HWPOISON: Add soft page offline support
[linux-2.6/linux-2.6-openrd.git] / drivers / md / dm-mpath.c
blobe81345a1d08fa2f1fec21437c420ccacbbdcd360
1 /*
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.
6 */
8 #include <linux/device-mapper.h>
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <asm/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
27 /* Path properties */
28 struct pgpath {
29 struct list_head list;
31 struct priority_group *pg; /* Owning PG */
32 unsigned is_active; /* Path status */
33 unsigned fail_count; /* Cumulative failure count */
35 struct dm_path path;
36 struct work_struct deactivate_path;
37 struct work_struct activate_path;
40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
44 * Each has a path selector which controls which path gets used.
46 struct priority_group {
47 struct list_head list;
49 struct multipath *m; /* Owning multipath instance */
50 struct path_selector ps;
52 unsigned pg_num; /* Reference number */
53 unsigned bypassed; /* Temporarily bypass this PG? */
55 unsigned nr_pgpaths; /* Number of paths in PG */
56 struct list_head pgpaths;
59 /* Multipath context */
60 struct multipath {
61 struct list_head list;
62 struct dm_target *ti;
64 spinlock_t lock;
66 const char *hw_handler_name;
67 char *hw_handler_params;
68 unsigned nr_priority_groups;
69 struct list_head priority_groups;
70 unsigned pg_init_required; /* pg_init needs calling? */
71 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
73 unsigned nr_valid_paths; /* Total number of usable paths */
74 struct pgpath *current_pgpath;
75 struct priority_group *current_pg;
76 struct priority_group *next_pg; /* Switch to this PG if set */
77 unsigned repeat_count; /* I/Os left before calling PS again */
79 unsigned queue_io; /* Must we queue all I/O? */
80 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
81 unsigned saved_queue_if_no_path;/* Saved state during suspension */
82 unsigned pg_init_retries; /* Number of times to retry pg_init */
83 unsigned pg_init_count; /* Number of times pg_init called */
85 struct work_struct process_queued_ios;
86 struct list_head queued_ios;
87 unsigned queue_size;
89 struct work_struct trigger_event;
92 * We must use a mempool of dm_mpath_io structs so that we
93 * can resubmit bios on error.
95 mempool_t *mpio_pool;
97 struct mutex work_mutex;
99 unsigned suspended; /* Don't create new I/O internally when set. */
103 * Context information attached to each bio we process.
105 struct dm_mpath_io {
106 struct pgpath *pgpath;
107 size_t nr_bytes;
110 typedef int (*action_fn) (struct pgpath *pgpath);
112 #define MIN_IOS 256 /* Mempool size */
114 static struct kmem_cache *_mpio_cache;
116 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
117 static void process_queued_ios(struct work_struct *work);
118 static void trigger_event(struct work_struct *work);
119 static void activate_path(struct work_struct *work);
120 static void deactivate_path(struct work_struct *work);
123 /*-----------------------------------------------
124 * Allocation routines
125 *-----------------------------------------------*/
127 static struct pgpath *alloc_pgpath(void)
129 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
131 if (pgpath) {
132 pgpath->is_active = 1;
133 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
134 INIT_WORK(&pgpath->activate_path, activate_path);
137 return pgpath;
140 static void free_pgpath(struct pgpath *pgpath)
142 kfree(pgpath);
145 static void deactivate_path(struct work_struct *work)
147 struct pgpath *pgpath =
148 container_of(work, struct pgpath, deactivate_path);
150 blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
153 static struct priority_group *alloc_priority_group(void)
155 struct priority_group *pg;
157 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
159 if (pg)
160 INIT_LIST_HEAD(&pg->pgpaths);
162 return pg;
165 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
167 struct pgpath *pgpath, *tmp;
168 struct multipath *m = ti->private;
170 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
171 list_del(&pgpath->list);
172 if (m->hw_handler_name)
173 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
174 dm_put_device(ti, pgpath->path.dev);
175 free_pgpath(pgpath);
179 static void free_priority_group(struct priority_group *pg,
180 struct dm_target *ti)
182 struct path_selector *ps = &pg->ps;
184 if (ps->type) {
185 ps->type->destroy(ps);
186 dm_put_path_selector(ps->type);
189 free_pgpaths(&pg->pgpaths, ti);
190 kfree(pg);
193 static struct multipath *alloc_multipath(struct dm_target *ti)
195 struct multipath *m;
197 m = kzalloc(sizeof(*m), GFP_KERNEL);
198 if (m) {
199 INIT_LIST_HEAD(&m->priority_groups);
200 INIT_LIST_HEAD(&m->queued_ios);
201 spin_lock_init(&m->lock);
202 m->queue_io = 1;
203 INIT_WORK(&m->process_queued_ios, process_queued_ios);
204 INIT_WORK(&m->trigger_event, trigger_event);
205 mutex_init(&m->work_mutex);
206 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
207 if (!m->mpio_pool) {
208 kfree(m);
209 return NULL;
211 m->ti = ti;
212 ti->private = m;
215 return m;
218 static void free_multipath(struct multipath *m)
220 struct priority_group *pg, *tmp;
222 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
223 list_del(&pg->list);
224 free_priority_group(pg, m->ti);
227 kfree(m->hw_handler_name);
228 kfree(m->hw_handler_params);
229 mempool_destroy(m->mpio_pool);
230 kfree(m);
234 /*-----------------------------------------------
235 * Path selection
236 *-----------------------------------------------*/
238 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
240 m->current_pg = pgpath->pg;
242 /* Must we initialise the PG first, and queue I/O till it's ready? */
243 if (m->hw_handler_name) {
244 m->pg_init_required = 1;
245 m->queue_io = 1;
246 } else {
247 m->pg_init_required = 0;
248 m->queue_io = 0;
251 m->pg_init_count = 0;
254 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
255 size_t nr_bytes)
257 struct dm_path *path;
259 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
260 if (!path)
261 return -ENXIO;
263 m->current_pgpath = path_to_pgpath(path);
265 if (m->current_pg != pg)
266 __switch_pg(m, m->current_pgpath);
268 return 0;
271 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
273 struct priority_group *pg;
274 unsigned bypassed = 1;
276 if (!m->nr_valid_paths)
277 goto failed;
279 /* Were we instructed to switch PG? */
280 if (m->next_pg) {
281 pg = m->next_pg;
282 m->next_pg = NULL;
283 if (!__choose_path_in_pg(m, pg, nr_bytes))
284 return;
287 /* Don't change PG until it has no remaining paths */
288 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
289 return;
292 * Loop through priority groups until we find a valid path.
293 * First time we skip PGs marked 'bypassed'.
294 * Second time we only try the ones we skipped.
296 do {
297 list_for_each_entry(pg, &m->priority_groups, list) {
298 if (pg->bypassed == bypassed)
299 continue;
300 if (!__choose_path_in_pg(m, pg, nr_bytes))
301 return;
303 } while (bypassed--);
305 failed:
306 m->current_pgpath = NULL;
307 m->current_pg = NULL;
311 * Check whether bios must be queued in the device-mapper core rather
312 * than here in the target.
314 * m->lock must be held on entry.
316 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
317 * same value then we are not between multipath_presuspend()
318 * and multipath_resume() calls and we have no need to check
319 * for the DMF_NOFLUSH_SUSPENDING flag.
321 static int __must_push_back(struct multipath *m)
323 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
324 dm_noflush_suspending(m->ti));
327 static int map_io(struct multipath *m, struct request *clone,
328 struct dm_mpath_io *mpio, unsigned was_queued)
330 int r = DM_MAPIO_REMAPPED;
331 size_t nr_bytes = blk_rq_bytes(clone);
332 unsigned long flags;
333 struct pgpath *pgpath;
334 struct block_device *bdev;
336 spin_lock_irqsave(&m->lock, flags);
338 /* Do we need to select a new pgpath? */
339 if (!m->current_pgpath ||
340 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
341 __choose_pgpath(m, nr_bytes);
343 pgpath = m->current_pgpath;
345 if (was_queued)
346 m->queue_size--;
348 if ((pgpath && m->queue_io) ||
349 (!pgpath && m->queue_if_no_path)) {
350 /* Queue for the daemon to resubmit */
351 list_add_tail(&clone->queuelist, &m->queued_ios);
352 m->queue_size++;
353 if ((m->pg_init_required && !m->pg_init_in_progress) ||
354 !m->queue_io)
355 queue_work(kmultipathd, &m->process_queued_ios);
356 pgpath = NULL;
357 r = DM_MAPIO_SUBMITTED;
358 } else if (pgpath) {
359 bdev = pgpath->path.dev->bdev;
360 clone->q = bdev_get_queue(bdev);
361 clone->rq_disk = bdev->bd_disk;
362 } else if (__must_push_back(m))
363 r = DM_MAPIO_REQUEUE;
364 else
365 r = -EIO; /* Failed */
367 mpio->pgpath = pgpath;
368 mpio->nr_bytes = nr_bytes;
370 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
371 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
372 nr_bytes);
374 spin_unlock_irqrestore(&m->lock, flags);
376 return r;
380 * If we run out of usable paths, should we queue I/O or error it?
382 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
383 unsigned save_old_value)
385 unsigned long flags;
387 spin_lock_irqsave(&m->lock, flags);
389 if (save_old_value)
390 m->saved_queue_if_no_path = m->queue_if_no_path;
391 else
392 m->saved_queue_if_no_path = queue_if_no_path;
393 m->queue_if_no_path = queue_if_no_path;
394 if (!m->queue_if_no_path && m->queue_size)
395 queue_work(kmultipathd, &m->process_queued_ios);
397 spin_unlock_irqrestore(&m->lock, flags);
399 return 0;
402 /*-----------------------------------------------------------------
403 * The multipath daemon is responsible for resubmitting queued ios.
404 *---------------------------------------------------------------*/
406 static void dispatch_queued_ios(struct multipath *m)
408 int r;
409 unsigned long flags;
410 struct dm_mpath_io *mpio;
411 union map_info *info;
412 struct request *clone, *n;
413 LIST_HEAD(cl);
415 spin_lock_irqsave(&m->lock, flags);
416 list_splice_init(&m->queued_ios, &cl);
417 spin_unlock_irqrestore(&m->lock, flags);
419 list_for_each_entry_safe(clone, n, &cl, queuelist) {
420 list_del_init(&clone->queuelist);
422 info = dm_get_rq_mapinfo(clone);
423 mpio = info->ptr;
425 r = map_io(m, clone, mpio, 1);
426 if (r < 0) {
427 mempool_free(mpio, m->mpio_pool);
428 dm_kill_unmapped_request(clone, r);
429 } else if (r == DM_MAPIO_REMAPPED)
430 dm_dispatch_request(clone);
431 else if (r == DM_MAPIO_REQUEUE) {
432 mempool_free(mpio, m->mpio_pool);
433 dm_requeue_unmapped_request(clone);
438 static void process_queued_ios(struct work_struct *work)
440 struct multipath *m =
441 container_of(work, struct multipath, process_queued_ios);
442 struct pgpath *pgpath = NULL, *tmp;
443 unsigned must_queue = 1;
444 unsigned long flags;
446 spin_lock_irqsave(&m->lock, flags);
448 if (!m->queue_size)
449 goto out;
451 if (!m->current_pgpath)
452 __choose_pgpath(m, 0);
454 pgpath = m->current_pgpath;
456 if ((pgpath && !m->queue_io) ||
457 (!pgpath && !m->queue_if_no_path))
458 must_queue = 0;
460 if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
461 m->pg_init_count++;
462 m->pg_init_required = 0;
463 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
464 if (queue_work(kmpath_handlerd, &tmp->activate_path))
465 m->pg_init_in_progress++;
468 out:
469 spin_unlock_irqrestore(&m->lock, flags);
470 if (!must_queue)
471 dispatch_queued_ios(m);
475 * An event is triggered whenever a path is taken out of use.
476 * Includes path failure and PG bypass.
478 static void trigger_event(struct work_struct *work)
480 struct multipath *m =
481 container_of(work, struct multipath, trigger_event);
483 dm_table_event(m->ti->table);
486 /*-----------------------------------------------------------------
487 * Constructor/argument parsing:
488 * <#multipath feature args> [<arg>]*
489 * <#hw_handler args> [hw_handler [<arg>]*]
490 * <#priority groups>
491 * <initial priority group>
492 * [<selector> <#selector args> [<arg>]*
493 * <#paths> <#per-path selector args>
494 * [<path> [<arg>]* ]+ ]+
495 *---------------------------------------------------------------*/
496 struct param {
497 unsigned min;
498 unsigned max;
499 char *error;
502 static int read_param(struct param *param, char *str, unsigned *v, char **error)
504 if (!str ||
505 (sscanf(str, "%u", v) != 1) ||
506 (*v < param->min) ||
507 (*v > param->max)) {
508 *error = param->error;
509 return -EINVAL;
512 return 0;
515 struct arg_set {
516 unsigned argc;
517 char **argv;
520 static char *shift(struct arg_set *as)
522 char *r;
524 if (as->argc) {
525 as->argc--;
526 r = *as->argv;
527 as->argv++;
528 return r;
531 return NULL;
534 static void consume(struct arg_set *as, unsigned n)
536 BUG_ON (as->argc < n);
537 as->argc -= n;
538 as->argv += n;
541 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
542 struct dm_target *ti)
544 int r;
545 struct path_selector_type *pst;
546 unsigned ps_argc;
548 static struct param _params[] = {
549 {0, 1024, "invalid number of path selector args"},
552 pst = dm_get_path_selector(shift(as));
553 if (!pst) {
554 ti->error = "unknown path selector type";
555 return -EINVAL;
558 r = read_param(_params, shift(as), &ps_argc, &ti->error);
559 if (r) {
560 dm_put_path_selector(pst);
561 return -EINVAL;
564 if (ps_argc > as->argc) {
565 dm_put_path_selector(pst);
566 ti->error = "not enough arguments for path selector";
567 return -EINVAL;
570 r = pst->create(&pg->ps, ps_argc, as->argv);
571 if (r) {
572 dm_put_path_selector(pst);
573 ti->error = "path selector constructor failed";
574 return r;
577 pg->ps.type = pst;
578 consume(as, ps_argc);
580 return 0;
583 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
584 struct dm_target *ti)
586 int r;
587 struct pgpath *p;
588 struct multipath *m = ti->private;
590 /* we need at least a path arg */
591 if (as->argc < 1) {
592 ti->error = "no device given";
593 return ERR_PTR(-EINVAL);
596 p = alloc_pgpath();
597 if (!p)
598 return ERR_PTR(-ENOMEM);
600 r = dm_get_device(ti, shift(as), ti->begin, ti->len,
601 dm_table_get_mode(ti->table), &p->path.dev);
602 if (r) {
603 ti->error = "error getting device";
604 goto bad;
607 if (m->hw_handler_name) {
608 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
610 r = scsi_dh_attach(q, m->hw_handler_name);
611 if (r == -EBUSY) {
613 * Already attached to different hw_handler,
614 * try to reattach with correct one.
616 scsi_dh_detach(q);
617 r = scsi_dh_attach(q, m->hw_handler_name);
620 if (r < 0) {
621 ti->error = "error attaching hardware handler";
622 dm_put_device(ti, p->path.dev);
623 goto bad;
626 if (m->hw_handler_params) {
627 r = scsi_dh_set_params(q, m->hw_handler_params);
628 if (r < 0) {
629 ti->error = "unable to set hardware "
630 "handler parameters";
631 scsi_dh_detach(q);
632 dm_put_device(ti, p->path.dev);
633 goto bad;
638 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
639 if (r) {
640 dm_put_device(ti, p->path.dev);
641 goto bad;
644 return p;
646 bad:
647 free_pgpath(p);
648 return ERR_PTR(r);
651 static struct priority_group *parse_priority_group(struct arg_set *as,
652 struct multipath *m)
654 static struct param _params[] = {
655 {1, 1024, "invalid number of paths"},
656 {0, 1024, "invalid number of selector args"}
659 int r;
660 unsigned i, nr_selector_args, nr_params;
661 struct priority_group *pg;
662 struct dm_target *ti = m->ti;
664 if (as->argc < 2) {
665 as->argc = 0;
666 ti->error = "not enough priority group arguments";
667 return ERR_PTR(-EINVAL);
670 pg = alloc_priority_group();
671 if (!pg) {
672 ti->error = "couldn't allocate priority group";
673 return ERR_PTR(-ENOMEM);
675 pg->m = m;
677 r = parse_path_selector(as, pg, ti);
678 if (r)
679 goto bad;
682 * read the paths
684 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
685 if (r)
686 goto bad;
688 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
689 if (r)
690 goto bad;
692 nr_params = 1 + nr_selector_args;
693 for (i = 0; i < pg->nr_pgpaths; i++) {
694 struct pgpath *pgpath;
695 struct arg_set path_args;
697 if (as->argc < nr_params) {
698 ti->error = "not enough path parameters";
699 goto bad;
702 path_args.argc = nr_params;
703 path_args.argv = as->argv;
705 pgpath = parse_path(&path_args, &pg->ps, ti);
706 if (IS_ERR(pgpath)) {
707 r = PTR_ERR(pgpath);
708 goto bad;
711 pgpath->pg = pg;
712 list_add_tail(&pgpath->list, &pg->pgpaths);
713 consume(as, nr_params);
716 return pg;
718 bad:
719 free_priority_group(pg, ti);
720 return ERR_PTR(r);
723 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
725 unsigned hw_argc;
726 int ret;
727 struct dm_target *ti = m->ti;
729 static struct param _params[] = {
730 {0, 1024, "invalid number of hardware handler args"},
733 if (read_param(_params, shift(as), &hw_argc, &ti->error))
734 return -EINVAL;
736 if (!hw_argc)
737 return 0;
739 if (hw_argc > as->argc) {
740 ti->error = "not enough arguments for hardware handler";
741 return -EINVAL;
744 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
745 request_module("scsi_dh_%s", m->hw_handler_name);
746 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
747 ti->error = "unknown hardware handler type";
748 ret = -EINVAL;
749 goto fail;
752 if (hw_argc > 1) {
753 char *p;
754 int i, j, len = 4;
756 for (i = 0; i <= hw_argc - 2; i++)
757 len += strlen(as->argv[i]) + 1;
758 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
759 if (!p) {
760 ti->error = "memory allocation failed";
761 ret = -ENOMEM;
762 goto fail;
764 j = sprintf(p, "%d", hw_argc - 1);
765 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
766 j = sprintf(p, "%s", as->argv[i]);
768 consume(as, hw_argc - 1);
770 return 0;
771 fail:
772 kfree(m->hw_handler_name);
773 m->hw_handler_name = NULL;
774 return ret;
777 static int parse_features(struct arg_set *as, struct multipath *m)
779 int r;
780 unsigned argc;
781 struct dm_target *ti = m->ti;
782 const char *param_name;
784 static struct param _params[] = {
785 {0, 3, "invalid number of feature args"},
786 {1, 50, "pg_init_retries must be between 1 and 50"},
789 r = read_param(_params, shift(as), &argc, &ti->error);
790 if (r)
791 return -EINVAL;
793 if (!argc)
794 return 0;
796 do {
797 param_name = shift(as);
798 argc--;
800 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
801 r = queue_if_no_path(m, 1, 0);
802 continue;
805 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
806 (argc >= 1)) {
807 r = read_param(_params + 1, shift(as),
808 &m->pg_init_retries, &ti->error);
809 argc--;
810 continue;
813 ti->error = "Unrecognised multipath feature request";
814 r = -EINVAL;
815 } while (argc && !r);
817 return r;
820 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
821 char **argv)
823 /* target parameters */
824 static struct param _params[] = {
825 {1, 1024, "invalid number of priority groups"},
826 {1, 1024, "invalid initial priority group number"},
829 int r;
830 struct multipath *m;
831 struct arg_set as;
832 unsigned pg_count = 0;
833 unsigned next_pg_num;
835 as.argc = argc;
836 as.argv = argv;
838 m = alloc_multipath(ti);
839 if (!m) {
840 ti->error = "can't allocate multipath";
841 return -EINVAL;
844 r = parse_features(&as, m);
845 if (r)
846 goto bad;
848 r = parse_hw_handler(&as, m);
849 if (r)
850 goto bad;
852 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
853 if (r)
854 goto bad;
856 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
857 if (r)
858 goto bad;
860 /* parse the priority groups */
861 while (as.argc) {
862 struct priority_group *pg;
864 pg = parse_priority_group(&as, m);
865 if (IS_ERR(pg)) {
866 r = PTR_ERR(pg);
867 goto bad;
870 m->nr_valid_paths += pg->nr_pgpaths;
871 list_add_tail(&pg->list, &m->priority_groups);
872 pg_count++;
873 pg->pg_num = pg_count;
874 if (!--next_pg_num)
875 m->next_pg = pg;
878 if (pg_count != m->nr_priority_groups) {
879 ti->error = "priority group count mismatch";
880 r = -EINVAL;
881 goto bad;
884 ti->num_flush_requests = 1;
886 return 0;
888 bad:
889 free_multipath(m);
890 return r;
893 static void flush_multipath_work(void)
895 flush_workqueue(kmpath_handlerd);
896 flush_workqueue(kmultipathd);
897 flush_scheduled_work();
900 static void multipath_dtr(struct dm_target *ti)
902 struct multipath *m = ti->private;
904 flush_multipath_work();
905 free_multipath(m);
909 * Map cloned requests
911 static int multipath_map(struct dm_target *ti, struct request *clone,
912 union map_info *map_context)
914 int r;
915 struct dm_mpath_io *mpio;
916 struct multipath *m = (struct multipath *) ti->private;
918 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
919 if (!mpio)
920 /* ENOMEM, requeue */
921 return DM_MAPIO_REQUEUE;
922 memset(mpio, 0, sizeof(*mpio));
924 map_context->ptr = mpio;
925 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
926 r = map_io(m, clone, mpio, 0);
927 if (r < 0 || r == DM_MAPIO_REQUEUE)
928 mempool_free(mpio, m->mpio_pool);
930 return r;
934 * Take a path out of use.
936 static int fail_path(struct pgpath *pgpath)
938 unsigned long flags;
939 struct multipath *m = pgpath->pg->m;
941 spin_lock_irqsave(&m->lock, flags);
943 if (!pgpath->is_active)
944 goto out;
946 DMWARN("Failing path %s.", pgpath->path.dev->name);
948 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
949 pgpath->is_active = 0;
950 pgpath->fail_count++;
952 m->nr_valid_paths--;
954 if (pgpath == m->current_pgpath)
955 m->current_pgpath = NULL;
957 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
958 pgpath->path.dev->name, m->nr_valid_paths);
960 schedule_work(&m->trigger_event);
961 queue_work(kmultipathd, &pgpath->deactivate_path);
963 out:
964 spin_unlock_irqrestore(&m->lock, flags);
966 return 0;
970 * Reinstate a previously-failed path
972 static int reinstate_path(struct pgpath *pgpath)
974 int r = 0;
975 unsigned long flags;
976 struct multipath *m = pgpath->pg->m;
978 spin_lock_irqsave(&m->lock, flags);
980 if (pgpath->is_active)
981 goto out;
983 if (!pgpath->pg->ps.type->reinstate_path) {
984 DMWARN("Reinstate path not supported by path selector %s",
985 pgpath->pg->ps.type->name);
986 r = -EINVAL;
987 goto out;
990 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
991 if (r)
992 goto out;
994 pgpath->is_active = 1;
996 if (!m->nr_valid_paths++ && m->queue_size) {
997 m->current_pgpath = NULL;
998 queue_work(kmultipathd, &m->process_queued_ios);
999 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1000 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
1001 m->pg_init_in_progress++;
1004 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1005 pgpath->path.dev->name, m->nr_valid_paths);
1007 schedule_work(&m->trigger_event);
1009 out:
1010 spin_unlock_irqrestore(&m->lock, flags);
1012 return r;
1016 * Fail or reinstate all paths that match the provided struct dm_dev.
1018 static int action_dev(struct multipath *m, struct dm_dev *dev,
1019 action_fn action)
1021 int r = 0;
1022 struct pgpath *pgpath;
1023 struct priority_group *pg;
1025 list_for_each_entry(pg, &m->priority_groups, list) {
1026 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1027 if (pgpath->path.dev == dev)
1028 r = action(pgpath);
1032 return r;
1036 * Temporarily try to avoid having to use the specified PG
1038 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1039 int bypassed)
1041 unsigned long flags;
1043 spin_lock_irqsave(&m->lock, flags);
1045 pg->bypassed = bypassed;
1046 m->current_pgpath = NULL;
1047 m->current_pg = NULL;
1049 spin_unlock_irqrestore(&m->lock, flags);
1051 schedule_work(&m->trigger_event);
1055 * Switch to using the specified PG from the next I/O that gets mapped
1057 static int switch_pg_num(struct multipath *m, const char *pgstr)
1059 struct priority_group *pg;
1060 unsigned pgnum;
1061 unsigned long flags;
1063 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1064 (pgnum > m->nr_priority_groups)) {
1065 DMWARN("invalid PG number supplied to switch_pg_num");
1066 return -EINVAL;
1069 spin_lock_irqsave(&m->lock, flags);
1070 list_for_each_entry(pg, &m->priority_groups, list) {
1071 pg->bypassed = 0;
1072 if (--pgnum)
1073 continue;
1075 m->current_pgpath = NULL;
1076 m->current_pg = NULL;
1077 m->next_pg = pg;
1079 spin_unlock_irqrestore(&m->lock, flags);
1081 schedule_work(&m->trigger_event);
1082 return 0;
1086 * Set/clear bypassed status of a PG.
1087 * PGs are numbered upwards from 1 in the order they were declared.
1089 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1091 struct priority_group *pg;
1092 unsigned pgnum;
1094 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1095 (pgnum > m->nr_priority_groups)) {
1096 DMWARN("invalid PG number supplied to bypass_pg");
1097 return -EINVAL;
1100 list_for_each_entry(pg, &m->priority_groups, list) {
1101 if (!--pgnum)
1102 break;
1105 bypass_pg(m, pg, bypassed);
1106 return 0;
1110 * Should we retry pg_init immediately?
1112 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1114 unsigned long flags;
1115 int limit_reached = 0;
1117 spin_lock_irqsave(&m->lock, flags);
1119 if (m->pg_init_count <= m->pg_init_retries)
1120 m->pg_init_required = 1;
1121 else
1122 limit_reached = 1;
1124 spin_unlock_irqrestore(&m->lock, flags);
1126 return limit_reached;
1129 static void pg_init_done(void *data, int errors)
1131 struct dm_path *path = data;
1132 struct pgpath *pgpath = path_to_pgpath(path);
1133 struct priority_group *pg = pgpath->pg;
1134 struct multipath *m = pg->m;
1135 unsigned long flags;
1137 /* device or driver problems */
1138 switch (errors) {
1139 case SCSI_DH_OK:
1140 break;
1141 case SCSI_DH_NOSYS:
1142 if (!m->hw_handler_name) {
1143 errors = 0;
1144 break;
1146 DMERR("Cannot failover device because scsi_dh_%s was not "
1147 "loaded.", m->hw_handler_name);
1149 * Fail path for now, so we do not ping pong
1151 fail_path(pgpath);
1152 break;
1153 case SCSI_DH_DEV_TEMP_BUSY:
1155 * Probably doing something like FW upgrade on the
1156 * controller so try the other pg.
1158 bypass_pg(m, pg, 1);
1159 break;
1160 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1161 case SCSI_DH_RETRY:
1162 case SCSI_DH_IMM_RETRY:
1163 case SCSI_DH_RES_TEMP_UNAVAIL:
1164 if (pg_init_limit_reached(m, pgpath))
1165 fail_path(pgpath);
1166 errors = 0;
1167 break;
1168 default:
1170 * We probably do not want to fail the path for a device
1171 * error, but this is what the old dm did. In future
1172 * patches we can do more advanced handling.
1174 fail_path(pgpath);
1177 spin_lock_irqsave(&m->lock, flags);
1178 if (errors) {
1179 if (pgpath == m->current_pgpath) {
1180 DMERR("Could not failover device. Error %d.", errors);
1181 m->current_pgpath = NULL;
1182 m->current_pg = NULL;
1184 } else if (!m->pg_init_required) {
1185 m->queue_io = 0;
1186 pg->bypassed = 0;
1189 m->pg_init_in_progress--;
1190 if (!m->pg_init_in_progress)
1191 queue_work(kmultipathd, &m->process_queued_ios);
1192 spin_unlock_irqrestore(&m->lock, flags);
1195 static void activate_path(struct work_struct *work)
1197 struct pgpath *pgpath =
1198 container_of(work, struct pgpath, activate_path);
1200 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1201 pg_init_done, &pgpath->path);
1205 * end_io handling
1207 static int do_end_io(struct multipath *m, struct request *clone,
1208 int error, struct dm_mpath_io *mpio)
1211 * We don't queue any clone request inside the multipath target
1212 * during end I/O handling, since those clone requests don't have
1213 * bio clones. If we queue them inside the multipath target,
1214 * we need to make bio clones, that requires memory allocation.
1215 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1216 * don't have bio clones.)
1217 * Instead of queueing the clone request here, we queue the original
1218 * request into dm core, which will remake a clone request and
1219 * clone bios for it and resubmit it later.
1221 int r = DM_ENDIO_REQUEUE;
1222 unsigned long flags;
1224 if (!error && !clone->errors)
1225 return 0; /* I/O complete */
1227 if (error == -EOPNOTSUPP)
1228 return error;
1230 if (mpio->pgpath)
1231 fail_path(mpio->pgpath);
1233 spin_lock_irqsave(&m->lock, flags);
1234 if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1235 r = -EIO;
1236 spin_unlock_irqrestore(&m->lock, flags);
1238 return r;
1241 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1242 int error, union map_info *map_context)
1244 struct multipath *m = ti->private;
1245 struct dm_mpath_io *mpio = map_context->ptr;
1246 struct pgpath *pgpath = mpio->pgpath;
1247 struct path_selector *ps;
1248 int r;
1250 r = do_end_io(m, clone, error, mpio);
1251 if (pgpath) {
1252 ps = &pgpath->pg->ps;
1253 if (ps->type->end_io)
1254 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1256 mempool_free(mpio, m->mpio_pool);
1258 return r;
1262 * Suspend can't complete until all the I/O is processed so if
1263 * the last path fails we must error any remaining I/O.
1264 * Note that if the freeze_bdev fails while suspending, the
1265 * queue_if_no_path state is lost - userspace should reset it.
1267 static void multipath_presuspend(struct dm_target *ti)
1269 struct multipath *m = (struct multipath *) ti->private;
1271 queue_if_no_path(m, 0, 1);
1274 static void multipath_postsuspend(struct dm_target *ti)
1276 struct multipath *m = ti->private;
1278 mutex_lock(&m->work_mutex);
1279 m->suspended = 1;
1280 flush_multipath_work();
1281 mutex_unlock(&m->work_mutex);
1285 * Restore the queue_if_no_path setting.
1287 static void multipath_resume(struct dm_target *ti)
1289 struct multipath *m = (struct multipath *) ti->private;
1290 unsigned long flags;
1292 mutex_lock(&m->work_mutex);
1293 m->suspended = 0;
1294 mutex_unlock(&m->work_mutex);
1296 spin_lock_irqsave(&m->lock, flags);
1297 m->queue_if_no_path = m->saved_queue_if_no_path;
1298 spin_unlock_irqrestore(&m->lock, flags);
1302 * Info output has the following format:
1303 * num_multipath_feature_args [multipath_feature_args]*
1304 * num_handler_status_args [handler_status_args]*
1305 * num_groups init_group_number
1306 * [A|D|E num_ps_status_args [ps_status_args]*
1307 * num_paths num_selector_args
1308 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1310 * Table output has the following format (identical to the constructor string):
1311 * num_feature_args [features_args]*
1312 * num_handler_args hw_handler [hw_handler_args]*
1313 * num_groups init_group_number
1314 * [priority selector-name num_ps_args [ps_args]*
1315 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1317 static int multipath_status(struct dm_target *ti, status_type_t type,
1318 char *result, unsigned int maxlen)
1320 int sz = 0;
1321 unsigned long flags;
1322 struct multipath *m = (struct multipath *) ti->private;
1323 struct priority_group *pg;
1324 struct pgpath *p;
1325 unsigned pg_num;
1326 char state;
1328 spin_lock_irqsave(&m->lock, flags);
1330 /* Features */
1331 if (type == STATUSTYPE_INFO)
1332 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1333 else {
1334 DMEMIT("%u ", m->queue_if_no_path +
1335 (m->pg_init_retries > 0) * 2);
1336 if (m->queue_if_no_path)
1337 DMEMIT("queue_if_no_path ");
1338 if (m->pg_init_retries)
1339 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1342 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1343 DMEMIT("0 ");
1344 else
1345 DMEMIT("1 %s ", m->hw_handler_name);
1347 DMEMIT("%u ", m->nr_priority_groups);
1349 if (m->next_pg)
1350 pg_num = m->next_pg->pg_num;
1351 else if (m->current_pg)
1352 pg_num = m->current_pg->pg_num;
1353 else
1354 pg_num = 1;
1356 DMEMIT("%u ", pg_num);
1358 switch (type) {
1359 case STATUSTYPE_INFO:
1360 list_for_each_entry(pg, &m->priority_groups, list) {
1361 if (pg->bypassed)
1362 state = 'D'; /* Disabled */
1363 else if (pg == m->current_pg)
1364 state = 'A'; /* Currently Active */
1365 else
1366 state = 'E'; /* Enabled */
1368 DMEMIT("%c ", state);
1370 if (pg->ps.type->status)
1371 sz += pg->ps.type->status(&pg->ps, NULL, type,
1372 result + sz,
1373 maxlen - sz);
1374 else
1375 DMEMIT("0 ");
1377 DMEMIT("%u %u ", pg->nr_pgpaths,
1378 pg->ps.type->info_args);
1380 list_for_each_entry(p, &pg->pgpaths, list) {
1381 DMEMIT("%s %s %u ", p->path.dev->name,
1382 p->is_active ? "A" : "F",
1383 p->fail_count);
1384 if (pg->ps.type->status)
1385 sz += pg->ps.type->status(&pg->ps,
1386 &p->path, type, result + sz,
1387 maxlen - sz);
1390 break;
1392 case STATUSTYPE_TABLE:
1393 list_for_each_entry(pg, &m->priority_groups, list) {
1394 DMEMIT("%s ", pg->ps.type->name);
1396 if (pg->ps.type->status)
1397 sz += pg->ps.type->status(&pg->ps, NULL, type,
1398 result + sz,
1399 maxlen - sz);
1400 else
1401 DMEMIT("0 ");
1403 DMEMIT("%u %u ", pg->nr_pgpaths,
1404 pg->ps.type->table_args);
1406 list_for_each_entry(p, &pg->pgpaths, list) {
1407 DMEMIT("%s ", p->path.dev->name);
1408 if (pg->ps.type->status)
1409 sz += pg->ps.type->status(&pg->ps,
1410 &p->path, type, result + sz,
1411 maxlen - sz);
1414 break;
1417 spin_unlock_irqrestore(&m->lock, flags);
1419 return 0;
1422 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1424 int r = -EINVAL;
1425 struct dm_dev *dev;
1426 struct multipath *m = (struct multipath *) ti->private;
1427 action_fn action;
1429 mutex_lock(&m->work_mutex);
1431 if (m->suspended) {
1432 r = -EBUSY;
1433 goto out;
1436 if (dm_suspended(ti)) {
1437 r = -EBUSY;
1438 goto out;
1441 if (argc == 1) {
1442 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1443 r = queue_if_no_path(m, 1, 0);
1444 goto out;
1445 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1446 r = queue_if_no_path(m, 0, 0);
1447 goto out;
1451 if (argc != 2) {
1452 DMWARN("Unrecognised multipath message received.");
1453 goto out;
1456 if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1457 r = bypass_pg_num(m, argv[1], 1);
1458 goto out;
1459 } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1460 r = bypass_pg_num(m, argv[1], 0);
1461 goto out;
1462 } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1463 r = switch_pg_num(m, argv[1]);
1464 goto out;
1465 } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1466 action = reinstate_path;
1467 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1468 action = fail_path;
1469 else {
1470 DMWARN("Unrecognised multipath message received.");
1471 goto out;
1474 r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1475 dm_table_get_mode(ti->table), &dev);
1476 if (r) {
1477 DMWARN("message: error getting device %s",
1478 argv[1]);
1479 goto out;
1482 r = action_dev(m, dev, action);
1484 dm_put_device(ti, dev);
1486 out:
1487 mutex_unlock(&m->work_mutex);
1488 return r;
1491 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1492 unsigned long arg)
1494 struct multipath *m = (struct multipath *) ti->private;
1495 struct block_device *bdev = NULL;
1496 fmode_t mode = 0;
1497 unsigned long flags;
1498 int r = 0;
1500 spin_lock_irqsave(&m->lock, flags);
1502 if (!m->current_pgpath)
1503 __choose_pgpath(m, 0);
1505 if (m->current_pgpath) {
1506 bdev = m->current_pgpath->path.dev->bdev;
1507 mode = m->current_pgpath->path.dev->mode;
1510 if (m->queue_io)
1511 r = -EAGAIN;
1512 else if (!bdev)
1513 r = -EIO;
1515 spin_unlock_irqrestore(&m->lock, flags);
1517 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1520 static int multipath_iterate_devices(struct dm_target *ti,
1521 iterate_devices_callout_fn fn, void *data)
1523 struct multipath *m = ti->private;
1524 struct priority_group *pg;
1525 struct pgpath *p;
1526 int ret = 0;
1528 list_for_each_entry(pg, &m->priority_groups, list) {
1529 list_for_each_entry(p, &pg->pgpaths, list) {
1530 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1531 if (ret)
1532 goto out;
1536 out:
1537 return ret;
1540 static int __pgpath_busy(struct pgpath *pgpath)
1542 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1544 return dm_underlying_device_busy(q);
1548 * We return "busy", only when we can map I/Os but underlying devices
1549 * are busy (so even if we map I/Os now, the I/Os will wait on
1550 * the underlying queue).
1551 * In other words, if we want to kill I/Os or queue them inside us
1552 * due to map unavailability, we don't return "busy". Otherwise,
1553 * dm core won't give us the I/Os and we can't do what we want.
1555 static int multipath_busy(struct dm_target *ti)
1557 int busy = 0, has_active = 0;
1558 struct multipath *m = ti->private;
1559 struct priority_group *pg;
1560 struct pgpath *pgpath;
1561 unsigned long flags;
1563 spin_lock_irqsave(&m->lock, flags);
1565 /* Guess which priority_group will be used at next mapping time */
1566 if (unlikely(!m->current_pgpath && m->next_pg))
1567 pg = m->next_pg;
1568 else if (likely(m->current_pg))
1569 pg = m->current_pg;
1570 else
1572 * We don't know which pg will be used at next mapping time.
1573 * We don't call __choose_pgpath() here to avoid to trigger
1574 * pg_init just by busy checking.
1575 * So we don't know whether underlying devices we will be using
1576 * at next mapping time are busy or not. Just try mapping.
1578 goto out;
1581 * If there is one non-busy active path at least, the path selector
1582 * will be able to select it. So we consider such a pg as not busy.
1584 busy = 1;
1585 list_for_each_entry(pgpath, &pg->pgpaths, list)
1586 if (pgpath->is_active) {
1587 has_active = 1;
1589 if (!__pgpath_busy(pgpath)) {
1590 busy = 0;
1591 break;
1595 if (!has_active)
1597 * No active path in this pg, so this pg won't be used and
1598 * the current_pg will be changed at next mapping time.
1599 * We need to try mapping to determine it.
1601 busy = 0;
1603 out:
1604 spin_unlock_irqrestore(&m->lock, flags);
1606 return busy;
1609 /*-----------------------------------------------------------------
1610 * Module setup
1611 *---------------------------------------------------------------*/
1612 static struct target_type multipath_target = {
1613 .name = "multipath",
1614 .version = {1, 1, 1},
1615 .module = THIS_MODULE,
1616 .ctr = multipath_ctr,
1617 .dtr = multipath_dtr,
1618 .map_rq = multipath_map,
1619 .rq_end_io = multipath_end_io,
1620 .presuspend = multipath_presuspend,
1621 .postsuspend = multipath_postsuspend,
1622 .resume = multipath_resume,
1623 .status = multipath_status,
1624 .message = multipath_message,
1625 .ioctl = multipath_ioctl,
1626 .iterate_devices = multipath_iterate_devices,
1627 .busy = multipath_busy,
1630 static int __init dm_multipath_init(void)
1632 int r;
1634 /* allocate a slab for the dm_ios */
1635 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1636 if (!_mpio_cache)
1637 return -ENOMEM;
1639 r = dm_register_target(&multipath_target);
1640 if (r < 0) {
1641 DMERR("register failed %d", r);
1642 kmem_cache_destroy(_mpio_cache);
1643 return -EINVAL;
1646 kmultipathd = create_workqueue("kmpathd");
1647 if (!kmultipathd) {
1648 DMERR("failed to create workqueue kmpathd");
1649 dm_unregister_target(&multipath_target);
1650 kmem_cache_destroy(_mpio_cache);
1651 return -ENOMEM;
1655 * A separate workqueue is used to handle the device handlers
1656 * to avoid overloading existing workqueue. Overloading the
1657 * old workqueue would also create a bottleneck in the
1658 * path of the storage hardware device activation.
1660 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1661 if (!kmpath_handlerd) {
1662 DMERR("failed to create workqueue kmpath_handlerd");
1663 destroy_workqueue(kmultipathd);
1664 dm_unregister_target(&multipath_target);
1665 kmem_cache_destroy(_mpio_cache);
1666 return -ENOMEM;
1669 DMINFO("version %u.%u.%u loaded",
1670 multipath_target.version[0], multipath_target.version[1],
1671 multipath_target.version[2]);
1673 return r;
1676 static void __exit dm_multipath_exit(void)
1678 destroy_workqueue(kmpath_handlerd);
1679 destroy_workqueue(kmultipathd);
1681 dm_unregister_target(&multipath_target);
1682 kmem_cache_destroy(_mpio_cache);
1685 module_init(dm_multipath_init);
1686 module_exit(dm_multipath_exit);
1688 MODULE_DESCRIPTION(DM_NAME " multipath target");
1689 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1690 MODULE_LICENSE("GPL");