xref: /openbmc/linux/drivers/md/dm-mpath.c (revision c922d5f7f5457da9e9b5a26dd53e2dcef6ca2f7d)
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 DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
26 
27 /* Path properties */
28 struct pgpath {
29 	struct list_head list;
30 
31 	struct priority_group *pg;	/* Owning PG */
32 	unsigned fail_count;		/* Cumulative failure count */
33 
34 	struct dm_path path;
35 };
36 
37 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
38 
39 /*
40  * Paths are grouped into Priority Groups and numbered from 1 upwards.
41  * Each has a path selector which controls which path gets used.
42  */
43 struct priority_group {
44 	struct list_head list;
45 
46 	struct multipath *m;		/* Owning multipath instance */
47 	struct path_selector ps;
48 
49 	unsigned pg_num;		/* Reference number */
50 	unsigned bypassed;		/* Temporarily bypass this PG? */
51 
52 	unsigned nr_pgpaths;		/* Number of paths in PG */
53 	struct list_head pgpaths;
54 };
55 
56 /* Multipath context */
57 struct multipath {
58 	struct list_head list;
59 	struct dm_target *ti;
60 
61 	spinlock_t lock;
62 
63 	struct hw_handler hw_handler;
64 	unsigned nr_priority_groups;
65 	struct list_head priority_groups;
66 	unsigned pg_init_required;	/* pg_init needs calling? */
67 	unsigned pg_init_in_progress;	/* Only one pg_init allowed at once */
68 
69 	unsigned nr_valid_paths;	/* Total number of usable paths */
70 	struct pgpath *current_pgpath;
71 	struct priority_group *current_pg;
72 	struct priority_group *next_pg;	/* Switch to this PG if set */
73 	unsigned repeat_count;		/* I/Os left before calling PS again */
74 
75 	unsigned queue_io;		/* Must we queue all I/O? */
76 	unsigned queue_if_no_path;	/* Queue I/O if last path fails? */
77 	unsigned saved_queue_if_no_path;/* Saved state during suspension */
78 
79 	struct work_struct process_queued_ios;
80 	struct bio_list queued_ios;
81 	unsigned queue_size;
82 
83 	struct work_struct trigger_event;
84 
85 	/*
86 	 * We must use a mempool of mpath_io structs so that we
87 	 * can resubmit bios on error.
88 	 */
89 	mempool_t *mpio_pool;
90 };
91 
92 /*
93  * Context information attached to each bio we process.
94  */
95 struct mpath_io {
96 	struct pgpath *pgpath;
97 	struct dm_bio_details details;
98 };
99 
100 typedef int (*action_fn) (struct pgpath *pgpath);
101 
102 #define MIN_IOS 256	/* Mempool size */
103 
104 static struct kmem_cache *_mpio_cache;
105 
106 struct workqueue_struct *kmultipathd;
107 static void process_queued_ios(struct work_struct *work);
108 static void trigger_event(struct work_struct *work);
109 
110 
111 /*-----------------------------------------------
112  * Allocation routines
113  *-----------------------------------------------*/
114 
115 static struct pgpath *alloc_pgpath(void)
116 {
117 	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
118 
119 	if (pgpath)
120 		pgpath->path.is_active = 1;
121 
122 	return pgpath;
123 }
124 
125 static inline void free_pgpath(struct pgpath *pgpath)
126 {
127 	kfree(pgpath);
128 }
129 
130 static struct priority_group *alloc_priority_group(void)
131 {
132 	struct priority_group *pg;
133 
134 	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
135 
136 	if (pg)
137 		INIT_LIST_HEAD(&pg->pgpaths);
138 
139 	return pg;
140 }
141 
142 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
143 {
144 	struct pgpath *pgpath, *tmp;
145 
146 	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
147 		list_del(&pgpath->list);
148 		dm_put_device(ti, pgpath->path.dev);
149 		free_pgpath(pgpath);
150 	}
151 }
152 
153 static void free_priority_group(struct priority_group *pg,
154 				struct dm_target *ti)
155 {
156 	struct path_selector *ps = &pg->ps;
157 
158 	if (ps->type) {
159 		ps->type->destroy(ps);
160 		dm_put_path_selector(ps->type);
161 	}
162 
163 	free_pgpaths(&pg->pgpaths, ti);
164 	kfree(pg);
165 }
166 
167 static struct multipath *alloc_multipath(struct dm_target *ti)
168 {
169 	struct multipath *m;
170 
171 	m = kzalloc(sizeof(*m), GFP_KERNEL);
172 	if (m) {
173 		INIT_LIST_HEAD(&m->priority_groups);
174 		spin_lock_init(&m->lock);
175 		m->queue_io = 1;
176 		INIT_WORK(&m->process_queued_ios, process_queued_ios);
177 		INIT_WORK(&m->trigger_event, trigger_event);
178 		m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
179 		if (!m->mpio_pool) {
180 			kfree(m);
181 			return NULL;
182 		}
183 		m->ti = ti;
184 		ti->private = m;
185 	}
186 
187 	return m;
188 }
189 
190 static void free_multipath(struct multipath *m)
191 {
192 	struct priority_group *pg, *tmp;
193 	struct hw_handler *hwh = &m->hw_handler;
194 
195 	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
196 		list_del(&pg->list);
197 		free_priority_group(pg, m->ti);
198 	}
199 
200 	if (hwh->type) {
201 		hwh->type->destroy(hwh);
202 		dm_put_hw_handler(hwh->type);
203 	}
204 
205 	mempool_destroy(m->mpio_pool);
206 	kfree(m);
207 }
208 
209 
210 /*-----------------------------------------------
211  * Path selection
212  *-----------------------------------------------*/
213 
214 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
215 {
216 	struct hw_handler *hwh = &m->hw_handler;
217 
218 	m->current_pg = pgpath->pg;
219 
220 	/* Must we initialise the PG first, and queue I/O till it's ready? */
221 	if (hwh->type && hwh->type->pg_init) {
222 		m->pg_init_required = 1;
223 		m->queue_io = 1;
224 	} else {
225 		m->pg_init_required = 0;
226 		m->queue_io = 0;
227 	}
228 }
229 
230 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
231 {
232 	struct dm_path *path;
233 
234 	path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
235 	if (!path)
236 		return -ENXIO;
237 
238 	m->current_pgpath = path_to_pgpath(path);
239 
240 	if (m->current_pg != pg)
241 		__switch_pg(m, m->current_pgpath);
242 
243 	return 0;
244 }
245 
246 static void __choose_pgpath(struct multipath *m)
247 {
248 	struct priority_group *pg;
249 	unsigned bypassed = 1;
250 
251 	if (!m->nr_valid_paths)
252 		goto failed;
253 
254 	/* Were we instructed to switch PG? */
255 	if (m->next_pg) {
256 		pg = m->next_pg;
257 		m->next_pg = NULL;
258 		if (!__choose_path_in_pg(m, pg))
259 			return;
260 	}
261 
262 	/* Don't change PG until it has no remaining paths */
263 	if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
264 		return;
265 
266 	/*
267 	 * Loop through priority groups until we find a valid path.
268 	 * First time we skip PGs marked 'bypassed'.
269 	 * Second time we only try the ones we skipped.
270 	 */
271 	do {
272 		list_for_each_entry(pg, &m->priority_groups, list) {
273 			if (pg->bypassed == bypassed)
274 				continue;
275 			if (!__choose_path_in_pg(m, pg))
276 				return;
277 		}
278 	} while (bypassed--);
279 
280 failed:
281 	m->current_pgpath = NULL;
282 	m->current_pg = NULL;
283 }
284 
285 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
286 		  unsigned was_queued)
287 {
288 	int r = 1;
289 	unsigned long flags;
290 	struct pgpath *pgpath;
291 
292 	spin_lock_irqsave(&m->lock, flags);
293 
294 	/* Do we need to select a new pgpath? */
295 	if (!m->current_pgpath ||
296 	    (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
297 		__choose_pgpath(m);
298 
299 	pgpath = m->current_pgpath;
300 
301 	if (was_queued)
302 		m->queue_size--;
303 
304 	if ((pgpath && m->queue_io) ||
305 	    (!pgpath && m->queue_if_no_path)) {
306 		/* Queue for the daemon to resubmit */
307 		bio_list_add(&m->queued_ios, bio);
308 		m->queue_size++;
309 		if ((m->pg_init_required && !m->pg_init_in_progress) ||
310 		    !m->queue_io)
311 			queue_work(kmultipathd, &m->process_queued_ios);
312 		pgpath = NULL;
313 		r = 0;
314 	} else if (!pgpath)
315 		r = -EIO;		/* Failed */
316 	else
317 		bio->bi_bdev = pgpath->path.dev->bdev;
318 
319 	mpio->pgpath = pgpath;
320 
321 	spin_unlock_irqrestore(&m->lock, flags);
322 
323 	return r;
324 }
325 
326 /*
327  * If we run out of usable paths, should we queue I/O or error it?
328  */
329 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
330 			    unsigned save_old_value)
331 {
332 	unsigned long flags;
333 
334 	spin_lock_irqsave(&m->lock, flags);
335 
336 	if (save_old_value)
337 		m->saved_queue_if_no_path = m->queue_if_no_path;
338 	else
339 		m->saved_queue_if_no_path = queue_if_no_path;
340 	m->queue_if_no_path = queue_if_no_path;
341 	if (!m->queue_if_no_path && m->queue_size)
342 		queue_work(kmultipathd, &m->process_queued_ios);
343 
344 	spin_unlock_irqrestore(&m->lock, flags);
345 
346 	return 0;
347 }
348 
349 /*-----------------------------------------------------------------
350  * The multipath daemon is responsible for resubmitting queued ios.
351  *---------------------------------------------------------------*/
352 
353 static void dispatch_queued_ios(struct multipath *m)
354 {
355 	int r;
356 	unsigned long flags;
357 	struct bio *bio = NULL, *next;
358 	struct mpath_io *mpio;
359 	union map_info *info;
360 
361 	spin_lock_irqsave(&m->lock, flags);
362 	bio = bio_list_get(&m->queued_ios);
363 	spin_unlock_irqrestore(&m->lock, flags);
364 
365 	while (bio) {
366 		next = bio->bi_next;
367 		bio->bi_next = NULL;
368 
369 		info = dm_get_mapinfo(bio);
370 		mpio = info->ptr;
371 
372 		r = map_io(m, bio, mpio, 1);
373 		if (r < 0)
374 			bio_endio(bio, bio->bi_size, r);
375 		else if (r == 1)
376 			generic_make_request(bio);
377 
378 		bio = next;
379 	}
380 }
381 
382 static void process_queued_ios(struct work_struct *work)
383 {
384 	struct multipath *m =
385 		container_of(work, struct multipath, process_queued_ios);
386 	struct hw_handler *hwh = &m->hw_handler;
387 	struct pgpath *pgpath = NULL;
388 	unsigned init_required = 0, must_queue = 1;
389 	unsigned long flags;
390 
391 	spin_lock_irqsave(&m->lock, flags);
392 
393 	if (!m->queue_size)
394 		goto out;
395 
396 	if (!m->current_pgpath)
397 		__choose_pgpath(m);
398 
399 	pgpath = m->current_pgpath;
400 
401 	if ((pgpath && !m->queue_io) ||
402 	    (!pgpath && !m->queue_if_no_path))
403 		must_queue = 0;
404 
405 	if (m->pg_init_required && !m->pg_init_in_progress) {
406 		m->pg_init_required = 0;
407 		m->pg_init_in_progress = 1;
408 		init_required = 1;
409 	}
410 
411 out:
412 	spin_unlock_irqrestore(&m->lock, flags);
413 
414 	if (init_required)
415 		hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
416 
417 	if (!must_queue)
418 		dispatch_queued_ios(m);
419 }
420 
421 /*
422  * An event is triggered whenever a path is taken out of use.
423  * Includes path failure and PG bypass.
424  */
425 static void trigger_event(struct work_struct *work)
426 {
427 	struct multipath *m =
428 		container_of(work, struct multipath, trigger_event);
429 
430 	dm_table_event(m->ti->table);
431 }
432 
433 /*-----------------------------------------------------------------
434  * Constructor/argument parsing:
435  * <#multipath feature args> [<arg>]*
436  * <#hw_handler args> [hw_handler [<arg>]*]
437  * <#priority groups>
438  * <initial priority group>
439  *     [<selector> <#selector args> [<arg>]*
440  *      <#paths> <#per-path selector args>
441  *         [<path> [<arg>]* ]+ ]+
442  *---------------------------------------------------------------*/
443 struct param {
444 	unsigned min;
445 	unsigned max;
446 	char *error;
447 };
448 
449 static int read_param(struct param *param, char *str, unsigned *v, char **error)
450 {
451 	if (!str ||
452 	    (sscanf(str, "%u", v) != 1) ||
453 	    (*v < param->min) ||
454 	    (*v > param->max)) {
455 		*error = param->error;
456 		return -EINVAL;
457 	}
458 
459 	return 0;
460 }
461 
462 struct arg_set {
463 	unsigned argc;
464 	char **argv;
465 };
466 
467 static char *shift(struct arg_set *as)
468 {
469 	char *r;
470 
471 	if (as->argc) {
472 		as->argc--;
473 		r = *as->argv;
474 		as->argv++;
475 		return r;
476 	}
477 
478 	return NULL;
479 }
480 
481 static void consume(struct arg_set *as, unsigned n)
482 {
483 	BUG_ON (as->argc < n);
484 	as->argc -= n;
485 	as->argv += n;
486 }
487 
488 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
489 			       struct dm_target *ti)
490 {
491 	int r;
492 	struct path_selector_type *pst;
493 	unsigned ps_argc;
494 
495 	static struct param _params[] = {
496 		{0, 1024, "invalid number of path selector args"},
497 	};
498 
499 	pst = dm_get_path_selector(shift(as));
500 	if (!pst) {
501 		ti->error = "unknown path selector type";
502 		return -EINVAL;
503 	}
504 
505 	r = read_param(_params, shift(as), &ps_argc, &ti->error);
506 	if (r)
507 		return -EINVAL;
508 
509 	r = pst->create(&pg->ps, ps_argc, as->argv);
510 	if (r) {
511 		dm_put_path_selector(pst);
512 		ti->error = "path selector constructor failed";
513 		return r;
514 	}
515 
516 	pg->ps.type = pst;
517 	consume(as, ps_argc);
518 
519 	return 0;
520 }
521 
522 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
523 			       struct dm_target *ti)
524 {
525 	int r;
526 	struct pgpath *p;
527 
528 	/* we need at least a path arg */
529 	if (as->argc < 1) {
530 		ti->error = "no device given";
531 		return NULL;
532 	}
533 
534 	p = alloc_pgpath();
535 	if (!p)
536 		return NULL;
537 
538 	r = dm_get_device(ti, shift(as), ti->begin, ti->len,
539 			  dm_table_get_mode(ti->table), &p->path.dev);
540 	if (r) {
541 		ti->error = "error getting device";
542 		goto bad;
543 	}
544 
545 	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
546 	if (r) {
547 		dm_put_device(ti, p->path.dev);
548 		goto bad;
549 	}
550 
551 	return p;
552 
553  bad:
554 	free_pgpath(p);
555 	return NULL;
556 }
557 
558 static struct priority_group *parse_priority_group(struct arg_set *as,
559 						   struct multipath *m)
560 {
561 	static struct param _params[] = {
562 		{1, 1024, "invalid number of paths"},
563 		{0, 1024, "invalid number of selector args"}
564 	};
565 
566 	int r;
567 	unsigned i, nr_selector_args, nr_params;
568 	struct priority_group *pg;
569 	struct dm_target *ti = m->ti;
570 
571 	if (as->argc < 2) {
572 		as->argc = 0;
573 		ti->error = "not enough priority group aruments";
574 		return NULL;
575 	}
576 
577 	pg = alloc_priority_group();
578 	if (!pg) {
579 		ti->error = "couldn't allocate priority group";
580 		return NULL;
581 	}
582 	pg->m = m;
583 
584 	r = parse_path_selector(as, pg, ti);
585 	if (r)
586 		goto bad;
587 
588 	/*
589 	 * read the paths
590 	 */
591 	r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
592 	if (r)
593 		goto bad;
594 
595 	r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
596 	if (r)
597 		goto bad;
598 
599 	nr_params = 1 + nr_selector_args;
600 	for (i = 0; i < pg->nr_pgpaths; i++) {
601 		struct pgpath *pgpath;
602 		struct arg_set path_args;
603 
604 		if (as->argc < nr_params)
605 			goto bad;
606 
607 		path_args.argc = nr_params;
608 		path_args.argv = as->argv;
609 
610 		pgpath = parse_path(&path_args, &pg->ps, ti);
611 		if (!pgpath)
612 			goto bad;
613 
614 		pgpath->pg = pg;
615 		list_add_tail(&pgpath->list, &pg->pgpaths);
616 		consume(as, nr_params);
617 	}
618 
619 	return pg;
620 
621  bad:
622 	free_priority_group(pg, ti);
623 	return NULL;
624 }
625 
626 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
627 {
628 	int r;
629 	struct hw_handler_type *hwht;
630 	unsigned hw_argc;
631 	struct dm_target *ti = m->ti;
632 
633 	static struct param _params[] = {
634 		{0, 1024, "invalid number of hardware handler args"},
635 	};
636 
637 	r = read_param(_params, shift(as), &hw_argc, &ti->error);
638 	if (r)
639 		return -EINVAL;
640 
641 	if (!hw_argc)
642 		return 0;
643 
644 	hwht = dm_get_hw_handler(shift(as));
645 	if (!hwht) {
646 		ti->error = "unknown hardware handler type";
647 		return -EINVAL;
648 	}
649 
650 	r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
651 	if (r) {
652 		dm_put_hw_handler(hwht);
653 		ti->error = "hardware handler constructor failed";
654 		return r;
655 	}
656 
657 	m->hw_handler.type = hwht;
658 	consume(as, hw_argc - 1);
659 
660 	return 0;
661 }
662 
663 static int parse_features(struct arg_set *as, struct multipath *m)
664 {
665 	int r;
666 	unsigned argc;
667 	struct dm_target *ti = m->ti;
668 
669 	static struct param _params[] = {
670 		{0, 1, "invalid number of feature args"},
671 	};
672 
673 	r = read_param(_params, shift(as), &argc, &ti->error);
674 	if (r)
675 		return -EINVAL;
676 
677 	if (!argc)
678 		return 0;
679 
680 	if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
681 		return queue_if_no_path(m, 1, 0);
682 	else {
683 		ti->error = "Unrecognised multipath feature request";
684 		return -EINVAL;
685 	}
686 }
687 
688 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
689 			 char **argv)
690 {
691 	/* target parameters */
692 	static struct param _params[] = {
693 		{1, 1024, "invalid number of priority groups"},
694 		{1, 1024, "invalid initial priority group number"},
695 	};
696 
697 	int r;
698 	struct multipath *m;
699 	struct arg_set as;
700 	unsigned pg_count = 0;
701 	unsigned next_pg_num;
702 
703 	as.argc = argc;
704 	as.argv = argv;
705 
706 	m = alloc_multipath(ti);
707 	if (!m) {
708 		ti->error = "can't allocate multipath";
709 		return -EINVAL;
710 	}
711 
712 	r = parse_features(&as, m);
713 	if (r)
714 		goto bad;
715 
716 	r = parse_hw_handler(&as, m);
717 	if (r)
718 		goto bad;
719 
720 	r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
721 	if (r)
722 		goto bad;
723 
724 	r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
725 	if (r)
726 		goto bad;
727 
728 	/* parse the priority groups */
729 	while (as.argc) {
730 		struct priority_group *pg;
731 
732 		pg = parse_priority_group(&as, m);
733 		if (!pg) {
734 			r = -EINVAL;
735 			goto bad;
736 		}
737 
738 		m->nr_valid_paths += pg->nr_pgpaths;
739 		list_add_tail(&pg->list, &m->priority_groups);
740 		pg_count++;
741 		pg->pg_num = pg_count;
742 		if (!--next_pg_num)
743 			m->next_pg = pg;
744 	}
745 
746 	if (pg_count != m->nr_priority_groups) {
747 		ti->error = "priority group count mismatch";
748 		r = -EINVAL;
749 		goto bad;
750 	}
751 
752 	return 0;
753 
754  bad:
755 	free_multipath(m);
756 	return r;
757 }
758 
759 static void multipath_dtr(struct dm_target *ti)
760 {
761 	struct multipath *m = (struct multipath *) ti->private;
762 
763 	flush_workqueue(kmultipathd);
764 	free_multipath(m);
765 }
766 
767 /*
768  * Map bios, recording original fields for later in case we have to resubmit
769  */
770 static int multipath_map(struct dm_target *ti, struct bio *bio,
771 			 union map_info *map_context)
772 {
773 	int r;
774 	struct mpath_io *mpio;
775 	struct multipath *m = (struct multipath *) ti->private;
776 
777 	if (bio_barrier(bio))
778 		return -EOPNOTSUPP;
779 
780 	mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
781 	dm_bio_record(&mpio->details, bio);
782 
783 	map_context->ptr = mpio;
784 	bio->bi_rw |= (1 << BIO_RW_FAILFAST);
785 	r = map_io(m, bio, mpio, 0);
786 	if (r < 0)
787 		mempool_free(mpio, m->mpio_pool);
788 
789 	return r;
790 }
791 
792 /*
793  * Take a path out of use.
794  */
795 static int fail_path(struct pgpath *pgpath)
796 {
797 	unsigned long flags;
798 	struct multipath *m = pgpath->pg->m;
799 
800 	spin_lock_irqsave(&m->lock, flags);
801 
802 	if (!pgpath->path.is_active)
803 		goto out;
804 
805 	DMWARN("Failing path %s.", pgpath->path.dev->name);
806 
807 	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
808 	pgpath->path.is_active = 0;
809 	pgpath->fail_count++;
810 
811 	m->nr_valid_paths--;
812 
813 	if (pgpath == m->current_pgpath)
814 		m->current_pgpath = NULL;
815 
816 	queue_work(kmultipathd, &m->trigger_event);
817 
818 out:
819 	spin_unlock_irqrestore(&m->lock, flags);
820 
821 	return 0;
822 }
823 
824 /*
825  * Reinstate a previously-failed path
826  */
827 static int reinstate_path(struct pgpath *pgpath)
828 {
829 	int r = 0;
830 	unsigned long flags;
831 	struct multipath *m = pgpath->pg->m;
832 
833 	spin_lock_irqsave(&m->lock, flags);
834 
835 	if (pgpath->path.is_active)
836 		goto out;
837 
838 	if (!pgpath->pg->ps.type) {
839 		DMWARN("Reinstate path not supported by path selector %s",
840 		       pgpath->pg->ps.type->name);
841 		r = -EINVAL;
842 		goto out;
843 	}
844 
845 	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
846 	if (r)
847 		goto out;
848 
849 	pgpath->path.is_active = 1;
850 
851 	m->current_pgpath = NULL;
852 	if (!m->nr_valid_paths++ && m->queue_size)
853 		queue_work(kmultipathd, &m->process_queued_ios);
854 
855 	queue_work(kmultipathd, &m->trigger_event);
856 
857 out:
858 	spin_unlock_irqrestore(&m->lock, flags);
859 
860 	return r;
861 }
862 
863 /*
864  * Fail or reinstate all paths that match the provided struct dm_dev.
865  */
866 static int action_dev(struct multipath *m, struct dm_dev *dev,
867 		      action_fn action)
868 {
869 	int r = 0;
870 	struct pgpath *pgpath;
871 	struct priority_group *pg;
872 
873 	list_for_each_entry(pg, &m->priority_groups, list) {
874 		list_for_each_entry(pgpath, &pg->pgpaths, list) {
875 			if (pgpath->path.dev == dev)
876 				r = action(pgpath);
877 		}
878 	}
879 
880 	return r;
881 }
882 
883 /*
884  * Temporarily try to avoid having to use the specified PG
885  */
886 static void bypass_pg(struct multipath *m, struct priority_group *pg,
887 		      int bypassed)
888 {
889 	unsigned long flags;
890 
891 	spin_lock_irqsave(&m->lock, flags);
892 
893 	pg->bypassed = bypassed;
894 	m->current_pgpath = NULL;
895 	m->current_pg = NULL;
896 
897 	spin_unlock_irqrestore(&m->lock, flags);
898 
899 	queue_work(kmultipathd, &m->trigger_event);
900 }
901 
902 /*
903  * Switch to using the specified PG from the next I/O that gets mapped
904  */
905 static int switch_pg_num(struct multipath *m, const char *pgstr)
906 {
907 	struct priority_group *pg;
908 	unsigned pgnum;
909 	unsigned long flags;
910 
911 	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
912 	    (pgnum > m->nr_priority_groups)) {
913 		DMWARN("invalid PG number supplied to switch_pg_num");
914 		return -EINVAL;
915 	}
916 
917 	spin_lock_irqsave(&m->lock, flags);
918 	list_for_each_entry(pg, &m->priority_groups, list) {
919 		pg->bypassed = 0;
920 		if (--pgnum)
921 			continue;
922 
923 		m->current_pgpath = NULL;
924 		m->current_pg = NULL;
925 		m->next_pg = pg;
926 	}
927 	spin_unlock_irqrestore(&m->lock, flags);
928 
929 	queue_work(kmultipathd, &m->trigger_event);
930 	return 0;
931 }
932 
933 /*
934  * Set/clear bypassed status of a PG.
935  * PGs are numbered upwards from 1 in the order they were declared.
936  */
937 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
938 {
939 	struct priority_group *pg;
940 	unsigned pgnum;
941 
942 	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
943 	    (pgnum > m->nr_priority_groups)) {
944 		DMWARN("invalid PG number supplied to bypass_pg");
945 		return -EINVAL;
946 	}
947 
948 	list_for_each_entry(pg, &m->priority_groups, list) {
949 		if (!--pgnum)
950 			break;
951 	}
952 
953 	bypass_pg(m, pg, bypassed);
954 	return 0;
955 }
956 
957 /*
958  * pg_init must call this when it has completed its initialisation
959  */
960 void dm_pg_init_complete(struct dm_path *path, unsigned err_flags)
961 {
962 	struct pgpath *pgpath = path_to_pgpath(path);
963 	struct priority_group *pg = pgpath->pg;
964 	struct multipath *m = pg->m;
965 	unsigned long flags;
966 
967 	/* We insist on failing the path if the PG is already bypassed. */
968 	if (err_flags && pg->bypassed)
969 		err_flags |= MP_FAIL_PATH;
970 
971 	if (err_flags & MP_FAIL_PATH)
972 		fail_path(pgpath);
973 
974 	if (err_flags & MP_BYPASS_PG)
975 		bypass_pg(m, pg, 1);
976 
977 	spin_lock_irqsave(&m->lock, flags);
978 	if (err_flags) {
979 		m->current_pgpath = NULL;
980 		m->current_pg = NULL;
981 	} else if (!m->pg_init_required)
982 		m->queue_io = 0;
983 
984 	m->pg_init_in_progress = 0;
985 	queue_work(kmultipathd, &m->process_queued_ios);
986 	spin_unlock_irqrestore(&m->lock, flags);
987 }
988 
989 /*
990  * end_io handling
991  */
992 static int do_end_io(struct multipath *m, struct bio *bio,
993 		     int error, struct mpath_io *mpio)
994 {
995 	struct hw_handler *hwh = &m->hw_handler;
996 	unsigned err_flags = MP_FAIL_PATH;	/* Default behavior */
997 	unsigned long flags;
998 
999 	if (!error)
1000 		return 0;	/* I/O complete */
1001 
1002 	if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1003 		return error;
1004 
1005 	if (error == -EOPNOTSUPP)
1006 		return error;
1007 
1008 	spin_lock_irqsave(&m->lock, flags);
1009 	if (!m->nr_valid_paths) {
1010 		if (!m->queue_if_no_path) {
1011 			spin_unlock_irqrestore(&m->lock, flags);
1012 			return -EIO;
1013 		} else {
1014 			spin_unlock_irqrestore(&m->lock, flags);
1015 			goto requeue;
1016 		}
1017 	}
1018 	spin_unlock_irqrestore(&m->lock, flags);
1019 
1020 	if (hwh->type && hwh->type->error)
1021 		err_flags = hwh->type->error(hwh, bio);
1022 
1023 	if (mpio->pgpath) {
1024 		if (err_flags & MP_FAIL_PATH)
1025 			fail_path(mpio->pgpath);
1026 
1027 		if (err_flags & MP_BYPASS_PG)
1028 			bypass_pg(m, mpio->pgpath->pg, 1);
1029 	}
1030 
1031 	if (err_flags & MP_ERROR_IO)
1032 		return -EIO;
1033 
1034       requeue:
1035 	dm_bio_restore(&mpio->details, bio);
1036 
1037 	/* queue for the daemon to resubmit or fail */
1038 	spin_lock_irqsave(&m->lock, flags);
1039 	bio_list_add(&m->queued_ios, bio);
1040 	m->queue_size++;
1041 	if (!m->queue_io)
1042 		queue_work(kmultipathd, &m->process_queued_ios);
1043 	spin_unlock_irqrestore(&m->lock, flags);
1044 
1045 	return 1;	/* io not complete */
1046 }
1047 
1048 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1049 			    int error, union map_info *map_context)
1050 {
1051 	struct multipath *m = (struct multipath *) ti->private;
1052 	struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1053 	struct pgpath *pgpath = mpio->pgpath;
1054 	struct path_selector *ps;
1055 	int r;
1056 
1057 	r  = do_end_io(m, bio, error, mpio);
1058 	if (pgpath) {
1059 		ps = &pgpath->pg->ps;
1060 		if (ps->type->end_io)
1061 			ps->type->end_io(ps, &pgpath->path);
1062 	}
1063 	if (r <= 0)
1064 		mempool_free(mpio, m->mpio_pool);
1065 
1066 	return r;
1067 }
1068 
1069 /*
1070  * Suspend can't complete until all the I/O is processed so if
1071  * the last path fails we must error any remaining I/O.
1072  * Note that if the freeze_bdev fails while suspending, the
1073  * queue_if_no_path state is lost - userspace should reset it.
1074  */
1075 static void multipath_presuspend(struct dm_target *ti)
1076 {
1077 	struct multipath *m = (struct multipath *) ti->private;
1078 
1079 	queue_if_no_path(m, 0, 1);
1080 }
1081 
1082 /*
1083  * Restore the queue_if_no_path setting.
1084  */
1085 static void multipath_resume(struct dm_target *ti)
1086 {
1087 	struct multipath *m = (struct multipath *) ti->private;
1088 	unsigned long flags;
1089 
1090 	spin_lock_irqsave(&m->lock, flags);
1091 	m->queue_if_no_path = m->saved_queue_if_no_path;
1092 	spin_unlock_irqrestore(&m->lock, flags);
1093 }
1094 
1095 /*
1096  * Info output has the following format:
1097  * num_multipath_feature_args [multipath_feature_args]*
1098  * num_handler_status_args [handler_status_args]*
1099  * num_groups init_group_number
1100  *            [A|D|E num_ps_status_args [ps_status_args]*
1101  *             num_paths num_selector_args
1102  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1103  *
1104  * Table output has the following format (identical to the constructor string):
1105  * num_feature_args [features_args]*
1106  * num_handler_args hw_handler [hw_handler_args]*
1107  * num_groups init_group_number
1108  *     [priority selector-name num_ps_args [ps_args]*
1109  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1110  */
1111 static int multipath_status(struct dm_target *ti, status_type_t type,
1112 			    char *result, unsigned int maxlen)
1113 {
1114 	int sz = 0;
1115 	unsigned long flags;
1116 	struct multipath *m = (struct multipath *) ti->private;
1117 	struct hw_handler *hwh = &m->hw_handler;
1118 	struct priority_group *pg;
1119 	struct pgpath *p;
1120 	unsigned pg_num;
1121 	char state;
1122 
1123 	spin_lock_irqsave(&m->lock, flags);
1124 
1125 	/* Features */
1126 	if (type == STATUSTYPE_INFO)
1127 		DMEMIT("1 %u ", m->queue_size);
1128 	else if (m->queue_if_no_path)
1129 		DMEMIT("1 queue_if_no_path ");
1130 	else
1131 		DMEMIT("0 ");
1132 
1133 	if (hwh->type && hwh->type->status)
1134 		sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1135 	else if (!hwh->type || type == STATUSTYPE_INFO)
1136 		DMEMIT("0 ");
1137 	else
1138 		DMEMIT("1 %s ", hwh->type->name);
1139 
1140 	DMEMIT("%u ", m->nr_priority_groups);
1141 
1142 	if (m->next_pg)
1143 		pg_num = m->next_pg->pg_num;
1144 	else if (m->current_pg)
1145 		pg_num = m->current_pg->pg_num;
1146 	else
1147 			pg_num = 1;
1148 
1149 	DMEMIT("%u ", pg_num);
1150 
1151 	switch (type) {
1152 	case STATUSTYPE_INFO:
1153 		list_for_each_entry(pg, &m->priority_groups, list) {
1154 			if (pg->bypassed)
1155 				state = 'D';	/* Disabled */
1156 			else if (pg == m->current_pg)
1157 				state = 'A';	/* Currently Active */
1158 			else
1159 				state = 'E';	/* Enabled */
1160 
1161 			DMEMIT("%c ", state);
1162 
1163 			if (pg->ps.type->status)
1164 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1165 							  result + sz,
1166 							  maxlen - sz);
1167 			else
1168 				DMEMIT("0 ");
1169 
1170 			DMEMIT("%u %u ", pg->nr_pgpaths,
1171 			       pg->ps.type->info_args);
1172 
1173 			list_for_each_entry(p, &pg->pgpaths, list) {
1174 				DMEMIT("%s %s %u ", p->path.dev->name,
1175 				       p->path.is_active ? "A" : "F",
1176 				       p->fail_count);
1177 				if (pg->ps.type->status)
1178 					sz += pg->ps.type->status(&pg->ps,
1179 					      &p->path, type, result + sz,
1180 					      maxlen - sz);
1181 			}
1182 		}
1183 		break;
1184 
1185 	case STATUSTYPE_TABLE:
1186 		list_for_each_entry(pg, &m->priority_groups, list) {
1187 			DMEMIT("%s ", pg->ps.type->name);
1188 
1189 			if (pg->ps.type->status)
1190 				sz += pg->ps.type->status(&pg->ps, NULL, type,
1191 							  result + sz,
1192 							  maxlen - sz);
1193 			else
1194 				DMEMIT("0 ");
1195 
1196 			DMEMIT("%u %u ", pg->nr_pgpaths,
1197 			       pg->ps.type->table_args);
1198 
1199 			list_for_each_entry(p, &pg->pgpaths, list) {
1200 				DMEMIT("%s ", p->path.dev->name);
1201 				if (pg->ps.type->status)
1202 					sz += pg->ps.type->status(&pg->ps,
1203 					      &p->path, type, result + sz,
1204 					      maxlen - sz);
1205 			}
1206 		}
1207 		break;
1208 	}
1209 
1210 	spin_unlock_irqrestore(&m->lock, flags);
1211 
1212 	return 0;
1213 }
1214 
1215 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1216 {
1217 	int r;
1218 	struct dm_dev *dev;
1219 	struct multipath *m = (struct multipath *) ti->private;
1220 	action_fn action;
1221 
1222 	if (argc == 1) {
1223 		if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1224 			return queue_if_no_path(m, 1, 0);
1225 		else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1226 			return queue_if_no_path(m, 0, 0);
1227 	}
1228 
1229 	if (argc != 2)
1230 		goto error;
1231 
1232 	if (!strnicmp(argv[0], MESG_STR("disable_group")))
1233 		return bypass_pg_num(m, argv[1], 1);
1234 	else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1235 		return bypass_pg_num(m, argv[1], 0);
1236 	else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1237 		return switch_pg_num(m, argv[1]);
1238 	else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1239 		action = reinstate_path;
1240 	else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1241 		action = fail_path;
1242 	else
1243 		goto error;
1244 
1245 	r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1246 			  dm_table_get_mode(ti->table), &dev);
1247 	if (r) {
1248 		DMWARN("message: error getting device %s",
1249 		       argv[1]);
1250 		return -EINVAL;
1251 	}
1252 
1253 	r = action_dev(m, dev, action);
1254 
1255 	dm_put_device(ti, dev);
1256 
1257 	return r;
1258 
1259 error:
1260 	DMWARN("Unrecognised multipath message received.");
1261 	return -EINVAL;
1262 }
1263 
1264 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1265 			   struct file *filp, unsigned int cmd,
1266 			   unsigned long arg)
1267 {
1268 	struct multipath *m = (struct multipath *) ti->private;
1269 	struct block_device *bdev = NULL;
1270 	unsigned long flags;
1271 	struct file fake_file = {};
1272 	struct dentry fake_dentry = {};
1273 	int r = 0;
1274 
1275 	fake_file.f_dentry = &fake_dentry;
1276 
1277 	spin_lock_irqsave(&m->lock, flags);
1278 
1279 	if (!m->current_pgpath)
1280 		__choose_pgpath(m);
1281 
1282 	if (m->current_pgpath) {
1283 		bdev = m->current_pgpath->path.dev->bdev;
1284 		fake_dentry.d_inode = bdev->bd_inode;
1285 		fake_file.f_mode = m->current_pgpath->path.dev->mode;
1286 	}
1287 
1288 	if (m->queue_io)
1289 		r = -EAGAIN;
1290 	else if (!bdev)
1291 		r = -EIO;
1292 
1293 	spin_unlock_irqrestore(&m->lock, flags);
1294 
1295 	return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1296 					 bdev->bd_disk, cmd, arg);
1297 }
1298 
1299 /*-----------------------------------------------------------------
1300  * Module setup
1301  *---------------------------------------------------------------*/
1302 static struct target_type multipath_target = {
1303 	.name = "multipath",
1304 	.version = {1, 0, 5},
1305 	.module = THIS_MODULE,
1306 	.ctr = multipath_ctr,
1307 	.dtr = multipath_dtr,
1308 	.map = multipath_map,
1309 	.end_io = multipath_end_io,
1310 	.presuspend = multipath_presuspend,
1311 	.resume = multipath_resume,
1312 	.status = multipath_status,
1313 	.message = multipath_message,
1314 	.ioctl  = multipath_ioctl,
1315 };
1316 
1317 static int __init dm_multipath_init(void)
1318 {
1319 	int r;
1320 
1321 	/* allocate a slab for the dm_ios */
1322 	_mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1323 					0, 0, NULL, NULL);
1324 	if (!_mpio_cache)
1325 		return -ENOMEM;
1326 
1327 	r = dm_register_target(&multipath_target);
1328 	if (r < 0) {
1329 		DMERR("%s: register failed %d", multipath_target.name, r);
1330 		kmem_cache_destroy(_mpio_cache);
1331 		return -EINVAL;
1332 	}
1333 
1334 	kmultipathd = create_workqueue("kmpathd");
1335 	if (!kmultipathd) {
1336 		DMERR("%s: failed to create workqueue kmpathd",
1337 				multipath_target.name);
1338 		dm_unregister_target(&multipath_target);
1339 		kmem_cache_destroy(_mpio_cache);
1340 		return -ENOMEM;
1341 	}
1342 
1343 	DMINFO("version %u.%u.%u loaded",
1344 	       multipath_target.version[0], multipath_target.version[1],
1345 	       multipath_target.version[2]);
1346 
1347 	return r;
1348 }
1349 
1350 static void __exit dm_multipath_exit(void)
1351 {
1352 	int r;
1353 
1354 	destroy_workqueue(kmultipathd);
1355 
1356 	r = dm_unregister_target(&multipath_target);
1357 	if (r < 0)
1358 		DMERR("%s: target unregister failed %d",
1359 		      multipath_target.name, r);
1360 	kmem_cache_destroy(_mpio_cache);
1361 }
1362 
1363 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1364 
1365 module_init(dm_multipath_init);
1366 module_exit(dm_multipath_exit);
1367 
1368 MODULE_DESCRIPTION(DM_NAME " multipath target");
1369 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1370 MODULE_LICENSE("GPL");
1371