Merge master.kernel.org:/home/rmk/linux-2.6-arm
[linux-2.6] / drivers / md / dm-mpath.c
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_slab_pool(MIN_IOS, _mpio_cache);
183                 if (!m->mpio_pool) {
184                         kfree(m);
185                         return NULL;
186                 }
187         }
188
189         return m;
190 }
191
192 static void free_multipath(struct multipath *m)
193 {
194         struct priority_group *pg, *tmp;
195         struct hw_handler *hwh = &m->hw_handler;
196
197         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
198                 list_del(&pg->list);
199                 free_priority_group(pg, m->ti);
200         }
201
202         if (hwh->type) {
203                 hwh->type->destroy(hwh);
204                 dm_put_hw_handler(hwh->type);
205         }
206
207         mempool_destroy(m->mpio_pool);
208         kfree(m);
209 }
210
211
212 /*-----------------------------------------------
213  * Path selection
214  *-----------------------------------------------*/
215
216 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
217 {
218         struct hw_handler *hwh = &m->hw_handler;
219
220         m->current_pg = pgpath->pg;
221
222         /* Must we initialise the PG first, and queue I/O till it's ready? */
223         if (hwh->type && hwh->type->pg_init) {
224                 m->pg_init_required = 1;
225                 m->queue_io = 1;
226         } else {
227                 m->pg_init_required = 0;
228                 m->queue_io = 0;
229         }
230 }
231
232 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
233 {
234         struct path *path;
235
236         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
237         if (!path)
238                 return -ENXIO;
239
240         m->current_pgpath = path_to_pgpath(path);
241
242         if (m->current_pg != pg)
243                 __switch_pg(m, m->current_pgpath);
244
245         return 0;
246 }
247
248 static void __choose_pgpath(struct multipath *m)
249 {
250         struct priority_group *pg;
251         unsigned bypassed = 1;
252
253         if (!m->nr_valid_paths)
254                 goto failed;
255
256         /* Were we instructed to switch PG? */
257         if (m->next_pg) {
258                 pg = m->next_pg;
259                 m->next_pg = NULL;
260                 if (!__choose_path_in_pg(m, pg))
261                         return;
262         }
263
264         /* Don't change PG until it has no remaining paths */
265         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
266                 return;
267
268         /*
269          * Loop through priority groups until we find a valid path.
270          * First time we skip PGs marked 'bypassed'.
271          * Second time we only try the ones we skipped.
272          */
273         do {
274                 list_for_each_entry(pg, &m->priority_groups, list) {
275                         if (pg->bypassed == bypassed)
276                                 continue;
277                         if (!__choose_path_in_pg(m, pg))
278                                 return;
279                 }
280         } while (bypassed--);
281
282 failed:
283         m->current_pgpath = NULL;
284         m->current_pg = NULL;
285 }
286
287 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
288                   unsigned was_queued)
289 {
290         int r = 1;
291         unsigned long flags;
292         struct pgpath *pgpath;
293
294         spin_lock_irqsave(&m->lock, flags);
295
296         /* Do we need to select a new pgpath? */
297         if (!m->current_pgpath ||
298             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
299                 __choose_pgpath(m);
300
301         pgpath = m->current_pgpath;
302
303         if (was_queued)
304                 m->queue_size--;
305
306         if ((pgpath && m->queue_io) ||
307             (!pgpath && m->queue_if_no_path)) {
308                 /* Queue for the daemon to resubmit */
309                 bio_list_add(&m->queued_ios, bio);
310                 m->queue_size++;
311                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
312                     !m->queue_io)
313                         queue_work(kmultipathd, &m->process_queued_ios);
314                 pgpath = NULL;
315                 r = 0;
316         } else if (!pgpath)
317                 r = -EIO;               /* Failed */
318         else
319                 bio->bi_bdev = pgpath->path.dev->bdev;
320
321         mpio->pgpath = pgpath;
322
323         spin_unlock_irqrestore(&m->lock, flags);
324
325         return r;
326 }
327
328 /*
329  * If we run out of usable paths, should we queue I/O or error it?
330  */
331 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
332                             unsigned save_old_value)
333 {
334         unsigned long flags;
335
336         spin_lock_irqsave(&m->lock, flags);
337
338         if (save_old_value)
339                 m->saved_queue_if_no_path = m->queue_if_no_path;
340         else
341                 m->saved_queue_if_no_path = queue_if_no_path;
342         m->queue_if_no_path = queue_if_no_path;
343         if (!m->queue_if_no_path && m->queue_size)
344                 queue_work(kmultipathd, &m->process_queued_ios);
345
346         spin_unlock_irqrestore(&m->lock, flags);
347
348         return 0;
349 }
350
351 /*-----------------------------------------------------------------
352  * The multipath daemon is responsible for resubmitting queued ios.
353  *---------------------------------------------------------------*/
354
355 static void dispatch_queued_ios(struct multipath *m)
356 {
357         int r;
358         unsigned long flags;
359         struct bio *bio = NULL, *next;
360         struct mpath_io *mpio;
361         union map_info *info;
362
363         spin_lock_irqsave(&m->lock, flags);
364         bio = bio_list_get(&m->queued_ios);
365         spin_unlock_irqrestore(&m->lock, flags);
366
367         while (bio) {
368                 next = bio->bi_next;
369                 bio->bi_next = NULL;
370
371                 info = dm_get_mapinfo(bio);
372                 mpio = info->ptr;
373
374                 r = map_io(m, bio, mpio, 1);
375                 if (r < 0)
376                         bio_endio(bio, bio->bi_size, r);
377                 else if (r == 1)
378                         generic_make_request(bio);
379
380                 bio = next;
381         }
382 }
383
384 static void process_queued_ios(void *data)
385 {
386         struct multipath *m = (struct multipath *) data;
387         struct hw_handler *hwh = &m->hw_handler;
388         struct pgpath *pgpath = NULL;
389         unsigned init_required = 0, must_queue = 1;
390         unsigned long flags;
391
392         spin_lock_irqsave(&m->lock, flags);
393
394         if (!m->queue_size)
395                 goto out;
396
397         if (!m->current_pgpath)
398                 __choose_pgpath(m);
399
400         pgpath = m->current_pgpath;
401
402         if ((pgpath && !m->queue_io) ||
403             (!pgpath && !m->queue_if_no_path))
404                 must_queue = 0;
405
406         if (m->pg_init_required && !m->pg_init_in_progress) {
407                 m->pg_init_required = 0;
408                 m->pg_init_in_progress = 1;
409                 init_required = 1;
410         }
411
412 out:
413         spin_unlock_irqrestore(&m->lock, flags);
414
415         if (init_required)
416                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
417
418         if (!must_queue)
419                 dispatch_queued_ios(m);
420 }
421
422 /*
423  * An event is triggered whenever a path is taken out of use.
424  * Includes path failure and PG bypass.
425  */
426 static void trigger_event(void *data)
427 {
428         struct multipath *m = (struct multipath *) data;
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 #define ESTR(s) ("dm-multipath: " s)
450
451 static int read_param(struct param *param, char *str, unsigned *v, char **error)
452 {
453         if (!str ||
454             (sscanf(str, "%u", v) != 1) ||
455             (*v < param->min) ||
456             (*v > param->max)) {
457                 *error = param->error;
458                 return -EINVAL;
459         }
460
461         return 0;
462 }
463
464 struct arg_set {
465         unsigned argc;
466         char **argv;
467 };
468
469 static char *shift(struct arg_set *as)
470 {
471         char *r;
472
473         if (as->argc) {
474                 as->argc--;
475                 r = *as->argv;
476                 as->argv++;
477                 return r;
478         }
479
480         return NULL;
481 }
482
483 static void consume(struct arg_set *as, unsigned n)
484 {
485         BUG_ON (as->argc < n);
486         as->argc -= n;
487         as->argv += n;
488 }
489
490 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
491                                struct dm_target *ti)
492 {
493         int r;
494         struct path_selector_type *pst;
495         unsigned ps_argc;
496
497         static struct param _params[] = {
498                 {0, 1024, ESTR("invalid number of path selector args")},
499         };
500
501         pst = dm_get_path_selector(shift(as));
502         if (!pst) {
503                 ti->error = ESTR("unknown path selector type");
504                 return -EINVAL;
505         }
506
507         r = read_param(_params, shift(as), &ps_argc, &ti->error);
508         if (r)
509                 return -EINVAL;
510
511         r = pst->create(&pg->ps, ps_argc, as->argv);
512         if (r) {
513                 dm_put_path_selector(pst);
514                 ti->error = ESTR("path selector constructor failed");
515                 return r;
516         }
517
518         pg->ps.type = pst;
519         consume(as, ps_argc);
520
521         return 0;
522 }
523
524 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
525                                struct dm_target *ti)
526 {
527         int r;
528         struct pgpath *p;
529
530         /* we need at least a path arg */
531         if (as->argc < 1) {
532                 ti->error = ESTR("no device given");
533                 return NULL;
534         }
535
536         p = alloc_pgpath();
537         if (!p)
538                 return NULL;
539
540         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
541                           dm_table_get_mode(ti->table), &p->path.dev);
542         if (r) {
543                 ti->error = ESTR("error getting device");
544                 goto bad;
545         }
546
547         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
548         if (r) {
549                 dm_put_device(ti, p->path.dev);
550                 goto bad;
551         }
552
553         return p;
554
555  bad:
556         free_pgpath(p);
557         return NULL;
558 }
559
560 static struct priority_group *parse_priority_group(struct arg_set *as,
561                                                    struct multipath *m,
562                                                    struct dm_target *ti)
563 {
564         static struct param _params[] = {
565                 {1, 1024, ESTR("invalid number of paths")},
566                 {0, 1024, ESTR("invalid number of selector args")}
567         };
568
569         int r;
570         unsigned i, nr_selector_args, nr_params;
571         struct priority_group *pg;
572
573         if (as->argc < 2) {
574                 as->argc = 0;
575                 ti->error = ESTR("not enough priority group aruments");
576                 return NULL;
577         }
578
579         pg = alloc_priority_group();
580         if (!pg) {
581                 ti->error = ESTR("couldn't allocate priority group");
582                 return NULL;
583         }
584         pg->m = m;
585
586         r = parse_path_selector(as, pg, ti);
587         if (r)
588                 goto bad;
589
590         /*
591          * read the paths
592          */
593         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
594         if (r)
595                 goto bad;
596
597         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
598         if (r)
599                 goto bad;
600
601         nr_params = 1 + nr_selector_args;
602         for (i = 0; i < pg->nr_pgpaths; i++) {
603                 struct pgpath *pgpath;
604                 struct arg_set path_args;
605
606                 if (as->argc < nr_params)
607                         goto bad;
608
609                 path_args.argc = nr_params;
610                 path_args.argv = as->argv;
611
612                 pgpath = parse_path(&path_args, &pg->ps, ti);
613                 if (!pgpath)
614                         goto bad;
615
616                 pgpath->pg = pg;
617                 list_add_tail(&pgpath->list, &pg->pgpaths);
618                 consume(as, nr_params);
619         }
620
621         return pg;
622
623  bad:
624         free_priority_group(pg, ti);
625         return NULL;
626 }
627
628 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
629                             struct dm_target *ti)
630 {
631         int r;
632         struct hw_handler_type *hwht;
633         unsigned hw_argc;
634
635         static struct param _params[] = {
636                 {0, 1024, ESTR("invalid number of hardware handler args")},
637         };
638
639         r = read_param(_params, shift(as), &hw_argc, &ti->error);
640         if (r)
641                 return -EINVAL;
642
643         if (!hw_argc)
644                 return 0;
645
646         hwht = dm_get_hw_handler(shift(as));
647         if (!hwht) {
648                 ti->error = ESTR("unknown hardware handler type");
649                 return -EINVAL;
650         }
651
652         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
653         if (r) {
654                 dm_put_hw_handler(hwht);
655                 ti->error = ESTR("hardware handler constructor failed");
656                 return r;
657         }
658
659         m->hw_handler.type = hwht;
660         consume(as, hw_argc - 1);
661
662         return 0;
663 }
664
665 static int parse_features(struct arg_set *as, struct multipath *m,
666                           struct dm_target *ti)
667 {
668         int r;
669         unsigned argc;
670
671         static struct param _params[] = {
672                 {0, 1, ESTR("invalid number of feature args")},
673         };
674
675         r = read_param(_params, shift(as), &argc, &ti->error);
676         if (r)
677                 return -EINVAL;
678
679         if (!argc)
680                 return 0;
681
682         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
683                 return queue_if_no_path(m, 1, 0);
684         else {
685                 ti->error = "Unrecognised multipath feature request";
686                 return -EINVAL;
687         }
688 }
689
690 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
691                          char **argv)
692 {
693         /* target parameters */
694         static struct param _params[] = {
695                 {1, 1024, ESTR("invalid number of priority groups")},
696                 {1, 1024, ESTR("invalid initial priority group number")},
697         };
698
699         int r;
700         struct multipath *m;
701         struct arg_set as;
702         unsigned pg_count = 0;
703         unsigned next_pg_num;
704
705         as.argc = argc;
706         as.argv = argv;
707
708         m = alloc_multipath();
709         if (!m) {
710                 ti->error = ESTR("can't allocate multipath");
711                 return -EINVAL;
712         }
713
714         r = parse_features(&as, m, ti);
715         if (r)
716                 goto bad;
717
718         r = parse_hw_handler(&as, m, ti);
719         if (r)
720                 goto bad;
721
722         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
723         if (r)
724                 goto bad;
725
726         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
727         if (r)
728                 goto bad;
729
730         /* parse the priority groups */
731         while (as.argc) {
732                 struct priority_group *pg;
733
734                 pg = parse_priority_group(&as, m, ti);
735                 if (!pg) {
736                         r = -EINVAL;
737                         goto bad;
738                 }
739
740                 m->nr_valid_paths += pg->nr_pgpaths;
741                 list_add_tail(&pg->list, &m->priority_groups);
742                 pg_count++;
743                 pg->pg_num = pg_count;
744                 if (!--next_pg_num)
745                         m->next_pg = pg;
746         }
747
748         if (pg_count != m->nr_priority_groups) {
749                 ti->error = ESTR("priority group count mismatch");
750                 r = -EINVAL;
751                 goto bad;
752         }
753
754         ti->private = m;
755         m->ti = ti;
756
757         return 0;
758
759  bad:
760         free_multipath(m);
761         return r;
762 }
763
764 static void multipath_dtr(struct dm_target *ti)
765 {
766         struct multipath *m = (struct multipath *) ti->private;
767
768         flush_workqueue(kmultipathd);
769         free_multipath(m);
770 }
771
772 /*
773  * Map bios, recording original fields for later in case we have to resubmit
774  */
775 static int multipath_map(struct dm_target *ti, struct bio *bio,
776                          union map_info *map_context)
777 {
778         int r;
779         struct mpath_io *mpio;
780         struct multipath *m = (struct multipath *) ti->private;
781
782         if (bio_barrier(bio))
783                 return -EOPNOTSUPP;
784
785         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
786         dm_bio_record(&mpio->details, bio);
787
788         map_context->ptr = mpio;
789         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
790         r = map_io(m, bio, mpio, 0);
791         if (r < 0)
792                 mempool_free(mpio, m->mpio_pool);
793
794         return r;
795 }
796
797 /*
798  * Take a path out of use.
799  */
800 static int fail_path(struct pgpath *pgpath)
801 {
802         unsigned long flags;
803         struct multipath *m = pgpath->pg->m;
804
805         spin_lock_irqsave(&m->lock, flags);
806
807         if (!pgpath->path.is_active)
808                 goto out;
809
810         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
811
812         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
813         pgpath->path.is_active = 0;
814         pgpath->fail_count++;
815
816         m->nr_valid_paths--;
817
818         if (pgpath == m->current_pgpath)
819                 m->current_pgpath = NULL;
820
821         queue_work(kmultipathd, &m->trigger_event);
822
823 out:
824         spin_unlock_irqrestore(&m->lock, flags);
825
826         return 0;
827 }
828
829 /*
830  * Reinstate a previously-failed path
831  */
832 static int reinstate_path(struct pgpath *pgpath)
833 {
834         int r = 0;
835         unsigned long flags;
836         struct multipath *m = pgpath->pg->m;
837
838         spin_lock_irqsave(&m->lock, flags);
839
840         if (pgpath->path.is_active)
841                 goto out;
842
843         if (!pgpath->pg->ps.type) {
844                 DMWARN("Reinstate path not supported by path selector %s",
845                        pgpath->pg->ps.type->name);
846                 r = -EINVAL;
847                 goto out;
848         }
849
850         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
851         if (r)
852                 goto out;
853
854         pgpath->path.is_active = 1;
855
856         m->current_pgpath = NULL;
857         if (!m->nr_valid_paths++ && m->queue_size)
858                 queue_work(kmultipathd, &m->process_queued_ios);
859
860         queue_work(kmultipathd, &m->trigger_event);
861
862 out:
863         spin_unlock_irqrestore(&m->lock, flags);
864
865         return r;
866 }
867
868 /*
869  * Fail or reinstate all paths that match the provided struct dm_dev.
870  */
871 static int action_dev(struct multipath *m, struct dm_dev *dev,
872                       action_fn action)
873 {
874         int r = 0;
875         struct pgpath *pgpath;
876         struct priority_group *pg;
877
878         list_for_each_entry(pg, &m->priority_groups, list) {
879                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
880                         if (pgpath->path.dev == dev)
881                                 r = action(pgpath);
882                 }
883         }
884
885         return r;
886 }
887
888 /*
889  * Temporarily try to avoid having to use the specified PG
890  */
891 static void bypass_pg(struct multipath *m, struct priority_group *pg,
892                       int bypassed)
893 {
894         unsigned long flags;
895
896         spin_lock_irqsave(&m->lock, flags);
897
898         pg->bypassed = bypassed;
899         m->current_pgpath = NULL;
900         m->current_pg = NULL;
901
902         spin_unlock_irqrestore(&m->lock, flags);
903
904         queue_work(kmultipathd, &m->trigger_event);
905 }
906
907 /*
908  * Switch to using the specified PG from the next I/O that gets mapped
909  */
910 static int switch_pg_num(struct multipath *m, const char *pgstr)
911 {
912         struct priority_group *pg;
913         unsigned pgnum;
914         unsigned long flags;
915
916         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
917             (pgnum > m->nr_priority_groups)) {
918                 DMWARN("invalid PG number supplied to switch_pg_num");
919                 return -EINVAL;
920         }
921
922         spin_lock_irqsave(&m->lock, flags);
923         list_for_each_entry(pg, &m->priority_groups, list) {
924                 pg->bypassed = 0;
925                 if (--pgnum)
926                         continue;
927
928                 m->current_pgpath = NULL;
929                 m->current_pg = NULL;
930                 m->next_pg = pg;
931         }
932         spin_unlock_irqrestore(&m->lock, flags);
933
934         queue_work(kmultipathd, &m->trigger_event);
935         return 0;
936 }
937
938 /*
939  * Set/clear bypassed status of a PG.
940  * PGs are numbered upwards from 1 in the order they were declared.
941  */
942 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
943 {
944         struct priority_group *pg;
945         unsigned pgnum;
946
947         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
948             (pgnum > m->nr_priority_groups)) {
949                 DMWARN("invalid PG number supplied to bypass_pg");
950                 return -EINVAL;
951         }
952
953         list_for_each_entry(pg, &m->priority_groups, list) {
954                 if (!--pgnum)
955                         break;
956         }
957
958         bypass_pg(m, pg, bypassed);
959         return 0;
960 }
961
962 /*
963  * pg_init must call this when it has completed its initialisation
964  */
965 void dm_pg_init_complete(struct path *path, unsigned err_flags)
966 {
967         struct pgpath *pgpath = path_to_pgpath(path);
968         struct priority_group *pg = pgpath->pg;
969         struct multipath *m = pg->m;
970         unsigned long flags;
971
972         /* We insist on failing the path if the PG is already bypassed. */
973         if (err_flags && pg->bypassed)
974                 err_flags |= MP_FAIL_PATH;
975
976         if (err_flags & MP_FAIL_PATH)
977                 fail_path(pgpath);
978
979         if (err_flags & MP_BYPASS_PG)
980                 bypass_pg(m, pg, 1);
981
982         spin_lock_irqsave(&m->lock, flags);
983         if (err_flags) {
984                 m->current_pgpath = NULL;
985                 m->current_pg = NULL;
986         } else if (!m->pg_init_required)
987                 m->queue_io = 0;
988
989         m->pg_init_in_progress = 0;
990         queue_work(kmultipathd, &m->process_queued_ios);
991         spin_unlock_irqrestore(&m->lock, flags);
992 }
993
994 /*
995  * end_io handling
996  */
997 static int do_end_io(struct multipath *m, struct bio *bio,
998                      int error, struct mpath_io *mpio)
999 {
1000         struct hw_handler *hwh = &m->hw_handler;
1001         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1002         unsigned long flags;
1003
1004         if (!error)
1005                 return 0;       /* I/O complete */
1006
1007         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1008                 return error;
1009
1010         if (error == -EOPNOTSUPP)
1011                 return error;
1012
1013         spin_lock_irqsave(&m->lock, flags);
1014         if (!m->nr_valid_paths) {
1015                 if (!m->queue_if_no_path) {
1016                         spin_unlock_irqrestore(&m->lock, flags);
1017                         return -EIO;
1018                 } else {
1019                         spin_unlock_irqrestore(&m->lock, flags);
1020                         goto requeue;
1021                 }
1022         }
1023         spin_unlock_irqrestore(&m->lock, flags);
1024
1025         if (hwh->type && hwh->type->error)
1026                 err_flags = hwh->type->error(hwh, bio);
1027
1028         if (mpio->pgpath) {
1029                 if (err_flags & MP_FAIL_PATH)
1030                         fail_path(mpio->pgpath);
1031
1032                 if (err_flags & MP_BYPASS_PG)
1033                         bypass_pg(m, mpio->pgpath->pg, 1);
1034         }
1035
1036         if (err_flags & MP_ERROR_IO)
1037                 return -EIO;
1038
1039       requeue:
1040         dm_bio_restore(&mpio->details, bio);
1041
1042         /* queue for the daemon to resubmit or fail */
1043         spin_lock_irqsave(&m->lock, flags);
1044         bio_list_add(&m->queued_ios, bio);
1045         m->queue_size++;
1046         if (!m->queue_io)
1047                 queue_work(kmultipathd, &m->process_queued_ios);
1048         spin_unlock_irqrestore(&m->lock, flags);
1049
1050         return 1;       /* io not complete */
1051 }
1052
1053 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1054                             int error, union map_info *map_context)
1055 {
1056         struct multipath *m = (struct multipath *) ti->private;
1057         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1058         struct pgpath *pgpath = mpio->pgpath;
1059         struct path_selector *ps;
1060         int r;
1061
1062         r  = do_end_io(m, bio, error, mpio);
1063         if (pgpath) {
1064                 ps = &pgpath->pg->ps;
1065                 if (ps->type->end_io)
1066                         ps->type->end_io(ps, &pgpath->path);
1067         }
1068         if (r <= 0)
1069                 mempool_free(mpio, m->mpio_pool);
1070
1071         return r;
1072 }
1073
1074 /*
1075  * Suspend can't complete until all the I/O is processed so if
1076  * the last path fails we must error any remaining I/O.
1077  * Note that if the freeze_bdev fails while suspending, the
1078  * queue_if_no_path state is lost - userspace should reset it.
1079  */
1080 static void multipath_presuspend(struct dm_target *ti)
1081 {
1082         struct multipath *m = (struct multipath *) ti->private;
1083
1084         queue_if_no_path(m, 0, 1);
1085 }
1086
1087 /*
1088  * Restore the queue_if_no_path setting.
1089  */
1090 static void multipath_resume(struct dm_target *ti)
1091 {
1092         struct multipath *m = (struct multipath *) ti->private;
1093         unsigned long flags;
1094
1095         spin_lock_irqsave(&m->lock, flags);
1096         m->queue_if_no_path = m->saved_queue_if_no_path;
1097         spin_unlock_irqrestore(&m->lock, flags);
1098 }
1099
1100 /*
1101  * Info output has the following format:
1102  * num_multipath_feature_args [multipath_feature_args]*
1103  * num_handler_status_args [handler_status_args]*
1104  * num_groups init_group_number
1105  *            [A|D|E num_ps_status_args [ps_status_args]*
1106  *             num_paths num_selector_args
1107  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1108  *
1109  * Table output has the following format (identical to the constructor string):
1110  * num_feature_args [features_args]*
1111  * num_handler_args hw_handler [hw_handler_args]*
1112  * num_groups init_group_number
1113  *     [priority selector-name num_ps_args [ps_args]*
1114  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1115  */
1116 static int multipath_status(struct dm_target *ti, status_type_t type,
1117                             char *result, unsigned int maxlen)
1118 {
1119         int sz = 0;
1120         unsigned long flags;
1121         struct multipath *m = (struct multipath *) ti->private;
1122         struct hw_handler *hwh = &m->hw_handler;
1123         struct priority_group *pg;
1124         struct pgpath *p;
1125         unsigned pg_num;
1126         char state;
1127
1128         spin_lock_irqsave(&m->lock, flags);
1129
1130         /* Features */
1131         if (type == STATUSTYPE_INFO)
1132                 DMEMIT("1 %u ", m->queue_size);
1133         else if (m->queue_if_no_path)
1134                 DMEMIT("1 queue_if_no_path ");
1135         else
1136                 DMEMIT("0 ");
1137
1138         if (hwh->type && hwh->type->status)
1139                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1140         else if (!hwh->type || type == STATUSTYPE_INFO)
1141                 DMEMIT("0 ");
1142         else
1143                 DMEMIT("1 %s ", hwh->type->name);
1144
1145         DMEMIT("%u ", m->nr_priority_groups);
1146
1147         if (m->next_pg)
1148                 pg_num = m->next_pg->pg_num;
1149         else if (m->current_pg)
1150                 pg_num = m->current_pg->pg_num;
1151         else
1152                         pg_num = 1;
1153
1154         DMEMIT("%u ", pg_num);
1155
1156         switch (type) {
1157         case STATUSTYPE_INFO:
1158                 list_for_each_entry(pg, &m->priority_groups, list) {
1159                         if (pg->bypassed)
1160                                 state = 'D';    /* Disabled */
1161                         else if (pg == m->current_pg)
1162                                 state = 'A';    /* Currently Active */
1163                         else
1164                                 state = 'E';    /* Enabled */
1165
1166                         DMEMIT("%c ", state);
1167
1168                         if (pg->ps.type->status)
1169                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1170                                                           result + sz,
1171                                                           maxlen - sz);
1172                         else
1173                                 DMEMIT("0 ");
1174
1175                         DMEMIT("%u %u ", pg->nr_pgpaths,
1176                                pg->ps.type->info_args);
1177
1178                         list_for_each_entry(p, &pg->pgpaths, list) {
1179                                 DMEMIT("%s %s %u ", p->path.dev->name,
1180                                        p->path.is_active ? "A" : "F",
1181                                        p->fail_count);
1182                                 if (pg->ps.type->status)
1183                                         sz += pg->ps.type->status(&pg->ps,
1184                                               &p->path, type, result + sz,
1185                                               maxlen - sz);
1186                         }
1187                 }
1188                 break;
1189
1190         case STATUSTYPE_TABLE:
1191                 list_for_each_entry(pg, &m->priority_groups, list) {
1192                         DMEMIT("%s ", pg->ps.type->name);
1193
1194                         if (pg->ps.type->status)
1195                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1196                                                           result + sz,
1197                                                           maxlen - sz);
1198                         else
1199                                 DMEMIT("0 ");
1200
1201                         DMEMIT("%u %u ", pg->nr_pgpaths,
1202                                pg->ps.type->table_args);
1203
1204                         list_for_each_entry(p, &pg->pgpaths, list) {
1205                                 DMEMIT("%s ", p->path.dev->name);
1206                                 if (pg->ps.type->status)
1207                                         sz += pg->ps.type->status(&pg->ps,
1208                                               &p->path, type, result + sz,
1209                                               maxlen - sz);
1210                         }
1211                 }
1212                 break;
1213         }
1214
1215         spin_unlock_irqrestore(&m->lock, flags);
1216
1217         return 0;
1218 }
1219
1220 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1221 {
1222         int r;
1223         struct dm_dev *dev;
1224         struct multipath *m = (struct multipath *) ti->private;
1225         action_fn action;
1226
1227         if (argc == 1) {
1228                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1229                         return queue_if_no_path(m, 1, 0);
1230                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1231                         return queue_if_no_path(m, 0, 0);
1232         }
1233
1234         if (argc != 2)
1235                 goto error;
1236
1237         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1238                 return bypass_pg_num(m, argv[1], 1);
1239         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1240                 return bypass_pg_num(m, argv[1], 0);
1241         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1242                 return switch_pg_num(m, argv[1]);
1243         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1244                 action = reinstate_path;
1245         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1246                 action = fail_path;
1247         else
1248                 goto error;
1249
1250         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1251                           dm_table_get_mode(ti->table), &dev);
1252         if (r) {
1253                 DMWARN("dm-multipath message: error getting device %s",
1254                        argv[1]);
1255                 return -EINVAL;
1256         }
1257
1258         r = action_dev(m, dev, action);
1259
1260         dm_put_device(ti, dev);
1261
1262         return r;
1263
1264 error:
1265         DMWARN("Unrecognised multipath message received.");
1266         return -EINVAL;
1267 }
1268
1269 /*-----------------------------------------------------------------
1270  * Module setup
1271  *---------------------------------------------------------------*/
1272 static struct target_type multipath_target = {
1273         .name = "multipath",
1274         .version = {1, 0, 4},
1275         .module = THIS_MODULE,
1276         .ctr = multipath_ctr,
1277         .dtr = multipath_dtr,
1278         .map = multipath_map,
1279         .end_io = multipath_end_io,
1280         .presuspend = multipath_presuspend,
1281         .resume = multipath_resume,
1282         .status = multipath_status,
1283         .message = multipath_message,
1284 };
1285
1286 static int __init dm_multipath_init(void)
1287 {
1288         int r;
1289
1290         /* allocate a slab for the dm_ios */
1291         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1292                                         0, 0, NULL, NULL);
1293         if (!_mpio_cache)
1294                 return -ENOMEM;
1295
1296         r = dm_register_target(&multipath_target);
1297         if (r < 0) {
1298                 DMERR("%s: register failed %d", multipath_target.name, r);
1299                 kmem_cache_destroy(_mpio_cache);
1300                 return -EINVAL;
1301         }
1302
1303         kmultipathd = create_workqueue("kmpathd");
1304         if (!kmultipathd) {
1305                 DMERR("%s: failed to create workqueue kmpathd",
1306                                 multipath_target.name);
1307                 dm_unregister_target(&multipath_target);
1308                 kmem_cache_destroy(_mpio_cache);
1309                 return -ENOMEM;
1310         }
1311
1312         DMINFO("dm-multipath version %u.%u.%u loaded",
1313                multipath_target.version[0], multipath_target.version[1],
1314                multipath_target.version[2]);
1315
1316         return r;
1317 }
1318
1319 static void __exit dm_multipath_exit(void)
1320 {
1321         int r;
1322
1323         destroy_workqueue(kmultipathd);
1324
1325         r = dm_unregister_target(&multipath_target);
1326         if (r < 0)
1327                 DMERR("%s: target unregister failed %d",
1328                       multipath_target.name, r);
1329         kmem_cache_destroy(_mpio_cache);
1330 }
1331
1332 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1333
1334 module_init(dm_multipath_init);
1335 module_exit(dm_multipath_exit);
1336
1337 MODULE_DESCRIPTION(DM_NAME " multipath target");
1338 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1339 MODULE_LICENSE("GPL");