xref: /openbmc/linux/drivers/md/dm-mpath.c (revision 87c2ce3b)
1 /*
2  * Copyright (C) 2003 Sistina Software Limited.
3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm.h"
9 #include "dm-path-selector.h"
10 #include "dm-hw-handler.h"
11 #include "dm-bio-list.h"
12 #include "dm-bio-record.h"
13 
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <asm/atomic.h>
23 
24 #define MESG_STR(x) x, sizeof(x)
25 
26 /* Path properties */
27 struct pgpath {
28 	struct list_head list;
29 
30 	struct priority_group *pg;	/* Owning PG */
31 	unsigned fail_count;		/* Cumulative failure count */
32 
33 	struct path path;
34 };
35 
36 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
37 
38 /*
39  * Paths are grouped into Priority Groups and numbered from 1 upwards.
40  * Each has a path selector which controls which path gets used.
41  */
42 struct priority_group {
43 	struct list_head list;
44 
45 	struct multipath *m;		/* Owning multipath instance */
46 	struct path_selector ps;
47 
48 	unsigned pg_num;		/* Reference number */
49 	unsigned bypassed;		/* Temporarily bypass this PG? */
50 
51 	unsigned nr_pgpaths;		/* Number of paths in PG */
52 	struct list_head pgpaths;
53 };
54 
55 /* Multipath context */
56 struct multipath {
57 	struct list_head list;
58 	struct dm_target *ti;
59 
60 	spinlock_t lock;
61 
62 	struct hw_handler hw_handler;
63 	unsigned nr_priority_groups;
64 	struct list_head priority_groups;
65 	unsigned pg_init_required;	/* pg_init needs calling? */
66 	unsigned pg_init_in_progress;	/* Only one pg_init allowed at once */
67 
68 	unsigned nr_valid_paths;	/* Total number of usable paths */
69 	struct pgpath *current_pgpath;
70 	struct priority_group *current_pg;
71 	struct priority_group *next_pg;	/* Switch to this PG if set */
72 	unsigned repeat_count;		/* I/Os left before calling PS again */
73 
74 	unsigned queue_io;		/* Must we queue all I/O? */
75 	unsigned queue_if_no_path;	/* Queue I/O if last path fails? */
76 	unsigned saved_queue_if_no_path;/* Saved state during suspension */
77 
78 	struct work_struct process_queued_ios;
79 	struct bio_list queued_ios;
80 	unsigned queue_size;
81 
82 	struct work_struct trigger_event;
83 
84 	/*
85 	 * We must use a mempool of mpath_io structs so that we
86 	 * can resubmit bios on error.
87 	 */
88 	mempool_t *mpio_pool;
89 };
90 
91 /*
92  * Context information attached to each bio we process.
93  */
94 struct mpath_io {
95 	struct pgpath *pgpath;
96 	struct dm_bio_details details;
97 };
98 
99 typedef int (*action_fn) (struct pgpath *pgpath);
100 
101 #define MIN_IOS 256	/* Mempool size */
102 
103 static kmem_cache_t *_mpio_cache;
104 
105 struct workqueue_struct *kmultipathd;
106 static void process_queued_ios(void *data);
107 static void trigger_event(void *data);
108 
109 
110 /*-----------------------------------------------
111  * Allocation routines
112  *-----------------------------------------------*/
113 
114 static struct pgpath *alloc_pgpath(void)
115 {
116 	struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
117 
118 	if (pgpath) {
119 		memset(pgpath, 0, sizeof(*pgpath));
120 		pgpath->path.is_active = 1;
121 	}
122 
123 	return pgpath;
124 }
125 
126 static inline void free_pgpath(struct pgpath *pgpath)
127 {
128 	kfree(pgpath);
129 }
130 
131 static struct priority_group *alloc_priority_group(void)
132 {
133 	struct priority_group *pg;
134 
135 	pg = kmalloc(sizeof(*pg), GFP_KERNEL);
136 	if (!pg)
137 		return NULL;
138 
139 	memset(pg, 0, sizeof(*pg));
140 	INIT_LIST_HEAD(&pg->pgpaths);
141 
142 	return pg;
143 }
144 
145 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
146 {
147 	struct pgpath *pgpath, *tmp;
148 
149 	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
150 		list_del(&pgpath->list);
151 		dm_put_device(ti, pgpath->path.dev);
152 		free_pgpath(pgpath);
153 	}
154 }
155 
156 static void free_priority_group(struct priority_group *pg,
157 				struct dm_target *ti)
158 {
159 	struct path_selector *ps = &pg->ps;
160 
161 	if (ps->type) {
162 		ps->type->destroy(ps);
163 		dm_put_path_selector(ps->type);
164 	}
165 
166 	free_pgpaths(&pg->pgpaths, ti);
167 	kfree(pg);
168 }
169 
170 static struct multipath *alloc_multipath(void)
171 {
172 	struct multipath *m;
173 
174 	m = kmalloc(sizeof(*m), GFP_KERNEL);
175 	if (m) {
176 		memset(m, 0, sizeof(*m));
177 		INIT_LIST_HEAD(&m->priority_groups);
178 		spin_lock_init(&m->lock);
179 		m->queue_io = 1;
180 		INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
181 		INIT_WORK(&m->trigger_event, trigger_event, m);
182 		m->mpio_pool = mempool_create(MIN_IOS, mempool_alloc_slab,
183 					      mempool_free_slab, _mpio_cache);
184 		if (!m->mpio_pool) {
185 			kfree(m);
186 			return NULL;
187 		}
188 	}
189 
190 	return m;
191 }
192 
193 static void free_multipath(struct multipath *m)
194 {
195 	struct priority_group *pg, *tmp;
196 	struct hw_handler *hwh = &m->hw_handler;
197 
198 	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
199 		list_del(&pg->list);
200 		free_priority_group(pg, m->ti);
201 	}
202 
203 	if (hwh->type) {
204 		hwh->type->destroy(hwh);
205 		dm_put_hw_handler(hwh->type);
206 	}
207 
208 	mempool_destroy(m->mpio_pool);
209 	kfree(m);
210 }
211 
212 
213 /*-----------------------------------------------
214  * Path selection
215  *-----------------------------------------------*/
216 
217 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
218 {
219 	struct hw_handler *hwh = &m->hw_handler;
220 
221 	m->current_pg = pgpath->pg;
222 
223 	/* Must we initialise the PG first, and queue I/O till it's ready? */
224 	if (hwh->type && hwh->type->pg_init) {
225 		m->pg_init_required = 1;
226 		m->queue_io = 1;
227 	} else {
228 		m->pg_init_required = 0;
229 		m->queue_io = 0;
230 	}
231 }
232 
233 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
234 {
235 	struct path *path;
236 
237 	path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
238 	if (!path)
239 		return -ENXIO;
240 
241 	m->current_pgpath = path_to_pgpath(path);
242 
243 	if (m->current_pg != pg)
244 		__switch_pg(m, m->current_pgpath);
245 
246 	return 0;
247 }
248 
249 static void __choose_pgpath(struct multipath *m)
250 {
251 	struct priority_group *pg;
252 	unsigned bypassed = 1;
253 
254 	if (!m->nr_valid_paths)
255 		goto failed;
256 
257 	/* Were we instructed to switch PG? */
258 	if (m->next_pg) {
259 		pg = m->next_pg;
260 		m->next_pg = NULL;
261 		if (!__choose_path_in_pg(m, pg))
262 			return;
263 	}
264 
265 	/* Don't change PG until it has no remaining paths */
266 	if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
267 		return;
268 
269 	/*
270 	 * Loop through priority groups until we find a valid path.
271 	 * First time we skip PGs marked 'bypassed'.
272 	 * Second time we only try the ones we skipped.
273 	 */
274 	do {
275 		list_for_each_entry(pg, &m->priority_groups, list) {
276 			if (pg->bypassed == bypassed)
277 				continue;
278 			if (!__choose_path_in_pg(m, pg))
279 				return;
280 		}
281 	} while (bypassed--);
282 
283 failed:
284 	m->current_pgpath = NULL;
285 	m->current_pg = NULL;
286 }
287 
288 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
289 		  unsigned was_queued)
290 {
291 	int r = 1;
292 	unsigned long flags;
293 	struct pgpath *pgpath;
294 
295 	spin_lock_irqsave(&m->lock, flags);
296 
297 	/* Do we need to select a new pgpath? */
298 	if (!m->current_pgpath ||
299 	    (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
300 		__choose_pgpath(m);
301 
302 	pgpath = m->current_pgpath;
303 
304 	if (was_queued)
305 		m->queue_size--;
306 
307 	if ((pgpath && m->queue_io) ||
308 	    (!pgpath && m->queue_if_no_path)) {
309 		/* Queue for the daemon to resubmit */
310 		bio_list_add(&m->queued_ios, bio);
311 		m->queue_size++;
312 		if ((m->pg_init_required && !m->pg_init_in_progress) ||
313 		    !m->queue_io)
314 			queue_work(kmultipathd, &m->process_queued_ios);
315 		pgpath = NULL;
316 		r = 0;
317 	} else if (!pgpath)
318 		r = -EIO;		/* Failed */
319 	else
320 		bio->bi_bdev = pgpath->path.dev->bdev;
321 
322 	mpio->pgpath = pgpath;
323 
324 	spin_unlock_irqrestore(&m->lock, flags);
325 
326 	return r;
327 }
328 
329 /*
330  * If we run out of usable paths, should we queue I/O or error it?
331  */
332 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
333 			    unsigned save_old_value)
334 {
335 	unsigned long flags;
336 
337 	spin_lock_irqsave(&m->lock, flags);
338 
339 	if (save_old_value)
340 		m->saved_queue_if_no_path = m->queue_if_no_path;
341 	else
342 		m->saved_queue_if_no_path = queue_if_no_path;
343 	m->queue_if_no_path = queue_if_no_path;
344 	if (!m->queue_if_no_path && m->queue_size)
345 		queue_work(kmultipathd, &m->process_queued_ios);
346 
347 	spin_unlock_irqrestore(&m->lock, flags);
348 
349 	return 0;
350 }
351 
352 /*-----------------------------------------------------------------
353  * The multipath daemon is responsible for resubmitting queued ios.
354  *---------------------------------------------------------------*/
355 
356 static void dispatch_queued_ios(struct multipath *m)
357 {
358 	int r;
359 	unsigned long flags;
360 	struct bio *bio = NULL, *next;
361 	struct mpath_io *mpio;
362 	union map_info *info;
363 
364 	spin_lock_irqsave(&m->lock, flags);
365 	bio = bio_list_get(&m->queued_ios);
366 	spin_unlock_irqrestore(&m->lock, flags);
367 
368 	while (bio) {
369 		next = bio->bi_next;
370 		bio->bi_next = NULL;
371 
372 		info = dm_get_mapinfo(bio);
373 		mpio = info->ptr;
374 
375 		r = map_io(m, bio, mpio, 1);
376 		if (r < 0)
377 			bio_endio(bio, bio->bi_size, r);
378 		else if (r == 1)
379 			generic_make_request(bio);
380 
381 		bio = next;
382 	}
383 }
384 
385 static void process_queued_ios(void *data)
386 {
387 	struct multipath *m = (struct multipath *) data;
388 	struct hw_handler *hwh = &m->hw_handler;
389 	struct pgpath *pgpath = NULL;
390 	unsigned init_required = 0, must_queue = 1;
391 	unsigned long flags;
392 
393 	spin_lock_irqsave(&m->lock, flags);
394 
395 	if (!m->queue_size)
396 		goto out;
397 
398 	if (!m->current_pgpath)
399 		__choose_pgpath(m);
400 
401 	pgpath = m->current_pgpath;
402 
403 	if ((pgpath && !m->queue_io) ||
404 	    (!pgpath && !m->queue_if_no_path))
405 		must_queue = 0;
406 
407 	if (m->pg_init_required && !m->pg_init_in_progress) {
408 		m->pg_init_required = 0;
409 		m->pg_init_in_progress = 1;
410 		init_required = 1;
411 	}
412 
413 out:
414 	spin_unlock_irqrestore(&m->lock, flags);
415 
416 	if (init_required)
417 		hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
418 
419 	if (!must_queue)
420 		dispatch_queued_ios(m);
421 }
422 
423 /*
424  * An event is triggered whenever a path is taken out of use.
425  * Includes path failure and PG bypass.
426  */
427 static void trigger_event(void *data)
428 {
429 	struct multipath *m = (struct multipath *) data;
430 
431 	dm_table_event(m->ti->table);
432 }
433 
434 /*-----------------------------------------------------------------
435  * Constructor/argument parsing:
436  * <#multipath feature args> [<arg>]*
437  * <#hw_handler args> [hw_handler [<arg>]*]
438  * <#priority groups>
439  * <initial priority group>
440  *     [<selector> <#selector args> [<arg>]*
441  *      <#paths> <#per-path selector args>
442  *         [<path> [<arg>]* ]+ ]+
443  *---------------------------------------------------------------*/
444 struct param {
445 	unsigned min;
446 	unsigned max;
447 	char *error;
448 };
449 
450 #define ESTR(s) ("dm-multipath: " s)
451 
452 static int read_param(struct param *param, char *str, unsigned *v, char **error)
453 {
454 	if (!str ||
455 	    (sscanf(str, "%u", v) != 1) ||
456 	    (*v < param->min) ||
457 	    (*v > param->max)) {
458 		*error = param->error;
459 		return -EINVAL;
460 	}
461 
462 	return 0;
463 }
464 
465 struct arg_set {
466 	unsigned argc;
467 	char **argv;
468 };
469 
470 static char *shift(struct arg_set *as)
471 {
472 	char *r;
473 
474 	if (as->argc) {
475 		as->argc--;
476 		r = *as->argv;
477 		as->argv++;
478 		return r;
479 	}
480 
481 	return NULL;
482 }
483 
484 static void consume(struct arg_set *as, unsigned n)
485 {
486 	BUG_ON (as->argc < n);
487 	as->argc -= n;
488 	as->argv += n;
489 }
490 
491 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
492 			       struct dm_target *ti)
493 {
494 	int r;
495 	struct path_selector_type *pst;
496 	unsigned ps_argc;
497 
498 	static struct param _params[] = {
499 		{0, 1024, ESTR("invalid number of path selector args")},
500 	};
501 
502 	pst = dm_get_path_selector(shift(as));
503 	if (!pst) {
504 		ti->error = ESTR("unknown path selector type");
505 		return -EINVAL;
506 	}
507 
508 	r = read_param(_params, shift(as), &ps_argc, &ti->error);
509 	if (r)
510 		return -EINVAL;
511 
512 	r = pst->create(&pg->ps, ps_argc, as->argv);
513 	if (r) {
514 		dm_put_path_selector(pst);
515 		ti->error = ESTR("path selector constructor failed");
516 		return r;
517 	}
518 
519 	pg->ps.type = pst;
520 	consume(as, ps_argc);
521 
522 	return 0;
523 }
524 
525 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
526 			       struct dm_target *ti)
527 {
528 	int r;
529 	struct pgpath *p;
530 
531 	/* we need at least a path arg */
532 	if (as->argc < 1) {
533 		ti->error = ESTR("no device given");
534 		return NULL;
535 	}
536 
537 	p = alloc_pgpath();
538 	if (!p)
539 		return NULL;
540 
541 	r = dm_get_device(ti, shift(as), ti->begin, ti->len,
542 			  dm_table_get_mode(ti->table), &p->path.dev);
543 	if (r) {
544 		ti->error = ESTR("error getting device");
545 		goto bad;
546 	}
547 
548 	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
549 	if (r) {
550 		dm_put_device(ti, p->path.dev);
551 		goto bad;
552 	}
553 
554 	return p;
555 
556  bad:
557 	free_pgpath(p);
558 	return NULL;
559 }
560 
561 static struct priority_group *parse_priority_group(struct arg_set *as,
562 						   struct multipath *m,
563 						   struct dm_target *ti)
564 {
565 	static struct param _params[] = {
566 		{1, 1024, ESTR("invalid number of paths")},
567 		{0, 1024, ESTR("invalid number of selector args")}
568 	};
569 
570 	int r;
571 	unsigned i, nr_selector_args, nr_params;
572 	struct priority_group *pg;
573 
574 	if (as->argc < 2) {
575 		as->argc = 0;
576 		ti->error = ESTR("not enough priority group aruments");
577 		return NULL;
578 	}
579 
580 	pg = alloc_priority_group();
581 	if (!pg) {
582 		ti->error = ESTR("couldn't allocate priority group");
583 		return NULL;
584 	}
585 	pg->m = m;
586 
587 	r = parse_path_selector(as, pg, ti);
588 	if (r)
589 		goto bad;
590 
591 	/*
592 	 * read the paths
593 	 */
594 	r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
595 	if (r)
596 		goto bad;
597 
598 	r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
599 	if (r)
600 		goto bad;
601 
602 	nr_params = 1 + nr_selector_args;
603 	for (i = 0; i < pg->nr_pgpaths; i++) {
604 		struct pgpath *pgpath;
605 		struct arg_set path_args;
606 
607 		if (as->argc < nr_params)
608 			goto bad;
609 
610 		path_args.argc = nr_params;
611 		path_args.argv = as->argv;
612 
613 		pgpath = parse_path(&path_args, &pg->ps, ti);
614 		if (!pgpath)
615 			goto bad;
616 
617 		pgpath->pg = pg;
618 		list_add_tail(&pgpath->list, &pg->pgpaths);
619 		consume(as, nr_params);
620 	}
621 
622 	return pg;
623 
624  bad:
625 	free_priority_group(pg, ti);
626 	return NULL;
627 }
628 
629 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
630 			    struct dm_target *ti)
631 {
632 	int r;
633 	struct hw_handler_type *hwht;
634 	unsigned hw_argc;
635 
636 	static struct param _params[] = {
637 		{0, 1024, ESTR("invalid number of hardware handler args")},
638 	};
639 
640 	r = read_param(_params, shift(as), &hw_argc, &ti->error);
641 	if (r)
642 		return -EINVAL;
643 
644 	if (!hw_argc)
645 		return 0;
646 
647 	hwht = dm_get_hw_handler(shift(as));
648 	if (!hwht) {
649 		ti->error = ESTR("unknown hardware handler type");
650 		return -EINVAL;
651 	}
652 
653 	r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
654 	if (r) {
655 		dm_put_hw_handler(hwht);
656 		ti->error = ESTR("hardware handler constructor failed");
657 		return r;
658 	}
659 
660 	m->hw_handler.type = hwht;
661 	consume(as, hw_argc - 1);
662 
663 	return 0;
664 }
665 
666 static int parse_features(struct arg_set *as, struct multipath *m,
667 			  struct dm_target *ti)
668 {
669 	int r;
670 	unsigned argc;
671 
672 	static struct param _params[] = {
673 		{0, 1, ESTR("invalid number of feature args")},
674 	};
675 
676 	r = read_param(_params, shift(as), &argc, &ti->error);
677 	if (r)
678 		return -EINVAL;
679 
680 	if (!argc)
681 		return 0;
682 
683 	if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
684 		return queue_if_no_path(m, 1, 0);
685 	else {
686 		ti->error = "Unrecognised multipath feature request";
687 		return -EINVAL;
688 	}
689 }
690 
691 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
692 			 char **argv)
693 {
694 	/* target parameters */
695 	static struct param _params[] = {
696 		{1, 1024, ESTR("invalid number of priority groups")},
697 		{1, 1024, ESTR("invalid initial priority group number")},
698 	};
699 
700 	int r;
701 	struct multipath *m;
702 	struct arg_set as;
703 	unsigned pg_count = 0;
704 	unsigned next_pg_num;
705 
706 	as.argc = argc;
707 	as.argv = argv;
708 
709 	m = alloc_multipath();
710 	if (!m) {
711 		ti->error = ESTR("can't allocate multipath");
712 		return -EINVAL;
713 	}
714 
715 	r = parse_features(&as, m, ti);
716 	if (r)
717 		goto bad;
718 
719 	r = parse_hw_handler(&as, m, ti);
720 	if (r)
721 		goto bad;
722 
723 	r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
724 	if (r)
725 		goto bad;
726 
727 	r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
728 	if (r)
729 		goto bad;
730 
731 	/* parse the priority groups */
732 	while (as.argc) {
733 		struct priority_group *pg;
734 
735 		pg = parse_priority_group(&as, m, ti);
736 		if (!pg) {
737 			r = -EINVAL;
738 			goto bad;
739 		}
740 
741 		m->nr_valid_paths += pg->nr_pgpaths;
742 		list_add_tail(&pg->list, &m->priority_groups);
743 		pg_count++;
744 		pg->pg_num = pg_count;
745 		if (!--next_pg_num)
746 			m->next_pg = pg;
747 	}
748 
749 	if (pg_count != m->nr_priority_groups) {
750 		ti->error = ESTR("priority group count mismatch");
751 		r = -EINVAL;
752 		goto bad;
753 	}
754 
755 	ti->private = m;
756 	m->ti = ti;
757 
758 	return 0;
759 
760  bad:
761 	free_multipath(m);
762 	return r;
763 }
764 
765 static void multipath_dtr(struct dm_target *ti)
766 {
767 	struct multipath *m = (struct multipath *) ti->private;
768 
769 	flush_workqueue(kmultipathd);
770 	free_multipath(m);
771 }
772 
773 /*
774  * Map bios, recording original fields for later in case we have to resubmit
775  */
776 static int multipath_map(struct dm_target *ti, struct bio *bio,
777 			 union map_info *map_context)
778 {
779 	int r;
780 	struct mpath_io *mpio;
781 	struct multipath *m = (struct multipath *) ti->private;
782 
783 	if (bio_barrier(bio))
784 		return -EOPNOTSUPP;
785 
786 	mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
787 	dm_bio_record(&mpio->details, bio);
788 
789 	map_context->ptr = mpio;
790 	bio->bi_rw |= (1 << BIO_RW_FAILFAST);
791 	r = map_io(m, bio, mpio, 0);
792 	if (r < 0)
793 		mempool_free(mpio, m->mpio_pool);
794 
795 	return r;
796 }
797 
798 /*
799  * Take a path out of use.
800  */
801 static int fail_path(struct pgpath *pgpath)
802 {
803 	unsigned long flags;
804 	struct multipath *m = pgpath->pg->m;
805 
806 	spin_lock_irqsave(&m->lock, flags);
807 
808 	if (!pgpath->path.is_active)
809 		goto out;
810 
811 	DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
812 
813 	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
814 	pgpath->path.is_active = 0;
815 	pgpath->fail_count++;
816 
817 	m->nr_valid_paths--;
818 
819 	if (pgpath == m->current_pgpath)
820 		m->current_pgpath = NULL;
821 
822 	queue_work(kmultipathd, &m->trigger_event);
823 
824 out:
825 	spin_unlock_irqrestore(&m->lock, flags);
826 
827 	return 0;
828 }
829 
830 /*
831  * Reinstate a previously-failed path
832  */
833 static int reinstate_path(struct pgpath *pgpath)
834 {
835 	int r = 0;
836 	unsigned long flags;
837 	struct multipath *m = pgpath->pg->m;
838 
839 	spin_lock_irqsave(&m->lock, flags);
840 
841 	if (pgpath->path.is_active)
842 		goto out;
843 
844 	if (!pgpath->pg->ps.type) {
845 		DMWARN("Reinstate path not supported by path selector %s",
846 		       pgpath->pg->ps.type->name);
847 		r = -EINVAL;
848 		goto out;
849 	}
850 
851 	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
852 	if (r)
853 		goto out;
854 
855 	pgpath->path.is_active = 1;
856 
857 	m->current_pgpath = NULL;
858 	if (!m->nr_valid_paths++ && m->queue_size)
859 		queue_work(kmultipathd, &m->process_queued_ios);
860 
861 	queue_work(kmultipathd, &m->trigger_event);
862 
863 out:
864 	spin_unlock_irqrestore(&m->lock, flags);
865 
866 	return r;
867 }
868 
869 /*
870  * Fail or reinstate all paths that match the provided struct dm_dev.
871  */
872 static int action_dev(struct multipath *m, struct dm_dev *dev,
873 		      action_fn action)
874 {
875 	int r = 0;
876 	struct pgpath *pgpath;
877 	struct priority_group *pg;
878 
879 	list_for_each_entry(pg, &m->priority_groups, list) {
880 		list_for_each_entry(pgpath, &pg->pgpaths, list) {
881 			if (pgpath->path.dev == dev)
882 				r = action(pgpath);
883 		}
884 	}
885 
886 	return r;
887 }
888 
889 /*
890  * Temporarily try to avoid having to use the specified PG
891  */
892 static void bypass_pg(struct multipath *m, struct priority_group *pg,
893 		      int bypassed)
894 {
895 	unsigned long flags;
896 
897 	spin_lock_irqsave(&m->lock, flags);
898 
899 	pg->bypassed = bypassed;
900 	m->current_pgpath = NULL;
901 	m->current_pg = NULL;
902 
903 	spin_unlock_irqrestore(&m->lock, flags);
904 
905 	queue_work(kmultipathd, &m->trigger_event);
906 }
907 
908 /*
909  * Switch to using the specified PG from the next I/O that gets mapped
910  */
911 static int switch_pg_num(struct multipath *m, const char *pgstr)
912 {
913 	struct priority_group *pg;
914 	unsigned pgnum;
915 	unsigned long flags;
916 
917 	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
918 	    (pgnum > m->nr_priority_groups)) {
919 		DMWARN("invalid PG number supplied to switch_pg_num");
920 		return -EINVAL;
921 	}
922 
923 	spin_lock_irqsave(&m->lock, flags);
924 	list_for_each_entry(pg, &m->priority_groups, list) {
925 		pg->bypassed = 0;
926 		if (--pgnum)
927 			continue;
928 
929 		m->current_pgpath = NULL;
930 		m->current_pg = NULL;
931 		m->next_pg = pg;
932 	}
933 	spin_unlock_irqrestore(&m->lock, flags);
934 
935 	queue_work(kmultipathd, &m->trigger_event);
936 	return 0;
937 }
938 
939 /*
940  * Set/clear bypassed status of a PG.
941  * PGs are numbered upwards from 1 in the order they were declared.
942  */
943 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
944 {
945 	struct priority_group *pg;
946 	unsigned pgnum;
947 
948 	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
949 	    (pgnum > m->nr_priority_groups)) {
950 		DMWARN("invalid PG number supplied to bypass_pg");
951 		return -EINVAL;
952 	}
953 
954 	list_for_each_entry(pg, &m->priority_groups, list) {
955 		if (!--pgnum)
956 			break;
957 	}
958 
959 	bypass_pg(m, pg, bypassed);
960 	return 0;
961 }
962 
963 /*
964  * pg_init must call this when it has completed its initialisation
965  */
966 void dm_pg_init_complete(struct path *path, unsigned err_flags)
967 {
968 	struct pgpath *pgpath = path_to_pgpath(path);
969 	struct priority_group *pg = pgpath->pg;
970 	struct multipath *m = pg->m;
971 	unsigned long flags;
972 
973 	/* We insist on failing the path if the PG is already bypassed. */
974 	if (err_flags && pg->bypassed)
975 		err_flags |= MP_FAIL_PATH;
976 
977 	if (err_flags & MP_FAIL_PATH)
978 		fail_path(pgpath);
979 
980 	if (err_flags & MP_BYPASS_PG)
981 		bypass_pg(m, pg, 1);
982 
983 	spin_lock_irqsave(&m->lock, flags);
984 	if (err_flags) {
985 		m->current_pgpath = NULL;
986 		m->current_pg = NULL;
987 	} else if (!m->pg_init_required)
988 		m->queue_io = 0;
989 
990 	m->pg_init_in_progress = 0;
991 	queue_work(kmultipathd, &m->process_queued_ios);
992 	spin_unlock_irqrestore(&m->lock, flags);
993 }
994 
995 /*
996  * end_io handling
997  */
998 static int do_end_io(struct multipath *m, struct bio *bio,
999 		     int error, struct mpath_io *mpio)
1000 {
1001 	struct hw_handler *hwh = &m->hw_handler;
1002 	unsigned err_flags = MP_FAIL_PATH;	/* Default behavior */
1003 	unsigned long flags;
1004 
1005 	if (!error)
1006 		return 0;	/* I/O complete */
1007 
1008 	if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1009 		return error;
1010 
1011 	if (error == -EOPNOTSUPP)
1012 		return error;
1013 
1014 	spin_lock_irqsave(&m->lock, flags);
1015 	if (!m->nr_valid_paths) {
1016 		if (!m->queue_if_no_path) {
1017 			spin_unlock_irqrestore(&m->lock, flags);
1018 			return -EIO;
1019 		} else {
1020 			spin_unlock_irqrestore(&m->lock, flags);
1021 			goto requeue;
1022 		}
1023 	}
1024 	spin_unlock_irqrestore(&m->lock, flags);
1025 
1026 	if (hwh->type && hwh->type->error)
1027 		err_flags = hwh->type->error(hwh, bio);
1028 
1029 	if (mpio->pgpath) {
1030 		if (err_flags & MP_FAIL_PATH)
1031 			fail_path(mpio->pgpath);
1032 
1033 		if (err_flags & MP_BYPASS_PG)
1034 			bypass_pg(m, mpio->pgpath->pg, 1);
1035 	}
1036 
1037 	if (err_flags & MP_ERROR_IO)
1038 		return -EIO;
1039 
1040       requeue:
1041 	dm_bio_restore(&mpio->details, bio);
1042 
1043 	/* queue for the daemon to resubmit or fail */
1044 	spin_lock_irqsave(&m->lock, flags);
1045 	bio_list_add(&m->queued_ios, bio);
1046 	m->queue_size++;
1047 	if (!m->queue_io)
1048 		queue_work(kmultipathd, &m->process_queued_ios);
1049 	spin_unlock_irqrestore(&m->lock, flags);
1050 
1051 	return 1;	/* io not complete */
1052 }
1053 
1054 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1055 			    int error, union map_info *map_context)
1056 {
1057 	struct multipath *m = (struct multipath *) ti->private;
1058 	struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1059 	struct pgpath *pgpath = mpio->pgpath;
1060 	struct path_selector *ps;
1061 	int r;
1062 
1063 	r  = do_end_io(m, bio, error, mpio);
1064 	if (pgpath) {
1065 		ps = &pgpath->pg->ps;
1066 		if (ps->type->end_io)
1067 			ps->type->end_io(ps, &pgpath->path);
1068 	}
1069 	if (r <= 0)
1070 		mempool_free(mpio, m->mpio_pool);
1071 
1072 	return r;
1073 }
1074 
1075 /*
1076  * Suspend can't complete until all the I/O is processed so if
1077  * the last path fails we must error any remaining I/O.
1078  * Note that if the freeze_bdev fails while suspending, the
1079  * queue_if_no_path state is lost - userspace should reset it.
1080  */
1081 static void multipath_presuspend(struct dm_target *ti)
1082 {
1083 	struct multipath *m = (struct multipath *) ti->private;
1084 
1085 	queue_if_no_path(m, 0, 1);
1086 }
1087 
1088 /*
1089  * Restore the queue_if_no_path setting.
1090  */
1091 static void multipath_resume(struct dm_target *ti)
1092 {
1093 	struct multipath *m = (struct multipath *) ti->private;
1094 	unsigned long flags;
1095 
1096 	spin_lock_irqsave(&m->lock, flags);
1097 	m->queue_if_no_path = m->saved_queue_if_no_path;
1098 	spin_unlock_irqrestore(&m->lock, flags);
1099 }
1100 
1101 /*
1102  * Info output has the following format:
1103  * num_multipath_feature_args [multipath_feature_args]*
1104  * num_handler_status_args [handler_status_args]*
1105  * num_groups init_group_number
1106  *            [A|D|E num_ps_status_args [ps_status_args]*
1107  *             num_paths num_selector_args
1108  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1109  *
1110  * Table output has the following format (identical to the constructor string):
1111  * num_feature_args [features_args]*
1112  * num_handler_args hw_handler [hw_handler_args]*
1113  * num_groups init_group_number
1114  *     [priority selector-name num_ps_args [ps_args]*
1115  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1116  */
1117 static int multipath_status(struct dm_target *ti, status_type_t type,
1118 			    char *result, unsigned int maxlen)
1119 {
1120 	int sz = 0;
1121 	unsigned long flags;
1122 	struct multipath *m = (struct multipath *) ti->private;
1123 	struct hw_handler *hwh = &m->hw_handler;
1124 	struct priority_group *pg;
1125 	struct pgpath *p;
1126 	unsigned pg_num;
1127 	char state;
1128 
1129 	spin_lock_irqsave(&m->lock, flags);
1130 
1131 	/* Features */
1132 	if (type == STATUSTYPE_INFO)
1133 		DMEMIT("1 %u ", m->queue_size);
1134 	else if (m->queue_if_no_path)
1135 		DMEMIT("1 queue_if_no_path ");
1136 	else
1137 		DMEMIT("0 ");
1138 
1139 	if (hwh->type && hwh->type->status)
1140 		sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1141 	else if (!hwh->type || type == STATUSTYPE_INFO)
1142 		DMEMIT("0 ");
1143 	else
1144 		DMEMIT("1 %s ", hwh->type->name);
1145 
1146 	DMEMIT("%u ", m->nr_priority_groups);
1147 
1148 	if (m->next_pg)
1149 		pg_num = m->next_pg->pg_num;
1150 	else if (m->current_pg)
1151 		pg_num = m->current_pg->pg_num;
1152 	else
1153 			pg_num = 1;
1154 
1155 	DMEMIT("%u ", pg_num);
1156 
1157 	switch (type) {
1158 	case STATUSTYPE_INFO:
1159 		list_for_each_entry(pg, &m->priority_groups, list) {
1160 			if (pg->bypassed)
1161 				state = 'D';	/* Disabled */
1162 			else if (pg == m->current_pg)
1163 				state = 'A';	/* Currently Active */
1164 			else
1165 				state = 'E';	/* Enabled */
1166 
1167 			DMEMIT("%c ", state);
1168 
1169 			if (pg->ps.type->status)
1170 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1171 							  result + sz,
1172 							  maxlen - sz);
1173 			else
1174 				DMEMIT("0 ");
1175 
1176 			DMEMIT("%u %u ", pg->nr_pgpaths,
1177 			       pg->ps.type->info_args);
1178 
1179 			list_for_each_entry(p, &pg->pgpaths, list) {
1180 				DMEMIT("%s %s %u ", p->path.dev->name,
1181 				       p->path.is_active ? "A" : "F",
1182 				       p->fail_count);
1183 				if (pg->ps.type->status)
1184 					sz += pg->ps.type->status(&pg->ps,
1185 					      &p->path, type, result + sz,
1186 					      maxlen - sz);
1187 			}
1188 		}
1189 		break;
1190 
1191 	case STATUSTYPE_TABLE:
1192 		list_for_each_entry(pg, &m->priority_groups, list) {
1193 			DMEMIT("%s ", pg->ps.type->name);
1194 
1195 			if (pg->ps.type->status)
1196 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1197 							  result + sz,
1198 							  maxlen - sz);
1199 			else
1200 				DMEMIT("0 ");
1201 
1202 			DMEMIT("%u %u ", pg->nr_pgpaths,
1203 			       pg->ps.type->table_args);
1204 
1205 			list_for_each_entry(p, &pg->pgpaths, list) {
1206 				DMEMIT("%s ", p->path.dev->name);
1207 				if (pg->ps.type->status)
1208 					sz += pg->ps.type->status(&pg->ps,
1209 					      &p->path, type, result + sz,
1210 					      maxlen - sz);
1211 			}
1212 		}
1213 		break;
1214 	}
1215 
1216 	spin_unlock_irqrestore(&m->lock, flags);
1217 
1218 	return 0;
1219 }
1220 
1221 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1222 {
1223 	int r;
1224 	struct dm_dev *dev;
1225 	struct multipath *m = (struct multipath *) ti->private;
1226 	action_fn action;
1227 
1228 	if (argc == 1) {
1229 		if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1230 			return queue_if_no_path(m, 1, 0);
1231 		else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1232 			return queue_if_no_path(m, 0, 0);
1233 	}
1234 
1235 	if (argc != 2)
1236 		goto error;
1237 
1238 	if (!strnicmp(argv[0], MESG_STR("disable_group")))
1239 		return bypass_pg_num(m, argv[1], 1);
1240 	else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1241 		return bypass_pg_num(m, argv[1], 0);
1242 	else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1243 		return switch_pg_num(m, argv[1]);
1244 	else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1245 		action = reinstate_path;
1246 	else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1247 		action = fail_path;
1248 	else
1249 		goto error;
1250 
1251 	r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1252 			  dm_table_get_mode(ti->table), &dev);
1253 	if (r) {
1254 		DMWARN("dm-multipath message: error getting device %s",
1255 		       argv[1]);
1256 		return -EINVAL;
1257 	}
1258 
1259 	r = action_dev(m, dev, action);
1260 
1261 	dm_put_device(ti, dev);
1262 
1263 	return r;
1264 
1265 error:
1266 	DMWARN("Unrecognised multipath message received.");
1267 	return -EINVAL;
1268 }
1269 
1270 /*-----------------------------------------------------------------
1271  * Module setup
1272  *---------------------------------------------------------------*/
1273 static struct target_type multipath_target = {
1274 	.name = "multipath",
1275 	.version = {1, 0, 4},
1276 	.module = THIS_MODULE,
1277 	.ctr = multipath_ctr,
1278 	.dtr = multipath_dtr,
1279 	.map = multipath_map,
1280 	.end_io = multipath_end_io,
1281 	.presuspend = multipath_presuspend,
1282 	.resume = multipath_resume,
1283 	.status = multipath_status,
1284 	.message = multipath_message,
1285 };
1286 
1287 static int __init dm_multipath_init(void)
1288 {
1289 	int r;
1290 
1291 	/* allocate a slab for the dm_ios */
1292 	_mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1293 					0, 0, NULL, NULL);
1294 	if (!_mpio_cache)
1295 		return -ENOMEM;
1296 
1297 	r = dm_register_target(&multipath_target);
1298 	if (r < 0) {
1299 		DMERR("%s: register failed %d", multipath_target.name, r);
1300 		kmem_cache_destroy(_mpio_cache);
1301 		return -EINVAL;
1302 	}
1303 
1304 	kmultipathd = create_workqueue("kmpathd");
1305 	if (!kmultipathd) {
1306 		DMERR("%s: failed to create workqueue kmpathd",
1307 				multipath_target.name);
1308 		dm_unregister_target(&multipath_target);
1309 		kmem_cache_destroy(_mpio_cache);
1310 		return -ENOMEM;
1311 	}
1312 
1313 	DMINFO("dm-multipath version %u.%u.%u loaded",
1314 	       multipath_target.version[0], multipath_target.version[1],
1315 	       multipath_target.version[2]);
1316 
1317 	return r;
1318 }
1319 
1320 static void __exit dm_multipath_exit(void)
1321 {
1322 	int r;
1323 
1324 	destroy_workqueue(kmultipathd);
1325 
1326 	r = dm_unregister_target(&multipath_target);
1327 	if (r < 0)
1328 		DMERR("%s: target unregister failed %d",
1329 		      multipath_target.name, r);
1330 	kmem_cache_destroy(_mpio_cache);
1331 }
1332 
1333 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1334 
1335 module_init(dm_multipath_init);
1336 module_exit(dm_multipath_exit);
1337 
1338 MODULE_DESCRIPTION(DM_NAME " multipath target");
1339 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1340 MODULE_LICENSE("GPL");
1341