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 DMINFO("retaining handler on device %pg", bdev);
906 *error = "error attaching hardware handler";
910 if (m->hw_handler_params) {
911 r = scsi_dh_set_params(q, m->hw_handler_params);
913 *error = "unable to set hardware handler parameters";
922 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
923 struct dm_target *ti)
927 struct multipath *m = ti->private;
928 struct request_queue *q;
929 const char *attached_handler_name = NULL;
931 /* we need at least a path arg */
933 ti->error = "no device given";
934 return ERR_PTR(-EINVAL);
939 return ERR_PTR(-ENOMEM);
941 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
944 ti->error = "error getting device";
948 q = bdev_get_queue(p->path.dev->bdev);
949 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
950 if (attached_handler_name || m->hw_handler_name) {
951 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
952 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
953 kfree(attached_handler_name);
955 dm_put_device(ti, p->path.dev);
960 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
962 dm_put_device(ti, p->path.dev);
972 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
975 static const struct dm_arg _args[] = {
976 {1, 1024, "invalid number of paths"},
977 {0, 1024, "invalid number of selector args"}
981 unsigned i, nr_selector_args, nr_args;
982 struct priority_group *pg;
983 struct dm_target *ti = m->ti;
987 ti->error = "not enough priority group arguments";
988 return ERR_PTR(-EINVAL);
991 pg = alloc_priority_group();
993 ti->error = "couldn't allocate priority group";
994 return ERR_PTR(-ENOMEM);
998 r = parse_path_selector(as, pg, ti);
1005 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1009 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1013 nr_args = 1 + nr_selector_args;
1014 for (i = 0; i < pg->nr_pgpaths; i++) {
1015 struct pgpath *pgpath;
1016 struct dm_arg_set path_args;
1018 if (as->argc < nr_args) {
1019 ti->error = "not enough path parameters";
1024 path_args.argc = nr_args;
1025 path_args.argv = as->argv;
1027 pgpath = parse_path(&path_args, &pg->ps, ti);
1028 if (IS_ERR(pgpath)) {
1029 r = PTR_ERR(pgpath);
1034 list_add_tail(&pgpath->list, &pg->pgpaths);
1035 dm_consume_args(as, nr_args);
1041 free_priority_group(pg, ti);
1045 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1049 struct dm_target *ti = m->ti;
1051 static const struct dm_arg _args[] = {
1052 {0, 1024, "invalid number of hardware handler args"},
1055 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1061 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1062 dm_consume_args(as, hw_argc);
1063 DMERR("bio-based multipath doesn't allow hardware handler args");
1067 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1068 if (!m->hw_handler_name)
1075 for (i = 0; i <= hw_argc - 2; i++)
1076 len += strlen(as->argv[i]) + 1;
1077 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1079 ti->error = "memory allocation failed";
1083 j = sprintf(p, "%d", hw_argc - 1);
1084 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1085 j = sprintf(p, "%s", as->argv[i]);
1087 dm_consume_args(as, hw_argc - 1);
1091 kfree(m->hw_handler_name);
1092 m->hw_handler_name = NULL;
1096 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1100 struct dm_target *ti = m->ti;
1101 const char *arg_name;
1103 static const struct dm_arg _args[] = {
1104 {0, 8, "invalid number of feature args"},
1105 {1, 50, "pg_init_retries must be between 1 and 50"},
1106 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1109 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1117 arg_name = dm_shift_arg(as);
1120 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1121 r = queue_if_no_path(m, true, false, __func__);
1125 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1126 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1130 if (!strcasecmp(arg_name, "pg_init_retries") &&
1132 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1137 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1139 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1144 if (!strcasecmp(arg_name, "queue_mode") &&
1146 const char *queue_mode_name = dm_shift_arg(as);
1148 if (!strcasecmp(queue_mode_name, "bio"))
1149 m->queue_mode = DM_TYPE_BIO_BASED;
1150 else if (!strcasecmp(queue_mode_name, "rq") ||
1151 !strcasecmp(queue_mode_name, "mq"))
1152 m->queue_mode = DM_TYPE_REQUEST_BASED;
1154 ti->error = "Unknown 'queue_mode' requested";
1161 ti->error = "Unrecognised multipath feature request";
1163 } while (argc && !r);
1168 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1170 /* target arguments */
1171 static const struct dm_arg _args[] = {
1172 {0, 1024, "invalid number of priority groups"},
1173 {0, 1024, "invalid initial priority group number"},
1177 struct multipath *m;
1178 struct dm_arg_set as;
1179 unsigned pg_count = 0;
1180 unsigned next_pg_num;
1181 unsigned long flags;
1186 m = alloc_multipath(ti);
1188 ti->error = "can't allocate multipath";
1192 r = parse_features(&as, m);
1196 r = alloc_multipath_stage2(ti, m);
1200 r = parse_hw_handler(&as, m);
1204 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1208 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1212 if ((!m->nr_priority_groups && next_pg_num) ||
1213 (m->nr_priority_groups && !next_pg_num)) {
1214 ti->error = "invalid initial priority group";
1219 /* parse the priority groups */
1221 struct priority_group *pg;
1222 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1224 pg = parse_priority_group(&as, m);
1230 nr_valid_paths += pg->nr_pgpaths;
1231 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1233 list_add_tail(&pg->list, &m->priority_groups);
1235 pg->pg_num = pg_count;
1240 if (pg_count != m->nr_priority_groups) {
1241 ti->error = "priority group count mismatch";
1246 spin_lock_irqsave(&m->lock, flags);
1247 enable_nopath_timeout(m);
1248 spin_unlock_irqrestore(&m->lock, flags);
1250 ti->num_flush_bios = 1;
1251 ti->num_discard_bios = 1;
1252 ti->num_write_zeroes_bios = 1;
1253 if (m->queue_mode == DM_TYPE_BIO_BASED)
1254 ti->per_io_data_size = multipath_per_bio_data_size();
1256 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1265 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1270 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1272 if (!atomic_read(&m->pg_init_in_progress))
1277 finish_wait(&m->pg_init_wait, &wait);
1280 static void flush_multipath_work(struct multipath *m)
1282 if (m->hw_handler_name) {
1283 unsigned long flags;
1285 if (!atomic_read(&m->pg_init_in_progress))
1288 spin_lock_irqsave(&m->lock, flags);
1289 if (atomic_read(&m->pg_init_in_progress) &&
1290 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1291 spin_unlock_irqrestore(&m->lock, flags);
1293 flush_workqueue(kmpath_handlerd);
1294 multipath_wait_for_pg_init_completion(m);
1296 spin_lock_irqsave(&m->lock, flags);
1297 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1299 spin_unlock_irqrestore(&m->lock, flags);
1302 if (m->queue_mode == DM_TYPE_BIO_BASED)
1303 flush_work(&m->process_queued_bios);
1304 flush_work(&m->trigger_event);
1307 static void multipath_dtr(struct dm_target *ti)
1309 struct multipath *m = ti->private;
1311 disable_nopath_timeout(m);
1312 flush_multipath_work(m);
1317 * Take a path out of use.
1319 static int fail_path(struct pgpath *pgpath)
1321 unsigned long flags;
1322 struct multipath *m = pgpath->pg->m;
1324 spin_lock_irqsave(&m->lock, flags);
1326 if (!pgpath->is_active)
1329 DMWARN("%s: Failing path %s.",
1330 dm_table_device_name(m->ti->table),
1331 pgpath->path.dev->name);
1333 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1334 pgpath->is_active = false;
1335 pgpath->fail_count++;
1337 atomic_dec(&m->nr_valid_paths);
1339 if (pgpath == m->current_pgpath)
1340 m->current_pgpath = NULL;
1342 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1343 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1345 schedule_work(&m->trigger_event);
1347 enable_nopath_timeout(m);
1350 spin_unlock_irqrestore(&m->lock, flags);
1356 * Reinstate a previously-failed path
1358 static int reinstate_path(struct pgpath *pgpath)
1360 int r = 0, run_queue = 0;
1361 unsigned long flags;
1362 struct multipath *m = pgpath->pg->m;
1363 unsigned nr_valid_paths;
1365 spin_lock_irqsave(&m->lock, flags);
1367 if (pgpath->is_active)
1370 DMWARN("%s: Reinstating path %s.",
1371 dm_table_device_name(m->ti->table),
1372 pgpath->path.dev->name);
1374 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1378 pgpath->is_active = true;
1380 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1381 if (nr_valid_paths == 1) {
1382 m->current_pgpath = NULL;
1384 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1385 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1386 atomic_inc(&m->pg_init_in_progress);
1389 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1390 pgpath->path.dev->name, nr_valid_paths);
1392 schedule_work(&m->trigger_event);
1395 spin_unlock_irqrestore(&m->lock, flags);
1397 dm_table_run_md_queue_async(m->ti->table);
1398 process_queued_io_list(m);
1401 if (pgpath->is_active)
1402 disable_nopath_timeout(m);
1408 * Fail or reinstate all paths that match the provided struct dm_dev.
1410 static int action_dev(struct multipath *m, struct dm_dev *dev,
1414 struct pgpath *pgpath;
1415 struct priority_group *pg;
1417 list_for_each_entry(pg, &m->priority_groups, list) {
1418 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1419 if (pgpath->path.dev == dev)
1428 * Temporarily try to avoid having to use the specified PG
1430 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1433 unsigned long flags;
1435 spin_lock_irqsave(&m->lock, flags);
1437 pg->bypassed = bypassed;
1438 m->current_pgpath = NULL;
1439 m->current_pg = NULL;
1441 spin_unlock_irqrestore(&m->lock, flags);
1443 schedule_work(&m->trigger_event);
1447 * Switch to using the specified PG from the next I/O that gets mapped
1449 static int switch_pg_num(struct multipath *m, const char *pgstr)
1451 struct priority_group *pg;
1453 unsigned long flags;
1456 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1457 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1458 DMWARN("invalid PG number supplied to switch_pg_num");
1462 spin_lock_irqsave(&m->lock, flags);
1463 list_for_each_entry(pg, &m->priority_groups, list) {
1464 pg->bypassed = false;
1468 m->current_pgpath = NULL;
1469 m->current_pg = NULL;
1472 spin_unlock_irqrestore(&m->lock, flags);
1474 schedule_work(&m->trigger_event);
1479 * Set/clear bypassed status of a PG.
1480 * PGs are numbered upwards from 1 in the order they were declared.
1482 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1484 struct priority_group *pg;
1488 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1489 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1490 DMWARN("invalid PG number supplied to bypass_pg");
1494 list_for_each_entry(pg, &m->priority_groups, list) {
1499 bypass_pg(m, pg, bypassed);
1504 * Should we retry pg_init immediately?
1506 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1508 unsigned long flags;
1509 bool limit_reached = false;
1511 spin_lock_irqsave(&m->lock, flags);
1513 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1514 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1515 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1517 limit_reached = true;
1519 spin_unlock_irqrestore(&m->lock, flags);
1521 return limit_reached;
1524 static void pg_init_done(void *data, int errors)
1526 struct pgpath *pgpath = data;
1527 struct priority_group *pg = pgpath->pg;
1528 struct multipath *m = pg->m;
1529 unsigned long flags;
1530 bool delay_retry = false;
1532 /* device or driver problems */
1537 if (!m->hw_handler_name) {
1541 DMERR("Could not failover the device: Handler scsi_dh_%s "
1542 "Error %d.", m->hw_handler_name, errors);
1544 * Fail path for now, so we do not ping pong
1548 case SCSI_DH_DEV_TEMP_BUSY:
1550 * Probably doing something like FW upgrade on the
1551 * controller so try the other pg.
1553 bypass_pg(m, pg, true);
1556 /* Wait before retrying. */
1559 case SCSI_DH_IMM_RETRY:
1560 case SCSI_DH_RES_TEMP_UNAVAIL:
1561 if (pg_init_limit_reached(m, pgpath))
1565 case SCSI_DH_DEV_OFFLINED:
1568 * We probably do not want to fail the path for a device
1569 * error, but this is what the old dm did. In future
1570 * patches we can do more advanced handling.
1575 spin_lock_irqsave(&m->lock, flags);
1577 if (pgpath == m->current_pgpath) {
1578 DMERR("Could not failover device. Error %d.", errors);
1579 m->current_pgpath = NULL;
1580 m->current_pg = NULL;
1582 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1583 pg->bypassed = false;
1585 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1586 /* Activations of other paths are still on going */
1589 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1591 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1593 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1595 if (__pg_init_all_paths(m))
1598 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1600 process_queued_io_list(m);
1603 * Wake up any thread waiting to suspend.
1605 wake_up(&m->pg_init_wait);
1608 spin_unlock_irqrestore(&m->lock, flags);
1611 static void activate_or_offline_path(struct pgpath *pgpath)
1613 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1615 if (pgpath->is_active && !blk_queue_dying(q))
1616 scsi_dh_activate(q, pg_init_done, pgpath);
1618 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1621 static void activate_path_work(struct work_struct *work)
1623 struct pgpath *pgpath =
1624 container_of(work, struct pgpath, activate_path.work);
1626 activate_or_offline_path(pgpath);
1629 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1630 blk_status_t error, union map_info *map_context)
1632 struct dm_mpath_io *mpio = get_mpio(map_context);
1633 struct pgpath *pgpath = mpio->pgpath;
1634 int r = DM_ENDIO_DONE;
1637 * We don't queue any clone request inside the multipath target
1638 * during end I/O handling, since those clone requests don't have
1639 * bio clones. If we queue them inside the multipath target,
1640 * we need to make bio clones, that requires memory allocation.
1641 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1642 * don't have bio clones.)
1643 * Instead of queueing the clone request here, we queue the original
1644 * request into dm core, which will remake a clone request and
1645 * clone bios for it and resubmit it later.
1647 if (error && blk_path_error(error)) {
1648 struct multipath *m = ti->private;
1650 if (error == BLK_STS_RESOURCE)
1651 r = DM_ENDIO_DELAY_REQUEUE;
1653 r = DM_ENDIO_REQUEUE;
1658 if (!atomic_read(&m->nr_valid_paths) &&
1659 !must_push_back_rq(m)) {
1660 if (error == BLK_STS_IOERR)
1662 /* complete with the original error */
1668 struct path_selector *ps = &pgpath->pg->ps;
1670 if (ps->type->end_io)
1671 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1672 clone->io_start_time_ns);
1678 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1679 blk_status_t *error)
1681 struct multipath *m = ti->private;
1682 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1683 struct pgpath *pgpath = mpio->pgpath;
1684 unsigned long flags;
1685 int r = DM_ENDIO_DONE;
1687 if (!*error || !blk_path_error(*error))
1693 if (!atomic_read(&m->nr_valid_paths)) {
1694 spin_lock_irqsave(&m->lock, flags);
1695 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1696 if (__must_push_back(m)) {
1697 r = DM_ENDIO_REQUEUE;
1700 *error = BLK_STS_IOERR;
1702 spin_unlock_irqrestore(&m->lock, flags);
1705 spin_unlock_irqrestore(&m->lock, flags);
1708 multipath_queue_bio(m, clone);
1709 r = DM_ENDIO_INCOMPLETE;
1712 struct path_selector *ps = &pgpath->pg->ps;
1714 if (ps->type->end_io)
1715 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1716 dm_start_time_ns_from_clone(clone));
1723 * Suspend with flush can't complete until all the I/O is processed
1724 * so if the last path fails we must error any remaining I/O.
1725 * - Note that if the freeze_bdev fails while suspending, the
1726 * queue_if_no_path state is lost - userspace should reset it.
1727 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1729 static void multipath_presuspend(struct dm_target *ti)
1731 struct multipath *m = ti->private;
1733 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1734 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1735 queue_if_no_path(m, false, true, __func__);
1738 static void multipath_postsuspend(struct dm_target *ti)
1740 struct multipath *m = ti->private;
1742 mutex_lock(&m->work_mutex);
1743 flush_multipath_work(m);
1744 mutex_unlock(&m->work_mutex);
1748 * Restore the queue_if_no_path setting.
1750 static void multipath_resume(struct dm_target *ti)
1752 struct multipath *m = ti->private;
1753 unsigned long flags;
1755 spin_lock_irqsave(&m->lock, flags);
1756 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1757 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1758 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1761 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1762 dm_table_device_name(m->ti->table), __func__,
1763 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1764 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1766 spin_unlock_irqrestore(&m->lock, flags);
1770 * Info output has the following format:
1771 * num_multipath_feature_args [multipath_feature_args]*
1772 * num_handler_status_args [handler_status_args]*
1773 * num_groups init_group_number
1774 * [A|D|E num_ps_status_args [ps_status_args]*
1775 * num_paths num_selector_args
1776 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1778 * Table output has the following format (identical to the constructor string):
1779 * num_feature_args [features_args]*
1780 * num_handler_args hw_handler [hw_handler_args]*
1781 * num_groups init_group_number
1782 * [priority selector-name num_ps_args [ps_args]*
1783 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1785 static void multipath_status(struct dm_target *ti, status_type_t type,
1786 unsigned status_flags, char *result, unsigned maxlen)
1788 int sz = 0, pg_counter, pgpath_counter;
1789 unsigned long flags;
1790 struct multipath *m = ti->private;
1791 struct priority_group *pg;
1796 spin_lock_irqsave(&m->lock, flags);
1799 if (type == STATUSTYPE_INFO)
1800 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1801 atomic_read(&m->pg_init_count));
1803 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1804 (m->pg_init_retries > 0) * 2 +
1805 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1806 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1807 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1809 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1810 DMEMIT("queue_if_no_path ");
1811 if (m->pg_init_retries)
1812 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1813 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1814 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1815 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1816 DMEMIT("retain_attached_hw_handler ");
1817 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1818 switch(m->queue_mode) {
1819 case DM_TYPE_BIO_BASED:
1820 DMEMIT("queue_mode bio ");
1829 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1832 DMEMIT("1 %s ", m->hw_handler_name);
1834 DMEMIT("%u ", m->nr_priority_groups);
1837 pg_num = m->next_pg->pg_num;
1838 else if (m->current_pg)
1839 pg_num = m->current_pg->pg_num;
1841 pg_num = (m->nr_priority_groups ? 1 : 0);
1843 DMEMIT("%u ", pg_num);
1846 case STATUSTYPE_INFO:
1847 list_for_each_entry(pg, &m->priority_groups, list) {
1849 state = 'D'; /* Disabled */
1850 else if (pg == m->current_pg)
1851 state = 'A'; /* Currently Active */
1853 state = 'E'; /* Enabled */
1855 DMEMIT("%c ", state);
1857 if (pg->ps.type->status)
1858 sz += pg->ps.type->status(&pg->ps, NULL, type,
1864 DMEMIT("%u %u ", pg->nr_pgpaths,
1865 pg->ps.type->info_args);
1867 list_for_each_entry(p, &pg->pgpaths, list) {
1868 DMEMIT("%s %s %u ", p->path.dev->name,
1869 p->is_active ? "A" : "F",
1871 if (pg->ps.type->status)
1872 sz += pg->ps.type->status(&pg->ps,
1873 &p->path, type, result + sz,
1879 case STATUSTYPE_TABLE:
1880 list_for_each_entry(pg, &m->priority_groups, list) {
1881 DMEMIT("%s ", pg->ps.type->name);
1883 if (pg->ps.type->status)
1884 sz += pg->ps.type->status(&pg->ps, NULL, type,
1890 DMEMIT("%u %u ", pg->nr_pgpaths,
1891 pg->ps.type->table_args);
1893 list_for_each_entry(p, &pg->pgpaths, list) {
1894 DMEMIT("%s ", p->path.dev->name);
1895 if (pg->ps.type->status)
1896 sz += pg->ps.type->status(&pg->ps,
1897 &p->path, type, result + sz,
1903 case STATUSTYPE_IMA:
1904 sz = 0; /*reset the result pointer*/
1906 DMEMIT_TARGET_NAME_VERSION(ti->type);
1907 DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1910 list_for_each_entry(pg, &m->priority_groups, list) {
1912 state = 'D'; /* Disabled */
1913 else if (pg == m->current_pg)
1914 state = 'A'; /* Currently Active */
1916 state = 'E'; /* Enabled */
1917 DMEMIT(",pg_state_%d=%c", pg_counter, state);
1918 DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1919 DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1922 list_for_each_entry(p, &pg->pgpaths, list) {
1923 DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1924 pg_counter, pgpath_counter, p->path.dev->name,
1925 pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1926 pg_counter, pgpath_counter, p->fail_count);
1927 if (pg->ps.type->status) {
1928 DMEMIT(",path_selector_status_%d_%d=",
1929 pg_counter, pgpath_counter);
1930 sz += pg->ps.type->status(&pg->ps, &p->path,
1942 spin_unlock_irqrestore(&m->lock, flags);
1945 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1946 char *result, unsigned maxlen)
1950 struct multipath *m = ti->private;
1952 unsigned long flags;
1954 mutex_lock(&m->work_mutex);
1956 if (dm_suspended(ti)) {
1962 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1963 r = queue_if_no_path(m, true, false, __func__);
1964 spin_lock_irqsave(&m->lock, flags);
1965 enable_nopath_timeout(m);
1966 spin_unlock_irqrestore(&m->lock, flags);
1968 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1969 r = queue_if_no_path(m, false, false, __func__);
1970 disable_nopath_timeout(m);
1976 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1980 if (!strcasecmp(argv[0], "disable_group")) {
1981 r = bypass_pg_num(m, argv[1], true);
1983 } else if (!strcasecmp(argv[0], "enable_group")) {
1984 r = bypass_pg_num(m, argv[1], false);
1986 } else if (!strcasecmp(argv[0], "switch_group")) {
1987 r = switch_pg_num(m, argv[1]);
1989 } else if (!strcasecmp(argv[0], "reinstate_path"))
1990 action = reinstate_path;
1991 else if (!strcasecmp(argv[0], "fail_path"))
1994 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1998 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
2000 DMWARN("message: error getting device %s",
2005 r = action_dev(m, dev, action);
2007 dm_put_device(ti, dev);
2010 mutex_unlock(&m->work_mutex);
2014 static int multipath_prepare_ioctl(struct dm_target *ti,
2015 struct block_device **bdev)
2017 struct multipath *m = ti->private;
2018 struct pgpath *pgpath;
2019 unsigned long flags;
2022 pgpath = READ_ONCE(m->current_pgpath);
2023 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2024 pgpath = choose_pgpath(m, 0);
2027 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2028 *bdev = pgpath->path.dev->bdev;
2031 /* pg_init has not started or completed */
2035 /* No path is available */
2037 spin_lock_irqsave(&m->lock, flags);
2038 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2040 spin_unlock_irqrestore(&m->lock, flags);
2043 if (r == -ENOTCONN) {
2044 if (!READ_ONCE(m->current_pg)) {
2045 /* Path status changed, redo selection */
2046 (void) choose_pgpath(m, 0);
2048 spin_lock_irqsave(&m->lock, flags);
2049 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2050 (void) __pg_init_all_paths(m);
2051 spin_unlock_irqrestore(&m->lock, flags);
2052 dm_table_run_md_queue_async(m->ti->table);
2053 process_queued_io_list(m);
2057 * Only pass ioctls through if the device sizes match exactly.
2059 if (!r && ti->len != bdev_nr_sectors((*bdev)))
2064 static int multipath_iterate_devices(struct dm_target *ti,
2065 iterate_devices_callout_fn fn, void *data)
2067 struct multipath *m = ti->private;
2068 struct priority_group *pg;
2072 list_for_each_entry(pg, &m->priority_groups, list) {
2073 list_for_each_entry(p, &pg->pgpaths, list) {
2074 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2084 static int pgpath_busy(struct pgpath *pgpath)
2086 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2088 return blk_lld_busy(q);
2092 * We return "busy", only when we can map I/Os but underlying devices
2093 * are busy (so even if we map I/Os now, the I/Os will wait on
2094 * the underlying queue).
2095 * In other words, if we want to kill I/Os or queue them inside us
2096 * due to map unavailability, we don't return "busy". Otherwise,
2097 * dm core won't give us the I/Os and we can't do what we want.
2099 static int multipath_busy(struct dm_target *ti)
2101 bool busy = false, has_active = false;
2102 struct multipath *m = ti->private;
2103 struct priority_group *pg, *next_pg;
2104 struct pgpath *pgpath;
2106 /* pg_init in progress */
2107 if (atomic_read(&m->pg_init_in_progress))
2110 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2111 if (!atomic_read(&m->nr_valid_paths)) {
2112 unsigned long flags;
2113 spin_lock_irqsave(&m->lock, flags);
2114 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2115 spin_unlock_irqrestore(&m->lock, flags);
2116 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2118 spin_unlock_irqrestore(&m->lock, flags);
2121 /* Guess which priority_group will be used at next mapping time */
2122 pg = READ_ONCE(m->current_pg);
2123 next_pg = READ_ONCE(m->next_pg);
2124 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2129 * We don't know which pg will be used at next mapping time.
2130 * We don't call choose_pgpath() here to avoid to trigger
2131 * pg_init just by busy checking.
2132 * So we don't know whether underlying devices we will be using
2133 * at next mapping time are busy or not. Just try mapping.
2139 * If there is one non-busy active path at least, the path selector
2140 * will be able to select it. So we consider such a pg as not busy.
2143 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2144 if (pgpath->is_active) {
2146 if (!pgpath_busy(pgpath)) {
2155 * No active path in this pg, so this pg won't be used and
2156 * the current_pg will be changed at next mapping time.
2157 * We need to try mapping to determine it.
2165 /*-----------------------------------------------------------------
2167 *---------------------------------------------------------------*/
2168 static struct target_type multipath_target = {
2169 .name = "multipath",
2170 .version = {1, 14, 0},
2171 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2172 DM_TARGET_PASSES_INTEGRITY,
2173 .module = THIS_MODULE,
2174 .ctr = multipath_ctr,
2175 .dtr = multipath_dtr,
2176 .clone_and_map_rq = multipath_clone_and_map,
2177 .release_clone_rq = multipath_release_clone,
2178 .rq_end_io = multipath_end_io,
2179 .map = multipath_map_bio,
2180 .end_io = multipath_end_io_bio,
2181 .presuspend = multipath_presuspend,
2182 .postsuspend = multipath_postsuspend,
2183 .resume = multipath_resume,
2184 .status = multipath_status,
2185 .message = multipath_message,
2186 .prepare_ioctl = multipath_prepare_ioctl,
2187 .iterate_devices = multipath_iterate_devices,
2188 .busy = multipath_busy,
2191 static int __init dm_multipath_init(void)
2195 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2197 DMERR("failed to create workqueue kmpathd");
2199 goto bad_alloc_kmultipathd;
2203 * A separate workqueue is used to handle the device handlers
2204 * to avoid overloading existing workqueue. Overloading the
2205 * old workqueue would also create a bottleneck in the
2206 * path of the storage hardware device activation.
2208 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2210 if (!kmpath_handlerd) {
2211 DMERR("failed to create workqueue kmpath_handlerd");
2213 goto bad_alloc_kmpath_handlerd;
2216 r = dm_register_target(&multipath_target);
2218 DMERR("request-based register failed %d", r);
2220 goto bad_register_target;
2225 bad_register_target:
2226 destroy_workqueue(kmpath_handlerd);
2227 bad_alloc_kmpath_handlerd:
2228 destroy_workqueue(kmultipathd);
2229 bad_alloc_kmultipathd:
2233 static void __exit dm_multipath_exit(void)
2235 destroy_workqueue(kmpath_handlerd);
2236 destroy_workqueue(kmultipathd);
2238 dm_unregister_target(&multipath_target);
2241 module_init(dm_multipath_init);
2242 module_exit(dm_multipath_exit);
2244 module_param_named(queue_if_no_path_timeout_secs,
2245 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2246 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2248 MODULE_DESCRIPTION(DM_NAME " multipath target");
2249 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2250 MODULE_LICENSE("GPL");