = bdev_get_queue(p->path.dev->bdev);
952 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
953 if (attached_handler_name || m->hw_handler_name) {
954 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
955 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
956 kfree(attached_handler_name);
958 dm_put_device(ti, p->path.dev);
963 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
965 dm_put_device(ti, p->path.dev);
975 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
978 static const struct dm_arg _args[] = {
979 {1, 1024, "invalid number of paths"},
980 {0, 1024, "invalid number of selector args"}
984 unsigned i, nr_selector_args, nr_args;
985 struct priority_group *pg;
986 struct dm_target *ti = m->ti;
990 ti->error = "not enough priority group arguments";
991 return ERR_PTR(-EINVAL);
994 pg = alloc_priority_group();
996 ti->error = "couldn't allocate priority group";
997 return ERR_PTR(-ENOMEM);
1001 r = parse_path_selector(as, pg, ti);
1008 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1012 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1016 nr_args = 1 + nr_selector_args;
1017 for (i = 0; i < pg->nr_pgpaths; i++) {
1018 struct pgpath *pgpath;
1019 struct dm_arg_set path_args;
1021 if (as->argc < nr_args) {
1022 ti->error = "not enough path parameters";
1027 path_args.argc = nr_args;
1028 path_args.argv = as->argv;
1030 pgpath = parse_path(&path_args, &pg->ps, ti);
1031 if (IS_ERR(pgpath)) {
1032 r = PTR_ERR(pgpath);
1037 list_add_tail(&pgpath->list, &pg->pgpaths);
1038 dm_consume_args(as, nr_args);
1044 free_priority_group(pg, ti);
1048 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1052 struct dm_target *ti = m->ti;
1054 static const struct dm_arg _args[] = {
1055 {0, 1024, "invalid number of hardware handler args"},
1058 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1064 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1065 dm_consume_args(as, hw_argc);
1066 DMERR("bio-based multipath doesn't allow hardware handler args");
1070 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1071 if (!m->hw_handler_name)
1078 for (i = 0; i <= hw_argc - 2; i++)
1079 len += strlen(as->argv[i]) + 1;
1080 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1082 ti->error = "memory allocation failed";
1086 j = sprintf(p, "%d", hw_argc - 1);
1087 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1088 j = sprintf(p, "%s", as->argv[i]);
1090 dm_consume_args(as, hw_argc - 1);
1094 kfree(m->hw_handler_name);
1095 m->hw_handler_name = NULL;
1099 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1103 struct dm_target *ti = m->ti;
1104 const char *arg_name;
1106 static const struct dm_arg _args[] = {
1107 {0, 8, "invalid number of feature args"},
1108 {1, 50, "pg_init_retries must be between 1 and 50"},
1109 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1112 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1120 arg_name = dm_shift_arg(as);
1123 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1124 r = queue_if_no_path(m, true, false, __func__);
1128 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1129 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1133 if (!strcasecmp(arg_name, "pg_init_retries") &&
1135 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1140 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1142 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1147 if (!strcasecmp(arg_name, "queue_mode") &&
1149 const char *queue_mode_name = dm_shift_arg(as);
1151 if (!strcasecmp(queue_mode_name, "bio"))
1152 m->queue_mode = DM_TYPE_BIO_BASED;
1153 else if (!strcasecmp(queue_mode_name, "rq") ||
1154 !strcasecmp(queue_mode_name, "mq"))
1155 m->queue_mode = DM_TYPE_REQUEST_BASED;
1157 ti->error = "Unknown 'queue_mode' requested";
1164 ti->error = "Unrecognised multipath feature request";
1166 } while (argc && !r);
1171 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1173 /* target arguments */
1174 static const struct dm_arg _args[] = {
1175 {0, 1024, "invalid number of priority groups"},
1176 {0, 1024, "invalid initial priority group number"},
1180 struct multipath *m;
1181 struct dm_arg_set as;
1182 unsigned pg_count = 0;
1183 unsigned next_pg_num;
1184 unsigned long flags;
1189 m = alloc_multipath(ti);
1191 ti->error = "can't allocate multipath";
1195 r = parse_features(&as, m);
1199 r = alloc_multipath_stage2(ti, m);
1203 r = parse_hw_handler(&as, m);
1207 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1211 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1215 if ((!m->nr_priority_groups && next_pg_num) ||
1216 (m->nr_priority_groups && !next_pg_num)) {
1217 ti->error = "invalid initial priority group";
1222 /* parse the priority groups */
1224 struct priority_group *pg;
1225 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1227 pg = parse_priority_group(&as, m);
1233 nr_valid_paths += pg->nr_pgpaths;
1234 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1236 list_add_tail(&pg->list, &m->priority_groups);
1238 pg->pg_num = pg_count;
1243 if (pg_count != m->nr_priority_groups) {
1244 ti->error = "priority group count mismatch";
1249 spin_lock_irqsave(&m->lock, flags);
1250 enable_nopath_timeout(m);
1251 spin_unlock_irqrestore(&m->lock, flags);
1253 ti->num_flush_bios = 1;
1254 ti->num_discard_bios = 1;
1255 ti->num_write_same_bios = 1;
1256 ti->num_write_zeroes_bios = 1;
1257 if (m->queue_mode == DM_TYPE_BIO_BASED)
1258 ti->per_io_data_size = multipath_per_bio_data_size();
1260 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1269 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1274 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1276 if (!atomic_read(&m->pg_init_in_progress))
1281 finish_wait(&m->pg_init_wait, &wait);
1284 static void flush_multipath_work(struct multipath *m)
1286 if (m->hw_handler_name) {
1287 unsigned long flags;
1289 if (!atomic_read(&m->pg_init_in_progress))
1292 spin_lock_irqsave(&m->lock, flags);
1293 if (atomic_read(&m->pg_init_in_progress) &&
1294 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1295 spin_unlock_irqrestore(&m->lock, flags);
1297 flush_workqueue(kmpath_handlerd);
1298 multipath_wait_for_pg_init_completion(m);
1300 spin_lock_irqsave(&m->lock, flags);
1301 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1303 spin_unlock_irqrestore(&m->lock, flags);
1306 if (m->queue_mode == DM_TYPE_BIO_BASED)
1307 flush_work(&m->process_queued_bios);
1308 flush_work(&m->trigger_event);
1311 static void multipath_dtr(struct dm_target *ti)
1313 struct multipath *m = ti->private;
1315 disable_nopath_timeout(m);
1316 flush_multipath_work(m);
1321 * Take a path out of use.
1323 static int fail_path(struct pgpath *pgpath)
1325 unsigned long flags;
1326 struct multipath *m = pgpath->pg->m;
1328 spin_lock_irqsave(&m->lock, flags);
1330 if (!pgpath->is_active)
1333 DMWARN("%s: Failing path %s.",
1334 dm_table_device_name(m->ti->table),
1335 pgpath->path.dev->name);
1337 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1338 pgpath->is_active = false;
1339 pgpath->fail_count++;
1341 atomic_dec(&m->nr_valid_paths);
1343 if (pgpath == m->current_pgpath)
1344 m->current_pgpath = NULL;
1346 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1347 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1349 schedule_work(&m->trigger_event);
1351 enable_nopath_timeout(m);
1354 spin_unlock_irqrestore(&m->lock, flags);
1360 * Reinstate a previously-failed path
1362 static int reinstate_path(struct pgpath *pgpath)
1364 int r = 0, run_queue = 0;
1365 unsigned long flags;
1366 struct multipath *m = pgpath->pg->m;
1367 unsigned nr_valid_paths;
1369 spin_lock_irqsave(&m->lock, flags);
1371 if (pgpath->is_active)
1374 DMWARN("%s: Reinstating path %s.",
1375 dm_table_device_name(m->ti->table),
1376 pgpath->path.dev->name);
1378 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1382 pgpath->is_active = true;
1384 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1385 if (nr_valid_paths == 1) {
1386 m->current_pgpath = NULL;
1388 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1389 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1390 atomic_inc(&m->pg_init_in_progress);
1393 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1394 pgpath->path.dev->name, nr_valid_paths);
1396 schedule_work(&m->trigger_event);
1399 spin_unlock_irqrestore(&m->lock, flags);
1401 dm_table_run_md_queue_async(m->ti->table);
1402 process_queued_io_list(m);
1405 if (pgpath->is_active)
1406 disable_nopath_timeout(m);
1412 * Fail or reinstate all paths that match the provided struct dm_dev.
1414 static int action_dev(struct multipath *m, struct dm_dev *dev,
1418 struct pgpath *pgpath;
1419 struct priority_group *pg;
1421 list_for_each_entry(pg, &m->priority_groups, list) {
1422 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1423 if (pgpath->path.dev == dev)
1432 * Temporarily try to avoid having to use the specified PG
1434 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1437 unsigned long flags;
1439 spin_lock_irqsave(&m->lock, flags);
1441 pg->bypassed = bypassed;
1442 m->current_pgpath = NULL;
1443 m->current_pg = NULL;
1445 spin_unlock_irqrestore(&m->lock, flags);
1447 schedule_work(&m->trigger_event);
1451 * Switch to using the specified PG from the next I/O that gets mapped
1453 static int switch_pg_num(struct multipath *m, const char *pgstr)
1455 struct priority_group *pg;
1457 unsigned long flags;
1460 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1461 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1462 DMWARN("invalid PG number supplied to switch_pg_num");
1466 spin_lock_irqsave(&m->lock, flags);
1467 list_for_each_entry(pg, &m->priority_groups, list) {
1468 pg->bypassed = false;
1472 m->current_pgpath = NULL;
1473 m->current_pg = NULL;
1476 spin_unlock_irqrestore(&m->lock, flags);
1478 schedule_work(&m->trigger_event);
1483 * Set/clear bypassed status of a PG.
1484 * PGs are numbered upwards from 1 in the order they were declared.
1486 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1488 struct priority_group *pg;
1492 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1493 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1494 DMWARN("invalid PG number supplied to bypass_pg");
1498 list_for_each_entry(pg, &m->priority_groups, list) {
1503 bypass_pg(m, pg, bypassed);
1508 * Should we retry pg_init immediately?
1510 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1512 unsigned long flags;
1513 bool limit_reached = false;
1515 spin_lock_irqsave(&m->lock, flags);
1517 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1518 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1519 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1521 limit_reached = true;
1523 spin_unlock_irqrestore(&m->lock, flags);
1525 return limit_reached;
1528 static void pg_init_done(void *data, int errors)
1530 struct pgpath *pgpath = data;
1531 struct priority_group *pg = pgpath->pg;
1532 struct multipath *m = pg->m;
1533 unsigned long flags;
1534 bool delay_retry = false;
1536 /* device or driver problems */
1541 if (!m->hw_handler_name) {
1545 DMERR("Could not failover the device: Handler scsi_dh_%s "
1546 "Error %d.", m->hw_handler_name, errors);
1548 * Fail path for now, so we do not ping pong
1552 case SCSI_DH_DEV_TEMP_BUSY:
1554 * Probably doing something like FW upgrade on the
1555 * controller so try the other pg.
1557 bypass_pg(m, pg, true);
1560 /* Wait before retrying. */
1563 case SCSI_DH_IMM_RETRY:
1564 case SCSI_DH_RES_TEMP_UNAVAIL:
1565 if (pg_init_limit_reached(m, pgpath))
1569 case SCSI_DH_DEV_OFFLINED:
1572 * We probably do not want to fail the path for a device
1573 * error, but this is what the old dm did. In future
1574 * patches we can do more advanced handling.
1579 spin_lock_irqsave(&m->lock, flags);
1581 if (pgpath == m->current_pgpath) {
1582 DMERR("Could not failover device. Error %d.", errors);
1583 m->current_pgpath = NULL;
1584 m->current_pg = NULL;
1586 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1587 pg->bypassed = false;
1589 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1590 /* Activations of other paths are still on going */
1593 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1595 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1597 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1599 if (__pg_init_all_paths(m))
1602 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1604 process_queued_io_list(m);
1607 * Wake up any thread waiting to suspend.
1609 wake_up(&m->pg_init_wait);
1612 spin_unlock_irqrestore(&m->lock, flags);
1615 static void activate_or_offline_path(struct pgpath *pgpath)
1617 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1619 if (pgpath->is_active && !blk_queue_dying(q))
1620 scsi_dh_activate(q, pg_init_done, pgpath);
1622 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1625 static void activate_path_work(struct work_struct *work)
1627 struct pgpath *pgpath =
1628 container_of(work, struct pgpath, activate_path.work);
1630 activate_or_offline_path(pgpath);
1633 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1634 blk_status_t error, union map_info *map_context)
1636 struct dm_mpath_io *mpio = get_mpio(map_context);
1637 struct pgpath *pgpath = mpio->pgpath;
1638 int r = DM_ENDIO_DONE;
1641 * We don't queue any clone request inside the multipath target
1642 * during end I/O handling, since those clone requests don't have
1643 * bio clones. If we queue them inside the multipath target,
1644 * we need to make bio clones, that requires memory allocation.
1645 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1646 * don't have bio clones.)
1647 * Instead of queueing the clone request here, we queue the original
1648 * request into dm core, which will remake a clone request and
1649 * clone bios for it and resubmit it later.
1651 if (error && blk_path_error(error)) {
1652 struct multipath *m = ti->private;
1654 if (error == BLK_STS_RESOURCE)
1655 r = DM_ENDIO_DELAY_REQUEUE;
1657 r = DM_ENDIO_REQUEUE;
1662 if (!atomic_read(&m->nr_valid_paths) &&
1663 !must_push_back_rq(m)) {
1664 if (error == BLK_STS_IOERR)
1666 /* complete with the original error */
1672 struct path_selector *ps = &pgpath->pg->ps;
1674 if (ps->type->end_io)
1675 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1676 clone->io_start_time_ns);
1682 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1683 blk_status_t *error)
1685 struct multipath *m = ti->private;
1686 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1687 struct pgpath *pgpath = mpio->pgpath;
1688 unsigned long flags;
1689 int r = DM_ENDIO_DONE;
1691 if (!*error || !blk_path_error(*error))
1697 if (!atomic_read(&m->nr_valid_paths)) {
1698 spin_lock_irqsave(&m->lock, flags);
1699 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1700 if (__must_push_back(m)) {
1701 r = DM_ENDIO_REQUEUE;
1704 *error = BLK_STS_IOERR;
1706 spin_unlock_irqrestore(&m->lock, flags);
1709 spin_unlock_irqrestore(&m->lock, flags);
1712 multipath_queue_bio(m, clone);
1713 r = DM_ENDIO_INCOMPLETE;
1716 struct path_selector *ps = &pgpath->pg->ps;
1718 if (ps->type->end_io)
1719 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1720 dm_start_time_ns_from_clone(clone));
1727 * Suspend with flush can't complete until all the I/O is processed
1728 * so if the last path fails we must error any remaining I/O.
1729 * - Note that if the freeze_bdev fails while suspending, the
1730 * queue_if_no_path state is lost - userspace should reset it.
1731 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1733 static void multipath_presuspend(struct dm_target *ti)
1735 struct multipath *m = ti->private;
1737 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1738 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1739 queue_if_no_path(m, false, true, __func__);
1742 static void multipath_postsuspend(struct dm_target *ti)
1744 struct multipath *m = ti->private;
1746 mutex_lock(&m->work_mutex);
1747 flush_multipath_work(m);
1748 mutex_unlock(&m->work_mutex);
1752 * Restore the queue_if_no_path setting.
1754 static void multipath_resume(struct dm_target *ti)
1756 struct multipath *m = ti->private;
1757 unsigned long flags;
1759 spin_lock_irqsave(&m->lock, flags);
1760 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1761 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1762 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1765 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1766 dm_table_device_name(m->ti->table), __func__,
1767 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1768 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1770 spin_unlock_irqrestore(&m->lock, flags);
1774 * Info output has the following format:
1775 * num_multipath_feature_args [multipath_feature_args]*
1776 * num_handler_status_args [handler_status_args]*
1777 * num_groups init_group_number
1778 * [A|D|E num_ps_status_args [ps_status_args]*
1779 * num_paths num_selector_args
1780 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1782 * Table output has the following format (identical to the constructor string):
1783 * num_feature_args [features_args]*
1784 * num_handler_args hw_handler [hw_handler_args]*
1785 * num_groups init_group_number
1786 * [priority selector-name num_ps_args [ps_args]*
1787 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1789 static void multipath_status(struct dm_target *ti, status_type_t type,
1790 unsigned status_flags, char *result, unsigned maxlen)
1792 int sz = 0, pg_counter, pgpath_counter;
1793 unsigned long flags;
1794 struct multipath *m = ti->private;
1795 struct priority_group *pg;
1800 spin_lock_irqsave(&m->lock, flags);
1803 if (type == STATUSTYPE_INFO)
1804 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1805 atomic_read(&m->pg_init_count));
1807 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1808 (m->pg_init_retries > 0) * 2 +
1809 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1810 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1811 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1813 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1814 DMEMIT("queue_if_no_path ");
1815 if (m->pg_init_retries)
1816 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1817 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1818 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1819 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1820 DMEMIT("retain_attached_hw_handler ");
1821 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1822 switch(m->queue_mode) {
1823 case DM_TYPE_BIO_BASED:
1824 DMEMIT("queue_mode bio ");
1833 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1836 DMEMIT("1 %s ", m->hw_handler_name);
1838 DMEMIT("%u ", m->nr_priority_groups);
1841 pg_num = m->next_pg->pg_num;
1842 else if (m->current_pg)
1843 pg_num = m->current_pg->pg_num;
1845 pg_num = (m->nr_priority_groups ? 1 : 0);
1847 DMEMIT("%u ", pg_num);
1850 case STATUSTYPE_INFO:
1851 list_for_each_entry(pg, &m->priority_groups, list) {
1853 state = 'D'; /* Disabled */
1854 else if (pg == m->current_pg)
1855 state = 'A'; /* Currently Active */
1857 state = 'E'; /* Enabled */
1859 DMEMIT("%c ", state);
1861 if (pg->ps.type->status)
1862 sz += pg->ps.type->status(&pg->ps, NULL, type,
1868 DMEMIT("%u %u ", pg->nr_pgpaths,
1869 pg->ps.type->info_args);
1871 list_for_each_entry(p, &pg->pgpaths, list) {
1872 DMEMIT("%s %s %u ", p->path.dev->name,
1873 p->is_active ? "A" : "F",
1875 if (pg->ps.type->status)
1876 sz += pg->ps.type->status(&pg->ps,
1877 &p->path, type, result + sz,
1883 case STATUSTYPE_TABLE:
1884 list_for_each_entry(pg, &m->priority_groups, list) {
1885 DMEMIT("%s ", pg->ps.type->name);
1887 if (pg->ps.type->status)
1888 sz += pg->ps.type->status(&pg->ps, NULL, type,
1894 DMEMIT("%u %u ", pg->nr_pgpaths,
1895 pg->ps.type->table_args);
1897 list_for_each_entry(p, &pg->pgpaths, list) {
1898 DMEMIT("%s ", p->path.dev->name);
1899 if (pg->ps.type->status)
1900 sz += pg->ps.type->status(&pg->ps,
1901 &p->path, type, result + sz,
1907 case STATUSTYPE_IMA:
1908 sz = 0; /*reset the result pointer*/
1910 DMEMIT_TARGET_NAME_VERSION(ti->type);
1911 DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1914 list_for_each_entry(pg, &m->priority_groups, list) {
1916 state = 'D'; /* Disabled */
1917 else if (pg == m->current_pg)
1918 state = 'A'; /* Currently Active */
1920 state = 'E'; /* Enabled */
1921 DMEMIT(",pg_state_%d=%c", pg_counter, state);
1922 DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1923 DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1926 list_for_each_entry(p, &pg->pgpaths, list) {
1927 DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1928 pg_counter, pgpath_counter, p->path.dev->name,
1929 pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1930 pg_counter, pgpath_counter, p->fail_count);
1931 if (pg->ps.type->status) {
1932 DMEMIT(",path_selector_status_%d_%d=",
1933 pg_counter, pgpath_counter);
1934 sz += pg->ps.type->status(&pg->ps, &p->path,
1946 spin_unlock_irqrestore(&m->lock, flags);
1949 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1950 char *result, unsigned maxlen)
1954 struct multipath *m = ti->private;
1956 unsigned long flags;
1958 mutex_lock(&m->work_mutex);
1960 if (dm_suspended(ti)) {
1966 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1967 r = queue_if_no_path(m, true, false, __func__);
1968 spin_lock_irqsave(&m->lock, flags);
1969 enable_nopath_timeout(m);
1970 spin_unlock_irqrestore(&m->lock, flags);
1972 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1973 r = queue_if_no_path(m, false, false, __func__);
1974 disable_nopath_timeout(m);
1980 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1984 if (!strcasecmp(argv[0], "disable_group")) {
1985 r = bypass_pg_num(m, argv[1], true);
1987 } else if (!strcasecmp(argv[0], "enable_group")) {
1988 r = bypass_pg_num(m, argv[1], false);
1990 } else if (!strcasecmp(argv[0], "switch_group")) {
1991 r = switch_pg_num(m, argv[1]);
1993 } else if (!strcasecmp(argv[0], "reinstate_path"))
1994 action = reinstate_path;
1995 else if (!strcasecmp(argv[0], "fail_path"))
1998 DMWARN("Unrecognised multipath message received: %s", argv[0]);
2002 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
2004 DMWARN("message: error getting device %s",
2009 r = action_dev(m, dev, action);
2011 dm_put_device(ti, dev);
2014 mutex_unlock(&m->work_mutex);
2018 static int multipath_prepare_ioctl(struct dm_target *ti,
2019 struct block_device **bdev)
2021 struct multipath *m = ti->private;
2022 struct pgpath *pgpath;
2023 unsigned long flags;
2026 pgpath = READ_ONCE(m->current_pgpath);
2027 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2028 pgpath = choose_pgpath(m, 0);
2031 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2032 *bdev = pgpath->path.dev->bdev;
2035 /* pg_init has not started or completed */
2039 /* No path is available */
2041 spin_lock_irqsave(&m->lock, flags);
2042 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2044 spin_unlock_irqrestore(&m->lock, flags);
2047 if (r == -ENOTCONN) {
2048 if (!READ_ONCE(m->current_pg)) {
2049 /* Path status changed, redo selection */
2050 (void) choose_pgpath(m, 0);
2052 spin_lock_irqsave(&m->lock, flags);
2053 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2054 (void) __pg_init_all_paths(m);
2055 spin_unlock_irqrestore(&m->lock, flags);
2056 dm_table_run_md_queue_async(m->ti->table);
2057 process_queued_io_list(m);
2061 * Only pass ioctls through if the device sizes match exactly.
2063 if (!r && ti->len != bdev_nr_sectors((*bdev)))
2068 static int multipath_iterate_devices(struct dm_target *ti,
2069 iterate_devices_callout_fn fn, void *data)
2071 struct multipath *m = ti->private;
2072 struct priority_group *pg;
2076 list_for_each_entry(pg, &m->priority_groups, list) {
2077 list_for_each_entry(p, &pg->pgpaths, list) {
2078 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2088 static int pgpath_busy(struct pgpath *pgpath)
2090 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2092 return blk_lld_busy(q);
2096 * We return "busy", only when we can map I/Os but underlying devices
2097 * are busy (so even if we map I/Os now, the I/Os will wait on
2098 * the underlying queue).
2099 * In other words, if we want to kill I/Os or queue them inside us
2100 * due to map unavailability, we don't return "busy". Otherwise,
2101 * dm core won't give us the I/Os and we can't do what we want.
2103 static int multipath_busy(struct dm_target *ti)
2105 bool busy = false, has_active = false;
2106 struct multipath *m = ti->private;
2107 struct priority_group *pg, *next_pg;
2108 struct pgpath *pgpath;
2110 /* pg_init in progress */
2111 if (atomic_read(&m->pg_init_in_progress))
2114 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2115 if (!atomic_read(&m->nr_valid_paths)) {
2116 unsigned long flags;
2117 spin_lock_irqsave(&m->lock, flags);
2118 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2119 spin_unlock_irqrestore(&m->lock, flags);
2120 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2122 spin_unlock_irqrestore(&m->lock, flags);
2125 /* Guess which priority_group will be used at next mapping time */
2126 pg = READ_ONCE(m->current_pg);
2127 next_pg = READ_ONCE(m->next_pg);
2128 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2133 * We don't know which pg will be used at next mapping time.
2134 * We don't call choose_pgpath() here to avoid to trigger
2135 * pg_init just by busy checking.
2136 * So we don't know whether underlying devices we will be using
2137 * at next mapping time are busy or not. Just try mapping.
2143 * If there is one non-busy active path at least, the path selector
2144 * will be able to select it. So we consider such a pg as not busy.
2147 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2148 if (pgpath->is_active) {
2150 if (!pgpath_busy(pgpath)) {
2159 * No active path in this pg, so this pg won't be used and
2160 * the current_pg will be changed at next mapping time.
2161 * We need to try mapping to determine it.
2169 /*-----------------------------------------------------------------
2171 *---------------------------------------------------------------*/
2172 static struct target_type multipath_target = {
2173 .name = "multipath",
2174 .version = {1, 14, 0},
2175 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2176 DM_TARGET_PASSES_INTEGRITY,
2177 .module = THIS_MODULE,
2178 .ctr = multipath_ctr,
2179 .dtr = multipath_dtr,
2180 .clone_and_map_rq = multipath_clone_and_map,
2181 .release_clone_rq = multipath_release_clone,
2182 .rq_end_io = multipath_end_io,
2183 .map = multipath_map_bio,
2184 .end_io = multipath_end_io_bio,
2185 .presuspend = multipath_presuspend,
2186 .postsuspend = multipath_postsuspend,
2187 .resume = multipath_resume,
2188 .status = multipath_status,
2189 .message = multipath_message,
2190 .prepare_ioctl = multipath_prepare_ioctl,
2191 .iterate_devices = multipath_iterate_devices,
2192 .busy = multipath_busy,
2195 static int __init dm_multipath_init(void)
2199 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2201 DMERR("failed to create workqueue kmpathd");
2203 goto bad_alloc_kmultipathd;
2207 * A separate workqueue is used to handle the device handlers
2208 * to avoid overloading existing workqueue. Overloading the
2209 * old workqueue would also create a bottleneck in the
2210 * path of the storage hardware device activation.
2212 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2214 if (!kmpath_handlerd) {
2215 DMERR("failed to create workqueue kmpath_handlerd");
2217 goto bad_alloc_kmpath_handlerd;
2220 r = dm_register_target(&multipath_target);
2222 DMERR("request-based register failed %d", r);
2224 goto bad_register_target;
2229 bad_register_target:
2230 destroy_workqueue(kmpath_handlerd);
2231 bad_alloc_kmpath_handlerd:
2232 destroy_workqueue(kmultipathd);
2233 bad_alloc_kmultipathd:
2237 static void __exit dm_multipath_exit(void)
2239 destroy_workqueue(kmpath_handlerd);
2240 destroy_workqueue(kmultipathd);
2242 dm_unregister_target(&multipath_target);
2245 module_init(dm_multipath_init);
2246 module_exit(dm_multipath_exit);
2248 module_param_named(queue_if_no_path_timeout_secs,
2249 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2250 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2252 MODULE_DESCRIPTION(DM_NAME " multipath target");
2253 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2254 MODULE_LICENSE("GPL");