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 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
471 DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
472 dm_device_name(md), \
473 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
474 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
475 dm_noflush_suspending((m)->ti)); \
479 * Check whether bios must be queued in the device-mapper core rather
480 * than here in the target.
482 static bool __must_push_back(struct multipath *m)
484 return dm_noflush_suspending(m->ti);
487 static bool must_push_back_rq(struct multipath *m)
492 spin_lock_irqsave(&m->lock, flags);
493 ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
494 spin_unlock_irqrestore(&m->lock, flags);
500 * Map cloned requests (request-based multipath)
502 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
503 union map_info *map_context,
504 struct request **__clone)
506 struct multipath *m = ti->private;
507 size_t nr_bytes = blk_rq_bytes(rq);
508 struct pgpath *pgpath;
509 struct block_device *bdev;
510 struct dm_mpath_io *mpio = get_mpio(map_context);
511 struct request_queue *q;
512 struct request *clone;
514 /* Do we need to select a new pgpath? */
515 pgpath = READ_ONCE(m->current_pgpath);
516 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
517 pgpath = choose_pgpath(m, nr_bytes);
520 if (must_push_back_rq(m))
521 return DM_MAPIO_DELAY_REQUEUE;
522 dm_report_EIO(m); /* Failed */
523 return DM_MAPIO_KILL;
524 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
525 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
526 pg_init_all_paths(m);
527 return DM_MAPIO_DELAY_REQUEUE;
530 mpio->pgpath = pgpath;
531 mpio->nr_bytes = nr_bytes;
533 bdev = pgpath->path.dev->bdev;
534 q = bdev_get_queue(bdev);
535 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
538 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
539 if (blk_queue_dying(q)) {
540 atomic_inc(&m->pg_init_in_progress);
541 activate_or_offline_path(pgpath);
542 return DM_MAPIO_DELAY_REQUEUE;
546 * blk-mq's SCHED_RESTART can cover this requeue, so we
547 * needn't deal with it by DELAY_REQUEUE. More importantly,
548 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
549 * get the queue busy feedback (via BLK_STS_RESOURCE),
550 * otherwise I/O merging can suffer.
552 return DM_MAPIO_REQUEUE;
554 clone->bio = clone->biotail = NULL;
555 clone->rq_disk = bdev->bd_disk;
556 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
559 if (pgpath->pg->ps.type->start_io)
560 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
563 return DM_MAPIO_REMAPPED;
566 static void multipath_release_clone(struct request *clone,
567 union map_info *map_context)
569 if (unlikely(map_context)) {
571 * non-NULL map_context means caller is still map
572 * method; must undo multipath_clone_and_map()
574 struct dm_mpath_io *mpio = get_mpio(map_context);
575 struct pgpath *pgpath = mpio->pgpath;
577 if (pgpath && pgpath->pg->ps.type->end_io)
578 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
581 clone->io_start_time_ns);
584 blk_put_request(clone);
588 * Map cloned bios (bio-based multipath)
591 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
593 /* Queue for the daemon to resubmit */
594 bio_list_add(&m->queued_bios, bio);
595 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
596 queue_work(kmultipathd, &m->process_queued_bios);
599 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
603 spin_lock_irqsave(&m->lock, flags);
604 __multipath_queue_bio(m, bio);
605 spin_unlock_irqrestore(&m->lock, flags);
608 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
610 struct pgpath *pgpath;
613 /* Do we need to select a new pgpath? */
614 pgpath = READ_ONCE(m->current_pgpath);
615 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
616 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
619 spin_lock_irqsave(&m->lock, flags);
620 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
621 __multipath_queue_bio(m, bio);
622 pgpath = ERR_PTR(-EAGAIN);
624 spin_unlock_irqrestore(&m->lock, flags);
626 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
627 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
628 multipath_queue_bio(m, bio);
629 pg_init_all_paths(m);
630 return ERR_PTR(-EAGAIN);
636 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
637 struct dm_mpath_io *mpio)
639 struct pgpath *pgpath = __map_bio(m, bio);
642 return DM_MAPIO_SUBMITTED;
645 if (__must_push_back(m))
646 return DM_MAPIO_REQUEUE;
648 return DM_MAPIO_KILL;
651 mpio->pgpath = pgpath;
654 bio_set_dev(bio, pgpath->path.dev->bdev);
655 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
657 if (pgpath->pg->ps.type->start_io)
658 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
661 return DM_MAPIO_REMAPPED;
664 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
666 struct multipath *m = ti->private;
667 struct dm_mpath_io *mpio = NULL;
669 multipath_init_per_bio_data(bio, &mpio);
670 return __multipath_map_bio(m, bio, mpio);
673 static void process_queued_io_list(struct multipath *m)
675 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
676 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
677 else if (m->queue_mode == DM_TYPE_BIO_BASED)
678 queue_work(kmultipathd, &m->process_queued_bios);
681 static void process_queued_bios(struct work_struct *work)
686 struct bio_list bios;
687 struct blk_plug plug;
688 struct multipath *m =
689 container_of(work, struct multipath, process_queued_bios);
691 bio_list_init(&bios);
693 spin_lock_irqsave(&m->lock, flags);
695 if (bio_list_empty(&m->queued_bios)) {
696 spin_unlock_irqrestore(&m->lock, flags);
700 bio_list_merge(&bios, &m->queued_bios);
701 bio_list_init(&m->queued_bios);
703 spin_unlock_irqrestore(&m->lock, flags);
705 blk_start_plug(&plug);
706 while ((bio = bio_list_pop(&bios))) {
707 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
708 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
709 r = __multipath_map_bio(m, bio, mpio);
712 bio->bi_status = BLK_STS_IOERR;
715 case DM_MAPIO_REQUEUE:
716 bio->bi_status = BLK_STS_DM_REQUEUE;
719 case DM_MAPIO_REMAPPED:
720 submit_bio_noacct(bio);
722 case DM_MAPIO_SUBMITTED:
725 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
728 blk_finish_plug(&plug);
732 * If we run out of usable paths, should we queue I/O or error it?
734 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
735 bool save_old_value, const char *caller)
738 bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
739 const char *dm_dev_name = dm_device_name(dm_table_get_md(m->ti->table));
741 DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
742 dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
744 spin_lock_irqsave(&m->lock, flags);
746 queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
747 saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
749 if (save_old_value) {
750 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
751 DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
754 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
755 } else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
756 /* due to "fail_if_no_path" message, need to honor it. */
757 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
759 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
761 DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
762 dm_dev_name, __func__,
763 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
764 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
765 dm_noflush_suspending(m->ti));
767 spin_unlock_irqrestore(&m->lock, flags);
769 if (!queue_if_no_path) {
770 dm_table_run_md_queue_async(m->ti->table);
771 process_queued_io_list(m);
778 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
779 * process any queued I/O.
781 static void queue_if_no_path_timeout_work(struct timer_list *t)
783 struct multipath *m = from_timer(m, t, nopath_timer);
784 struct mapped_device *md = dm_table_get_md(m->ti->table);
786 DMWARN("queue_if_no_path timeout on %s, failing queued IO", dm_device_name(md));
787 queue_if_no_path(m, false, false, __func__);
791 * Enable the queue_if_no_path timeout if necessary.
792 * Called with m->lock held.
794 static void enable_nopath_timeout(struct multipath *m)
796 unsigned long queue_if_no_path_timeout =
797 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
799 lockdep_assert_held(&m->lock);
801 if (queue_if_no_path_timeout > 0 &&
802 atomic_read(&m->nr_valid_paths) == 0 &&
803 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
804 mod_timer(&m->nopath_timer,
805 jiffies + queue_if_no_path_timeout);
809 static void disable_nopath_timeout(struct multipath *m)
811 del_timer_sync(&m->nopath_timer);
815 * An event is triggered whenever a path is taken out of use.
816 * Includes path failure and PG bypass.
818 static void trigger_event(struct work_struct *work)
820 struct multipath *m =
821 container_of(work, struct multipath, trigger_event);
823 dm_table_event(m->ti->table);
826 /*-----------------------------------------------------------------
827 * Constructor/argument parsing:
828 * <#multipath feature args> [<arg>]*
829 * <#hw_handler args> [hw_handler [<arg>]*]
831 * <initial priority group>
832 * [<selector> <#selector args> [<arg>]*
833 * <#paths> <#per-path selector args>
834 * [<path> [<arg>]* ]+ ]+
835 *---------------------------------------------------------------*/
836 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
837 struct dm_target *ti)
840 struct path_selector_type *pst;
843 static const struct dm_arg _args[] = {
844 {0, 1024, "invalid number of path selector args"},
847 pst = dm_get_path_selector(dm_shift_arg(as));
849 ti->error = "unknown path selector type";
853 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
855 dm_put_path_selector(pst);
859 r = pst->create(&pg->ps, ps_argc, as->argv);
861 dm_put_path_selector(pst);
862 ti->error = "path selector constructor failed";
867 dm_consume_args(as, ps_argc);
872 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
873 const char **attached_handler_name, char **error)
875 struct request_queue *q = bdev_get_queue(bdev);
878 if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
880 if (*attached_handler_name) {
882 * Clear any hw_handler_params associated with a
883 * handler that isn't already attached.
885 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
886 kfree(m->hw_handler_params);
887 m->hw_handler_params = NULL;
891 * Reset hw_handler_name to match the attached handler
893 * NB. This modifies the table line to show the actual
894 * handler instead of the original table passed in.
896 kfree(m->hw_handler_name);
897 m->hw_handler_name = *attached_handler_name;
898 *attached_handler_name = NULL;
902 if (m->hw_handler_name) {
903 r = scsi_dh_attach(q, m->hw_handler_name);
905 char b[BDEVNAME_SIZE];
907 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
912 *error = "error attaching hardware handler";
916 if (m->hw_handler_params) {
917 r = scsi_dh_set_params(q, m->hw_handler_params);
919 *error = "unable to set hardware handler parameters";
928 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
929 struct dm_target *ti)
933 struct multipath *m = ti->private;
934 struct request_queue *q;
935 const char *attached_handler_name = NULL;
937 /* we need at least a path arg */
939 ti->error = "no device given";
940 return ERR_PTR(-EINVAL);
945 return ERR_PTR(-ENOMEM);
947 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
950 ti->error = "error getting device";
954 q = bdev_get_queue(p->path.dev->bdev);
955 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
956 if (attached_handler_name || m->hw_handler_name) {
957 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
958 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
959 kfree(attached_handler_name);
961 dm_put_device(ti, p->path.dev);
966 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
968 dm_put_device(ti, p->path.dev);
978 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
981 static const struct dm_arg _args[] = {
982 {1, 1024, "invalid number of paths"},
983 {0, 1024, "invalid number of selector args"}
987 unsigned i, nr_selector_args, nr_args;
988 struct priority_group *pg;
989 struct dm_target *ti = m->ti;
993 ti->error = "not enough priority group arguments";
994 return ERR_PTR(-EINVAL);
997 pg = alloc_priority_group();
999 ti->error = "couldn't allocate priority group";
1000 return ERR_PTR(-ENOMEM);
1004 r = parse_path_selector(as, pg, ti);
1011 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1015 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1019 nr_args = 1 + nr_selector_args;
1020 for (i = 0; i < pg->nr_pgpaths; i++) {
1021 struct pgpath *pgpath;
1022 struct dm_arg_set path_args;
1024 if (as->argc < nr_args) {
1025 ti->error = "not enough path parameters";
1030 path_args.argc = nr_args;
1031 path_args.argv = as->argv;
1033 pgpath = parse_path(&path_args, &pg->ps, ti);
1034 if (IS_ERR(pgpath)) {
1035 r = PTR_ERR(pgpath);
1040 list_add_tail(&pgpath->list, &pg->pgpaths);
1041 dm_consume_args(as, nr_args);
1047 free_priority_group(pg, ti);
1051 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1055 struct dm_target *ti = m->ti;
1057 static const struct dm_arg _args[] = {
1058 {0, 1024, "invalid number of hardware handler args"},
1061 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1067 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1068 dm_consume_args(as, hw_argc);
1069 DMERR("bio-based multipath doesn't allow hardware handler args");
1073 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1074 if (!m->hw_handler_name)
1081 for (i = 0; i <= hw_argc - 2; i++)
1082 len += strlen(as->argv[i]) + 1;
1083 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1085 ti->error = "memory allocation failed";
1089 j = sprintf(p, "%d", hw_argc - 1);
1090 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1091 j = sprintf(p, "%s", as->argv[i]);
1093 dm_consume_args(as, hw_argc - 1);
1097 kfree(m->hw_handler_name);
1098 m->hw_handler_name = NULL;
1102 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1106 struct dm_target *ti = m->ti;
1107 const char *arg_name;
1109 static const struct dm_arg _args[] = {
1110 {0, 8, "invalid number of feature args"},
1111 {1, 50, "pg_init_retries must be between 1 and 50"},
1112 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1115 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1123 arg_name = dm_shift_arg(as);
1126 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1127 r = queue_if_no_path(m, true, false, __func__);
1131 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1132 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1136 if (!strcasecmp(arg_name, "pg_init_retries") &&
1138 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1143 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1145 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1150 if (!strcasecmp(arg_name, "queue_mode") &&
1152 const char *queue_mode_name = dm_shift_arg(as);
1154 if (!strcasecmp(queue_mode_name, "bio"))
1155 m->queue_mode = DM_TYPE_BIO_BASED;
1156 else if (!strcasecmp(queue_mode_name, "rq") ||
1157 !strcasecmp(queue_mode_name, "mq"))
1158 m->queue_mode = DM_TYPE_REQUEST_BASED;
1160 ti->error = "Unknown 'queue_mode' requested";
1167 ti->error = "Unrecognised multipath feature request";
1169 } while (argc && !r);
1174 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1176 /* target arguments */
1177 static const struct dm_arg _args[] = {
1178 {0, 1024, "invalid number of priority groups"},
1179 {0, 1024, "invalid initial priority group number"},
1183 struct multipath *m;
1184 struct dm_arg_set as;
1185 unsigned pg_count = 0;
1186 unsigned next_pg_num;
1187 unsigned long flags;
1192 m = alloc_multipath(ti);
1194 ti->error = "can't allocate multipath";
1198 r = parse_features(&as, m);
1202 r = alloc_multipath_stage2(ti, m);
1206 r = parse_hw_handler(&as, m);
1210 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1214 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1218 if ((!m->nr_priority_groups && next_pg_num) ||
1219 (m->nr_priority_groups && !next_pg_num)) {
1220 ti->error = "invalid initial priority group";
1225 /* parse the priority groups */
1227 struct priority_group *pg;
1228 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1230 pg = parse_priority_group(&as, m);
1236 nr_valid_paths += pg->nr_pgpaths;
1237 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1239 list_add_tail(&pg->list, &m->priority_groups);
1241 pg->pg_num = pg_count;
1246 if (pg_count != m->nr_priority_groups) {
1247 ti->error = "priority group count mismatch";
1252 spin_lock_irqsave(&m->lock, flags);
1253 enable_nopath_timeout(m);
1254 spin_unlock_irqrestore(&m->lock, flags);
1256 ti->num_flush_bios = 1;
1257 ti->num_discard_bios = 1;
1258 ti->num_write_same_bios = 1;
1259 ti->num_write_zeroes_bios = 1;
1260 if (m->queue_mode == DM_TYPE_BIO_BASED)
1261 ti->per_io_data_size = multipath_per_bio_data_size();
1263 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1272 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1277 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1279 if (!atomic_read(&m->pg_init_in_progress))
1284 finish_wait(&m->pg_init_wait, &wait);
1287 static void flush_multipath_work(struct multipath *m)
1289 if (m->hw_handler_name) {
1290 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1291 smp_mb__after_atomic();
1293 if (atomic_read(&m->pg_init_in_progress))
1294 flush_workqueue(kmpath_handlerd);
1295 multipath_wait_for_pg_init_completion(m);
1297 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1298 smp_mb__after_atomic();
1301 if (m->queue_mode == DM_TYPE_BIO_BASED)
1302 flush_work(&m->process_queued_bios);
1303 flush_work(&m->trigger_event);
1306 static void multipath_dtr(struct dm_target *ti)
1308 struct multipath *m = ti->private;
1310 disable_nopath_timeout(m);
1311 flush_multipath_work(m);
1316 * Take a path out of use.
1318 static int fail_path(struct pgpath *pgpath)
1320 unsigned long flags;
1321 struct multipath *m = pgpath->pg->m;
1323 spin_lock_irqsave(&m->lock, flags);
1325 if (!pgpath->is_active)
1328 DMWARN("%s: Failing path %s.",
1329 dm_device_name(dm_table_get_md(m->ti->table)),
1330 pgpath->path.dev->name);
1332 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1333 pgpath->is_active = false;
1334 pgpath->fail_count++;
1336 atomic_dec(&m->nr_valid_paths);
1338 if (pgpath == m->current_pgpath)
1339 m->current_pgpath = NULL;
1341 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1342 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1344 schedule_work(&m->trigger_event);
1346 enable_nopath_timeout(m);
1349 spin_unlock_irqrestore(&m->lock, flags);
1355 * Reinstate a previously-failed path
1357 static int reinstate_path(struct pgpath *pgpath)
1359 int r = 0, run_queue = 0;
1360 unsigned long flags;
1361 struct multipath *m = pgpath->pg->m;
1362 unsigned nr_valid_paths;
1364 spin_lock_irqsave(&m->lock, flags);
1366 if (pgpath->is_active)
1369 DMWARN("%s: Reinstating path %s.",
1370 dm_device_name(dm_table_get_md(m->ti->table)),
1371 pgpath->path.dev->name);
1373 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1377 pgpath->is_active = true;
1379 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1380 if (nr_valid_paths == 1) {
1381 m->current_pgpath = NULL;
1383 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1384 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1385 atomic_inc(&m->pg_init_in_progress);
1388 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1389 pgpath->path.dev->name, nr_valid_paths);
1391 schedule_work(&m->trigger_event);
1394 spin_unlock_irqrestore(&m->lock, flags);
1396 dm_table_run_md_queue_async(m->ti->table);
1397 process_queued_io_list(m);
1400 if (pgpath->is_active)
1401 disable_nopath_timeout(m);
1407 * Fail or reinstate all paths that match the provided struct dm_dev.
1409 static int action_dev(struct multipath *m, struct dm_dev *dev,
1413 struct pgpath *pgpath;
1414 struct priority_group *pg;
1416 list_for_each_entry(pg, &m->priority_groups, list) {
1417 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1418 if (pgpath->path.dev == dev)
1427 * Temporarily try to avoid having to use the specified PG
1429 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1432 unsigned long flags;
1434 spin_lock_irqsave(&m->lock, flags);
1436 pg->bypassed = bypassed;
1437 m->current_pgpath = NULL;
1438 m->current_pg = NULL;
1440 spin_unlock_irqrestore(&m->lock, flags);
1442 schedule_work(&m->trigger_event);
1446 * Switch to using the specified PG from the next I/O that gets mapped
1448 static int switch_pg_num(struct multipath *m, const char *pgstr)
1450 struct priority_group *pg;
1452 unsigned long flags;
1455 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1456 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1457 DMWARN("invalid PG number supplied to switch_pg_num");
1461 spin_lock_irqsave(&m->lock, flags);
1462 list_for_each_entry(pg, &m->priority_groups, list) {
1463 pg->bypassed = false;
1467 m->current_pgpath = NULL;
1468 m->current_pg = NULL;
1471 spin_unlock_irqrestore(&m->lock, flags);
1473 schedule_work(&m->trigger_event);
1478 * Set/clear bypassed status of a PG.
1479 * PGs are numbered upwards from 1 in the order they were declared.
1481 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1483 struct priority_group *pg;
1487 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1488 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1489 DMWARN("invalid PG number supplied to bypass_pg");
1493 list_for_each_entry(pg, &m->priority_groups, list) {
1498 bypass_pg(m, pg, bypassed);
1503 * Should we retry pg_init immediately?
1505 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1507 unsigned long flags;
1508 bool limit_reached = false;
1510 spin_lock_irqsave(&m->lock, flags);
1512 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1513 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1514 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1516 limit_reached = true;
1518 spin_unlock_irqrestore(&m->lock, flags);
1520 return limit_reached;
1523 static void pg_init_done(void *data, int errors)
1525 struct pgpath *pgpath = data;
1526 struct priority_group *pg = pgpath->pg;
1527 struct multipath *m = pg->m;
1528 unsigned long flags;
1529 bool delay_retry = false;
1531 /* device or driver problems */
1536 if (!m->hw_handler_name) {
1540 DMERR("Could not failover the device: Handler scsi_dh_%s "
1541 "Error %d.", m->hw_handler_name, errors);
1543 * Fail path for now, so we do not ping pong
1547 case SCSI_DH_DEV_TEMP_BUSY:
1549 * Probably doing something like FW upgrade on the
1550 * controller so try the other pg.
1552 bypass_pg(m, pg, true);
1555 /* Wait before retrying. */
1558 case SCSI_DH_IMM_RETRY:
1559 case SCSI_DH_RES_TEMP_UNAVAIL:
1560 if (pg_init_limit_reached(m, pgpath))
1564 case SCSI_DH_DEV_OFFLINED:
1567 * We probably do not want to fail the path for a device
1568 * error, but this is what the old dm did. In future
1569 * patches we can do more advanced handling.
1574 spin_lock_irqsave(&m->lock, flags);
1576 if (pgpath == m->current_pgpath) {
1577 DMERR("Could not failover device. Error %d.", errors);
1578 m->current_pgpath = NULL;
1579 m->current_pg = NULL;
1581 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1582 pg->bypassed = false;
1584 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1585 /* Activations of other paths are still on going */
1588 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1590 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1592 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1594 if (__pg_init_all_paths(m))
1597 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1599 process_queued_io_list(m);
1602 * Wake up any thread waiting to suspend.
1604 wake_up(&m->pg_init_wait);
1607 spin_unlock_irqrestore(&m->lock, flags);
1610 static void activate_or_offline_path(struct pgpath *pgpath)
1612 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1614 if (pgpath->is_active && !blk_queue_dying(q))
1615 scsi_dh_activate(q, pg_init_done, pgpath);
1617 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1620 static void activate_path_work(struct work_struct *work)
1622 struct pgpath *pgpath =
1623 container_of(work, struct pgpath, activate_path.work);
1625 activate_or_offline_path(pgpath);
1628 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1629 blk_status_t error, union map_info *map_context)
1631 struct dm_mpath_io *mpio = get_mpio(map_context);
1632 struct pgpath *pgpath = mpio->pgpath;
1633 int r = DM_ENDIO_DONE;
1636 * We don't queue any clone request inside the multipath target
1637 * during end I/O handling, since those clone requests don't have
1638 * bio clones. If we queue them inside the multipath target,
1639 * we need to make bio clones, that requires memory allocation.
1640 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1641 * don't have bio clones.)
1642 * Instead of queueing the clone request here, we queue the original
1643 * request into dm core, which will remake a clone request and
1644 * clone bios for it and resubmit it later.
1646 if (error && blk_path_error(error)) {
1647 struct multipath *m = ti->private;
1649 if (error == BLK_STS_RESOURCE)
1650 r = DM_ENDIO_DELAY_REQUEUE;
1652 r = DM_ENDIO_REQUEUE;
1657 if (!atomic_read(&m->nr_valid_paths) &&
1658 !must_push_back_rq(m)) {
1659 if (error == BLK_STS_IOERR)
1661 /* complete with the original error */
1667 struct path_selector *ps = &pgpath->pg->ps;
1669 if (ps->type->end_io)
1670 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1671 clone->io_start_time_ns);
1677 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1678 blk_status_t *error)
1680 struct multipath *m = ti->private;
1681 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1682 struct pgpath *pgpath = mpio->pgpath;
1683 unsigned long flags;
1684 int r = DM_ENDIO_DONE;
1686 if (!*error || !blk_path_error(*error))
1692 if (!atomic_read(&m->nr_valid_paths)) {
1693 spin_lock_irqsave(&m->lock, flags);
1694 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1695 if (__must_push_back(m)) {
1696 r = DM_ENDIO_REQUEUE;
1699 *error = BLK_STS_IOERR;
1701 spin_unlock_irqrestore(&m->lock, flags);
1704 spin_unlock_irqrestore(&m->lock, flags);
1707 multipath_queue_bio(m, clone);
1708 r = DM_ENDIO_INCOMPLETE;
1711 struct path_selector *ps = &pgpath->pg->ps;
1713 if (ps->type->end_io)
1714 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1715 dm_start_time_ns_from_clone(clone));
1722 * Suspend with flush can't complete until all the I/O is processed
1723 * so if the last path fails we must error any remaining I/O.
1724 * - Note that if the freeze_bdev fails while suspending, the
1725 * queue_if_no_path state is lost - userspace should reset it.
1726 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1728 static void multipath_presuspend(struct dm_target *ti)
1730 struct multipath *m = ti->private;
1732 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1733 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1734 queue_if_no_path(m, false, true, __func__);
1737 static void multipath_postsuspend(struct dm_target *ti)
1739 struct multipath *m = ti->private;
1741 mutex_lock(&m->work_mutex);
1742 flush_multipath_work(m);
1743 mutex_unlock(&m->work_mutex);
1747 * Restore the queue_if_no_path setting.
1749 static void multipath_resume(struct dm_target *ti)
1751 struct multipath *m = ti->private;
1752 unsigned long flags;
1754 spin_lock_irqsave(&m->lock, flags);
1755 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1756 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1757 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1760 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1761 dm_device_name(dm_table_get_md(m->ti->table)), __func__,
1762 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1763 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1765 spin_unlock_irqrestore(&m->lock, flags);
1769 * Info output has the following format:
1770 * num_multipath_feature_args [multipath_feature_args]*
1771 * num_handler_status_args [handler_status_args]*
1772 * num_groups init_group_number
1773 * [A|D|E num_ps_status_args [ps_status_args]*
1774 * num_paths num_selector_args
1775 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1777 * Table output has the following format (identical to the constructor string):
1778 * num_feature_args [features_args]*
1779 * num_handler_args hw_handler [hw_handler_args]*
1780 * num_groups init_group_number
1781 * [priority selector-name num_ps_args [ps_args]*
1782 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1784 static void multipath_status(struct dm_target *ti, status_type_t type,
1785 unsigned status_flags, char *result, unsigned maxlen)
1788 unsigned long flags;
1789 struct multipath *m = ti->private;
1790 struct priority_group *pg;
1795 spin_lock_irqsave(&m->lock, flags);
1798 if (type == STATUSTYPE_INFO)
1799 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1800 atomic_read(&m->pg_init_count));
1802 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1803 (m->pg_init_retries > 0) * 2 +
1804 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1805 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1806 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1808 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1809 DMEMIT("queue_if_no_path ");
1810 if (m->pg_init_retries)
1811 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1812 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1813 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1814 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1815 DMEMIT("retain_attached_hw_handler ");
1816 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1817 switch(m->queue_mode) {
1818 case DM_TYPE_BIO_BASED:
1819 DMEMIT("queue_mode bio ");
1828 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1831 DMEMIT("1 %s ", m->hw_handler_name);
1833 DMEMIT("%u ", m->nr_priority_groups);
1836 pg_num = m->next_pg->pg_num;
1837 else if (m->current_pg)
1838 pg_num = m->current_pg->pg_num;
1840 pg_num = (m->nr_priority_groups ? 1 : 0);
1842 DMEMIT("%u ", pg_num);
1845 case STATUSTYPE_INFO:
1846 list_for_each_entry(pg, &m->priority_groups, list) {
1848 state = 'D'; /* Disabled */
1849 else if (pg == m->current_pg)
1850 state = 'A'; /* Currently Active */
1852 state = 'E'; /* Enabled */
1854 DMEMIT("%c ", state);
1856 if (pg->ps.type->status)
1857 sz += pg->ps.type->status(&pg->ps, NULL, type,
1863 DMEMIT("%u %u ", pg->nr_pgpaths,
1864 pg->ps.type->info_args);
1866 list_for_each_entry(p, &pg->pgpaths, list) {
1867 DMEMIT("%s %s %u ", p->path.dev->name,
1868 p->is_active ? "A" : "F",
1870 if (pg->ps.type->status)
1871 sz += pg->ps.type->status(&pg->ps,
1872 &p->path, type, result + sz,
1878 case STATUSTYPE_TABLE:
1879 list_for_each_entry(pg, &m->priority_groups, list) {
1880 DMEMIT("%s ", pg->ps.type->name);
1882 if (pg->ps.type->status)
1883 sz += pg->ps.type->status(&pg->ps, NULL, type,
1889 DMEMIT("%u %u ", pg->nr_pgpaths,
1890 pg->ps.type->table_args);
1892 list_for_each_entry(p, &pg->pgpaths, list) {
1893 DMEMIT("%s ", p->path.dev->name);
1894 if (pg->ps.type->status)
1895 sz += pg->ps.type->status(&pg->ps,
1896 &p->path, type, result + sz,
1903 spin_unlock_irqrestore(&m->lock, flags);
1906 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1907 char *result, unsigned maxlen)
1911 struct multipath *m = ti->private;
1913 unsigned long flags;
1915 mutex_lock(&m->work_mutex);
1917 if (dm_suspended(ti)) {
1923 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1924 r = queue_if_no_path(m, true, false, __func__);
1925 spin_lock_irqsave(&m->lock, flags);
1926 enable_nopath_timeout(m);
1927 spin_unlock_irqrestore(&m->lock, flags);
1929 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1930 r = queue_if_no_path(m, false, false, __func__);
1931 disable_nopath_timeout(m);
1937 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1941 if (!strcasecmp(argv[0], "disable_group")) {
1942 r = bypass_pg_num(m, argv[1], true);
1944 } else if (!strcasecmp(argv[0], "enable_group")) {
1945 r = bypass_pg_num(m, argv[1], false);
1947 } else if (!strcasecmp(argv[0], "switch_group")) {
1948 r = switch_pg_num(m, argv[1]);
1950 } else if (!strcasecmp(argv[0], "reinstate_path"))
1951 action = reinstate_path;
1952 else if (!strcasecmp(argv[0], "fail_path"))
1955 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1959 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1961 DMWARN("message: error getting device %s",
1966 r = action_dev(m, dev, action);
1968 dm_put_device(ti, dev);
1971 mutex_unlock(&m->work_mutex);
1975 static int multipath_prepare_ioctl(struct dm_target *ti,
1976 struct block_device **bdev)
1978 struct multipath *m = ti->private;
1979 struct pgpath *pgpath;
1980 unsigned long flags;
1983 pgpath = READ_ONCE(m->current_pgpath);
1984 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
1985 pgpath = choose_pgpath(m, 0);
1988 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
1989 *bdev = pgpath->path.dev->bdev;
1992 /* pg_init has not started or completed */
1996 /* No path is available */
1998 spin_lock_irqsave(&m->lock, flags);
1999 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2001 spin_unlock_irqrestore(&m->lock, flags);
2004 if (r == -ENOTCONN) {
2005 if (!READ_ONCE(m->current_pg)) {
2006 /* Path status changed, redo selection */
2007 (void) choose_pgpath(m, 0);
2009 spin_lock_irqsave(&m->lock, flags);
2010 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2011 (void) __pg_init_all_paths(m);
2012 spin_unlock_irqrestore(&m->lock, flags);
2013 dm_table_run_md_queue_async(m->ti->table);
2014 process_queued_io_list(m);
2018 * Only pass ioctls through if the device sizes match exactly.
2020 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
2025 static int multipath_iterate_devices(struct dm_target *ti,
2026 iterate_devices_callout_fn fn, void *data)
2028 struct multipath *m = ti->private;
2029 struct priority_group *pg;
2033 list_for_each_entry(pg, &m->priority_groups, list) {
2034 list_for_each_entry(p, &pg->pgpaths, list) {
2035 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2045 static int pgpath_busy(struct pgpath *pgpath)
2047 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2049 return blk_lld_busy(q);
2053 * We return "busy", only when we can map I/Os but underlying devices
2054 * are busy (so even if we map I/Os now, the I/Os will wait on
2055 * the underlying queue).
2056 * In other words, if we want to kill I/Os or queue them inside us
2057 * due to map unavailability, we don't return "busy". Otherwise,
2058 * dm core won't give us the I/Os and we can't do what we want.
2060 static int multipath_busy(struct dm_target *ti)
2062 bool busy = false, has_active = false;
2063 struct multipath *m = ti->private;
2064 struct priority_group *pg, *next_pg;
2065 struct pgpath *pgpath;
2067 /* pg_init in progress */
2068 if (atomic_read(&m->pg_init_in_progress))
2071 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2072 if (!atomic_read(&m->nr_valid_paths)) {
2073 unsigned long flags;
2074 spin_lock_irqsave(&m->lock, flags);
2075 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2076 spin_unlock_irqrestore(&m->lock, flags);
2077 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2079 spin_unlock_irqrestore(&m->lock, flags);
2082 /* Guess which priority_group will be used at next mapping time */
2083 pg = READ_ONCE(m->current_pg);
2084 next_pg = READ_ONCE(m->next_pg);
2085 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2090 * We don't know which pg will be used at next mapping time.
2091 * We don't call choose_pgpath() here to avoid to trigger
2092 * pg_init just by busy checking.
2093 * So we don't know whether underlying devices we will be using
2094 * at next mapping time are busy or not. Just try mapping.
2100 * If there is one non-busy active path at least, the path selector
2101 * will be able to select it. So we consider such a pg as not busy.
2104 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2105 if (pgpath->is_active) {
2107 if (!pgpath_busy(pgpath)) {
2116 * No active path in this pg, so this pg won't be used and
2117 * the current_pg will be changed at next mapping time.
2118 * We need to try mapping to determine it.
2126 /*-----------------------------------------------------------------
2128 *---------------------------------------------------------------*/
2129 static struct target_type multipath_target = {
2130 .name = "multipath",
2131 .version = {1, 14, 0},
2132 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2133 DM_TARGET_PASSES_INTEGRITY,
2134 .module = THIS_MODULE,
2135 .ctr = multipath_ctr,
2136 .dtr = multipath_dtr,
2137 .clone_and_map_rq = multipath_clone_and_map,
2138 .release_clone_rq = multipath_release_clone,
2139 .rq_end_io = multipath_end_io,
2140 .map = multipath_map_bio,
2141 .end_io = multipath_end_io_bio,
2142 .presuspend = multipath_presuspend,
2143 .postsuspend = multipath_postsuspend,
2144 .resume = multipath_resume,
2145 .status = multipath_status,
2146 .message = multipath_message,
2147 .prepare_ioctl = multipath_prepare_ioctl,
2148 .iterate_devices = multipath_iterate_devices,
2149 .busy = multipath_busy,
2152 static int __init dm_multipath_init(void)
2156 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2158 DMERR("failed to create workqueue kmpathd");
2160 goto bad_alloc_kmultipathd;
2164 * A separate workqueue is used to handle the device handlers
2165 * to avoid overloading existing workqueue. Overloading the
2166 * old workqueue would also create a bottleneck in the
2167 * path of the storage hardware device activation.
2169 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2171 if (!kmpath_handlerd) {
2172 DMERR("failed to create workqueue kmpath_handlerd");
2174 goto bad_alloc_kmpath_handlerd;
2177 r = dm_register_target(&multipath_target);
2179 DMERR("request-based register failed %d", r);
2181 goto bad_register_target;
2186 bad_register_target:
2187 destroy_workqueue(kmpath_handlerd);
2188 bad_alloc_kmpath_handlerd:
2189 destroy_workqueue(kmultipathd);
2190 bad_alloc_kmultipathd:
2194 static void __exit dm_multipath_exit(void)
2196 destroy_workqueue(kmpath_handlerd);
2197 destroy_workqueue(kmultipathd);
2199 dm_unregister_target(&multipath_target);
2202 module_init(dm_multipath_init);
2203 module_exit(dm_multipath_exit);
2205 module_param_named(queue_if_no_path_timeout_secs,
2206 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2207 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2209 MODULE_DESCRIPTION(DM_NAME " multipath target");
2210 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2211 MODULE_LICENSE("GPL");